1 // SPDX-License-Identifier: GPL-2.0-or-later 1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* 2 /*
3 * fs/eventpoll.c (Efficient event retrieval 3 * fs/eventpoll.c (Efficient event retrieval implementation)
4 * Copyright (C) 2001,...,2009 Davide Libenz 4 * Copyright (C) 2001,...,2009 Davide Libenzi
5 * 5 *
6 * Davide Libenzi <davidel@xmailserver.org> 6 * Davide Libenzi <davidel@xmailserver.org>
7 */ 7 */
8 8
9 #include <linux/init.h> 9 #include <linux/init.h>
10 #include <linux/kernel.h> 10 #include <linux/kernel.h>
11 #include <linux/sched/signal.h> 11 #include <linux/sched/signal.h>
12 #include <linux/fs.h> 12 #include <linux/fs.h>
13 #include <linux/file.h> 13 #include <linux/file.h>
14 #include <linux/signal.h> 14 #include <linux/signal.h>
15 #include <linux/errno.h> 15 #include <linux/errno.h>
16 #include <linux/mm.h> 16 #include <linux/mm.h>
17 #include <linux/slab.h> 17 #include <linux/slab.h>
18 #include <linux/poll.h> 18 #include <linux/poll.h>
19 #include <linux/string.h> 19 #include <linux/string.h>
20 #include <linux/list.h> 20 #include <linux/list.h>
21 #include <linux/hash.h> 21 #include <linux/hash.h>
22 #include <linux/spinlock.h> 22 #include <linux/spinlock.h>
23 #include <linux/syscalls.h> 23 #include <linux/syscalls.h>
24 #include <linux/rbtree.h> 24 #include <linux/rbtree.h>
25 #include <linux/wait.h> 25 #include <linux/wait.h>
26 #include <linux/eventpoll.h> 26 #include <linux/eventpoll.h>
27 #include <linux/mount.h> 27 #include <linux/mount.h>
28 #include <linux/bitops.h> 28 #include <linux/bitops.h>
29 #include <linux/mutex.h> 29 #include <linux/mutex.h>
30 #include <linux/anon_inodes.h> 30 #include <linux/anon_inodes.h>
31 #include <linux/device.h> 31 #include <linux/device.h>
32 #include <linux/uaccess.h> 32 #include <linux/uaccess.h>
33 #include <asm/io.h> 33 #include <asm/io.h>
34 #include <asm/mman.h> 34 #include <asm/mman.h>
35 #include <linux/atomic.h> 35 #include <linux/atomic.h>
36 #include <linux/proc_fs.h> 36 #include <linux/proc_fs.h>
37 #include <linux/seq_file.h> 37 #include <linux/seq_file.h>
38 #include <linux/compat.h> 38 #include <linux/compat.h>
39 #include <linux/rculist.h> 39 #include <linux/rculist.h>
40 #include <linux/capability.h> <<
41 #include <net/busy_poll.h> 40 #include <net/busy_poll.h>
42 41
43 /* 42 /*
44 * LOCKING: 43 * LOCKING:
45 * There are three level of locking required b 44 * There are three level of locking required by epoll :
46 * 45 *
47 * 1) epnested_mutex (mutex) !! 46 * 1) epmutex (mutex)
48 * 2) ep->mtx (mutex) 47 * 2) ep->mtx (mutex)
49 * 3) ep->lock (rwlock) 48 * 3) ep->lock (rwlock)
50 * 49 *
51 * The acquire order is the one listed above, 50 * The acquire order is the one listed above, from 1 to 3.
52 * We need a rwlock (ep->lock) because we mani 51 * We need a rwlock (ep->lock) because we manipulate objects
53 * from inside the poll callback, that might b 52 * from inside the poll callback, that might be triggered from
54 * a wake_up() that in turn might be called fr 53 * a wake_up() that in turn might be called from IRQ context.
55 * So we can't sleep inside the poll callback 54 * So we can't sleep inside the poll callback and hence we need
56 * a spinlock. During the event transfer loop 55 * a spinlock. During the event transfer loop (from kernel to
57 * user space) we could end up sleeping due a 56 * user space) we could end up sleeping due a copy_to_user(), so
58 * we need a lock that will allow us to sleep. 57 * we need a lock that will allow us to sleep. This lock is a
59 * mutex (ep->mtx). It is acquired during the 58 * mutex (ep->mtx). It is acquired during the event transfer loop,
60 * during epoll_ctl(EPOLL_CTL_DEL) and during 59 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
61 * The epnested_mutex is acquired when inserti !! 60 * Then we also need a global mutex to serialize eventpoll_release_file()
62 * epoll fd. We do this so that we walk the ep !! 61 * and ep_free().
>> 62 * This mutex is acquired by ep_free() during the epoll file
>> 63 * cleanup path and it is also acquired by eventpoll_release_file()
>> 64 * if a file has been pushed inside an epoll set and it is then
>> 65 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
>> 66 * It is also acquired when inserting an epoll fd onto another epoll
>> 67 * fd. We do this so that we walk the epoll tree and ensure that this
63 * insertion does not create a cycle of epoll 68 * insertion does not create a cycle of epoll file descriptors, which
64 * could lead to deadlock. We need a global mu 69 * could lead to deadlock. We need a global mutex to prevent two
65 * simultaneous inserts (A into B and B into A 70 * simultaneous inserts (A into B and B into A) from racing and
66 * constructing a cycle without either insert 71 * constructing a cycle without either insert observing that it is
67 * going to. 72 * going to.
68 * It is necessary to acquire multiple "ep->mt 73 * It is necessary to acquire multiple "ep->mtx"es at once in the
69 * case when one epoll fd is added to another. 74 * case when one epoll fd is added to another. In this case, we
70 * always acquire the locks in the order of ne 75 * always acquire the locks in the order of nesting (i.e. after
71 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx w 76 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
72 * before e2->mtx). Since we disallow cycles o 77 * before e2->mtx). Since we disallow cycles of epoll file
73 * descriptors, this ensures that the mutexes 78 * descriptors, this ensures that the mutexes are well-ordered. In
74 * order to communicate this nesting to lockde 79 * order to communicate this nesting to lockdep, when walking a tree
75 * of epoll file descriptors, we use the curre 80 * of epoll file descriptors, we use the current recursion depth as
76 * the lockdep subkey. 81 * the lockdep subkey.
77 * It is possible to drop the "ep->mtx" and to 82 * It is possible to drop the "ep->mtx" and to use the global
78 * mutex "epnested_mutex" (together with "ep-> !! 83 * mutex "epmutex" (together with "ep->lock") to have it working,
79 * but having "ep->mtx" will make the interfac 84 * but having "ep->mtx" will make the interface more scalable.
80 * Events that require holding "epnested_mutex !! 85 * Events that require holding "epmutex" are very rare, while for
81 * normal operations the epoll private "ep->mt 86 * normal operations the epoll private "ep->mtx" will guarantee
82 * a better scalability. 87 * a better scalability.
83 */ 88 */
84 89
85 /* Epoll private bits inside the event mask */ 90 /* Epoll private bits inside the event mask */
86 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLON 91 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
87 92
88 #define EPOLLINOUT_BITS (EPOLLIN | EPOLLOUT) 93 #define EPOLLINOUT_BITS (EPOLLIN | EPOLLOUT)
89 94
90 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BIT 95 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | EPOLLERR | EPOLLHUP | \
91 EPOLLWAKEUP | 96 EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
92 97
93 /* Maximum number of nesting allowed inside ep 98 /* Maximum number of nesting allowed inside epoll sets */
94 #define EP_MAX_NESTS 4 99 #define EP_MAX_NESTS 4
95 100
96 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct 101 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
97 102
98 #define EP_UNACTIVE_PTR ((void *) -1L) 103 #define EP_UNACTIVE_PTR ((void *) -1L)
99 104
100 #define EP_ITEM_COST (sizeof(struct epitem) + 105 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
101 106
102 struct epoll_filefd { 107 struct epoll_filefd {
103 struct file *file; 108 struct file *file;
104 int fd; 109 int fd;
105 } __packed; 110 } __packed;
106 111
107 /* Wait structure used by the poll hooks */ !! 112 /*
108 struct eppoll_entry { !! 113 * Structure used to track possible nested calls, for too deep recursions
109 /* List header used to link this struc !! 114 * and loop cycles.
110 struct eppoll_entry *next; !! 115 */
111 !! 116 struct nested_call_node {
112 /* The "base" pointer is set to the co !! 117 struct list_head llink;
113 struct epitem *base; !! 118 void *cookie;
114 !! 119 void *ctx;
115 /* !! 120 };
116 * Wait queue item that will be linked <<
117 * queue head. <<
118 */ <<
119 wait_queue_entry_t wait; <<
120 121
121 /* The wait queue head that linked the !! 122 /*
122 wait_queue_head_t *whead; !! 123 * This structure is used as collector for nested calls, to check for
>> 124 * maximum recursion dept and loop cycles.
>> 125 */
>> 126 struct nested_calls {
>> 127 struct list_head tasks_call_list;
>> 128 spinlock_t lock;
123 }; 129 };
124 130
125 /* 131 /*
126 * Each file descriptor added to the eventpoll 132 * Each file descriptor added to the eventpoll interface will
127 * have an entry of this type linked to the "r 133 * have an entry of this type linked to the "rbr" RB tree.
128 * Avoid increasing the size of this struct, t 134 * Avoid increasing the size of this struct, there can be many thousands
129 * of these on a server and we do not want thi 135 * of these on a server and we do not want this to take another cache line.
130 */ 136 */
131 struct epitem { 137 struct epitem {
132 union { 138 union {
133 /* RB tree node links this str 139 /* RB tree node links this structure to the eventpoll RB tree */
134 struct rb_node rbn; 140 struct rb_node rbn;
135 /* Used to free the struct epi 141 /* Used to free the struct epitem */
136 struct rcu_head rcu; 142 struct rcu_head rcu;
137 }; 143 };
138 144
139 /* List header used to link this struc 145 /* List header used to link this structure to the eventpoll ready list */
140 struct list_head rdllink; 146 struct list_head rdllink;
141 147
142 /* 148 /*
143 * Works together "struct eventpoll"-> 149 * Works together "struct eventpoll"->ovflist in keeping the
144 * single linked chain of items. 150 * single linked chain of items.
145 */ 151 */
146 struct epitem *next; 152 struct epitem *next;
147 153
148 /* The file descriptor information thi 154 /* The file descriptor information this item refers to */
149 struct epoll_filefd ffd; 155 struct epoll_filefd ffd;
150 156
151 /* !! 157 /* Number of active wait queue attached to poll operations */
152 * Protected by file->f_lock, true for !! 158 int nwait;
153 * removed from the "struct file" item <<
154 * eventpoll->refcount orchestrates "s <<
155 */ <<
156 bool dying; <<
157 159
158 /* List containing poll wait queues */ 160 /* List containing poll wait queues */
159 struct eppoll_entry *pwqlist; !! 161 struct list_head pwqlist;
160 162
161 /* The "container" of this item */ 163 /* The "container" of this item */
162 struct eventpoll *ep; 164 struct eventpoll *ep;
163 165
164 /* List header used to link this item 166 /* List header used to link this item to the "struct file" items list */
165 struct hlist_node fllink; !! 167 struct list_head fllink;
166 168
167 /* wakeup_source used when EPOLLWAKEUP 169 /* wakeup_source used when EPOLLWAKEUP is set */
168 struct wakeup_source __rcu *ws; 170 struct wakeup_source __rcu *ws;
169 171
170 /* The structure that describe the int 172 /* The structure that describe the interested events and the source fd */
171 struct epoll_event event; 173 struct epoll_event event;
172 }; 174 };
173 175
174 /* 176 /*
175 * This structure is stored inside the "privat 177 * This structure is stored inside the "private_data" member of the file
176 * structure and represents the main data stru 178 * structure and represents the main data structure for the eventpoll
177 * interface. 179 * interface.
178 */ 180 */
179 struct eventpoll { 181 struct eventpoll {
180 /* 182 /*
181 * This mutex is used to ensure that f 183 * This mutex is used to ensure that files are not removed
182 * while epoll is using them. This is 184 * while epoll is using them. This is held during the event
183 * collection loop, the file cleanup p 185 * collection loop, the file cleanup path, the epoll file exit
184 * code and the ctl operations. 186 * code and the ctl operations.
185 */ 187 */
186 struct mutex mtx; 188 struct mutex mtx;
187 189
188 /* Wait queue used by sys_epoll_wait() 190 /* Wait queue used by sys_epoll_wait() */
189 wait_queue_head_t wq; 191 wait_queue_head_t wq;
190 192
191 /* Wait queue used by file->poll() */ 193 /* Wait queue used by file->poll() */
192 wait_queue_head_t poll_wait; 194 wait_queue_head_t poll_wait;
193 195
194 /* List of ready file descriptors */ 196 /* List of ready file descriptors */
195 struct list_head rdllist; 197 struct list_head rdllist;
196 198
197 /* Lock which protects rdllist and ovf 199 /* Lock which protects rdllist and ovflist */
198 rwlock_t lock; 200 rwlock_t lock;
199 201
200 /* RB tree root used to store monitore 202 /* RB tree root used to store monitored fd structs */
201 struct rb_root_cached rbr; 203 struct rb_root_cached rbr;
202 204
203 /* 205 /*
204 * This is a single linked list that c 206 * This is a single linked list that chains all the "struct epitem" that
205 * happened while transferring ready e 207 * happened while transferring ready events to userspace w/out
206 * holding ->lock. 208 * holding ->lock.
207 */ 209 */
208 struct epitem *ovflist; 210 struct epitem *ovflist;
209 211
210 /* wakeup_source used when ep_send_eve !! 212 /* wakeup_source used when ep_scan_ready_list is running */
211 struct wakeup_source *ws; 213 struct wakeup_source *ws;
212 214
213 /* The user that created the eventpoll 215 /* The user that created the eventpoll descriptor */
214 struct user_struct *user; 216 struct user_struct *user;
215 217
216 struct file *file; 218 struct file *file;
217 219
218 /* used to optimize loop detection che 220 /* used to optimize loop detection check */
219 u64 gen; 221 u64 gen;
220 struct hlist_head refs; <<
221 <<
222 /* <<
223 * usage count, used together with epi <<
224 * orchestrate the disposal of this st <<
225 */ <<
226 refcount_t refcount; <<
227 222
228 #ifdef CONFIG_NET_RX_BUSY_POLL 223 #ifdef CONFIG_NET_RX_BUSY_POLL
229 /* used to track busy poll napi_id */ 224 /* used to track busy poll napi_id */
230 unsigned int napi_id; 225 unsigned int napi_id;
231 /* busy poll timeout */ <<
232 u32 busy_poll_usecs; <<
233 /* busy poll packet budget */ <<
234 u16 busy_poll_budget; <<
235 bool prefer_busy_poll; <<
236 #endif 226 #endif
237 227
238 #ifdef CONFIG_DEBUG_LOCK_ALLOC 228 #ifdef CONFIG_DEBUG_LOCK_ALLOC
239 /* tracks wakeup nests for lockdep val 229 /* tracks wakeup nests for lockdep validation */
240 u8 nests; 230 u8 nests;
241 #endif 231 #endif
242 }; 232 };
243 233
>> 234 /* Wait structure used by the poll hooks */
>> 235 struct eppoll_entry {
>> 236 /* List header used to link this structure to the "struct epitem" */
>> 237 struct list_head llink;
>> 238
>> 239 /* The "base" pointer is set to the container "struct epitem" */
>> 240 struct epitem *base;
>> 241
>> 242 /*
>> 243 * Wait queue item that will be linked to the target file wait
>> 244 * queue head.
>> 245 */
>> 246 wait_queue_entry_t wait;
>> 247
>> 248 /* The wait queue head that linked the "wait" wait queue item */
>> 249 wait_queue_head_t *whead;
>> 250 };
>> 251
244 /* Wrapper struct used by poll queueing */ 252 /* Wrapper struct used by poll queueing */
245 struct ep_pqueue { 253 struct ep_pqueue {
246 poll_table pt; 254 poll_table pt;
247 struct epitem *epi; 255 struct epitem *epi;
248 }; 256 };
249 257
>> 258 /* Used by the ep_send_events() function as callback private data */
>> 259 struct ep_send_events_data {
>> 260 int maxevents;
>> 261 struct epoll_event __user *events;
>> 262 int res;
>> 263 };
>> 264
250 /* 265 /*
251 * Configuration options available inside /pro 266 * Configuration options available inside /proc/sys/fs/epoll/
252 */ 267 */
253 /* Maximum number of epoll watched descriptors 268 /* Maximum number of epoll watched descriptors, per user */
254 static long max_user_watches __read_mostly; 269 static long max_user_watches __read_mostly;
255 270
256 /* Used for cycles detection */ !! 271 /*
257 static DEFINE_MUTEX(epnested_mutex); !! 272 * This mutex is used to serialize ep_free() and eventpoll_release_file().
>> 273 */
>> 274 static DEFINE_MUTEX(epmutex);
258 275
259 static u64 loop_check_gen = 0; 276 static u64 loop_check_gen = 0;
260 277
261 /* Used to check for epoll file descriptor inc 278 /* Used to check for epoll file descriptor inclusion loops */
262 static struct eventpoll *inserting_into; !! 279 static struct nested_calls poll_loop_ncalls;
263 280
264 /* Slab cache used to allocate "struct epitem" 281 /* Slab cache used to allocate "struct epitem" */
265 static struct kmem_cache *epi_cache __ro_after !! 282 static struct kmem_cache *epi_cache __read_mostly;
266 283
267 /* Slab cache used to allocate "struct eppoll_ 284 /* Slab cache used to allocate "struct eppoll_entry" */
268 static struct kmem_cache *pwq_cache __ro_after !! 285 static struct kmem_cache *pwq_cache __read_mostly;
269 286
270 /* 287 /*
271 * List of files with newly added links, where 288 * List of files with newly added links, where we may need to limit the number
272 * of emanating paths. Protected by the epnest !! 289 * of emanating paths. Protected by the epmutex.
273 */ 290 */
274 struct epitems_head { !! 291 static LIST_HEAD(tfile_check_list);
275 struct hlist_head epitems; <<
276 struct epitems_head *next; <<
277 }; <<
278 static struct epitems_head *tfile_check_list = <<
279 <<
280 static struct kmem_cache *ephead_cache __ro_af <<
281 <<
282 static inline void free_ephead(struct epitems_ <<
283 { <<
284 if (head) <<
285 kmem_cache_free(ephead_cache, <<
286 } <<
287 <<
288 static void list_file(struct file *file) <<
289 { <<
290 struct epitems_head *head; <<
291 <<
292 head = container_of(file->f_ep, struct <<
293 if (!head->next) { <<
294 head->next = tfile_check_list; <<
295 tfile_check_list = head; <<
296 } <<
297 } <<
298 <<
299 static void unlist_file(struct epitems_head *h <<
300 { <<
301 struct epitems_head *to_free = head; <<
302 struct hlist_node *p = rcu_dereference <<
303 if (p) { <<
304 struct epitem *epi= container_ <<
305 spin_lock(&epi->ffd.file->f_lo <<
306 if (!hlist_empty(&head->epitem <<
307 to_free = NULL; <<
308 head->next = NULL; <<
309 spin_unlock(&epi->ffd.file->f_ <<
310 } <<
311 free_ephead(to_free); <<
312 } <<
313 292
314 #ifdef CONFIG_SYSCTL 293 #ifdef CONFIG_SYSCTL
315 294
316 #include <linux/sysctl.h> 295 #include <linux/sysctl.h>
317 296
318 static long long_zero; 297 static long long_zero;
319 static long long_max = LONG_MAX; 298 static long long_max = LONG_MAX;
320 299
321 static struct ctl_table epoll_table[] = { !! 300 struct ctl_table epoll_table[] = {
322 { 301 {
323 .procname = "max_user_wa 302 .procname = "max_user_watches",
324 .data = &max_user_wa 303 .data = &max_user_watches,
325 .maxlen = sizeof(max_u 304 .maxlen = sizeof(max_user_watches),
326 .mode = 0644, 305 .mode = 0644,
327 .proc_handler = proc_doulong 306 .proc_handler = proc_doulongvec_minmax,
328 .extra1 = &long_zero, 307 .extra1 = &long_zero,
329 .extra2 = &long_max, 308 .extra2 = &long_max,
330 }, 309 },
>> 310 { }
331 }; 311 };
332 <<
333 static void __init epoll_sysctls_init(void) <<
334 { <<
335 register_sysctl("fs/epoll", epoll_tabl <<
336 } <<
337 #else <<
338 #define epoll_sysctls_init() do { } while (0) <<
339 #endif /* CONFIG_SYSCTL */ 312 #endif /* CONFIG_SYSCTL */
340 313
341 static const struct file_operations eventpoll_ 314 static const struct file_operations eventpoll_fops;
342 315
343 static inline int is_file_epoll(struct file *f 316 static inline int is_file_epoll(struct file *f)
344 { 317 {
345 return f->f_op == &eventpoll_fops; 318 return f->f_op == &eventpoll_fops;
346 } 319 }
347 320
348 /* Setup the structure that is used as key for 321 /* Setup the structure that is used as key for the RB tree */
349 static inline void ep_set_ffd(struct epoll_fil 322 static inline void ep_set_ffd(struct epoll_filefd *ffd,
350 struct file *fil 323 struct file *file, int fd)
351 { 324 {
352 ffd->file = file; 325 ffd->file = file;
353 ffd->fd = fd; 326 ffd->fd = fd;
354 } 327 }
355 328
356 /* Compare RB tree keys */ 329 /* Compare RB tree keys */
357 static inline int ep_cmp_ffd(struct epoll_file 330 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
358 struct epoll_file 331 struct epoll_filefd *p2)
359 { 332 {
360 return (p1->file > p2->file ? +1: 333 return (p1->file > p2->file ? +1:
361 (p1->file < p2->file ? -1 : p1 334 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
362 } 335 }
363 336
364 /* Tells us if the item is currently linked */ 337 /* Tells us if the item is currently linked */
365 static inline int ep_is_linked(struct epitem * 338 static inline int ep_is_linked(struct epitem *epi)
366 { 339 {
367 return !list_empty(&epi->rdllink); 340 return !list_empty(&epi->rdllink);
368 } 341 }
369 342
370 static inline struct eppoll_entry *ep_pwq_from 343 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
371 { 344 {
372 return container_of(p, struct eppoll_e 345 return container_of(p, struct eppoll_entry, wait);
373 } 346 }
374 347
375 /* Get the "struct epitem" from a wait queue p 348 /* Get the "struct epitem" from a wait queue pointer */
376 static inline struct epitem *ep_item_from_wait 349 static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
377 { 350 {
378 return container_of(p, struct eppoll_e 351 return container_of(p, struct eppoll_entry, wait)->base;
379 } 352 }
380 353
>> 354 /* Get the "struct epitem" from an epoll queue wrapper */
>> 355 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
>> 356 {
>> 357 return container_of(p, struct ep_pqueue, pt)->epi;
>> 358 }
>> 359
>> 360 /* Initialize the poll safe wake up structure */
>> 361 static void ep_nested_calls_init(struct nested_calls *ncalls)
>> 362 {
>> 363 INIT_LIST_HEAD(&ncalls->tasks_call_list);
>> 364 spin_lock_init(&ncalls->lock);
>> 365 }
>> 366
381 /** 367 /**
382 * ep_events_available - Checks if ready event 368 * ep_events_available - Checks if ready events might be available.
383 * 369 *
384 * @ep: Pointer to the eventpoll context. 370 * @ep: Pointer to the eventpoll context.
385 * 371 *
386 * Return: a value different than %zero if rea !! 372 * Returns: Returns a value different than zero if ready events are available,
387 * or %zero otherwise. !! 373 * or zero otherwise.
388 */ 374 */
389 static inline int ep_events_available(struct e 375 static inline int ep_events_available(struct eventpoll *ep)
390 { 376 {
391 return !list_empty_careful(&ep->rdllis 377 return !list_empty_careful(&ep->rdllist) ||
392 READ_ONCE(ep->ovflist) != EP_U 378 READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR;
393 } 379 }
394 380
395 #ifdef CONFIG_NET_RX_BUSY_POLL 381 #ifdef CONFIG_NET_RX_BUSY_POLL
396 /** <<
397 * busy_loop_ep_timeout - check if busy poll h <<
398 * from the epoll instance ep is preferred, bu <<
399 * the system-wide global via busy_loop_timeou <<
400 * <<
401 * @start_time: The start time used to compute <<
402 * @ep: Pointer to the eventpoll context. <<
403 * <<
404 * Return: true if the timeout has expired, fa <<
405 */ <<
406 static bool busy_loop_ep_timeout(unsigned long <<
407 struct eventp <<
408 { <<
409 unsigned long bp_usec = READ_ONCE(ep-> <<
410 <<
411 if (bp_usec) { <<
412 unsigned long end_time = start <<
413 unsigned long now = busy_loop_ <<
414 <<
415 return time_after(now, end_tim <<
416 } else { <<
417 return busy_loop_timeout(start <<
418 } <<
419 } <<
420 <<
421 static bool ep_busy_loop_on(struct eventpoll * <<
422 { <<
423 return !!READ_ONCE(ep->busy_poll_usecs <<
424 } <<
425 <<
426 static bool ep_busy_loop_end(void *p, unsigned 382 static bool ep_busy_loop_end(void *p, unsigned long start_time)
427 { 383 {
428 struct eventpoll *ep = p; 384 struct eventpoll *ep = p;
429 385
430 return ep_events_available(ep) || busy !! 386 return ep_events_available(ep) || busy_loop_timeout(start_time);
431 } 387 }
432 388
433 /* 389 /*
434 * Busy poll if globally on and supporting soc 390 * Busy poll if globally on and supporting sockets found && no events,
435 * busy loop will return if need_resched or ep 391 * busy loop will return if need_resched or ep_events_available.
436 * 392 *
437 * we must do our busy polling with irqs enabl 393 * we must do our busy polling with irqs enabled
438 */ 394 */
439 static bool ep_busy_loop(struct eventpoll *ep, !! 395 static void ep_busy_loop(struct eventpoll *ep, int nonblock)
440 { 396 {
441 unsigned int napi_id = READ_ONCE(ep->n 397 unsigned int napi_id = READ_ONCE(ep->napi_id);
442 u16 budget = READ_ONCE(ep->busy_poll_b <<
443 bool prefer_busy_poll = READ_ONCE(ep-> <<
444 398
445 if (!budget) !! 399 if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on())
446 budget = BUSY_POLL_BUDGET; !! 400 napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep);
>> 401 }
447 402
448 if (napi_id >= MIN_NAPI_ID && ep_busy_ !! 403 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
449 napi_busy_loop(napi_id, nonblo !! 404 {
450 ep, prefer_busy !! 405 if (ep->napi_id)
451 if (ep_events_available(ep)) <<
452 return true; <<
453 /* <<
454 * Busy poll timed out. Drop <<
455 * it back in when we have mov <<
456 * ID onto the ready list. <<
457 */ <<
458 ep->napi_id = 0; 406 ep->napi_id = 0;
459 return false; <<
460 } <<
461 return false; <<
462 } 407 }
463 408
464 /* 409 /*
465 * Set epoll busy poll NAPI ID from sk. 410 * Set epoll busy poll NAPI ID from sk.
466 */ 411 */
467 static inline void ep_set_busy_poll_napi_id(st 412 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
468 { 413 {
469 struct eventpoll *ep = epi->ep; !! 414 struct eventpoll *ep;
470 unsigned int napi_id; 415 unsigned int napi_id;
471 struct socket *sock; 416 struct socket *sock;
472 struct sock *sk; 417 struct sock *sk;
>> 418 int err;
473 419
474 if (!ep_busy_loop_on(ep)) !! 420 if (!net_busy_loop_on())
475 return; 421 return;
476 422
477 sock = sock_from_file(epi->ffd.file); !! 423 sock = sock_from_file(epi->ffd.file, &err);
478 if (!sock) 424 if (!sock)
479 return; 425 return;
480 426
481 sk = sock->sk; 427 sk = sock->sk;
482 if (!sk) 428 if (!sk)
483 return; 429 return;
484 430
485 napi_id = READ_ONCE(sk->sk_napi_id); 431 napi_id = READ_ONCE(sk->sk_napi_id);
>> 432 ep = epi->ep;
486 433
487 /* Non-NAPI IDs can be rejected 434 /* Non-NAPI IDs can be rejected
488 * or 435 * or
489 * Nothing to do if we already have th 436 * Nothing to do if we already have this ID
490 */ 437 */
491 if (napi_id < MIN_NAPI_ID || napi_id = 438 if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)
492 return; 439 return;
493 440
494 /* record NAPI ID for use in next busy 441 /* record NAPI ID for use in next busy poll */
495 ep->napi_id = napi_id; 442 ep->napi_id = napi_id;
496 } 443 }
497 444
498 static long ep_eventpoll_bp_ioctl(struct file <<
499 unsigned lon <<
500 { <<
501 struct eventpoll *ep = file->private_d <<
502 void __user *uarg = (void __user *)arg <<
503 struct epoll_params epoll_params; <<
504 <<
505 switch (cmd) { <<
506 case EPIOCSPARAMS: <<
507 if (copy_from_user(&epoll_para <<
508 return -EFAULT; <<
509 <<
510 /* pad byte must be zero */ <<
511 if (epoll_params.__pad) <<
512 return -EINVAL; <<
513 <<
514 if (epoll_params.busy_poll_use <<
515 return -EINVAL; <<
516 <<
517 if (epoll_params.prefer_busy_p <<
518 return -EINVAL; <<
519 <<
520 if (epoll_params.busy_poll_bud <<
521 !capable(CAP_NET_ADMIN)) <<
522 return -EPERM; <<
523 <<
524 WRITE_ONCE(ep->busy_poll_usecs <<
525 WRITE_ONCE(ep->busy_poll_budge <<
526 WRITE_ONCE(ep->prefer_busy_pol <<
527 return 0; <<
528 case EPIOCGPARAMS: <<
529 memset(&epoll_params, 0, sizeo <<
530 epoll_params.busy_poll_usecs = <<
531 epoll_params.busy_poll_budget <<
532 epoll_params.prefer_busy_poll <<
533 if (copy_to_user(uarg, &epoll_ <<
534 return -EFAULT; <<
535 return 0; <<
536 default: <<
537 return -ENOIOCTLCMD; <<
538 } <<
539 } <<
540 <<
541 #else 445 #else
542 446
543 static inline bool ep_busy_loop(struct eventpo !! 447 static inline void ep_busy_loop(struct eventpoll *ep, int nonblock)
544 { 448 {
545 return false; <<
546 } 449 }
547 450
548 static inline void ep_set_busy_poll_napi_id(st !! 451 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
549 { 452 {
550 } 453 }
551 454
552 static long ep_eventpoll_bp_ioctl(struct file !! 455 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
553 unsigned lon <<
554 { 456 {
555 return -EOPNOTSUPP; <<
556 } 457 }
557 458
558 #endif /* CONFIG_NET_RX_BUSY_POLL */ 459 #endif /* CONFIG_NET_RX_BUSY_POLL */
559 460
>> 461 /**
>> 462 * ep_call_nested - Perform a bound (possibly) nested call, by checking
>> 463 * that the recursion limit is not exceeded, and that
>> 464 * the same nested call (by the meaning of same cookie) is
>> 465 * no re-entered.
>> 466 *
>> 467 * @ncalls: Pointer to the nested_calls structure to be used for this call.
>> 468 * @nproc: Nested call core function pointer.
>> 469 * @priv: Opaque data to be passed to the @nproc callback.
>> 470 * @cookie: Cookie to be used to identify this nested call.
>> 471 * @ctx: This instance context.
>> 472 *
>> 473 * Returns: Returns the code returned by the @nproc callback, or -1 if
>> 474 * the maximum recursion limit has been exceeded.
>> 475 */
>> 476 static int ep_call_nested(struct nested_calls *ncalls,
>> 477 int (*nproc)(void *, void *, int), void *priv,
>> 478 void *cookie, void *ctx)
>> 479 {
>> 480 int error, call_nests = 0;
>> 481 unsigned long flags;
>> 482 struct list_head *lsthead = &ncalls->tasks_call_list;
>> 483 struct nested_call_node *tncur;
>> 484 struct nested_call_node tnode;
>> 485
>> 486 spin_lock_irqsave(&ncalls->lock, flags);
>> 487
>> 488 /*
>> 489 * Try to see if the current task is already inside this wakeup call.
>> 490 * We use a list here, since the population inside this set is always
>> 491 * very much limited.
>> 492 */
>> 493 list_for_each_entry(tncur, lsthead, llink) {
>> 494 if (tncur->ctx == ctx &&
>> 495 (tncur->cookie == cookie || ++call_nests > EP_MAX_NESTS)) {
>> 496 /*
>> 497 * Ops ... loop detected or maximum nest level reached.
>> 498 * We abort this wake by breaking the cycle itself.
>> 499 */
>> 500 error = -1;
>> 501 goto out_unlock;
>> 502 }
>> 503 }
>> 504
>> 505 /* Add the current task and cookie to the list */
>> 506 tnode.ctx = ctx;
>> 507 tnode.cookie = cookie;
>> 508 list_add(&tnode.llink, lsthead);
>> 509
>> 510 spin_unlock_irqrestore(&ncalls->lock, flags);
>> 511
>> 512 /* Call the nested function */
>> 513 error = (*nproc)(priv, cookie, call_nests);
>> 514
>> 515 /* Remove the current task from the list */
>> 516 spin_lock_irqsave(&ncalls->lock, flags);
>> 517 list_del(&tnode.llink);
>> 518 out_unlock:
>> 519 spin_unlock_irqrestore(&ncalls->lock, flags);
>> 520
>> 521 return error;
>> 522 }
>> 523
560 /* 524 /*
561 * As described in commit 0ccf831cb lockdep: a 525 * As described in commit 0ccf831cb lockdep: annotate epoll
562 * the use of wait queues used by epoll is don 526 * the use of wait queues used by epoll is done in a very controlled
563 * manner. Wake ups can nest inside each other 527 * manner. Wake ups can nest inside each other, but are never done
564 * with the same locking. For example: 528 * with the same locking. For example:
565 * 529 *
566 * dfd = socket(...); 530 * dfd = socket(...);
567 * efd1 = epoll_create(); 531 * efd1 = epoll_create();
568 * efd2 = epoll_create(); 532 * efd2 = epoll_create();
569 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...); 533 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
570 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...) 534 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
571 * 535 *
572 * When a packet arrives to the device underne 536 * When a packet arrives to the device underneath "dfd", the net code will
573 * issue a wake_up() on its poll wake list. Ep 537 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
574 * callback wakeup entry on that queue, and th 538 * callback wakeup entry on that queue, and the wake_up() performed by the
575 * "dfd" net code will end up in ep_poll_callb 539 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
576 * (efd1) notices that it may have some event 540 * (efd1) notices that it may have some event ready, so it needs to wake up
577 * the waiters on its poll wait list (efd2). S 541 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
578 * that ends up in another wake_up(), after ha 542 * that ends up in another wake_up(), after having checked about the
579 * recursion constraints. That are, no more th !! 543 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
580 * stack blasting. !! 544 * avoid stack blasting.
581 * 545 *
582 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, ma 546 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
583 * this special case of epoll. 547 * this special case of epoll.
584 */ 548 */
585 #ifdef CONFIG_DEBUG_LOCK_ALLOC 549 #ifdef CONFIG_DEBUG_LOCK_ALLOC
586 550
587 static void ep_poll_safewake(struct eventpoll 551 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi,
588 unsigned pollflag 552 unsigned pollflags)
589 { 553 {
590 struct eventpoll *ep_src; 554 struct eventpoll *ep_src;
591 unsigned long flags; 555 unsigned long flags;
592 u8 nests = 0; 556 u8 nests = 0;
593 557
594 /* 558 /*
595 * To set the subclass or nesting leve 559 * To set the subclass or nesting level for spin_lock_irqsave_nested()
596 * it might be natural to create a per 560 * it might be natural to create a per-cpu nest count. However, since
597 * we can recurse on ep->poll_wait.loc 561 * we can recurse on ep->poll_wait.lock, and a non-raw spinlock can
598 * schedule() in the -rt kernel, the p 562 * schedule() in the -rt kernel, the per-cpu variable are no longer
599 * protected. Thus, we are introducing 563 * protected. Thus, we are introducing a per eventpoll nest field.
600 * If we are not being call from ep_po 564 * If we are not being call from ep_poll_callback(), epi is NULL and
601 * we are at the first level of nestin 565 * we are at the first level of nesting, 0. Otherwise, we are being
602 * called from ep_poll_callback() and 566 * called from ep_poll_callback() and if a previous wakeup source is
603 * not an epoll file itself, we are at 567 * not an epoll file itself, we are at depth 1 since the wakeup source
604 * is depth 0. If the wakeup source is 568 * is depth 0. If the wakeup source is a previous epoll file in the
605 * wakeup chain then we use its nests 569 * wakeup chain then we use its nests value and record ours as
606 * nests + 1. The previous epoll file 570 * nests + 1. The previous epoll file nests value is stable since its
607 * already holding its own poll_wait.l 571 * already holding its own poll_wait.lock.
608 */ 572 */
609 if (epi) { 573 if (epi) {
610 if ((is_file_epoll(epi->ffd.fi 574 if ((is_file_epoll(epi->ffd.file))) {
611 ep_src = epi->ffd.file 575 ep_src = epi->ffd.file->private_data;
612 nests = ep_src->nests; 576 nests = ep_src->nests;
613 } else { 577 } else {
614 nests = 1; 578 nests = 1;
615 } 579 }
616 } 580 }
617 spin_lock_irqsave_nested(&ep->poll_wai 581 spin_lock_irqsave_nested(&ep->poll_wait.lock, flags, nests);
618 ep->nests = nests + 1; 582 ep->nests = nests + 1;
619 wake_up_locked_poll(&ep->poll_wait, EP 583 wake_up_locked_poll(&ep->poll_wait, EPOLLIN | pollflags);
620 ep->nests = 0; 584 ep->nests = 0;
621 spin_unlock_irqrestore(&ep->poll_wait. 585 spin_unlock_irqrestore(&ep->poll_wait.lock, flags);
622 } 586 }
623 587
624 #else 588 #else
625 589
626 static void ep_poll_safewake(struct eventpoll 590 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi,
627 __poll_t pollflag !! 591 unsigned pollflags)
628 { 592 {
629 wake_up_poll(&ep->poll_wait, EPOLLIN | 593 wake_up_poll(&ep->poll_wait, EPOLLIN | pollflags);
630 } 594 }
631 595
632 #endif 596 #endif
633 597
634 static void ep_remove_wait_queue(struct eppoll 598 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
635 { 599 {
636 wait_queue_head_t *whead; 600 wait_queue_head_t *whead;
637 601
638 rcu_read_lock(); 602 rcu_read_lock();
639 /* 603 /*
640 * If it is cleared by POLLFREE, it sh 604 * If it is cleared by POLLFREE, it should be rcu-safe.
641 * If we read NULL we need a barrier p 605 * If we read NULL we need a barrier paired with
642 * smp_store_release() in ep_poll_call 606 * smp_store_release() in ep_poll_callback(), otherwise
643 * we rely on whead->lock. 607 * we rely on whead->lock.
644 */ 608 */
645 whead = smp_load_acquire(&pwq->whead); 609 whead = smp_load_acquire(&pwq->whead);
646 if (whead) 610 if (whead)
647 remove_wait_queue(whead, &pwq- 611 remove_wait_queue(whead, &pwq->wait);
648 rcu_read_unlock(); 612 rcu_read_unlock();
649 } 613 }
650 614
651 /* 615 /*
652 * This function unregisters poll callbacks fr 616 * This function unregisters poll callbacks from the associated file
653 * descriptor. Must be called with "mtx" held !! 617 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
>> 618 * ep_free).
654 */ 619 */
655 static void ep_unregister_pollwait(struct even 620 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
656 { 621 {
657 struct eppoll_entry **p = &epi->pwqlis !! 622 struct list_head *lsthead = &epi->pwqlist;
658 struct eppoll_entry *pwq; 623 struct eppoll_entry *pwq;
659 624
660 while ((pwq = *p) != NULL) { !! 625 while (!list_empty(lsthead)) {
661 *p = pwq->next; !! 626 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
>> 627
>> 628 list_del(&pwq->llink);
662 ep_remove_wait_queue(pwq); 629 ep_remove_wait_queue(pwq);
663 kmem_cache_free(pwq_cache, pwq 630 kmem_cache_free(pwq_cache, pwq);
664 } 631 }
665 } 632 }
666 633
667 /* call only when ep->mtx is held */ 634 /* call only when ep->mtx is held */
668 static inline struct wakeup_source *ep_wakeup_ 635 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
669 { 636 {
670 return rcu_dereference_check(epi->ws, 637 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
671 } 638 }
672 639
673 /* call only when ep->mtx is held */ 640 /* call only when ep->mtx is held */
674 static inline void ep_pm_stay_awake(struct epi 641 static inline void ep_pm_stay_awake(struct epitem *epi)
675 { 642 {
676 struct wakeup_source *ws = ep_wakeup_s 643 struct wakeup_source *ws = ep_wakeup_source(epi);
677 644
678 if (ws) 645 if (ws)
679 __pm_stay_awake(ws); 646 __pm_stay_awake(ws);
680 } 647 }
681 648
682 static inline bool ep_has_wakeup_source(struct 649 static inline bool ep_has_wakeup_source(struct epitem *epi)
683 { 650 {
684 return rcu_access_pointer(epi->ws) ? t 651 return rcu_access_pointer(epi->ws) ? true : false;
685 } 652 }
686 653
687 /* call when ep->mtx cannot be held (ep_poll_c 654 /* call when ep->mtx cannot be held (ep_poll_callback) */
688 static inline void ep_pm_stay_awake_rcu(struct 655 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
689 { 656 {
690 struct wakeup_source *ws; 657 struct wakeup_source *ws;
691 658
692 rcu_read_lock(); 659 rcu_read_lock();
693 ws = rcu_dereference(epi->ws); 660 ws = rcu_dereference(epi->ws);
694 if (ws) 661 if (ws)
695 __pm_stay_awake(ws); 662 __pm_stay_awake(ws);
696 rcu_read_unlock(); 663 rcu_read_unlock();
697 } 664 }
698 665
699 !! 666 /**
700 /* !! 667 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
701 * ep->mutex needs to be held because we could !! 668 * the scan code, to call f_op->poll(). Also allows for
702 * eventpoll_release_file() and epoll_ctl(). !! 669 * O(NumReady) performance.
703 */ !! 670 *
704 static void ep_start_scan(struct eventpoll *ep !! 671 * @ep: Pointer to the epoll private data structure.
>> 672 * @sproc: Pointer to the scan callback.
>> 673 * @priv: Private opaque data passed to the @sproc callback.
>> 674 * @depth: The current depth of recursive f_op->poll calls.
>> 675 * @ep_locked: caller already holds ep->mtx
>> 676 *
>> 677 * Returns: The same integer error code returned by the @sproc callback.
>> 678 */
>> 679 static __poll_t ep_scan_ready_list(struct eventpoll *ep,
>> 680 __poll_t (*sproc)(struct eventpoll *,
>> 681 struct list_head *, void *),
>> 682 void *priv, int depth, bool ep_locked)
705 { 683 {
>> 684 __poll_t res;
>> 685 struct epitem *epi, *nepi;
>> 686 LIST_HEAD(txlist);
>> 687
>> 688 lockdep_assert_irqs_enabled();
>> 689
>> 690 /*
>> 691 * We need to lock this because we could be hit by
>> 692 * eventpoll_release_file() and epoll_ctl().
>> 693 */
>> 694
>> 695 if (!ep_locked)
>> 696 mutex_lock_nested(&ep->mtx, depth);
>> 697
706 /* 698 /*
707 * Steal the ready list, and re-init t 699 * Steal the ready list, and re-init the original one to the
708 * empty list. Also, set ep->ovflist t 700 * empty list. Also, set ep->ovflist to NULL so that events
709 * happening while looping w/out locks 701 * happening while looping w/out locks, are not lost. We cannot
710 * have the poll callback to queue dir 702 * have the poll callback to queue directly on ep->rdllist,
711 * because we want the "sproc" callbac 703 * because we want the "sproc" callback to be able to do it
712 * in a lockless way. 704 * in a lockless way.
713 */ 705 */
714 lockdep_assert_irqs_enabled(); <<
715 write_lock_irq(&ep->lock); 706 write_lock_irq(&ep->lock);
716 list_splice_init(&ep->rdllist, txlist) !! 707 list_splice_init(&ep->rdllist, &txlist);
717 WRITE_ONCE(ep->ovflist, NULL); 708 WRITE_ONCE(ep->ovflist, NULL);
718 write_unlock_irq(&ep->lock); 709 write_unlock_irq(&ep->lock);
719 } <<
720 710
721 static void ep_done_scan(struct eventpoll *ep, !! 711 /*
722 struct list_head *txl !! 712 * Now call the callback function.
723 { !! 713 */
724 struct epitem *epi, *nepi; !! 714 res = (*sproc)(ep, &txlist, priv);
725 715
726 write_lock_irq(&ep->lock); 716 write_lock_irq(&ep->lock);
727 /* 717 /*
728 * During the time we spent inside the 718 * During the time we spent inside the "sproc" callback, some
729 * other events might have been queued 719 * other events might have been queued by the poll callback.
730 * We re-insert them inside the main r 720 * We re-insert them inside the main ready-list here.
731 */ 721 */
732 for (nepi = READ_ONCE(ep->ovflist); (e 722 for (nepi = READ_ONCE(ep->ovflist); (epi = nepi) != NULL;
733 nepi = epi->next, epi->next = EP_ 723 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
734 /* 724 /*
735 * We need to check if the ite 725 * We need to check if the item is already in the list.
736 * During the "sproc" callback 726 * During the "sproc" callback execution time, items are
737 * queued into ->ovflist but t 727 * queued into ->ovflist but the "txlist" might already
738 * contain them, and the list_ 728 * contain them, and the list_splice() below takes care of them.
739 */ 729 */
740 if (!ep_is_linked(epi)) { 730 if (!ep_is_linked(epi)) {
741 /* 731 /*
742 * ->ovflist is LIFO, 732 * ->ovflist is LIFO, so we have to reverse it in order
743 * to keep in FIFO. 733 * to keep in FIFO.
744 */ 734 */
745 list_add(&epi->rdllink 735 list_add(&epi->rdllink, &ep->rdllist);
746 ep_pm_stay_awake(epi); 736 ep_pm_stay_awake(epi);
747 } 737 }
748 } 738 }
749 /* 739 /*
750 * We need to set back ep->ovflist to 740 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
751 * releasing the lock, events will be 741 * releasing the lock, events will be queued in the normal way inside
752 * ep->rdllist. 742 * ep->rdllist.
753 */ 743 */
754 WRITE_ONCE(ep->ovflist, EP_UNACTIVE_PT 744 WRITE_ONCE(ep->ovflist, EP_UNACTIVE_PTR);
755 745
756 /* 746 /*
757 * Quickly re-inject items left on "tx 747 * Quickly re-inject items left on "txlist".
758 */ 748 */
759 list_splice(txlist, &ep->rdllist); !! 749 list_splice(&txlist, &ep->rdllist);
760 __pm_relax(ep->ws); 750 __pm_relax(ep->ws);
761 751
762 if (!list_empty(&ep->rdllist)) { 752 if (!list_empty(&ep->rdllist)) {
763 if (waitqueue_active(&ep->wq)) 753 if (waitqueue_active(&ep->wq))
764 wake_up(&ep->wq); 754 wake_up(&ep->wq);
765 } 755 }
766 756
767 write_unlock_irq(&ep->lock); 757 write_unlock_irq(&ep->lock);
768 } <<
769 <<
770 static void ep_get(struct eventpoll *ep) <<
771 { <<
772 refcount_inc(&ep->refcount); <<
773 } <<
774 758
775 /* !! 759 if (!ep_locked)
776 * Returns true if the event poll can be dispo !! 760 mutex_unlock(&ep->mtx);
777 */ <<
778 static bool ep_refcount_dec_and_test(struct ev <<
779 { <<
780 if (!refcount_dec_and_test(&ep->refcou <<
781 return false; <<
782 761
783 WARN_ON_ONCE(!RB_EMPTY_ROOT(&ep->rbr.r !! 762 return res;
784 return true; <<
785 } 763 }
786 764
787 static void ep_free(struct eventpoll *ep) !! 765 static void epi_rcu_free(struct rcu_head *head)
788 { 766 {
789 mutex_destroy(&ep->mtx); !! 767 struct epitem *epi = container_of(head, struct epitem, rcu);
790 free_uid(ep->user); !! 768 kmem_cache_free(epi_cache, epi);
791 wakeup_source_unregister(ep->ws); <<
792 kfree(ep); <<
793 } 769 }
794 770
795 /* 771 /*
796 * Removes a "struct epitem" from the eventpol 772 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
797 * all the associated resources. Must be calle 773 * all the associated resources. Must be called with "mtx" held.
798 * If the dying flag is set, do the removal on <<
799 * This prevents ep_clear_and_put() from dropp <<
800 * while running concurrently with eventpoll_r <<
801 * Returns true if the eventpoll can be dispos <<
802 */ 774 */
803 static bool __ep_remove(struct eventpoll *ep, !! 775 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
804 { 776 {
805 struct file *file = epi->ffd.file; 777 struct file *file = epi->ffd.file;
806 struct epitems_head *to_free; <<
807 struct hlist_head *head; <<
808 778
809 lockdep_assert_irqs_enabled(); 779 lockdep_assert_irqs_enabled();
810 780
811 /* 781 /*
812 * Removes poll wait queue hooks. 782 * Removes poll wait queue hooks.
813 */ 783 */
814 ep_unregister_pollwait(ep, epi); 784 ep_unregister_pollwait(ep, epi);
815 785
816 /* Remove the current item from the li 786 /* Remove the current item from the list of epoll hooks */
817 spin_lock(&file->f_lock); 787 spin_lock(&file->f_lock);
818 if (epi->dying && !force) { !! 788 list_del_rcu(&epi->fllink);
819 spin_unlock(&file->f_lock); <<
820 return false; <<
821 } <<
822 <<
823 to_free = NULL; <<
824 head = file->f_ep; <<
825 if (head->first == &epi->fllink && !ep <<
826 file->f_ep = NULL; <<
827 if (!is_file_epoll(file)) { <<
828 struct epitems_head *v <<
829 v = container_of(head, <<
830 if (!smp_load_acquire( <<
831 to_free = v; <<
832 } <<
833 } <<
834 hlist_del_rcu(&epi->fllink); <<
835 spin_unlock(&file->f_lock); 789 spin_unlock(&file->f_lock);
836 free_ephead(to_free); <<
837 790
838 rb_erase_cached(&epi->rbn, &ep->rbr); 791 rb_erase_cached(&epi->rbn, &ep->rbr);
839 792
840 write_lock_irq(&ep->lock); 793 write_lock_irq(&ep->lock);
841 if (ep_is_linked(epi)) 794 if (ep_is_linked(epi))
842 list_del_init(&epi->rdllink); 795 list_del_init(&epi->rdllink);
843 write_unlock_irq(&ep->lock); 796 write_unlock_irq(&ep->lock);
844 797
845 wakeup_source_unregister(ep_wakeup_sou 798 wakeup_source_unregister(ep_wakeup_source(epi));
846 /* 799 /*
847 * At this point it is safe to free th 800 * At this point it is safe to free the eventpoll item. Use the union
848 * field epi->rcu, since we are trying 801 * field epi->rcu, since we are trying to minimize the size of
849 * 'struct epitem'. The 'rbn' field is 802 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
850 * ep->mtx. The rcu read side, reverse 803 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
851 * use of the rbn field. 804 * use of the rbn field.
852 */ 805 */
853 kfree_rcu(epi, rcu); !! 806 call_rcu(&epi->rcu, epi_rcu_free);
854 807
855 percpu_counter_dec(&ep->user->epoll_wa !! 808 atomic_long_dec(&ep->user->epoll_watches);
856 return ep_refcount_dec_and_test(ep); <<
857 } <<
858 809
859 /* !! 810 return 0;
860 * ep_remove variant for callers owing an addi <<
861 */ <<
862 static void ep_remove_safe(struct eventpoll *e <<
863 { <<
864 WARN_ON_ONCE(__ep_remove(ep, epi, fals <<
865 } 811 }
866 812
867 static void ep_clear_and_put(struct eventpoll !! 813 static void ep_free(struct eventpoll *ep)
868 { 814 {
869 struct rb_node *rbp, *next; !! 815 struct rb_node *rbp;
870 struct epitem *epi; 816 struct epitem *epi;
871 bool dispose; <<
872 817
873 /* We need to release all tasks waitin 818 /* We need to release all tasks waiting for these file */
874 if (waitqueue_active(&ep->poll_wait)) 819 if (waitqueue_active(&ep->poll_wait))
875 ep_poll_safewake(ep, NULL, 0); 820 ep_poll_safewake(ep, NULL, 0);
876 821
877 mutex_lock(&ep->mtx); !! 822 /*
>> 823 * We need to lock this because we could be hit by
>> 824 * eventpoll_release_file() while we're freeing the "struct eventpoll".
>> 825 * We do not need to hold "ep->mtx" here because the epoll file
>> 826 * is on the way to be removed and no one has references to it
>> 827 * anymore. The only hit might come from eventpoll_release_file() but
>> 828 * holding "epmutex" is sufficient here.
>> 829 */
>> 830 mutex_lock(&epmutex);
878 831
879 /* 832 /*
880 * Walks through the whole tree by unr 833 * Walks through the whole tree by unregistering poll callbacks.
881 */ 834 */
882 for (rbp = rb_first_cached(&ep->rbr); 835 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
883 epi = rb_entry(rbp, struct epi 836 epi = rb_entry(rbp, struct epitem, rbn);
884 837
885 ep_unregister_pollwait(ep, epi 838 ep_unregister_pollwait(ep, epi);
886 cond_resched(); 839 cond_resched();
887 } 840 }
888 841
889 /* 842 /*
890 * Walks through the whole tree and tr !! 843 * Walks through the whole tree by freeing each "struct epitem". At this
891 * Note that ep_remove_safe() will not !! 844 * point we are sure no poll callbacks will be lingering around, and also by
892 * racing eventpoll_release_file(); th !! 845 * holding "epmutex" we can be sure that no file cleanup code will hit
893 * At this point we are sure no poll c !! 846 * us during this operation. So we can avoid the lock on "ep->lock".
894 * Since we still own a reference to t !! 847 * We do not need to lock ep->mtx, either, we only do it to prevent
895 * dispose it. !! 848 * a lockdep warning.
896 */ 849 */
897 for (rbp = rb_first_cached(&ep->rbr); !! 850 mutex_lock(&ep->mtx);
898 next = rb_next(rbp); !! 851 while ((rbp = rb_first_cached(&ep->rbr)) != NULL) {
899 epi = rb_entry(rbp, struct epi 852 epi = rb_entry(rbp, struct epitem, rbn);
900 ep_remove_safe(ep, epi); !! 853 ep_remove(ep, epi);
901 cond_resched(); 854 cond_resched();
902 } 855 }
903 <<
904 dispose = ep_refcount_dec_and_test(ep) <<
905 mutex_unlock(&ep->mtx); 856 mutex_unlock(&ep->mtx);
906 857
907 if (dispose) !! 858 mutex_unlock(&epmutex);
908 ep_free(ep); !! 859 mutex_destroy(&ep->mtx);
909 } !! 860 free_uid(ep->user);
910 !! 861 wakeup_source_unregister(ep->ws);
911 static long ep_eventpoll_ioctl(struct file *fi !! 862 kfree(ep);
912 unsigned long a <<
913 { <<
914 int ret; <<
915 <<
916 if (!is_file_epoll(file)) <<
917 return -EINVAL; <<
918 <<
919 switch (cmd) { <<
920 case EPIOCSPARAMS: <<
921 case EPIOCGPARAMS: <<
922 ret = ep_eventpoll_bp_ioctl(fi <<
923 break; <<
924 default: <<
925 ret = -EINVAL; <<
926 break; <<
927 } <<
928 <<
929 return ret; <<
930 } 863 }
931 864
932 static int ep_eventpoll_release(struct inode * 865 static int ep_eventpoll_release(struct inode *inode, struct file *file)
933 { 866 {
934 struct eventpoll *ep = file->private_d 867 struct eventpoll *ep = file->private_data;
935 868
936 if (ep) 869 if (ep)
937 ep_clear_and_put(ep); !! 870 ep_free(ep);
938 871
939 return 0; 872 return 0;
940 } 873 }
941 874
942 static __poll_t ep_item_poll(const struct epit !! 875 static __poll_t ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
>> 876 void *priv);
>> 877 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
>> 878 poll_table *pt);
943 879
944 static __poll_t __ep_eventpoll_poll(struct fil !! 880 /*
>> 881 * Differs from ep_eventpoll_poll() in that internal callers already have
>> 882 * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
>> 883 * is correctly annotated.
>> 884 */
>> 885 static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt,
>> 886 int depth)
>> 887 {
>> 888 struct eventpoll *ep;
>> 889 bool locked;
>> 890
>> 891 pt->_key = epi->event.events;
>> 892 if (!is_file_epoll(epi->ffd.file))
>> 893 return vfs_poll(epi->ffd.file, pt) & epi->event.events;
>> 894
>> 895 ep = epi->ffd.file->private_data;
>> 896 poll_wait(epi->ffd.file, &ep->poll_wait, pt);
>> 897 locked = pt && (pt->_qproc == ep_ptable_queue_proc);
>> 898
>> 899 return ep_scan_ready_list(epi->ffd.file->private_data,
>> 900 ep_read_events_proc, &depth, depth,
>> 901 locked) & epi->event.events;
>> 902 }
>> 903
>> 904 static __poll_t ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
>> 905 void *priv)
945 { 906 {
946 struct eventpoll *ep = file->private_d <<
947 LIST_HEAD(txlist); <<
948 struct epitem *epi, *tmp; 907 struct epitem *epi, *tmp;
949 poll_table pt; 908 poll_table pt;
950 __poll_t res = 0; !! 909 int depth = *(int *)priv;
951 910
952 init_poll_funcptr(&pt, NULL); 911 init_poll_funcptr(&pt, NULL);
>> 912 depth++;
953 913
954 /* Insert inside our poll wait queue * !! 914 list_for_each_entry_safe(epi, tmp, head, rdllink) {
955 poll_wait(file, &ep->poll_wait, wait); !! 915 if (ep_item_poll(epi, &pt, depth)) {
956 !! 916 return EPOLLIN | EPOLLRDNORM;
957 /* <<
958 * Proceed to find out if wanted event <<
959 * the ready list. <<
960 */ <<
961 mutex_lock_nested(&ep->mtx, depth); <<
962 ep_start_scan(ep, &txlist); <<
963 list_for_each_entry_safe(epi, tmp, &tx <<
964 if (ep_item_poll(epi, &pt, dep <<
965 res = EPOLLIN | EPOLLR <<
966 break; <<
967 } else { 917 } else {
968 /* 918 /*
969 * Item has been dropp 919 * Item has been dropped into the ready list by the poll
970 * callback, but it's 920 * callback, but it's not actually ready, as far as
971 * caller requested ev 921 * caller requested events goes. We can remove it here.
972 */ 922 */
973 __pm_relax(ep_wakeup_s 923 __pm_relax(ep_wakeup_source(epi));
974 list_del_init(&epi->rd 924 list_del_init(&epi->rdllink);
975 } 925 }
976 } 926 }
977 ep_done_scan(ep, &txlist); <<
978 mutex_unlock(&ep->mtx); <<
979 return res; <<
980 } <<
981 927
982 /* !! 928 return 0;
983 * The ffd.file pointer may be in the process <<
984 * being closed, but we may not have finished <<
985 * <<
986 * Normally, even with the atomic_long_inc_not <<
987 * been free'd and then gotten re-allocated to <<
988 * files are not RCU-delayed, they are SLAB_TY <<
989 * <<
990 * But for epoll, users hold the ep->mtx mutex <<
991 * the process of being free'd will block in e <<
992 * and thus the underlying file allocation wil <<
993 * file re-use cannot happen. <<
994 * <<
995 * For the same reason we can avoid a rcu_read <<
996 * operation - 'ffd.file' cannot go away even <<
997 * reached zero (but we must still not call ou <<
998 * etc). <<
999 */ <<
1000 static struct file *epi_fget(const struct epi <<
1001 { <<
1002 struct file *file; <<
1003 <<
1004 file = epi->ffd.file; <<
1005 if (!atomic_long_inc_not_zero(&file-> <<
1006 file = NULL; <<
1007 return file; <<
1008 } 929 }
1009 930
1010 /* !! 931 static __poll_t ep_eventpoll_poll(struct file *file, poll_table *wait)
1011 * Differs from ep_eventpoll_poll() in that i <<
1012 * the ep->mtx so we need to start from depth <<
1013 * is correctly annotated. <<
1014 */ <<
1015 static __poll_t ep_item_poll(const struct epi <<
1016 int depth) <<
1017 { 932 {
1018 struct file *file = epi_fget(epi); !! 933 struct eventpoll *ep = file->private_data;
1019 __poll_t res; !! 934 int depth = 0;
>> 935
>> 936 /* Insert inside our poll wait queue */
>> 937 poll_wait(file, &ep->poll_wait, wait);
1020 938
1021 /* 939 /*
1022 * We could return EPOLLERR | EPOLLHU !! 940 * Proceed to find out if wanted events are really available inside
1023 * treat this more as "file doesn't e !! 941 * the ready list.
1024 */ 942 */
1025 if (!file) !! 943 return ep_scan_ready_list(ep, ep_read_events_proc,
1026 return 0; !! 944 &depth, depth, false);
1027 <<
1028 pt->_key = epi->event.events; <<
1029 if (!is_file_epoll(file)) <<
1030 res = vfs_poll(file, pt); <<
1031 else <<
1032 res = __ep_eventpoll_poll(fil <<
1033 fput(file); <<
1034 return res & epi->event.events; <<
1035 } <<
1036 <<
1037 static __poll_t ep_eventpoll_poll(struct file <<
1038 { <<
1039 return __ep_eventpoll_poll(file, wait <<
1040 } 945 }
1041 946
1042 #ifdef CONFIG_PROC_FS 947 #ifdef CONFIG_PROC_FS
1043 static void ep_show_fdinfo(struct seq_file *m 948 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
1044 { 949 {
1045 struct eventpoll *ep = f->private_dat 950 struct eventpoll *ep = f->private_data;
1046 struct rb_node *rbp; 951 struct rb_node *rbp;
1047 952
1048 mutex_lock(&ep->mtx); 953 mutex_lock(&ep->mtx);
1049 for (rbp = rb_first_cached(&ep->rbr); 954 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1050 struct epitem *epi = rb_entry 955 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
1051 struct inode *inode = file_in 956 struct inode *inode = file_inode(epi->ffd.file);
1052 957
1053 seq_printf(m, "tfd: %8d event 958 seq_printf(m, "tfd: %8d events: %8x data: %16llx "
1054 " pos:%lli ino:%lx 959 " pos:%lli ino:%lx sdev:%x\n",
1055 epi->ffd.fd, epi-> 960 epi->ffd.fd, epi->event.events,
1056 (long long)epi->ev 961 (long long)epi->event.data,
1057 (long long)epi->ff 962 (long long)epi->ffd.file->f_pos,
1058 inode->i_ino, inod 963 inode->i_ino, inode->i_sb->s_dev);
1059 if (seq_has_overflowed(m)) 964 if (seq_has_overflowed(m))
1060 break; 965 break;
1061 } 966 }
1062 mutex_unlock(&ep->mtx); 967 mutex_unlock(&ep->mtx);
1063 } 968 }
1064 #endif 969 #endif
1065 970
1066 /* File callbacks that implement the eventpol 971 /* File callbacks that implement the eventpoll file behaviour */
1067 static const struct file_operations eventpoll 972 static const struct file_operations eventpoll_fops = {
1068 #ifdef CONFIG_PROC_FS 973 #ifdef CONFIG_PROC_FS
1069 .show_fdinfo = ep_show_fdinfo, 974 .show_fdinfo = ep_show_fdinfo,
1070 #endif 975 #endif
1071 .release = ep_eventpoll_releas 976 .release = ep_eventpoll_release,
1072 .poll = ep_eventpoll_poll, 977 .poll = ep_eventpoll_poll,
1073 .llseek = noop_llseek, 978 .llseek = noop_llseek,
1074 .unlocked_ioctl = ep_eventpoll_ioctl, <<
1075 .compat_ioctl = compat_ptr_ioctl, <<
1076 }; 979 };
1077 980
1078 /* 981 /*
1079 * This is called from eventpoll_release() to 982 * This is called from eventpoll_release() to unlink files from the eventpoll
1080 * interface. We need to have this facility t 983 * interface. We need to have this facility to cleanup correctly files that are
1081 * closed without being removed from the even 984 * closed without being removed from the eventpoll interface.
1082 */ 985 */
1083 void eventpoll_release_file(struct file *file 986 void eventpoll_release_file(struct file *file)
1084 { 987 {
1085 struct eventpoll *ep; 988 struct eventpoll *ep;
1086 struct epitem *epi; !! 989 struct epitem *epi, *next;
1087 bool dispose; <<
1088 990
1089 /* 991 /*
1090 * Use the 'dying' flag to prevent a !! 992 * We don't want to get "file->f_lock" because it is not
1091 * touching the epitems list before e !! 993 * necessary. It is not necessary because we're in the "struct file"
1092 * the ep->mtx. !! 994 * cleanup path, and this means that no one is using this file anymore.
>> 995 * So, for example, epoll_ctl() cannot hit here since if we reach this
>> 996 * point, the file counter already went to zero and fget() would fail.
>> 997 * The only hit might come from ep_free() but by holding the mutex
>> 998 * will correctly serialize the operation. We do need to acquire
>> 999 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
>> 1000 * from anywhere but ep_free().
>> 1001 *
>> 1002 * Besides, ep_remove() acquires the lock, so we can't hold it here.
1093 */ 1003 */
1094 again: !! 1004 mutex_lock(&epmutex);
1095 spin_lock(&file->f_lock); !! 1005 list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
1096 if (file->f_ep && file->f_ep->first) <<
1097 epi = hlist_entry(file->f_ep- <<
1098 epi->dying = true; <<
1099 spin_unlock(&file->f_lock); <<
1100 <<
1101 /* <<
1102 * ep access is safe as we st <<
1103 * struct <<
1104 */ <<
1105 ep = epi->ep; 1006 ep = epi->ep;
1106 mutex_lock(&ep->mtx); !! 1007 mutex_lock_nested(&ep->mtx, 0);
1107 dispose = __ep_remove(ep, epi !! 1008 ep_remove(ep, epi);
1108 mutex_unlock(&ep->mtx); 1009 mutex_unlock(&ep->mtx);
1109 <<
1110 if (dispose) <<
1111 ep_free(ep); <<
1112 goto again; <<
1113 } 1010 }
1114 spin_unlock(&file->f_lock); !! 1011 mutex_unlock(&epmutex);
1115 } 1012 }
1116 1013
1117 static int ep_alloc(struct eventpoll **pep) 1014 static int ep_alloc(struct eventpoll **pep)
1118 { 1015 {
>> 1016 int error;
>> 1017 struct user_struct *user;
1119 struct eventpoll *ep; 1018 struct eventpoll *ep;
1120 1019
>> 1020 user = get_current_user();
>> 1021 error = -ENOMEM;
1121 ep = kzalloc(sizeof(*ep), GFP_KERNEL) 1022 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
1122 if (unlikely(!ep)) 1023 if (unlikely(!ep))
1123 return -ENOMEM; !! 1024 goto free_uid;
1124 1025
1125 mutex_init(&ep->mtx); 1026 mutex_init(&ep->mtx);
1126 rwlock_init(&ep->lock); 1027 rwlock_init(&ep->lock);
1127 init_waitqueue_head(&ep->wq); 1028 init_waitqueue_head(&ep->wq);
1128 init_waitqueue_head(&ep->poll_wait); 1029 init_waitqueue_head(&ep->poll_wait);
1129 INIT_LIST_HEAD(&ep->rdllist); 1030 INIT_LIST_HEAD(&ep->rdllist);
1130 ep->rbr = RB_ROOT_CACHED; 1031 ep->rbr = RB_ROOT_CACHED;
1131 ep->ovflist = EP_UNACTIVE_PTR; 1032 ep->ovflist = EP_UNACTIVE_PTR;
1132 ep->user = get_current_user(); !! 1033 ep->user = user;
1133 refcount_set(&ep->refcount, 1); <<
1134 1034
1135 *pep = ep; 1035 *pep = ep;
1136 1036
1137 return 0; 1037 return 0;
>> 1038
>> 1039 free_uid:
>> 1040 free_uid(user);
>> 1041 return error;
1138 } 1042 }
1139 1043
1140 /* 1044 /*
1141 * Search the file inside the eventpoll tree. 1045 * Search the file inside the eventpoll tree. The RB tree operations
1142 * are protected by the "mtx" mutex, and ep_f 1046 * are protected by the "mtx" mutex, and ep_find() must be called with
1143 * "mtx" held. 1047 * "mtx" held.
1144 */ 1048 */
1145 static struct epitem *ep_find(struct eventpol 1049 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
1146 { 1050 {
1147 int kcmp; 1051 int kcmp;
1148 struct rb_node *rbp; 1052 struct rb_node *rbp;
1149 struct epitem *epi, *epir = NULL; 1053 struct epitem *epi, *epir = NULL;
1150 struct epoll_filefd ffd; 1054 struct epoll_filefd ffd;
1151 1055
1152 ep_set_ffd(&ffd, file, fd); 1056 ep_set_ffd(&ffd, file, fd);
1153 for (rbp = ep->rbr.rb_root.rb_node; r 1057 for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {
1154 epi = rb_entry(rbp, struct ep 1058 epi = rb_entry(rbp, struct epitem, rbn);
1155 kcmp = ep_cmp_ffd(&ffd, &epi- 1059 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
1156 if (kcmp > 0) 1060 if (kcmp > 0)
1157 rbp = rbp->rb_right; 1061 rbp = rbp->rb_right;
1158 else if (kcmp < 0) 1062 else if (kcmp < 0)
1159 rbp = rbp->rb_left; 1063 rbp = rbp->rb_left;
1160 else { 1064 else {
1161 epir = epi; 1065 epir = epi;
1162 break; 1066 break;
1163 } 1067 }
1164 } 1068 }
1165 1069
1166 return epir; 1070 return epir;
1167 } 1071 }
1168 1072
1169 #ifdef CONFIG_KCMP 1073 #ifdef CONFIG_KCMP
1170 static struct epitem *ep_find_tfd(struct even 1074 static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
1171 { 1075 {
1172 struct rb_node *rbp; 1076 struct rb_node *rbp;
1173 struct epitem *epi; 1077 struct epitem *epi;
1174 1078
1175 for (rbp = rb_first_cached(&ep->rbr); 1079 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1176 epi = rb_entry(rbp, struct ep 1080 epi = rb_entry(rbp, struct epitem, rbn);
1177 if (epi->ffd.fd == tfd) { 1081 if (epi->ffd.fd == tfd) {
1178 if (toff == 0) 1082 if (toff == 0)
1179 return epi; 1083 return epi;
1180 else 1084 else
1181 toff--; 1085 toff--;
1182 } 1086 }
1183 cond_resched(); 1087 cond_resched();
1184 } 1088 }
1185 1089
1186 return NULL; 1090 return NULL;
1187 } 1091 }
1188 1092
1189 struct file *get_epoll_tfile_raw_ptr(struct f 1093 struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
1190 unsigned 1094 unsigned long toff)
1191 { 1095 {
1192 struct file *file_raw; 1096 struct file *file_raw;
1193 struct eventpoll *ep; 1097 struct eventpoll *ep;
1194 struct epitem *epi; 1098 struct epitem *epi;
1195 1099
1196 if (!is_file_epoll(file)) 1100 if (!is_file_epoll(file))
1197 return ERR_PTR(-EINVAL); 1101 return ERR_PTR(-EINVAL);
1198 1102
1199 ep = file->private_data; 1103 ep = file->private_data;
1200 1104
1201 mutex_lock(&ep->mtx); 1105 mutex_lock(&ep->mtx);
1202 epi = ep_find_tfd(ep, tfd, toff); 1106 epi = ep_find_tfd(ep, tfd, toff);
1203 if (epi) 1107 if (epi)
1204 file_raw = epi->ffd.file; 1108 file_raw = epi->ffd.file;
1205 else 1109 else
1206 file_raw = ERR_PTR(-ENOENT); 1110 file_raw = ERR_PTR(-ENOENT);
1207 mutex_unlock(&ep->mtx); 1111 mutex_unlock(&ep->mtx);
1208 1112
1209 return file_raw; 1113 return file_raw;
1210 } 1114 }
1211 #endif /* CONFIG_KCMP */ 1115 #endif /* CONFIG_KCMP */
1212 1116
1213 /* !! 1117 /**
1214 * Adds a new entry to the tail of the list i 1118 * Adds a new entry to the tail of the list in a lockless way, i.e.
1215 * multiple CPUs are allowed to call this fun 1119 * multiple CPUs are allowed to call this function concurrently.
1216 * 1120 *
1217 * Beware: it is necessary to prevent any oth 1121 * Beware: it is necessary to prevent any other modifications of the
1218 * existing list until all changes ar 1122 * existing list until all changes are completed, in other words
1219 * concurrent list_add_tail_lockless( 1123 * concurrent list_add_tail_lockless() calls should be protected
1220 * with a read lock, where write lock 1124 * with a read lock, where write lock acts as a barrier which
1221 * makes sure all list_add_tail_lockl 1125 * makes sure all list_add_tail_lockless() calls are fully
1222 * completed. 1126 * completed.
1223 * 1127 *
1224 * Also an element can be locklessly a 1128 * Also an element can be locklessly added to the list only in one
1225 * direction i.e. either to the tail o !! 1129 * direction i.e. either to the tail either to the head, otherwise
1226 * concurrent access will corrupt the 1130 * concurrent access will corrupt the list.
1227 * 1131 *
1228 * Return: %false if element has been already !! 1132 * Returns %false if element has been already added to the list, %true
1229 * otherwise. 1133 * otherwise.
1230 */ 1134 */
1231 static inline bool list_add_tail_lockless(str 1135 static inline bool list_add_tail_lockless(struct list_head *new,
1232 str 1136 struct list_head *head)
1233 { 1137 {
1234 struct list_head *prev; 1138 struct list_head *prev;
1235 1139
1236 /* 1140 /*
1237 * This is simple 'new->next = head' 1141 * This is simple 'new->next = head' operation, but cmpxchg()
1238 * is used in order to detect that sa 1142 * is used in order to detect that same element has been just
1239 * added to the list from another CPU 1143 * added to the list from another CPU: the winner observes
1240 * new->next == new. 1144 * new->next == new.
1241 */ 1145 */
1242 if (!try_cmpxchg(&new->next, &new, he !! 1146 if (cmpxchg(&new->next, new, head) != new)
1243 return false; 1147 return false;
1244 1148
1245 /* 1149 /*
1246 * Initially ->next of a new element 1150 * Initially ->next of a new element must be updated with the head
1247 * (we are inserting to the tail) and 1151 * (we are inserting to the tail) and only then pointers are atomically
1248 * exchanged. XCHG guarantees memory 1152 * exchanged. XCHG guarantees memory ordering, thus ->next should be
1249 * updated before pointers are actual 1153 * updated before pointers are actually swapped and pointers are
1250 * swapped before prev->next is updat 1154 * swapped before prev->next is updated.
1251 */ 1155 */
1252 1156
1253 prev = xchg(&head->prev, new); 1157 prev = xchg(&head->prev, new);
1254 1158
1255 /* 1159 /*
1256 * It is safe to modify prev->next an 1160 * It is safe to modify prev->next and new->prev, because a new element
1257 * is added only to the tail and new- 1161 * is added only to the tail and new->next is updated before XCHG.
1258 */ 1162 */
1259 1163
1260 prev->next = new; 1164 prev->next = new;
1261 new->prev = prev; 1165 new->prev = prev;
1262 1166
1263 return true; 1167 return true;
1264 } 1168 }
1265 1169
1266 /* !! 1170 /**
1267 * Chains a new epi entry to the tail of the 1171 * Chains a new epi entry to the tail of the ep->ovflist in a lockless way,
1268 * i.e. multiple CPUs are allowed to call thi 1172 * i.e. multiple CPUs are allowed to call this function concurrently.
1269 * 1173 *
1270 * Return: %false if epi element has been alr !! 1174 * Returns %false if epi element has been already chained, %true otherwise.
1271 */ 1175 */
1272 static inline bool chain_epi_lockless(struct 1176 static inline bool chain_epi_lockless(struct epitem *epi)
1273 { 1177 {
1274 struct eventpoll *ep = epi->ep; 1178 struct eventpoll *ep = epi->ep;
1275 1179
1276 /* Fast preliminary check */ 1180 /* Fast preliminary check */
1277 if (epi->next != EP_UNACTIVE_PTR) 1181 if (epi->next != EP_UNACTIVE_PTR)
1278 return false; 1182 return false;
1279 1183
1280 /* Check that the same epi has not be 1184 /* Check that the same epi has not been just chained from another CPU */
1281 if (cmpxchg(&epi->next, EP_UNACTIVE_P 1185 if (cmpxchg(&epi->next, EP_UNACTIVE_PTR, NULL) != EP_UNACTIVE_PTR)
1282 return false; 1186 return false;
1283 1187
1284 /* Atomically exchange tail */ 1188 /* Atomically exchange tail */
1285 epi->next = xchg(&ep->ovflist, epi); 1189 epi->next = xchg(&ep->ovflist, epi);
1286 1190
1287 return true; 1191 return true;
1288 } 1192 }
1289 1193
1290 /* 1194 /*
1291 * This is the callback that is passed to the 1195 * This is the callback that is passed to the wait queue wakeup
1292 * mechanism. It is called by the stored file 1196 * mechanism. It is called by the stored file descriptors when they
1293 * have events to report. 1197 * have events to report.
1294 * 1198 *
1295 * This callback takes a read lock in order n !! 1199 * This callback takes a read lock in order not to content with concurrent
1296 * events from another file descriptor, thus !! 1200 * events from another file descriptors, thus all modifications to ->rdllist
1297 * or ->ovflist are lockless. Read lock is p 1201 * or ->ovflist are lockless. Read lock is paired with the write lock from
1298 * ep_start/done_scan(), which stops all list !! 1202 * ep_scan_ready_list(), which stops all list modifications and guarantees
1299 * that lists state is seen correctly. 1203 * that lists state is seen correctly.
1300 * 1204 *
1301 * Another thing worth to mention is that ep_ 1205 * Another thing worth to mention is that ep_poll_callback() can be called
1302 * concurrently for the same @epi from differ 1206 * concurrently for the same @epi from different CPUs if poll table was inited
1303 * with several wait queues entries. Plural 1207 * with several wait queues entries. Plural wakeup from different CPUs of a
1304 * single wait queue is serialized by wq.lock 1208 * single wait queue is serialized by wq.lock, but the case when multiple wait
1305 * queues are used should be detected accordi 1209 * queues are used should be detected accordingly. This is detected using
1306 * cmpxchg() operation. 1210 * cmpxchg() operation.
1307 */ 1211 */
1308 static int ep_poll_callback(wait_queue_entry_ 1212 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1309 { 1213 {
1310 int pwake = 0; 1214 int pwake = 0;
1311 struct epitem *epi = ep_item_from_wai 1215 struct epitem *epi = ep_item_from_wait(wait);
1312 struct eventpoll *ep = epi->ep; 1216 struct eventpoll *ep = epi->ep;
1313 __poll_t pollflags = key_to_poll(key) 1217 __poll_t pollflags = key_to_poll(key);
1314 unsigned long flags; 1218 unsigned long flags;
1315 int ewake = 0; 1219 int ewake = 0;
1316 1220
1317 read_lock_irqsave(&ep->lock, flags); 1221 read_lock_irqsave(&ep->lock, flags);
1318 1222
1319 ep_set_busy_poll_napi_id(epi); 1223 ep_set_busy_poll_napi_id(epi);
1320 1224
1321 /* 1225 /*
1322 * If the event mask does not contain 1226 * If the event mask does not contain any poll(2) event, we consider the
1323 * descriptor to be disabled. This co 1227 * descriptor to be disabled. This condition is likely the effect of the
1324 * EPOLLONESHOT bit that disables the 1228 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1325 * until the next EPOLL_CTL_MOD will 1229 * until the next EPOLL_CTL_MOD will be issued.
1326 */ 1230 */
1327 if (!(epi->event.events & ~EP_PRIVATE 1231 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1328 goto out_unlock; 1232 goto out_unlock;
1329 1233
1330 /* 1234 /*
1331 * Check the events coming with the c 1235 * Check the events coming with the callback. At this stage, not
1332 * every device reports the events in 1236 * every device reports the events in the "key" parameter of the
1333 * callback. We need to be able to ha 1237 * callback. We need to be able to handle both cases here, hence the
1334 * test for "key" != NULL before the 1238 * test for "key" != NULL before the event match test.
1335 */ 1239 */
1336 if (pollflags && !(pollflags & epi->e 1240 if (pollflags && !(pollflags & epi->event.events))
1337 goto out_unlock; 1241 goto out_unlock;
1338 1242
1339 /* 1243 /*
1340 * If we are transferring events to u 1244 * If we are transferring events to userspace, we can hold no locks
1341 * (because we're accessing user memo 1245 * (because we're accessing user memory, and because of linux f_op->poll()
1342 * semantics). All the events that ha 1246 * semantics). All the events that happen during that period of time are
1343 * chained in ep->ovflist and requeue 1247 * chained in ep->ovflist and requeued later on.
1344 */ 1248 */
1345 if (READ_ONCE(ep->ovflist) != EP_UNAC 1249 if (READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR) {
1346 if (chain_epi_lockless(epi)) 1250 if (chain_epi_lockless(epi))
1347 ep_pm_stay_awake_rcu( 1251 ep_pm_stay_awake_rcu(epi);
1348 } else if (!ep_is_linked(epi)) { 1252 } else if (!ep_is_linked(epi)) {
1349 /* In the usual case, add eve 1253 /* In the usual case, add event to ready list. */
1350 if (list_add_tail_lockless(&e 1254 if (list_add_tail_lockless(&epi->rdllink, &ep->rdllist))
1351 ep_pm_stay_awake_rcu( 1255 ep_pm_stay_awake_rcu(epi);
1352 } 1256 }
1353 1257
1354 /* 1258 /*
1355 * Wake up ( if active ) both the eve 1259 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1356 * wait list. 1260 * wait list.
1357 */ 1261 */
1358 if (waitqueue_active(&ep->wq)) { 1262 if (waitqueue_active(&ep->wq)) {
1359 if ((epi->event.events & EPOL 1263 if ((epi->event.events & EPOLLEXCLUSIVE) &&
1360 !(pol 1264 !(pollflags & POLLFREE)) {
1361 switch (pollflags & E 1265 switch (pollflags & EPOLLINOUT_BITS) {
1362 case EPOLLIN: 1266 case EPOLLIN:
1363 if (epi->even 1267 if (epi->event.events & EPOLLIN)
1364 ewake 1268 ewake = 1;
1365 break; 1269 break;
1366 case EPOLLOUT: 1270 case EPOLLOUT:
1367 if (epi->even 1271 if (epi->event.events & EPOLLOUT)
1368 ewake 1272 ewake = 1;
1369 break; 1273 break;
1370 case 0: 1274 case 0:
1371 ewake = 1; 1275 ewake = 1;
1372 break; 1276 break;
1373 } 1277 }
1374 } 1278 }
1375 wake_up(&ep->wq); 1279 wake_up(&ep->wq);
1376 } 1280 }
1377 if (waitqueue_active(&ep->poll_wait)) 1281 if (waitqueue_active(&ep->poll_wait))
1378 pwake++; 1282 pwake++;
1379 1283
1380 out_unlock: 1284 out_unlock:
1381 read_unlock_irqrestore(&ep->lock, fla 1285 read_unlock_irqrestore(&ep->lock, flags);
1382 1286
1383 /* We have to call this outside the l 1287 /* We have to call this outside the lock */
1384 if (pwake) 1288 if (pwake)
1385 ep_poll_safewake(ep, epi, pol 1289 ep_poll_safewake(ep, epi, pollflags & EPOLL_URING_WAKE);
1386 1290
1387 if (!(epi->event.events & EPOLLEXCLUS 1291 if (!(epi->event.events & EPOLLEXCLUSIVE))
1388 ewake = 1; 1292 ewake = 1;
1389 1293
1390 if (pollflags & POLLFREE) { 1294 if (pollflags & POLLFREE) {
1391 /* 1295 /*
1392 * If we race with ep_remove_ 1296 * If we race with ep_remove_wait_queue() it can miss
1393 * ->whead = NULL and do anot 1297 * ->whead = NULL and do another remove_wait_queue() after
1394 * us, so we can't use __remo 1298 * us, so we can't use __remove_wait_queue().
1395 */ 1299 */
1396 list_del_init(&wait->entry); 1300 list_del_init(&wait->entry);
1397 /* 1301 /*
1398 * ->whead != NULL protects u !! 1302 * ->whead != NULL protects us from the race with ep_free()
1399 * ep_clear_and_put() or ep_r !! 1303 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1400 * takes whead->lock held by !! 1304 * held by the caller. Once we nullify it, nothing protects
1401 * nothing protects ep/epi or !! 1305 * ep/epi or even wait.
1402 */ 1306 */
1403 smp_store_release(&ep_pwq_fro 1307 smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1404 } 1308 }
1405 1309
1406 return ewake; 1310 return ewake;
1407 } 1311 }
1408 1312
1409 /* 1313 /*
1410 * This is the callback that is used to add o 1314 * This is the callback that is used to add our wait queue to the
1411 * target file wakeup lists. 1315 * target file wakeup lists.
1412 */ 1316 */
1413 static void ep_ptable_queue_proc(struct file 1317 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1414 poll_table * 1318 poll_table *pt)
1415 { 1319 {
1416 struct ep_pqueue *epq = container_of( !! 1320 struct epitem *epi = ep_item_from_epqueue(pt);
1417 struct epitem *epi = epq->epi; <<
1418 struct eppoll_entry *pwq; 1321 struct eppoll_entry *pwq;
1419 1322
1420 if (unlikely(!epi)) // an earlier !! 1323 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1421 return; !! 1324 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1422 !! 1325 pwq->whead = whead;
1423 pwq = kmem_cache_alloc(pwq_cache, GFP !! 1326 pwq->base = epi;
1424 if (unlikely(!pwq)) { !! 1327 if (epi->event.events & EPOLLEXCLUSIVE)
1425 epq->epi = NULL; !! 1328 add_wait_queue_exclusive(whead, &pwq->wait);
1426 return; !! 1329 else
>> 1330 add_wait_queue(whead, &pwq->wait);
>> 1331 list_add_tail(&pwq->llink, &epi->pwqlist);
>> 1332 epi->nwait++;
>> 1333 } else {
>> 1334 /* We have to signal that an error occurred */
>> 1335 epi->nwait = -1;
1427 } 1336 }
1428 <<
1429 init_waitqueue_func_entry(&pwq->wait, <<
1430 pwq->whead = whead; <<
1431 pwq->base = epi; <<
1432 if (epi->event.events & EPOLLEXCLUSIV <<
1433 add_wait_queue_exclusive(whea <<
1434 else <<
1435 add_wait_queue(whead, &pwq->w <<
1436 pwq->next = epi->pwqlist; <<
1437 epi->pwqlist = pwq; <<
1438 } 1337 }
1439 1338
1440 static void ep_rbtree_insert(struct eventpoll 1339 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1441 { 1340 {
1442 int kcmp; 1341 int kcmp;
1443 struct rb_node **p = &ep->rbr.rb_root 1342 struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;
1444 struct epitem *epic; 1343 struct epitem *epic;
1445 bool leftmost = true; 1344 bool leftmost = true;
1446 1345
1447 while (*p) { 1346 while (*p) {
1448 parent = *p; 1347 parent = *p;
1449 epic = rb_entry(parent, struc 1348 epic = rb_entry(parent, struct epitem, rbn);
1450 kcmp = ep_cmp_ffd(&epi->ffd, 1349 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1451 if (kcmp > 0) { 1350 if (kcmp > 0) {
1452 p = &parent->rb_right 1351 p = &parent->rb_right;
1453 leftmost = false; 1352 leftmost = false;
1454 } else 1353 } else
1455 p = &parent->rb_left; 1354 p = &parent->rb_left;
1456 } 1355 }
1457 rb_link_node(&epi->rbn, parent, p); 1356 rb_link_node(&epi->rbn, parent, p);
1458 rb_insert_color_cached(&epi->rbn, &ep 1357 rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);
1459 } 1358 }
1460 1359
1461 1360
1462 1361
1463 #define PATH_ARR_SIZE 5 1362 #define PATH_ARR_SIZE 5
1464 /* 1363 /*
1465 * These are the number paths of length 1 to 1364 * These are the number paths of length 1 to 5, that we are allowing to emanate
1466 * from a single file of interest. For exampl 1365 * from a single file of interest. For example, we allow 1000 paths of length
1467 * 1, to emanate from each file of interest. 1366 * 1, to emanate from each file of interest. This essentially represents the
1468 * potential wakeup paths, which need to be l 1367 * potential wakeup paths, which need to be limited in order to avoid massive
1469 * uncontrolled wakeup storms. The common use 1368 * uncontrolled wakeup storms. The common use case should be a single ep which
1470 * is connected to n file sources. In this ca 1369 * is connected to n file sources. In this case each file source has 1 path
1471 * of length 1. Thus, the numbers below shoul 1370 * of length 1. Thus, the numbers below should be more than sufficient. These
1472 * path limits are enforced during an EPOLL_C 1371 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1473 * and delete can't add additional paths. Pro !! 1372 * and delete can't add additional paths. Protected by the epmutex.
1474 */ 1373 */
1475 static const int path_limits[PATH_ARR_SIZE] = 1374 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1476 static int path_count[PATH_ARR_SIZE]; 1375 static int path_count[PATH_ARR_SIZE];
1477 1376
1478 static int path_count_inc(int nests) 1377 static int path_count_inc(int nests)
1479 { 1378 {
1480 /* Allow an arbitrary number of depth 1379 /* Allow an arbitrary number of depth 1 paths */
1481 if (nests == 0) 1380 if (nests == 0)
1482 return 0; 1381 return 0;
1483 1382
1484 if (++path_count[nests] > path_limits 1383 if (++path_count[nests] > path_limits[nests])
1485 return -1; 1384 return -1;
1486 return 0; 1385 return 0;
1487 } 1386 }
1488 1387
1489 static void path_count_init(void) 1388 static void path_count_init(void)
1490 { 1389 {
1491 int i; 1390 int i;
1492 1391
1493 for (i = 0; i < PATH_ARR_SIZE; i++) 1392 for (i = 0; i < PATH_ARR_SIZE; i++)
1494 path_count[i] = 0; 1393 path_count[i] = 0;
1495 } 1394 }
1496 1395
1497 static int reverse_path_check_proc(struct hli !! 1396 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1498 { 1397 {
1499 int error = 0; 1398 int error = 0;
>> 1399 struct file *file = priv;
>> 1400 struct file *child_file;
1500 struct epitem *epi; 1401 struct epitem *epi;
1501 1402
1502 if (depth > EP_MAX_NESTS) /* too deep <<
1503 return -1; <<
1504 <<
1505 /* CTL_DEL can remove links here, but 1403 /* CTL_DEL can remove links here, but that can't increase our count */
1506 hlist_for_each_entry_rcu(epi, refs, f !! 1404 rcu_read_lock();
1507 struct hlist_head *refs = &ep !! 1405 list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1508 if (hlist_empty(refs)) !! 1406 child_file = epi->ep->file;
1509 error = path_count_in !! 1407 if (is_file_epoll(child_file)) {
1510 else !! 1408 if (list_empty(&child_file->f_ep_links)) {
1511 error = reverse_path_ !! 1409 if (path_count_inc(call_nests)) {
1512 if (error != 0) !! 1410 error = -1;
1513 break; !! 1411 break;
>> 1412 }
>> 1413 } else {
>> 1414 error = ep_call_nested(&poll_loop_ncalls,
>> 1415 reverse_path_check_proc,
>> 1416 child_file, child_file,
>> 1417 current);
>> 1418 }
>> 1419 if (error != 0)
>> 1420 break;
>> 1421 } else {
>> 1422 printk(KERN_ERR "reverse_path_check_proc: "
>> 1423 "file is not an ep!\n");
>> 1424 }
1514 } 1425 }
>> 1426 rcu_read_unlock();
1515 return error; 1427 return error;
1516 } 1428 }
1517 1429
1518 /** 1430 /**
1519 * reverse_path_check - The tfile_check_list !! 1431 * reverse_path_check - The tfile_check_list is list of file *, which have
1520 * links that are propos 1432 * links that are proposed to be newly added. We need to
1521 * make sure that those 1433 * make sure that those added links don't add too many
1522 * paths such that we wi 1434 * paths such that we will spend all our time waking up
1523 * eventpoll objects. 1435 * eventpoll objects.
1524 * 1436 *
1525 * Return: %zero if the proposed links don't !! 1437 * Returns: Returns zero if the proposed links don't create too many paths,
1526 * %-1 otherwise. !! 1438 * -1 otherwise.
1527 */ 1439 */
1528 static int reverse_path_check(void) 1440 static int reverse_path_check(void)
1529 { 1441 {
1530 struct epitems_head *p; !! 1442 int error = 0;
>> 1443 struct file *current_file;
1531 1444
1532 for (p = tfile_check_list; p != EP_UN !! 1445 /* let's call this for all tfiles */
1533 int error; !! 1446 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1534 path_count_init(); 1447 path_count_init();
1535 rcu_read_lock(); !! 1448 error = ep_call_nested(&poll_loop_ncalls,
1536 error = reverse_path_check_pr !! 1449 reverse_path_check_proc, current_file,
1537 rcu_read_unlock(); !! 1450 current_file, current);
1538 if (error) 1451 if (error)
1539 return error; !! 1452 break;
1540 } 1453 }
1541 return 0; !! 1454 return error;
1542 } 1455 }
1543 1456
1544 static int ep_create_wakeup_source(struct epi 1457 static int ep_create_wakeup_source(struct epitem *epi)
1545 { 1458 {
1546 struct name_snapshot n; 1459 struct name_snapshot n;
1547 struct wakeup_source *ws; 1460 struct wakeup_source *ws;
1548 1461
1549 if (!epi->ep->ws) { 1462 if (!epi->ep->ws) {
1550 epi->ep->ws = wakeup_source_r 1463 epi->ep->ws = wakeup_source_register(NULL, "eventpoll");
1551 if (!epi->ep->ws) 1464 if (!epi->ep->ws)
1552 return -ENOMEM; 1465 return -ENOMEM;
1553 } 1466 }
1554 1467
1555 take_dentry_name_snapshot(&n, epi->ff 1468 take_dentry_name_snapshot(&n, epi->ffd.file->f_path.dentry);
1556 ws = wakeup_source_register(NULL, n.n 1469 ws = wakeup_source_register(NULL, n.name.name);
1557 release_dentry_name_snapshot(&n); 1470 release_dentry_name_snapshot(&n);
1558 1471
1559 if (!ws) 1472 if (!ws)
1560 return -ENOMEM; 1473 return -ENOMEM;
1561 rcu_assign_pointer(epi->ws, ws); 1474 rcu_assign_pointer(epi->ws, ws);
1562 1475
1563 return 0; 1476 return 0;
1564 } 1477 }
1565 1478
1566 /* rare code path, only used when EPOLL_CTL_M 1479 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1567 static noinline void ep_destroy_wakeup_source 1480 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1568 { 1481 {
1569 struct wakeup_source *ws = ep_wakeup_ 1482 struct wakeup_source *ws = ep_wakeup_source(epi);
1570 1483
1571 RCU_INIT_POINTER(epi->ws, NULL); 1484 RCU_INIT_POINTER(epi->ws, NULL);
1572 1485
1573 /* 1486 /*
1574 * wait for ep_pm_stay_awake_rcu to f 1487 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1575 * used internally by wakeup_source_r 1488 * used internally by wakeup_source_remove, too (called by
1576 * wakeup_source_unregister), so we c 1489 * wakeup_source_unregister), so we cannot use call_rcu
1577 */ 1490 */
1578 synchronize_rcu(); 1491 synchronize_rcu();
1579 wakeup_source_unregister(ws); 1492 wakeup_source_unregister(ws);
1580 } 1493 }
1581 1494
1582 static int attach_epitem(struct file *file, s <<
1583 { <<
1584 struct epitems_head *to_free = NULL; <<
1585 struct hlist_head *head = NULL; <<
1586 struct eventpoll *ep = NULL; <<
1587 <<
1588 if (is_file_epoll(file)) <<
1589 ep = file->private_data; <<
1590 <<
1591 if (ep) { <<
1592 head = &ep->refs; <<
1593 } else if (!READ_ONCE(file->f_ep)) { <<
1594 allocate: <<
1595 to_free = kmem_cache_zalloc(e <<
1596 if (!to_free) <<
1597 return -ENOMEM; <<
1598 head = &to_free->epitems; <<
1599 } <<
1600 spin_lock(&file->f_lock); <<
1601 if (!file->f_ep) { <<
1602 if (unlikely(!head)) { <<
1603 spin_unlock(&file->f_ <<
1604 goto allocate; <<
1605 } <<
1606 file->f_ep = head; <<
1607 to_free = NULL; <<
1608 } <<
1609 hlist_add_head_rcu(&epi->fllink, file <<
1610 spin_unlock(&file->f_lock); <<
1611 free_ephead(to_free); <<
1612 return 0; <<
1613 } <<
1614 <<
1615 /* 1495 /*
1616 * Must be called with "mtx" held. 1496 * Must be called with "mtx" held.
1617 */ 1497 */
1618 static int ep_insert(struct eventpoll *ep, co 1498 static int ep_insert(struct eventpoll *ep, const struct epoll_event *event,
1619 struct file *tfile, int 1499 struct file *tfile, int fd, int full_check)
1620 { 1500 {
1621 int error, pwake = 0; 1501 int error, pwake = 0;
1622 __poll_t revents; 1502 __poll_t revents;
>> 1503 long user_watches;
1623 struct epitem *epi; 1504 struct epitem *epi;
1624 struct ep_pqueue epq; 1505 struct ep_pqueue epq;
1625 struct eventpoll *tep = NULL; <<
1626 <<
1627 if (is_file_epoll(tfile)) <<
1628 tep = tfile->private_data; <<
1629 1506
1630 lockdep_assert_irqs_enabled(); 1507 lockdep_assert_irqs_enabled();
1631 1508
1632 if (unlikely(percpu_counter_compare(& !! 1509 user_watches = atomic_long_read(&ep->user->epoll_watches);
1633 m !! 1510 if (unlikely(user_watches >= max_user_watches))
1634 return -ENOSPC; 1511 return -ENOSPC;
1635 percpu_counter_inc(&ep->user->epoll_w !! 1512 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1636 <<
1637 if (!(epi = kmem_cache_zalloc(epi_cac <<
1638 percpu_counter_dec(&ep->user- <<
1639 return -ENOMEM; 1513 return -ENOMEM;
1640 } <<
1641 1514
1642 /* Item initialization follow here .. 1515 /* Item initialization follow here ... */
1643 INIT_LIST_HEAD(&epi->rdllink); 1516 INIT_LIST_HEAD(&epi->rdllink);
>> 1517 INIT_LIST_HEAD(&epi->fllink);
>> 1518 INIT_LIST_HEAD(&epi->pwqlist);
1644 epi->ep = ep; 1519 epi->ep = ep;
1645 ep_set_ffd(&epi->ffd, tfile, fd); 1520 ep_set_ffd(&epi->ffd, tfile, fd);
1646 epi->event = *event; 1521 epi->event = *event;
>> 1522 epi->nwait = 0;
1647 epi->next = EP_UNACTIVE_PTR; 1523 epi->next = EP_UNACTIVE_PTR;
1648 !! 1524 if (epi->event.events & EPOLLWAKEUP) {
1649 if (tep) !! 1525 error = ep_create_wakeup_source(epi);
1650 mutex_lock_nested(&tep->mtx, !! 1526 if (error)
1651 /* Add the current item to the list o !! 1527 goto error_create_wakeup_source;
1652 if (unlikely(attach_epitem(tfile, epi !! 1528 } else {
1653 if (tep) !! 1529 RCU_INIT_POINTER(epi->ws, NULL);
1654 mutex_unlock(&tep->mt <<
1655 kmem_cache_free(epi_cache, ep <<
1656 percpu_counter_dec(&ep->user- <<
1657 return -ENOMEM; <<
1658 } 1530 }
1659 1531
1660 if (full_check && !tep) !! 1532 /* Add the current item to the list of active epoll hook for this file */
1661 list_file(tfile); !! 1533 spin_lock(&tfile->f_lock);
>> 1534 list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
>> 1535 spin_unlock(&tfile->f_lock);
1662 1536
1663 /* 1537 /*
1664 * Add the current item to the RB tre 1538 * Add the current item to the RB tree. All RB tree operations are
1665 * protected by "mtx", and ep_insert( 1539 * protected by "mtx", and ep_insert() is called with "mtx" held.
1666 */ 1540 */
1667 ep_rbtree_insert(ep, epi); 1541 ep_rbtree_insert(ep, epi);
1668 if (tep) <<
1669 mutex_unlock(&tep->mtx); <<
1670 <<
1671 /* <<
1672 * ep_remove_safe() calls in the late <<
1673 * ep_free() as the ep file itself st <<
1674 */ <<
1675 ep_get(ep); <<
1676 1542
1677 /* now check if we've created too man 1543 /* now check if we've created too many backpaths */
1678 if (unlikely(full_check && reverse_pa !! 1544 error = -EINVAL;
1679 ep_remove_safe(ep, epi); !! 1545 if (full_check && reverse_path_check())
1680 return -EINVAL; !! 1546 goto error_remove_epi;
1681 } <<
1682 <<
1683 if (epi->event.events & EPOLLWAKEUP) <<
1684 error = ep_create_wakeup_sour <<
1685 if (error) { <<
1686 ep_remove_safe(ep, ep <<
1687 return error; <<
1688 } <<
1689 } <<
1690 1547
1691 /* Initialize the poll table using th 1548 /* Initialize the poll table using the queue callback */
1692 epq.epi = epi; 1549 epq.epi = epi;
1693 init_poll_funcptr(&epq.pt, ep_ptable_ 1550 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1694 1551
1695 /* 1552 /*
1696 * Attach the item to the poll hooks 1553 * Attach the item to the poll hooks and get current event bits.
1697 * We can safely use the file* here b 1554 * We can safely use the file* here because its usage count has
1698 * been increased by the caller of th 1555 * been increased by the caller of this function. Note that after
1699 * this operation completes, the poll 1556 * this operation completes, the poll callback can start hitting
1700 * the new item. 1557 * the new item.
1701 */ 1558 */
1702 revents = ep_item_poll(epi, &epq.pt, 1559 revents = ep_item_poll(epi, &epq.pt, 1);
1703 1560
1704 /* 1561 /*
1705 * We have to check if something went 1562 * We have to check if something went wrong during the poll wait queue
1706 * install process. Namely an allocat 1563 * install process. Namely an allocation for a wait queue failed due
1707 * high memory pressure. 1564 * high memory pressure.
1708 */ 1565 */
1709 if (unlikely(!epq.epi)) { !! 1566 error = -ENOMEM;
1710 ep_remove_safe(ep, epi); !! 1567 if (epi->nwait < 0)
1711 return -ENOMEM; !! 1568 goto error_unregister;
1712 } <<
1713 1569
1714 /* We have to drop the new item insid 1570 /* We have to drop the new item inside our item list to keep track of it */
1715 write_lock_irq(&ep->lock); 1571 write_lock_irq(&ep->lock);
1716 1572
1717 /* record NAPI ID of new item if pres 1573 /* record NAPI ID of new item if present */
1718 ep_set_busy_poll_napi_id(epi); 1574 ep_set_busy_poll_napi_id(epi);
1719 1575
1720 /* If the file is already "ready" we 1576 /* If the file is already "ready" we drop it inside the ready list */
1721 if (revents && !ep_is_linked(epi)) { 1577 if (revents && !ep_is_linked(epi)) {
1722 list_add_tail(&epi->rdllink, 1578 list_add_tail(&epi->rdllink, &ep->rdllist);
1723 ep_pm_stay_awake(epi); 1579 ep_pm_stay_awake(epi);
1724 1580
1725 /* Notify waiting tasks that 1581 /* Notify waiting tasks that events are available */
1726 if (waitqueue_active(&ep->wq) 1582 if (waitqueue_active(&ep->wq))
1727 wake_up(&ep->wq); 1583 wake_up(&ep->wq);
1728 if (waitqueue_active(&ep->pol 1584 if (waitqueue_active(&ep->poll_wait))
1729 pwake++; 1585 pwake++;
1730 } 1586 }
1731 1587
1732 write_unlock_irq(&ep->lock); 1588 write_unlock_irq(&ep->lock);
1733 1589
>> 1590 atomic_long_inc(&ep->user->epoll_watches);
>> 1591
1734 /* We have to call this outside the l 1592 /* We have to call this outside the lock */
1735 if (pwake) 1593 if (pwake)
1736 ep_poll_safewake(ep, NULL, 0) 1594 ep_poll_safewake(ep, NULL, 0);
1737 1595
1738 return 0; 1596 return 0;
>> 1597
>> 1598 error_unregister:
>> 1599 ep_unregister_pollwait(ep, epi);
>> 1600 error_remove_epi:
>> 1601 spin_lock(&tfile->f_lock);
>> 1602 list_del_rcu(&epi->fllink);
>> 1603 spin_unlock(&tfile->f_lock);
>> 1604
>> 1605 rb_erase_cached(&epi->rbn, &ep->rbr);
>> 1606
>> 1607 /*
>> 1608 * We need to do this because an event could have been arrived on some
>> 1609 * allocated wait queue. Note that we don't care about the ep->ovflist
>> 1610 * list, since that is used/cleaned only inside a section bound by "mtx".
>> 1611 * And ep_insert() is called with "mtx" held.
>> 1612 */
>> 1613 write_lock_irq(&ep->lock);
>> 1614 if (ep_is_linked(epi))
>> 1615 list_del_init(&epi->rdllink);
>> 1616 write_unlock_irq(&ep->lock);
>> 1617
>> 1618 wakeup_source_unregister(ep_wakeup_source(epi));
>> 1619
>> 1620 error_create_wakeup_source:
>> 1621 kmem_cache_free(epi_cache, epi);
>> 1622
>> 1623 return error;
1739 } 1624 }
1740 1625
1741 /* 1626 /*
1742 * Modify the interest event mask by dropping 1627 * Modify the interest event mask by dropping an event if the new mask
1743 * has a match in the current file status. Mu 1628 * has a match in the current file status. Must be called with "mtx" held.
1744 */ 1629 */
1745 static int ep_modify(struct eventpoll *ep, st 1630 static int ep_modify(struct eventpoll *ep, struct epitem *epi,
1746 const struct epoll_event 1631 const struct epoll_event *event)
1747 { 1632 {
1748 int pwake = 0; 1633 int pwake = 0;
1749 poll_table pt; 1634 poll_table pt;
1750 1635
1751 lockdep_assert_irqs_enabled(); 1636 lockdep_assert_irqs_enabled();
1752 1637
1753 init_poll_funcptr(&pt, NULL); 1638 init_poll_funcptr(&pt, NULL);
1754 1639
1755 /* 1640 /*
1756 * Set the new event interest mask be 1641 * Set the new event interest mask before calling f_op->poll();
1757 * otherwise we might miss an event t 1642 * otherwise we might miss an event that happens between the
1758 * f_op->poll() call and the new even 1643 * f_op->poll() call and the new event set registering.
1759 */ 1644 */
1760 epi->event.events = event->events; /* 1645 epi->event.events = event->events; /* need barrier below */
1761 epi->event.data = event->data; /* pro 1646 epi->event.data = event->data; /* protected by mtx */
1762 if (epi->event.events & EPOLLWAKEUP) 1647 if (epi->event.events & EPOLLWAKEUP) {
1763 if (!ep_has_wakeup_source(epi 1648 if (!ep_has_wakeup_source(epi))
1764 ep_create_wakeup_sour 1649 ep_create_wakeup_source(epi);
1765 } else if (ep_has_wakeup_source(epi)) 1650 } else if (ep_has_wakeup_source(epi)) {
1766 ep_destroy_wakeup_source(epi) 1651 ep_destroy_wakeup_source(epi);
1767 } 1652 }
1768 1653
1769 /* 1654 /*
1770 * The following barrier has two effe 1655 * The following barrier has two effects:
1771 * 1656 *
1772 * 1) Flush epi changes above to othe 1657 * 1) Flush epi changes above to other CPUs. This ensures
1773 * we do not miss events from ep_p 1658 * we do not miss events from ep_poll_callback if an
1774 * event occurs immediately after 1659 * event occurs immediately after we call f_op->poll().
1775 * We need this because we did not 1660 * We need this because we did not take ep->lock while
1776 * changing epi above (but ep_poll 1661 * changing epi above (but ep_poll_callback does take
1777 * ep->lock). 1662 * ep->lock).
1778 * 1663 *
1779 * 2) We also need to ensure we do no 1664 * 2) We also need to ensure we do not miss _past_ events
1780 * when calling f_op->poll(). Thi 1665 * when calling f_op->poll(). This barrier also
1781 * pairs with the barrier in wq_ha 1666 * pairs with the barrier in wq_has_sleeper (see
1782 * comments for wq_has_sleeper). 1667 * comments for wq_has_sleeper).
1783 * 1668 *
1784 * This barrier will now guarantee ep 1669 * This barrier will now guarantee ep_poll_callback or f_op->poll
1785 * (or both) will notice the readines 1670 * (or both) will notice the readiness of an item.
1786 */ 1671 */
1787 smp_mb(); 1672 smp_mb();
1788 1673
1789 /* 1674 /*
1790 * Get current event bits. We can saf 1675 * Get current event bits. We can safely use the file* here because
1791 * its usage count has been increased 1676 * its usage count has been increased by the caller of this function.
1792 * If the item is "hot" and it is not 1677 * If the item is "hot" and it is not registered inside the ready
1793 * list, push it inside. 1678 * list, push it inside.
1794 */ 1679 */
1795 if (ep_item_poll(epi, &pt, 1)) { 1680 if (ep_item_poll(epi, &pt, 1)) {
1796 write_lock_irq(&ep->lock); 1681 write_lock_irq(&ep->lock);
1797 if (!ep_is_linked(epi)) { 1682 if (!ep_is_linked(epi)) {
1798 list_add_tail(&epi->r 1683 list_add_tail(&epi->rdllink, &ep->rdllist);
1799 ep_pm_stay_awake(epi) 1684 ep_pm_stay_awake(epi);
1800 1685
1801 /* Notify waiting tas 1686 /* Notify waiting tasks that events are available */
1802 if (waitqueue_active( 1687 if (waitqueue_active(&ep->wq))
1803 wake_up(&ep-> 1688 wake_up(&ep->wq);
1804 if (waitqueue_active( 1689 if (waitqueue_active(&ep->poll_wait))
1805 pwake++; 1690 pwake++;
1806 } 1691 }
1807 write_unlock_irq(&ep->lock); 1692 write_unlock_irq(&ep->lock);
1808 } 1693 }
1809 1694
1810 /* We have to call this outside the l 1695 /* We have to call this outside the lock */
1811 if (pwake) 1696 if (pwake)
1812 ep_poll_safewake(ep, NULL, 0) 1697 ep_poll_safewake(ep, NULL, 0);
1813 1698
1814 return 0; 1699 return 0;
1815 } 1700 }
1816 1701
1817 static int ep_send_events(struct eventpoll *e !! 1702 static __poll_t ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1818 struct epoll_event !! 1703 void *priv)
1819 { 1704 {
>> 1705 struct ep_send_events_data *esed = priv;
>> 1706 __poll_t revents;
1820 struct epitem *epi, *tmp; 1707 struct epitem *epi, *tmp;
1821 LIST_HEAD(txlist); !! 1708 struct epoll_event __user *uevent = esed->events;
>> 1709 struct wakeup_source *ws;
1822 poll_table pt; 1710 poll_table pt;
1823 int res = 0; <<
1824 <<
1825 /* <<
1826 * Always short-circuit for fatal sig <<
1827 * timely exit without the chance of <<
1828 * fetching repeatedly. <<
1829 */ <<
1830 if (fatal_signal_pending(current)) <<
1831 return -EINTR; <<
1832 1711
1833 init_poll_funcptr(&pt, NULL); 1712 init_poll_funcptr(&pt, NULL);
1834 !! 1713 esed->res = 0;
1835 mutex_lock(&ep->mtx); <<
1836 ep_start_scan(ep, &txlist); <<
1837 1714
1838 /* 1715 /*
1839 * We can loop without lock because w 1716 * We can loop without lock because we are passed a task private list.
1840 * Items cannot vanish during the loo !! 1717 * Items cannot vanish during the loop because ep_scan_ready_list() is
>> 1718 * holding "mtx" during this call.
1841 */ 1719 */
1842 list_for_each_entry_safe(epi, tmp, &t !! 1720 lockdep_assert_held(&ep->mtx);
1843 struct wakeup_source *ws; <<
1844 __poll_t revents; <<
1845 1721
1846 if (res >= maxevents) !! 1722 list_for_each_entry_safe(epi, tmp, head, rdllink) {
>> 1723 if (esed->res >= esed->maxevents)
1847 break; 1724 break;
1848 1725
1849 /* 1726 /*
1850 * Activate ep->ws before dea 1727 * Activate ep->ws before deactivating epi->ws to prevent
1851 * triggering auto-suspend he 1728 * triggering auto-suspend here (in case we reactive epi->ws
1852 * below). 1729 * below).
1853 * 1730 *
1854 * This could be rearranged t 1731 * This could be rearranged to delay the deactivation of epi->ws
1855 * instead, but then epi->ws 1732 * instead, but then epi->ws would temporarily be out of sync
1856 * with ep_is_linked(). 1733 * with ep_is_linked().
1857 */ 1734 */
1858 ws = ep_wakeup_source(epi); 1735 ws = ep_wakeup_source(epi);
1859 if (ws) { 1736 if (ws) {
1860 if (ws->active) 1737 if (ws->active)
1861 __pm_stay_awa 1738 __pm_stay_awake(ep->ws);
1862 __pm_relax(ws); 1739 __pm_relax(ws);
1863 } 1740 }
1864 1741
1865 list_del_init(&epi->rdllink); 1742 list_del_init(&epi->rdllink);
1866 1743
1867 /* 1744 /*
1868 * If the event mask intersec 1745 * If the event mask intersect the caller-requested one,
1869 * deliver the event to users !! 1746 * deliver the event to userspace. Again, ep_scan_ready_list()
1870 * so no operations coming fr !! 1747 * is holding ep->mtx, so no operations coming from userspace
>> 1748 * can change the item.
1871 */ 1749 */
1872 revents = ep_item_poll(epi, & 1750 revents = ep_item_poll(epi, &pt, 1);
1873 if (!revents) 1751 if (!revents)
1874 continue; 1752 continue;
1875 1753
1876 events = epoll_put_uevent(rev !! 1754 if (__put_user(revents, &uevent->events) ||
1877 if (!events) { !! 1755 __put_user(epi->event.data, &uevent->data)) {
1878 list_add(&epi->rdllin !! 1756 list_add(&epi->rdllink, head);
1879 ep_pm_stay_awake(epi) 1757 ep_pm_stay_awake(epi);
1880 if (!res) !! 1758 if (!esed->res)
1881 res = -EFAULT !! 1759 esed->res = -EFAULT;
1882 break; !! 1760 return 0;
1883 } 1761 }
1884 res++; !! 1762 esed->res++;
>> 1763 uevent++;
1885 if (epi->event.events & EPOLL 1764 if (epi->event.events & EPOLLONESHOT)
1886 epi->event.events &= 1765 epi->event.events &= EP_PRIVATE_BITS;
1887 else if (!(epi->event.events 1766 else if (!(epi->event.events & EPOLLET)) {
1888 /* 1767 /*
1889 * If this file has b 1768 * If this file has been added with Level
1890 * Trigger mode, we n 1769 * Trigger mode, we need to insert back inside
1891 * the ready list, so 1770 * the ready list, so that the next call to
1892 * epoll_wait() will 1771 * epoll_wait() will check again the events
1893 * availability. At t 1772 * availability. At this point, no one can insert
1894 * into ep->rdllist b 1773 * into ep->rdllist besides us. The epoll_ctl()
1895 * callers are locked 1774 * callers are locked out by
1896 * ep_send_events() h !! 1775 * ep_scan_ready_list() holding "mtx" and the
1897 * poll callback will 1776 * poll callback will queue them in ep->ovflist.
1898 */ 1777 */
1899 list_add_tail(&epi->r 1778 list_add_tail(&epi->rdllink, &ep->rdllist);
1900 ep_pm_stay_awake(epi) 1779 ep_pm_stay_awake(epi);
1901 } 1780 }
1902 } 1781 }
1903 ep_done_scan(ep, &txlist); <<
1904 mutex_unlock(&ep->mtx); <<
1905 1782
1906 return res; !! 1783 return 0;
1907 } 1784 }
1908 1785
1909 static struct timespec64 *ep_timeout_to_times !! 1786 static int ep_send_events(struct eventpoll *ep,
>> 1787 struct epoll_event __user *events, int maxevents)
1910 { 1788 {
1911 struct timespec64 now; !! 1789 struct ep_send_events_data esed;
1912 1790
1913 if (ms < 0) !! 1791 esed.maxevents = maxevents;
1914 return NULL; !! 1792 esed.events = events;
1915 1793
1916 if (!ms) { !! 1794 ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1917 to->tv_sec = 0; !! 1795 return esed.res;
1918 to->tv_nsec = 0; !! 1796 }
1919 return to; <<
1920 } <<
1921 1797
1922 to->tv_sec = ms / MSEC_PER_SEC; !! 1798 static inline struct timespec64 ep_set_mstimeout(long ms)
1923 to->tv_nsec = NSEC_PER_MSEC * (ms % M !! 1799 {
>> 1800 struct timespec64 now, ts = {
>> 1801 .tv_sec = ms / MSEC_PER_SEC,
>> 1802 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
>> 1803 };
1924 1804
1925 ktime_get_ts64(&now); 1805 ktime_get_ts64(&now);
1926 *to = timespec64_add_safe(now, *to); !! 1806 return timespec64_add_safe(now, ts);
1927 return to; <<
1928 } 1807 }
1929 1808
1930 /* 1809 /*
1931 * autoremove_wake_function, but remove even 1810 * autoremove_wake_function, but remove even on failure to wake up, because we
1932 * know that default_wake_function/ttwu will 1811 * know that default_wake_function/ttwu will only fail if the thread is already
1933 * woken, and in that case the ep_poll loop w 1812 * woken, and in that case the ep_poll loop will remove the entry anyways, not
1934 * try to reuse it. 1813 * try to reuse it.
1935 */ 1814 */
1936 static int ep_autoremove_wake_function(struct 1815 static int ep_autoremove_wake_function(struct wait_queue_entry *wq_entry,
1937 unsign 1816 unsigned int mode, int sync, void *key)
1938 { 1817 {
1939 int ret = default_wake_function(wq_en 1818 int ret = default_wake_function(wq_entry, mode, sync, key);
1940 1819
1941 /* 1820 /*
1942 * Pairs with list_empty_careful in e 1821 * Pairs with list_empty_careful in ep_poll, and ensures future loop
1943 * iterations see the cause of this w 1822 * iterations see the cause of this wakeup.
1944 */ 1823 */
1945 list_del_init_careful(&wq_entry->entr 1824 list_del_init_careful(&wq_entry->entry);
1946 return ret; 1825 return ret;
1947 } 1826 }
1948 1827
1949 /** 1828 /**
1950 * ep_poll - Retrieves ready events, and deli !! 1829 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1951 * event buffer. 1830 * event buffer.
1952 * 1831 *
1953 * @ep: Pointer to the eventpoll context. 1832 * @ep: Pointer to the eventpoll context.
1954 * @events: Pointer to the userspace buffer w 1833 * @events: Pointer to the userspace buffer where the ready events should be
1955 * stored. 1834 * stored.
1956 * @maxevents: Size (in terms of number of ev 1835 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1957 * @timeout: Maximum timeout for the ready ev 1836 * @timeout: Maximum timeout for the ready events fetch operation, in
1958 * timespec. If the timeout is zero !! 1837 * milliseconds. If the @timeout is zero, the function will not block,
1959 * while if the @timeout ptr is NUL !! 1838 * while if the @timeout is less than zero, the function will block
1960 * until at least one event has bee 1839 * until at least one event has been retrieved (or an error
1961 * occurred). 1840 * occurred).
1962 * 1841 *
1963 * Return: the number of ready events which h !! 1842 * Returns: Returns the number of ready events which have been fetched, or an
1964 * error code, in case of error. 1843 * error code, in case of error.
1965 */ 1844 */
1966 static int ep_poll(struct eventpoll *ep, stru 1845 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1967 int maxevents, struct time !! 1846 int maxevents, long timeout)
1968 { 1847 {
1969 int res, eavail, timed_out = 0; !! 1848 int res = 0, eavail, timed_out = 0;
1970 u64 slack = 0; 1849 u64 slack = 0;
1971 wait_queue_entry_t wait; 1850 wait_queue_entry_t wait;
1972 ktime_t expires, *to = NULL; 1851 ktime_t expires, *to = NULL;
1973 1852
1974 lockdep_assert_irqs_enabled(); 1853 lockdep_assert_irqs_enabled();
1975 1854
1976 if (timeout && (timeout->tv_sec | tim !! 1855 if (timeout > 0) {
1977 slack = select_estimate_accur !! 1856 struct timespec64 end_time = ep_set_mstimeout(timeout);
>> 1857
>> 1858 slack = select_estimate_accuracy(&end_time);
1978 to = &expires; 1859 to = &expires;
1979 *to = timespec64_to_ktime(*ti !! 1860 *to = timespec64_to_ktime(end_time);
1980 } else if (timeout) { !! 1861 } else if (timeout == 0) {
1981 /* 1862 /*
1982 * Avoid the unnecessary trip 1863 * Avoid the unnecessary trip to the wait queue loop, if the
1983 * caller specified a non blo !! 1864 * caller specified a non blocking operation. We still need
>> 1865 * lock because we could race and not see an epi being added
>> 1866 * to the ready list while in irq callback. Thus incorrectly
>> 1867 * returning 0 back to userspace.
1984 */ 1868 */
1985 timed_out = 1; 1869 timed_out = 1;
1986 } <<
1987 1870
1988 /* !! 1871 write_lock_irq(&ep->lock);
1989 * This call is racy: We may or may n !! 1872 eavail = ep_events_available(ep);
1990 * to the ready list under the lock ( !! 1873 write_unlock_irq(&ep->lock);
1991 * with a non-zero timeout, this thre <<
1992 * lock and will add to the wait queu <<
1993 * timeout, the user by definition sh <<
1994 * recheck again. <<
1995 */ <<
1996 eavail = ep_events_available(ep); <<
1997 1874
1998 while (1) { !! 1875 goto send_events;
1999 if (eavail) { !! 1876 }
2000 /* <<
2001 * Try to transfer ev <<
2002 * 0 events and there <<
2003 * trying again in se <<
2004 */ <<
2005 res = ep_send_events( <<
2006 if (res) <<
2007 return res; <<
2008 } <<
2009 1877
2010 if (timed_out) !! 1878 fetch_events:
2011 return 0; <<
2012 1879
2013 eavail = ep_busy_loop(ep, tim !! 1880 if (!ep_events_available(ep))
2014 if (eavail) !! 1881 ep_busy_loop(ep, timed_out);
2015 continue; <<
2016 1882
2017 if (signal_pending(current)) !! 1883 eavail = ep_events_available(ep);
2018 return -EINTR; !! 1884 if (eavail)
>> 1885 goto send_events;
2019 1886
>> 1887 /*
>> 1888 * Busy poll timed out. Drop NAPI ID for now, we can add
>> 1889 * it back in when we have moved a socket with a valid NAPI
>> 1890 * ID onto the ready list.
>> 1891 */
>> 1892 ep_reset_busy_poll_napi_id(ep);
>> 1893
>> 1894 do {
2020 /* 1895 /*
2021 * Internally init_wait() use 1896 * Internally init_wait() uses autoremove_wake_function(),
2022 * thus wait entry is removed 1897 * thus wait entry is removed from the wait queue on each
2023 * wakeup. Why it is importan 1898 * wakeup. Why it is important? In case of several waiters
2024 * each new wakeup will hit t 1899 * each new wakeup will hit the next waiter, giving it the
2025 * chance to harvest new even 1900 * chance to harvest new event. Otherwise wakeup can be
2026 * lost. This is also good pe 1901 * lost. This is also good performance-wise, because on
2027 * normal wakeup path no need 1902 * normal wakeup path no need to call __remove_wait_queue()
2028 * explicitly, thus ep->lock 1903 * explicitly, thus ep->lock is not taken, which halts the
2029 * event delivery. 1904 * event delivery.
2030 * 1905 *
2031 * In fact, we now use an eve 1906 * In fact, we now use an even more aggressive function that
2032 * unconditionally removes, b 1907 * unconditionally removes, because we don't reuse the wait
2033 * entry between loop iterati 1908 * entry between loop iterations. This lets us also avoid the
2034 * performance issue if a pro 1909 * performance issue if a process is killed, causing all of its
2035 * threads to wake up without 1910 * threads to wake up without being removed normally.
2036 */ 1911 */
2037 init_wait(&wait); 1912 init_wait(&wait);
2038 wait.func = ep_autoremove_wak 1913 wait.func = ep_autoremove_wake_function;
2039 1914
2040 write_lock_irq(&ep->lock); 1915 write_lock_irq(&ep->lock);
2041 /* 1916 /*
2042 * Barrierless variant, waitq 1917 * Barrierless variant, waitqueue_active() is called under
2043 * the same lock on wakeup ep 1918 * the same lock on wakeup ep_poll_callback() side, so it
2044 * is safe to avoid an explic 1919 * is safe to avoid an explicit barrier.
2045 */ 1920 */
2046 __set_current_state(TASK_INTE 1921 __set_current_state(TASK_INTERRUPTIBLE);
2047 1922
2048 /* 1923 /*
2049 * Do the final check under t !! 1924 * Do the final check under the lock. ep_scan_ready_list()
2050 * plays with two lists (->rd 1925 * plays with two lists (->rdllist and ->ovflist) and there
2051 * is always a race when both 1926 * is always a race when both lists are empty for short
2052 * period of time although ev 1927 * period of time although events are pending, so lock is
2053 * important. 1928 * important.
2054 */ 1929 */
2055 eavail = ep_events_available( 1930 eavail = ep_events_available(ep);
2056 if (!eavail) !! 1931 if (!eavail) {
2057 __add_wait_queue_excl !! 1932 if (signal_pending(current))
2058 !! 1933 res = -EINTR;
>> 1934 else
>> 1935 __add_wait_queue_exclusive(&ep->wq, &wait);
>> 1936 }
2059 write_unlock_irq(&ep->lock); 1937 write_unlock_irq(&ep->lock);
2060 1938
2061 if (!eavail) !! 1939 if (!eavail && !res)
2062 timed_out = !schedule 1940 timed_out = !schedule_hrtimeout_range(to, slack,
2063 1941 HRTIMER_MODE_ABS);
2064 __set_current_state(TASK_RUNN <<
2065 1942
2066 /* 1943 /*
2067 * We were woken up, thus go 1944 * We were woken up, thus go and try to harvest some events.
2068 * If timed out and still on 1945 * If timed out and still on the wait queue, recheck eavail
2069 * carefully under lock, belo 1946 * carefully under lock, below.
2070 */ 1947 */
2071 eavail = 1; 1948 eavail = 1;
>> 1949 } while (0);
2072 1950
2073 if (!list_empty_careful(&wait !! 1951 __set_current_state(TASK_RUNNING);
2074 write_lock_irq(&ep->l !! 1952
2075 /* !! 1953 if (!list_empty_careful(&wait.entry)) {
2076 * If the thread time !! 1954 write_lock_irq(&ep->lock);
2077 * it means that the !! 1955 /*
2078 * timeout expired be !! 1956 * If the thread timed out and is not on the wait queue, it
2079 * Thus, when wait.en !! 1957 * means that the thread was woken up after its timeout expired
2080 * events. !! 1958 * before it could reacquire the lock. Thus, when wait.entry is
2081 */ !! 1959 * empty, it needs to harvest events.
2082 if (timed_out) !! 1960 */
2083 eavail = list !! 1961 if (timed_out)
2084 __remove_wait_queue(& !! 1962 eavail = list_empty(&wait.entry);
2085 write_unlock_irq(&ep- !! 1963 __remove_wait_queue(&ep->wq, &wait);
2086 } !! 1964 write_unlock_irq(&ep->lock);
2087 } 1965 }
>> 1966
>> 1967 send_events:
>> 1968 if (fatal_signal_pending(current)) {
>> 1969 /*
>> 1970 * Always short-circuit for fatal signals to allow
>> 1971 * threads to make a timely exit without the chance of
>> 1972 * finding more events available and fetching
>> 1973 * repeatedly.
>> 1974 */
>> 1975 res = -EINTR;
>> 1976 }
>> 1977 /*
>> 1978 * Try to transfer events to user space. In case we get 0 events and
>> 1979 * there's still timeout left over, we go trying again in search of
>> 1980 * more luck.
>> 1981 */
>> 1982 if (!res && eavail &&
>> 1983 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
>> 1984 goto fetch_events;
>> 1985
>> 1986 return res;
2088 } 1987 }
2089 1988
2090 /** 1989 /**
2091 * ep_loop_check_proc - verify that adding an !! 1990 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
2092 * epoll structure does !! 1991 * API, to verify that adding an epoll file inside another
>> 1992 * epoll structure, does not violate the constraints, in
2093 * terms of closed loops 1993 * terms of closed loops, or too deep chains (which can
2094 * result in excessive s 1994 * result in excessive stack usage).
2095 * 1995 *
2096 * @ep: the &struct eventpoll to be currently !! 1996 * @priv: Pointer to the epoll file to be currently checked.
2097 * @depth: Current depth of the path being ch !! 1997 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
>> 1998 * data structure pointer.
>> 1999 * @call_nests: Current dept of the @ep_call_nested() call stack.
2098 * 2000 *
2099 * Return: %zero if adding the epoll @file in !! 2001 * Returns: Returns zero if adding the epoll @file inside current epoll
2100 * structure @ep does not violate th !! 2002 * structure @ep does not violate the constraints, or -1 otherwise.
2101 */ 2003 */
2102 static int ep_loop_check_proc(struct eventpol !! 2004 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
2103 { 2005 {
2104 int error = 0; 2006 int error = 0;
>> 2007 struct file *file = priv;
>> 2008 struct eventpoll *ep = file->private_data;
>> 2009 struct eventpoll *ep_tovisit;
2105 struct rb_node *rbp; 2010 struct rb_node *rbp;
2106 struct epitem *epi; 2011 struct epitem *epi;
2107 2012
2108 mutex_lock_nested(&ep->mtx, depth + 1 !! 2013 mutex_lock_nested(&ep->mtx, call_nests + 1);
2109 ep->gen = loop_check_gen; 2014 ep->gen = loop_check_gen;
2110 for (rbp = rb_first_cached(&ep->rbr); 2015 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
2111 epi = rb_entry(rbp, struct ep 2016 epi = rb_entry(rbp, struct epitem, rbn);
2112 if (unlikely(is_file_epoll(ep 2017 if (unlikely(is_file_epoll(epi->ffd.file))) {
2113 struct eventpoll *ep_ <<
2114 ep_tovisit = epi->ffd 2018 ep_tovisit = epi->ffd.file->private_data;
2115 if (ep_tovisit->gen = 2019 if (ep_tovisit->gen == loop_check_gen)
2116 continue; 2020 continue;
2117 if (ep_tovisit == ins !! 2021 error = ep_call_nested(&poll_loop_ncalls,
2118 error = -1; !! 2022 ep_loop_check_proc, epi->ffd.file,
2119 else !! 2023 ep_tovisit, current);
2120 error = ep_lo <<
2121 if (error != 0) 2024 if (error != 0)
2122 break; 2025 break;
2123 } else { 2026 } else {
2124 /* 2027 /*
2125 * If we've reached a 2028 * If we've reached a file that is not associated with
2126 * an ep, then we nee 2029 * an ep, then we need to check if the newly added
2127 * links are going to 2030 * links are going to add too many wakeup paths. We do
2128 * this by adding it 2031 * this by adding it to the tfile_check_list, if it's
2129 * not already there, 2032 * not already there, and calling reverse_path_check()
2130 * during ep_insert() 2033 * during ep_insert().
2131 */ 2034 */
2132 list_file(epi->ffd.fi !! 2035 if (list_empty(&epi->ffd.file->f_tfile_llink)) {
>> 2036 if (get_file_rcu(epi->ffd.file))
>> 2037 list_add(&epi->ffd.file->f_tfile_llink,
>> 2038 &tfile_check_list);
>> 2039 }
2133 } 2040 }
2134 } 2041 }
2135 mutex_unlock(&ep->mtx); 2042 mutex_unlock(&ep->mtx);
2136 2043
2137 return error; 2044 return error;
2138 } 2045 }
2139 2046
2140 /** 2047 /**
2141 * ep_loop_check - Performs a check to verify !! 2048 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
2142 * into another epoll file (r !! 2049 * another epoll file (represented by @ep) does not create
2143 * closed loops or too deep c 2050 * closed loops or too deep chains.
2144 * 2051 *
2145 * @ep: Pointer to the epoll we are inserting !! 2052 * @ep: Pointer to the epoll private data structure.
2146 * @to: Pointer to the epoll to be inserted. !! 2053 * @file: Pointer to the epoll file to be checked.
2147 * 2054 *
2148 * Return: %zero if adding the epoll @to insi !! 2055 * Returns: Returns zero if adding the epoll @file inside current epoll
2149 * does not violate the constraints, or %-1 o !! 2056 * structure @ep does not violate the constraints, or -1 otherwise.
2150 */ 2057 */
2151 static int ep_loop_check(struct eventpoll *ep !! 2058 static int ep_loop_check(struct eventpoll *ep, struct file *file)
2152 { 2059 {
2153 inserting_into = ep; !! 2060 return ep_call_nested(&poll_loop_ncalls,
2154 return ep_loop_check_proc(to, 0); !! 2061 ep_loop_check_proc, file, ep, current);
2155 } 2062 }
2156 2063
2157 static void clear_tfile_check_list(void) 2064 static void clear_tfile_check_list(void)
2158 { 2065 {
2159 rcu_read_lock(); !! 2066 struct file *file;
2160 while (tfile_check_list != EP_UNACTIV !! 2067
2161 struct epitems_head *head = t !! 2068 /* first clear the tfile_check_list */
2162 tfile_check_list = head->next !! 2069 while (!list_empty(&tfile_check_list)) {
2163 unlist_file(head); !! 2070 file = list_first_entry(&tfile_check_list, struct file,
>> 2071 f_tfile_llink);
>> 2072 list_del_init(&file->f_tfile_llink);
>> 2073 fput(file);
2164 } 2074 }
2165 rcu_read_unlock(); !! 2075 INIT_LIST_HEAD(&tfile_check_list);
2166 } 2076 }
2167 2077
2168 /* 2078 /*
2169 * Open an eventpoll file descriptor. 2079 * Open an eventpoll file descriptor.
2170 */ 2080 */
2171 static int do_epoll_create(int flags) 2081 static int do_epoll_create(int flags)
2172 { 2082 {
2173 int error, fd; 2083 int error, fd;
2174 struct eventpoll *ep = NULL; 2084 struct eventpoll *ep = NULL;
2175 struct file *file; 2085 struct file *file;
2176 2086
2177 /* Check the EPOLL_* constant for con 2087 /* Check the EPOLL_* constant for consistency. */
2178 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEX 2088 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
2179 2089
2180 if (flags & ~EPOLL_CLOEXEC) 2090 if (flags & ~EPOLL_CLOEXEC)
2181 return -EINVAL; 2091 return -EINVAL;
2182 /* 2092 /*
2183 * Create the internal data structure 2093 * Create the internal data structure ("struct eventpoll").
2184 */ 2094 */
2185 error = ep_alloc(&ep); 2095 error = ep_alloc(&ep);
2186 if (error < 0) 2096 if (error < 0)
2187 return error; 2097 return error;
2188 /* 2098 /*
2189 * Creates all the items needed to se 2099 * Creates all the items needed to setup an eventpoll file. That is,
2190 * a file structure and a free file d 2100 * a file structure and a free file descriptor.
2191 */ 2101 */
2192 fd = get_unused_fd_flags(O_RDWR | (fl 2102 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
2193 if (fd < 0) { 2103 if (fd < 0) {
2194 error = fd; 2104 error = fd;
2195 goto out_free_ep; 2105 goto out_free_ep;
2196 } 2106 }
2197 file = anon_inode_getfile("[eventpoll 2107 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
2198 O_RDWR | (fl 2108 O_RDWR | (flags & O_CLOEXEC));
2199 if (IS_ERR(file)) { 2109 if (IS_ERR(file)) {
2200 error = PTR_ERR(file); 2110 error = PTR_ERR(file);
2201 goto out_free_fd; 2111 goto out_free_fd;
2202 } 2112 }
2203 #ifdef CONFIG_NET_RX_BUSY_POLL <<
2204 ep->busy_poll_usecs = 0; <<
2205 ep->busy_poll_budget = 0; <<
2206 ep->prefer_busy_poll = false; <<
2207 #endif <<
2208 ep->file = file; 2113 ep->file = file;
2209 fd_install(fd, file); 2114 fd_install(fd, file);
2210 return fd; 2115 return fd;
2211 2116
2212 out_free_fd: 2117 out_free_fd:
2213 put_unused_fd(fd); 2118 put_unused_fd(fd);
2214 out_free_ep: 2119 out_free_ep:
2215 ep_clear_and_put(ep); !! 2120 ep_free(ep);
2216 return error; 2121 return error;
2217 } 2122 }
2218 2123
2219 SYSCALL_DEFINE1(epoll_create1, int, flags) 2124 SYSCALL_DEFINE1(epoll_create1, int, flags)
2220 { 2125 {
2221 return do_epoll_create(flags); 2126 return do_epoll_create(flags);
2222 } 2127 }
2223 2128
2224 SYSCALL_DEFINE1(epoll_create, int, size) 2129 SYSCALL_DEFINE1(epoll_create, int, size)
2225 { 2130 {
2226 if (size <= 0) 2131 if (size <= 0)
2227 return -EINVAL; 2132 return -EINVAL;
2228 2133
2229 return do_epoll_create(0); 2134 return do_epoll_create(0);
2230 } 2135 }
2231 2136
2232 #ifdef CONFIG_PM_SLEEP <<
2233 static inline void ep_take_care_of_epollwakeu <<
2234 { <<
2235 if ((epev->events & EPOLLWAKEUP) && ! <<
2236 epev->events &= ~EPOLLWAKEUP; <<
2237 } <<
2238 #else <<
2239 static inline void ep_take_care_of_epollwakeu <<
2240 { <<
2241 epev->events &= ~EPOLLWAKEUP; <<
2242 } <<
2243 #endif <<
2244 <<
2245 static inline int epoll_mutex_lock(struct mut 2137 static inline int epoll_mutex_lock(struct mutex *mutex, int depth,
2246 bool nonbl 2138 bool nonblock)
2247 { 2139 {
2248 if (!nonblock) { 2140 if (!nonblock) {
2249 mutex_lock_nested(mutex, dept 2141 mutex_lock_nested(mutex, depth);
2250 return 0; 2142 return 0;
2251 } 2143 }
2252 if (mutex_trylock(mutex)) 2144 if (mutex_trylock(mutex))
2253 return 0; 2145 return 0;
2254 return -EAGAIN; 2146 return -EAGAIN;
2255 } 2147 }
2256 2148
2257 int do_epoll_ctl(int epfd, int op, int fd, st 2149 int do_epoll_ctl(int epfd, int op, int fd, struct epoll_event *epds,
2258 bool nonblock) 2150 bool nonblock)
2259 { 2151 {
2260 int error; 2152 int error;
2261 int full_check = 0; 2153 int full_check = 0;
2262 struct fd f, tf; 2154 struct fd f, tf;
2263 struct eventpoll *ep; 2155 struct eventpoll *ep;
2264 struct epitem *epi; 2156 struct epitem *epi;
2265 struct eventpoll *tep = NULL; 2157 struct eventpoll *tep = NULL;
2266 2158
2267 error = -EBADF; 2159 error = -EBADF;
2268 f = fdget(epfd); 2160 f = fdget(epfd);
2269 if (!f.file) 2161 if (!f.file)
2270 goto error_return; 2162 goto error_return;
2271 2163
2272 /* Get the "struct file *" for the ta 2164 /* Get the "struct file *" for the target file */
2273 tf = fdget(fd); 2165 tf = fdget(fd);
2274 if (!tf.file) 2166 if (!tf.file)
2275 goto error_fput; 2167 goto error_fput;
2276 2168
2277 /* The target file descriptor must su 2169 /* The target file descriptor must support poll */
2278 error = -EPERM; 2170 error = -EPERM;
2279 if (!file_can_poll(tf.file)) 2171 if (!file_can_poll(tf.file))
2280 goto error_tgt_fput; 2172 goto error_tgt_fput;
2281 2173
2282 /* Check if EPOLLWAKEUP is allowed */ 2174 /* Check if EPOLLWAKEUP is allowed */
2283 if (ep_op_has_event(op)) 2175 if (ep_op_has_event(op))
2284 ep_take_care_of_epollwakeup(e 2176 ep_take_care_of_epollwakeup(epds);
2285 2177
2286 /* 2178 /*
2287 * We have to check that the file str 2179 * We have to check that the file structure underneath the file descriptor
2288 * the user passed to us _is_ an even 2180 * the user passed to us _is_ an eventpoll file. And also we do not permit
2289 * adding an epoll file descriptor in 2181 * adding an epoll file descriptor inside itself.
2290 */ 2182 */
2291 error = -EINVAL; 2183 error = -EINVAL;
2292 if (f.file == tf.file || !is_file_epo 2184 if (f.file == tf.file || !is_file_epoll(f.file))
2293 goto error_tgt_fput; 2185 goto error_tgt_fput;
2294 2186
2295 /* 2187 /*
2296 * epoll adds to the wakeup queue at 2188 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2297 * so EPOLLEXCLUSIVE is not allowed f 2189 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2298 * Also, we do not currently supporte 2190 * Also, we do not currently supported nested exclusive wakeups.
2299 */ 2191 */
2300 if (ep_op_has_event(op) && (epds->eve 2192 if (ep_op_has_event(op) && (epds->events & EPOLLEXCLUSIVE)) {
2301 if (op == EPOLL_CTL_MOD) 2193 if (op == EPOLL_CTL_MOD)
2302 goto error_tgt_fput; 2194 goto error_tgt_fput;
2303 if (op == EPOLL_CTL_ADD && (i 2195 if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2304 (epds->events 2196 (epds->events & ~EPOLLEXCLUSIVE_OK_BITS)))
2305 goto error_tgt_fput; 2197 goto error_tgt_fput;
2306 } 2198 }
2307 2199
2308 /* 2200 /*
2309 * At this point it is safe to assume 2201 * At this point it is safe to assume that the "private_data" contains
2310 * our own data structure. 2202 * our own data structure.
2311 */ 2203 */
2312 ep = f.file->private_data; 2204 ep = f.file->private_data;
2313 2205
2314 /* 2206 /*
2315 * When we insert an epoll file descr !! 2207 * When we insert an epoll file descriptor, inside another epoll file
2316 * descriptor, there is the chance of !! 2208 * descriptor, there is the change of creating closed loops, which are
2317 * better be handled here, than in mo 2209 * better be handled here, than in more critical paths. While we are
2318 * checking for loops we also determi 2210 * checking for loops we also determine the list of files reachable
2319 * and hang them on the tfile_check_l 2211 * and hang them on the tfile_check_list, so we can check that we
2320 * haven't created too many possible 2212 * haven't created too many possible wakeup paths.
2321 * 2213 *
2322 * We do not need to take the global 2214 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2323 * the epoll file descriptor is attac 2215 * the epoll file descriptor is attaching directly to a wakeup source,
2324 * unless the epoll file descriptor i 2216 * unless the epoll file descriptor is nested. The purpose of taking the
2325 * 'epnested_mutex' on add is to prev !! 2217 * 'epmutex' on add is to prevent complex toplogies such as loops and
2326 * deep wakeup paths from forming in 2218 * deep wakeup paths from forming in parallel through multiple
2327 * EPOLL_CTL_ADD operations. 2219 * EPOLL_CTL_ADD operations.
2328 */ 2220 */
2329 error = epoll_mutex_lock(&ep->mtx, 0, 2221 error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2330 if (error) 2222 if (error)
2331 goto error_tgt_fput; 2223 goto error_tgt_fput;
2332 if (op == EPOLL_CTL_ADD) { 2224 if (op == EPOLL_CTL_ADD) {
2333 if (READ_ONCE(f.file->f_ep) | !! 2225 if (!list_empty(&f.file->f_ep_links) ||
2334 is_file_epoll(tf.file)) { !! 2226 ep->gen == loop_check_gen ||
>> 2227 is_file_epoll(tf.file)) {
2335 mutex_unlock(&ep->mtx 2228 mutex_unlock(&ep->mtx);
2336 error = epoll_mutex_l !! 2229 error = epoll_mutex_lock(&epmutex, 0, nonblock);
2337 if (error) 2230 if (error)
2338 goto error_tg 2231 goto error_tgt_fput;
2339 loop_check_gen++; 2232 loop_check_gen++;
2340 full_check = 1; 2233 full_check = 1;
2341 if (is_file_epoll(tf. 2234 if (is_file_epoll(tf.file)) {
2342 tep = tf.file <<
2343 error = -ELOO 2235 error = -ELOOP;
2344 if (ep_loop_c !! 2236 if (ep_loop_check(ep, tf.file) != 0)
2345 goto 2237 goto error_tgt_fput;
>> 2238 } else {
>> 2239 get_file(tf.file);
>> 2240 list_add(&tf.file->f_tfile_llink,
>> 2241 &tfile_check_list);
2346 } 2242 }
2347 error = epoll_mutex_l 2243 error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2348 if (error) 2244 if (error)
2349 goto error_tg 2245 goto error_tgt_fput;
>> 2246 if (is_file_epoll(tf.file)) {
>> 2247 tep = tf.file->private_data;
>> 2248 error = epoll_mutex_lock(&tep->mtx, 1, nonblock);
>> 2249 if (error) {
>> 2250 mutex_unlock(&ep->mtx);
>> 2251 goto error_tgt_fput;
>> 2252 }
>> 2253 }
2350 } 2254 }
2351 } 2255 }
2352 2256
2353 /* 2257 /*
2354 * Try to lookup the file inside our !! 2258 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
2355 * above, we can be sure to be able t 2259 * above, we can be sure to be able to use the item looked up by
2356 * ep_find() till we release the mute 2260 * ep_find() till we release the mutex.
2357 */ 2261 */
2358 epi = ep_find(ep, tf.file, fd); 2262 epi = ep_find(ep, tf.file, fd);
2359 2263
2360 error = -EINVAL; 2264 error = -EINVAL;
2361 switch (op) { 2265 switch (op) {
2362 case EPOLL_CTL_ADD: 2266 case EPOLL_CTL_ADD:
2363 if (!epi) { 2267 if (!epi) {
2364 epds->events |= EPOLL 2268 epds->events |= EPOLLERR | EPOLLHUP;
2365 error = ep_insert(ep, 2269 error = ep_insert(ep, epds, tf.file, fd, full_check);
2366 } else 2270 } else
2367 error = -EEXIST; 2271 error = -EEXIST;
2368 break; 2272 break;
2369 case EPOLL_CTL_DEL: 2273 case EPOLL_CTL_DEL:
2370 if (epi) { !! 2274 if (epi)
2371 /* !! 2275 error = ep_remove(ep, epi);
2372 * The eventpoll itse !! 2276 else
2373 * can't go to zero h <<
2374 */ <<
2375 ep_remove_safe(ep, ep <<
2376 error = 0; <<
2377 } else { <<
2378 error = -ENOENT; 2277 error = -ENOENT;
2379 } <<
2380 break; 2278 break;
2381 case EPOLL_CTL_MOD: 2279 case EPOLL_CTL_MOD:
2382 if (epi) { 2280 if (epi) {
2383 if (!(epi->event.even 2281 if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2384 epds->events 2282 epds->events |= EPOLLERR | EPOLLHUP;
2385 error = ep_mo 2283 error = ep_modify(ep, epi, epds);
2386 } 2284 }
2387 } else 2285 } else
2388 error = -ENOENT; 2286 error = -ENOENT;
2389 break; 2287 break;
2390 } 2288 }
>> 2289 if (tep != NULL)
>> 2290 mutex_unlock(&tep->mtx);
2391 mutex_unlock(&ep->mtx); 2291 mutex_unlock(&ep->mtx);
2392 2292
2393 error_tgt_fput: 2293 error_tgt_fput:
2394 if (full_check) { 2294 if (full_check) {
2395 clear_tfile_check_list(); 2295 clear_tfile_check_list();
2396 loop_check_gen++; 2296 loop_check_gen++;
2397 mutex_unlock(&epnested_mutex) !! 2297 mutex_unlock(&epmutex);
2398 } 2298 }
2399 2299
2400 fdput(tf); 2300 fdput(tf);
2401 error_fput: 2301 error_fput:
2402 fdput(f); 2302 fdput(f);
2403 error_return: 2303 error_return:
2404 2304
2405 return error; 2305 return error;
2406 } 2306 }
2407 2307
2408 /* 2308 /*
2409 * The following function implements the cont 2309 * The following function implements the controller interface for
2410 * the eventpoll file that enables the insert 2310 * the eventpoll file that enables the insertion/removal/change of
2411 * file descriptors inside the interest set. 2311 * file descriptors inside the interest set.
2412 */ 2312 */
2413 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op 2313 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
2414 struct epoll_event __user *, 2314 struct epoll_event __user *, event)
2415 { 2315 {
2416 struct epoll_event epds; 2316 struct epoll_event epds;
2417 2317
2418 if (ep_op_has_event(op) && 2318 if (ep_op_has_event(op) &&
2419 copy_from_user(&epds, event, size 2319 copy_from_user(&epds, event, sizeof(struct epoll_event)))
2420 return -EFAULT; 2320 return -EFAULT;
2421 2321
2422 return do_epoll_ctl(epfd, op, fd, &ep 2322 return do_epoll_ctl(epfd, op, fd, &epds, false);
2423 } 2323 }
2424 2324
2425 /* 2325 /*
2426 * Implement the event wait interface for the 2326 * Implement the event wait interface for the eventpoll file. It is the kernel
2427 * part of the user space epoll_wait(2). 2327 * part of the user space epoll_wait(2).
2428 */ 2328 */
2429 static int do_epoll_wait(int epfd, struct epo 2329 static int do_epoll_wait(int epfd, struct epoll_event __user *events,
2430 int maxevents, struc !! 2330 int maxevents, int timeout)
2431 { 2331 {
2432 int error; 2332 int error;
2433 struct fd f; 2333 struct fd f;
2434 struct eventpoll *ep; 2334 struct eventpoll *ep;
2435 2335
2436 /* The maximum number of event must b 2336 /* The maximum number of event must be greater than zero */
2437 if (maxevents <= 0 || maxevents > EP_ 2337 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2438 return -EINVAL; 2338 return -EINVAL;
2439 2339
2440 /* Verify that the area passed by the 2340 /* Verify that the area passed by the user is writeable */
2441 if (!access_ok(events, maxevents * si 2341 if (!access_ok(events, maxevents * sizeof(struct epoll_event)))
2442 return -EFAULT; 2342 return -EFAULT;
2443 2343
2444 /* Get the "struct file *" for the ev 2344 /* Get the "struct file *" for the eventpoll file */
2445 f = fdget(epfd); 2345 f = fdget(epfd);
2446 if (!f.file) 2346 if (!f.file)
2447 return -EBADF; 2347 return -EBADF;
2448 2348
2449 /* 2349 /*
2450 * We have to check that the file str 2350 * We have to check that the file structure underneath the fd
2451 * the user passed to us _is_ an even 2351 * the user passed to us _is_ an eventpoll file.
2452 */ 2352 */
2453 error = -EINVAL; 2353 error = -EINVAL;
2454 if (!is_file_epoll(f.file)) 2354 if (!is_file_epoll(f.file))
2455 goto error_fput; 2355 goto error_fput;
2456 2356
2457 /* 2357 /*
2458 * At this point it is safe to assume 2358 * At this point it is safe to assume that the "private_data" contains
2459 * our own data structure. 2359 * our own data structure.
2460 */ 2360 */
2461 ep = f.file->private_data; 2361 ep = f.file->private_data;
2462 2362
2463 /* Time to fish for events ... */ 2363 /* Time to fish for events ... */
2464 error = ep_poll(ep, events, maxevents !! 2364 error = ep_poll(ep, events, maxevents, timeout);
2465 2365
2466 error_fput: 2366 error_fput:
2467 fdput(f); 2367 fdput(f);
2468 return error; 2368 return error;
2469 } 2369 }
2470 2370
2471 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct 2371 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2472 int, maxevents, int, timeout) 2372 int, maxevents, int, timeout)
2473 { 2373 {
2474 struct timespec64 to; !! 2374 return do_epoll_wait(epfd, events, maxevents, timeout);
2475 <<
2476 return do_epoll_wait(epfd, events, ma <<
2477 ep_timeout_to_ti <<
2478 } 2375 }
2479 2376
2480 /* 2377 /*
2481 * Implement the event wait interface for the 2378 * Implement the event wait interface for the eventpoll file. It is the kernel
2482 * part of the user space epoll_pwait(2). 2379 * part of the user space epoll_pwait(2).
2483 */ 2380 */
2484 static int do_epoll_pwait(int epfd, struct ep !! 2381 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2485 int maxevents, stru !! 2382 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2486 const sigset_t __us !! 2383 size_t, sigsetsize)
2487 { 2384 {
2488 int error; 2385 int error;
2489 2386
2490 /* 2387 /*
2491 * If the caller wants a certain sign 2388 * If the caller wants a certain signal mask to be set during the wait,
2492 * we apply it here. 2389 * we apply it here.
2493 */ 2390 */
2494 error = set_user_sigmask(sigmask, sig 2391 error = set_user_sigmask(sigmask, sigsetsize);
2495 if (error) 2392 if (error)
2496 return error; 2393 return error;
2497 2394
2498 error = do_epoll_wait(epfd, events, m !! 2395 error = do_epoll_wait(epfd, events, maxevents, timeout);
2499 <<
2500 restore_saved_sigmask_unless(error == 2396 restore_saved_sigmask_unless(error == -EINTR);
2501 2397
2502 return error; 2398 return error;
2503 } 2399 }
2504 2400
2505 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struc <<
2506 int, maxevents, int, timeout, <<
2507 size_t, sigsetsize) <<
2508 { <<
2509 struct timespec64 to; <<
2510 <<
2511 return do_epoll_pwait(epfd, events, m <<
2512 ep_timeout_to_t <<
2513 sigmask, sigset <<
2514 } <<
2515 <<
2516 SYSCALL_DEFINE6(epoll_pwait2, int, epfd, stru <<
2517 int, maxevents, const struct <<
2518 const sigset_t __user *, sigm <<
2519 { <<
2520 struct timespec64 ts, *to = NULL; <<
2521 <<
2522 if (timeout) { <<
2523 if (get_timespec64(&ts, timeo <<
2524 return -EFAULT; <<
2525 to = &ts; <<
2526 if (poll_select_set_timeout(t <<
2527 return -EINVAL; <<
2528 } <<
2529 <<
2530 return do_epoll_pwait(epfd, events, m <<
2531 sigmask, sigset <<
2532 } <<
2533 <<
2534 #ifdef CONFIG_COMPAT 2401 #ifdef CONFIG_COMPAT
2535 static int do_compat_epoll_pwait(int epfd, st !! 2402 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2536 int maxevent !! 2403 struct epoll_event __user *, events,
2537 const compat !! 2404 int, maxevents, int, timeout,
2538 compat_size_ !! 2405 const compat_sigset_t __user *, sigmask,
>> 2406 compat_size_t, sigsetsize)
2539 { 2407 {
2540 long err; 2408 long err;
2541 2409
2542 /* 2410 /*
2543 * If the caller wants a certain sign 2411 * If the caller wants a certain signal mask to be set during the wait,
2544 * we apply it here. 2412 * we apply it here.
2545 */ 2413 */
2546 err = set_compat_user_sigmask(sigmask 2414 err = set_compat_user_sigmask(sigmask, sigsetsize);
2547 if (err) 2415 if (err)
2548 return err; 2416 return err;
2549 2417
2550 err = do_epoll_wait(epfd, events, max 2418 err = do_epoll_wait(epfd, events, maxevents, timeout);
2551 <<
2552 restore_saved_sigmask_unless(err == - 2419 restore_saved_sigmask_unless(err == -EINTR);
2553 2420
2554 return err; 2421 return err;
2555 } 2422 }
2556 <<
2557 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd <<
2558 struct epoll_event __u <<
2559 int, maxevents, int, t <<
2560 const compat_sigset_t <<
2561 compat_size_t, sigsets <<
2562 { <<
2563 struct timespec64 to; <<
2564 <<
2565 return do_compat_epoll_pwait(epfd, ev <<
2566 ep_timeo <<
2567 sigmask, <<
2568 } <<
2569 <<
2570 COMPAT_SYSCALL_DEFINE6(epoll_pwait2, int, epf <<
2571 struct epoll_event __u <<
2572 int, maxevents, <<
2573 const struct __kernel_ <<
2574 const compat_sigset_t <<
2575 compat_size_t, sigsets <<
2576 { <<
2577 struct timespec64 ts, *to = NULL; <<
2578 <<
2579 if (timeout) { <<
2580 if (get_timespec64(&ts, timeo <<
2581 return -EFAULT; <<
2582 to = &ts; <<
2583 if (poll_select_set_timeout(t <<
2584 return -EINVAL; <<
2585 } <<
2586 <<
2587 return do_compat_epoll_pwait(epfd, ev <<
2588 sigmask, <<
2589 } <<
2590 <<
2591 #endif 2423 #endif
2592 2424
2593 static int __init eventpoll_init(void) 2425 static int __init eventpoll_init(void)
2594 { 2426 {
2595 struct sysinfo si; 2427 struct sysinfo si;
2596 2428
2597 si_meminfo(&si); 2429 si_meminfo(&si);
2598 /* 2430 /*
2599 * Allows top 4% of lomem to be alloc 2431 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2600 */ 2432 */
2601 max_user_watches = (((si.totalram - s 2433 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2602 EP_ITEM_COST; 2434 EP_ITEM_COST;
2603 BUG_ON(max_user_watches < 0); 2435 BUG_ON(max_user_watches < 0);
2604 2436
2605 /* 2437 /*
>> 2438 * Initialize the structure used to perform epoll file descriptor
>> 2439 * inclusion loops checks.
>> 2440 */
>> 2441 ep_nested_calls_init(&poll_loop_ncalls);
>> 2442
>> 2443 /*
2606 * We can have many thousands of epit 2444 * We can have many thousands of epitems, so prevent this from
2607 * using an extra cache line on 64-bi 2445 * using an extra cache line on 64-bit (and smaller) CPUs
2608 */ 2446 */
2609 BUILD_BUG_ON(sizeof(void *) <= 8 && s 2447 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2610 2448
2611 /* Allocates slab cache used to alloc 2449 /* Allocates slab cache used to allocate "struct epitem" items */
2612 epi_cache = kmem_cache_create("eventp 2450 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2613 0, SLAB_HWCACHE_ALIGN 2451 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC|SLAB_ACCOUNT, NULL);
2614 2452
2615 /* Allocates slab cache used to alloc 2453 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2616 pwq_cache = kmem_cache_create("eventp 2454 pwq_cache = kmem_cache_create("eventpoll_pwq",
2617 sizeof(struct eppoll_entry), 2455 sizeof(struct eppoll_entry), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
2618 epoll_sysctls_init(); <<
2619 <<
2620 ephead_cache = kmem_cache_create("ep_ <<
2621 sizeof(struct epitems_head), <<
2622 2456
2623 return 0; 2457 return 0;
2624 } 2458 }
2625 fs_initcall(eventpoll_init); 2459 fs_initcall(eventpoll_init);
2626 2460
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.