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