1 ================================= 1 =================================
2 A Tour Through RCU's Requirements 2 A Tour Through RCU's Requirements
3 ================================= 3 =================================
4 4
5 Copyright IBM Corporation, 2015 5 Copyright IBM Corporation, 2015
6 6
7 Author: Paul E. McKenney 7 Author: Paul E. McKenney
8 8
9 The initial version of this document appeared 9 The initial version of this document appeared in the
10 `LWN <https://lwn.net/>`_ on those articles: 10 `LWN <https://lwn.net/>`_ on those articles:
11 `part 1 <https://lwn.net/Articles/652156/>`_, 11 `part 1 <https://lwn.net/Articles/652156/>`_,
12 `part 2 <https://lwn.net/Articles/652677/>`_, 12 `part 2 <https://lwn.net/Articles/652677/>`_, and
13 `part 3 <https://lwn.net/Articles/653326/>`_. 13 `part 3 <https://lwn.net/Articles/653326/>`_.
14 14
15 Introduction 15 Introduction
16 ------------ 16 ------------
17 17
18 Read-copy update (RCU) is a synchronization me 18 Read-copy update (RCU) is a synchronization mechanism that is often used
19 as a replacement for reader-writer locking. RC 19 as a replacement for reader-writer locking. RCU is unusual in that
20 updaters do not block readers, which means tha 20 updaters do not block readers, which means that RCU's read-side
21 primitives can be exceedingly fast and scalabl 21 primitives can be exceedingly fast and scalable. In addition, updaters
22 can make useful forward progress concurrently 22 can make useful forward progress concurrently with readers. However, all
23 this concurrency between RCU readers and updat 23 this concurrency between RCU readers and updaters does raise the
24 question of exactly what RCU readers are doing 24 question of exactly what RCU readers are doing, which in turn raises the
25 question of exactly what RCU's requirements ar 25 question of exactly what RCU's requirements are.
26 26
27 This document therefore summarizes RCU's requi 27 This document therefore summarizes RCU's requirements, and can be
28 thought of as an informal, high-level specific 28 thought of as an informal, high-level specification for RCU. It is
29 important to understand that RCU's specificati 29 important to understand that RCU's specification is primarily empirical
30 in nature; in fact, I learned about many of th 30 in nature; in fact, I learned about many of these requirements the hard
31 way. This situation might cause some consterna 31 way. This situation might cause some consternation, however, not only
32 has this learning process been a lot of fun, b 32 has this learning process been a lot of fun, but it has also been a
33 great privilege to work with so many people wi 33 great privilege to work with so many people willing to apply
34 technologies in interesting new ways. 34 technologies in interesting new ways.
35 35
36 All that aside, here are the categories of cur 36 All that aside, here are the categories of currently known RCU
37 requirements: 37 requirements:
38 38
39 #. `Fundamental Requirements`_ 39 #. `Fundamental Requirements`_
40 #. `Fundamental Non-Requirements`_ 40 #. `Fundamental Non-Requirements`_
41 #. `Parallelism Facts of Life`_ 41 #. `Parallelism Facts of Life`_
42 #. `Quality-of-Implementation Requirements`_ 42 #. `Quality-of-Implementation Requirements`_
43 #. `Linux Kernel Complications`_ 43 #. `Linux Kernel Complications`_
44 #. `Software-Engineering Requirements`_ 44 #. `Software-Engineering Requirements`_
45 #. `Other RCU Flavors`_ 45 #. `Other RCU Flavors`_
46 #. `Possible Future Changes`_ 46 #. `Possible Future Changes`_
47 47
48 This is followed by a summary_, however, the a 48 This is followed by a summary_, however, the answers to
49 each quick quiz immediately follows the quiz. 49 each quick quiz immediately follows the quiz. Select the big white space
50 with your mouse to see the answer. 50 with your mouse to see the answer.
51 51
52 Fundamental Requirements 52 Fundamental Requirements
53 ------------------------ 53 ------------------------
54 54
55 RCU's fundamental requirements are the closest 55 RCU's fundamental requirements are the closest thing RCU has to hard
56 mathematical requirements. These are: 56 mathematical requirements. These are:
57 57
58 #. `Grace-Period Guarantee`_ 58 #. `Grace-Period Guarantee`_
59 #. `Publish/Subscribe Guarantee`_ 59 #. `Publish/Subscribe Guarantee`_
60 #. `Memory-Barrier Guarantees`_ 60 #. `Memory-Barrier Guarantees`_
61 #. `RCU Primitives Guaranteed to Execute Uncon 61 #. `RCU Primitives Guaranteed to Execute Unconditionally`_
62 #. `Guaranteed Read-to-Write Upgrade`_ 62 #. `Guaranteed Read-to-Write Upgrade`_
63 63
64 Grace-Period Guarantee 64 Grace-Period Guarantee
65 ~~~~~~~~~~~~~~~~~~~~~~ 65 ~~~~~~~~~~~~~~~~~~~~~~
66 66
67 RCU's grace-period guarantee is unusual in bei 67 RCU's grace-period guarantee is unusual in being premeditated: Jack
68 Slingwine and I had this guarantee firmly in m 68 Slingwine and I had this guarantee firmly in mind when we started work
69 on RCU (then called “rclock”) in the early 69 on RCU (then called “rclock”) in the early 1990s. That said, the past
70 two decades of experience with RCU have produc 70 two decades of experience with RCU have produced a much more detailed
71 understanding of this guarantee. 71 understanding of this guarantee.
72 72
73 RCU's grace-period guarantee allows updaters t 73 RCU's grace-period guarantee allows updaters to wait for the completion
74 of all pre-existing RCU read-side critical sec 74 of all pre-existing RCU read-side critical sections. An RCU read-side
75 critical section begins with the marker rcu_re 75 critical section begins with the marker rcu_read_lock() and ends
76 with the marker rcu_read_unlock(). These marke 76 with the marker rcu_read_unlock(). These markers may be nested, and
77 RCU treats a nested set as one big RCU read-si 77 RCU treats a nested set as one big RCU read-side critical section.
78 Production-quality implementations of rcu_read 78 Production-quality implementations of rcu_read_lock() and
79 rcu_read_unlock() are extremely lightweight, a 79 rcu_read_unlock() are extremely lightweight, and in fact have
80 exactly zero overhead in Linux kernels built f 80 exactly zero overhead in Linux kernels built for production use with
81 ``CONFIG_PREEMPTION=n``. 81 ``CONFIG_PREEMPTION=n``.
82 82
83 This guarantee allows ordering to be enforced 83 This guarantee allows ordering to be enforced with extremely low
84 overhead to readers, for example: 84 overhead to readers, for example:
85 85
86 :: 86 ::
87 87
88 1 int x, y; 88 1 int x, y;
89 2 89 2
90 3 void thread0(void) 90 3 void thread0(void)
91 4 { 91 4 {
92 5 rcu_read_lock(); 92 5 rcu_read_lock();
93 6 r1 = READ_ONCE(x); 93 6 r1 = READ_ONCE(x);
94 7 r2 = READ_ONCE(y); 94 7 r2 = READ_ONCE(y);
95 8 rcu_read_unlock(); 95 8 rcu_read_unlock();
96 9 } 96 9 }
97 10 97 10
98 11 void thread1(void) 98 11 void thread1(void)
99 12 { 99 12 {
100 13 WRITE_ONCE(x, 1); 100 13 WRITE_ONCE(x, 1);
101 14 synchronize_rcu(); 101 14 synchronize_rcu();
102 15 WRITE_ONCE(y, 1); 102 15 WRITE_ONCE(y, 1);
103 16 } 103 16 }
104 104
105 Because the synchronize_rcu() on line 14 wait 105 Because the synchronize_rcu() on line 14 waits for all pre-existing
106 readers, any instance of thread0() that loads 106 readers, any instance of thread0() that loads a value of zero from
107 ``x`` must complete before thread1() stores to 107 ``x`` must complete before thread1() stores to ``y``, so that
108 instance must also load a value of zero from ` 108 instance must also load a value of zero from ``y``. Similarly, any
109 instance of thread0() that loads a value of on 109 instance of thread0() that loads a value of one from ``y`` must have
110 started after the synchronize_rcu() started, a 110 started after the synchronize_rcu() started, and must therefore also
111 load a value of one from ``x``. Therefore, the 111 load a value of one from ``x``. Therefore, the outcome:
112 112
113 :: 113 ::
114 114
115 (r1 == 0 && r2 == 1) 115 (r1 == 0 && r2 == 1)
116 116
117 cannot happen. 117 cannot happen.
118 118
119 +--------------------------------------------- 119 +-----------------------------------------------------------------------+
120 | **Quick Quiz**: 120 | **Quick Quiz**: |
121 +--------------------------------------------- 121 +-----------------------------------------------------------------------+
122 | Wait a minute! You said that updaters can ma 122 | Wait a minute! You said that updaters can make useful forward |
123 | progress concurrently with readers, but pre- 123 | progress concurrently with readers, but pre-existing readers will |
124 | block synchronize_rcu()!!! 124 | block synchronize_rcu()!!! |
125 | Just who are you trying to fool??? 125 | Just who are you trying to fool??? |
126 +--------------------------------------------- 126 +-----------------------------------------------------------------------+
127 | **Answer**: 127 | **Answer**: |
128 +--------------------------------------------- 128 +-----------------------------------------------------------------------+
129 | First, if updaters do not wish to be blocked 129 | First, if updaters do not wish to be blocked by readers, they can use |
130 | call_rcu() or kfree_rcu(), which will be dis 130 | call_rcu() or kfree_rcu(), which will be discussed later. |
131 | Second, even when using synchronize_rcu(), t 131 | Second, even when using synchronize_rcu(), the other update-side |
132 | code does run concurrently with readers, whe 132 | code does run concurrently with readers, whether pre-existing or not. |
133 +--------------------------------------------- 133 +-----------------------------------------------------------------------+
134 134
135 This scenario resembles one of the first uses 135 This scenario resembles one of the first uses of RCU in
136 `DYNIX/ptx <https://en.wikipedia.org/wiki/DYNI 136 `DYNIX/ptx <https://en.wikipedia.org/wiki/DYNIX>`__, which managed a
137 distributed lock manager's transition into a s 137 distributed lock manager's transition into a state suitable for handling
138 recovery from node failure, more or less as fo 138 recovery from node failure, more or less as follows:
139 139
140 :: 140 ::
141 141
142 1 #define STATE_NORMAL 0 142 1 #define STATE_NORMAL 0
143 2 #define STATE_WANT_RECOVERY 1 143 2 #define STATE_WANT_RECOVERY 1
144 3 #define STATE_RECOVERING 2 144 3 #define STATE_RECOVERING 2
145 4 #define STATE_WANT_NORMAL 3 145 4 #define STATE_WANT_NORMAL 3
146 5 146 5
147 6 int state = STATE_NORMAL; 147 6 int state = STATE_NORMAL;
148 7 148 7
149 8 void do_something_dlm(void) 149 8 void do_something_dlm(void)
150 9 { 150 9 {
151 10 int state_snap; 151 10 int state_snap;
152 11 152 11
153 12 rcu_read_lock(); 153 12 rcu_read_lock();
154 13 state_snap = READ_ONCE(state); 154 13 state_snap = READ_ONCE(state);
155 14 if (state_snap == STATE_NORMAL) 155 14 if (state_snap == STATE_NORMAL)
156 15 do_something(); 156 15 do_something();
157 16 else 157 16 else
158 17 do_something_carefully(); 158 17 do_something_carefully();
159 18 rcu_read_unlock(); 159 18 rcu_read_unlock();
160 19 } 160 19 }
161 20 161 20
162 21 void start_recovery(void) 162 21 void start_recovery(void)
163 22 { 163 22 {
164 23 WRITE_ONCE(state, STATE_WANT_RECOVE 164 23 WRITE_ONCE(state, STATE_WANT_RECOVERY);
165 24 synchronize_rcu(); 165 24 synchronize_rcu();
166 25 WRITE_ONCE(state, STATE_RECOVERING) 166 25 WRITE_ONCE(state, STATE_RECOVERING);
167 26 recovery(); 167 26 recovery();
168 27 WRITE_ONCE(state, STATE_WANT_NORMAL 168 27 WRITE_ONCE(state, STATE_WANT_NORMAL);
169 28 synchronize_rcu(); 169 28 synchronize_rcu();
170 29 WRITE_ONCE(state, STATE_NORMAL); 170 29 WRITE_ONCE(state, STATE_NORMAL);
171 30 } 171 30 }
172 172
173 The RCU read-side critical section in do_somet 173 The RCU read-side critical section in do_something_dlm() works with
174 the synchronize_rcu() in start_recovery() to g 174 the synchronize_rcu() in start_recovery() to guarantee that
175 do_something() never runs concurrently with re 175 do_something() never runs concurrently with recovery(), but with
176 little or no synchronization overhead in do_so 176 little or no synchronization overhead in do_something_dlm().
177 177
178 +--------------------------------------------- 178 +-----------------------------------------------------------------------+
179 | **Quick Quiz**: 179 | **Quick Quiz**: |
180 +--------------------------------------------- 180 +-----------------------------------------------------------------------+
181 | Why is the synchronize_rcu() on line 28 nee 181 | Why is the synchronize_rcu() on line 28 needed? |
182 +--------------------------------------------- 182 +-----------------------------------------------------------------------+
183 | **Answer**: 183 | **Answer**: |
184 +--------------------------------------------- 184 +-----------------------------------------------------------------------+
185 | Without that extra grace period, memory reor 185 | Without that extra grace period, memory reordering could result in |
186 | do_something_dlm() executing do_something() 186 | do_something_dlm() executing do_something() concurrently with |
187 | the last bits of recovery(). 187 | the last bits of recovery(). |
188 +--------------------------------------------- 188 +-----------------------------------------------------------------------+
189 189
190 In order to avoid fatal problems such as deadl 190 In order to avoid fatal problems such as deadlocks, an RCU read-side
191 critical section must not contain calls to syn 191 critical section must not contain calls to synchronize_rcu().
192 Similarly, an RCU read-side critical section m 192 Similarly, an RCU read-side critical section must not contain anything
193 that waits, directly or indirectly, on complet 193 that waits, directly or indirectly, on completion of an invocation of
194 synchronize_rcu(). 194 synchronize_rcu().
195 195
196 Although RCU's grace-period guarantee is usefu 196 Although RCU's grace-period guarantee is useful in and of itself, with
197 `quite a few use cases <https://lwn.net/Articl 197 `quite a few use cases <https://lwn.net/Articles/573497/>`__, it would
198 be good to be able to use RCU to coordinate re 198 be good to be able to use RCU to coordinate read-side access to linked
199 data structures. For this, the grace-period gu 199 data structures. For this, the grace-period guarantee is not sufficient,
200 as can be seen in function add_gp_buggy() belo 200 as can be seen in function add_gp_buggy() below. We will look at the
201 reader's code later, but in the meantime, just 201 reader's code later, but in the meantime, just think of the reader as
202 locklessly picking up the ``gp`` pointer, and, 202 locklessly picking up the ``gp`` pointer, and, if the value loaded is
203 non-\ ``NULL``, locklessly accessing the ``->a 203 non-\ ``NULL``, locklessly accessing the ``->a`` and ``->b`` fields.
204 204
205 :: 205 ::
206 206
207 1 bool add_gp_buggy(int a, int b) 207 1 bool add_gp_buggy(int a, int b)
208 2 { 208 2 {
209 3 p = kmalloc(sizeof(*p), GFP_KERNEL) 209 3 p = kmalloc(sizeof(*p), GFP_KERNEL);
210 4 if (!p) 210 4 if (!p)
211 5 return -ENOMEM; 211 5 return -ENOMEM;
212 6 spin_lock(&gp_lock); 212 6 spin_lock(&gp_lock);
213 7 if (rcu_access_pointer(gp)) { 213 7 if (rcu_access_pointer(gp)) {
214 8 spin_unlock(&gp_lock); 214 8 spin_unlock(&gp_lock);
215 9 return false; 215 9 return false;
216 10 } 216 10 }
217 11 p->a = a; 217 11 p->a = a;
218 12 p->b = a; 218 12 p->b = a;
219 13 gp = p; /* ORDERING BUG */ 219 13 gp = p; /* ORDERING BUG */
220 14 spin_unlock(&gp_lock); 220 14 spin_unlock(&gp_lock);
221 15 return true; 221 15 return true;
222 16 } 222 16 }
223 223
224 The problem is that both the compiler and weak 224 The problem is that both the compiler and weakly ordered CPUs are within
225 their rights to reorder this code as follows: 225 their rights to reorder this code as follows:
226 226
227 :: 227 ::
228 228
229 1 bool add_gp_buggy_optimized(int a, in 229 1 bool add_gp_buggy_optimized(int a, int b)
230 2 { 230 2 {
231 3 p = kmalloc(sizeof(*p), GFP_KERNEL) 231 3 p = kmalloc(sizeof(*p), GFP_KERNEL);
232 4 if (!p) 232 4 if (!p)
233 5 return -ENOMEM; 233 5 return -ENOMEM;
234 6 spin_lock(&gp_lock); 234 6 spin_lock(&gp_lock);
235 7 if (rcu_access_pointer(gp)) { 235 7 if (rcu_access_pointer(gp)) {
236 8 spin_unlock(&gp_lock); 236 8 spin_unlock(&gp_lock);
237 9 return false; 237 9 return false;
238 10 } 238 10 }
239 11 gp = p; /* ORDERING BUG */ 239 11 gp = p; /* ORDERING BUG */
240 12 p->a = a; 240 12 p->a = a;
241 13 p->b = a; 241 13 p->b = a;
242 14 spin_unlock(&gp_lock); 242 14 spin_unlock(&gp_lock);
243 15 return true; 243 15 return true;
244 16 } 244 16 }
245 245
246 If an RCU reader fetches ``gp`` just after ``a 246 If an RCU reader fetches ``gp`` just after ``add_gp_buggy_optimized``
247 executes line 11, it will see garbage in the 247 executes line 11, it will see garbage in the ``->a`` and ``->b`` fields.
248 And this is but one of many ways in which comp 248 And this is but one of many ways in which compiler and hardware
249 optimizations could cause trouble. Therefore, 249 optimizations could cause trouble. Therefore, we clearly need some way
250 to prevent the compiler and the CPU from reord 250 to prevent the compiler and the CPU from reordering in this manner,
251 which brings us to the publish-subscribe guara 251 which brings us to the publish-subscribe guarantee discussed in the next
252 section. 252 section.
253 253
254 Publish/Subscribe Guarantee 254 Publish/Subscribe Guarantee
255 ~~~~~~~~~~~~~~~~~~~~~~~~~~~ 255 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
256 256
257 RCU's publish-subscribe guarantee allows data 257 RCU's publish-subscribe guarantee allows data to be inserted into a
258 linked data structure without disrupting RCU r 258 linked data structure without disrupting RCU readers. The updater uses
259 rcu_assign_pointer() to insert the new data, a 259 rcu_assign_pointer() to insert the new data, and readers use
260 rcu_dereference() to access data, whether new 260 rcu_dereference() to access data, whether new or old. The following
261 shows an example of insertion: 261 shows an example of insertion:
262 262
263 :: 263 ::
264 264
265 1 bool add_gp(int a, int b) 265 1 bool add_gp(int a, int b)
266 2 { 266 2 {
267 3 p = kmalloc(sizeof(*p), GFP_KERNEL) 267 3 p = kmalloc(sizeof(*p), GFP_KERNEL);
268 4 if (!p) 268 4 if (!p)
269 5 return -ENOMEM; 269 5 return -ENOMEM;
270 6 spin_lock(&gp_lock); 270 6 spin_lock(&gp_lock);
271 7 if (rcu_access_pointer(gp)) { 271 7 if (rcu_access_pointer(gp)) {
272 8 spin_unlock(&gp_lock); 272 8 spin_unlock(&gp_lock);
273 9 return false; 273 9 return false;
274 10 } 274 10 }
275 11 p->a = a; 275 11 p->a = a;
276 12 p->b = a; 276 12 p->b = a;
277 13 rcu_assign_pointer(gp, p); 277 13 rcu_assign_pointer(gp, p);
278 14 spin_unlock(&gp_lock); 278 14 spin_unlock(&gp_lock);
279 15 return true; 279 15 return true;
280 16 } 280 16 }
281 281
282 The rcu_assign_pointer() on line 13 is concep 282 The rcu_assign_pointer() on line 13 is conceptually equivalent to a
283 simple assignment statement, but also guarante 283 simple assignment statement, but also guarantees that its assignment
284 will happen after the two assignments in lines 284 will happen after the two assignments in lines 11 and 12, similar to the
285 C11 ``memory_order_release`` store operation. 285 C11 ``memory_order_release`` store operation. It also prevents any
286 number of “interesting” compiler optimizat 286 number of “interesting” compiler optimizations, for example, the use of
287 ``gp`` as a scratch location immediately prece 287 ``gp`` as a scratch location immediately preceding the assignment.
288 288
289 +--------------------------------------------- 289 +-----------------------------------------------------------------------+
290 | **Quick Quiz**: 290 | **Quick Quiz**: |
291 +--------------------------------------------- 291 +-----------------------------------------------------------------------+
292 | But rcu_assign_pointer() does nothing to pre 292 | But rcu_assign_pointer() does nothing to prevent the two |
293 | assignments to ``p->a`` and ``p->b`` from be 293 | assignments to ``p->a`` and ``p->b`` from being reordered. Can't that |
294 | also cause problems? 294 | also cause problems? |
295 +--------------------------------------------- 295 +-----------------------------------------------------------------------+
296 | **Answer**: 296 | **Answer**: |
297 +--------------------------------------------- 297 +-----------------------------------------------------------------------+
298 | No, it cannot. The readers cannot see either 298 | No, it cannot. The readers cannot see either of these two fields |
299 | until the assignment to ``gp``, by which tim 299 | until the assignment to ``gp``, by which time both fields are fully |
300 | initialized. So reordering the assignments t 300 | initialized. So reordering the assignments to ``p->a`` and ``p->b`` |
301 | cannot possibly cause any problems. 301 | cannot possibly cause any problems. |
302 +--------------------------------------------- 302 +-----------------------------------------------------------------------+
303 303
304 It is tempting to assume that the reader need 304 It is tempting to assume that the reader need not do anything special to
305 control its accesses to the RCU-protected data 305 control its accesses to the RCU-protected data, as shown in
306 do_something_gp_buggy() below: 306 do_something_gp_buggy() below:
307 307
308 :: 308 ::
309 309
310 1 bool do_something_gp_buggy(void) 310 1 bool do_something_gp_buggy(void)
311 2 { 311 2 {
312 3 rcu_read_lock(); 312 3 rcu_read_lock();
313 4 p = gp; /* OPTIMIZATIONS GALORE!!! 313 4 p = gp; /* OPTIMIZATIONS GALORE!!! */
314 5 if (p) { 314 5 if (p) {
315 6 do_something(p->a, p->b); 315 6 do_something(p->a, p->b);
316 7 rcu_read_unlock(); 316 7 rcu_read_unlock();
317 8 return true; 317 8 return true;
318 9 } 318 9 }
319 10 rcu_read_unlock(); 319 10 rcu_read_unlock();
320 11 return false; 320 11 return false;
321 12 } 321 12 }
322 322
323 However, this temptation must be resisted beca 323 However, this temptation must be resisted because there are a
324 surprisingly large number of ways that the com 324 surprisingly large number of ways that the compiler (or weak ordering
325 CPUs like the DEC Alpha) can trip this code up 325 CPUs like the DEC Alpha) can trip this code up. For but one example, if
326 the compiler were short of registers, it might 326 the compiler were short of registers, it might choose to refetch from
327 ``gp`` rather than keeping a separate copy in 327 ``gp`` rather than keeping a separate copy in ``p`` as follows:
328 328
329 :: 329 ::
330 330
331 1 bool do_something_gp_buggy_optimized( 331 1 bool do_something_gp_buggy_optimized(void)
332 2 { 332 2 {
333 3 rcu_read_lock(); 333 3 rcu_read_lock();
334 4 if (gp) { /* OPTIMIZATIONS GALORE!! 334 4 if (gp) { /* OPTIMIZATIONS GALORE!!! */
335 5 do_something(gp->a, gp->b); 335 5 do_something(gp->a, gp->b);
336 6 rcu_read_unlock(); 336 6 rcu_read_unlock();
337 7 return true; 337 7 return true;
338 8 } 338 8 }
339 9 rcu_read_unlock(); 339 9 rcu_read_unlock();
340 10 return false; 340 10 return false;
341 11 } 341 11 }
342 342
343 If this function ran concurrently with a serie 343 If this function ran concurrently with a series of updates that replaced
344 the current structure with a new one, the fetc 344 the current structure with a new one, the fetches of ``gp->a`` and
345 ``gp->b`` might well come from two different s 345 ``gp->b`` might well come from two different structures, which could
346 cause serious confusion. To prevent this (and 346 cause serious confusion. To prevent this (and much else besides),
347 do_something_gp() uses rcu_dereference() to fe 347 do_something_gp() uses rcu_dereference() to fetch from ``gp``:
348 348
349 :: 349 ::
350 350
351 1 bool do_something_gp(void) 351 1 bool do_something_gp(void)
352 2 { 352 2 {
353 3 rcu_read_lock(); 353 3 rcu_read_lock();
354 4 p = rcu_dereference(gp); 354 4 p = rcu_dereference(gp);
355 5 if (p) { 355 5 if (p) {
356 6 do_something(p->a, p->b); 356 6 do_something(p->a, p->b);
357 7 rcu_read_unlock(); 357 7 rcu_read_unlock();
358 8 return true; 358 8 return true;
359 9 } 359 9 }
360 10 rcu_read_unlock(); 360 10 rcu_read_unlock();
361 11 return false; 361 11 return false;
362 12 } 362 12 }
363 363
364 The rcu_dereference() uses volatile casts and 364 The rcu_dereference() uses volatile casts and (for DEC Alpha) memory
365 barriers in the Linux kernel. Should a |high-q 365 barriers in the Linux kernel. Should a |high-quality implementation of
366 C11 memory_order_consume [PDF]|_ 366 C11 memory_order_consume [PDF]|_
367 ever appear, then rcu_dereference() could be i 367 ever appear, then rcu_dereference() could be implemented as a
368 ``memory_order_consume`` load. Regardless of t 368 ``memory_order_consume`` load. Regardless of the exact implementation, a
369 pointer fetched by rcu_dereference() may not b 369 pointer fetched by rcu_dereference() may not be used outside of the
370 outermost RCU read-side critical section conta 370 outermost RCU read-side critical section containing that
371 rcu_dereference(), unless protection of the co 371 rcu_dereference(), unless protection of the corresponding data
372 element has been passed from RCU to some other 372 element has been passed from RCU to some other synchronization
373 mechanism, most commonly locking or reference 373 mechanism, most commonly locking or reference counting
374 (see ../../rcuref.rst). 374 (see ../../rcuref.rst).
375 375
376 .. |high-quality implementation of C11 memory_ 376 .. |high-quality implementation of C11 memory_order_consume [PDF]| replace:: high-quality implementation of C11 ``memory_order_consume`` [PDF]
377 .. _high-quality implementation of C11 memory_ 377 .. _high-quality implementation of C11 memory_order_consume [PDF]: http://www.rdrop.com/users/paulmck/RCU/consume.2015.07.13a.pdf
378 378
379 In short, updaters use rcu_assign_pointer() an 379 In short, updaters use rcu_assign_pointer() and readers use
380 rcu_dereference(), and these two RCU API eleme 380 rcu_dereference(), and these two RCU API elements work together to
381 ensure that readers have a consistent view of 381 ensure that readers have a consistent view of newly added data elements.
382 382
383 Of course, it is also necessary to remove elem 383 Of course, it is also necessary to remove elements from RCU-protected
384 data structures, for example, using the follow 384 data structures, for example, using the following process:
385 385
386 #. Remove the data element from the enclosing 386 #. Remove the data element from the enclosing structure.
387 #. Wait for all pre-existing RCU read-side cri 387 #. Wait for all pre-existing RCU read-side critical sections to complete
388 (because only pre-existing readers can poss 388 (because only pre-existing readers can possibly have a reference to
389 the newly removed data element). 389 the newly removed data element).
390 #. At this point, only the updater has a refer 390 #. At this point, only the updater has a reference to the newly removed
391 data element, so it can safely reclaim the 391 data element, so it can safely reclaim the data element, for example,
392 by passing it to kfree(). 392 by passing it to kfree().
393 393
394 This process is implemented by remove_gp_synch 394 This process is implemented by remove_gp_synchronous():
395 395
396 :: 396 ::
397 397
398 1 bool remove_gp_synchronous(void) 398 1 bool remove_gp_synchronous(void)
399 2 { 399 2 {
400 3 struct foo *p; 400 3 struct foo *p;
401 4 401 4
402 5 spin_lock(&gp_lock); 402 5 spin_lock(&gp_lock);
403 6 p = rcu_access_pointer(gp); 403 6 p = rcu_access_pointer(gp);
404 7 if (!p) { 404 7 if (!p) {
405 8 spin_unlock(&gp_lock); 405 8 spin_unlock(&gp_lock);
406 9 return false; 406 9 return false;
407 10 } 407 10 }
408 11 rcu_assign_pointer(gp, NULL); 408 11 rcu_assign_pointer(gp, NULL);
409 12 spin_unlock(&gp_lock); 409 12 spin_unlock(&gp_lock);
410 13 synchronize_rcu(); 410 13 synchronize_rcu();
411 14 kfree(p); 411 14 kfree(p);
412 15 return true; 412 15 return true;
413 16 } 413 16 }
414 414
415 This function is straightforward, with line 1 415 This function is straightforward, with line 13 waiting for a grace
416 period before line 14 frees the old data elem 416 period before line 14 frees the old data element. This waiting ensures
417 that readers will reach line 7 of do_somethin 417 that readers will reach line 7 of do_something_gp() before the data
418 element referenced by ``p`` is freed. The rcu_ 418 element referenced by ``p`` is freed. The rcu_access_pointer() on
419 line 6 is similar to rcu_dereference(), excep 419 line 6 is similar to rcu_dereference(), except that:
420 420
421 #. The value returned by rcu_access_pointer() 421 #. The value returned by rcu_access_pointer() cannot be
422 dereferenced. If you want to access the val 422 dereferenced. If you want to access the value pointed to as well as
423 the pointer itself, use rcu_dereference() i 423 the pointer itself, use rcu_dereference() instead of
424 rcu_access_pointer(). 424 rcu_access_pointer().
425 #. The call to rcu_access_pointer() need not b 425 #. The call to rcu_access_pointer() need not be protected. In
426 contrast, rcu_dereference() must either be 426 contrast, rcu_dereference() must either be within an RCU
427 read-side critical section or in a code seg 427 read-side critical section or in a code segment where the pointer
428 cannot change, for example, in code protect 428 cannot change, for example, in code protected by the corresponding
429 update-side lock. 429 update-side lock.
430 430
431 +--------------------------------------------- 431 +-----------------------------------------------------------------------+
432 | **Quick Quiz**: 432 | **Quick Quiz**: |
433 +--------------------------------------------- 433 +-----------------------------------------------------------------------+
434 | Without the rcu_dereference() or the rcu_acc 434 | Without the rcu_dereference() or the rcu_access_pointer(), |
435 | what destructive optimizations might the com 435 | what destructive optimizations might the compiler make use of? |
436 +--------------------------------------------- 436 +-----------------------------------------------------------------------+
437 | **Answer**: 437 | **Answer**: |
438 +--------------------------------------------- 438 +-----------------------------------------------------------------------+
439 | Let's start with what happens to do_somethin 439 | Let's start with what happens to do_something_gp() if it fails to |
440 | use rcu_dereference(). It could reuse a valu 440 | use rcu_dereference(). It could reuse a value formerly fetched |
441 | from this same pointer. It could also fetch 441 | from this same pointer. It could also fetch the pointer from ``gp`` |
442 | in a byte-at-a-time manner, resulting in *lo 442 | in a byte-at-a-time manner, resulting in *load tearing*, in turn |
443 | resulting a bytewise mash-up of two distinct 443 | resulting a bytewise mash-up of two distinct pointer values. It might |
444 | even use value-speculation optimizations, wh 444 | even use value-speculation optimizations, where it makes a wrong |
445 | guess, but by the time it gets around to che 445 | guess, but by the time it gets around to checking the value, an |
446 | update has changed the pointer to match the 446 | update has changed the pointer to match the wrong guess. Too bad |
447 | about any dereferences that returned pre-ini 447 | about any dereferences that returned pre-initialization garbage in |
448 | the meantime! 448 | the meantime! |
449 | For remove_gp_synchronous(), as long as all 449 | For remove_gp_synchronous(), as long as all modifications to |
450 | ``gp`` are carried out while holding ``gp_lo 450 | ``gp`` are carried out while holding ``gp_lock``, the above |
451 | optimizations are harmless. However, ``spars 451 | optimizations are harmless. However, ``sparse`` will complain if you |
452 | define ``gp`` with ``__rcu`` and then access 452 | define ``gp`` with ``__rcu`` and then access it without using either |
453 | rcu_access_pointer() or rcu_dereference(). 453 | rcu_access_pointer() or rcu_dereference(). |
454 +--------------------------------------------- 454 +-----------------------------------------------------------------------+
455 455
456 In short, RCU's publish-subscribe guarantee is 456 In short, RCU's publish-subscribe guarantee is provided by the
457 combination of rcu_assign_pointer() and rcu_de 457 combination of rcu_assign_pointer() and rcu_dereference(). This
458 guarantee allows data elements to be safely ad 458 guarantee allows data elements to be safely added to RCU-protected
459 linked data structures without disrupting RCU 459 linked data structures without disrupting RCU readers. This guarantee
460 can be used in combination with the grace-peri 460 can be used in combination with the grace-period guarantee to also allow
461 data elements to be removed from RCU-protected 461 data elements to be removed from RCU-protected linked data structures,
462 again without disrupting RCU readers. 462 again without disrupting RCU readers.
463 463
464 This guarantee was only partially premeditated 464 This guarantee was only partially premeditated. DYNIX/ptx used an
465 explicit memory barrier for publication, but h 465 explicit memory barrier for publication, but had nothing resembling
466 rcu_dereference() for subscription, nor did it 466 rcu_dereference() for subscription, nor did it have anything
467 resembling the dependency-ordering barrier tha 467 resembling the dependency-ordering barrier that was later subsumed
468 into rcu_dereference() and later still into RE 468 into rcu_dereference() and later still into READ_ONCE(). The
469 need for these operations made itself known qu 469 need for these operations made itself known quite suddenly at a
470 late-1990s meeting with the DEC Alpha architec 470 late-1990s meeting with the DEC Alpha architects, back in the days when
471 DEC was still a free-standing company. It took 471 DEC was still a free-standing company. It took the Alpha architects a
472 good hour to convince me that any sort of barr 472 good hour to convince me that any sort of barrier would ever be needed,
473 and it then took me a good *two* hours to conv 473 and it then took me a good *two* hours to convince them that their
474 documentation did not make this point clear. M 474 documentation did not make this point clear. More recent work with the C
475 and C++ standards committees have provided muc 475 and C++ standards committees have provided much education on tricks and
476 traps from the compiler. In short, compilers w 476 traps from the compiler. In short, compilers were much less tricky in
477 the early 1990s, but in 2015, don't even think 477 the early 1990s, but in 2015, don't even think about omitting
478 rcu_dereference()! 478 rcu_dereference()!
479 479
480 Memory-Barrier Guarantees 480 Memory-Barrier Guarantees
481 ~~~~~~~~~~~~~~~~~~~~~~~~~ 481 ~~~~~~~~~~~~~~~~~~~~~~~~~
482 482
483 The previous section's simple linked-data-stru 483 The previous section's simple linked-data-structure scenario clearly
484 demonstrates the need for RCU's stringent memo 484 demonstrates the need for RCU's stringent memory-ordering guarantees on
485 systems with more than one CPU: 485 systems with more than one CPU:
486 486
487 #. Each CPU that has an RCU read-side critical 487 #. Each CPU that has an RCU read-side critical section that begins
488 before synchronize_rcu() starts is guarante 488 before synchronize_rcu() starts is guaranteed to execute a full
489 memory barrier between the time that the RC 489 memory barrier between the time that the RCU read-side critical
490 section ends and the time that synchronize_ 490 section ends and the time that synchronize_rcu() returns. Without
491 this guarantee, a pre-existing RCU read-sid 491 this guarantee, a pre-existing RCU read-side critical section might
492 hold a reference to the newly removed ``str 492 hold a reference to the newly removed ``struct foo`` after the
493 kfree() on line 14 of remove_gp_synchronou 493 kfree() on line 14 of remove_gp_synchronous().
494 #. Each CPU that has an RCU read-side critical 494 #. Each CPU that has an RCU read-side critical section that ends after
495 synchronize_rcu() returns is guaranteed to 495 synchronize_rcu() returns is guaranteed to execute a full memory
496 barrier between the time that synchronize_r 496 barrier between the time that synchronize_rcu() begins and the
497 time that the RCU read-side critical sectio 497 time that the RCU read-side critical section begins. Without this
498 guarantee, a later RCU read-side critical s 498 guarantee, a later RCU read-side critical section running after the
499 kfree() on line 14 of remove_gp_synchronou 499 kfree() on line 14 of remove_gp_synchronous() might later run
500 do_something_gp() and find the newly delete 500 do_something_gp() and find the newly deleted ``struct foo``.
501 #. If the task invoking synchronize_rcu() rema 501 #. If the task invoking synchronize_rcu() remains on a given CPU,
502 then that CPU is guaranteed to execute a fu 502 then that CPU is guaranteed to execute a full memory barrier sometime
503 during the execution of synchronize_rcu(). 503 during the execution of synchronize_rcu(). This guarantee ensures
504 that the kfree() on line 14 of remove_gp_s 504 that the kfree() on line 14 of remove_gp_synchronous() really
505 does execute after the removal on line 11. 505 does execute after the removal on line 11.
506 #. If the task invoking synchronize_rcu() migr 506 #. If the task invoking synchronize_rcu() migrates among a group of
507 CPUs during that invocation, then each of t 507 CPUs during that invocation, then each of the CPUs in that group is
508 guaranteed to execute a full memory barrier 508 guaranteed to execute a full memory barrier sometime during the
509 execution of synchronize_rcu(). This guaran 509 execution of synchronize_rcu(). This guarantee also ensures that
510 the kfree() on line 14 of remove_gp_synchr 510 the kfree() on line 14 of remove_gp_synchronous() really does
511 execute after the removal on line 11, but 511 execute after the removal on line 11, but also in the case where the
512 thread executing the synchronize_rcu() migr 512 thread executing the synchronize_rcu() migrates in the meantime.
513 513
514 +--------------------------------------------- 514 +-----------------------------------------------------------------------+
515 | **Quick Quiz**: 515 | **Quick Quiz**: |
516 +--------------------------------------------- 516 +-----------------------------------------------------------------------+
517 | Given that multiple CPUs can start RCU read- 517 | Given that multiple CPUs can start RCU read-side critical sections at |
518 | any time without any ordering whatsoever, ho 518 | any time without any ordering whatsoever, how can RCU possibly tell |
519 | whether or not a given RCU read-side critica 519 | whether or not a given RCU read-side critical section starts before a |
520 | given instance of synchronize_rcu()? 520 | given instance of synchronize_rcu()? |
521 +--------------------------------------------- 521 +-----------------------------------------------------------------------+
522 | **Answer**: 522 | **Answer**: |
523 +--------------------------------------------- 523 +-----------------------------------------------------------------------+
524 | If RCU cannot tell whether or not a given RC 524 | If RCU cannot tell whether or not a given RCU read-side critical |
525 | section starts before a given instance of sy 525 | section starts before a given instance of synchronize_rcu(), then |
526 | it must assume that the RCU read-side critic 526 | it must assume that the RCU read-side critical section started first. |
527 | In other words, a given instance of synchron 527 | In other words, a given instance of synchronize_rcu() can avoid |
528 | waiting on a given RCU read-side critical se 528 | waiting on a given RCU read-side critical section only if it can |
529 | prove that synchronize_rcu() started first. 529 | prove that synchronize_rcu() started first. |
530 | A related question is “When rcu_read_lock( 530 | A related question is “When rcu_read_lock() doesn't generate any |
531 | code, why does it matter how it relates to a 531 | code, why does it matter how it relates to a grace period?” The |
532 | answer is that it is not the relationship of 532 | answer is that it is not the relationship of rcu_read_lock() |
533 | itself that is important, but rather the rel 533 | itself that is important, but rather the relationship of the code |
534 | within the enclosed RCU read-side critical s 534 | within the enclosed RCU read-side critical section to the code |
535 | preceding and following the grace period. If 535 | preceding and following the grace period. If we take this viewpoint, |
536 | then a given RCU read-side critical section 536 | then a given RCU read-side critical section begins before a given |
537 | grace period when some access preceding the 537 | grace period when some access preceding the grace period observes the |
538 | effect of some access within the critical se 538 | effect of some access within the critical section, in which case none |
539 | of the accesses within the critical section 539 | of the accesses within the critical section may observe the effects |
540 | of any access following the grace period. 540 | of any access following the grace period. |
541 | 541 | |
542 | As of late 2016, mathematical models of RCU 542 | As of late 2016, mathematical models of RCU take this viewpoint, for |
543 | example, see slides 62 and 63 of the `2016 543 | example, see slides 62 and 63 of the `2016 LinuxCon |
544 | EU <http://www2.rdrop.com/users/paulmck/scal 544 | EU <http://www2.rdrop.com/users/paulmck/scalability/paper/LinuxMM.201 |
545 | 6.10.04c.LCE.pdf>`__ 545 | 6.10.04c.LCE.pdf>`__ |
546 | presentation. 546 | presentation. |
547 +--------------------------------------------- 547 +-----------------------------------------------------------------------+
548 548
549 +--------------------------------------------- 549 +-----------------------------------------------------------------------+
550 | **Quick Quiz**: 550 | **Quick Quiz**: |
551 +--------------------------------------------- 551 +-----------------------------------------------------------------------+
552 | The first and second guarantees require unbe 552 | The first and second guarantees require unbelievably strict ordering! |
553 | Are all these memory barriers *really* requi 553 | Are all these memory barriers *really* required? |
554 +--------------------------------------------- 554 +-----------------------------------------------------------------------+
555 | **Answer**: 555 | **Answer**: |
556 +--------------------------------------------- 556 +-----------------------------------------------------------------------+
557 | Yes, they really are required. To see why th 557 | Yes, they really are required. To see why the first guarantee is |
558 | required, consider the following sequence of 558 | required, consider the following sequence of events: |
559 | 559 | |
560 | #. CPU 1: rcu_read_lock() 560 | #. CPU 1: rcu_read_lock() |
561 | #. CPU 1: ``q = rcu_dereference(gp); /* Very 561 | #. CPU 1: ``q = rcu_dereference(gp); /* Very likely to return p. */`` |
562 | #. CPU 0: ``list_del_rcu(p);`` 562 | #. CPU 0: ``list_del_rcu(p);`` |
563 | #. CPU 0: synchronize_rcu() starts. 563 | #. CPU 0: synchronize_rcu() starts. |
564 | #. CPU 1: ``do_something_with(q->a);`` 564 | #. CPU 1: ``do_something_with(q->a);`` |
565 | ``/* No smp_mb(), so might happen after k 565 | ``/* No smp_mb(), so might happen after kfree(). */`` |
566 | #. CPU 1: rcu_read_unlock() 566 | #. CPU 1: rcu_read_unlock() |
567 | #. CPU 0: synchronize_rcu() returns. 567 | #. CPU 0: synchronize_rcu() returns. |
568 | #. CPU 0: ``kfree(p);`` 568 | #. CPU 0: ``kfree(p);`` |
569 | 569 | |
570 | Therefore, there absolutely must be a full m 570 | Therefore, there absolutely must be a full memory barrier between the |
571 | end of the RCU read-side critical section an 571 | end of the RCU read-side critical section and the end of the grace |
572 | period. 572 | period. |
573 | 573 | |
574 | The sequence of events demonstrating the nec 574 | The sequence of events demonstrating the necessity of the second rule |
575 | is roughly similar: 575 | is roughly similar: |
576 | 576 | |
577 | #. CPU 0: ``list_del_rcu(p);`` 577 | #. CPU 0: ``list_del_rcu(p);`` |
578 | #. CPU 0: synchronize_rcu() starts. 578 | #. CPU 0: synchronize_rcu() starts. |
579 | #. CPU 1: rcu_read_lock() 579 | #. CPU 1: rcu_read_lock() |
580 | #. CPU 1: ``q = rcu_dereference(gp);`` 580 | #. CPU 1: ``q = rcu_dereference(gp);`` |
581 | ``/* Might return p if no memory barrier. 581 | ``/* Might return p if no memory barrier. */`` |
582 | #. CPU 0: synchronize_rcu() returns. 582 | #. CPU 0: synchronize_rcu() returns. |
583 | #. CPU 0: ``kfree(p);`` 583 | #. CPU 0: ``kfree(p);`` |
584 | #. CPU 1: ``do_something_with(q->a); /* Boom 584 | #. CPU 1: ``do_something_with(q->a); /* Boom!!! */`` |
585 | #. CPU 1: rcu_read_unlock() 585 | #. CPU 1: rcu_read_unlock() |
586 | 586 | |
587 | And similarly, without a memory barrier betw 587 | And similarly, without a memory barrier between the beginning of the |
588 | grace period and the beginning of the RCU re 588 | grace period and the beginning of the RCU read-side critical section, |
589 | CPU 1 might end up accessing the freelist. 589 | CPU 1 might end up accessing the freelist. |
590 | 590 | |
591 | The “as if” rule of course applies, so t 591 | The “as if” rule of course applies, so that any implementation that |
592 | acts as if the appropriate memory barriers w 592 | acts as if the appropriate memory barriers were in place is a correct |
593 | implementation. That said, it is much easier 593 | implementation. That said, it is much easier to fool yourself into |
594 | believing that you have adhered to the as-if 594 | believing that you have adhered to the as-if rule than it is to |
595 | actually adhere to it! 595 | actually adhere to it! |
596 +--------------------------------------------- 596 +-----------------------------------------------------------------------+
597 597
598 +--------------------------------------------- 598 +-----------------------------------------------------------------------+
599 | **Quick Quiz**: 599 | **Quick Quiz**: |
600 +--------------------------------------------- 600 +-----------------------------------------------------------------------+
601 | You claim that rcu_read_lock() and rcu_read_ 601 | You claim that rcu_read_lock() and rcu_read_unlock() generate |
602 | absolutely no code in some kernel builds. Th 602 | absolutely no code in some kernel builds. This means that the |
603 | compiler might arbitrarily rearrange consecu 603 | compiler might arbitrarily rearrange consecutive RCU read-side |
604 | critical sections. Given such rearrangement, 604 | critical sections. Given such rearrangement, if a given RCU read-side |
605 | critical section is done, how can you be sur 605 | critical section is done, how can you be sure that all prior RCU |
606 | read-side critical sections are done? Won't 606 | read-side critical sections are done? Won't the compiler |
607 | rearrangements make that impossible to deter 607 | rearrangements make that impossible to determine? |
608 +--------------------------------------------- 608 +-----------------------------------------------------------------------+
609 | **Answer**: 609 | **Answer**: |
610 +--------------------------------------------- 610 +-----------------------------------------------------------------------+
611 | In cases where rcu_read_lock() and rcu_read_ 611 | In cases where rcu_read_lock() and rcu_read_unlock() generate |
612 | absolutely no code, RCU infers quiescent sta 612 | absolutely no code, RCU infers quiescent states only at special |
613 | locations, for example, within the scheduler 613 | locations, for example, within the scheduler. Because calls to |
614 | schedule() had better prevent calling-code a 614 | schedule() had better prevent calling-code accesses to shared |
615 | variables from being rearranged across the c 615 | variables from being rearranged across the call to schedule(), if |
616 | RCU detects the end of a given RCU read-side 616 | RCU detects the end of a given RCU read-side critical section, it |
617 | will necessarily detect the end of all prior 617 | will necessarily detect the end of all prior RCU read-side critical |
618 | sections, no matter how aggressively the com 618 | sections, no matter how aggressively the compiler scrambles the code. |
619 | Again, this all assumes that the compiler ca 619 | Again, this all assumes that the compiler cannot scramble code across |
620 | calls to the scheduler, out of interrupt han 620 | calls to the scheduler, out of interrupt handlers, into the idle |
621 | loop, into user-mode code, and so on. But if 621 | loop, into user-mode code, and so on. But if your kernel build allows |
622 | that sort of scrambling, you have broken far 622 | that sort of scrambling, you have broken far more than just RCU! |
623 +--------------------------------------------- 623 +-----------------------------------------------------------------------+
624 624
625 Note that these memory-barrier requirements do 625 Note that these memory-barrier requirements do not replace the
626 fundamental RCU requirement that a grace perio 626 fundamental RCU requirement that a grace period wait for all
627 pre-existing readers. On the contrary, the mem 627 pre-existing readers. On the contrary, the memory barriers called out in
628 this section must operate in such a way as to 628 this section must operate in such a way as to *enforce* this fundamental
629 requirement. Of course, different implementati 629 requirement. Of course, different implementations enforce this
630 requirement in different ways, but enforce it 630 requirement in different ways, but enforce it they must.
631 631
632 RCU Primitives Guaranteed to Execute Unconditi 632 RCU Primitives Guaranteed to Execute Unconditionally
633 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 633 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
634 634
635 The common-case RCU primitives are uncondition 635 The common-case RCU primitives are unconditional. They are invoked, they
636 do their job, and they return, with no possibi 636 do their job, and they return, with no possibility of error, and no need
637 to retry. This is a key RCU design philosophy. 637 to retry. This is a key RCU design philosophy.
638 638
639 However, this philosophy is pragmatic rather t 639 However, this philosophy is pragmatic rather than pigheaded. If someone
640 comes up with a good justification for a parti 640 comes up with a good justification for a particular conditional RCU
641 primitive, it might well be implemented and ad 641 primitive, it might well be implemented and added. After all, this
642 guarantee was reverse-engineered, not premedit 642 guarantee was reverse-engineered, not premeditated. The unconditional
643 nature of the RCU primitives was initially an 643 nature of the RCU primitives was initially an accident of
644 implementation, and later experience with sync 644 implementation, and later experience with synchronization primitives
645 with conditional primitives caused me to eleva 645 with conditional primitives caused me to elevate this accident to a
646 guarantee. Therefore, the justification for ad 646 guarantee. Therefore, the justification for adding a conditional
647 primitive to RCU would need to be based on det 647 primitive to RCU would need to be based on detailed and compelling use
648 cases. 648 cases.
649 649
650 Guaranteed Read-to-Write Upgrade 650 Guaranteed Read-to-Write Upgrade
651 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 651 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
652 652
653 As far as RCU is concerned, it is always possi 653 As far as RCU is concerned, it is always possible to carry out an update
654 within an RCU read-side critical section. For 654 within an RCU read-side critical section. For example, that RCU
655 read-side critical section might search for a 655 read-side critical section might search for a given data element, and
656 then might acquire the update-side spinlock in 656 then might acquire the update-side spinlock in order to update that
657 element, all while remaining in that RCU read- 657 element, all while remaining in that RCU read-side critical section. Of
658 course, it is necessary to exit the RCU read-s 658 course, it is necessary to exit the RCU read-side critical section
659 before invoking synchronize_rcu(), however, th 659 before invoking synchronize_rcu(), however, this inconvenience can
660 be avoided through use of the call_rcu() and k 660 be avoided through use of the call_rcu() and kfree_rcu() API
661 members described later in this document. 661 members described later in this document.
662 662
663 +--------------------------------------------- 663 +-----------------------------------------------------------------------+
664 | **Quick Quiz**: 664 | **Quick Quiz**: |
665 +--------------------------------------------- 665 +-----------------------------------------------------------------------+
666 | But how does the upgrade-to-write operation 666 | But how does the upgrade-to-write operation exclude other readers? |
667 +--------------------------------------------- 667 +-----------------------------------------------------------------------+
668 | **Answer**: 668 | **Answer**: |
669 +--------------------------------------------- 669 +-----------------------------------------------------------------------+
670 | It doesn't, just like normal RCU updates, wh 670 | It doesn't, just like normal RCU updates, which also do not exclude |
671 | RCU readers. 671 | RCU readers. |
672 +--------------------------------------------- 672 +-----------------------------------------------------------------------+
673 673
674 This guarantee allows lookup code to be shared 674 This guarantee allows lookup code to be shared between read-side and
675 update-side code, and was premeditated, appear 675 update-side code, and was premeditated, appearing in the earliest
676 DYNIX/ptx RCU documentation. 676 DYNIX/ptx RCU documentation.
677 677
678 Fundamental Non-Requirements 678 Fundamental Non-Requirements
679 ---------------------------- 679 ----------------------------
680 680
681 RCU provides extremely lightweight readers, an 681 RCU provides extremely lightweight readers, and its read-side
682 guarantees, though quite useful, are correspon 682 guarantees, though quite useful, are correspondingly lightweight. It is
683 therefore all too easy to assume that RCU is g 683 therefore all too easy to assume that RCU is guaranteeing more than it
684 really is. Of course, the list of things that 684 really is. Of course, the list of things that RCU does not guarantee is
685 infinitely long, however, the following sectio 685 infinitely long, however, the following sections list a few
686 non-guarantees that have caused confusion. Exc 686 non-guarantees that have caused confusion. Except where otherwise noted,
687 these non-guarantees were premeditated. 687 these non-guarantees were premeditated.
688 688
689 #. `Readers Impose Minimal Ordering`_ 689 #. `Readers Impose Minimal Ordering`_
690 #. `Readers Do Not Exclude Updaters`_ 690 #. `Readers Do Not Exclude Updaters`_
691 #. `Updaters Only Wait For Old Readers`_ 691 #. `Updaters Only Wait For Old Readers`_
692 #. `Grace Periods Don't Partition Read-Side Cr 692 #. `Grace Periods Don't Partition Read-Side Critical Sections`_
693 #. `Read-Side Critical Sections Don't Partitio 693 #. `Read-Side Critical Sections Don't Partition Grace Periods`_
694 694
695 Readers Impose Minimal Ordering 695 Readers Impose Minimal Ordering
696 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 696 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
697 697
698 Reader-side markers such as rcu_read_lock() an 698 Reader-side markers such as rcu_read_lock() and
699 rcu_read_unlock() provide absolutely no orderi 699 rcu_read_unlock() provide absolutely no ordering guarantees except
700 through their interaction with the grace-perio 700 through their interaction with the grace-period APIs such as
701 synchronize_rcu(). To see this, consider the f 701 synchronize_rcu(). To see this, consider the following pair of
702 threads: 702 threads:
703 703
704 :: 704 ::
705 705
706 1 void thread0(void) 706 1 void thread0(void)
707 2 { 707 2 {
708 3 rcu_read_lock(); 708 3 rcu_read_lock();
709 4 WRITE_ONCE(x, 1); 709 4 WRITE_ONCE(x, 1);
710 5 rcu_read_unlock(); 710 5 rcu_read_unlock();
711 6 rcu_read_lock(); 711 6 rcu_read_lock();
712 7 WRITE_ONCE(y, 1); 712 7 WRITE_ONCE(y, 1);
713 8 rcu_read_unlock(); 713 8 rcu_read_unlock();
714 9 } 714 9 }
715 10 715 10
716 11 void thread1(void) 716 11 void thread1(void)
717 12 { 717 12 {
718 13 rcu_read_lock(); 718 13 rcu_read_lock();
719 14 r1 = READ_ONCE(y); 719 14 r1 = READ_ONCE(y);
720 15 rcu_read_unlock(); 720 15 rcu_read_unlock();
721 16 rcu_read_lock(); 721 16 rcu_read_lock();
722 17 r2 = READ_ONCE(x); 722 17 r2 = READ_ONCE(x);
723 18 rcu_read_unlock(); 723 18 rcu_read_unlock();
724 19 } 724 19 }
725 725
726 After thread0() and thread1() execute concurre 726 After thread0() and thread1() execute concurrently, it is quite
727 possible to have 727 possible to have
728 728
729 :: 729 ::
730 730
731 (r1 == 1 && r2 == 0) 731 (r1 == 1 && r2 == 0)
732 732
733 (that is, ``y`` appears to have been assigned 733 (that is, ``y`` appears to have been assigned before ``x``), which would
734 not be possible if rcu_read_lock() and rcu_rea 734 not be possible if rcu_read_lock() and rcu_read_unlock() had
735 much in the way of ordering properties. But th 735 much in the way of ordering properties. But they do not, so the CPU is
736 within its rights to do significant reordering 736 within its rights to do significant reordering. This is by design: Any
737 significant ordering constraints would slow do 737 significant ordering constraints would slow down these fast-path APIs.
738 738
739 +--------------------------------------------- 739 +-----------------------------------------------------------------------+
740 | **Quick Quiz**: 740 | **Quick Quiz**: |
741 +--------------------------------------------- 741 +-----------------------------------------------------------------------+
742 | Can't the compiler also reorder this code? 742 | Can't the compiler also reorder this code? |
743 +--------------------------------------------- 743 +-----------------------------------------------------------------------+
744 | **Answer**: 744 | **Answer**: |
745 +--------------------------------------------- 745 +-----------------------------------------------------------------------+
746 | No, the volatile casts in READ_ONCE() and WR 746 | No, the volatile casts in READ_ONCE() and WRITE_ONCE() |
747 | prevent the compiler from reordering in this 747 | prevent the compiler from reordering in this particular case. |
748 +--------------------------------------------- 748 +-----------------------------------------------------------------------+
749 749
750 Readers Do Not Exclude Updaters 750 Readers Do Not Exclude Updaters
751 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 751 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
752 752
753 Neither rcu_read_lock() nor rcu_read_unlock() 753 Neither rcu_read_lock() nor rcu_read_unlock() exclude updates.
754 All they do is to prevent grace periods from e 754 All they do is to prevent grace periods from ending. The following
755 example illustrates this: 755 example illustrates this:
756 756
757 :: 757 ::
758 758
759 1 void thread0(void) 759 1 void thread0(void)
760 2 { 760 2 {
761 3 rcu_read_lock(); 761 3 rcu_read_lock();
762 4 r1 = READ_ONCE(y); 762 4 r1 = READ_ONCE(y);
763 5 if (r1) { 763 5 if (r1) {
764 6 do_something_with_nonzero_x(); 764 6 do_something_with_nonzero_x();
765 7 r2 = READ_ONCE(x); 765 7 r2 = READ_ONCE(x);
766 8 WARN_ON(!r2); /* BUG!!! */ 766 8 WARN_ON(!r2); /* BUG!!! */
767 9 } 767 9 }
768 10 rcu_read_unlock(); 768 10 rcu_read_unlock();
769 11 } 769 11 }
770 12 770 12
771 13 void thread1(void) 771 13 void thread1(void)
772 14 { 772 14 {
773 15 spin_lock(&my_lock); 773 15 spin_lock(&my_lock);
774 16 WRITE_ONCE(x, 1); 774 16 WRITE_ONCE(x, 1);
775 17 WRITE_ONCE(y, 1); 775 17 WRITE_ONCE(y, 1);
776 18 spin_unlock(&my_lock); 776 18 spin_unlock(&my_lock);
777 19 } 777 19 }
778 778
779 If the thread0() function's rcu_read_lock() ex 779 If the thread0() function's rcu_read_lock() excluded the
780 thread1() function's update, the WARN_ON() cou 780 thread1() function's update, the WARN_ON() could never fire. But
781 the fact is that rcu_read_lock() does not excl 781 the fact is that rcu_read_lock() does not exclude much of anything
782 aside from subsequent grace periods, of which 782 aside from subsequent grace periods, of which thread1() has none, so
783 the WARN_ON() can and does fire. 783 the WARN_ON() can and does fire.
784 784
785 Updaters Only Wait For Old Readers 785 Updaters Only Wait For Old Readers
786 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 786 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
787 787
788 It might be tempting to assume that after sync 788 It might be tempting to assume that after synchronize_rcu()
789 completes, there are no readers executing. Thi 789 completes, there are no readers executing. This temptation must be
790 avoided because new readers can start immediat 790 avoided because new readers can start immediately after
791 synchronize_rcu() starts, and synchronize_rcu( 791 synchronize_rcu() starts, and synchronize_rcu() is under no
792 obligation to wait for these new readers. 792 obligation to wait for these new readers.
793 793
794 +--------------------------------------------- 794 +-----------------------------------------------------------------------+
795 | **Quick Quiz**: 795 | **Quick Quiz**: |
796 +--------------------------------------------- 796 +-----------------------------------------------------------------------+
797 | Suppose that synchronize_rcu() did wait unti 797 | Suppose that synchronize_rcu() did wait until *all* readers had |
798 | completed instead of waiting only on pre-exi 798 | completed instead of waiting only on pre-existing readers. For how |
799 | long would the updater be able to rely on th 799 | long would the updater be able to rely on there being no readers? |
800 +--------------------------------------------- 800 +-----------------------------------------------------------------------+
801 | **Answer**: 801 | **Answer**: |
802 +--------------------------------------------- 802 +-----------------------------------------------------------------------+
803 | For no time at all. Even if synchronize_rcu( 803 | For no time at all. Even if synchronize_rcu() were to wait until |
804 | all readers had completed, a new reader migh 804 | all readers had completed, a new reader might start immediately after |
805 | synchronize_rcu() completed. Therefore, the 805 | synchronize_rcu() completed. Therefore, the code following |
806 | synchronize_rcu() can *never* rely on there 806 | synchronize_rcu() can *never* rely on there being no readers. |
807 +--------------------------------------------- 807 +-----------------------------------------------------------------------+
808 808
809 Grace Periods Don't Partition Read-Side Critic 809 Grace Periods Don't Partition Read-Side Critical Sections
810 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 810 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
811 811
812 It is tempting to assume that if any part of o 812 It is tempting to assume that if any part of one RCU read-side critical
813 section precedes a given grace period, and if 813 section precedes a given grace period, and if any part of another RCU
814 read-side critical section follows that same g 814 read-side critical section follows that same grace period, then all of
815 the first RCU read-side critical section must 815 the first RCU read-side critical section must precede all of the second.
816 However, this just isn't the case: A single gr 816 However, this just isn't the case: A single grace period does not
817 partition the set of RCU read-side critical se 817 partition the set of RCU read-side critical sections. An example of this
818 situation can be illustrated as follows, where 818 situation can be illustrated as follows, where ``x``, ``y``, and ``z``
819 are initially all zero: 819 are initially all zero:
820 820
821 :: 821 ::
822 822
823 1 void thread0(void) 823 1 void thread0(void)
824 2 { 824 2 {
825 3 rcu_read_lock(); 825 3 rcu_read_lock();
826 4 WRITE_ONCE(a, 1); 826 4 WRITE_ONCE(a, 1);
827 5 WRITE_ONCE(b, 1); 827 5 WRITE_ONCE(b, 1);
828 6 rcu_read_unlock(); 828 6 rcu_read_unlock();
829 7 } 829 7 }
830 8 830 8
831 9 void thread1(void) 831 9 void thread1(void)
832 10 { 832 10 {
833 11 r1 = READ_ONCE(a); 833 11 r1 = READ_ONCE(a);
834 12 synchronize_rcu(); 834 12 synchronize_rcu();
835 13 WRITE_ONCE(c, 1); 835 13 WRITE_ONCE(c, 1);
836 14 } 836 14 }
837 15 837 15
838 16 void thread2(void) 838 16 void thread2(void)
839 17 { 839 17 {
840 18 rcu_read_lock(); 840 18 rcu_read_lock();
841 19 r2 = READ_ONCE(b); 841 19 r2 = READ_ONCE(b);
842 20 r3 = READ_ONCE(c); 842 20 r3 = READ_ONCE(c);
843 21 rcu_read_unlock(); 843 21 rcu_read_unlock();
844 22 } 844 22 }
845 845
846 It turns out that the outcome: 846 It turns out that the outcome:
847 847
848 :: 848 ::
849 849
850 (r1 == 1 && r2 == 0 && r3 == 1) 850 (r1 == 1 && r2 == 0 && r3 == 1)
851 851
852 is entirely possible. The following figure sho 852 is entirely possible. The following figure show how this can happen,
853 with each circled ``QS`` indicating the point 853 with each circled ``QS`` indicating the point at which RCU recorded a
854 *quiescent state* for each thread, that is, a 854 *quiescent state* for each thread, that is, a state in which RCU knows
855 that the thread cannot be in the midst of an R 855 that the thread cannot be in the midst of an RCU read-side critical
856 section that started before the current grace 856 section that started before the current grace period:
857 857
858 .. kernel-figure:: GPpartitionReaders1.svg 858 .. kernel-figure:: GPpartitionReaders1.svg
859 859
860 If it is necessary to partition RCU read-side 860 If it is necessary to partition RCU read-side critical sections in this
861 manner, it is necessary to use two grace perio 861 manner, it is necessary to use two grace periods, where the first grace
862 period is known to end before the second grace 862 period is known to end before the second grace period starts:
863 863
864 :: 864 ::
865 865
866 1 void thread0(void) 866 1 void thread0(void)
867 2 { 867 2 {
868 3 rcu_read_lock(); 868 3 rcu_read_lock();
869 4 WRITE_ONCE(a, 1); 869 4 WRITE_ONCE(a, 1);
870 5 WRITE_ONCE(b, 1); 870 5 WRITE_ONCE(b, 1);
871 6 rcu_read_unlock(); 871 6 rcu_read_unlock();
872 7 } 872 7 }
873 8 873 8
874 9 void thread1(void) 874 9 void thread1(void)
875 10 { 875 10 {
876 11 r1 = READ_ONCE(a); 876 11 r1 = READ_ONCE(a);
877 12 synchronize_rcu(); 877 12 synchronize_rcu();
878 13 WRITE_ONCE(c, 1); 878 13 WRITE_ONCE(c, 1);
879 14 } 879 14 }
880 15 880 15
881 16 void thread2(void) 881 16 void thread2(void)
882 17 { 882 17 {
883 18 r2 = READ_ONCE(c); 883 18 r2 = READ_ONCE(c);
884 19 synchronize_rcu(); 884 19 synchronize_rcu();
885 20 WRITE_ONCE(d, 1); 885 20 WRITE_ONCE(d, 1);
886 21 } 886 21 }
887 22 887 22
888 23 void thread3(void) 888 23 void thread3(void)
889 24 { 889 24 {
890 25 rcu_read_lock(); 890 25 rcu_read_lock();
891 26 r3 = READ_ONCE(b); 891 26 r3 = READ_ONCE(b);
892 27 r4 = READ_ONCE(d); 892 27 r4 = READ_ONCE(d);
893 28 rcu_read_unlock(); 893 28 rcu_read_unlock();
894 29 } 894 29 }
895 895
896 Here, if ``(r1 == 1)``, then thread0()'s write 896 Here, if ``(r1 == 1)``, then thread0()'s write to ``b`` must happen
897 before the end of thread1()'s grace period. If 897 before the end of thread1()'s grace period. If in addition
898 ``(r4 == 1)``, then thread3()'s read from ``b` 898 ``(r4 == 1)``, then thread3()'s read from ``b`` must happen after
899 the beginning of thread2()'s grace period. If 899 the beginning of thread2()'s grace period. If it is also the case
900 that ``(r2 == 1)``, then the end of thread1()' 900 that ``(r2 == 1)``, then the end of thread1()'s grace period must
901 precede the beginning of thread2()'s grace per 901 precede the beginning of thread2()'s grace period. This mean that
902 the two RCU read-side critical sections cannot 902 the two RCU read-side critical sections cannot overlap, guaranteeing
903 that ``(r3 == 1)``. As a result, the outcome: 903 that ``(r3 == 1)``. As a result, the outcome:
904 904
905 :: 905 ::
906 906
907 (r1 == 1 && r2 == 1 && r3 == 0 && r4 == 907 (r1 == 1 && r2 == 1 && r3 == 0 && r4 == 1)
908 908
909 cannot happen. 909 cannot happen.
910 910
911 This non-requirement was also non-premeditated 911 This non-requirement was also non-premeditated, but became apparent when
912 studying RCU's interaction with memory orderin 912 studying RCU's interaction with memory ordering.
913 913
914 Read-Side Critical Sections Don't Partition Gr 914 Read-Side Critical Sections Don't Partition Grace Periods
915 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 915 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
916 916
917 It is also tempting to assume that if an RCU r 917 It is also tempting to assume that if an RCU read-side critical section
918 happens between a pair of grace periods, then 918 happens between a pair of grace periods, then those grace periods cannot
919 overlap. However, this temptation leads nowher 919 overlap. However, this temptation leads nowhere good, as can be
920 illustrated by the following, with all variabl 920 illustrated by the following, with all variables initially zero:
921 921
922 :: 922 ::
923 923
924 1 void thread0(void) 924 1 void thread0(void)
925 2 { 925 2 {
926 3 rcu_read_lock(); 926 3 rcu_read_lock();
927 4 WRITE_ONCE(a, 1); 927 4 WRITE_ONCE(a, 1);
928 5 WRITE_ONCE(b, 1); 928 5 WRITE_ONCE(b, 1);
929 6 rcu_read_unlock(); 929 6 rcu_read_unlock();
930 7 } 930 7 }
931 8 931 8
932 9 void thread1(void) 932 9 void thread1(void)
933 10 { 933 10 {
934 11 r1 = READ_ONCE(a); 934 11 r1 = READ_ONCE(a);
935 12 synchronize_rcu(); 935 12 synchronize_rcu();
936 13 WRITE_ONCE(c, 1); 936 13 WRITE_ONCE(c, 1);
937 14 } 937 14 }
938 15 938 15
939 16 void thread2(void) 939 16 void thread2(void)
940 17 { 940 17 {
941 18 rcu_read_lock(); 941 18 rcu_read_lock();
942 19 WRITE_ONCE(d, 1); 942 19 WRITE_ONCE(d, 1);
943 20 r2 = READ_ONCE(c); 943 20 r2 = READ_ONCE(c);
944 21 rcu_read_unlock(); 944 21 rcu_read_unlock();
945 22 } 945 22 }
946 23 946 23
947 24 void thread3(void) 947 24 void thread3(void)
948 25 { 948 25 {
949 26 r3 = READ_ONCE(d); 949 26 r3 = READ_ONCE(d);
950 27 synchronize_rcu(); 950 27 synchronize_rcu();
951 28 WRITE_ONCE(e, 1); 951 28 WRITE_ONCE(e, 1);
952 29 } 952 29 }
953 30 953 30
954 31 void thread4(void) 954 31 void thread4(void)
955 32 { 955 32 {
956 33 rcu_read_lock(); 956 33 rcu_read_lock();
957 34 r4 = READ_ONCE(b); 957 34 r4 = READ_ONCE(b);
958 35 r5 = READ_ONCE(e); 958 35 r5 = READ_ONCE(e);
959 36 rcu_read_unlock(); 959 36 rcu_read_unlock();
960 37 } 960 37 }
961 961
962 In this case, the outcome: 962 In this case, the outcome:
963 963
964 :: 964 ::
965 965
966 (r1 == 1 && r2 == 1 && r3 == 1 && r4 == 966 (r1 == 1 && r2 == 1 && r3 == 1 && r4 == 0 && r5 == 1)
967 967
968 is entirely possible, as illustrated below: 968 is entirely possible, as illustrated below:
969 969
970 .. kernel-figure:: ReadersPartitionGP1.svg 970 .. kernel-figure:: ReadersPartitionGP1.svg
971 971
972 Again, an RCU read-side critical section can o 972 Again, an RCU read-side critical section can overlap almost all of a
973 given grace period, just so long as it does no 973 given grace period, just so long as it does not overlap the entire grace
974 period. As a result, an RCU read-side critical 974 period. As a result, an RCU read-side critical section cannot partition
975 a pair of RCU grace periods. 975 a pair of RCU grace periods.
976 976
977 +--------------------------------------------- 977 +-----------------------------------------------------------------------+
978 | **Quick Quiz**: 978 | **Quick Quiz**: |
979 +--------------------------------------------- 979 +-----------------------------------------------------------------------+
980 | How long a sequence of grace periods, each s 980 | How long a sequence of grace periods, each separated by an RCU |
981 | read-side critical section, would be require 981 | read-side critical section, would be required to partition the RCU |
982 | read-side critical sections at the beginning 982 | read-side critical sections at the beginning and end of the chain? |
983 +--------------------------------------------- 983 +-----------------------------------------------------------------------+
984 | **Answer**: 984 | **Answer**: |
985 +--------------------------------------------- 985 +-----------------------------------------------------------------------+
986 | In theory, an infinite number. In practice, 986 | In theory, an infinite number. In practice, an unknown number that is |
987 | sensitive to both implementation details and 987 | sensitive to both implementation details and timing considerations. |
988 | Therefore, even in practice, RCU users must 988 | Therefore, even in practice, RCU users must abide by the theoretical |
989 | rather than the practical answer. 989 | rather than the practical answer. |
990 +--------------------------------------------- 990 +-----------------------------------------------------------------------+
991 991
992 Parallelism Facts of Life 992 Parallelism Facts of Life
993 ------------------------- 993 -------------------------
994 994
995 These parallelism facts of life are by no mean 995 These parallelism facts of life are by no means specific to RCU, but the
996 RCU implementation must abide by them. They th 996 RCU implementation must abide by them. They therefore bear repeating:
997 997
998 #. Any CPU or task may be delayed at any time, 998 #. Any CPU or task may be delayed at any time, and any attempts to avoid
999 these delays by disabling preemption, inter 999 these delays by disabling preemption, interrupts, or whatever are
1000 completely futile. This is most obvious in 1000 completely futile. This is most obvious in preemptible user-level
1001 environments and in virtualized environmen 1001 environments and in virtualized environments (where a given guest
1002 OS's VCPUs can be preempted at any time by 1002 OS's VCPUs can be preempted at any time by the underlying
1003 hypervisor), but can also happen in bare-m 1003 hypervisor), but can also happen in bare-metal environments due to
1004 ECC errors, NMIs, and other hardware event 1004 ECC errors, NMIs, and other hardware events. Although a delay of more
1005 than about 20 seconds can result in splats 1005 than about 20 seconds can result in splats, the RCU implementation is
1006 obligated to use algorithms that can toler 1006 obligated to use algorithms that can tolerate extremely long delays,
1007 but where “extremely long” is not long 1007 but where “extremely long” is not long enough to allow wrap-around
1008 when incrementing a 64-bit counter. 1008 when incrementing a 64-bit counter.
1009 #. Both the compiler and the CPU can reorder 1009 #. Both the compiler and the CPU can reorder memory accesses. Where it
1010 matters, RCU must use compiler directives 1010 matters, RCU must use compiler directives and memory-barrier
1011 instructions to preserve ordering. 1011 instructions to preserve ordering.
1012 #. Conflicting writes to memory locations in 1012 #. Conflicting writes to memory locations in any given cache line will
1013 result in expensive cache misses. Greater 1013 result in expensive cache misses. Greater numbers of concurrent
1014 writes and more-frequent concurrent writes 1014 writes and more-frequent concurrent writes will result in more
1015 dramatic slowdowns. RCU is therefore oblig 1015 dramatic slowdowns. RCU is therefore obligated to use algorithms that
1016 have sufficient locality to avoid signific 1016 have sufficient locality to avoid significant performance and
1017 scalability problems. 1017 scalability problems.
1018 #. As a rough rule of thumb, only one CPU's w 1018 #. As a rough rule of thumb, only one CPU's worth of processing may be
1019 carried out under the protection of any gi 1019 carried out under the protection of any given exclusive lock. RCU
1020 must therefore use scalable locking design 1020 must therefore use scalable locking designs.
1021 #. Counters are finite, especially on 32-bit 1021 #. Counters are finite, especially on 32-bit systems. RCU's use of
1022 counters must therefore tolerate counter w 1022 counters must therefore tolerate counter wrap, or be designed such
1023 that counter wrap would take way more time 1023 that counter wrap would take way more time than a single system is
1024 likely to run. An uptime of ten years is q 1024 likely to run. An uptime of ten years is quite possible, a runtime of
1025 a century much less so. As an example of t 1025 a century much less so. As an example of the latter, RCU's
1026 dyntick-idle nesting counter allows 54 bit 1026 dyntick-idle nesting counter allows 54 bits for interrupt nesting
1027 level (this counter is 64 bits even on a 3 1027 level (this counter is 64 bits even on a 32-bit system). Overflowing
1028 this counter requires 2\ :sup:`54` half-in 1028 this counter requires 2\ :sup:`54` half-interrupts on a given CPU
1029 without that CPU ever going idle. If a hal 1029 without that CPU ever going idle. If a half-interrupt happened every
1030 microsecond, it would take 570 years of ru 1030 microsecond, it would take 570 years of runtime to overflow this
1031 counter, which is currently believed to be 1031 counter, which is currently believed to be an acceptably long time.
1032 #. Linux systems can have thousands of CPUs r 1032 #. Linux systems can have thousands of CPUs running a single Linux
1033 kernel in a single shared-memory environme 1033 kernel in a single shared-memory environment. RCU must therefore pay
1034 close attention to high-end scalability. 1034 close attention to high-end scalability.
1035 1035
1036 This last parallelism fact of life means that 1036 This last parallelism fact of life means that RCU must pay special
1037 attention to the preceding facts of life. The 1037 attention to the preceding facts of life. The idea that Linux might
1038 scale to systems with thousands of CPUs would 1038 scale to systems with thousands of CPUs would have been met with some
1039 skepticism in the 1990s, but these requiremen 1039 skepticism in the 1990s, but these requirements would have otherwise
1040 have been unsurprising, even in the early 199 1040 have been unsurprising, even in the early 1990s.
1041 1041
1042 Quality-of-Implementation Requirements 1042 Quality-of-Implementation Requirements
1043 -------------------------------------- 1043 --------------------------------------
1044 1044
1045 These sections list quality-of-implementation 1045 These sections list quality-of-implementation requirements. Although an
1046 RCU implementation that ignores these require 1046 RCU implementation that ignores these requirements could still be used,
1047 it would likely be subject to limitations tha 1047 it would likely be subject to limitations that would make it
1048 inappropriate for industrial-strength product 1048 inappropriate for industrial-strength production use. Classes of
1049 quality-of-implementation requirements are as 1049 quality-of-implementation requirements are as follows:
1050 1050
1051 #. `Specialization`_ 1051 #. `Specialization`_
1052 #. `Performance and Scalability`_ 1052 #. `Performance and Scalability`_
1053 #. `Forward Progress`_ 1053 #. `Forward Progress`_
1054 #. `Composability`_ 1054 #. `Composability`_
1055 #. `Corner Cases`_ 1055 #. `Corner Cases`_
1056 1056
1057 These classes is covered in the following sec 1057 These classes is covered in the following sections.
1058 1058
1059 Specialization 1059 Specialization
1060 ~~~~~~~~~~~~~~ 1060 ~~~~~~~~~~~~~~
1061 1061
1062 RCU is and always has been intended primarily 1062 RCU is and always has been intended primarily for read-mostly
1063 situations, which means that RCU's read-side 1063 situations, which means that RCU's read-side primitives are optimized,
1064 often at the expense of its update-side primi 1064 often at the expense of its update-side primitives. Experience thus far
1065 is captured by the following list of situatio 1065 is captured by the following list of situations:
1066 1066
1067 #. Read-mostly data, where stale and inconsis 1067 #. Read-mostly data, where stale and inconsistent data is not a problem:
1068 RCU works great! 1068 RCU works great!
1069 #. Read-mostly data, where data must be consi 1069 #. Read-mostly data, where data must be consistent: RCU works well.
1070 #. Read-write data, where data must be consis 1070 #. Read-write data, where data must be consistent: RCU *might* work OK.
1071 Or not. 1071 Or not.
1072 #. Write-mostly data, where data must be cons 1072 #. Write-mostly data, where data must be consistent: RCU is very
1073 unlikely to be the right tool for the job, 1073 unlikely to be the right tool for the job, with the following
1074 exceptions, where RCU can provide: 1074 exceptions, where RCU can provide:
1075 1075
1076 a. Existence guarantees for update-friendl 1076 a. Existence guarantees for update-friendly mechanisms.
1077 b. Wait-free read-side primitives for real 1077 b. Wait-free read-side primitives for real-time use.
1078 1078
1079 This focus on read-mostly situations means th 1079 This focus on read-mostly situations means that RCU must interoperate
1080 with other synchronization primitives. For ex 1080 with other synchronization primitives. For example, the add_gp() and
1081 remove_gp_synchronous() examples discussed ea 1081 remove_gp_synchronous() examples discussed earlier use RCU to
1082 protect readers and locking to coordinate upd 1082 protect readers and locking to coordinate updaters. However, the need
1083 extends much farther, requiring that a variet 1083 extends much farther, requiring that a variety of synchronization
1084 primitives be legal within RCU read-side crit 1084 primitives be legal within RCU read-side critical sections, including
1085 spinlocks, sequence locks, atomic operations, 1085 spinlocks, sequence locks, atomic operations, reference counters, and
1086 memory barriers. 1086 memory barriers.
1087 1087
1088 +-------------------------------------------- 1088 +-----------------------------------------------------------------------+
1089 | **Quick Quiz**: 1089 | **Quick Quiz**: |
1090 +-------------------------------------------- 1090 +-----------------------------------------------------------------------+
1091 | What about sleeping locks? 1091 | What about sleeping locks? |
1092 +-------------------------------------------- 1092 +-----------------------------------------------------------------------+
1093 | **Answer**: 1093 | **Answer**: |
1094 +-------------------------------------------- 1094 +-----------------------------------------------------------------------+
1095 | These are forbidden within Linux-kernel RCU 1095 | These are forbidden within Linux-kernel RCU read-side critical |
1096 | sections because it is not legal to place a 1096 | sections because it is not legal to place a quiescent state (in this |
1097 | case, voluntary context switch) within an R 1097 | case, voluntary context switch) within an RCU read-side critical |
1098 | section. However, sleeping locks may be use 1098 | section. However, sleeping locks may be used within userspace RCU |
1099 | read-side critical sections, and also withi 1099 | read-side critical sections, and also within Linux-kernel sleepable |
1100 | RCU `(SRCU) <Sleepable RCU_>`__ read-side c 1100 | RCU `(SRCU) <Sleepable RCU_>`__ read-side critical sections. In |
1101 | addition, the -rt patchset turns spinlocks 1101 | addition, the -rt patchset turns spinlocks into a sleeping locks so |
1102 | that the corresponding critical sections ca 1102 | that the corresponding critical sections can be preempted, which also |
1103 | means that these sleeplockified spinlocks ( 1103 | means that these sleeplockified spinlocks (but not other sleeping |
1104 | locks!) may be acquire within -rt-Linux-ker 1104 | locks!) may be acquire within -rt-Linux-kernel RCU read-side critical |
1105 | sections. 1105 | sections. |
1106 | Note that it *is* legal for a normal RCU re 1106 | Note that it *is* legal for a normal RCU read-side critical section |
1107 | to conditionally acquire a sleeping locks ( 1107 | to conditionally acquire a sleeping locks (as in |
1108 | mutex_trylock()), but only as long as it do 1108 | mutex_trylock()), but only as long as it does not loop |
1109 | indefinitely attempting to conditionally ac 1109 | indefinitely attempting to conditionally acquire that sleeping locks. |
1110 | The key point is that things like mutex_try 1110 | The key point is that things like mutex_trylock() either return |
1111 | with the mutex held, or return an error ind 1111 | with the mutex held, or return an error indication if the mutex was |
1112 | not immediately available. Either way, mute 1112 | not immediately available. Either way, mutex_trylock() returns |
1113 | immediately without sleeping. 1113 | immediately without sleeping. |
1114 +-------------------------------------------- 1114 +-----------------------------------------------------------------------+
1115 1115
1116 It often comes as a surprise that many algori 1116 It often comes as a surprise that many algorithms do not require a
1117 consistent view of data, but many can functio 1117 consistent view of data, but many can function in that mode, with
1118 network routing being the poster child. Inter 1118 network routing being the poster child. Internet routing algorithms take
1119 significant time to propagate updates, so tha 1119 significant time to propagate updates, so that by the time an update
1120 arrives at a given system, that system has be 1120 arrives at a given system, that system has been sending network traffic
1121 the wrong way for a considerable length of ti 1121 the wrong way for a considerable length of time. Having a few threads
1122 continue to send traffic the wrong way for a 1122 continue to send traffic the wrong way for a few more milliseconds is
1123 clearly not a problem: In the worst case, TCP 1123 clearly not a problem: In the worst case, TCP retransmissions will
1124 eventually get the data where it needs to go. 1124 eventually get the data where it needs to go. In general, when tracking
1125 the state of the universe outside of the comp 1125 the state of the universe outside of the computer, some level of
1126 inconsistency must be tolerated due to speed- 1126 inconsistency must be tolerated due to speed-of-light delays if nothing
1127 else. 1127 else.
1128 1128
1129 Furthermore, uncertainty about external state 1129 Furthermore, uncertainty about external state is inherent in many cases.
1130 For example, a pair of veterinarians might us 1130 For example, a pair of veterinarians might use heartbeat to determine
1131 whether or not a given cat was alive. But how 1131 whether or not a given cat was alive. But how long should they wait
1132 after the last heartbeat to decide that the c 1132 after the last heartbeat to decide that the cat is in fact dead? Waiting
1133 less than 400 milliseconds makes no sense bec 1133 less than 400 milliseconds makes no sense because this would mean that a
1134 relaxed cat would be considered to cycle betw 1134 relaxed cat would be considered to cycle between death and life more
1135 than 100 times per minute. Moreover, just as 1135 than 100 times per minute. Moreover, just as with human beings, a cat's
1136 heart might stop for some period of time, so 1136 heart might stop for some period of time, so the exact wait period is a
1137 judgment call. One of our pair of veterinaria 1137 judgment call. One of our pair of veterinarians might wait 30 seconds
1138 before pronouncing the cat dead, while the ot 1138 before pronouncing the cat dead, while the other might insist on waiting
1139 a full minute. The two veterinarians would th 1139 a full minute. The two veterinarians would then disagree on the state of
1140 the cat during the final 30 seconds of the mi 1140 the cat during the final 30 seconds of the minute following the last
1141 heartbeat. 1141 heartbeat.
1142 1142
1143 Interestingly enough, this same situation app 1143 Interestingly enough, this same situation applies to hardware. When push
1144 comes to shove, how do we tell whether or not 1144 comes to shove, how do we tell whether or not some external server has
1145 failed? We send messages to it periodically, 1145 failed? We send messages to it periodically, and declare it failed if we
1146 don't receive a response within a given perio 1146 don't receive a response within a given period of time. Policy decisions
1147 can usually tolerate short periods of inconsi 1147 can usually tolerate short periods of inconsistency. The policy was
1148 decided some time ago, and is only now being 1148 decided some time ago, and is only now being put into effect, so a few
1149 milliseconds of delay is normally inconsequen 1149 milliseconds of delay is normally inconsequential.
1150 1150
1151 However, there are algorithms that absolutely 1151 However, there are algorithms that absolutely must see consistent data.
1152 For example, the translation between a user-l 1152 For example, the translation between a user-level SystemV semaphore ID
1153 to the corresponding in-kernel data structure 1153 to the corresponding in-kernel data structure is protected by RCU, but
1154 it is absolutely forbidden to update a semaph 1154 it is absolutely forbidden to update a semaphore that has just been
1155 removed. In the Linux kernel, this need for c 1155 removed. In the Linux kernel, this need for consistency is accommodated
1156 by acquiring spinlocks located in the in-kern 1156 by acquiring spinlocks located in the in-kernel data structure from
1157 within the RCU read-side critical section, an 1157 within the RCU read-side critical section, and this is indicated by the
1158 green box in the figure above. Many other tec 1158 green box in the figure above. Many other techniques may be used, and
1159 are in fact used within the Linux kernel. 1159 are in fact used within the Linux kernel.
1160 1160
1161 In short, RCU is not required to maintain con 1161 In short, RCU is not required to maintain consistency, and other
1162 mechanisms may be used in concert with RCU wh 1162 mechanisms may be used in concert with RCU when consistency is required.
1163 RCU's specialization allows it to do its job 1163 RCU's specialization allows it to do its job extremely well, and its
1164 ability to interoperate with other synchroniz 1164 ability to interoperate with other synchronization mechanisms allows the
1165 right mix of synchronization tools to be used 1165 right mix of synchronization tools to be used for a given job.
1166 1166
1167 Performance and Scalability 1167 Performance and Scalability
1168 ~~~~~~~~~~~~~~~~~~~~~~~~~~~ 1168 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
1169 1169
1170 Energy efficiency is a critical component of 1170 Energy efficiency is a critical component of performance today, and
1171 Linux-kernel RCU implementations must therefo 1171 Linux-kernel RCU implementations must therefore avoid unnecessarily
1172 awakening idle CPUs. I cannot claim that this 1172 awakening idle CPUs. I cannot claim that this requirement was
1173 premeditated. In fact, I learned of it during 1173 premeditated. In fact, I learned of it during a telephone conversation
1174 in which I was given “frank and open” fee 1174 in which I was given “frank and open” feedback on the importance of
1175 energy efficiency in battery-powered systems 1175 energy efficiency in battery-powered systems and on specific
1176 energy-efficiency shortcomings of the Linux-k 1176 energy-efficiency shortcomings of the Linux-kernel RCU implementation.
1177 In my experience, the battery-powered embedde 1177 In my experience, the battery-powered embedded community will consider
1178 any unnecessary wakeups to be extremely unfri 1178 any unnecessary wakeups to be extremely unfriendly acts. So much so that
1179 mere Linux-kernel-mailing-list posts are insu 1179 mere Linux-kernel-mailing-list posts are insufficient to vent their ire.
1180 1180
1181 Memory consumption is not particularly import 1181 Memory consumption is not particularly important for in most situations,
1182 and has become decreasingly so as memory size 1182 and has become decreasingly so as memory sizes have expanded and memory
1183 costs have plummeted. However, as I learned f 1183 costs have plummeted. However, as I learned from Matt Mackall's
1184 `bloatwatch <http://elinux.org/Linux_Tiny-FAQ 1184 `bloatwatch <http://elinux.org/Linux_Tiny-FAQ>`__ efforts, memory
1185 footprint is critically important on single-C 1185 footprint is critically important on single-CPU systems with
1186 non-preemptible (``CONFIG_PREEMPTION=n``) ker 1186 non-preemptible (``CONFIG_PREEMPTION=n``) kernels, and thus `tiny
1187 RCU <https://lore.kernel.org/r/20090113221724. 1187 RCU <https://lore.kernel.org/r/20090113221724.GA15307@linux.vnet.ibm.com">https://lore.kernel.org/r/20090113221724.GA15307@linux.vnet.ibm.com>`__
1188 was born. Josh Triplett has since taken over 1188 was born. Josh Triplett has since taken over the small-memory banner
1189 with his `Linux kernel tinification <https:// 1189 with his `Linux kernel tinification <https://tiny.wiki.kernel.org/>`__
1190 project, which resulted in `SRCU <Sleepable R 1190 project, which resulted in `SRCU <Sleepable RCU_>`__ becoming optional
1191 for those kernels not needing it. 1191 for those kernels not needing it.
1192 1192
1193 The remaining performance requirements are, f 1193 The remaining performance requirements are, for the most part,
1194 unsurprising. For example, in keeping with RC 1194 unsurprising. For example, in keeping with RCU's read-side
1195 specialization, rcu_dereference() should have 1195 specialization, rcu_dereference() should have negligible overhead
1196 (for example, suppression of a few minor comp 1196 (for example, suppression of a few minor compiler optimizations).
1197 Similarly, in non-preemptible environments, r 1197 Similarly, in non-preemptible environments, rcu_read_lock() and
1198 rcu_read_unlock() should have exactly zero ov 1198 rcu_read_unlock() should have exactly zero overhead.
1199 1199
1200 In preemptible environments, in the case wher 1200 In preemptible environments, in the case where the RCU read-side
1201 critical section was not preempted (as will b 1201 critical section was not preempted (as will be the case for the
1202 highest-priority real-time process), rcu_read 1202 highest-priority real-time process), rcu_read_lock() and
1203 rcu_read_unlock() should have minimal overhea 1203 rcu_read_unlock() should have minimal overhead. In particular, they
1204 should not contain atomic read-modify-write o 1204 should not contain atomic read-modify-write operations, memory-barrier
1205 instructions, preemption disabling, interrupt 1205 instructions, preemption disabling, interrupt disabling, or backwards
1206 branches. However, in the case where the RCU 1206 branches. However, in the case where the RCU read-side critical section
1207 was preempted, rcu_read_unlock() may acquire 1207 was preempted, rcu_read_unlock() may acquire spinlocks and disable
1208 interrupts. This is why it is better to nest 1208 interrupts. This is why it is better to nest an RCU read-side critical
1209 section within a preempt-disable region than 1209 section within a preempt-disable region than vice versa, at least in
1210 cases where that critical section is short en 1210 cases where that critical section is short enough to avoid unduly
1211 degrading real-time latencies. 1211 degrading real-time latencies.
1212 1212
1213 The synchronize_rcu() grace-period-wait primi 1213 The synchronize_rcu() grace-period-wait primitive is optimized for
1214 throughput. It may therefore incur several mi 1214 throughput. It may therefore incur several milliseconds of latency in
1215 addition to the duration of the longest RCU r 1215 addition to the duration of the longest RCU read-side critical section.
1216 On the other hand, multiple concurrent invoca 1216 On the other hand, multiple concurrent invocations of
1217 synchronize_rcu() are required to use batchin 1217 synchronize_rcu() are required to use batching optimizations so that
1218 they can be satisfied by a single underlying 1218 they can be satisfied by a single underlying grace-period-wait
1219 operation. For example, in the Linux kernel, 1219 operation. For example, in the Linux kernel, it is not unusual for a
1220 single grace-period-wait operation to serve m 1220 single grace-period-wait operation to serve more than `1,000 separate
1221 invocations <https://www.usenix.org/conferenc 1221 invocations <https://www.usenix.org/conference/2004-usenix-annual-technical-conference/making-rcu-safe-deep-sub-millisecond-response>`__
1222 of synchronize_rcu(), thus amortizing the per 1222 of synchronize_rcu(), thus amortizing the per-invocation overhead
1223 down to nearly zero. However, the grace-perio 1223 down to nearly zero. However, the grace-period optimization is also
1224 required to avoid measurable degradation of r 1224 required to avoid measurable degradation of real-time scheduling and
1225 interrupt latencies. 1225 interrupt latencies.
1226 1226
1227 In some cases, the multi-millisecond synchron 1227 In some cases, the multi-millisecond synchronize_rcu() latencies are
1228 unacceptable. In these cases, synchronize_rcu 1228 unacceptable. In these cases, synchronize_rcu_expedited() may be
1229 used instead, reducing the grace-period laten 1229 used instead, reducing the grace-period latency down to a few tens of
1230 microseconds on small systems, at least in ca 1230 microseconds on small systems, at least in cases where the RCU read-side
1231 critical sections are short. There are curren 1231 critical sections are short. There are currently no special latency
1232 requirements for synchronize_rcu_expedited() 1232 requirements for synchronize_rcu_expedited() on large systems, but,
1233 consistent with the empirical nature of the R 1233 consistent with the empirical nature of the RCU specification, that is
1234 subject to change. However, there most defini 1234 subject to change. However, there most definitely are scalability
1235 requirements: A storm of synchronize_rcu_expe 1235 requirements: A storm of synchronize_rcu_expedited() invocations on
1236 4096 CPUs should at least make reasonable for 1236 4096 CPUs should at least make reasonable forward progress. In return
1237 for its shorter latencies, synchronize_rcu_ex 1237 for its shorter latencies, synchronize_rcu_expedited() is permitted
1238 to impose modest degradation of real-time lat 1238 to impose modest degradation of real-time latency on non-idle online
1239 CPUs. Here, “modest” means roughly the sa 1239 CPUs. Here, “modest” means roughly the same latency degradation as a
1240 scheduling-clock interrupt. 1240 scheduling-clock interrupt.
1241 1241
1242 There are a number of situations where even 1242 There are a number of situations where even
1243 synchronize_rcu_expedited()'s reduced grace-p 1243 synchronize_rcu_expedited()'s reduced grace-period latency is
1244 unacceptable. In these situations, the asynch 1244 unacceptable. In these situations, the asynchronous call_rcu() can
1245 be used in place of synchronize_rcu() as foll 1245 be used in place of synchronize_rcu() as follows:
1246 1246
1247 :: 1247 ::
1248 1248
1249 1 struct foo { 1249 1 struct foo {
1250 2 int a; 1250 2 int a;
1251 3 int b; 1251 3 int b;
1252 4 struct rcu_head rh; 1252 4 struct rcu_head rh;
1253 5 }; 1253 5 };
1254 6 1254 6
1255 7 static void remove_gp_cb(struct rcu_ 1255 7 static void remove_gp_cb(struct rcu_head *rhp)
1256 8 { 1256 8 {
1257 9 struct foo *p = container_of(rhp, 1257 9 struct foo *p = container_of(rhp, struct foo, rh);
1258 10 1258 10
1259 11 kfree(p); 1259 11 kfree(p);
1260 12 } 1260 12 }
1261 13 1261 13
1262 14 bool remove_gp_asynchronous(void) 1262 14 bool remove_gp_asynchronous(void)
1263 15 { 1263 15 {
1264 16 struct foo *p; 1264 16 struct foo *p;
1265 17 1265 17
1266 18 spin_lock(&gp_lock); 1266 18 spin_lock(&gp_lock);
1267 19 p = rcu_access_pointer(gp); 1267 19 p = rcu_access_pointer(gp);
1268 20 if (!p) { 1268 20 if (!p) {
1269 21 spin_unlock(&gp_lock); 1269 21 spin_unlock(&gp_lock);
1270 22 return false; 1270 22 return false;
1271 23 } 1271 23 }
1272 24 rcu_assign_pointer(gp, NULL); 1272 24 rcu_assign_pointer(gp, NULL);
1273 25 call_rcu(&p->rh, remove_gp_cb); 1273 25 call_rcu(&p->rh, remove_gp_cb);
1274 26 spin_unlock(&gp_lock); 1274 26 spin_unlock(&gp_lock);
1275 27 return true; 1275 27 return true;
1276 28 } 1276 28 }
1277 1277
1278 A definition of ``struct foo`` is finally nee 1278 A definition of ``struct foo`` is finally needed, and appears on
1279 lines 1-5. The function remove_gp_cb() is pa 1279 lines 1-5. The function remove_gp_cb() is passed to call_rcu()
1280 on line 25, and will be invoked after the en 1280 on line 25, and will be invoked after the end of a subsequent grace
1281 period. This gets the same effect as remove_g 1281 period. This gets the same effect as remove_gp_synchronous(), but
1282 without forcing the updater to wait for a gra 1282 without forcing the updater to wait for a grace period to elapse. The
1283 call_rcu() function may be used in a number o 1283 call_rcu() function may be used in a number of situations where
1284 neither synchronize_rcu() nor synchronize_rcu 1284 neither synchronize_rcu() nor synchronize_rcu_expedited() would
1285 be legal, including within preempt-disable co 1285 be legal, including within preempt-disable code, local_bh_disable()
1286 code, interrupt-disable code, and interrupt h 1286 code, interrupt-disable code, and interrupt handlers. However, even
1287 call_rcu() is illegal within NMI handlers and 1287 call_rcu() is illegal within NMI handlers and from idle and offline
1288 CPUs. The callback function (remove_gp_cb() i 1288 CPUs. The callback function (remove_gp_cb() in this case) will be
1289 executed within softirq (software interrupt) 1289 executed within softirq (software interrupt) environment within the
1290 Linux kernel, either within a real softirq ha 1290 Linux kernel, either within a real softirq handler or under the
1291 protection of local_bh_disable(). In both the 1291 protection of local_bh_disable(). In both the Linux kernel and in
1292 userspace, it is bad practice to write an RCU 1292 userspace, it is bad practice to write an RCU callback function that
1293 takes too long. Long-running operations shoul 1293 takes too long. Long-running operations should be relegated to separate
1294 threads or (in the Linux kernel) workqueues. 1294 threads or (in the Linux kernel) workqueues.
1295 1295
1296 +-------------------------------------------- 1296 +-----------------------------------------------------------------------+
1297 | **Quick Quiz**: 1297 | **Quick Quiz**: |
1298 +-------------------------------------------- 1298 +-----------------------------------------------------------------------+
1299 | Why does line 19 use rcu_access_pointer()? 1299 | Why does line 19 use rcu_access_pointer()? After all, |
1300 | call_rcu() on line 25 stores into the stru 1300 | call_rcu() on line 25 stores into the structure, which would |
1301 | interact badly with concurrent insertions. 1301 | interact badly with concurrent insertions. Doesn't this mean that |
1302 | rcu_dereference() is required? 1302 | rcu_dereference() is required? |
1303 +-------------------------------------------- 1303 +-----------------------------------------------------------------------+
1304 | **Answer**: 1304 | **Answer**: |
1305 +-------------------------------------------- 1305 +-----------------------------------------------------------------------+
1306 | Presumably the ``->gp_lock`` acquired on li 1306 | Presumably the ``->gp_lock`` acquired on line 18 excludes any |
1307 | changes, including any insertions that rcu_ 1307 | changes, including any insertions that rcu_dereference() would |
1308 | protect against. Therefore, any insertions 1308 | protect against. Therefore, any insertions will be delayed until |
1309 | after ``->gp_lock`` is released on line 25 1309 | after ``->gp_lock`` is released on line 25, which in turn means that |
1310 | rcu_access_pointer() suffices. 1310 | rcu_access_pointer() suffices. |
1311 +-------------------------------------------- 1311 +-----------------------------------------------------------------------+
1312 1312
1313 However, all that remove_gp_cb() is doing is 1313 However, all that remove_gp_cb() is doing is invoking kfree() on
1314 the data element. This is a common idiom, and 1314 the data element. This is a common idiom, and is supported by
1315 kfree_rcu(), which allows “fire and forget 1315 kfree_rcu(), which allows “fire and forget” operation as shown
1316 below: 1316 below:
1317 1317
1318 :: 1318 ::
1319 1319
1320 1 struct foo { 1320 1 struct foo {
1321 2 int a; 1321 2 int a;
1322 3 int b; 1322 3 int b;
1323 4 struct rcu_head rh; 1323 4 struct rcu_head rh;
1324 5 }; 1324 5 };
1325 6 1325 6
1326 7 bool remove_gp_faf(void) 1326 7 bool remove_gp_faf(void)
1327 8 { 1327 8 {
1328 9 struct foo *p; 1328 9 struct foo *p;
1329 10 1329 10
1330 11 spin_lock(&gp_lock); 1330 11 spin_lock(&gp_lock);
1331 12 p = rcu_dereference(gp); 1331 12 p = rcu_dereference(gp);
1332 13 if (!p) { 1332 13 if (!p) {
1333 14 spin_unlock(&gp_lock); 1333 14 spin_unlock(&gp_lock);
1334 15 return false; 1334 15 return false;
1335 16 } 1335 16 }
1336 17 rcu_assign_pointer(gp, NULL); 1336 17 rcu_assign_pointer(gp, NULL);
1337 18 kfree_rcu(p, rh); 1337 18 kfree_rcu(p, rh);
1338 19 spin_unlock(&gp_lock); 1338 19 spin_unlock(&gp_lock);
1339 20 return true; 1339 20 return true;
1340 21 } 1340 21 }
1341 1341
1342 Note that remove_gp_faf() simply invokes kfre 1342 Note that remove_gp_faf() simply invokes kfree_rcu() and
1343 proceeds, without any need to pay any further 1343 proceeds, without any need to pay any further attention to the
1344 subsequent grace period and kfree(). It is pe 1344 subsequent grace period and kfree(). It is permissible to invoke
1345 kfree_rcu() from the same environments as for 1345 kfree_rcu() from the same environments as for call_rcu().
1346 Interestingly enough, DYNIX/ptx had the equiv 1346 Interestingly enough, DYNIX/ptx had the equivalents of call_rcu()
1347 and kfree_rcu(), but not synchronize_rcu(). T 1347 and kfree_rcu(), but not synchronize_rcu(). This was due to the
1348 fact that RCU was not heavily used within DYN 1348 fact that RCU was not heavily used within DYNIX/ptx, so the very few
1349 places that needed something like synchronize 1349 places that needed something like synchronize_rcu() simply
1350 open-coded it. 1350 open-coded it.
1351 1351
1352 +-------------------------------------------- 1352 +-----------------------------------------------------------------------+
1353 | **Quick Quiz**: 1353 | **Quick Quiz**: |
1354 +-------------------------------------------- 1354 +-----------------------------------------------------------------------+
1355 | Earlier it was claimed that call_rcu() and 1355 | Earlier it was claimed that call_rcu() and kfree_rcu() |
1356 | allowed updaters to avoid being blocked by 1356 | allowed updaters to avoid being blocked by readers. But how can that |
1357 | be correct, given that the invocation of th 1357 | be correct, given that the invocation of the callback and the freeing |
1358 | of the memory (respectively) must still wai 1358 | of the memory (respectively) must still wait for a grace period to |
1359 | elapse? 1359 | elapse? |
1360 +-------------------------------------------- 1360 +-----------------------------------------------------------------------+
1361 | **Answer**: 1361 | **Answer**: |
1362 +-------------------------------------------- 1362 +-----------------------------------------------------------------------+
1363 | We could define things this way, but keep i 1363 | We could define things this way, but keep in mind that this sort of |
1364 | definition would say that updates in garbag 1364 | definition would say that updates in garbage-collected languages |
1365 | cannot complete until the next time the gar 1365 | cannot complete until the next time the garbage collector runs, which |
1366 | does not seem at all reasonable. The key po 1366 | does not seem at all reasonable. The key point is that in most cases, |
1367 | an updater using either call_rcu() or kfree 1367 | an updater using either call_rcu() or kfree_rcu() can proceed |
1368 | to the next update as soon as it has invoke 1368 | to the next update as soon as it has invoked call_rcu() or |
1369 | kfree_rcu(), without having to wait for a s 1369 | kfree_rcu(), without having to wait for a subsequent grace |
1370 | period. 1370 | period. |
1371 +-------------------------------------------- 1371 +-----------------------------------------------------------------------+
1372 1372
1373 But what if the updater must wait for the com 1373 But what if the updater must wait for the completion of code to be
1374 executed after the end of the grace period, b 1374 executed after the end of the grace period, but has other tasks that can
1375 be carried out in the meantime? The polling-s 1375 be carried out in the meantime? The polling-style
1376 get_state_synchronize_rcu() and cond_synchron 1376 get_state_synchronize_rcu() and cond_synchronize_rcu() functions
1377 may be used for this purpose, as shown below: 1377 may be used for this purpose, as shown below:
1378 1378
1379 :: 1379 ::
1380 1380
1381 1 bool remove_gp_poll(void) 1381 1 bool remove_gp_poll(void)
1382 2 { 1382 2 {
1383 3 struct foo *p; 1383 3 struct foo *p;
1384 4 unsigned long s; 1384 4 unsigned long s;
1385 5 1385 5
1386 6 spin_lock(&gp_lock); 1386 6 spin_lock(&gp_lock);
1387 7 p = rcu_access_pointer(gp); 1387 7 p = rcu_access_pointer(gp);
1388 8 if (!p) { 1388 8 if (!p) {
1389 9 spin_unlock(&gp_lock); 1389 9 spin_unlock(&gp_lock);
1390 10 return false; 1390 10 return false;
1391 11 } 1391 11 }
1392 12 rcu_assign_pointer(gp, NULL); 1392 12 rcu_assign_pointer(gp, NULL);
1393 13 spin_unlock(&gp_lock); 1393 13 spin_unlock(&gp_lock);
1394 14 s = get_state_synchronize_rcu(); 1394 14 s = get_state_synchronize_rcu();
1395 15 do_something_while_waiting(); 1395 15 do_something_while_waiting();
1396 16 cond_synchronize_rcu(s); 1396 16 cond_synchronize_rcu(s);
1397 17 kfree(p); 1397 17 kfree(p);
1398 18 return true; 1398 18 return true;
1399 19 } 1399 19 }
1400 1400
1401 On line 14, get_state_synchronize_rcu() obta 1401 On line 14, get_state_synchronize_rcu() obtains a “cookie” from RCU,
1402 then line 15 carries out other tasks, and fi 1402 then line 15 carries out other tasks, and finally, line 16 returns
1403 immediately if a grace period has elapsed in 1403 immediately if a grace period has elapsed in the meantime, but otherwise
1404 waits as required. The need for ``get_state_s 1404 waits as required. The need for ``get_state_synchronize_rcu`` and
1405 cond_synchronize_rcu() has appeared quite rec 1405 cond_synchronize_rcu() has appeared quite recently, so it is too
1406 early to tell whether they will stand the tes 1406 early to tell whether they will stand the test of time.
1407 1407
1408 RCU thus provides a range of tools to allow u 1408 RCU thus provides a range of tools to allow updaters to strike the
1409 required tradeoff between latency, flexibilit 1409 required tradeoff between latency, flexibility and CPU overhead.
1410 1410
1411 Forward Progress 1411 Forward Progress
1412 ~~~~~~~~~~~~~~~~ 1412 ~~~~~~~~~~~~~~~~
1413 1413
1414 In theory, delaying grace-period completion a 1414 In theory, delaying grace-period completion and callback invocation is
1415 harmless. In practice, not only are memory si 1415 harmless. In practice, not only are memory sizes finite but also
1416 callbacks sometimes do wakeups, and sufficien 1416 callbacks sometimes do wakeups, and sufficiently deferred wakeups can be
1417 difficult to distinguish from system hangs. T 1417 difficult to distinguish from system hangs. Therefore, RCU must provide
1418 a number of mechanisms to promote forward pro 1418 a number of mechanisms to promote forward progress.
1419 1419
1420 These mechanisms are not foolproof, nor can t 1420 These mechanisms are not foolproof, nor can they be. For one simple
1421 example, an infinite loop in an RCU read-side 1421 example, an infinite loop in an RCU read-side critical section must by
1422 definition prevent later grace periods from e 1422 definition prevent later grace periods from ever completing. For a more
1423 involved example, consider a 64-CPU system bu 1423 involved example, consider a 64-CPU system built with
1424 ``CONFIG_RCU_NOCB_CPU=y`` and booted with ``r 1424 ``CONFIG_RCU_NOCB_CPU=y`` and booted with ``rcu_nocbs=1-63``, where
1425 CPUs 1 through 63 spin in tight loops that 1425 CPUs 1 through 63 spin in tight loops that invoke call_rcu(). Even
1426 if these tight loops also contain calls to co 1426 if these tight loops also contain calls to cond_resched() (thus
1427 allowing grace periods to complete), CPU 0 s 1427 allowing grace periods to complete), CPU 0 simply will not be able to
1428 invoke callbacks as fast as the other 63 CPUs 1428 invoke callbacks as fast as the other 63 CPUs can register them, at
1429 least not until the system runs out of memory 1429 least not until the system runs out of memory. In both of these
1430 examples, the Spiderman principle applies: Wi 1430 examples, the Spiderman principle applies: With great power comes great
1431 responsibility. However, short of this level 1431 responsibility. However, short of this level of abuse, RCU is required
1432 to ensure timely completion of grace periods 1432 to ensure timely completion of grace periods and timely invocation of
1433 callbacks. 1433 callbacks.
1434 1434
1435 RCU takes the following steps to encourage ti 1435 RCU takes the following steps to encourage timely completion of grace
1436 periods: 1436 periods:
1437 1437
1438 #. If a grace period fails to complete within 1438 #. If a grace period fails to complete within 100 milliseconds, RCU
1439 causes future invocations of cond_resched( 1439 causes future invocations of cond_resched() on the holdout CPUs
1440 to provide an RCU quiescent state. RCU als 1440 to provide an RCU quiescent state. RCU also causes those CPUs'
1441 need_resched() invocations to return ``tru 1441 need_resched() invocations to return ``true``, but only after the
1442 corresponding CPU's next scheduling-clock. 1442 corresponding CPU's next scheduling-clock.
1443 #. CPUs mentioned in the ``nohz_full`` kernel 1443 #. CPUs mentioned in the ``nohz_full`` kernel boot parameter can run
1444 indefinitely in the kernel without schedul 1444 indefinitely in the kernel without scheduling-clock interrupts, which
1445 defeats the above need_resched() strategem 1445 defeats the above need_resched() strategem. RCU will therefore
1446 invoke resched_cpu() on any ``nohz_full`` 1446 invoke resched_cpu() on any ``nohz_full`` CPUs still holding out
1447 after 109 milliseconds. 1447 after 109 milliseconds.
1448 #. In kernels built with ``CONFIG_RCU_BOOST=y 1448 #. In kernels built with ``CONFIG_RCU_BOOST=y``, if a given task that
1449 has been preempted within an RCU read-side 1449 has been preempted within an RCU read-side critical section is
1450 holding out for more than 500 millisecond 1450 holding out for more than 500 milliseconds, RCU will resort to
1451 priority boosting. 1451 priority boosting.
1452 #. If a CPU is still holding out 10 seconds 1452 #. If a CPU is still holding out 10 seconds into the grace period, RCU
1453 will invoke resched_cpu() on it regardless 1453 will invoke resched_cpu() on it regardless of its ``nohz_full``
1454 state. 1454 state.
1455 1455
1456 The above values are defaults for systems run 1456 The above values are defaults for systems running with ``HZ=1000``. They
1457 will vary as the value of ``HZ`` varies, and 1457 will vary as the value of ``HZ`` varies, and can also be changed using
1458 the relevant Kconfig options and kernel boot 1458 the relevant Kconfig options and kernel boot parameters. RCU currently
1459 does not do much sanity checking of these par 1459 does not do much sanity checking of these parameters, so please use
1460 caution when changing them. Note that these f 1460 caution when changing them. Note that these forward-progress measures
1461 are provided only for RCU, not for `SRCU <Sle 1461 are provided only for RCU, not for `SRCU <Sleepable RCU_>`__ or `Tasks
1462 RCU`_. 1462 RCU`_.
1463 1463
1464 RCU takes the following steps in call_rcu() t 1464 RCU takes the following steps in call_rcu() to encourage timely
1465 invocation of callbacks when any given non-\ 1465 invocation of callbacks when any given non-\ ``rcu_nocbs`` CPU has
1466 10,000 callbacks, or has 10,000 more callback 1466 10,000 callbacks, or has 10,000 more callbacks than it had the last time
1467 encouragement was provided: 1467 encouragement was provided:
1468 1468
1469 #. Starts a grace period, if one is not alrea 1469 #. Starts a grace period, if one is not already in progress.
1470 #. Forces immediate checking for quiescent st 1470 #. Forces immediate checking for quiescent states, rather than waiting
1471 for three milliseconds to have elapsed sin 1471 for three milliseconds to have elapsed since the beginning of the
1472 grace period. 1472 grace period.
1473 #. Immediately tags the CPU's callbacks with 1473 #. Immediately tags the CPU's callbacks with their grace period
1474 completion numbers, rather than waiting fo 1474 completion numbers, rather than waiting for the ``RCU_SOFTIRQ``
1475 handler to get around to it. 1475 handler to get around to it.
1476 #. Lifts callback-execution batch limits, whi 1476 #. Lifts callback-execution batch limits, which speeds up callback
1477 invocation at the expense of degrading rea 1477 invocation at the expense of degrading realtime response.
1478 1478
1479 Again, these are default values when running 1479 Again, these are default values when running at ``HZ=1000``, and can be
1480 overridden. Again, these forward-progress mea 1480 overridden. Again, these forward-progress measures are provided only for
1481 RCU, not for `SRCU <Sleepable RCU_>`__ or `Ta 1481 RCU, not for `SRCU <Sleepable RCU_>`__ or `Tasks
1482 RCU`_. Even for RCU, callback-invocation forw 1482 RCU`_. Even for RCU, callback-invocation forward
1483 progress for ``rcu_nocbs`` CPUs is much less 1483 progress for ``rcu_nocbs`` CPUs is much less well-developed, in part
1484 because workloads benefiting from ``rcu_nocbs 1484 because workloads benefiting from ``rcu_nocbs`` CPUs tend to invoke
1485 call_rcu() relatively infrequently. If worklo 1485 call_rcu() relatively infrequently. If workloads emerge that need
1486 both ``rcu_nocbs`` CPUs and high call_rcu() i 1486 both ``rcu_nocbs`` CPUs and high call_rcu() invocation rates, then
1487 additional forward-progress work will be requ 1487 additional forward-progress work will be required.
1488 1488
1489 Composability 1489 Composability
1490 ~~~~~~~~~~~~~ 1490 ~~~~~~~~~~~~~
1491 1491
1492 Composability has received much attention in 1492 Composability has received much attention in recent years, perhaps in
1493 part due to the collision of multicore hardwa 1493 part due to the collision of multicore hardware with object-oriented
1494 techniques designed in single-threaded enviro 1494 techniques designed in single-threaded environments for single-threaded
1495 use. And in theory, RCU read-side critical se 1495 use. And in theory, RCU read-side critical sections may be composed, and
1496 in fact may be nested arbitrarily deeply. In 1496 in fact may be nested arbitrarily deeply. In practice, as with all
1497 real-world implementations of composable cons 1497 real-world implementations of composable constructs, there are
1498 limitations. 1498 limitations.
1499 1499
1500 Implementations of RCU for which rcu_read_loc 1500 Implementations of RCU for which rcu_read_lock() and
1501 rcu_read_unlock() generate no code, such as L 1501 rcu_read_unlock() generate no code, such as Linux-kernel RCU when
1502 ``CONFIG_PREEMPTION=n``, can be nested arbitr 1502 ``CONFIG_PREEMPTION=n``, can be nested arbitrarily deeply. After all, there
1503 is no overhead. Except that if all these inst 1503 is no overhead. Except that if all these instances of
1504 rcu_read_lock() and rcu_read_unlock() are vis 1504 rcu_read_lock() and rcu_read_unlock() are visible to the
1505 compiler, compilation will eventually fail du 1505 compiler, compilation will eventually fail due to exhausting memory,
1506 mass storage, or user patience, whichever com 1506 mass storage, or user patience, whichever comes first. If the nesting is
1507 not visible to the compiler, as is the case w 1507 not visible to the compiler, as is the case with mutually recursive
1508 functions each in its own translation unit, s 1508 functions each in its own translation unit, stack overflow will result.
1509 If the nesting takes the form of loops, perha 1509 If the nesting takes the form of loops, perhaps in the guise of tail
1510 recursion, either the control variable will o 1510 recursion, either the control variable will overflow or (in the Linux
1511 kernel) you will get an RCU CPU stall warning 1511 kernel) you will get an RCU CPU stall warning. Nevertheless, this class
1512 of RCU implementations is one of the most com 1512 of RCU implementations is one of the most composable constructs in
1513 existence. 1513 existence.
1514 1514
1515 RCU implementations that explicitly track nes 1515 RCU implementations that explicitly track nesting depth are limited by
1516 the nesting-depth counter. For example, the L 1516 the nesting-depth counter. For example, the Linux kernel's preemptible
1517 RCU limits nesting to ``INT_MAX``. This shoul 1517 RCU limits nesting to ``INT_MAX``. This should suffice for almost all
1518 practical purposes. That said, a consecutive 1518 practical purposes. That said, a consecutive pair of RCU read-side
1519 critical sections between which there is an o 1519 critical sections between which there is an operation that waits for a
1520 grace period cannot be enclosed in another RC 1520 grace period cannot be enclosed in another RCU read-side critical
1521 section. This is because it is not legal to w 1521 section. This is because it is not legal to wait for a grace period
1522 within an RCU read-side critical section: To 1522 within an RCU read-side critical section: To do so would result either
1523 in deadlock or in RCU implicitly splitting th 1523 in deadlock or in RCU implicitly splitting the enclosing RCU read-side
1524 critical section, neither of which is conduci 1524 critical section, neither of which is conducive to a long-lived and
1525 prosperous kernel. 1525 prosperous kernel.
1526 1526
1527 It is worth noting that RCU is not alone in l 1527 It is worth noting that RCU is not alone in limiting composability. For
1528 example, many transactional-memory implementa 1528 example, many transactional-memory implementations prohibit composing a
1529 pair of transactions separated by an irrevoca 1529 pair of transactions separated by an irrevocable operation (for example,
1530 a network receive operation). For another exa 1530 a network receive operation). For another example, lock-based critical
1531 sections can be composed surprisingly freely, 1531 sections can be composed surprisingly freely, but only if deadlock is
1532 avoided. 1532 avoided.
1533 1533
1534 In short, although RCU read-side critical sec 1534 In short, although RCU read-side critical sections are highly
1535 composable, care is required in some situatio 1535 composable, care is required in some situations, just as is the case for
1536 any other composable synchronization mechanis 1536 any other composable synchronization mechanism.
1537 1537
1538 Corner Cases 1538 Corner Cases
1539 ~~~~~~~~~~~~ 1539 ~~~~~~~~~~~~
1540 1540
1541 A given RCU workload might have an endless an 1541 A given RCU workload might have an endless and intense stream of RCU
1542 read-side critical sections, perhaps even so 1542 read-side critical sections, perhaps even so intense that there was
1543 never a point in time during which there was 1543 never a point in time during which there was not at least one RCU
1544 read-side critical section in flight. RCU can 1544 read-side critical section in flight. RCU cannot allow this situation to
1545 block grace periods: As long as all the RCU r 1545 block grace periods: As long as all the RCU read-side critical sections
1546 are finite, grace periods must also be finite 1546 are finite, grace periods must also be finite.
1547 1547
1548 That said, preemptible RCU implementations co 1548 That said, preemptible RCU implementations could potentially result in
1549 RCU read-side critical sections being preempt 1549 RCU read-side critical sections being preempted for long durations,
1550 which has the effect of creating a long-durat 1550 which has the effect of creating a long-duration RCU read-side critical
1551 section. This situation can arise only in hea 1551 section. This situation can arise only in heavily loaded systems, but
1552 systems using real-time priorities are of cou 1552 systems using real-time priorities are of course more vulnerable.
1553 Therefore, RCU priority boosting is provided 1553 Therefore, RCU priority boosting is provided to help deal with this
1554 case. That said, the exact requirements on RC 1554 case. That said, the exact requirements on RCU priority boosting will
1555 likely evolve as more experience accumulates. 1555 likely evolve as more experience accumulates.
1556 1556
1557 Other workloads might have very high update r 1557 Other workloads might have very high update rates. Although one can
1558 argue that such workloads should instead use 1558 argue that such workloads should instead use something other than RCU,
1559 the fact remains that RCU must handle such wo 1559 the fact remains that RCU must handle such workloads gracefully. This
1560 requirement is another factor driving batchin 1560 requirement is another factor driving batching of grace periods, but it
1561 is also the driving force behind the checks f 1561 is also the driving force behind the checks for large numbers of queued
1562 RCU callbacks in the call_rcu() code path. Fi 1562 RCU callbacks in the call_rcu() code path. Finally, high update
1563 rates should not delay RCU read-side critical 1563 rates should not delay RCU read-side critical sections, although some
1564 small read-side delays can occur when using 1564 small read-side delays can occur when using
1565 synchronize_rcu_expedited(), courtesy of this 1565 synchronize_rcu_expedited(), courtesy of this function's use of
1566 smp_call_function_single(). 1566 smp_call_function_single().
1567 1567
1568 Although all three of these corner cases were 1568 Although all three of these corner cases were understood in the early
1569 1990s, a simple user-level test consisting of 1569 1990s, a simple user-level test consisting of ``close(open(path))`` in a
1570 tight loop in the early 2000s suddenly provid 1570 tight loop in the early 2000s suddenly provided a much deeper
1571 appreciation of the high-update-rate corner c 1571 appreciation of the high-update-rate corner case. This test also
1572 motivated addition of some RCU code to react 1572 motivated addition of some RCU code to react to high update rates, for
1573 example, if a given CPU finds itself with mor 1573 example, if a given CPU finds itself with more than 10,000 RCU callbacks
1574 queued, it will cause RCU to take evasive act 1574 queued, it will cause RCU to take evasive action by more aggressively
1575 starting grace periods and more aggressively 1575 starting grace periods and more aggressively forcing completion of
1576 grace-period processing. This evasive action 1576 grace-period processing. This evasive action causes the grace period to
1577 complete more quickly, but at the cost of res 1577 complete more quickly, but at the cost of restricting RCU's batching
1578 optimizations, thus increasing the CPU overhe 1578 optimizations, thus increasing the CPU overhead incurred by that grace
1579 period. 1579 period.
1580 1580
1581 Software-Engineering Requirements 1581 Software-Engineering Requirements
1582 --------------------------------- 1582 ---------------------------------
1583 1583
1584 Between Murphy's Law and “To err is human 1584 Between Murphy's Law and “To err is human”, it is necessary to guard
1585 against mishaps and misuse: 1585 against mishaps and misuse:
1586 1586
1587 #. It is all too easy to forget to use rcu_re 1587 #. It is all too easy to forget to use rcu_read_lock() everywhere
1588 that it is needed, so kernels built with ` 1588 that it is needed, so kernels built with ``CONFIG_PROVE_RCU=y`` will
1589 splat if rcu_dereference() is used outside 1589 splat if rcu_dereference() is used outside of an RCU read-side
1590 critical section. Update-side code can use 1590 critical section. Update-side code can use
1591 rcu_dereference_protected(), which takes a 1591 rcu_dereference_protected(), which takes a `lockdep
1592 expression <https://lwn.net/Articles/37198 1592 expression <https://lwn.net/Articles/371986/>`__ to indicate what is
1593 providing the protection. If the indicated 1593 providing the protection. If the indicated protection is not
1594 provided, a lockdep splat is emitted. 1594 provided, a lockdep splat is emitted.
1595 Code shared between readers and updaters c 1595 Code shared between readers and updaters can use
1596 rcu_dereference_check(), which also takes 1596 rcu_dereference_check(), which also takes a lockdep expression,
1597 and emits a lockdep splat if neither rcu_r 1597 and emits a lockdep splat if neither rcu_read_lock() nor the
1598 indicated protection is in place. In addit 1598 indicated protection is in place. In addition,
1599 rcu_dereference_raw() is used in those (ho 1599 rcu_dereference_raw() is used in those (hopefully rare) cases
1600 where the required protection cannot be ea 1600 where the required protection cannot be easily described. Finally,
1601 rcu_read_lock_held() is provided to allow 1601 rcu_read_lock_held() is provided to allow a function to verify
1602 that it has been invoked within an RCU rea 1602 that it has been invoked within an RCU read-side critical section. I
1603 was made aware of this set of requirements 1603 was made aware of this set of requirements shortly after Thomas
1604 Gleixner audited a number of RCU uses. 1604 Gleixner audited a number of RCU uses.
1605 #. A given function might wish to check for R 1605 #. A given function might wish to check for RCU-related preconditions
1606 upon entry, before using any other RCU API 1606 upon entry, before using any other RCU API. The
1607 rcu_lockdep_assert() does this job, assert 1607 rcu_lockdep_assert() does this job, asserting the expression in
1608 kernels having lockdep enabled and doing n 1608 kernels having lockdep enabled and doing nothing otherwise.
1609 #. It is also easy to forget to use rcu_assig 1609 #. It is also easy to forget to use rcu_assign_pointer() and
1610 rcu_dereference(), perhaps (incorrectly) s 1610 rcu_dereference(), perhaps (incorrectly) substituting a simple
1611 assignment. To catch this sort of error, a 1611 assignment. To catch this sort of error, a given RCU-protected
1612 pointer may be tagged with ``__rcu``, afte 1612 pointer may be tagged with ``__rcu``, after which sparse will
1613 complain about simple-assignment accesses 1613 complain about simple-assignment accesses to that pointer. Arnd
1614 Bergmann made me aware of this requirement 1614 Bergmann made me aware of this requirement, and also supplied the
1615 needed `patch series <https://lwn.net/Arti 1615 needed `patch series <https://lwn.net/Articles/376011/>`__.
1616 #. Kernels built with ``CONFIG_DEBUG_OBJECTS_ 1616 #. Kernels built with ``CONFIG_DEBUG_OBJECTS_RCU_HEAD=y`` will splat if
1617 a data element is passed to call_rcu() twi 1617 a data element is passed to call_rcu() twice in a row, without a
1618 grace period in between. (This error is si 1618 grace period in between. (This error is similar to a double free.)
1619 The corresponding ``rcu_head`` structures 1619 The corresponding ``rcu_head`` structures that are dynamically
1620 allocated are automatically tracked, but ` 1620 allocated are automatically tracked, but ``rcu_head`` structures
1621 allocated on the stack must be initialized 1621 allocated on the stack must be initialized with
1622 init_rcu_head_on_stack() and cleaned up wi 1622 init_rcu_head_on_stack() and cleaned up with
1623 destroy_rcu_head_on_stack(). Similarly, st 1623 destroy_rcu_head_on_stack(). Similarly, statically allocated
1624 non-stack ``rcu_head`` structures must be 1624 non-stack ``rcu_head`` structures must be initialized with
1625 init_rcu_head() and cleaned up with destro 1625 init_rcu_head() and cleaned up with destroy_rcu_head().
1626 Mathieu Desnoyers made me aware of this re 1626 Mathieu Desnoyers made me aware of this requirement, and also
1627 supplied the needed 1627 supplied the needed
1628 `patch <https://lore.kernel.org/r/201003190 1628 `patch <https://lore.kernel.org/r/20100319013024.GA28456@Krystal">https://lore.kernel.org/r/20100319013024.GA28456@Krystal>`__.
1629 #. An infinite loop in an RCU read-side criti 1629 #. An infinite loop in an RCU read-side critical section will eventually
1630 trigger an RCU CPU stall warning splat, wi 1630 trigger an RCU CPU stall warning splat, with the duration of
1631 “eventually” being controlled by the ` 1631 “eventually” being controlled by the ``RCU_CPU_STALL_TIMEOUT``
1632 ``Kconfig`` option, or, alternatively, by 1632 ``Kconfig`` option, or, alternatively, by the
1633 ``rcupdate.rcu_cpu_stall_timeout`` boot/sy 1633 ``rcupdate.rcu_cpu_stall_timeout`` boot/sysfs parameter. However, RCU
1634 is not obligated to produce this splat unl 1634 is not obligated to produce this splat unless there is a grace period
1635 waiting on that particular RCU read-side c 1635 waiting on that particular RCU read-side critical section.
1636 1636
1637 Some extreme workloads might intentionally 1637 Some extreme workloads might intentionally delay RCU grace periods,
1638 and systems running those workloads can be 1638 and systems running those workloads can be booted with
1639 ``rcupdate.rcu_cpu_stall_suppress`` to sup 1639 ``rcupdate.rcu_cpu_stall_suppress`` to suppress the splats. This
1640 kernel parameter may also be set via ``sys 1640 kernel parameter may also be set via ``sysfs``. Furthermore, RCU CPU
1641 stall warnings are counter-productive duri 1641 stall warnings are counter-productive during sysrq dumps and during
1642 panics. RCU therefore supplies the rcu_sys 1642 panics. RCU therefore supplies the rcu_sysrq_start() and
1643 rcu_sysrq_end() API members to be called b 1643 rcu_sysrq_end() API members to be called before and after long
1644 sysrq dumps. RCU also supplies the rcu_pan 1644 sysrq dumps. RCU also supplies the rcu_panic() notifier that is
1645 automatically invoked at the beginning of 1645 automatically invoked at the beginning of a panic to suppress further
1646 RCU CPU stall warnings. 1646 RCU CPU stall warnings.
1647 1647
1648 This requirement made itself known in the 1648 This requirement made itself known in the early 1990s, pretty much
1649 the first time that it was necessary to de 1649 the first time that it was necessary to debug a CPU stall. That said,
1650 the initial implementation in DYNIX/ptx wa 1650 the initial implementation in DYNIX/ptx was quite generic in
1651 comparison with that of Linux. 1651 comparison with that of Linux.
1652 1652
1653 #. Although it would be very good to detect p 1653 #. Although it would be very good to detect pointers leaking out of RCU
1654 read-side critical sections, there is curr 1654 read-side critical sections, there is currently no good way of doing
1655 this. One complication is the need to dist 1655 this. One complication is the need to distinguish between pointers
1656 leaking and pointers that have been handed 1656 leaking and pointers that have been handed off from RCU to some other
1657 synchronization mechanism, for example, re 1657 synchronization mechanism, for example, reference counting.
1658 #. In kernels built with ``CONFIG_RCU_TRACE=y 1658 #. In kernels built with ``CONFIG_RCU_TRACE=y``, RCU-related information
1659 is provided via event tracing. 1659 is provided via event tracing.
1660 #. Open-coded use of rcu_assign_pointer() and 1660 #. Open-coded use of rcu_assign_pointer() and rcu_dereference()
1661 to create typical linked data structures c 1661 to create typical linked data structures can be surprisingly
1662 error-prone. Therefore, RCU-protected `lin 1662 error-prone. Therefore, RCU-protected `linked
1663 lists <https://lwn.net/Articles/609973/#RC 1663 lists <https://lwn.net/Articles/609973/#RCU%20List%20APIs>`__ and,
1664 more recently, RCU-protected `hash 1664 more recently, RCU-protected `hash
1665 tables <https://lwn.net/Articles/612100/>` 1665 tables <https://lwn.net/Articles/612100/>`__ are available. Many
1666 other special-purpose RCU-protected data s 1666 other special-purpose RCU-protected data structures are available in
1667 the Linux kernel and the userspace RCU lib 1667 the Linux kernel and the userspace RCU library.
1668 #. Some linked structures are created at comp 1668 #. Some linked structures are created at compile time, but still require
1669 ``__rcu`` checking. The RCU_POINTER_INITIA 1669 ``__rcu`` checking. The RCU_POINTER_INITIALIZER() macro serves
1670 this purpose. 1670 this purpose.
1671 #. It is not necessary to use rcu_assign_poin 1671 #. It is not necessary to use rcu_assign_pointer() when creating
1672 linked structures that are to be published 1672 linked structures that are to be published via a single external
1673 pointer. The RCU_INIT_POINTER() macro is p 1673 pointer. The RCU_INIT_POINTER() macro is provided for this task.
1674 1674
1675 This not a hard-and-fast list: RCU's diagnost 1675 This not a hard-and-fast list: RCU's diagnostic capabilities will
1676 continue to be guided by the number and type 1676 continue to be guided by the number and type of usage bugs found in
1677 real-world RCU usage. 1677 real-world RCU usage.
1678 1678
1679 Linux Kernel Complications 1679 Linux Kernel Complications
1680 -------------------------- 1680 --------------------------
1681 1681
1682 The Linux kernel provides an interesting envi 1682 The Linux kernel provides an interesting environment for all kinds of
1683 software, including RCU. Some of the relevant 1683 software, including RCU. Some of the relevant points of interest are as
1684 follows: 1684 follows:
1685 1685
1686 #. `Configuration`_ 1686 #. `Configuration`_
1687 #. `Firmware Interface`_ 1687 #. `Firmware Interface`_
1688 #. `Early Boot`_ 1688 #. `Early Boot`_
1689 #. `Interrupts and NMIs`_ 1689 #. `Interrupts and NMIs`_
1690 #. `Loadable Modules`_ 1690 #. `Loadable Modules`_
1691 #. `Hotplug CPU`_ 1691 #. `Hotplug CPU`_
1692 #. `Scheduler and RCU`_ 1692 #. `Scheduler and RCU`_
1693 #. `Tracing and RCU`_ 1693 #. `Tracing and RCU`_
1694 #. `Accesses to User Memory and RCU`_ 1694 #. `Accesses to User Memory and RCU`_
1695 #. `Energy Efficiency`_ 1695 #. `Energy Efficiency`_
1696 #. `Scheduling-Clock Interrupts and RCU`_ 1696 #. `Scheduling-Clock Interrupts and RCU`_
1697 #. `Memory Efficiency`_ 1697 #. `Memory Efficiency`_
1698 #. `Performance, Scalability, Response Time, 1698 #. `Performance, Scalability, Response Time, and Reliability`_
1699 1699
1700 This list is probably incomplete, but it does 1700 This list is probably incomplete, but it does give a feel for the most
1701 notable Linux-kernel complications. Each of t 1701 notable Linux-kernel complications. Each of the following sections
1702 covers one of the above topics. 1702 covers one of the above topics.
1703 1703
1704 Configuration 1704 Configuration
1705 ~~~~~~~~~~~~~ 1705 ~~~~~~~~~~~~~
1706 1706
1707 RCU's goal is automatic configuration, so tha 1707 RCU's goal is automatic configuration, so that almost nobody needs to
1708 worry about RCU's ``Kconfig`` options. And fo 1708 worry about RCU's ``Kconfig`` options. And for almost all users, RCU
1709 does in fact work well “out of the box.” 1709 does in fact work well “out of the box.”
1710 1710
1711 However, there are specialized use cases that 1711 However, there are specialized use cases that are handled by kernel boot
1712 parameters and ``Kconfig`` options. Unfortuna 1712 parameters and ``Kconfig`` options. Unfortunately, the ``Kconfig``
1713 system will explicitly ask users about new `` 1713 system will explicitly ask users about new ``Kconfig`` options, which
1714 requires almost all of them be hidden behind 1714 requires almost all of them be hidden behind a ``CONFIG_RCU_EXPERT``
1715 ``Kconfig`` option. 1715 ``Kconfig`` option.
1716 1716
1717 This all should be quite obvious, but the fac 1717 This all should be quite obvious, but the fact remains that Linus
1718 Torvalds recently had to 1718 Torvalds recently had to
1719 `remind <https://lore.kernel.org/r/CA+55aFy4wc 1719 `remind <https://lore.kernel.org/r/CA+55aFy4wcCwaL4okTs8wXhGZ5h-ibecy_Meg9C4MNQrUnwMcg@mail.gmail.com">https://lore.kernel.org/r/CA+55aFy4wcCwaL4okTs8wXhGZ5h-ibecy_Meg9C4MNQrUnwMcg@mail.gmail.com>`__
1720 me of this requirement. 1720 me of this requirement.
1721 1721
1722 Firmware Interface 1722 Firmware Interface
1723 ~~~~~~~~~~~~~~~~~~ 1723 ~~~~~~~~~~~~~~~~~~
1724 1724
1725 In many cases, kernel obtains information abo 1725 In many cases, kernel obtains information about the system from the
1726 firmware, and sometimes things are lost in tr 1726 firmware, and sometimes things are lost in translation. Or the
1727 translation is accurate, but the original mes 1727 translation is accurate, but the original message is bogus.
1728 1728
1729 For example, some systems' firmware overrepor 1729 For example, some systems' firmware overreports the number of CPUs,
1730 sometimes by a large factor. If RCU naively b 1730 sometimes by a large factor. If RCU naively believed the firmware, as it
1731 used to do, it would create too many per-CPU 1731 used to do, it would create too many per-CPU kthreads. Although the
1732 resulting system will still run correctly, th 1732 resulting system will still run correctly, the extra kthreads needlessly
1733 consume memory and can cause confusion when t 1733 consume memory and can cause confusion when they show up in ``ps``
1734 listings. 1734 listings.
1735 1735
1736 RCU must therefore wait for a given CPU to ac 1736 RCU must therefore wait for a given CPU to actually come online before
1737 it can allow itself to believe that the CPU a 1737 it can allow itself to believe that the CPU actually exists. The
1738 resulting “ghost CPUs” (which are never g 1738 resulting “ghost CPUs” (which are never going to come online) cause a
1739 number of `interesting 1739 number of `interesting
1740 complications <https://paulmck.livejournal.co 1740 complications <https://paulmck.livejournal.com/37494.html>`__.
1741 1741
1742 Early Boot 1742 Early Boot
1743 ~~~~~~~~~~ 1743 ~~~~~~~~~~
1744 1744
1745 The Linux kernel's boot sequence is an intere 1745 The Linux kernel's boot sequence is an interesting process, and RCU is
1746 used early, even before rcu_init() is invoked 1746 used early, even before rcu_init() is invoked. In fact, a number of
1747 RCU's primitives can be used as soon as the i 1747 RCU's primitives can be used as soon as the initial task's
1748 ``task_struct`` is available and the boot CPU 1748 ``task_struct`` is available and the boot CPU's per-CPU variables are
1749 set up. The read-side primitives (rcu_read_lo 1749 set up. The read-side primitives (rcu_read_lock(),
1750 rcu_read_unlock(), rcu_dereference(), and 1750 rcu_read_unlock(), rcu_dereference(), and
1751 rcu_access_pointer()) will operate normally v 1751 rcu_access_pointer()) will operate normally very early on, as will
1752 rcu_assign_pointer(). 1752 rcu_assign_pointer().
1753 1753
1754 Although call_rcu() may be invoked at any tim 1754 Although call_rcu() may be invoked at any time during boot,
1755 callbacks are not guaranteed to be invoked un 1755 callbacks are not guaranteed to be invoked until after all of RCU's
1756 kthreads have been spawned, which occurs at e 1756 kthreads have been spawned, which occurs at early_initcall() time.
1757 This delay in callback invocation is due to t 1757 This delay in callback invocation is due to the fact that RCU does not
1758 invoke callbacks until it is fully initialize 1758 invoke callbacks until it is fully initialized, and this full
1759 initialization cannot occur until after the s 1759 initialization cannot occur until after the scheduler has initialized
1760 itself to the point where RCU can spawn and r 1760 itself to the point where RCU can spawn and run its kthreads. In theory,
1761 it would be possible to invoke callbacks earl 1761 it would be possible to invoke callbacks earlier, however, this is not a
1762 panacea because there would be severe restric 1762 panacea because there would be severe restrictions on what operations
1763 those callbacks could invoke. 1763 those callbacks could invoke.
1764 1764
1765 Perhaps surprisingly, synchronize_rcu() and 1765 Perhaps surprisingly, synchronize_rcu() and
1766 synchronize_rcu_expedited(), will operate nor 1766 synchronize_rcu_expedited(), will operate normally during very early
1767 boot, the reason being that there is only one 1767 boot, the reason being that there is only one CPU and preemption is
1768 disabled. This means that the call synchroniz 1768 disabled. This means that the call synchronize_rcu() (or friends)
1769 itself is a quiescent state and thus a grace 1769 itself is a quiescent state and thus a grace period, so the early-boot
1770 implementation can be a no-op. 1770 implementation can be a no-op.
1771 1771
1772 However, once the scheduler has spawned its f 1772 However, once the scheduler has spawned its first kthread, this early
1773 boot trick fails for synchronize_rcu() (as we 1773 boot trick fails for synchronize_rcu() (as well as for
1774 synchronize_rcu_expedited()) in ``CONFIG_PREE 1774 synchronize_rcu_expedited()) in ``CONFIG_PREEMPTION=y`` kernels. The
1775 reason is that an RCU read-side critical sect 1775 reason is that an RCU read-side critical section might be preempted,
1776 which means that a subsequent synchronize_rcu 1776 which means that a subsequent synchronize_rcu() really does have to
1777 wait for something, as opposed to simply retu 1777 wait for something, as opposed to simply returning immediately.
1778 Unfortunately, synchronize_rcu() can't do thi 1778 Unfortunately, synchronize_rcu() can't do this until all of its
1779 kthreads are spawned, which doesn't happen un 1779 kthreads are spawned, which doesn't happen until some time during
1780 early_initcalls() time. But this is no excuse 1780 early_initcalls() time. But this is no excuse: RCU is nevertheless
1781 required to correctly handle synchronous grac 1781 required to correctly handle synchronous grace periods during this time
1782 period. Once all of its kthreads are up and r 1782 period. Once all of its kthreads are up and running, RCU starts running
1783 normally. 1783 normally.
1784 1784
1785 +-------------------------------------------- 1785 +-----------------------------------------------------------------------+
1786 | **Quick Quiz**: 1786 | **Quick Quiz**: |
1787 +-------------------------------------------- 1787 +-----------------------------------------------------------------------+
1788 | How can RCU possibly handle grace periods b 1788 | How can RCU possibly handle grace periods before all of its kthreads |
1789 | have been spawned??? 1789 | have been spawned??? |
1790 +-------------------------------------------- 1790 +-----------------------------------------------------------------------+
1791 | **Answer**: 1791 | **Answer**: |
1792 +-------------------------------------------- 1792 +-----------------------------------------------------------------------+
1793 | Very carefully! 1793 | Very carefully! |
1794 | During the “dead zone” between the time 1794 | During the “dead zone” between the time that the scheduler spawns the |
1795 | first task and the time that all of RCU's k 1795 | first task and the time that all of RCU's kthreads have been spawned, |
1796 | all synchronous grace periods are handled b 1796 | all synchronous grace periods are handled by the expedited |
1797 | grace-period mechanism. At runtime, this ex 1797 | grace-period mechanism. At runtime, this expedited mechanism relies |
1798 | on workqueues, but during the dead zone the 1798 | on workqueues, but during the dead zone the requesting task itself |
1799 | drives the desired expedited grace period. 1799 | drives the desired expedited grace period. Because dead-zone |
1800 | execution takes place within task context, 1800 | execution takes place within task context, everything works. Once the |
1801 | dead zone ends, expedited grace periods go 1801 | dead zone ends, expedited grace periods go back to using workqueues, |
1802 | as is required to avoid problems that would 1802 | as is required to avoid problems that would otherwise occur when a |
1803 | user task received a POSIX signal while dri 1803 | user task received a POSIX signal while driving an expedited grace |
1804 | period. 1804 | period. |
1805 | 1805 | |
1806 | And yes, this does mean that it is unhelpfu 1806 | And yes, this does mean that it is unhelpful to send POSIX signals to |
1807 | random tasks between the time that the sche 1807 | random tasks between the time that the scheduler spawns its first |
1808 | kthread and the time that RCU's kthreads ha 1808 | kthread and the time that RCU's kthreads have all been spawned. If |
1809 | there ever turns out to be a good reason fo 1809 | there ever turns out to be a good reason for sending POSIX signals |
1810 | during that time, appropriate adjustments w 1810 | during that time, appropriate adjustments will be made. (If it turns |
1811 | out that POSIX signals are sent during this 1811 | out that POSIX signals are sent during this time for no good reason, |
1812 | other adjustments will be made, appropriate 1812 | other adjustments will be made, appropriate or otherwise.) |
1813 +-------------------------------------------- 1813 +-----------------------------------------------------------------------+
1814 1814
1815 I learned of these boot-time requirements as 1815 I learned of these boot-time requirements as a result of a series of
1816 system hangs. 1816 system hangs.
1817 1817
1818 Interrupts and NMIs 1818 Interrupts and NMIs
1819 ~~~~~~~~~~~~~~~~~~~ 1819 ~~~~~~~~~~~~~~~~~~~
1820 1820
1821 The Linux kernel has interrupts, and RCU read 1821 The Linux kernel has interrupts, and RCU read-side critical sections are
1822 legal within interrupt handlers and within in 1822 legal within interrupt handlers and within interrupt-disabled regions of
1823 code, as are invocations of call_rcu(). 1823 code, as are invocations of call_rcu().
1824 1824
1825 Some Linux-kernel architectures can enter an 1825 Some Linux-kernel architectures can enter an interrupt handler from
1826 non-idle process context, and then just never 1826 non-idle process context, and then just never leave it, instead
1827 stealthily transitioning back to process cont 1827 stealthily transitioning back to process context. This trick is
1828 sometimes used to invoke system calls from in 1828 sometimes used to invoke system calls from inside the kernel. These
1829 “half-interrupts” mean that RCU has to be 1829 “half-interrupts” mean that RCU has to be very careful about how it
1830 counts interrupt nesting levels. I learned of 1830 counts interrupt nesting levels. I learned of this requirement the hard
1831 way during a rewrite of RCU's dyntick-idle co 1831 way during a rewrite of RCU's dyntick-idle code.
1832 1832
1833 The Linux kernel has non-maskable interrupts 1833 The Linux kernel has non-maskable interrupts (NMIs), and RCU read-side
1834 critical sections are legal within NMI handle 1834 critical sections are legal within NMI handlers. Thankfully, RCU
1835 update-side primitives, including call_rcu(), 1835 update-side primitives, including call_rcu(), are prohibited within
1836 NMI handlers. 1836 NMI handlers.
1837 1837
1838 The name notwithstanding, some Linux-kernel a 1838 The name notwithstanding, some Linux-kernel architectures can have
1839 nested NMIs, which RCU must handle correctly. 1839 nested NMIs, which RCU must handle correctly. Andy Lutomirski `surprised
1840 me <https://lore.kernel.org/r/CALCETrXLq1y7e_d 1840 me <https://lore.kernel.org/r/CALCETrXLq1y7e_dKFPgou-FKHB6Pu-r8+t-6Ds+8=va7anBWDA@mail.gmail.com">https://lore.kernel.org/r/CALCETrXLq1y7e_dKFPgou-FKHB6Pu-r8+t-6Ds+8=va7anBWDA@mail.gmail.com>`__
1841 with this requirement; he also kindly surpris 1841 with this requirement; he also kindly surprised me with `an
1842 algorithm <https://lore.kernel.org/r/CALCETrXS 1842 algorithm <https://lore.kernel.org/r/CALCETrXSY9JpW3uE6H8WYk81sg56qasA2aqmjMPsq5dOtzso=g@mail.gmail.com">https://lore.kernel.org/r/CALCETrXSY9JpW3uE6H8WYk81sg56qasA2aqmjMPsq5dOtzso=g@mail.gmail.com>`__
1843 that meets this requirement. 1843 that meets this requirement.
1844 1844
1845 Furthermore, NMI handlers can be interrupted 1845 Furthermore, NMI handlers can be interrupted by what appear to RCU to be
1846 normal interrupts. One way that this can happ 1846 normal interrupts. One way that this can happen is for code that
1847 directly invokes ct_irq_enter() and ct_irq_ex 1847 directly invokes ct_irq_enter() and ct_irq_exit() to be called
1848 from an NMI handler. This astonishing fact of 1848 from an NMI handler. This astonishing fact of life prompted the current
1849 code structure, which has ct_irq_enter() invo 1849 code structure, which has ct_irq_enter() invoking
1850 ct_nmi_enter() and ct_irq_exit() invoking ct_ 1850 ct_nmi_enter() and ct_irq_exit() invoking ct_nmi_exit().
1851 And yes, I also learned of this requirement t 1851 And yes, I also learned of this requirement the hard way.
1852 1852
1853 Loadable Modules 1853 Loadable Modules
1854 ~~~~~~~~~~~~~~~~ 1854 ~~~~~~~~~~~~~~~~
1855 1855
1856 The Linux kernel has loadable modules, and th 1856 The Linux kernel has loadable modules, and these modules can also be
1857 unloaded. After a given module has been unloa 1857 unloaded. After a given module has been unloaded, any attempt to call
1858 one of its functions results in a segmentatio 1858 one of its functions results in a segmentation fault. The module-unload
1859 functions must therefore cancel any delayed c 1859 functions must therefore cancel any delayed calls to loadable-module
1860 functions, for example, any outstanding mod_t 1860 functions, for example, any outstanding mod_timer() must be dealt
1861 with via timer_shutdown_sync() or similar. 1861 with via timer_shutdown_sync() or similar.
1862 1862
1863 Unfortunately, there is no way to cancel an R 1863 Unfortunately, there is no way to cancel an RCU callback; once you
1864 invoke call_rcu(), the callback function is e 1864 invoke call_rcu(), the callback function is eventually going to be
1865 invoked, unless the system goes down first. B 1865 invoked, unless the system goes down first. Because it is normally
1866 considered socially irresponsible to crash th 1866 considered socially irresponsible to crash the system in response to a
1867 module unload request, we need some other way 1867 module unload request, we need some other way to deal with in-flight RCU
1868 callbacks. 1868 callbacks.
1869 1869
1870 RCU therefore provides rcu_barrier(), which w 1870 RCU therefore provides rcu_barrier(), which waits until all
1871 in-flight RCU callbacks have been invoked. If 1871 in-flight RCU callbacks have been invoked. If a module uses
1872 call_rcu(), its exit function should therefor 1872 call_rcu(), its exit function should therefore prevent any future
1873 invocation of call_rcu(), then invoke rcu_bar 1873 invocation of call_rcu(), then invoke rcu_barrier(). In theory,
1874 the underlying module-unload code could invok 1874 the underlying module-unload code could invoke rcu_barrier()
1875 unconditionally, but in practice this would i 1875 unconditionally, but in practice this would incur unacceptable
1876 latencies. 1876 latencies.
1877 1877
1878 Nikita Danilov noted this requirement for an 1878 Nikita Danilov noted this requirement for an analogous
1879 filesystem-unmount situation, and Dipankar Sa 1879 filesystem-unmount situation, and Dipankar Sarma incorporated
1880 rcu_barrier() into RCU. The need for rcu_barr 1880 rcu_barrier() into RCU. The need for rcu_barrier() for module
1881 unloading became apparent later. 1881 unloading became apparent later.
1882 1882
1883 .. important:: 1883 .. important::
1884 1884
1885 The rcu_barrier() function is not, repeat, 1885 The rcu_barrier() function is not, repeat,
1886 *not*, obligated to wait for a grace perio 1886 *not*, obligated to wait for a grace period. It is instead only required
1887 to wait for RCU callbacks that have alread 1887 to wait for RCU callbacks that have already been posted. Therefore, if
1888 there are no RCU callbacks posted anywhere 1888 there are no RCU callbacks posted anywhere in the system,
1889 rcu_barrier() is within its rights to retu 1889 rcu_barrier() is within its rights to return immediately. Even if
1890 there are callbacks posted, rcu_barrier() 1890 there are callbacks posted, rcu_barrier() does not necessarily need
1891 to wait for a grace period. 1891 to wait for a grace period.
1892 1892
1893 +-------------------------------------------- 1893 +-----------------------------------------------------------------------+
1894 | **Quick Quiz**: 1894 | **Quick Quiz**: |
1895 +-------------------------------------------- 1895 +-----------------------------------------------------------------------+
1896 | Wait a minute! Each RCU callbacks must wait 1896 | Wait a minute! Each RCU callbacks must wait for a grace period to |
1897 | complete, and rcu_barrier() must wait for e 1897 | complete, and rcu_barrier() must wait for each pre-existing |
1898 | callback to be invoked. Doesn't rcu_barrier 1898 | callback to be invoked. Doesn't rcu_barrier() therefore need to |
1899 | wait for a full grace period if there is ev 1899 | wait for a full grace period if there is even one callback posted |
1900 | anywhere in the system? 1900 | anywhere in the system? |
1901 +-------------------------------------------- 1901 +-----------------------------------------------------------------------+
1902 | **Answer**: 1902 | **Answer**: |
1903 +-------------------------------------------- 1903 +-----------------------------------------------------------------------+
1904 | Absolutely not!!! 1904 | Absolutely not!!! |
1905 | Yes, each RCU callbacks must wait for a gra 1905 | Yes, each RCU callbacks must wait for a grace period to complete, but |
1906 | it might well be partly (or even completely 1906 | it might well be partly (or even completely) finished waiting by the |
1907 | time rcu_barrier() is invoked. In that case 1907 | time rcu_barrier() is invoked. In that case, rcu_barrier() |
1908 | need only wait for the remaining portion of 1908 | need only wait for the remaining portion of the grace period to |
1909 | elapse. So even if there are quite a few ca 1909 | elapse. So even if there are quite a few callbacks posted, |
1910 | rcu_barrier() might well return quite quick 1910 | rcu_barrier() might well return quite quickly. |
1911 | 1911 | |
1912 | So if you need to wait for a grace period a 1912 | So if you need to wait for a grace period as well as for all |
1913 | pre-existing callbacks, you will need to in 1913 | pre-existing callbacks, you will need to invoke both |
1914 | synchronize_rcu() and rcu_barrier(). If lat 1914 | synchronize_rcu() and rcu_barrier(). If latency is a concern, |
1915 | you can always use workqueues to invoke the 1915 | you can always use workqueues to invoke them concurrently. |
1916 +-------------------------------------------- 1916 +-----------------------------------------------------------------------+
1917 1917
1918 Hotplug CPU 1918 Hotplug CPU
1919 ~~~~~~~~~~~ 1919 ~~~~~~~~~~~
1920 1920
1921 The Linux kernel supports CPU hotplug, which 1921 The Linux kernel supports CPU hotplug, which means that CPUs can come
1922 and go. It is of course illegal to use any RC 1922 and go. It is of course illegal to use any RCU API member from an
1923 offline CPU, with the exception of `SRCU <Sle 1923 offline CPU, with the exception of `SRCU <Sleepable RCU_>`__ read-side
1924 critical sections. This requirement was prese 1924 critical sections. This requirement was present from day one in
1925 DYNIX/ptx, but on the other hand, the Linux k 1925 DYNIX/ptx, but on the other hand, the Linux kernel's CPU-hotplug
1926 implementation is “interesting.” 1926 implementation is “interesting.”
1927 1927
1928 The Linux-kernel CPU-hotplug implementation h 1928 The Linux-kernel CPU-hotplug implementation has notifiers that are used
1929 to allow the various kernel subsystems (inclu 1929 to allow the various kernel subsystems (including RCU) to respond
1930 appropriately to a given CPU-hotplug operatio 1930 appropriately to a given CPU-hotplug operation. Most RCU operations may
1931 be invoked from CPU-hotplug notifiers, includ 1931 be invoked from CPU-hotplug notifiers, including even synchronous
1932 grace-period operations such as (synchronize_ 1932 grace-period operations such as (synchronize_rcu() and
1933 synchronize_rcu_expedited()). However, these 1933 synchronize_rcu_expedited()). However, these synchronous operations
1934 do block and therefore cannot be invoked from 1934 do block and therefore cannot be invoked from notifiers that execute via
1935 stop_machine(), specifically those between th 1935 stop_machine(), specifically those between the ``CPUHP_AP_OFFLINE``
1936 and ``CPUHP_AP_ONLINE`` states. 1936 and ``CPUHP_AP_ONLINE`` states.
1937 1937
1938 In addition, all-callback-wait operations suc 1938 In addition, all-callback-wait operations such as rcu_barrier() may
1939 not be invoked from any CPU-hotplug notifier. 1939 not be invoked from any CPU-hotplug notifier. This restriction is due
1940 to the fact that there are phases of CPU-hotp 1940 to the fact that there are phases of CPU-hotplug operations where the
1941 outgoing CPU's callbacks will not be invoked 1941 outgoing CPU's callbacks will not be invoked until after the CPU-hotplug
1942 operation ends, which could also result in de 1942 operation ends, which could also result in deadlock. Furthermore,
1943 rcu_barrier() blocks CPU-hotplug operations d 1943 rcu_barrier() blocks CPU-hotplug operations during its execution,
1944 which results in another type of deadlock whe 1944 which results in another type of deadlock when invoked from a CPU-hotplug
1945 notifier. 1945 notifier.
1946 1946
1947 Finally, RCU must avoid deadlocks due to inte 1947 Finally, RCU must avoid deadlocks due to interaction between hotplug,
1948 timers and grace period processing. It does s 1948 timers and grace period processing. It does so by maintaining its own set
1949 of books that duplicate the centrally maintai 1949 of books that duplicate the centrally maintained ``cpu_online_mask``,
1950 and also by reporting quiescent states explic 1950 and also by reporting quiescent states explicitly when a CPU goes
1951 offline. This explicit reporting of quiescen 1951 offline. This explicit reporting of quiescent states avoids any need
1952 for the force-quiescent-state loop (FQS) to r 1952 for the force-quiescent-state loop (FQS) to report quiescent states for
1953 offline CPUs. However, as a debugging measur 1953 offline CPUs. However, as a debugging measure, the FQS loop does splat
1954 if offline CPUs block an RCU grace period for 1954 if offline CPUs block an RCU grace period for too long.
1955 1955
1956 An offline CPU's quiescent state will be repo 1956 An offline CPU's quiescent state will be reported either:
1957 1957
1958 1. As the CPU goes offline using RCU's hotpl 1958 1. As the CPU goes offline using RCU's hotplug notifier (rcutree_report_cpu_dead()).
1959 2. When grace period initialization (rcu_gp_ 1959 2. When grace period initialization (rcu_gp_init()) detects a
1960 race either with CPU offlining or with a 1960 race either with CPU offlining or with a task unblocking on a leaf
1961 ``rcu_node`` structure whose CPUs are all 1961 ``rcu_node`` structure whose CPUs are all offline.
1962 1962
1963 The CPU-online path (rcutree_report_cpu_start 1963 The CPU-online path (rcutree_report_cpu_starting()) should never need to report
1964 a quiescent state for an offline CPU. Howeve 1964 a quiescent state for an offline CPU. However, as a debugging measure,
1965 it does emit a warning if a quiescent state w 1965 it does emit a warning if a quiescent state was not already reported
1966 for that CPU. 1966 for that CPU.
1967 1967
1968 During the checking/modification of RCU's hot 1968 During the checking/modification of RCU's hotplug bookkeeping, the
1969 corresponding CPU's leaf node lock is held. T 1969 corresponding CPU's leaf node lock is held. This avoids race conditions
1970 between RCU's hotplug notifier hooks, the gra 1970 between RCU's hotplug notifier hooks, the grace period initialization
1971 code, and the FQS loop, all of which refer to 1971 code, and the FQS loop, all of which refer to or modify this bookkeeping.
1972 1972
1973 Scheduler and RCU 1973 Scheduler and RCU
1974 ~~~~~~~~~~~~~~~~~ 1974 ~~~~~~~~~~~~~~~~~
1975 1975
1976 RCU makes use of kthreads, and it is necessar 1976 RCU makes use of kthreads, and it is necessary to avoid excessive CPU-time
1977 accumulation by these kthreads. This requirem 1977 accumulation by these kthreads. This requirement was no surprise, but
1978 RCU's violation of it when running context-sw 1978 RCU's violation of it when running context-switch-heavy workloads when
1979 built with ``CONFIG_NO_HZ_FULL=y`` `did come 1979 built with ``CONFIG_NO_HZ_FULL=y`` `did come as a surprise
1980 [PDF] <http://www.rdrop.com/users/paulmck/sca 1980 [PDF] <http://www.rdrop.com/users/paulmck/scalability/paper/BareMetal.2015.01.15b.pdf>`__.
1981 RCU has made good progress towards meeting th 1981 RCU has made good progress towards meeting this requirement, even for
1982 context-switch-heavy ``CONFIG_NO_HZ_FULL=y`` 1982 context-switch-heavy ``CONFIG_NO_HZ_FULL=y`` workloads, but there is
1983 room for further improvement. 1983 room for further improvement.
1984 1984
1985 There is no longer any prohibition against ho 1985 There is no longer any prohibition against holding any of
1986 scheduler's runqueue or priority-inheritance 1986 scheduler's runqueue or priority-inheritance spinlocks across an
1987 rcu_read_unlock(), even if interrupts and pre 1987 rcu_read_unlock(), even if interrupts and preemption were enabled
1988 somewhere within the corresponding RCU read-s 1988 somewhere within the corresponding RCU read-side critical section.
1989 Therefore, it is now perfectly legal to execu 1989 Therefore, it is now perfectly legal to execute rcu_read_lock()
1990 with preemption enabled, acquire one of the s 1990 with preemption enabled, acquire one of the scheduler locks, and hold
1991 that lock across the matching rcu_read_unlock 1991 that lock across the matching rcu_read_unlock().
1992 1992
1993 Similarly, the RCU flavor consolidation has r 1993 Similarly, the RCU flavor consolidation has removed the need for negative
1994 nesting. The fact that interrupt-disabled re 1994 nesting. The fact that interrupt-disabled regions of code act as RCU
1995 read-side critical sections implicitly avoids 1995 read-side critical sections implicitly avoids earlier issues that used
1996 to result in destructive recursion via interr 1996 to result in destructive recursion via interrupt handler's use of RCU.
1997 1997
1998 Tracing and RCU 1998 Tracing and RCU
1999 ~~~~~~~~~~~~~~~ 1999 ~~~~~~~~~~~~~~~
2000 2000
2001 It is possible to use tracing on RCU code, bu 2001 It is possible to use tracing on RCU code, but tracing itself uses RCU.
2002 For this reason, rcu_dereference_raw_check() 2002 For this reason, rcu_dereference_raw_check() is provided for use
2003 by tracing, which avoids the destructive recu 2003 by tracing, which avoids the destructive recursion that could otherwise
2004 ensue. This API is also used by virtualizatio 2004 ensue. This API is also used by virtualization in some architectures,
2005 where RCU readers execute in environments in 2005 where RCU readers execute in environments in which tracing cannot be
2006 used. The tracing folks both located the requ 2006 used. The tracing folks both located the requirement and provided the
2007 needed fix, so this surprise requirement was 2007 needed fix, so this surprise requirement was relatively painless.
2008 2008
2009 Accesses to User Memory and RCU 2009 Accesses to User Memory and RCU
2010 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 2010 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2011 2011
2012 The kernel needs to access user-space memory, 2012 The kernel needs to access user-space memory, for example, to access data
2013 referenced by system-call parameters. The ge 2013 referenced by system-call parameters. The get_user() macro does this job.
2014 2014
2015 However, user-space memory might well be page 2015 However, user-space memory might well be paged out, which means that
2016 get_user() might well page-fault and thus blo 2016 get_user() might well page-fault and thus block while waiting for the
2017 resulting I/O to complete. It would be a ver 2017 resulting I/O to complete. It would be a very bad thing for the compiler to
2018 reorder a get_user() invocation into an RCU r 2018 reorder a get_user() invocation into an RCU read-side critical section.
2019 2019
2020 For example, suppose that the source code loo 2020 For example, suppose that the source code looked like this:
2021 2021
2022 :: 2022 ::
2023 2023
2024 1 rcu_read_lock(); 2024 1 rcu_read_lock();
2025 2 p = rcu_dereference(gp); 2025 2 p = rcu_dereference(gp);
2026 3 v = p->value; 2026 3 v = p->value;
2027 4 rcu_read_unlock(); 2027 4 rcu_read_unlock();
2028 5 get_user(user_v, user_p); 2028 5 get_user(user_v, user_p);
2029 6 do_something_with(v, user_v); 2029 6 do_something_with(v, user_v);
2030 2030
2031 The compiler must not be permitted to transfo 2031 The compiler must not be permitted to transform this source code into
2032 the following: 2032 the following:
2033 2033
2034 :: 2034 ::
2035 2035
2036 1 rcu_read_lock(); 2036 1 rcu_read_lock();
2037 2 p = rcu_dereference(gp); 2037 2 p = rcu_dereference(gp);
2038 3 get_user(user_v, user_p); // BUG: PO 2038 3 get_user(user_v, user_p); // BUG: POSSIBLE PAGE FAULT!!!
2039 4 v = p->value; 2039 4 v = p->value;
2040 5 rcu_read_unlock(); 2040 5 rcu_read_unlock();
2041 6 do_something_with(v, user_v); 2041 6 do_something_with(v, user_v);
2042 2042
2043 If the compiler did make this transformation 2043 If the compiler did make this transformation in a ``CONFIG_PREEMPTION=n`` kernel
2044 build, and if get_user() did page fault, the 2044 build, and if get_user() did page fault, the result would be a quiescent
2045 state in the middle of an RCU read-side criti 2045 state in the middle of an RCU read-side critical section. This misplaced
2046 quiescent state could result in line 4 being 2046 quiescent state could result in line 4 being a use-after-free access,
2047 which could be bad for your kernel's actuaria 2047 which could be bad for your kernel's actuarial statistics. Similar examples
2048 can be constructed with the call to get_user( 2048 can be constructed with the call to get_user() preceding the
2049 rcu_read_lock(). 2049 rcu_read_lock().
2050 2050
2051 Unfortunately, get_user() doesn't have any pa 2051 Unfortunately, get_user() doesn't have any particular ordering properties,
2052 and in some architectures the underlying ``as 2052 and in some architectures the underlying ``asm`` isn't even marked
2053 ``volatile``. And even if it was marked ``vo 2053 ``volatile``. And even if it was marked ``volatile``, the above access to
2054 ``p->value`` is not volatile, so the compiler 2054 ``p->value`` is not volatile, so the compiler would not have any reason to keep
2055 those two accesses in order. 2055 those two accesses in order.
2056 2056
2057 Therefore, the Linux-kernel definitions of rc 2057 Therefore, the Linux-kernel definitions of rcu_read_lock() and
2058 rcu_read_unlock() must act as compiler barrie 2058 rcu_read_unlock() must act as compiler barriers, at least for outermost
2059 instances of rcu_read_lock() and rcu_read_unl 2059 instances of rcu_read_lock() and rcu_read_unlock() within a nested set
2060 of RCU read-side critical sections. 2060 of RCU read-side critical sections.
2061 2061
2062 Energy Efficiency 2062 Energy Efficiency
2063 ~~~~~~~~~~~~~~~~~ 2063 ~~~~~~~~~~~~~~~~~
2064 2064
2065 Interrupting idle CPUs is considered socially 2065 Interrupting idle CPUs is considered socially unacceptable, especially
2066 by people with battery-powered embedded syste 2066 by people with battery-powered embedded systems. RCU therefore conserves
2067 energy by detecting which CPUs are idle, incl 2067 energy by detecting which CPUs are idle, including tracking CPUs that
2068 have been interrupted from idle. This is a la 2068 have been interrupted from idle. This is a large part of the
2069 energy-efficiency requirement, so I learned o 2069 energy-efficiency requirement, so I learned of this via an irate phone
2070 call. 2070 call.
2071 2071
2072 Because RCU avoids interrupting idle CPUs, it 2072 Because RCU avoids interrupting idle CPUs, it is illegal to execute an
2073 RCU read-side critical section on an idle CPU 2073 RCU read-side critical section on an idle CPU. (Kernels built with
2074 ``CONFIG_PROVE_RCU=y`` will splat if you try 2074 ``CONFIG_PROVE_RCU=y`` will splat if you try it.)
2075 2075
2076 It is similarly socially unacceptable to inte 2076 It is similarly socially unacceptable to interrupt an ``nohz_full`` CPU
2077 running in userspace. RCU must therefore trac 2077 running in userspace. RCU must therefore track ``nohz_full`` userspace
2078 execution. RCU must therefore be able to samp 2078 execution. RCU must therefore be able to sample state at two points in
2079 time, and be able to determine whether or not 2079 time, and be able to determine whether or not some other CPU spent any
2080 time idle and/or executing in userspace. 2080 time idle and/or executing in userspace.
2081 2081
2082 These energy-efficiency requirements have pro 2082 These energy-efficiency requirements have proven quite difficult to
2083 understand and to meet, for example, there ha 2083 understand and to meet, for example, there have been more than five
2084 clean-sheet rewrites of RCU's energy-efficien 2084 clean-sheet rewrites of RCU's energy-efficiency code, the last of which
2085 was finally able to demonstrate `real energy 2085 was finally able to demonstrate `real energy savings running on real
2086 hardware 2086 hardware
2087 [PDF] <http://www.rdrop.com/users/paulmck/rea 2087 [PDF] <http://www.rdrop.com/users/paulmck/realtime/paper/AMPenergy.2013.04.19a.pdf>`__.
2088 As noted earlier, I learned of many of these 2088 As noted earlier, I learned of many of these requirements via angry
2089 phone calls: Flaming me on the Linux-kernel m 2089 phone calls: Flaming me on the Linux-kernel mailing list was apparently
2090 not sufficient to fully vent their ire at RCU 2090 not sufficient to fully vent their ire at RCU's energy-efficiency bugs!
2091 2091
2092 Scheduling-Clock Interrupts and RCU 2092 Scheduling-Clock Interrupts and RCU
2093 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 2093 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2094 2094
2095 The kernel transitions between in-kernel non- 2095 The kernel transitions between in-kernel non-idle execution, userspace
2096 execution, and the idle loop. Depending on ke 2096 execution, and the idle loop. Depending on kernel configuration, RCU
2097 handles these states differently: 2097 handles these states differently:
2098 2098
2099 +-----------------+------------------+------- 2099 +-----------------+------------------+------------------+-----------------+
2100 | ``HZ`` Kconfig | In-Kernel | Usermo 2100 | ``HZ`` Kconfig | In-Kernel | Usermode | Idle |
2101 +=================+==================+======= 2101 +=================+==================+==================+=================+
2102 | ``HZ_PERIODIC`` | Can rely on | Can re 2102 | ``HZ_PERIODIC`` | Can rely on | Can rely on | Can rely on |
2103 | | scheduling-clock | schedu 2103 | | scheduling-clock | scheduling-clock | RCU's |
2104 | | interrupt. | interr 2104 | | interrupt. | interrupt and | dyntick-idle |
2105 | | | its de 2105 | | | its detection | detection. |
2106 | | | of int 2106 | | | of interrupt | |
2107 | | | from u 2107 | | | from usermode. | |
2108 +-----------------+------------------+------- 2108 +-----------------+------------------+------------------+-----------------+
2109 | ``NO_HZ_IDLE`` | Can rely on | Can re 2109 | ``NO_HZ_IDLE`` | Can rely on | Can rely on | Can rely on |
2110 | | scheduling-clock | schedu 2110 | | scheduling-clock | scheduling-clock | RCU's |
2111 | | interrupt. | interr 2111 | | interrupt. | interrupt and | dyntick-idle |
2112 | | | its de 2112 | | | its detection | detection. |
2113 | | | of int 2113 | | | of interrupt | |
2114 | | | from u 2114 | | | from usermode. | |
2115 +-----------------+------------------+------- 2115 +-----------------+------------------+------------------+-----------------+
2116 | ``NO_HZ_FULL`` | Can only | Can re 2116 | ``NO_HZ_FULL`` | Can only | Can rely on | Can rely on |
2117 | | sometimes rely | RCU's 2117 | | sometimes rely | RCU's | RCU's |
2118 | | on | dyntic 2118 | | on | dyntick-idle | dyntick-idle |
2119 | | scheduling-clock | detect 2119 | | scheduling-clock | detection. | detection. |
2120 | | interrupt. In | 2120 | | interrupt. In | | |
2121 | | other cases, it | 2121 | | other cases, it | | |
2122 | | is necessary to | 2122 | | is necessary to | | |
2123 | | bound kernel | 2123 | | bound kernel | | |
2124 | | execution times | 2124 | | execution times | | |
2125 | | and/or use | 2125 | | and/or use | | |
2126 | | IPIs. | 2126 | | IPIs. | | |
2127 +-----------------+------------------+------- 2127 +-----------------+------------------+------------------+-----------------+
2128 2128
2129 +-------------------------------------------- 2129 +-----------------------------------------------------------------------+
2130 | **Quick Quiz**: 2130 | **Quick Quiz**: |
2131 +-------------------------------------------- 2131 +-----------------------------------------------------------------------+
2132 | Why can't ``NO_HZ_FULL`` in-kernel executio 2132 | Why can't ``NO_HZ_FULL`` in-kernel execution rely on the |
2133 | scheduling-clock interrupt, just like ``HZ_ 2133 | scheduling-clock interrupt, just like ``HZ_PERIODIC`` and |
2134 | ``NO_HZ_IDLE`` do? 2134 | ``NO_HZ_IDLE`` do? |
2135 +-------------------------------------------- 2135 +-----------------------------------------------------------------------+
2136 | **Answer**: 2136 | **Answer**: |
2137 +-------------------------------------------- 2137 +-----------------------------------------------------------------------+
2138 | Because, as a performance optimization, ``N 2138 | Because, as a performance optimization, ``NO_HZ_FULL`` does not |
2139 | necessarily re-enable the scheduling-clock 2139 | necessarily re-enable the scheduling-clock interrupt on entry to each |
2140 | and every system call. 2140 | and every system call. |
2141 +-------------------------------------------- 2141 +-----------------------------------------------------------------------+
2142 2142
2143 However, RCU must be reliably informed as to 2143 However, RCU must be reliably informed as to whether any given CPU is
2144 currently in the idle loop, and, for ``NO_HZ_ 2144 currently in the idle loop, and, for ``NO_HZ_FULL``, also whether that
2145 CPU is executing in usermode, as discussed 2145 CPU is executing in usermode, as discussed
2146 `earlier <Energy Efficiency_>`__. It also req 2146 `earlier <Energy Efficiency_>`__. It also requires that the
2147 scheduling-clock interrupt be enabled when RC 2147 scheduling-clock interrupt be enabled when RCU needs it to be:
2148 2148
2149 #. If a CPU is either idle or executing in us 2149 #. If a CPU is either idle or executing in usermode, and RCU believes it
2150 is non-idle, the scheduling-clock tick had 2150 is non-idle, the scheduling-clock tick had better be running.
2151 Otherwise, you will get RCU CPU stall warn 2151 Otherwise, you will get RCU CPU stall warnings. Or at best, very long
2152 (11-second) grace periods, with a pointles 2152 (11-second) grace periods, with a pointless IPI waking the CPU from
2153 time to time. 2153 time to time.
2154 #. If a CPU is in a portion of the kernel tha 2154 #. If a CPU is in a portion of the kernel that executes RCU read-side
2155 critical sections, and RCU believes this C 2155 critical sections, and RCU believes this CPU to be idle, you will get
2156 random memory corruption. **DON'T DO THIS! 2156 random memory corruption. **DON'T DO THIS!!!**
2157 This is one reason to test with lockdep, w 2157 This is one reason to test with lockdep, which will complain about
2158 this sort of thing. 2158 this sort of thing.
2159 #. If a CPU is in a portion of the kernel tha 2159 #. If a CPU is in a portion of the kernel that is absolutely positively
2160 no-joking guaranteed to never execute any 2160 no-joking guaranteed to never execute any RCU read-side critical
2161 sections, and RCU believes this CPU to be 2161 sections, and RCU believes this CPU to be idle, no problem. This
2162 sort of thing is used by some architecture 2162 sort of thing is used by some architectures for light-weight
2163 exception handlers, which can then avoid t 2163 exception handlers, which can then avoid the overhead of
2164 ct_irq_enter() and ct_irq_exit() at except 2164 ct_irq_enter() and ct_irq_exit() at exception entry and
2165 exit, respectively. Some go further and av 2165 exit, respectively. Some go further and avoid the entireties of
2166 irq_enter() and irq_exit(). 2166 irq_enter() and irq_exit().
2167 Just make very sure you are running some o 2167 Just make very sure you are running some of your tests with
2168 ``CONFIG_PROVE_RCU=y``, just in case one o 2168 ``CONFIG_PROVE_RCU=y``, just in case one of your code paths was in
2169 fact joking about not doing RCU read-side 2169 fact joking about not doing RCU read-side critical sections.
2170 #. If a CPU is executing in the kernel with t 2170 #. If a CPU is executing in the kernel with the scheduling-clock
2171 interrupt disabled and RCU believes this C 2171 interrupt disabled and RCU believes this CPU to be non-idle, and if
2172 the CPU goes idle (from an RCU perspective 2172 the CPU goes idle (from an RCU perspective) every few jiffies, no
2173 problem. It is usually OK for there to be 2173 problem. It is usually OK for there to be the occasional gap between
2174 idle periods of up to a second or so. 2174 idle periods of up to a second or so.
2175 If the gap grows too long, you get RCU CPU 2175 If the gap grows too long, you get RCU CPU stall warnings.
2176 #. If a CPU is either idle or executing in us 2176 #. If a CPU is either idle or executing in usermode, and RCU believes it
2177 to be idle, of course no problem. 2177 to be idle, of course no problem.
2178 #. If a CPU is executing in the kernel, the k 2178 #. If a CPU is executing in the kernel, the kernel code path is passing
2179 through quiescent states at a reasonable f 2179 through quiescent states at a reasonable frequency (preferably about
2180 once per few jiffies, but the occasional e 2180 once per few jiffies, but the occasional excursion to a second or so
2181 is usually OK) and the scheduling-clock in 2181 is usually OK) and the scheduling-clock interrupt is enabled, of
2182 course no problem. 2182 course no problem.
2183 If the gap between a successive pair of qu 2183 If the gap between a successive pair of quiescent states grows too
2184 long, you get RCU CPU stall warnings. 2184 long, you get RCU CPU stall warnings.
2185 2185
2186 +-------------------------------------------- 2186 +-----------------------------------------------------------------------+
2187 | **Quick Quiz**: 2187 | **Quick Quiz**: |
2188 +-------------------------------------------- 2188 +-----------------------------------------------------------------------+
2189 | But what if my driver has a hardware interr 2189 | But what if my driver has a hardware interrupt handler that can run |
2190 | for many seconds? I cannot invoke schedule( 2190 | for many seconds? I cannot invoke schedule() from an hardware |
2191 | interrupt handler, after all! 2191 | interrupt handler, after all! |
2192 +-------------------------------------------- 2192 +-----------------------------------------------------------------------+
2193 | **Answer**: 2193 | **Answer**: |
2194 +-------------------------------------------- 2194 +-----------------------------------------------------------------------+
2195 | One approach is to do ``ct_irq_exit();ct_ir 2195 | One approach is to do ``ct_irq_exit();ct_irq_enter();`` every so |
2196 | often. But given that long-running interrup 2196 | often. But given that long-running interrupt handlers can cause other |
2197 | problems, not least for response time, shou 2197 | problems, not least for response time, shouldn't you work to keep |
2198 | your interrupt handler's runtime within rea 2198 | your interrupt handler's runtime within reasonable bounds? |
2199 +-------------------------------------------- 2199 +-----------------------------------------------------------------------+
2200 2200
2201 But as long as RCU is properly informed of ke 2201 But as long as RCU is properly informed of kernel state transitions
2202 between in-kernel execution, usermode executi 2202 between in-kernel execution, usermode execution, and idle, and as long
2203 as the scheduling-clock interrupt is enabled 2203 as the scheduling-clock interrupt is enabled when RCU needs it to be,
2204 you can rest assured that the bugs you encoun 2204 you can rest assured that the bugs you encounter will be in some other
2205 part of RCU or some other part of the kernel! 2205 part of RCU or some other part of the kernel!
2206 2206
2207 Memory Efficiency 2207 Memory Efficiency
2208 ~~~~~~~~~~~~~~~~~ 2208 ~~~~~~~~~~~~~~~~~
2209 2209
2210 Although small-memory non-realtime systems ca 2210 Although small-memory non-realtime systems can simply use Tiny RCU, code
2211 size is only one aspect of memory efficiency. 2211 size is only one aspect of memory efficiency. Another aspect is the size
2212 of the ``rcu_head`` structure used by call_rc 2212 of the ``rcu_head`` structure used by call_rcu() and
2213 kfree_rcu(). Although this structure contains 2213 kfree_rcu(). Although this structure contains nothing more than a
2214 pair of pointers, it does appear in many RCU- 2214 pair of pointers, it does appear in many RCU-protected data structures,
2215 including some that are size critical. The `` 2215 including some that are size critical. The ``page`` structure is a case
2216 in point, as evidenced by the many occurrence 2216 in point, as evidenced by the many occurrences of the ``union`` keyword
2217 within that structure. 2217 within that structure.
2218 2218
2219 This need for memory efficiency is one reason 2219 This need for memory efficiency is one reason that RCU uses hand-crafted
2220 singly linked lists to track the ``rcu_head`` 2220 singly linked lists to track the ``rcu_head`` structures that are
2221 waiting for a grace period to elapse. It is a 2221 waiting for a grace period to elapse. It is also the reason why
2222 ``rcu_head`` structures do not contain debug 2222 ``rcu_head`` structures do not contain debug information, such as fields
2223 tracking the file and line of the call_rcu() 2223 tracking the file and line of the call_rcu() or kfree_rcu() that
2224 posted them. Although this information might 2224 posted them. Although this information might appear in debug-only kernel
2225 builds at some point, in the meantime, the `` 2225 builds at some point, in the meantime, the ``->func`` field will often
2226 provide the needed debug information. 2226 provide the needed debug information.
2227 2227
2228 However, in some cases, the need for memory e 2228 However, in some cases, the need for memory efficiency leads to even
2229 more extreme measures. Returning to the ``pag 2229 more extreme measures. Returning to the ``page`` structure, the
2230 ``rcu_head`` field shares storage with a grea 2230 ``rcu_head`` field shares storage with a great many other structures
2231 that are used at various points in the corres 2231 that are used at various points in the corresponding page's lifetime. In
2232 order to correctly resolve certain `race 2232 order to correctly resolve certain `race
2233 conditions <https://lore.kernel.org/r/14399761 2233 conditions <https://lore.kernel.org/r/1439976106-137226-1-git-send-email-kirill.shutemov@linux.intel.com">https://lore.kernel.org/r/1439976106-137226-1-git-send-email-kirill.shutemov@linux.intel.com>`__,
2234 the Linux kernel's memory-management subsyste 2234 the Linux kernel's memory-management subsystem needs a particular bit to
2235 remain zero during all phases of grace-period 2235 remain zero during all phases of grace-period processing, and that bit
2236 happens to map to the bottom bit of the ``rcu 2236 happens to map to the bottom bit of the ``rcu_head`` structure's
2237 ``->next`` field. RCU makes this guarantee as 2237 ``->next`` field. RCU makes this guarantee as long as call_rcu() is
2238 used to post the callback, as opposed to kfre 2238 used to post the callback, as opposed to kfree_rcu() or some future
2239 “lazy” variant of call_rcu() that might o 2239 “lazy” variant of call_rcu() that might one day be created for
2240 energy-efficiency purposes. 2240 energy-efficiency purposes.
2241 2241
2242 That said, there are limits. RCU requires tha 2242 That said, there are limits. RCU requires that the ``rcu_head``
2243 structure be aligned to a two-byte boundary, 2243 structure be aligned to a two-byte boundary, and passing a misaligned
2244 ``rcu_head`` structure to one of the call_rcu 2244 ``rcu_head`` structure to one of the call_rcu() family of functions
2245 will result in a splat. It is therefore neces 2245 will result in a splat. It is therefore necessary to exercise caution
2246 when packing structures containing fields of 2246 when packing structures containing fields of type ``rcu_head``. Why not
2247 a four-byte or even eight-byte alignment requ 2247 a four-byte or even eight-byte alignment requirement? Because the m68k
2248 architecture provides only two-byte alignment 2248 architecture provides only two-byte alignment, and thus acts as
2249 alignment's least common denominator. 2249 alignment's least common denominator.
2250 2250
2251 The reason for reserving the bottom bit of po 2251 The reason for reserving the bottom bit of pointers to ``rcu_head``
2252 structures is to leave the door open to “la 2252 structures is to leave the door open to “lazy” callbacks whose
2253 invocations can safely be deferred. Deferring 2253 invocations can safely be deferred. Deferring invocation could
2254 potentially have energy-efficiency benefits, 2254 potentially have energy-efficiency benefits, but only if the rate of
2255 non-lazy callbacks decreases significantly fo 2255 non-lazy callbacks decreases significantly for some important workload.
2256 In the meantime, reserving the bottom bit kee 2256 In the meantime, reserving the bottom bit keeps this option open in case
2257 it one day becomes useful. 2257 it one day becomes useful.
2258 2258
2259 Performance, Scalability, Response Time, and 2259 Performance, Scalability, Response Time, and Reliability
2260 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 2260 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2261 2261
2262 Expanding on the `earlier 2262 Expanding on the `earlier
2263 discussion <Performance and Scalability_>`__, 2263 discussion <Performance and Scalability_>`__, RCU is used heavily by
2264 hot code paths in performance-critical portio 2264 hot code paths in performance-critical portions of the Linux kernel's
2265 networking, security, virtualization, and sch 2265 networking, security, virtualization, and scheduling code paths. RCU
2266 must therefore use efficient implementations, 2266 must therefore use efficient implementations, especially in its
2267 read-side primitives. To that end, it would b 2267 read-side primitives. To that end, it would be good if preemptible RCU's
2268 implementation of rcu_read_lock() could be in 2268 implementation of rcu_read_lock() could be inlined, however, doing
2269 this requires resolving ``#include`` issues w 2269 this requires resolving ``#include`` issues with the ``task_struct``
2270 structure. 2270 structure.
2271 2271
2272 The Linux kernel supports hardware configurat 2272 The Linux kernel supports hardware configurations with up to 4096 CPUs,
2273 which means that RCU must be extremely scalab 2273 which means that RCU must be extremely scalable. Algorithms that involve
2274 frequent acquisitions of global locks or freq 2274 frequent acquisitions of global locks or frequent atomic operations on
2275 global variables simply cannot be tolerated w 2275 global variables simply cannot be tolerated within the RCU
2276 implementation. RCU therefore makes heavy use 2276 implementation. RCU therefore makes heavy use of a combining tree based
2277 on the ``rcu_node`` structure. RCU is require 2277 on the ``rcu_node`` structure. RCU is required to tolerate all CPUs
2278 continuously invoking any combination of RCU' 2278 continuously invoking any combination of RCU's runtime primitives with
2279 minimal per-operation overhead. In fact, in m 2279 minimal per-operation overhead. In fact, in many cases, increasing load
2280 must *decrease* the per-operation overhead, w 2280 must *decrease* the per-operation overhead, witness the batching
2281 optimizations for synchronize_rcu(), call_rcu 2281 optimizations for synchronize_rcu(), call_rcu(),
2282 synchronize_rcu_expedited(), and rcu_barrier( 2282 synchronize_rcu_expedited(), and rcu_barrier(). As a general
2283 rule, RCU must cheerfully accept whatever the 2283 rule, RCU must cheerfully accept whatever the rest of the Linux kernel
2284 decides to throw at it. 2284 decides to throw at it.
2285 2285
2286 The Linux kernel is used for real-time worklo 2286 The Linux kernel is used for real-time workloads, especially in
2287 conjunction with the `-rt 2287 conjunction with the `-rt
2288 patchset <https://wiki.linuxfoundation.org/re 2288 patchset <https://wiki.linuxfoundation.org/realtime/>`__. The
2289 real-time-latency response requirements are s 2289 real-time-latency response requirements are such that the traditional
2290 approach of disabling preemption across RCU r 2290 approach of disabling preemption across RCU read-side critical sections
2291 is inappropriate. Kernels built with ``CONFIG 2291 is inappropriate. Kernels built with ``CONFIG_PREEMPTION=y`` therefore use
2292 an RCU implementation that allows RCU read-si 2292 an RCU implementation that allows RCU read-side critical sections to be
2293 preempted. This requirement made its presence 2293 preempted. This requirement made its presence known after users made it
2294 clear that an earlier `real-time 2294 clear that an earlier `real-time
2295 patch <https://lwn.net/Articles/107930/>`__ d 2295 patch <https://lwn.net/Articles/107930/>`__ did not meet their needs, in
2296 conjunction with some `RCU 2296 conjunction with some `RCU
2297 issues <https://lore.kernel.org/r/200503180020 2297 issues <https://lore.kernel.org/r/20050318002026.GA2693@us.ibm.com">https://lore.kernel.org/r/20050318002026.GA2693@us.ibm.com>`__
2298 encountered by a very early version of the -r 2298 encountered by a very early version of the -rt patchset.
2299 2299
2300 In addition, RCU must make do with a sub-100- 2300 In addition, RCU must make do with a sub-100-microsecond real-time
2301 latency budget. In fact, on smaller systems w 2301 latency budget. In fact, on smaller systems with the -rt patchset, the
2302 Linux kernel provides sub-20-microsecond real 2302 Linux kernel provides sub-20-microsecond real-time latencies for the
2303 whole kernel, including RCU. RCU's scalabilit 2303 whole kernel, including RCU. RCU's scalability and latency must
2304 therefore be sufficient for these sorts of co 2304 therefore be sufficient for these sorts of configurations. To my
2305 surprise, the sub-100-microsecond real-time l 2305 surprise, the sub-100-microsecond real-time latency budget `applies to
2306 even the largest systems 2306 even the largest systems
2307 [PDF] <http://www.rdrop.com/users/paulmck/rea 2307 [PDF] <http://www.rdrop.com/users/paulmck/realtime/paper/bigrt.2013.01.31a.LCA.pdf>`__,
2308 up to and including systems with 4096 CPUs. T 2308 up to and including systems with 4096 CPUs. This real-time requirement
2309 motivated the grace-period kthread, which als 2309 motivated the grace-period kthread, which also simplified handling of a
2310 number of race conditions. 2310 number of race conditions.
2311 2311
2312 RCU must avoid degrading real-time response f 2312 RCU must avoid degrading real-time response for CPU-bound threads,
2313 whether executing in usermode (which is one u 2313 whether executing in usermode (which is one use case for
2314 ``CONFIG_NO_HZ_FULL=y``) or in the kernel. Th 2314 ``CONFIG_NO_HZ_FULL=y``) or in the kernel. That said, CPU-bound loops in
2315 the kernel must execute cond_resched() at lea 2315 the kernel must execute cond_resched() at least once per few tens of
2316 milliseconds in order to avoid receiving an I 2316 milliseconds in order to avoid receiving an IPI from RCU.
2317 2317
2318 Finally, RCU's status as a synchronization pr 2318 Finally, RCU's status as a synchronization primitive means that any RCU
2319 failure can result in arbitrary memory corrup 2319 failure can result in arbitrary memory corruption that can be extremely
2320 difficult to debug. This means that RCU must 2320 difficult to debug. This means that RCU must be extremely reliable,
2321 which in practice also means that RCU must ha 2321 which in practice also means that RCU must have an aggressive
2322 stress-test suite. This stress-test suite is 2322 stress-test suite. This stress-test suite is called ``rcutorture``.
2323 2323
2324 Although the need for ``rcutorture`` was no s 2324 Although the need for ``rcutorture`` was no surprise, the current
2325 immense popularity of the Linux kernel is pos 2325 immense popularity of the Linux kernel is posing interesting—and perhaps
2326 unprecedented—validation challenges. To see 2326 unprecedented—validation challenges. To see this, keep in mind that
2327 there are well over one billion instances of 2327 there are well over one billion instances of the Linux kernel running
2328 today, given Android smartphones, Linux-power 2328 today, given Android smartphones, Linux-powered televisions, and
2329 servers. This number can be expected to incre 2329 servers. This number can be expected to increase sharply with the advent
2330 of the celebrated Internet of Things. 2330 of the celebrated Internet of Things.
2331 2331
2332 Suppose that RCU contains a race condition th 2332 Suppose that RCU contains a race condition that manifests on average
2333 once per million years of runtime. This bug w 2333 once per million years of runtime. This bug will be occurring about
2334 three times per *day* across the installed ba 2334 three times per *day* across the installed base. RCU could simply hide
2335 behind hardware error rates, given that no on 2335 behind hardware error rates, given that no one should really expect
2336 their smartphone to last for a million years. 2336 their smartphone to last for a million years. However, anyone taking too
2337 much comfort from this thought should conside 2337 much comfort from this thought should consider the fact that in most
2338 jurisdictions, a successful multi-year test o 2338 jurisdictions, a successful multi-year test of a given mechanism, which
2339 might include a Linux kernel, suffices for a 2339 might include a Linux kernel, suffices for a number of types of
2340 safety-critical certifications. In fact, rumo 2340 safety-critical certifications. In fact, rumor has it that the Linux
2341 kernel is already being used in production fo 2341 kernel is already being used in production for safety-critical
2342 applications. I don't know about you, but I w 2342 applications. I don't know about you, but I would feel quite bad if a
2343 bug in RCU killed someone. Which might explai 2343 bug in RCU killed someone. Which might explain my recent focus on
2344 validation and verification. 2344 validation and verification.
2345 2345
2346 Other RCU Flavors 2346 Other RCU Flavors
2347 ----------------- 2347 -----------------
2348 2348
2349 One of the more surprising things about RCU i 2349 One of the more surprising things about RCU is that there are now no
2350 fewer than five *flavors*, or API families. I 2350 fewer than five *flavors*, or API families. In addition, the primary
2351 flavor that has been the sole focus up to thi 2351 flavor that has been the sole focus up to this point has two different
2352 implementations, non-preemptible and preempti 2352 implementations, non-preemptible and preemptible. The other four flavors
2353 are listed below, with requirements for each 2353 are listed below, with requirements for each described in a separate
2354 section. 2354 section.
2355 2355
2356 #. `Bottom-Half Flavor (Historical)`_ 2356 #. `Bottom-Half Flavor (Historical)`_
2357 #. `Sched Flavor (Historical)`_ 2357 #. `Sched Flavor (Historical)`_
2358 #. `Sleepable RCU`_ 2358 #. `Sleepable RCU`_
2359 #. `Tasks RCU`_ 2359 #. `Tasks RCU`_
2360 #. `Tasks Trace RCU`_ 2360 #. `Tasks Trace RCU`_
2361 2361
2362 Bottom-Half Flavor (Historical) 2362 Bottom-Half Flavor (Historical)
2363 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 2363 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2364 2364
2365 The RCU-bh flavor of RCU has since been expre 2365 The RCU-bh flavor of RCU has since been expressed in terms of the other
2366 RCU flavors as part of a consolidation of the 2366 RCU flavors as part of a consolidation of the three flavors into a
2367 single flavor. The read-side API remains, and 2367 single flavor. The read-side API remains, and continues to disable
2368 softirq and to be accounted for by lockdep. M 2368 softirq and to be accounted for by lockdep. Much of the material in this
2369 section is therefore strictly historical in n 2369 section is therefore strictly historical in nature.
2370 2370
2371 The softirq-disable (AKA “bottom-half”, h 2371 The softirq-disable (AKA “bottom-half”, hence the “_bh” abbreviations)
2372 flavor of RCU, or *RCU-bh*, was developed by 2372 flavor of RCU, or *RCU-bh*, was developed by Dipankar Sarma to provide a
2373 flavor of RCU that could withstand the networ 2373 flavor of RCU that could withstand the network-based denial-of-service
2374 attacks researched by Robert Olsson. These at 2374 attacks researched by Robert Olsson. These attacks placed so much
2375 networking load on the system that some of th 2375 networking load on the system that some of the CPUs never exited softirq
2376 execution, which in turn prevented those CPUs 2376 execution, which in turn prevented those CPUs from ever executing a
2377 context switch, which, in the RCU implementat 2377 context switch, which, in the RCU implementation of that time, prevented
2378 grace periods from ever ending. The result wa 2378 grace periods from ever ending. The result was an out-of-memory
2379 condition and a system hang. 2379 condition and a system hang.
2380 2380
2381 The solution was the creation of RCU-bh, whic 2381 The solution was the creation of RCU-bh, which does
2382 local_bh_disable() across its read-side criti 2382 local_bh_disable() across its read-side critical sections, and which
2383 uses the transition from one type of softirq 2383 uses the transition from one type of softirq processing to another as a
2384 quiescent state in addition to context switch 2384 quiescent state in addition to context switch, idle, user mode, and
2385 offline. This means that RCU-bh grace periods 2385 offline. This means that RCU-bh grace periods can complete even when
2386 some of the CPUs execute in softirq indefinit 2386 some of the CPUs execute in softirq indefinitely, thus allowing
2387 algorithms based on RCU-bh to withstand netwo 2387 algorithms based on RCU-bh to withstand network-based denial-of-service
2388 attacks. 2388 attacks.
2389 2389
2390 Because rcu_read_lock_bh() and rcu_read_unloc 2390 Because rcu_read_lock_bh() and rcu_read_unlock_bh() disable and
2391 re-enable softirq handlers, any attempt to st 2391 re-enable softirq handlers, any attempt to start a softirq handlers
2392 during the RCU-bh read-side critical section 2392 during the RCU-bh read-side critical section will be deferred. In this
2393 case, rcu_read_unlock_bh() will invoke softir 2393 case, rcu_read_unlock_bh() will invoke softirq processing, which can
2394 take considerable time. One can of course arg 2394 take considerable time. One can of course argue that this softirq
2395 overhead should be associated with the code f 2395 overhead should be associated with the code following the RCU-bh
2396 read-side critical section rather than rcu_re 2396 read-side critical section rather than rcu_read_unlock_bh(), but the
2397 fact is that most profiling tools cannot be e 2397 fact is that most profiling tools cannot be expected to make this sort
2398 of fine distinction. For example, suppose tha 2398 of fine distinction. For example, suppose that a three-millisecond-long
2399 RCU-bh read-side critical section executes du 2399 RCU-bh read-side critical section executes during a time of heavy
2400 networking load. There will very likely be an 2400 networking load. There will very likely be an attempt to invoke at least
2401 one softirq handler during that three millise 2401 one softirq handler during that three milliseconds, but any such
2402 invocation will be delayed until the time of 2402 invocation will be delayed until the time of the
2403 rcu_read_unlock_bh(). This can of course make 2403 rcu_read_unlock_bh(). This can of course make it appear at first
2404 glance as if rcu_read_unlock_bh() was executi 2404 glance as if rcu_read_unlock_bh() was executing very slowly.
2405 2405
2406 The `RCU-bh 2406 The `RCU-bh
2407 API <https://lwn.net/Articles/609973/#RCU%20P 2407 API <https://lwn.net/Articles/609973/#RCU%20Per-Flavor%20API%20Table>`__
2408 includes rcu_read_lock_bh(), rcu_read_unlock_ 2408 includes rcu_read_lock_bh(), rcu_read_unlock_bh(), rcu_dereference_bh(),
2409 rcu_dereference_bh_check(), and rcu_read_lock 2409 rcu_dereference_bh_check(), and rcu_read_lock_bh_held(). However, the
2410 old RCU-bh update-side APIs are now gone, rep 2410 old RCU-bh update-side APIs are now gone, replaced by synchronize_rcu(),
2411 synchronize_rcu_expedited(), call_rcu(), and 2411 synchronize_rcu_expedited(), call_rcu(), and rcu_barrier(). In addition,
2412 anything that disables bottom halves also mar 2412 anything that disables bottom halves also marks an RCU-bh read-side
2413 critical section, including local_bh_disable( 2413 critical section, including local_bh_disable() and local_bh_enable(),
2414 local_irq_save() and local_irq_restore(), and 2414 local_irq_save() and local_irq_restore(), and so on.
2415 2415
2416 Sched Flavor (Historical) 2416 Sched Flavor (Historical)
2417 ~~~~~~~~~~~~~~~~~~~~~~~~~ 2417 ~~~~~~~~~~~~~~~~~~~~~~~~~
2418 2418
2419 The RCU-sched flavor of RCU has since been ex 2419 The RCU-sched flavor of RCU has since been expressed in terms of the
2420 other RCU flavors as part of a consolidation 2420 other RCU flavors as part of a consolidation of the three flavors into a
2421 single flavor. The read-side API remains, and 2421 single flavor. The read-side API remains, and continues to disable
2422 preemption and to be accounted for by lockdep 2422 preemption and to be accounted for by lockdep. Much of the material in
2423 this section is therefore strictly historical 2423 this section is therefore strictly historical in nature.
2424 2424
2425 Before preemptible RCU, waiting for an RCU gr 2425 Before preemptible RCU, waiting for an RCU grace period had the side
2426 effect of also waiting for all pre-existing i 2426 effect of also waiting for all pre-existing interrupt and NMI handlers.
2427 However, there are legitimate preemptible-RCU 2427 However, there are legitimate preemptible-RCU implementations that do
2428 not have this property, given that any point 2428 not have this property, given that any point in the code outside of an
2429 RCU read-side critical section can be a quies 2429 RCU read-side critical section can be a quiescent state. Therefore,
2430 *RCU-sched* was created, which follows “cla 2430 *RCU-sched* was created, which follows “classic” RCU in that an
2431 RCU-sched grace period waits for pre-existing 2431 RCU-sched grace period waits for pre-existing interrupt and NMI
2432 handlers. In kernels built with ``CONFIG_PREE 2432 handlers. In kernels built with ``CONFIG_PREEMPTION=n``, the RCU and
2433 RCU-sched APIs have identical implementations 2433 RCU-sched APIs have identical implementations, while kernels built with
2434 ``CONFIG_PREEMPTION=y`` provide a separate im 2434 ``CONFIG_PREEMPTION=y`` provide a separate implementation for each.
2435 2435
2436 Note well that in ``CONFIG_PREEMPTION=y`` ker 2436 Note well that in ``CONFIG_PREEMPTION=y`` kernels,
2437 rcu_read_lock_sched() and rcu_read_unlock_sch 2437 rcu_read_lock_sched() and rcu_read_unlock_sched() disable and
2438 re-enable preemption, respectively. This mean 2438 re-enable preemption, respectively. This means that if there was a
2439 preemption attempt during the RCU-sched read- 2439 preemption attempt during the RCU-sched read-side critical section,
2440 rcu_read_unlock_sched() will enter the schedu 2440 rcu_read_unlock_sched() will enter the scheduler, with all the
2441 latency and overhead entailed. Just as with r 2441 latency and overhead entailed. Just as with rcu_read_unlock_bh(),
2442 this can make it look as if rcu_read_unlock_s 2442 this can make it look as if rcu_read_unlock_sched() was executing
2443 very slowly. However, the highest-priority ta 2443 very slowly. However, the highest-priority task won't be preempted, so
2444 that task will enjoy low-overhead rcu_read_un 2444 that task will enjoy low-overhead rcu_read_unlock_sched()
2445 invocations. 2445 invocations.
2446 2446
2447 The `RCU-sched 2447 The `RCU-sched
2448 API <https://lwn.net/Articles/609973/#RCU%20P 2448 API <https://lwn.net/Articles/609973/#RCU%20Per-Flavor%20API%20Table>`__
2449 includes rcu_read_lock_sched(), rcu_read_unlo 2449 includes rcu_read_lock_sched(), rcu_read_unlock_sched(),
2450 rcu_read_lock_sched_notrace(), rcu_read_unloc 2450 rcu_read_lock_sched_notrace(), rcu_read_unlock_sched_notrace(),
2451 rcu_dereference_sched(), rcu_dereference_sche 2451 rcu_dereference_sched(), rcu_dereference_sched_check(), and
2452 rcu_read_lock_sched_held(). However, the old 2452 rcu_read_lock_sched_held(). However, the old RCU-sched update-side APIs
2453 are now gone, replaced by synchronize_rcu(), 2453 are now gone, replaced by synchronize_rcu(), synchronize_rcu_expedited(),
2454 call_rcu(), and rcu_barrier(). In addition, 2454 call_rcu(), and rcu_barrier(). In addition, anything that disables
2455 preemption also marks an RCU-sched read-side 2455 preemption also marks an RCU-sched read-side critical section,
2456 including preempt_disable() and preempt_enabl 2456 including preempt_disable() and preempt_enable(), local_irq_save()
2457 and local_irq_restore(), and so on. 2457 and local_irq_restore(), and so on.
2458 2458
2459 Sleepable RCU 2459 Sleepable RCU
2460 ~~~~~~~~~~~~~ 2460 ~~~~~~~~~~~~~
2461 2461
2462 For well over a decade, someone saying “I n 2462 For well over a decade, someone saying “I need to block within an RCU
2463 read-side critical section” was a reliable 2463 read-side critical section” was a reliable indication that this someone
2464 did not understand RCU. After all, if you are 2464 did not understand RCU. After all, if you are always blocking in an RCU
2465 read-side critical section, you can probably 2465 read-side critical section, you can probably afford to use a
2466 higher-overhead synchronization mechanism. Ho 2466 higher-overhead synchronization mechanism. However, that changed with
2467 the advent of the Linux kernel's notifiers, w 2467 the advent of the Linux kernel's notifiers, whose RCU read-side critical
2468 sections almost never sleep, but sometimes ne 2468 sections almost never sleep, but sometimes need to. This resulted in the
2469 introduction of `sleepable RCU <https://lwn.n 2469 introduction of `sleepable RCU <https://lwn.net/Articles/202847/>`__, or
2470 *SRCU*. 2470 *SRCU*.
2471 2471
2472 SRCU allows different domains to be defined, 2472 SRCU allows different domains to be defined, with each such domain
2473 defined by an instance of an ``srcu_struct`` 2473 defined by an instance of an ``srcu_struct`` structure. A pointer to
2474 this structure must be passed in to each SRCU 2474 this structure must be passed in to each SRCU function, for example,
2475 ``synchronize_srcu(&ss)``, where ``ss`` is th 2475 ``synchronize_srcu(&ss)``, where ``ss`` is the ``srcu_struct``
2476 structure. The key benefit of these domains i 2476 structure. The key benefit of these domains is that a slow SRCU reader
2477 in one domain does not delay an SRCU grace pe 2477 in one domain does not delay an SRCU grace period in some other domain.
2478 That said, one consequence of these domains i 2478 That said, one consequence of these domains is that read-side code must
2479 pass a “cookie” from srcu_read_lock() to 2479 pass a “cookie” from srcu_read_lock() to srcu_read_unlock(), for
2480 example, as follows: 2480 example, as follows:
2481 2481
2482 :: 2482 ::
2483 2483
2484 1 int idx; 2484 1 int idx;
2485 2 2485 2
2486 3 idx = srcu_read_lock(&ss); 2486 3 idx = srcu_read_lock(&ss);
2487 4 do_something(); 2487 4 do_something();
2488 5 srcu_read_unlock(&ss, idx); 2488 5 srcu_read_unlock(&ss, idx);
2489 2489
2490 As noted above, it is legal to block within S 2490 As noted above, it is legal to block within SRCU read-side critical
2491 sections, however, with great power comes gre 2491 sections, however, with great power comes great responsibility. If you
2492 block forever in one of a given domain's SRCU 2492 block forever in one of a given domain's SRCU read-side critical
2493 sections, then that domain's grace periods wi 2493 sections, then that domain's grace periods will also be blocked forever.
2494 Of course, one good way to block forever is t 2494 Of course, one good way to block forever is to deadlock, which can
2495 happen if any operation in a given domain's S 2495 happen if any operation in a given domain's SRCU read-side critical
2496 section can wait, either directly or indirect 2496 section can wait, either directly or indirectly, for that domain's grace
2497 period to elapse. For example, this results i 2497 period to elapse. For example, this results in a self-deadlock:
2498 2498
2499 :: 2499 ::
2500 2500
2501 1 int idx; 2501 1 int idx;
2502 2 2502 2
2503 3 idx = srcu_read_lock(&ss); 2503 3 idx = srcu_read_lock(&ss);
2504 4 do_something(); 2504 4 do_something();
2505 5 synchronize_srcu(&ss); 2505 5 synchronize_srcu(&ss);
2506 6 srcu_read_unlock(&ss, idx); 2506 6 srcu_read_unlock(&ss, idx);
2507 2507
2508 However, if line 5 acquired a mutex that was 2508 However, if line 5 acquired a mutex that was held across a
2509 synchronize_srcu() for domain ``ss``, deadloc 2509 synchronize_srcu() for domain ``ss``, deadlock would still be
2510 possible. Furthermore, if line 5 acquired a 2510 possible. Furthermore, if line 5 acquired a mutex that was held across a
2511 synchronize_srcu() for some other domain ``ss 2511 synchronize_srcu() for some other domain ``ss1``, and if an
2512 ``ss1``-domain SRCU read-side critical sectio 2512 ``ss1``-domain SRCU read-side critical section acquired another mutex
2513 that was held across as ``ss``-domain synchro 2513 that was held across as ``ss``-domain synchronize_srcu(), deadlock
2514 would again be possible. Such a deadlock cycl 2514 would again be possible. Such a deadlock cycle could extend across an
2515 arbitrarily large number of different SRCU do 2515 arbitrarily large number of different SRCU domains. Again, with great
2516 power comes great responsibility. 2516 power comes great responsibility.
2517 2517
2518 Unlike the other RCU flavors, SRCU read-side 2518 Unlike the other RCU flavors, SRCU read-side critical sections can run
2519 on idle and even offline CPUs. This ability r 2519 on idle and even offline CPUs. This ability requires that
2520 srcu_read_lock() and srcu_read_unlock() conta 2520 srcu_read_lock() and srcu_read_unlock() contain memory barriers,
2521 which means that SRCU readers will run a bit 2521 which means that SRCU readers will run a bit slower than would RCU
2522 readers. It also motivates the smp_mb__after_ 2522 readers. It also motivates the smp_mb__after_srcu_read_unlock() API,
2523 which, in combination with srcu_read_unlock() 2523 which, in combination with srcu_read_unlock(), guarantees a full
2524 memory barrier. 2524 memory barrier.
2525 2525
2526 Also unlike other RCU flavors, synchronize_sr 2526 Also unlike other RCU flavors, synchronize_srcu() may **not** be
2527 invoked from CPU-hotplug notifiers, due to th 2527 invoked from CPU-hotplug notifiers, due to the fact that SRCU grace
2528 periods make use of timers and the possibilit 2528 periods make use of timers and the possibility of timers being
2529 temporarily “stranded” on the outgoing CP 2529 temporarily “stranded” on the outgoing CPU. This stranding of timers
2530 means that timers posted to the outgoing CPU 2530 means that timers posted to the outgoing CPU will not fire until late in
2531 the CPU-hotplug process. The problem is that 2531 the CPU-hotplug process. The problem is that if a notifier is waiting on
2532 an SRCU grace period, that grace period is wa 2532 an SRCU grace period, that grace period is waiting on a timer, and that
2533 timer is stranded on the outgoing CPU, then t 2533 timer is stranded on the outgoing CPU, then the notifier will never be
2534 awakened, in other words, deadlock has occurr 2534 awakened, in other words, deadlock has occurred. This same situation of
2535 course also prohibits srcu_barrier() from bei 2535 course also prohibits srcu_barrier() from being invoked from
2536 CPU-hotplug notifiers. 2536 CPU-hotplug notifiers.
2537 2537
2538 SRCU also differs from other RCU flavors in t 2538 SRCU also differs from other RCU flavors in that SRCU's expedited and
2539 non-expedited grace periods are implemented b 2539 non-expedited grace periods are implemented by the same mechanism. This
2540 means that in the current SRCU implementation 2540 means that in the current SRCU implementation, expediting a future grace
2541 period has the side effect of expediting all 2541 period has the side effect of expediting all prior grace periods that
2542 have not yet completed. (But please note that 2542 have not yet completed. (But please note that this is a property of the
2543 current implementation, not necessarily of fu 2543 current implementation, not necessarily of future implementations.) In
2544 addition, if SRCU has been idle for longer th 2544 addition, if SRCU has been idle for longer than the interval specified
2545 by the ``srcutree.exp_holdoff`` kernel boot p 2545 by the ``srcutree.exp_holdoff`` kernel boot parameter (25 microseconds
2546 by default), and if a synchronize_srcu() invo 2546 by default), and if a synchronize_srcu() invocation ends this idle
2547 period, that invocation will be automatically 2547 period, that invocation will be automatically expedited.
2548 2548
2549 As of v4.12, SRCU's callbacks are maintained 2549 As of v4.12, SRCU's callbacks are maintained per-CPU, eliminating a
2550 locking bottleneck present in prior kernel ve 2550 locking bottleneck present in prior kernel versions. Although this will
2551 allow users to put much heavier stress on cal 2551 allow users to put much heavier stress on call_srcu(), it is
2552 important to note that SRCU does not yet take 2552 important to note that SRCU does not yet take any special steps to deal
2553 with callback flooding. So if you are posting 2553 with callback flooding. So if you are posting (say) 10,000 SRCU
2554 callbacks per second per CPU, you are probabl 2554 callbacks per second per CPU, you are probably totally OK, but if you
2555 intend to post (say) 1,000,000 SRCU callbacks 2555 intend to post (say) 1,000,000 SRCU callbacks per second per CPU, please
2556 run some tests first. SRCU just might need a 2556 run some tests first. SRCU just might need a few adjustment to deal with
2557 that sort of load. Of course, your mileage ma 2557 that sort of load. Of course, your mileage may vary based on the speed
2558 of your CPUs and the size of your memory. 2558 of your CPUs and the size of your memory.
2559 2559
2560 The `SRCU 2560 The `SRCU
2561 API <https://lwn.net/Articles/609973/#RCU%20P 2561 API <https://lwn.net/Articles/609973/#RCU%20Per-Flavor%20API%20Table>`__
2562 includes srcu_read_lock(), srcu_read_unlock() 2562 includes srcu_read_lock(), srcu_read_unlock(),
2563 srcu_dereference(), srcu_dereference_check(), 2563 srcu_dereference(), srcu_dereference_check(),
2564 synchronize_srcu(), synchronize_srcu_expedite 2564 synchronize_srcu(), synchronize_srcu_expedited(),
2565 call_srcu(), srcu_barrier(), and srcu_read_lo 2565 call_srcu(), srcu_barrier(), and srcu_read_lock_held(). It
2566 also includes DEFINE_SRCU(), DEFINE_STATIC_SR 2566 also includes DEFINE_SRCU(), DEFINE_STATIC_SRCU(), and
2567 init_srcu_struct() APIs for defining and init 2567 init_srcu_struct() APIs for defining and initializing
2568 ``srcu_struct`` structures. 2568 ``srcu_struct`` structures.
2569 2569
2570 More recently, the SRCU API has added polling 2570 More recently, the SRCU API has added polling interfaces:
2571 2571
2572 #. start_poll_synchronize_srcu() returns a co 2572 #. start_poll_synchronize_srcu() returns a cookie identifying
2573 the completion of a future SRCU grace peri 2573 the completion of a future SRCU grace period and ensures
2574 that this grace period will be started. 2574 that this grace period will be started.
2575 #. poll_state_synchronize_srcu() returns ``tr 2575 #. poll_state_synchronize_srcu() returns ``true`` iff the
2576 specified cookie corresponds to an already 2576 specified cookie corresponds to an already-completed
2577 SRCU grace period. 2577 SRCU grace period.
2578 #. get_state_synchronize_srcu() returns a coo 2578 #. get_state_synchronize_srcu() returns a cookie just like
2579 start_poll_synchronize_srcu() does, but di 2579 start_poll_synchronize_srcu() does, but differs in that
2580 it does nothing to ensure that any future 2580 it does nothing to ensure that any future SRCU grace period
2581 will be started. 2581 will be started.
2582 2582
2583 These functions are used to avoid unnecessary 2583 These functions are used to avoid unnecessary SRCU grace periods in
2584 certain types of buffer-cache algorithms havi 2584 certain types of buffer-cache algorithms having multi-stage age-out
2585 mechanisms. The idea is that by the time the 2585 mechanisms. The idea is that by the time the block has aged completely
2586 from the cache, an SRCU grace period will be 2586 from the cache, an SRCU grace period will be very likely to have elapsed.
2587 2587
2588 Tasks RCU 2588 Tasks RCU
2589 ~~~~~~~~~ 2589 ~~~~~~~~~
2590 2590
2591 Some forms of tracing use “trampolines” t 2591 Some forms of tracing use “trampolines” to handle the binary rewriting
2592 required to install different types of probes 2592 required to install different types of probes. It would be good to be
2593 able to free old trampolines, which sounds li 2593 able to free old trampolines, which sounds like a job for some form of
2594 RCU. However, because it is necessary to be a 2594 RCU. However, because it is necessary to be able to install a trace
2595 anywhere in the code, it is not possible to u 2595 anywhere in the code, it is not possible to use read-side markers such
2596 as rcu_read_lock() and rcu_read_unlock(). In 2596 as rcu_read_lock() and rcu_read_unlock(). In addition, it does
2597 not work to have these markers in the trampol 2597 not work to have these markers in the trampoline itself, because there
2598 would need to be instructions following rcu_r 2598 would need to be instructions following rcu_read_unlock(). Although
2599 synchronize_rcu() would guarantee that execut 2599 synchronize_rcu() would guarantee that execution reached the
2600 rcu_read_unlock(), it would not be able to gu 2600 rcu_read_unlock(), it would not be able to guarantee that execution
2601 had completely left the trampoline. Worse yet 2601 had completely left the trampoline. Worse yet, in some situations
2602 the trampoline's protection must extend a few 2602 the trampoline's protection must extend a few instructions *prior* to
2603 execution reaching the trampoline. For examp 2603 execution reaching the trampoline. For example, these few instructions
2604 might calculate the address of the trampoline 2604 might calculate the address of the trampoline, so that entering the
2605 trampoline would be pre-ordained a surprising 2605 trampoline would be pre-ordained a surprisingly long time before execution
2606 actually reached the trampoline itself. 2606 actually reached the trampoline itself.
2607 2607
2608 The solution, in the form of `Tasks 2608 The solution, in the form of `Tasks
2609 RCU <https://lwn.net/Articles/607117/>`__, is 2609 RCU <https://lwn.net/Articles/607117/>`__, is to have implicit read-side
2610 critical sections that are delimited by volun 2610 critical sections that are delimited by voluntary context switches, that
2611 is, calls to schedule(), cond_resched(), and 2611 is, calls to schedule(), cond_resched(), and
2612 synchronize_rcu_tasks(). In addition, transit 2612 synchronize_rcu_tasks(). In addition, transitions to and from
2613 userspace execution also delimit tasks-RCU re 2613 userspace execution also delimit tasks-RCU read-side critical sections.
2614 Idle tasks are ignored by Tasks RCU, and Task 2614 Idle tasks are ignored by Tasks RCU, and Tasks Rude RCU may be used to
2615 interact with them. 2615 interact with them.
2616 2616
2617 Note well that involuntary context switches a 2617 Note well that involuntary context switches are *not* Tasks-RCU quiescent
2618 states. After all, in preemptible kernels, a 2618 states. After all, in preemptible kernels, a task executing code in a
2619 trampoline might be preempted. In this case, 2619 trampoline might be preempted. In this case, the Tasks-RCU grace period
2620 clearly cannot end until that task resumes an 2620 clearly cannot end until that task resumes and its execution leaves that
2621 trampoline. This means, among other things, 2621 trampoline. This means, among other things, that cond_resched() does
2622 not provide a Tasks RCU quiescent state. (In 2622 not provide a Tasks RCU quiescent state. (Instead, use rcu_softirq_qs()
2623 from softirq or rcu_tasks_classic_qs() otherw 2623 from softirq or rcu_tasks_classic_qs() otherwise.)
2624 2624
2625 The tasks-RCU API is quite compact, consistin 2625 The tasks-RCU API is quite compact, consisting only of
2626 call_rcu_tasks(), synchronize_rcu_tasks(), an 2626 call_rcu_tasks(), synchronize_rcu_tasks(), and
2627 rcu_barrier_tasks(). In ``CONFIG_PREEMPTION=n 2627 rcu_barrier_tasks(). In ``CONFIG_PREEMPTION=n`` kernels, trampolines
2628 cannot be preempted, so these APIs map to cal 2628 cannot be preempted, so these APIs map to call_rcu(),
2629 synchronize_rcu(), and rcu_barrier(), respect 2629 synchronize_rcu(), and rcu_barrier(), respectively. In
2630 ``CONFIG_PREEMPTION=y`` kernels, trampolines 2630 ``CONFIG_PREEMPTION=y`` kernels, trampolines can be preempted, and these
2631 three APIs are therefore implemented by separ 2631 three APIs are therefore implemented by separate functions that check
2632 for voluntary context switches. 2632 for voluntary context switches.
2633 2633
2634 Tasks Rude RCU 2634 Tasks Rude RCU
2635 ~~~~~~~~~~~~~~ 2635 ~~~~~~~~~~~~~~
2636 2636
2637 Some forms of tracing need to wait for all pr 2637 Some forms of tracing need to wait for all preemption-disabled regions
2638 of code running on any online CPU, including 2638 of code running on any online CPU, including those executed when RCU is
2639 not watching. This means that synchronize_rc 2639 not watching. This means that synchronize_rcu() is insufficient, and
2640 Tasks Rude RCU must be used instead. This fl 2640 Tasks Rude RCU must be used instead. This flavor of RCU does its work by
2641 forcing a workqueue to be scheduled on each o 2641 forcing a workqueue to be scheduled on each online CPU, hence the "Rude"
2642 moniker. And this operation is considered to 2642 moniker. And this operation is considered to be quite rude by real-time
2643 workloads that don't want their ``nohz_full`` 2643 workloads that don't want their ``nohz_full`` CPUs receiving IPIs and
2644 by battery-powered systems that don't want th 2644 by battery-powered systems that don't want their idle CPUs to be awakened.
2645 2645
2646 Once kernel entry/exit and deep-idle function 2646 Once kernel entry/exit and deep-idle functions have been properly tagged
2647 ``noinstr``, Tasks RCU can start paying atten 2647 ``noinstr``, Tasks RCU can start paying attention to idle tasks (except
2648 those that are idle from RCU's perspective) a 2648 those that are idle from RCU's perspective) and then Tasks Rude RCU can
2649 be removed from the kernel. 2649 be removed from the kernel.
2650 2650
2651 The tasks-rude-RCU API is also reader-marking 2651 The tasks-rude-RCU API is also reader-marking-free and thus quite compact,
2652 consisting of call_rcu_tasks_rude(), synchron 2652 consisting of call_rcu_tasks_rude(), synchronize_rcu_tasks_rude(),
2653 and rcu_barrier_tasks_rude(). 2653 and rcu_barrier_tasks_rude().
2654 2654
2655 Tasks Trace RCU 2655 Tasks Trace RCU
2656 ~~~~~~~~~~~~~~~ 2656 ~~~~~~~~~~~~~~~
2657 2657
2658 Some forms of tracing need to sleep in reader 2658 Some forms of tracing need to sleep in readers, but cannot tolerate
2659 SRCU's read-side overhead, which includes a f 2659 SRCU's read-side overhead, which includes a full memory barrier in both
2660 srcu_read_lock() and srcu_read_unlock(). Thi 2660 srcu_read_lock() and srcu_read_unlock(). This need is handled by a
2661 Tasks Trace RCU that uses scheduler locking a 2661 Tasks Trace RCU that uses scheduler locking and IPIs to synchronize with
2662 readers. Real-time systems that cannot toler 2662 readers. Real-time systems that cannot tolerate IPIs may build their
2663 kernels with ``CONFIG_TASKS_TRACE_RCU_READ_MB 2663 kernels with ``CONFIG_TASKS_TRACE_RCU_READ_MB=y``, which avoids the IPIs at
2664 the expense of adding full memory barriers to 2664 the expense of adding full memory barriers to the read-side primitives.
2665 2665
2666 The tasks-trace-RCU API is also reasonably co 2666 The tasks-trace-RCU API is also reasonably compact,
2667 consisting of rcu_read_lock_trace(), rcu_read 2667 consisting of rcu_read_lock_trace(), rcu_read_unlock_trace(),
2668 rcu_read_lock_trace_held(), call_rcu_tasks_tr 2668 rcu_read_lock_trace_held(), call_rcu_tasks_trace(),
2669 synchronize_rcu_tasks_trace(), and rcu_barrie 2669 synchronize_rcu_tasks_trace(), and rcu_barrier_tasks_trace().
2670 2670
2671 Possible Future Changes 2671 Possible Future Changes
2672 ----------------------- 2672 -----------------------
2673 2673
2674 One of the tricks that RCU uses to attain upd 2674 One of the tricks that RCU uses to attain update-side scalability is to
2675 increase grace-period latency with increasing 2675 increase grace-period latency with increasing numbers of CPUs. If this
2676 becomes a serious problem, it will be necessa 2676 becomes a serious problem, it will be necessary to rework the
2677 grace-period state machine so as to avoid the 2677 grace-period state machine so as to avoid the need for the additional
2678 latency. 2678 latency.
2679 2679
2680 RCU disables CPU hotplug in a few places, per 2680 RCU disables CPU hotplug in a few places, perhaps most notably in the
2681 rcu_barrier() operations. If there is a stron 2681 rcu_barrier() operations. If there is a strong reason to use
2682 rcu_barrier() in CPU-hotplug notifiers, it wi 2682 rcu_barrier() in CPU-hotplug notifiers, it will be necessary to
2683 avoid disabling CPU hotplug. This would intro 2683 avoid disabling CPU hotplug. This would introduce some complexity, so
2684 there had better be a *very* good reason. 2684 there had better be a *very* good reason.
2685 2685
2686 The tradeoff between grace-period latency on 2686 The tradeoff between grace-period latency on the one hand and
2687 interruptions of other CPUs on the other hand 2687 interruptions of other CPUs on the other hand may need to be
2688 re-examined. The desire is of course for zero 2688 re-examined. The desire is of course for zero grace-period latency as
2689 well as zero interprocessor interrupts undert 2689 well as zero interprocessor interrupts undertaken during an expedited
2690 grace period operation. While this ideal is u 2690 grace period operation. While this ideal is unlikely to be achievable,
2691 it is quite possible that further improvement 2691 it is quite possible that further improvements can be made.
2692 2692
2693 The multiprocessor implementations of RCU use 2693 The multiprocessor implementations of RCU use a combining tree that
2694 groups CPUs so as to reduce lock contention a 2694 groups CPUs so as to reduce lock contention and increase cache locality.
2695 However, this combining tree does not spread 2695 However, this combining tree does not spread its memory across NUMA
2696 nodes nor does it align the CPU groups with h 2696 nodes nor does it align the CPU groups with hardware features such as
2697 sockets or cores. Such spreading and alignmen 2697 sockets or cores. Such spreading and alignment is currently believed to
2698 be unnecessary because the hotpath read-side 2698 be unnecessary because the hotpath read-side primitives do not access
2699 the combining tree, nor does call_rcu() in th 2699 the combining tree, nor does call_rcu() in the common case. If you
2700 believe that your architecture needs such spr 2700 believe that your architecture needs such spreading and alignment, then
2701 your architecture should also benefit from th 2701 your architecture should also benefit from the
2702 ``rcutree.rcu_fanout_leaf`` boot parameter, w 2702 ``rcutree.rcu_fanout_leaf`` boot parameter, which can be set to the
2703 number of CPUs in a socket, NUMA node, or wha 2703 number of CPUs in a socket, NUMA node, or whatever. If the number of
2704 CPUs is too large, use a fraction of the numb 2704 CPUs is too large, use a fraction of the number of CPUs. If the number
2705 of CPUs is a large prime number, well, that c 2705 of CPUs is a large prime number, well, that certainly is an
2706 “interesting” architectural choice! More 2706 “interesting” architectural choice! More flexible arrangements might be
2707 considered, but only if ``rcutree.rcu_fanout_ 2707 considered, but only if ``rcutree.rcu_fanout_leaf`` has proven
2708 inadequate, and only if the inadequacy has be 2708 inadequate, and only if the inadequacy has been demonstrated by a
2709 carefully run and realistic system-level work 2709 carefully run and realistic system-level workload.
2710 2710
2711 Please note that arrangements that require RC 2711 Please note that arrangements that require RCU to remap CPU numbers will
2712 require extremely good demonstration of need 2712 require extremely good demonstration of need and full exploration of
2713 alternatives. 2713 alternatives.
2714 2714
2715 RCU's various kthreads are reasonably recent 2715 RCU's various kthreads are reasonably recent additions. It is quite
2716 likely that adjustments will be required to m 2716 likely that adjustments will be required to more gracefully handle
2717 extreme loads. It might also be necessary to 2717 extreme loads. It might also be necessary to be able to relate CPU
2718 utilization by RCU's kthreads and softirq han 2718 utilization by RCU's kthreads and softirq handlers to the code that
2719 instigated this CPU utilization. For example, 2719 instigated this CPU utilization. For example, RCU callback overhead
2720 might be charged back to the originating call 2720 might be charged back to the originating call_rcu() instance, though
2721 probably not in production kernels. 2721 probably not in production kernels.
2722 2722
2723 Additional work may be required to provide re 2723 Additional work may be required to provide reasonable forward-progress
2724 guarantees under heavy load for grace periods 2724 guarantees under heavy load for grace periods and for callback
2725 invocation. 2725 invocation.
2726 2726
2727 Summary 2727 Summary
2728 ------- 2728 -------
2729 2729
2730 This document has presented more than two dec 2730 This document has presented more than two decade's worth of RCU
2731 requirements. Given that the requirements kee 2731 requirements. Given that the requirements keep changing, this will not
2732 be the last word on this subject, but at leas 2732 be the last word on this subject, but at least it serves to get an
2733 important subset of the requirements set fort 2733 important subset of the requirements set forth.
2734 2734
2735 Acknowledgments 2735 Acknowledgments
2736 --------------- 2736 ---------------
2737 2737
2738 I am grateful to Steven Rostedt, Lai Jiangsha 2738 I am grateful to Steven Rostedt, Lai Jiangshan, Ingo Molnar, Oleg
2739 Nesterov, Borislav Petkov, Peter Zijlstra, Bo 2739 Nesterov, Borislav Petkov, Peter Zijlstra, Boqun Feng, and Andy
2740 Lutomirski for their help in rendering this a 2740 Lutomirski for their help in rendering this article human readable, and
2741 to Michelle Rankin for her support of this ef 2741 to Michelle Rankin for her support of this effort. Other contributions
2742 are acknowledged in the Linux kernel's git ar 2742 are acknowledged in the Linux kernel's git archive.
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.