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