1 ================================================= 2 A Tour Through TREE_RCU's Expedited Grace Periods 3 ================================================= 4 5 Introduction 6 ============ 7 8 This document describes RCU's expedited grace periods. 9 Unlike RCU's normal grace periods, which accept long latencies to attain 10 high efficiency and minimal disturbance, expedited grace periods accept 11 lower efficiency and significant disturbance to attain shorter latencies. 12 13 There are two flavors of RCU (RCU-preempt and RCU-sched), with an earlier 14 third RCU-bh flavor having been implemented in terms of the other two. 15 Each of the two implementations is covered in its own section. 16 17 Expedited Grace Period Design 18 ============================= 19 20 The expedited RCU grace periods cannot be accused of being subtle, 21 given that they for all intents and purposes hammer every CPU that 22 has not yet provided a quiescent state for the current expedited 23 grace period. 24 The one saving grace is that the hammer has grown a bit smaller 25 over time: The old call to ``try_stop_cpus()`` has been 26 replaced with a set of calls to ``smp_call_function_single()``, 27 each of which results in an IPI to the target CPU. 28 The corresponding handler function checks the CPU's state, motivating 29 a faster quiescent state where possible, and triggering a report 30 of that quiescent state. 31 As always for RCU, once everything has spent some time in a quiescent 32 state, the expedited grace period has completed. 33 34 The details of the ``smp_call_function_single()`` handler's 35 operation depend on the RCU flavor, as described in the following 36 sections. 37 38 RCU-preempt Expedited Grace Periods 39 =================================== 40 41 ``CONFIG_PREEMPTION=y`` kernels implement RCU-preempt. 42 The overall flow of the handling of a given CPU by an RCU-preempt 43 expedited grace period is shown in the following diagram: 44 45 .. kernel-figure:: ExpRCUFlow.svg 46 47 The solid arrows denote direct action, for example, a function call. 48 The dotted arrows denote indirect action, for example, an IPI 49 or a state that is reached after some time. 50 51 If a given CPU is offline or idle, ``synchronize_rcu_expedited()`` 52 will ignore it because idle and offline CPUs are already residing 53 in quiescent states. 54 Otherwise, the expedited grace period will use 55 ``smp_call_function_single()`` to send the CPU an IPI, which 56 is handled by ``rcu_exp_handler()``. 57 58 However, because this is preemptible RCU, ``rcu_exp_handler()`` 59 can check to see if the CPU is currently running in an RCU read-side 60 critical section. 61 If not, the handler can immediately report a quiescent state. 62 Otherwise, it sets flags so that the outermost ``rcu_read_unlock()`` 63 invocation will provide the needed quiescent-state report. 64 This flag-setting avoids the previous forced preemption of all 65 CPUs that might have RCU read-side critical sections. 66 In addition, this flag-setting is done so as to avoid increasing 67 the overhead of the common-case fastpath through the scheduler. 68 69 Again because this is preemptible RCU, an RCU read-side critical section 70 can be preempted. 71 When that happens, RCU will enqueue the task, which will the continue to 72 block the current expedited grace period until it resumes and finds its 73 outermost ``rcu_read_unlock()``. 74 The CPU will report a quiescent state just after enqueuing the task because 75 the CPU is no longer blocking the grace period. 76 It is instead the preempted task doing the blocking. 77 The list of blocked tasks is managed by ``rcu_preempt_ctxt_queue()``, 78 which is called from ``rcu_preempt_note_context_switch()``, which 79 in turn is called from ``rcu_note_context_switch()``, which in 80 turn is called from the scheduler. 81 82 83 +-----------------------------------------------------------------------+ 84 | **Quick Quiz**: | 85 +-----------------------------------------------------------------------+ 86 | Why not just have the expedited grace period check the state of all | 87 | the CPUs? After all, that would avoid all those real-time-unfriendly | 88 | IPIs. | 89 +-----------------------------------------------------------------------+ 90 | **Answer**: | 91 +-----------------------------------------------------------------------+ 92 | Because we want the RCU read-side critical sections to run fast, | 93 | which means no memory barriers. Therefore, it is not possible to | 94 | safely check the state from some other CPU. And even if it was | 95 | possible to safely check the state, it would still be necessary to | 96 | IPI the CPU to safely interact with the upcoming | 97 | ``rcu_read_unlock()`` invocation, which means that the remote state | 98 | testing would not help the worst-case latency that real-time | 99 | applications care about. | 100 | | 101 | One way to prevent your real-time application from getting hit with | 102 | these IPIs is to build your kernel with ``CONFIG_NO_HZ_FULL=y``. RCU | 103 | would then perceive the CPU running your application as being idle, | 104 | and it would be able to safely detect that state without needing to | 105 | IPI the CPU. | 106 +-----------------------------------------------------------------------+ 107 108 Please note that this is just the overall flow: Additional complications 109 can arise due to races with CPUs going idle or offline, among other 110 things. 111 112 RCU-sched Expedited Grace Periods 113 --------------------------------- 114 115 ``CONFIG_PREEMPTION=n`` kernels implement RCU-sched. The overall flow of 116 the handling of a given CPU by an RCU-sched expedited grace period is 117 shown in the following diagram: 118 119 .. kernel-figure:: ExpSchedFlow.svg 120 121 As with RCU-preempt, RCU-sched's ``synchronize_rcu_expedited()`` ignores 122 offline and idle CPUs, again because they are in remotely detectable 123 quiescent states. However, because the ``rcu_read_lock_sched()`` and 124 ``rcu_read_unlock_sched()`` leave no trace of their invocation, in 125 general it is not possible to tell whether or not the current CPU is in 126 an RCU read-side critical section. The best that RCU-sched's 127 ``rcu_exp_handler()`` can do is to check for idle, on the off-chance 128 that the CPU went idle while the IPI was in flight. If the CPU is idle, 129 then ``rcu_exp_handler()`` reports the quiescent state. 130 131 Otherwise, the handler forces a future context switch by setting the 132 NEED_RESCHED flag of the current task's thread flag and the CPU preempt 133 counter. At the time of the context switch, the CPU reports the 134 quiescent state. Should the CPU go offline first, it will report the 135 quiescent state at that time. 136 137 Expedited Grace Period and CPU Hotplug 138 -------------------------------------- 139 140 The expedited nature of expedited grace periods require a much tighter 141 interaction with CPU hotplug operations than is required for normal 142 grace periods. In addition, attempting to IPI offline CPUs will result 143 in splats, but failing to IPI online CPUs can result in too-short grace 144 periods. Neither option is acceptable in production kernels. 145 146 The interaction between expedited grace periods and CPU hotplug 147 operations is carried out at several levels: 148 149 #. The number of CPUs that have ever been online is tracked by the 150 ``rcu_state`` structure's ``->ncpus`` field. The ``rcu_state`` 151 structure's ``->ncpus_snap`` field tracks the number of CPUs that 152 have ever been online at the beginning of an RCU expedited grace 153 period. Note that this number never decreases, at least in the 154 absence of a time machine. 155 #. The identities of the CPUs that have ever been online is tracked by 156 the ``rcu_node`` structure's ``->expmaskinitnext`` field. The 157 ``rcu_node`` structure's ``->expmaskinit`` field tracks the 158 identities of the CPUs that were online at least once at the 159 beginning of the most recent RCU expedited grace period. The 160 ``rcu_state`` structure's ``->ncpus`` and ``->ncpus_snap`` fields are 161 used to detect when new CPUs have come online for the first time, 162 that is, when the ``rcu_node`` structure's ``->expmaskinitnext`` 163 field has changed since the beginning of the last RCU expedited grace 164 period, which triggers an update of each ``rcu_node`` structure's 165 ``->expmaskinit`` field from its ``->expmaskinitnext`` field. 166 #. Each ``rcu_node`` structure's ``->expmaskinit`` field is used to 167 initialize that structure's ``->expmask`` at the beginning of each 168 RCU expedited grace period. This means that only those CPUs that have 169 been online at least once will be considered for a given grace 170 period. 171 #. Any CPU that goes offline will clear its bit in its leaf ``rcu_node`` 172 structure's ``->qsmaskinitnext`` field, so any CPU with that bit 173 clear can safely be ignored. However, it is possible for a CPU coming 174 online or going offline to have this bit set for some time while 175 ``cpu_online`` returns ``false``. 176 #. For each non-idle CPU that RCU believes is currently online, the 177 grace period invokes ``smp_call_function_single()``. If this 178 succeeds, the CPU was fully online. Failure indicates that the CPU is 179 in the process of coming online or going offline, in which case it is 180 necessary to wait for a short time period and try again. The purpose 181 of this wait (or series of waits, as the case may be) is to permit a 182 concurrent CPU-hotplug operation to complete. 183 #. In the case of RCU-sched, one of the last acts of an outgoing CPU is 184 to invoke ``rcutree_report_cpu_dead()``, which reports a quiescent state for 185 that CPU. However, this is likely paranoia-induced redundancy. 186 187 +-----------------------------------------------------------------------+ 188 | **Quick Quiz**: | 189 +-----------------------------------------------------------------------+ 190 | Why all the dancing around with multiple counters and masks tracking | 191 | CPUs that were once online? Why not just have a single set of masks | 192 | tracking the currently online CPUs and be done with it? | 193 +-----------------------------------------------------------------------+ 194 | **Answer**: | 195 +-----------------------------------------------------------------------+ 196 | Maintaining single set of masks tracking the online CPUs *sounds* | 197 | easier, at least until you try working out all the race conditions | 198 | between grace-period initialization and CPU-hotplug operations. For | 199 | example, suppose initialization is progressing down the tree while a | 200 | CPU-offline operation is progressing up the tree. This situation can | 201 | result in bits set at the top of the tree that have no counterparts | 202 | at the bottom of the tree. Those bits will never be cleared, which | 203 | will result in grace-period hangs. In short, that way lies madness, | 204 | to say nothing of a great many bugs, hangs, and deadlocks. | 205 | In contrast, the current multi-mask multi-counter scheme ensures that | 206 | grace-period initialization will always see consistent masks up and | 207 | down the tree, which brings significant simplifications over the | 208 | single-mask method. | 209 | | 210 | This is an instance of `deferring work in order to avoid | 211 | synchronization <http://www.cs.columbia.edu/~library/TR-repository/re | 212 | ports/reports-1992/cucs-039-92.ps.gz>`__. | 213 | Lazily recording CPU-hotplug events at the beginning of the next | 214 | grace period greatly simplifies maintenance of the CPU-tracking | 215 | bitmasks in the ``rcu_node`` tree. | 216 +-----------------------------------------------------------------------+ 217 218 Expedited Grace Period Refinements 219 ---------------------------------- 220 221 Idle-CPU Checks 222 ~~~~~~~~~~~~~~~ 223 224 Each expedited grace period checks for idle CPUs when initially forming 225 the mask of CPUs to be IPIed and again just before IPIing a CPU (both 226 checks are carried out by ``sync_rcu_exp_select_cpus()``). If the CPU is 227 idle at any time between those two times, the CPU will not be IPIed. 228 Instead, the task pushing the grace period forward will include the idle 229 CPUs in the mask passed to ``rcu_report_exp_cpu_mult()``. 230 231 For RCU-sched, there is an additional check: If the IPI has interrupted 232 the idle loop, then ``rcu_exp_handler()`` invokes 233 ``rcu_report_exp_rdp()`` to report the corresponding quiescent state. 234 235 For RCU-preempt, there is no specific check for idle in the IPI handler 236 (``rcu_exp_handler()``), but because RCU read-side critical sections are 237 not permitted within the idle loop, if ``rcu_exp_handler()`` sees that 238 the CPU is within RCU read-side critical section, the CPU cannot 239 possibly be idle. Otherwise, ``rcu_exp_handler()`` invokes 240 ``rcu_report_exp_rdp()`` to report the corresponding quiescent state, 241 regardless of whether or not that quiescent state was due to the CPU 242 being idle. 243 244 In summary, RCU expedited grace periods check for idle when building the 245 bitmask of CPUs that must be IPIed, just before sending each IPI, and 246 (either explicitly or implicitly) within the IPI handler. 247 248 Batching via Sequence Counter 249 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 250 251 If each grace-period request was carried out separately, expedited grace 252 periods would have abysmal scalability and problematic high-load 253 characteristics. Because each grace-period operation can serve an 254 unlimited number of updates, it is important to *batch* requests, so 255 that a single expedited grace-period operation will cover all requests 256 in the corresponding batch. 257 258 This batching is controlled by a sequence counter named 259 ``->expedited_sequence`` in the ``rcu_state`` structure. This counter 260 has an odd value when there is an expedited grace period in progress and 261 an even value otherwise, so that dividing the counter value by two gives 262 the number of completed grace periods. During any given update request, 263 the counter must transition from even to odd and then back to even, thus 264 indicating that a grace period has elapsed. Therefore, if the initial 265 value of the counter is ``s``, the updater must wait until the counter 266 reaches at least the value ``(s+3)&~0x1``. This counter is managed by 267 the following access functions: 268 269 #. ``rcu_exp_gp_seq_start()``, which marks the start of an expedited 270 grace period. 271 #. ``rcu_exp_gp_seq_end()``, which marks the end of an expedited grace 272 period. 273 #. ``rcu_exp_gp_seq_snap()``, which obtains a snapshot of the counter. 274 #. ``rcu_exp_gp_seq_done()``, which returns ``true`` if a full expedited 275 grace period has elapsed since the corresponding call to 276 ``rcu_exp_gp_seq_snap()``. 277 278 Again, only one request in a given batch need actually carry out a 279 grace-period operation, which means there must be an efficient way to 280 identify which of many concurrent requests will initiate the grace 281 period, and that there be an efficient way for the remaining requests to 282 wait for that grace period to complete. However, that is the topic of 283 the next section. 284 285 Funnel Locking and Wait/Wakeup 286 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 287 288 The natural way to sort out which of a batch of updaters will initiate 289 the expedited grace period is to use the ``rcu_node`` combining tree, as 290 implemented by the ``exp_funnel_lock()`` function. The first updater 291 corresponding to a given grace period arriving at a given ``rcu_node`` 292 structure records its desired grace-period sequence number in the 293 ``->exp_seq_rq`` field and moves up to the next level in the tree. 294 Otherwise, if the ``->exp_seq_rq`` field already contains the sequence 295 number for the desired grace period or some later one, the updater 296 blocks on one of four wait queues in the ``->exp_wq[]`` array, using the 297 second-from-bottom and third-from bottom bits as an index. An 298 ``->exp_lock`` field in the ``rcu_node`` structure synchronizes access 299 to these fields. 300 301 An empty ``rcu_node`` tree is shown in the following diagram, with the 302 white cells representing the ``->exp_seq_rq`` field and the red cells 303 representing the elements of the ``->exp_wq[]`` array. 304 305 .. kernel-figure:: Funnel0.svg 306 307 The next diagram shows the situation after the arrival of Task A and 308 Task B at the leftmost and rightmost leaf ``rcu_node`` structures, 309 respectively. The current value of the ``rcu_state`` structure's 310 ``->expedited_sequence`` field is zero, so adding three and clearing the 311 bottom bit results in the value two, which both tasks record in the 312 ``->exp_seq_rq`` field of their respective ``rcu_node`` structures: 313 314 .. kernel-figure:: Funnel1.svg 315 316 Each of Tasks A and B will move up to the root ``rcu_node`` structure. 317 Suppose that Task A wins, recording its desired grace-period sequence 318 number and resulting in the state shown below: 319 320 .. kernel-figure:: Funnel2.svg 321 322 Task A now advances to initiate a new grace period, while Task B moves 323 up to the root ``rcu_node`` structure, and, seeing that its desired 324 sequence number is already recorded, blocks on ``->exp_wq[1]``. 325 326 +-----------------------------------------------------------------------+ 327 | **Quick Quiz**: | 328 +-----------------------------------------------------------------------+ 329 | Why ``->exp_wq[1]``? Given that the value of these tasks' desired | 330 | sequence number is two, so shouldn't they instead block on | 331 | ``->exp_wq[2]``? | 332 +-----------------------------------------------------------------------+ 333 | **Answer**: | 334 +-----------------------------------------------------------------------+ 335 | No. | 336 | Recall that the bottom bit of the desired sequence number indicates | 337 | whether or not a grace period is currently in progress. It is | 338 | therefore necessary to shift the sequence number right one bit | 339 | position to obtain the number of the grace period. This results in | 340 | ``->exp_wq[1]``. | 341 +-----------------------------------------------------------------------+ 342 343 If Tasks C and D also arrive at this point, they will compute the same 344 desired grace-period sequence number, and see that both leaf 345 ``rcu_node`` structures already have that value recorded. They will 346 therefore block on their respective ``rcu_node`` structures' 347 ``->exp_wq[1]`` fields, as shown below: 348 349 .. kernel-figure:: Funnel3.svg 350 351 Task A now acquires the ``rcu_state`` structure's ``->exp_mutex`` and 352 initiates the grace period, which increments ``->expedited_sequence``. 353 Therefore, if Tasks E and F arrive, they will compute a desired sequence 354 number of 4 and will record this value as shown below: 355 356 .. kernel-figure:: Funnel4.svg 357 358 Tasks E and F will propagate up the ``rcu_node`` combining tree, with 359 Task F blocking on the root ``rcu_node`` structure and Task E wait for 360 Task A to finish so that it can start the next grace period. The 361 resulting state is as shown below: 362 363 .. kernel-figure:: Funnel5.svg 364 365 Once the grace period completes, Task A starts waking up the tasks 366 waiting for this grace period to complete, increments the 367 ``->expedited_sequence``, acquires the ``->exp_wake_mutex`` and then 368 releases the ``->exp_mutex``. This results in the following state: 369 370 .. kernel-figure:: Funnel6.svg 371 372 Task E can then acquire ``->exp_mutex`` and increment 373 ``->expedited_sequence`` to the value three. If new tasks G and H arrive 374 and moves up the combining tree at the same time, the state will be as 375 follows: 376 377 .. kernel-figure:: Funnel7.svg 378 379 Note that three of the root ``rcu_node`` structure's waitqueues are now 380 occupied. However, at some point, Task A will wake up the tasks blocked 381 on the ``->exp_wq`` waitqueues, resulting in the following state: 382 383 .. kernel-figure:: Funnel8.svg 384 385 Execution will continue with Tasks E and H completing their grace 386 periods and carrying out their wakeups. 387 388 +-----------------------------------------------------------------------+ 389 | **Quick Quiz**: | 390 +-----------------------------------------------------------------------+ 391 | What happens if Task A takes so long to do its wakeups that Task E's | 392 | grace period completes? | 393 +-----------------------------------------------------------------------+ 394 | **Answer**: | 395 +-----------------------------------------------------------------------+ 396 | Then Task E will block on the ``->exp_wake_mutex``, which will also | 397 | prevent it from releasing ``->exp_mutex``, which in turn will prevent | 398 | the next grace period from starting. This last is important in | 399 | preventing overflow of the ``->exp_wq[]`` array. | 400 +-----------------------------------------------------------------------+ 401 402 Use of Workqueues 403 ~~~~~~~~~~~~~~~~~ 404 405 In earlier implementations, the task requesting the expedited grace 406 period also drove it to completion. This straightforward approach had 407 the disadvantage of needing to account for POSIX signals sent to user 408 tasks, so more recent implementations use the Linux kernel's 409 workqueues (see Documentation/core-api/workqueue.rst). 410 411 The requesting task still does counter snapshotting and funnel-lock 412 processing, but the task reaching the top of the funnel lock does a 413 ``schedule_work()`` (from ``_synchronize_rcu_expedited()`` so that a 414 workqueue kthread does the actual grace-period processing. Because 415 workqueue kthreads do not accept POSIX signals, grace-period-wait 416 processing need not allow for POSIX signals. In addition, this approach 417 allows wakeups for the previous expedited grace period to be overlapped 418 with processing for the next expedited grace period. Because there are 419 only four sets of waitqueues, it is necessary to ensure that the 420 previous grace period's wakeups complete before the next grace period's 421 wakeups start. This is handled by having the ``->exp_mutex`` guard 422 expedited grace-period processing and the ``->exp_wake_mutex`` guard 423 wakeups. The key point is that the ``->exp_mutex`` is not released until 424 the first wakeup is complete, which means that the ``->exp_wake_mutex`` 425 has already been acquired at that point. This approach ensures that the 426 previous grace period's wakeups can be carried out while the current 427 grace period is in process, but that these wakeups will complete before 428 the next grace period starts. This means that only three waitqueues are 429 required, guaranteeing that the four that are provided are sufficient. 430 431 Stall Warnings 432 ~~~~~~~~~~~~~~ 433 434 Expediting grace periods does nothing to speed things up when RCU 435 readers take too long, and therefore expedited grace periods check for 436 stalls just as normal grace periods do. 437 438 +-----------------------------------------------------------------------+ 439 | **Quick Quiz**: | 440 +-----------------------------------------------------------------------+ 441 | But why not just let the normal grace-period machinery detect the | 442 | stalls, given that a given reader must block both normal and | 443 | expedited grace periods? | 444 +-----------------------------------------------------------------------+ 445 | **Answer**: | 446 +-----------------------------------------------------------------------+ 447 | Because it is quite possible that at a given time there is no normal | 448 | grace period in progress, in which case the normal grace period | 449 | cannot emit a stall warning. | 450 +-----------------------------------------------------------------------+ 451 452 The ``synchronize_sched_expedited_wait()`` function loops waiting for 453 the expedited grace period to end, but with a timeout set to the current 454 RCU CPU stall-warning time. If this time is exceeded, any CPUs or 455 ``rcu_node`` structures blocking the current grace period are printed. 456 Each stall warning results in another pass through the loop, but the 457 second and subsequent passes use longer stall times. 458 459 Mid-boot operation 460 ~~~~~~~~~~~~~~~~~~ 461 462 The use of workqueues has the advantage that the expedited grace-period 463 code need not worry about POSIX signals. Unfortunately, it has the 464 corresponding disadvantage that workqueues cannot be used until they are 465 initialized, which does not happen until some time after the scheduler 466 spawns the first task. Given that there are parts of the kernel that 467 really do want to execute grace periods during this mid-boot “dead 468 zone”, expedited grace periods must do something else during this time. 469 470 What they do is to fall back to the old practice of requiring that the 471 requesting task drive the expedited grace period, as was the case before 472 the use of workqueues. However, the requesting task is only required to 473 drive the grace period during the mid-boot dead zone. Before mid-boot, a 474 synchronous grace period is a no-op. Some time after mid-boot, 475 workqueues are used. 476 477 Non-expedited non-SRCU synchronous grace periods must also operate 478 normally during mid-boot. This is handled by causing non-expedited grace 479 periods to take the expedited code path during mid-boot. 480 481 The current code assumes that there are no POSIX signals during the 482 mid-boot dead zone. However, if an overwhelming need for POSIX signals 483 somehow arises, appropriate adjustments can be made to the expedited 484 stall-warning code. One such adjustment would reinstate the 485 pre-workqueue stall-warning checks, but only during the mid-boot dead 486 zone. 487 488 With this refinement, synchronous grace periods can now be used from 489 task context pretty much any time during the life of the kernel. That 490 is, aside from some points in the suspend, hibernate, or shutdown code 491 path. 492 493 Summary 494 ~~~~~~~ 495 496 Expedited grace periods use a sequence-number approach to promote 497 batching, so that a single grace-period operation can serve numerous 498 requests. A funnel lock is used to efficiently identify the one task out 499 of a concurrent group that will request the grace period. All members of 500 the group will block on waitqueues provided in the ``rcu_node`` 501 structure. The actual grace-period processing is carried out by a 502 workqueue. 503 504 CPU-hotplug operations are noted lazily in order to prevent the need for 505 tight synchronization between expedited grace periods and CPU-hotplug 506 operations. The dyntick-idle counters are used to avoid sending IPIs to 507 idle CPUs, at least in the common case. RCU-preempt and RCU-sched use 508 different IPI handlers and different code to respond to the state 509 changes carried out by those handlers, but otherwise use common code. 510 511 Quiescent states are tracked using the ``rcu_node`` tree, and once all 512 necessary quiescent states have been reported, all tasks waiting on this 513 expedited grace period are awakened. A pair of mutexes are used to allow 514 one grace period's wakeups to proceed concurrently with the next grace 515 period's processing. 516 517 This combination of mechanisms allows expedited grace periods to run 518 reasonably efficiently. However, for non-time-critical tasks, normal 519 grace periods should be used instead because their longer duration 520 permits much higher degrees of batching, and thus much lower per-request 521 overheads.
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.