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