1 // SPDX-License-Identifier: GPL-2.0-or-later <<
2 /* 1 /*
3 * fs/eventpoll.c (Efficient event retrieval 2 * fs/eventpoll.c (Efficient event retrieval implementation)
4 * Copyright (C) 2001,...,2009 Davide Libenz 3 * Copyright (C) 2001,...,2009 Davide Libenzi
5 * 4 *
>> 5 * This program is free software; you can redistribute it and/or modify
>> 6 * it under the terms of the GNU General Public License as published by
>> 7 * the Free Software Foundation; either version 2 of the License, or
>> 8 * (at your option) any later version.
>> 9 *
6 * Davide Libenzi <davidel@xmailserver.org> 10 * Davide Libenzi <davidel@xmailserver.org>
>> 11 *
7 */ 12 */
8 13
9 #include <linux/init.h> 14 #include <linux/init.h>
10 #include <linux/kernel.h> 15 #include <linux/kernel.h>
11 #include <linux/sched/signal.h> 16 #include <linux/sched/signal.h>
12 #include <linux/fs.h> 17 #include <linux/fs.h>
13 #include <linux/file.h> 18 #include <linux/file.h>
14 #include <linux/signal.h> 19 #include <linux/signal.h>
15 #include <linux/errno.h> 20 #include <linux/errno.h>
16 #include <linux/mm.h> 21 #include <linux/mm.h>
17 #include <linux/slab.h> 22 #include <linux/slab.h>
18 #include <linux/poll.h> 23 #include <linux/poll.h>
19 #include <linux/string.h> 24 #include <linux/string.h>
20 #include <linux/list.h> 25 #include <linux/list.h>
21 #include <linux/hash.h> 26 #include <linux/hash.h>
22 #include <linux/spinlock.h> 27 #include <linux/spinlock.h>
23 #include <linux/syscalls.h> 28 #include <linux/syscalls.h>
24 #include <linux/rbtree.h> 29 #include <linux/rbtree.h>
25 #include <linux/wait.h> 30 #include <linux/wait.h>
26 #include <linux/eventpoll.h> 31 #include <linux/eventpoll.h>
27 #include <linux/mount.h> 32 #include <linux/mount.h>
28 #include <linux/bitops.h> 33 #include <linux/bitops.h>
29 #include <linux/mutex.h> 34 #include <linux/mutex.h>
30 #include <linux/anon_inodes.h> 35 #include <linux/anon_inodes.h>
31 #include <linux/device.h> 36 #include <linux/device.h>
32 #include <linux/uaccess.h> 37 #include <linux/uaccess.h>
33 #include <asm/io.h> 38 #include <asm/io.h>
34 #include <asm/mman.h> 39 #include <asm/mman.h>
35 #include <linux/atomic.h> 40 #include <linux/atomic.h>
36 #include <linux/proc_fs.h> 41 #include <linux/proc_fs.h>
37 #include <linux/seq_file.h> 42 #include <linux/seq_file.h>
38 #include <linux/compat.h> 43 #include <linux/compat.h>
39 #include <linux/rculist.h> 44 #include <linux/rculist.h>
40 #include <linux/capability.h> <<
41 #include <net/busy_poll.h> 45 #include <net/busy_poll.h>
42 46
43 /* 47 /*
44 * LOCKING: 48 * LOCKING:
45 * There are three level of locking required b 49 * There are three level of locking required by epoll :
46 * 50 *
47 * 1) epnested_mutex (mutex) !! 51 * 1) epmutex (mutex)
48 * 2) ep->mtx (mutex) 52 * 2) ep->mtx (mutex)
49 * 3) ep->lock (rwlock) !! 53 * 3) ep->lock (spinlock)
50 * 54 *
51 * The acquire order is the one listed above, 55 * The acquire order is the one listed above, from 1 to 3.
52 * We need a rwlock (ep->lock) because we mani !! 56 * We need a spinlock (ep->lock) because we manipulate objects
53 * from inside the poll callback, that might b 57 * from inside the poll callback, that might be triggered from
54 * a wake_up() that in turn might be called fr 58 * a wake_up() that in turn might be called from IRQ context.
55 * So we can't sleep inside the poll callback 59 * So we can't sleep inside the poll callback and hence we need
56 * a spinlock. During the event transfer loop 60 * a spinlock. During the event transfer loop (from kernel to
57 * user space) we could end up sleeping due a 61 * user space) we could end up sleeping due a copy_to_user(), so
58 * we need a lock that will allow us to sleep. 62 * we need a lock that will allow us to sleep. This lock is a
59 * mutex (ep->mtx). It is acquired during the 63 * mutex (ep->mtx). It is acquired during the event transfer loop,
60 * during epoll_ctl(EPOLL_CTL_DEL) and during 64 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
61 * The epnested_mutex is acquired when inserti !! 65 * Then we also need a global mutex to serialize eventpoll_release_file()
62 * epoll fd. We do this so that we walk the ep !! 66 * and ep_free().
>> 67 * This mutex is acquired by ep_free() during the epoll file
>> 68 * cleanup path and it is also acquired by eventpoll_release_file()
>> 69 * if a file has been pushed inside an epoll set and it is then
>> 70 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
>> 71 * It is also acquired when inserting an epoll fd onto another epoll
>> 72 * fd. We do this so that we walk the epoll tree and ensure that this
63 * insertion does not create a cycle of epoll 73 * insertion does not create a cycle of epoll file descriptors, which
64 * could lead to deadlock. We need a global mu 74 * could lead to deadlock. We need a global mutex to prevent two
65 * simultaneous inserts (A into B and B into A 75 * simultaneous inserts (A into B and B into A) from racing and
66 * constructing a cycle without either insert 76 * constructing a cycle without either insert observing that it is
67 * going to. 77 * going to.
68 * It is necessary to acquire multiple "ep->mt 78 * It is necessary to acquire multiple "ep->mtx"es at once in the
69 * case when one epoll fd is added to another. 79 * case when one epoll fd is added to another. In this case, we
70 * always acquire the locks in the order of ne 80 * always acquire the locks in the order of nesting (i.e. after
71 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx w 81 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
72 * before e2->mtx). Since we disallow cycles o 82 * before e2->mtx). Since we disallow cycles of epoll file
73 * descriptors, this ensures that the mutexes 83 * descriptors, this ensures that the mutexes are well-ordered. In
74 * order to communicate this nesting to lockde 84 * order to communicate this nesting to lockdep, when walking a tree
75 * of epoll file descriptors, we use the curre 85 * of epoll file descriptors, we use the current recursion depth as
76 * the lockdep subkey. 86 * the lockdep subkey.
77 * It is possible to drop the "ep->mtx" and to 87 * It is possible to drop the "ep->mtx" and to use the global
78 * mutex "epnested_mutex" (together with "ep-> !! 88 * mutex "epmutex" (together with "ep->lock") to have it working,
79 * but having "ep->mtx" will make the interfac 89 * but having "ep->mtx" will make the interface more scalable.
80 * Events that require holding "epnested_mutex !! 90 * Events that require holding "epmutex" are very rare, while for
81 * normal operations the epoll private "ep->mt 91 * normal operations the epoll private "ep->mtx" will guarantee
82 * a better scalability. 92 * a better scalability.
83 */ 93 */
84 94
85 /* Epoll private bits inside the event mask */ 95 /* Epoll private bits inside the event mask */
86 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLON 96 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
87 97
88 #define EPOLLINOUT_BITS (EPOLLIN | EPOLLOUT) !! 98 #define EPOLLINOUT_BITS (POLLIN | POLLOUT)
89 99
90 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BIT !! 100 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | POLLERR | POLLHUP | \
91 EPOLLWAKEUP | 101 EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
92 102
93 /* Maximum number of nesting allowed inside ep 103 /* Maximum number of nesting allowed inside epoll sets */
94 #define EP_MAX_NESTS 4 104 #define EP_MAX_NESTS 4
95 105
96 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct 106 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
97 107
98 #define EP_UNACTIVE_PTR ((void *) -1L) 108 #define EP_UNACTIVE_PTR ((void *) -1L)
99 109
100 #define EP_ITEM_COST (sizeof(struct epitem) + 110 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
101 111
102 struct epoll_filefd { 112 struct epoll_filefd {
103 struct file *file; 113 struct file *file;
104 int fd; 114 int fd;
105 } __packed; 115 } __packed;
106 116
107 /* Wait structure used by the poll hooks */ !! 117 /*
108 struct eppoll_entry { !! 118 * Structure used to track possible nested calls, for too deep recursions
109 /* List header used to link this struc !! 119 * and loop cycles.
110 struct eppoll_entry *next; !! 120 */
111 !! 121 struct nested_call_node {
112 /* The "base" pointer is set to the co !! 122 struct list_head llink;
113 struct epitem *base; !! 123 void *cookie;
114 !! 124 void *ctx;
115 /* !! 125 };
116 * Wait queue item that will be linked <<
117 * queue head. <<
118 */ <<
119 wait_queue_entry_t wait; <<
120 126
121 /* The wait queue head that linked the !! 127 /*
122 wait_queue_head_t *whead; !! 128 * This structure is used as collector for nested calls, to check for
>> 129 * maximum recursion dept and loop cycles.
>> 130 */
>> 131 struct nested_calls {
>> 132 struct list_head tasks_call_list;
>> 133 spinlock_t lock;
123 }; 134 };
124 135
125 /* 136 /*
126 * Each file descriptor added to the eventpoll 137 * Each file descriptor added to the eventpoll interface will
127 * have an entry of this type linked to the "r 138 * have an entry of this type linked to the "rbr" RB tree.
128 * Avoid increasing the size of this struct, t 139 * Avoid increasing the size of this struct, there can be many thousands
129 * of these on a server and we do not want thi 140 * of these on a server and we do not want this to take another cache line.
130 */ 141 */
131 struct epitem { 142 struct epitem {
132 union { 143 union {
133 /* RB tree node links this str 144 /* RB tree node links this structure to the eventpoll RB tree */
134 struct rb_node rbn; 145 struct rb_node rbn;
135 /* Used to free the struct epi 146 /* Used to free the struct epitem */
136 struct rcu_head rcu; 147 struct rcu_head rcu;
137 }; 148 };
138 149
139 /* List header used to link this struc 150 /* List header used to link this structure to the eventpoll ready list */
140 struct list_head rdllink; 151 struct list_head rdllink;
141 152
142 /* 153 /*
143 * Works together "struct eventpoll"-> 154 * Works together "struct eventpoll"->ovflist in keeping the
144 * single linked chain of items. 155 * single linked chain of items.
145 */ 156 */
146 struct epitem *next; 157 struct epitem *next;
147 158
148 /* The file descriptor information thi 159 /* The file descriptor information this item refers to */
149 struct epoll_filefd ffd; 160 struct epoll_filefd ffd;
150 161
151 /* !! 162 /* Number of active wait queue attached to poll operations */
152 * Protected by file->f_lock, true for !! 163 int nwait;
153 * removed from the "struct file" item <<
154 * eventpoll->refcount orchestrates "s <<
155 */ <<
156 bool dying; <<
157 164
158 /* List containing poll wait queues */ 165 /* List containing poll wait queues */
159 struct eppoll_entry *pwqlist; !! 166 struct list_head pwqlist;
160 167
161 /* The "container" of this item */ 168 /* The "container" of this item */
162 struct eventpoll *ep; 169 struct eventpoll *ep;
163 170
164 /* List header used to link this item 171 /* List header used to link this item to the "struct file" items list */
165 struct hlist_node fllink; !! 172 struct list_head fllink;
166 173
167 /* wakeup_source used when EPOLLWAKEUP 174 /* wakeup_source used when EPOLLWAKEUP is set */
168 struct wakeup_source __rcu *ws; 175 struct wakeup_source __rcu *ws;
169 176
170 /* The structure that describe the int 177 /* The structure that describe the interested events and the source fd */
171 struct epoll_event event; 178 struct epoll_event event;
172 }; 179 };
173 180
174 /* 181 /*
175 * This structure is stored inside the "privat 182 * This structure is stored inside the "private_data" member of the file
176 * structure and represents the main data stru 183 * structure and represents the main data structure for the eventpoll
177 * interface. 184 * interface.
178 */ 185 */
179 struct eventpoll { 186 struct eventpoll {
>> 187 /* Protect the access to this structure */
>> 188 spinlock_t lock;
>> 189
180 /* 190 /*
181 * This mutex is used to ensure that f 191 * This mutex is used to ensure that files are not removed
182 * while epoll is using them. This is 192 * while epoll is using them. This is held during the event
183 * collection loop, the file cleanup p 193 * collection loop, the file cleanup path, the epoll file exit
184 * code and the ctl operations. 194 * code and the ctl operations.
185 */ 195 */
186 struct mutex mtx; 196 struct mutex mtx;
187 197
188 /* Wait queue used by sys_epoll_wait() 198 /* Wait queue used by sys_epoll_wait() */
189 wait_queue_head_t wq; 199 wait_queue_head_t wq;
190 200
191 /* Wait queue used by file->poll() */ 201 /* Wait queue used by file->poll() */
192 wait_queue_head_t poll_wait; 202 wait_queue_head_t poll_wait;
193 203
194 /* List of ready file descriptors */ 204 /* List of ready file descriptors */
195 struct list_head rdllist; 205 struct list_head rdllist;
196 206
197 /* Lock which protects rdllist and ovf <<
198 rwlock_t lock; <<
199 <<
200 /* RB tree root used to store monitore 207 /* RB tree root used to store monitored fd structs */
201 struct rb_root_cached rbr; 208 struct rb_root_cached rbr;
202 209
203 /* 210 /*
204 * This is a single linked list that c 211 * This is a single linked list that chains all the "struct epitem" that
205 * happened while transferring ready e 212 * happened while transferring ready events to userspace w/out
206 * holding ->lock. 213 * holding ->lock.
207 */ 214 */
208 struct epitem *ovflist; 215 struct epitem *ovflist;
209 216
210 /* wakeup_source used when ep_send_eve !! 217 /* wakeup_source used when ep_scan_ready_list is running */
211 struct wakeup_source *ws; 218 struct wakeup_source *ws;
212 219
213 /* The user that created the eventpoll 220 /* The user that created the eventpoll descriptor */
214 struct user_struct *user; 221 struct user_struct *user;
215 222
216 struct file *file; 223 struct file *file;
217 224
218 /* used to optimize loop detection che 225 /* used to optimize loop detection check */
219 u64 gen; 226 u64 gen;
220 struct hlist_head refs; <<
221 <<
222 /* <<
223 * usage count, used together with epi <<
224 * orchestrate the disposal of this st <<
225 */ <<
226 refcount_t refcount; <<
227 227
228 #ifdef CONFIG_NET_RX_BUSY_POLL 228 #ifdef CONFIG_NET_RX_BUSY_POLL
229 /* used to track busy poll napi_id */ 229 /* used to track busy poll napi_id */
230 unsigned int napi_id; 230 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 231 #endif
>> 232 };
237 233
238 #ifdef CONFIG_DEBUG_LOCK_ALLOC !! 234 /* Wait structure used by the poll hooks */
239 /* tracks wakeup nests for lockdep val !! 235 struct eppoll_entry {
240 u8 nests; !! 236 /* List header used to link this structure to the "struct epitem" */
241 #endif !! 237 struct list_head llink;
>> 238
>> 239 /* The "base" pointer is set to the container "struct epitem" */
>> 240 struct epitem *base;
>> 241
>> 242 /*
>> 243 * Wait queue item that will be linked to the target file wait
>> 244 * queue head.
>> 245 */
>> 246 wait_queue_entry_t wait;
>> 247
>> 248 /* The wait queue head that linked the "wait" wait queue item */
>> 249 wait_queue_head_t *whead;
242 }; 250 };
243 251
244 /* Wrapper struct used by poll queueing */ 252 /* Wrapper struct used by poll queueing */
245 struct ep_pqueue { 253 struct ep_pqueue {
246 poll_table pt; 254 poll_table pt;
247 struct epitem *epi; 255 struct epitem *epi;
248 }; 256 };
249 257
>> 258 /* Used by the ep_send_events() function as callback private data */
>> 259 struct ep_send_events_data {
>> 260 int maxevents;
>> 261 struct epoll_event __user *events;
>> 262 };
>> 263
250 /* 264 /*
251 * Configuration options available inside /pro 265 * Configuration options available inside /proc/sys/fs/epoll/
252 */ 266 */
253 /* Maximum number of epoll watched descriptors 267 /* Maximum number of epoll watched descriptors, per user */
254 static long max_user_watches __read_mostly; 268 static long max_user_watches __read_mostly;
255 269
256 /* Used for cycles detection */ !! 270 /*
257 static DEFINE_MUTEX(epnested_mutex); !! 271 * This mutex is used to serialize ep_free() and eventpoll_release_file().
>> 272 */
>> 273 static DEFINE_MUTEX(epmutex);
258 274
259 static u64 loop_check_gen = 0; 275 static u64 loop_check_gen = 0;
260 276
261 /* Used to check for epoll file descriptor inc 277 /* Used to check for epoll file descriptor inclusion loops */
262 static struct eventpoll *inserting_into; !! 278 static struct nested_calls poll_loop_ncalls;
>> 279
>> 280 /* Used for safe wake up implementation */
>> 281 static struct nested_calls poll_safewake_ncalls;
>> 282
>> 283 /* Used to call file's f_op->poll() under the nested calls boundaries */
>> 284 static struct nested_calls poll_readywalk_ncalls;
263 285
264 /* Slab cache used to allocate "struct epitem" 286 /* Slab cache used to allocate "struct epitem" */
265 static struct kmem_cache *epi_cache __ro_after !! 287 static struct kmem_cache *epi_cache __read_mostly;
266 288
267 /* Slab cache used to allocate "struct eppoll_ 289 /* Slab cache used to allocate "struct eppoll_entry" */
268 static struct kmem_cache *pwq_cache __ro_after !! 290 static struct kmem_cache *pwq_cache __read_mostly;
269 291
270 /* 292 /*
271 * List of files with newly added links, where 293 * List of files with newly added links, where we may need to limit the number
272 * of emanating paths. Protected by the epnest !! 294 * of emanating paths. Protected by the epmutex.
273 */ 295 */
274 struct epitems_head { !! 296 static LIST_HEAD(tfile_check_list);
275 struct hlist_head epitems; <<
276 struct epitems_head *next; <<
277 }; <<
278 static struct epitems_head *tfile_check_list = <<
279 <<
280 static struct kmem_cache *ephead_cache __ro_af <<
281 <<
282 static inline void free_ephead(struct epitems_ <<
283 { <<
284 if (head) <<
285 kmem_cache_free(ephead_cache, <<
286 } <<
287 <<
288 static void list_file(struct file *file) <<
289 { <<
290 struct epitems_head *head; <<
291 <<
292 head = container_of(file->f_ep, struct <<
293 if (!head->next) { <<
294 head->next = tfile_check_list; <<
295 tfile_check_list = head; <<
296 } <<
297 } <<
298 <<
299 static void unlist_file(struct epitems_head *h <<
300 { <<
301 struct epitems_head *to_free = head; <<
302 struct hlist_node *p = rcu_dereference <<
303 if (p) { <<
304 struct epitem *epi= container_ <<
305 spin_lock(&epi->ffd.file->f_lo <<
306 if (!hlist_empty(&head->epitem <<
307 to_free = NULL; <<
308 head->next = NULL; <<
309 spin_unlock(&epi->ffd.file->f_ <<
310 } <<
311 free_ephead(to_free); <<
312 } <<
313 297
314 #ifdef CONFIG_SYSCTL 298 #ifdef CONFIG_SYSCTL
315 299
316 #include <linux/sysctl.h> 300 #include <linux/sysctl.h>
317 301
318 static long long_zero; !! 302 static long zero;
319 static long long_max = LONG_MAX; 303 static long long_max = LONG_MAX;
320 304
321 static struct ctl_table epoll_table[] = { !! 305 struct ctl_table epoll_table[] = {
322 { 306 {
323 .procname = "max_user_wa 307 .procname = "max_user_watches",
324 .data = &max_user_wa 308 .data = &max_user_watches,
325 .maxlen = sizeof(max_u 309 .maxlen = sizeof(max_user_watches),
326 .mode = 0644, 310 .mode = 0644,
327 .proc_handler = proc_doulong 311 .proc_handler = proc_doulongvec_minmax,
328 .extra1 = &long_zero, !! 312 .extra1 = &zero,
329 .extra2 = &long_max, 313 .extra2 = &long_max,
330 }, 314 },
>> 315 { }
331 }; 316 };
332 <<
333 static void __init epoll_sysctls_init(void) <<
334 { <<
335 register_sysctl("fs/epoll", epoll_tabl <<
336 } <<
337 #else <<
338 #define epoll_sysctls_init() do { } while (0) <<
339 #endif /* CONFIG_SYSCTL */ 317 #endif /* CONFIG_SYSCTL */
340 318
341 static const struct file_operations eventpoll_ 319 static const struct file_operations eventpoll_fops;
342 320
343 static inline int is_file_epoll(struct file *f 321 static inline int is_file_epoll(struct file *f)
344 { 322 {
345 return f->f_op == &eventpoll_fops; 323 return f->f_op == &eventpoll_fops;
346 } 324 }
347 325
348 /* Setup the structure that is used as key for 326 /* Setup the structure that is used as key for the RB tree */
349 static inline void ep_set_ffd(struct epoll_fil 327 static inline void ep_set_ffd(struct epoll_filefd *ffd,
350 struct file *fil 328 struct file *file, int fd)
351 { 329 {
352 ffd->file = file; 330 ffd->file = file;
353 ffd->fd = fd; 331 ffd->fd = fd;
354 } 332 }
355 333
356 /* Compare RB tree keys */ 334 /* Compare RB tree keys */
357 static inline int ep_cmp_ffd(struct epoll_file 335 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
358 struct epoll_file 336 struct epoll_filefd *p2)
359 { 337 {
360 return (p1->file > p2->file ? +1: 338 return (p1->file > p2->file ? +1:
361 (p1->file < p2->file ? -1 : p1 339 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
362 } 340 }
363 341
364 /* Tells us if the item is currently linked */ 342 /* Tells us if the item is currently linked */
365 static inline int ep_is_linked(struct epitem * !! 343 static inline int ep_is_linked(struct list_head *p)
366 { 344 {
367 return !list_empty(&epi->rdllink); !! 345 return !list_empty(p);
368 } 346 }
369 347
370 static inline struct eppoll_entry *ep_pwq_from 348 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
371 { 349 {
372 return container_of(p, struct eppoll_e 350 return container_of(p, struct eppoll_entry, wait);
373 } 351 }
374 352
375 /* Get the "struct epitem" from a wait queue p 353 /* Get the "struct epitem" from a wait queue pointer */
376 static inline struct epitem *ep_item_from_wait 354 static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
377 { 355 {
378 return container_of(p, struct eppoll_e 356 return container_of(p, struct eppoll_entry, wait)->base;
379 } 357 }
380 358
381 /** !! 359 /* Get the "struct epitem" from an epoll queue wrapper */
382 * ep_events_available - Checks if ready event !! 360 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
383 * <<
384 * @ep: Pointer to the eventpoll context. <<
385 * <<
386 * Return: a value different than %zero if rea <<
387 * or %zero otherwise. <<
388 */ <<
389 static inline int ep_events_available(struct e <<
390 { 361 {
391 return !list_empty_careful(&ep->rdllis !! 362 return container_of(p, struct ep_pqueue, pt)->epi;
392 READ_ONCE(ep->ovflist) != EP_U !! 363 }
>> 364
>> 365 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
>> 366 static inline int ep_op_has_event(int op)
>> 367 {
>> 368 return op != EPOLL_CTL_DEL;
>> 369 }
>> 370
>> 371 /* Initialize the poll safe wake up structure */
>> 372 static void ep_nested_calls_init(struct nested_calls *ncalls)
>> 373 {
>> 374 INIT_LIST_HEAD(&ncalls->tasks_call_list);
>> 375 spin_lock_init(&ncalls->lock);
393 } 376 }
394 377
395 #ifdef CONFIG_NET_RX_BUSY_POLL <<
396 /** 378 /**
397 * busy_loop_ep_timeout - check if busy poll h !! 379 * ep_events_available - Checks if ready events might be available.
398 * from the epoll instance ep is preferred, bu <<
399 * the system-wide global via busy_loop_timeou <<
400 * 380 *
401 * @start_time: The start time used to compute <<
402 * @ep: Pointer to the eventpoll context. 381 * @ep: Pointer to the eventpoll context.
403 * 382 *
404 * Return: true if the timeout has expired, fa !! 383 * Returns: Returns a value different than zero if ready events are available,
>> 384 * or zero otherwise.
405 */ 385 */
406 static bool busy_loop_ep_timeout(unsigned long !! 386 static inline int ep_events_available(struct eventpoll *ep)
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 { 387 {
423 return !!READ_ONCE(ep->busy_poll_usecs !! 388 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
424 } 389 }
425 390
>> 391 #ifdef CONFIG_NET_RX_BUSY_POLL
426 static bool ep_busy_loop_end(void *p, unsigned 392 static bool ep_busy_loop_end(void *p, unsigned long start_time)
427 { 393 {
428 struct eventpoll *ep = p; 394 struct eventpoll *ep = p;
429 395
430 return ep_events_available(ep) || busy !! 396 return ep_events_available(ep) || busy_loop_timeout(start_time);
431 } 397 }
>> 398 #endif /* CONFIG_NET_RX_BUSY_POLL */
432 399
433 /* 400 /*
434 * Busy poll if globally on and supporting soc 401 * Busy poll if globally on and supporting sockets found && no events,
435 * busy loop will return if need_resched or ep 402 * busy loop will return if need_resched or ep_events_available.
436 * 403 *
437 * we must do our busy polling with irqs enabl 404 * we must do our busy polling with irqs enabled
438 */ 405 */
439 static bool ep_busy_loop(struct eventpoll *ep, !! 406 static void ep_busy_loop(struct eventpoll *ep, int nonblock)
440 { 407 {
>> 408 #ifdef CONFIG_NET_RX_BUSY_POLL
441 unsigned int napi_id = READ_ONCE(ep->n 409 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-> <<
444 410
445 if (!budget) !! 411 if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on())
446 budget = BUSY_POLL_BUDGET; !! 412 napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep);
>> 413 #endif
>> 414 }
447 415
448 if (napi_id >= MIN_NAPI_ID && ep_busy_ !! 416 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
449 napi_busy_loop(napi_id, nonblo !! 417 {
450 ep, prefer_busy !! 418 #ifdef CONFIG_NET_RX_BUSY_POLL
451 if (ep_events_available(ep)) !! 419 if (ep->napi_id)
452 return true; <<
453 /* <<
454 * Busy poll timed out. Drop <<
455 * it back in when we have mov <<
456 * ID onto the ready list. <<
457 */ <<
458 ep->napi_id = 0; 420 ep->napi_id = 0;
459 return false; !! 421 #endif
460 } <<
461 return false; <<
462 } 422 }
463 423
464 /* 424 /*
465 * Set epoll busy poll NAPI ID from sk. 425 * Set epoll busy poll NAPI ID from sk.
466 */ 426 */
467 static inline void ep_set_busy_poll_napi_id(st 427 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
468 { 428 {
469 struct eventpoll *ep = epi->ep; !! 429 #ifdef CONFIG_NET_RX_BUSY_POLL
>> 430 struct eventpoll *ep;
470 unsigned int napi_id; 431 unsigned int napi_id;
471 struct socket *sock; 432 struct socket *sock;
472 struct sock *sk; 433 struct sock *sk;
>> 434 int err;
473 435
474 if (!ep_busy_loop_on(ep)) !! 436 if (!net_busy_loop_on())
475 return; 437 return;
476 438
477 sock = sock_from_file(epi->ffd.file); !! 439 sock = sock_from_file(epi->ffd.file, &err);
478 if (!sock) 440 if (!sock)
479 return; 441 return;
480 442
481 sk = sock->sk; 443 sk = sock->sk;
482 if (!sk) 444 if (!sk)
483 return; 445 return;
484 446
485 napi_id = READ_ONCE(sk->sk_napi_id); 447 napi_id = READ_ONCE(sk->sk_napi_id);
>> 448 ep = epi->ep;
486 449
487 /* Non-NAPI IDs can be rejected 450 /* Non-NAPI IDs can be rejected
488 * or 451 * or
489 * Nothing to do if we already have th 452 * Nothing to do if we already have this ID
490 */ 453 */
491 if (napi_id < MIN_NAPI_ID || napi_id = 454 if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)
492 return; 455 return;
493 456
494 /* record NAPI ID for use in next busy 457 /* record NAPI ID for use in next busy poll */
495 ep->napi_id = napi_id; 458 ep->napi_id = napi_id;
>> 459 #endif
496 } 460 }
497 461
498 static long ep_eventpoll_bp_ioctl(struct file !! 462 /**
499 unsigned lon !! 463 * ep_call_nested - Perform a bound (possibly) nested call, by checking
>> 464 * that the recursion limit is not exceeded, and that
>> 465 * the same nested call (by the meaning of same cookie) is
>> 466 * no re-entered.
>> 467 *
>> 468 * @ncalls: Pointer to the nested_calls structure to be used for this call.
>> 469 * @max_nests: Maximum number of allowed nesting calls.
>> 470 * @nproc: Nested call core function pointer.
>> 471 * @priv: Opaque data to be passed to the @nproc callback.
>> 472 * @cookie: Cookie to be used to identify this nested call.
>> 473 * @ctx: This instance context.
>> 474 *
>> 475 * Returns: Returns the code returned by the @nproc callback, or -1 if
>> 476 * the maximum recursion limit has been exceeded.
>> 477 */
>> 478 static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
>> 479 int (*nproc)(void *, void *, int), void *priv,
>> 480 void *cookie, void *ctx)
500 { 481 {
501 struct eventpoll *ep = file->private_d !! 482 int error, call_nests = 0;
502 void __user *uarg = (void __user *)arg !! 483 unsigned long flags;
503 struct epoll_params epoll_params; !! 484 struct list_head *lsthead = &ncalls->tasks_call_list;
504 !! 485 struct nested_call_node *tncur;
505 switch (cmd) { !! 486 struct nested_call_node tnode;
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; <<
513 487
514 if (epoll_params.busy_poll_use !! 488 spin_lock_irqsave(&ncalls->lock, flags);
515 return -EINVAL; <<
516 489
517 if (epoll_params.prefer_busy_p !! 490 /*
518 return -EINVAL; !! 491 * Try to see if the current task is already inside this wakeup call.
519 !! 492 * We use a list here, since the population inside this set is always
520 if (epoll_params.busy_poll_bud !! 493 * very much limited.
521 !capable(CAP_NET_ADMIN)) !! 494 */
522 return -EPERM; !! 495 list_for_each_entry(tncur, lsthead, llink) {
523 !! 496 if (tncur->ctx == ctx &&
524 WRITE_ONCE(ep->busy_poll_usecs !! 497 (tncur->cookie == cookie || ++call_nests > max_nests)) {
525 WRITE_ONCE(ep->busy_poll_budge !! 498 /*
526 WRITE_ONCE(ep->prefer_busy_pol !! 499 * Ops ... loop detected or maximum nest level reached.
527 return 0; !! 500 * We abort this wake by breaking the cycle itself.
528 case EPIOCGPARAMS: !! 501 */
529 memset(&epoll_params, 0, sizeo !! 502 error = -1;
530 epoll_params.busy_poll_usecs = !! 503 goto out_unlock;
531 epoll_params.busy_poll_budget !! 504 }
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 } 505 }
539 } <<
540 <<
541 #else <<
542 506
543 static inline bool ep_busy_loop(struct eventpo !! 507 /* Add the current task and cookie to the list */
544 { !! 508 tnode.ctx = ctx;
545 return false; !! 509 tnode.cookie = cookie;
546 } !! 510 list_add(&tnode.llink, lsthead);
547 !! 511
548 static inline void ep_set_busy_poll_napi_id(st !! 512 spin_unlock_irqrestore(&ncalls->lock, flags);
549 { !! 513
550 } !! 514 /* Call the nested function */
>> 515 error = (*nproc)(priv, cookie, call_nests);
>> 516
>> 517 /* Remove the current task from the list */
>> 518 spin_lock_irqsave(&ncalls->lock, flags);
>> 519 list_del(&tnode.llink);
>> 520 out_unlock:
>> 521 spin_unlock_irqrestore(&ncalls->lock, flags);
551 522
552 static long ep_eventpoll_bp_ioctl(struct file !! 523 return error;
553 unsigned lon <<
554 { <<
555 return -EOPNOTSUPP; <<
556 } 524 }
557 525
558 #endif /* CONFIG_NET_RX_BUSY_POLL */ <<
559 <<
560 /* 526 /*
561 * As described in commit 0ccf831cb lockdep: a 527 * As described in commit 0ccf831cb lockdep: annotate epoll
562 * the use of wait queues used by epoll is don 528 * the use of wait queues used by epoll is done in a very controlled
563 * manner. Wake ups can nest inside each other 529 * manner. Wake ups can nest inside each other, but are never done
564 * with the same locking. For example: 530 * with the same locking. For example:
565 * 531 *
566 * dfd = socket(...); 532 * dfd = socket(...);
567 * efd1 = epoll_create(); 533 * efd1 = epoll_create();
568 * efd2 = epoll_create(); 534 * efd2 = epoll_create();
569 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...); 535 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
570 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...) 536 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
571 * 537 *
572 * When a packet arrives to the device underne 538 * When a packet arrives to the device underneath "dfd", the net code will
573 * issue a wake_up() on its poll wake list. Ep 539 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
574 * callback wakeup entry on that queue, and th 540 * callback wakeup entry on that queue, and the wake_up() performed by the
575 * "dfd" net code will end up in ep_poll_callb 541 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
576 * (efd1) notices that it may have some event 542 * (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 543 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
578 * that ends up in another wake_up(), after ha 544 * that ends up in another wake_up(), after having checked about the
579 * recursion constraints. That are, no more th !! 545 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
580 * stack blasting. !! 546 * avoid stack blasting.
581 * 547 *
582 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, ma 548 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
583 * this special case of epoll. 549 * this special case of epoll.
584 */ 550 */
585 #ifdef CONFIG_DEBUG_LOCK_ALLOC 551 #ifdef CONFIG_DEBUG_LOCK_ALLOC
586 !! 552 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
587 static void ep_poll_safewake(struct eventpoll !! 553 unsigned long events, int subclass)
588 unsigned pollflag <<
589 { 554 {
590 struct eventpoll *ep_src; <<
591 unsigned long flags; 555 unsigned long flags;
592 u8 nests = 0; <<
593 556
594 /* !! 557 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
595 * To set the subclass or nesting leve !! 558 wake_up_locked_poll(wqueue, events);
596 * it might be natural to create a per !! 559 spin_unlock_irqrestore(&wqueue->lock, flags);
597 * we can recurse on ep->poll_wait.loc <<
598 * schedule() in the -rt kernel, the p <<
599 * protected. Thus, we are introducing <<
600 * If we are not being call from ep_po <<
601 * we are at the first level of nestin <<
602 * called from ep_poll_callback() and <<
603 * not an epoll file itself, we are at <<
604 * is depth 0. If the wakeup source is <<
605 * wakeup chain then we use its nests <<
606 * nests + 1. The previous epoll file <<
607 * already holding its own poll_wait.l <<
608 */ <<
609 if (epi) { <<
610 if ((is_file_epoll(epi->ffd.fi <<
611 ep_src = epi->ffd.file <<
612 nests = ep_src->nests; <<
613 } else { <<
614 nests = 1; <<
615 } <<
616 } <<
617 spin_lock_irqsave_nested(&ep->poll_wai <<
618 ep->nests = nests + 1; <<
619 wake_up_locked_poll(&ep->poll_wait, EP <<
620 ep->nests = 0; <<
621 spin_unlock_irqrestore(&ep->poll_wait. <<
622 } 560 }
623 <<
624 #else 561 #else
>> 562 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
>> 563 unsigned long events, int subclass)
>> 564 {
>> 565 wake_up_poll(wqueue, events);
>> 566 }
>> 567 #endif
625 568
626 static void ep_poll_safewake(struct eventpoll !! 569 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
627 __poll_t pollflag <<
628 { 570 {
629 wake_up_poll(&ep->poll_wait, EPOLLIN | !! 571 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
>> 572 1 + call_nests);
>> 573 return 0;
630 } 574 }
631 575
632 #endif !! 576 /*
>> 577 * Perform a safe wake up of the poll wait list. The problem is that
>> 578 * with the new callback'd wake up system, it is possible that the
>> 579 * poll callback is reentered from inside the call to wake_up() done
>> 580 * on the poll wait queue head. The rule is that we cannot reenter the
>> 581 * wake up code from the same task more than EP_MAX_NESTS times,
>> 582 * and we cannot reenter the same wait queue head at all. This will
>> 583 * enable to have a hierarchy of epoll file descriptor of no more than
>> 584 * EP_MAX_NESTS deep.
>> 585 */
>> 586 static void ep_poll_safewake(wait_queue_head_t *wq)
>> 587 {
>> 588 int this_cpu = get_cpu();
>> 589
>> 590 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
>> 591 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
>> 592
>> 593 put_cpu();
>> 594 }
633 595
634 static void ep_remove_wait_queue(struct eppoll 596 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
635 { 597 {
636 wait_queue_head_t *whead; 598 wait_queue_head_t *whead;
637 599
638 rcu_read_lock(); 600 rcu_read_lock();
639 /* 601 /*
640 * If it is cleared by POLLFREE, it sh 602 * If it is cleared by POLLFREE, it should be rcu-safe.
641 * If we read NULL we need a barrier p 603 * If we read NULL we need a barrier paired with
642 * smp_store_release() in ep_poll_call 604 * smp_store_release() in ep_poll_callback(), otherwise
643 * we rely on whead->lock. 605 * we rely on whead->lock.
644 */ 606 */
645 whead = smp_load_acquire(&pwq->whead); 607 whead = smp_load_acquire(&pwq->whead);
646 if (whead) 608 if (whead)
647 remove_wait_queue(whead, &pwq- 609 remove_wait_queue(whead, &pwq->wait);
648 rcu_read_unlock(); 610 rcu_read_unlock();
649 } 611 }
650 612
651 /* 613 /*
652 * This function unregisters poll callbacks fr 614 * This function unregisters poll callbacks from the associated file
653 * descriptor. Must be called with "mtx" held !! 615 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
>> 616 * ep_free).
654 */ 617 */
655 static void ep_unregister_pollwait(struct even 618 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
656 { 619 {
657 struct eppoll_entry **p = &epi->pwqlis !! 620 struct list_head *lsthead = &epi->pwqlist;
658 struct eppoll_entry *pwq; 621 struct eppoll_entry *pwq;
659 622
660 while ((pwq = *p) != NULL) { !! 623 while (!list_empty(lsthead)) {
661 *p = pwq->next; !! 624 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
>> 625
>> 626 list_del(&pwq->llink);
662 ep_remove_wait_queue(pwq); 627 ep_remove_wait_queue(pwq);
663 kmem_cache_free(pwq_cache, pwq 628 kmem_cache_free(pwq_cache, pwq);
664 } 629 }
665 } 630 }
666 631
667 /* call only when ep->mtx is held */ 632 /* call only when ep->mtx is held */
668 static inline struct wakeup_source *ep_wakeup_ 633 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
669 { 634 {
670 return rcu_dereference_check(epi->ws, 635 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
671 } 636 }
672 637
673 /* call only when ep->mtx is held */ 638 /* call only when ep->mtx is held */
674 static inline void ep_pm_stay_awake(struct epi 639 static inline void ep_pm_stay_awake(struct epitem *epi)
675 { 640 {
676 struct wakeup_source *ws = ep_wakeup_s 641 struct wakeup_source *ws = ep_wakeup_source(epi);
677 642
678 if (ws) 643 if (ws)
679 __pm_stay_awake(ws); 644 __pm_stay_awake(ws);
680 } 645 }
681 646
682 static inline bool ep_has_wakeup_source(struct 647 static inline bool ep_has_wakeup_source(struct epitem *epi)
683 { 648 {
684 return rcu_access_pointer(epi->ws) ? t 649 return rcu_access_pointer(epi->ws) ? true : false;
685 } 650 }
686 651
687 /* call when ep->mtx cannot be held (ep_poll_c 652 /* call when ep->mtx cannot be held (ep_poll_callback) */
688 static inline void ep_pm_stay_awake_rcu(struct 653 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
689 { 654 {
690 struct wakeup_source *ws; 655 struct wakeup_source *ws;
691 656
692 rcu_read_lock(); 657 rcu_read_lock();
693 ws = rcu_dereference(epi->ws); 658 ws = rcu_dereference(epi->ws);
694 if (ws) 659 if (ws)
695 __pm_stay_awake(ws); 660 __pm_stay_awake(ws);
696 rcu_read_unlock(); 661 rcu_read_unlock();
697 } 662 }
698 663
699 !! 664 /**
700 /* !! 665 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
701 * ep->mutex needs to be held because we could !! 666 * the scan code, to call f_op->poll(). Also allows for
702 * eventpoll_release_file() and epoll_ctl(). !! 667 * O(NumReady) performance.
703 */ !! 668 *
704 static void ep_start_scan(struct eventpoll *ep !! 669 * @ep: Pointer to the epoll private data structure.
>> 670 * @sproc: Pointer to the scan callback.
>> 671 * @priv: Private opaque data passed to the @sproc callback.
>> 672 * @depth: The current depth of recursive f_op->poll calls.
>> 673 * @ep_locked: caller already holds ep->mtx
>> 674 *
>> 675 * Returns: The same integer error code returned by the @sproc callback.
>> 676 */
>> 677 static int ep_scan_ready_list(struct eventpoll *ep,
>> 678 int (*sproc)(struct eventpoll *,
>> 679 struct list_head *, void *),
>> 680 void *priv, int depth, bool ep_locked)
705 { 681 {
>> 682 int error, pwake = 0;
>> 683 unsigned long flags;
>> 684 struct epitem *epi, *nepi;
>> 685 LIST_HEAD(txlist);
>> 686
>> 687 /*
>> 688 * We need to lock this because we could be hit by
>> 689 * eventpoll_release_file() and epoll_ctl().
>> 690 */
>> 691
>> 692 if (!ep_locked)
>> 693 mutex_lock_nested(&ep->mtx, depth);
>> 694
706 /* 695 /*
707 * Steal the ready list, and re-init t 696 * Steal the ready list, and re-init the original one to the
708 * empty list. Also, set ep->ovflist t 697 * empty list. Also, set ep->ovflist to NULL so that events
709 * happening while looping w/out locks 698 * happening while looping w/out locks, are not lost. We cannot
710 * have the poll callback to queue dir 699 * have the poll callback to queue directly on ep->rdllist,
711 * because we want the "sproc" callbac 700 * because we want the "sproc" callback to be able to do it
712 * in a lockless way. 701 * in a lockless way.
713 */ 702 */
714 lockdep_assert_irqs_enabled(); !! 703 spin_lock_irqsave(&ep->lock, flags);
715 write_lock_irq(&ep->lock); !! 704 list_splice_init(&ep->rdllist, &txlist);
716 list_splice_init(&ep->rdllist, txlist) !! 705 ep->ovflist = NULL;
717 WRITE_ONCE(ep->ovflist, NULL); !! 706 spin_unlock_irqrestore(&ep->lock, flags);
718 write_unlock_irq(&ep->lock); <<
719 } <<
720 707
721 static void ep_done_scan(struct eventpoll *ep, !! 708 /*
722 struct list_head *txl !! 709 * Now call the callback function.
723 { !! 710 */
724 struct epitem *epi, *nepi; !! 711 error = (*sproc)(ep, &txlist, priv);
725 712
726 write_lock_irq(&ep->lock); !! 713 spin_lock_irqsave(&ep->lock, flags);
727 /* 714 /*
728 * During the time we spent inside the 715 * During the time we spent inside the "sproc" callback, some
729 * other events might have been queued 716 * other events might have been queued by the poll callback.
730 * We re-insert them inside the main r 717 * We re-insert them inside the main ready-list here.
731 */ 718 */
732 for (nepi = READ_ONCE(ep->ovflist); (e !! 719 for (nepi = ep->ovflist; (epi = nepi) != NULL;
733 nepi = epi->next, epi->next = EP_ 720 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
734 /* 721 /*
735 * We need to check if the ite 722 * We need to check if the item is already in the list.
736 * During the "sproc" callback 723 * During the "sproc" callback execution time, items are
737 * queued into ->ovflist but t 724 * queued into ->ovflist but the "txlist" might already
738 * contain them, and the list_ 725 * contain them, and the list_splice() below takes care of them.
739 */ 726 */
740 if (!ep_is_linked(epi)) { !! 727 if (!ep_is_linked(&epi->rdllink)) {
741 /* !! 728 list_add_tail(&epi->rdllink, &ep->rdllist);
742 * ->ovflist is LIFO, <<
743 * to keep in FIFO. <<
744 */ <<
745 list_add(&epi->rdllink <<
746 ep_pm_stay_awake(epi); 729 ep_pm_stay_awake(epi);
747 } 730 }
748 } 731 }
749 /* 732 /*
750 * We need to set back ep->ovflist to 733 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
751 * releasing the lock, events will be 734 * releasing the lock, events will be queued in the normal way inside
752 * ep->rdllist. 735 * ep->rdllist.
753 */ 736 */
754 WRITE_ONCE(ep->ovflist, EP_UNACTIVE_PT !! 737 ep->ovflist = EP_UNACTIVE_PTR;
755 738
756 /* 739 /*
757 * Quickly re-inject items left on "tx 740 * Quickly re-inject items left on "txlist".
758 */ 741 */
759 list_splice(txlist, &ep->rdllist); !! 742 list_splice(&txlist, &ep->rdllist);
760 __pm_relax(ep->ws); 743 __pm_relax(ep->ws);
761 744
762 if (!list_empty(&ep->rdllist)) { 745 if (!list_empty(&ep->rdllist)) {
>> 746 /*
>> 747 * Wake up (if active) both the eventpoll wait list and
>> 748 * the ->poll() wait list (delayed after we release the lock).
>> 749 */
763 if (waitqueue_active(&ep->wq)) 750 if (waitqueue_active(&ep->wq))
764 wake_up(&ep->wq); !! 751 wake_up_locked(&ep->wq);
>> 752 if (waitqueue_active(&ep->poll_wait))
>> 753 pwake++;
765 } 754 }
>> 755 spin_unlock_irqrestore(&ep->lock, flags);
766 756
767 write_unlock_irq(&ep->lock); !! 757 if (!ep_locked)
768 } !! 758 mutex_unlock(&ep->mtx);
769 <<
770 static void ep_get(struct eventpoll *ep) <<
771 { <<
772 refcount_inc(&ep->refcount); <<
773 } <<
774 759
775 /* !! 760 /* We have to call this outside the lock */
776 * Returns true if the event poll can be dispo !! 761 if (pwake)
777 */ !! 762 ep_poll_safewake(&ep->poll_wait);
778 static bool ep_refcount_dec_and_test(struct ev <<
779 { <<
780 if (!refcount_dec_and_test(&ep->refcou <<
781 return false; <<
782 763
783 WARN_ON_ONCE(!RB_EMPTY_ROOT(&ep->rbr.r !! 764 return error;
784 return true; <<
785 } 765 }
786 766
787 static void ep_free(struct eventpoll *ep) !! 767 static void epi_rcu_free(struct rcu_head *head)
788 { 768 {
789 mutex_destroy(&ep->mtx); !! 769 struct epitem *epi = container_of(head, struct epitem, rcu);
790 free_uid(ep->user); !! 770 kmem_cache_free(epi_cache, epi);
791 wakeup_source_unregister(ep->ws); <<
792 kfree(ep); <<
793 } 771 }
794 772
795 /* 773 /*
796 * Removes a "struct epitem" from the eventpol 774 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
797 * all the associated resources. Must be calle 775 * 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 */ 776 */
803 static bool __ep_remove(struct eventpoll *ep, !! 777 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
804 { 778 {
>> 779 unsigned long flags;
805 struct file *file = epi->ffd.file; 780 struct file *file = epi->ffd.file;
806 struct epitems_head *to_free; <<
807 struct hlist_head *head; <<
808 <<
809 lockdep_assert_irqs_enabled(); <<
810 781
811 /* 782 /*
812 * Removes poll wait queue hooks. !! 783 * Removes poll wait queue hooks. We _have_ to do this without holding
>> 784 * the "ep->lock" otherwise a deadlock might occur. This because of the
>> 785 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
>> 786 * queue head lock when unregistering the wait queue. The wakeup callback
>> 787 * will run by holding the wait queue head lock and will call our callback
>> 788 * that will try to get "ep->lock".
813 */ 789 */
814 ep_unregister_pollwait(ep, epi); 790 ep_unregister_pollwait(ep, epi);
815 791
816 /* Remove the current item from the li 792 /* Remove the current item from the list of epoll hooks */
817 spin_lock(&file->f_lock); 793 spin_lock(&file->f_lock);
818 if (epi->dying && !force) { !! 794 list_del_rcu(&epi->fllink);
819 spin_unlock(&file->f_lock); <<
820 return false; <<
821 } <<
822 <<
823 to_free = NULL; <<
824 head = file->f_ep; <<
825 if (head->first == &epi->fllink && !ep <<
826 file->f_ep = NULL; <<
827 if (!is_file_epoll(file)) { <<
828 struct epitems_head *v <<
829 v = container_of(head, <<
830 if (!smp_load_acquire( <<
831 to_free = v; <<
832 } <<
833 } <<
834 hlist_del_rcu(&epi->fllink); <<
835 spin_unlock(&file->f_lock); 795 spin_unlock(&file->f_lock);
836 free_ephead(to_free); <<
837 796
838 rb_erase_cached(&epi->rbn, &ep->rbr); 797 rb_erase_cached(&epi->rbn, &ep->rbr);
839 798
840 write_lock_irq(&ep->lock); !! 799 spin_lock_irqsave(&ep->lock, flags);
841 if (ep_is_linked(epi)) !! 800 if (ep_is_linked(&epi->rdllink))
842 list_del_init(&epi->rdllink); 801 list_del_init(&epi->rdllink);
843 write_unlock_irq(&ep->lock); !! 802 spin_unlock_irqrestore(&ep->lock, flags);
844 803
845 wakeup_source_unregister(ep_wakeup_sou 804 wakeup_source_unregister(ep_wakeup_source(epi));
846 /* 805 /*
847 * At this point it is safe to free th 806 * At this point it is safe to free the eventpoll item. Use the union
848 * field epi->rcu, since we are trying 807 * field epi->rcu, since we are trying to minimize the size of
849 * 'struct epitem'. The 'rbn' field is 808 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
850 * ep->mtx. The rcu read side, reverse 809 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
851 * use of the rbn field. 810 * use of the rbn field.
852 */ 811 */
853 kfree_rcu(epi, rcu); !! 812 call_rcu(&epi->rcu, epi_rcu_free);
854 813
855 percpu_counter_dec(&ep->user->epoll_wa !! 814 atomic_long_dec(&ep->user->epoll_watches);
856 return ep_refcount_dec_and_test(ep); <<
857 } <<
858 815
859 /* !! 816 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 } 817 }
866 818
867 static void ep_clear_and_put(struct eventpoll !! 819 static void ep_free(struct eventpoll *ep)
868 { 820 {
869 struct rb_node *rbp, *next; !! 821 struct rb_node *rbp;
870 struct epitem *epi; 822 struct epitem *epi;
871 bool dispose; <<
872 823
873 /* We need to release all tasks waitin 824 /* We need to release all tasks waiting for these file */
874 if (waitqueue_active(&ep->poll_wait)) 825 if (waitqueue_active(&ep->poll_wait))
875 ep_poll_safewake(ep, NULL, 0); !! 826 ep_poll_safewake(&ep->poll_wait);
876 827
877 mutex_lock(&ep->mtx); !! 828 /*
>> 829 * We need to lock this because we could be hit by
>> 830 * eventpoll_release_file() while we're freeing the "struct eventpoll".
>> 831 * We do not need to hold "ep->mtx" here because the epoll file
>> 832 * is on the way to be removed and no one has references to it
>> 833 * anymore. The only hit might come from eventpoll_release_file() but
>> 834 * holding "epmutex" is sufficient here.
>> 835 */
>> 836 mutex_lock(&epmutex);
878 837
879 /* 838 /*
880 * Walks through the whole tree by unr 839 * Walks through the whole tree by unregistering poll callbacks.
881 */ 840 */
882 for (rbp = rb_first_cached(&ep->rbr); 841 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
883 epi = rb_entry(rbp, struct epi 842 epi = rb_entry(rbp, struct epitem, rbn);
884 843
885 ep_unregister_pollwait(ep, epi 844 ep_unregister_pollwait(ep, epi);
886 cond_resched(); 845 cond_resched();
887 } 846 }
888 847
889 /* 848 /*
890 * Walks through the whole tree and tr !! 849 * Walks through the whole tree by freeing each "struct epitem". At this
891 * Note that ep_remove_safe() will not !! 850 * point we are sure no poll callbacks will be lingering around, and also by
892 * racing eventpoll_release_file(); th !! 851 * holding "epmutex" we can be sure that no file cleanup code will hit
893 * At this point we are sure no poll c !! 852 * us during this operation. So we can avoid the lock on "ep->lock".
894 * Since we still own a reference to t !! 853 * We do not need to lock ep->mtx, either, we only do it to prevent
895 * dispose it. !! 854 * a lockdep warning.
896 */ 855 */
897 for (rbp = rb_first_cached(&ep->rbr); !! 856 mutex_lock(&ep->mtx);
898 next = rb_next(rbp); !! 857 while ((rbp = rb_first_cached(&ep->rbr)) != NULL) {
899 epi = rb_entry(rbp, struct epi 858 epi = rb_entry(rbp, struct epitem, rbn);
900 ep_remove_safe(ep, epi); !! 859 ep_remove(ep, epi);
901 cond_resched(); 860 cond_resched();
902 } 861 }
903 <<
904 dispose = ep_refcount_dec_and_test(ep) <<
905 mutex_unlock(&ep->mtx); 862 mutex_unlock(&ep->mtx);
906 863
907 if (dispose) !! 864 mutex_unlock(&epmutex);
908 ep_free(ep); !! 865 mutex_destroy(&ep->mtx);
909 } !! 866 free_uid(ep->user);
910 !! 867 wakeup_source_unregister(ep->ws);
911 static long ep_eventpoll_ioctl(struct file *fi !! 868 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 } 869 }
931 870
932 static int ep_eventpoll_release(struct inode * 871 static int ep_eventpoll_release(struct inode *inode, struct file *file)
933 { 872 {
934 struct eventpoll *ep = file->private_d 873 struct eventpoll *ep = file->private_data;
935 874
936 if (ep) 875 if (ep)
937 ep_clear_and_put(ep); !! 876 ep_free(ep);
938 877
939 return 0; 878 return 0;
940 } 879 }
941 880
942 static __poll_t ep_item_poll(const struct epit !! 881 static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt)
>> 882 {
>> 883 pt->_key = epi->event.events;
>> 884
>> 885 return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events;
>> 886 }
943 887
944 static __poll_t __ep_eventpoll_poll(struct fil !! 888 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
>> 889 void *priv)
945 { 890 {
946 struct eventpoll *ep = file->private_d <<
947 LIST_HEAD(txlist); <<
948 struct epitem *epi, *tmp; 891 struct epitem *epi, *tmp;
949 poll_table pt; 892 poll_table pt;
950 __poll_t res = 0; <<
951 893
952 init_poll_funcptr(&pt, NULL); 894 init_poll_funcptr(&pt, NULL);
953 895
954 /* Insert inside our poll wait queue * !! 896 list_for_each_entry_safe(epi, tmp, head, rdllink) {
955 poll_wait(file, &ep->poll_wait, wait); !! 897 if (ep_item_poll(epi, &pt))
956 !! 898 return POLLIN | POLLRDNORM;
957 /* !! 899 else {
958 * Proceed to find out if wanted event <<
959 * the ready list. <<
960 */ <<
961 mutex_lock_nested(&ep->mtx, depth); <<
962 ep_start_scan(ep, &txlist); <<
963 list_for_each_entry_safe(epi, tmp, &tx <<
964 if (ep_item_poll(epi, &pt, dep <<
965 res = EPOLLIN | EPOLLR <<
966 break; <<
967 } else { <<
968 /* 900 /*
969 * Item has been dropp 901 * Item has been dropped into the ready list by the poll
970 * callback, but it's 902 * callback, but it's not actually ready, as far as
971 * caller requested ev 903 * caller requested events goes. We can remove it here.
972 */ 904 */
973 __pm_relax(ep_wakeup_s 905 __pm_relax(ep_wakeup_source(epi));
974 list_del_init(&epi->rd 906 list_del_init(&epi->rdllink);
975 } 907 }
976 } 908 }
977 ep_done_scan(ep, &txlist); !! 909
978 mutex_unlock(&ep->mtx); !! 910 return 0;
979 return res; <<
980 } 911 }
981 912
982 /* !! 913 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
983 * The ffd.file pointer may be in the process !! 914 poll_table *pt);
984 * being closed, but we may not have finished !! 915
985 * !! 916 struct readyevents_arg {
986 * Normally, even with the atomic_long_inc_not !! 917 struct eventpoll *ep;
987 * been free'd and then gotten re-allocated to !! 918 bool locked;
988 * files are not RCU-delayed, they are SLAB_TY !! 919 };
989 * !! 920
990 * But for epoll, users hold the ep->mtx mutex !! 921 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
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 { 922 {
1002 struct file *file; !! 923 struct readyevents_arg *arg = priv;
1003 924
1004 file = epi->ffd.file; !! 925 return ep_scan_ready_list(arg->ep, ep_read_events_proc, NULL,
1005 if (!atomic_long_inc_not_zero(&file-> !! 926 call_nests + 1, arg->locked);
1006 file = NULL; <<
1007 return file; <<
1008 } 927 }
1009 928
1010 /* !! 929 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
1011 * Differs from ep_eventpoll_poll() in that i <<
1012 * the ep->mtx so we need to start from depth <<
1013 * is correctly annotated. <<
1014 */ <<
1015 static __poll_t ep_item_poll(const struct epi <<
1016 int depth) <<
1017 { 930 {
1018 struct file *file = epi_fget(epi); !! 931 int pollflags;
1019 __poll_t res; !! 932 struct eventpoll *ep = file->private_data;
>> 933 struct readyevents_arg arg;
1020 934
1021 /* 935 /*
1022 * We could return EPOLLERR | EPOLLHU !! 936 * During ep_insert() we already hold the ep->mtx for the tfile.
1023 * treat this more as "file doesn't e !! 937 * Prevent re-aquisition.
1024 */ 938 */
1025 if (!file) !! 939 arg.locked = wait && (wait->_qproc == ep_ptable_queue_proc);
1026 return 0; !! 940 arg.ep = ep;
1027 941
1028 pt->_key = epi->event.events; !! 942 /* Insert inside our poll wait queue */
1029 if (!is_file_epoll(file)) !! 943 poll_wait(file, &ep->poll_wait, wait);
1030 res = vfs_poll(file, pt); <<
1031 else <<
1032 res = __ep_eventpoll_poll(fil <<
1033 fput(file); <<
1034 return res & epi->event.events; <<
1035 } <<
1036 944
1037 static __poll_t ep_eventpoll_poll(struct file !! 945 /*
1038 { !! 946 * Proceed to find out if wanted events are really available inside
1039 return __ep_eventpoll_poll(file, wait !! 947 * the ready list. This need to be done under ep_call_nested()
>> 948 * supervision, since the call to f_op->poll() done on listed files
>> 949 * could re-enter here.
>> 950 */
>> 951 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
>> 952 ep_poll_readyevents_proc, &arg, ep, current);
>> 953
>> 954 return pollflags != -1 ? pollflags : 0;
1040 } 955 }
1041 956
1042 #ifdef CONFIG_PROC_FS 957 #ifdef CONFIG_PROC_FS
1043 static void ep_show_fdinfo(struct seq_file *m 958 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
1044 { 959 {
1045 struct eventpoll *ep = f->private_dat 960 struct eventpoll *ep = f->private_data;
1046 struct rb_node *rbp; 961 struct rb_node *rbp;
1047 962
1048 mutex_lock(&ep->mtx); 963 mutex_lock(&ep->mtx);
1049 for (rbp = rb_first_cached(&ep->rbr); 964 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1050 struct epitem *epi = rb_entry 965 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
1051 struct inode *inode = file_in 966 struct inode *inode = file_inode(epi->ffd.file);
1052 967
1053 seq_printf(m, "tfd: %8d event 968 seq_printf(m, "tfd: %8d events: %8x data: %16llx "
1054 " pos:%lli ino:%lx 969 " pos:%lli ino:%lx sdev:%x\n",
1055 epi->ffd.fd, epi-> 970 epi->ffd.fd, epi->event.events,
1056 (long long)epi->ev 971 (long long)epi->event.data,
1057 (long long)epi->ff 972 (long long)epi->ffd.file->f_pos,
1058 inode->i_ino, inod 973 inode->i_ino, inode->i_sb->s_dev);
1059 if (seq_has_overflowed(m)) 974 if (seq_has_overflowed(m))
1060 break; 975 break;
1061 } 976 }
1062 mutex_unlock(&ep->mtx); 977 mutex_unlock(&ep->mtx);
1063 } 978 }
1064 #endif 979 #endif
1065 980
1066 /* File callbacks that implement the eventpol 981 /* File callbacks that implement the eventpoll file behaviour */
1067 static const struct file_operations eventpoll 982 static const struct file_operations eventpoll_fops = {
1068 #ifdef CONFIG_PROC_FS 983 #ifdef CONFIG_PROC_FS
1069 .show_fdinfo = ep_show_fdinfo, 984 .show_fdinfo = ep_show_fdinfo,
1070 #endif 985 #endif
1071 .release = ep_eventpoll_releas 986 .release = ep_eventpoll_release,
1072 .poll = ep_eventpoll_poll, 987 .poll = ep_eventpoll_poll,
1073 .llseek = noop_llseek, 988 .llseek = noop_llseek,
1074 .unlocked_ioctl = ep_eventpoll_ioctl, <<
1075 .compat_ioctl = compat_ptr_ioctl, <<
1076 }; 989 };
1077 990
1078 /* 991 /*
1079 * This is called from eventpoll_release() to 992 * This is called from eventpoll_release() to unlink files from the eventpoll
1080 * interface. We need to have this facility t 993 * interface. We need to have this facility to cleanup correctly files that are
1081 * closed without being removed from the even 994 * closed without being removed from the eventpoll interface.
1082 */ 995 */
1083 void eventpoll_release_file(struct file *file 996 void eventpoll_release_file(struct file *file)
1084 { 997 {
1085 struct eventpoll *ep; 998 struct eventpoll *ep;
1086 struct epitem *epi; !! 999 struct epitem *epi, *next;
1087 bool dispose; <<
1088 1000
1089 /* 1001 /*
1090 * Use the 'dying' flag to prevent a !! 1002 * We don't want to get "file->f_lock" because it is not
1091 * touching the epitems list before e !! 1003 * necessary. It is not necessary because we're in the "struct file"
1092 * the ep->mtx. !! 1004 * cleanup path, and this means that no one is using this file anymore.
>> 1005 * So, for example, epoll_ctl() cannot hit here since if we reach this
>> 1006 * point, the file counter already went to zero and fget() would fail.
>> 1007 * The only hit might come from ep_free() but by holding the mutex
>> 1008 * will correctly serialize the operation. We do need to acquire
>> 1009 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
>> 1010 * from anywhere but ep_free().
>> 1011 *
>> 1012 * Besides, ep_remove() acquires the lock, so we can't hold it here.
1093 */ 1013 */
1094 again: !! 1014 mutex_lock(&epmutex);
1095 spin_lock(&file->f_lock); !! 1015 list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
1096 if (file->f_ep && file->f_ep->first) <<
1097 epi = hlist_entry(file->f_ep- <<
1098 epi->dying = true; <<
1099 spin_unlock(&file->f_lock); <<
1100 <<
1101 /* <<
1102 * ep access is safe as we st <<
1103 * struct <<
1104 */ <<
1105 ep = epi->ep; 1016 ep = epi->ep;
1106 mutex_lock(&ep->mtx); !! 1017 mutex_lock_nested(&ep->mtx, 0);
1107 dispose = __ep_remove(ep, epi !! 1018 ep_remove(ep, epi);
1108 mutex_unlock(&ep->mtx); 1019 mutex_unlock(&ep->mtx);
1109 <<
1110 if (dispose) <<
1111 ep_free(ep); <<
1112 goto again; <<
1113 } 1020 }
1114 spin_unlock(&file->f_lock); !! 1021 mutex_unlock(&epmutex);
1115 } 1022 }
1116 1023
1117 static int ep_alloc(struct eventpoll **pep) 1024 static int ep_alloc(struct eventpoll **pep)
1118 { 1025 {
>> 1026 int error;
>> 1027 struct user_struct *user;
1119 struct eventpoll *ep; 1028 struct eventpoll *ep;
1120 1029
>> 1030 user = get_current_user();
>> 1031 error = -ENOMEM;
1121 ep = kzalloc(sizeof(*ep), GFP_KERNEL) 1032 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
1122 if (unlikely(!ep)) 1033 if (unlikely(!ep))
1123 return -ENOMEM; !! 1034 goto free_uid;
1124 1035
>> 1036 spin_lock_init(&ep->lock);
1125 mutex_init(&ep->mtx); 1037 mutex_init(&ep->mtx);
1126 rwlock_init(&ep->lock); <<
1127 init_waitqueue_head(&ep->wq); 1038 init_waitqueue_head(&ep->wq);
1128 init_waitqueue_head(&ep->poll_wait); 1039 init_waitqueue_head(&ep->poll_wait);
1129 INIT_LIST_HEAD(&ep->rdllist); 1040 INIT_LIST_HEAD(&ep->rdllist);
1130 ep->rbr = RB_ROOT_CACHED; 1041 ep->rbr = RB_ROOT_CACHED;
1131 ep->ovflist = EP_UNACTIVE_PTR; 1042 ep->ovflist = EP_UNACTIVE_PTR;
1132 ep->user = get_current_user(); !! 1043 ep->user = user;
1133 refcount_set(&ep->refcount, 1); <<
1134 1044
1135 *pep = ep; 1045 *pep = ep;
1136 1046
1137 return 0; 1047 return 0;
>> 1048
>> 1049 free_uid:
>> 1050 free_uid(user);
>> 1051 return error;
1138 } 1052 }
1139 1053
1140 /* 1054 /*
1141 * Search the file inside the eventpoll tree. 1055 * Search the file inside the eventpoll tree. The RB tree operations
1142 * are protected by the "mtx" mutex, and ep_f 1056 * are protected by the "mtx" mutex, and ep_find() must be called with
1143 * "mtx" held. 1057 * "mtx" held.
1144 */ 1058 */
1145 static struct epitem *ep_find(struct eventpol 1059 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
1146 { 1060 {
1147 int kcmp; 1061 int kcmp;
1148 struct rb_node *rbp; 1062 struct rb_node *rbp;
1149 struct epitem *epi, *epir = NULL; 1063 struct epitem *epi, *epir = NULL;
1150 struct epoll_filefd ffd; 1064 struct epoll_filefd ffd;
1151 1065
1152 ep_set_ffd(&ffd, file, fd); 1066 ep_set_ffd(&ffd, file, fd);
1153 for (rbp = ep->rbr.rb_root.rb_node; r 1067 for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {
1154 epi = rb_entry(rbp, struct ep 1068 epi = rb_entry(rbp, struct epitem, rbn);
1155 kcmp = ep_cmp_ffd(&ffd, &epi- 1069 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
1156 if (kcmp > 0) 1070 if (kcmp > 0)
1157 rbp = rbp->rb_right; 1071 rbp = rbp->rb_right;
1158 else if (kcmp < 0) 1072 else if (kcmp < 0)
1159 rbp = rbp->rb_left; 1073 rbp = rbp->rb_left;
1160 else { 1074 else {
1161 epir = epi; 1075 epir = epi;
1162 break; 1076 break;
1163 } 1077 }
1164 } 1078 }
1165 1079
1166 return epir; 1080 return epir;
1167 } 1081 }
1168 1082
1169 #ifdef CONFIG_KCMP !! 1083 #ifdef CONFIG_CHECKPOINT_RESTORE
1170 static struct epitem *ep_find_tfd(struct even 1084 static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
1171 { 1085 {
1172 struct rb_node *rbp; 1086 struct rb_node *rbp;
1173 struct epitem *epi; 1087 struct epitem *epi;
1174 1088
1175 for (rbp = rb_first_cached(&ep->rbr); 1089 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1176 epi = rb_entry(rbp, struct ep 1090 epi = rb_entry(rbp, struct epitem, rbn);
1177 if (epi->ffd.fd == tfd) { 1091 if (epi->ffd.fd == tfd) {
1178 if (toff == 0) 1092 if (toff == 0)
1179 return epi; 1093 return epi;
1180 else 1094 else
1181 toff--; 1095 toff--;
1182 } 1096 }
1183 cond_resched(); 1097 cond_resched();
1184 } 1098 }
1185 1099
1186 return NULL; 1100 return NULL;
1187 } 1101 }
1188 1102
1189 struct file *get_epoll_tfile_raw_ptr(struct f 1103 struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
1190 unsigned 1104 unsigned long toff)
1191 { 1105 {
1192 struct file *file_raw; 1106 struct file *file_raw;
1193 struct eventpoll *ep; 1107 struct eventpoll *ep;
1194 struct epitem *epi; 1108 struct epitem *epi;
1195 1109
1196 if (!is_file_epoll(file)) 1110 if (!is_file_epoll(file))
1197 return ERR_PTR(-EINVAL); 1111 return ERR_PTR(-EINVAL);
1198 1112
1199 ep = file->private_data; 1113 ep = file->private_data;
1200 1114
1201 mutex_lock(&ep->mtx); 1115 mutex_lock(&ep->mtx);
1202 epi = ep_find_tfd(ep, tfd, toff); 1116 epi = ep_find_tfd(ep, tfd, toff);
1203 if (epi) 1117 if (epi)
1204 file_raw = epi->ffd.file; 1118 file_raw = epi->ffd.file;
1205 else 1119 else
1206 file_raw = ERR_PTR(-ENOENT); 1120 file_raw = ERR_PTR(-ENOENT);
1207 mutex_unlock(&ep->mtx); 1121 mutex_unlock(&ep->mtx);
1208 1122
1209 return file_raw; 1123 return file_raw;
1210 } 1124 }
1211 #endif /* CONFIG_KCMP */ !! 1125 #endif /* CONFIG_CHECKPOINT_RESTORE */
1212 <<
1213 /* <<
1214 * Adds a new entry to the tail of the list i <<
1215 * multiple CPUs are allowed to call this fun <<
1216 * <<
1217 * Beware: it is necessary to prevent any oth <<
1218 * existing list until all changes ar <<
1219 * concurrent list_add_tail_lockless( <<
1220 * with a read lock, where write lock <<
1221 * makes sure all list_add_tail_lockl <<
1222 * completed. <<
1223 * <<
1224 * Also an element can be locklessly a <<
1225 * direction i.e. either to the tail o <<
1226 * concurrent access will corrupt the <<
1227 * <<
1228 * Return: %false if element has been already <<
1229 * otherwise. <<
1230 */ <<
1231 static inline bool list_add_tail_lockless(str <<
1232 str <<
1233 { <<
1234 struct list_head *prev; <<
1235 <<
1236 /* <<
1237 * This is simple 'new->next = head' <<
1238 * is used in order to detect that sa <<
1239 * added to the list from another CPU <<
1240 * new->next == new. <<
1241 */ <<
1242 if (!try_cmpxchg(&new->next, &new, he <<
1243 return false; <<
1244 <<
1245 /* <<
1246 * Initially ->next of a new element <<
1247 * (we are inserting to the tail) and <<
1248 * exchanged. XCHG guarantees memory <<
1249 * updated before pointers are actual <<
1250 * swapped before prev->next is updat <<
1251 */ <<
1252 <<
1253 prev = xchg(&head->prev, new); <<
1254 <<
1255 /* <<
1256 * It is safe to modify prev->next an <<
1257 * is added only to the tail and new- <<
1258 */ <<
1259 <<
1260 prev->next = new; <<
1261 new->prev = prev; <<
1262 <<
1263 return true; <<
1264 } <<
1265 <<
1266 /* <<
1267 * Chains a new epi entry to the tail of the <<
1268 * i.e. multiple CPUs are allowed to call thi <<
1269 * <<
1270 * Return: %false if epi element has been alr <<
1271 */ <<
1272 static inline bool chain_epi_lockless(struct <<
1273 { <<
1274 struct eventpoll *ep = epi->ep; <<
1275 <<
1276 /* Fast preliminary check */ <<
1277 if (epi->next != EP_UNACTIVE_PTR) <<
1278 return false; <<
1279 <<
1280 /* Check that the same epi has not be <<
1281 if (cmpxchg(&epi->next, EP_UNACTIVE_P <<
1282 return false; <<
1283 <<
1284 /* Atomically exchange tail */ <<
1285 epi->next = xchg(&ep->ovflist, epi); <<
1286 <<
1287 return true; <<
1288 } <<
1289 1126
1290 /* 1127 /*
1291 * This is the callback that is passed to the 1128 * This is the callback that is passed to the wait queue wakeup
1292 * mechanism. It is called by the stored file 1129 * mechanism. It is called by the stored file descriptors when they
1293 * have events to report. 1130 * have events to report.
1294 * <<
1295 * This callback takes a read lock in order n <<
1296 * events from another file descriptor, thus <<
1297 * or ->ovflist are lockless. Read lock is p <<
1298 * ep_start/done_scan(), which stops all list <<
1299 * that lists state is seen correctly. <<
1300 * <<
1301 * Another thing worth to mention is that ep_ <<
1302 * concurrently for the same @epi from differ <<
1303 * with several wait queues entries. Plural <<
1304 * single wait queue is serialized by wq.lock <<
1305 * queues are used should be detected accordi <<
1306 * cmpxchg() operation. <<
1307 */ 1131 */
1308 static int ep_poll_callback(wait_queue_entry_ 1132 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1309 { 1133 {
1310 int pwake = 0; 1134 int pwake = 0;
>> 1135 unsigned long flags;
1311 struct epitem *epi = ep_item_from_wai 1136 struct epitem *epi = ep_item_from_wait(wait);
1312 struct eventpoll *ep = epi->ep; 1137 struct eventpoll *ep = epi->ep;
1313 __poll_t pollflags = key_to_poll(key) <<
1314 unsigned long flags; <<
1315 int ewake = 0; 1138 int ewake = 0;
1316 1139
1317 read_lock_irqsave(&ep->lock, flags); !! 1140 spin_lock_irqsave(&ep->lock, flags);
1318 1141
1319 ep_set_busy_poll_napi_id(epi); 1142 ep_set_busy_poll_napi_id(epi);
1320 1143
1321 /* 1144 /*
1322 * If the event mask does not contain 1145 * If the event mask does not contain any poll(2) event, we consider the
1323 * descriptor to be disabled. This co 1146 * descriptor to be disabled. This condition is likely the effect of the
1324 * EPOLLONESHOT bit that disables the 1147 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1325 * until the next EPOLL_CTL_MOD will 1148 * until the next EPOLL_CTL_MOD will be issued.
1326 */ 1149 */
1327 if (!(epi->event.events & ~EP_PRIVATE 1150 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1328 goto out_unlock; 1151 goto out_unlock;
1329 1152
1330 /* 1153 /*
1331 * Check the events coming with the c 1154 * Check the events coming with the callback. At this stage, not
1332 * every device reports the events in 1155 * every device reports the events in the "key" parameter of the
1333 * callback. We need to be able to ha 1156 * callback. We need to be able to handle both cases here, hence the
1334 * test for "key" != NULL before the 1157 * test for "key" != NULL before the event match test.
1335 */ 1158 */
1336 if (pollflags && !(pollflags & epi->e !! 1159 if (key && !((unsigned long) key & epi->event.events))
1337 goto out_unlock; 1160 goto out_unlock;
1338 1161
1339 /* 1162 /*
1340 * If we are transferring events to u 1163 * If we are transferring events to userspace, we can hold no locks
1341 * (because we're accessing user memo 1164 * (because we're accessing user memory, and because of linux f_op->poll()
1342 * semantics). All the events that ha 1165 * semantics). All the events that happen during that period of time are
1343 * chained in ep->ovflist and requeue 1166 * chained in ep->ovflist and requeued later on.
1344 */ 1167 */
1345 if (READ_ONCE(ep->ovflist) != EP_UNAC !! 1168 if (ep->ovflist != EP_UNACTIVE_PTR) {
1346 if (chain_epi_lockless(epi)) !! 1169 if (epi->next == EP_UNACTIVE_PTR) {
1347 ep_pm_stay_awake_rcu( !! 1170 epi->next = ep->ovflist;
1348 } else if (!ep_is_linked(epi)) { !! 1171 ep->ovflist = epi;
1349 /* In the usual case, add eve !! 1172 if (epi->ws) {
1350 if (list_add_tail_lockless(&e !! 1173 /*
1351 ep_pm_stay_awake_rcu( !! 1174 * Activate ep->ws since epi->ws may get
>> 1175 * deactivated at any time.
>> 1176 */
>> 1177 __pm_stay_awake(ep->ws);
>> 1178 }
>> 1179
>> 1180 }
>> 1181 goto out_unlock;
>> 1182 }
>> 1183
>> 1184 /* If this file is already in the ready list we exit soon */
>> 1185 if (!ep_is_linked(&epi->rdllink)) {
>> 1186 list_add_tail(&epi->rdllink, &ep->rdllist);
>> 1187 ep_pm_stay_awake_rcu(epi);
1352 } 1188 }
1353 1189
1354 /* 1190 /*
1355 * Wake up ( if active ) both the eve 1191 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1356 * wait list. 1192 * wait list.
1357 */ 1193 */
1358 if (waitqueue_active(&ep->wq)) { 1194 if (waitqueue_active(&ep->wq)) {
1359 if ((epi->event.events & EPOL 1195 if ((epi->event.events & EPOLLEXCLUSIVE) &&
1360 !(pol !! 1196 !((unsigned long)key & POLLFREE)) {
1361 switch (pollflags & E !! 1197 switch ((unsigned long)key & EPOLLINOUT_BITS) {
1362 case EPOLLIN: !! 1198 case POLLIN:
1363 if (epi->even !! 1199 if (epi->event.events & POLLIN)
1364 ewake 1200 ewake = 1;
1365 break; 1201 break;
1366 case EPOLLOUT: !! 1202 case POLLOUT:
1367 if (epi->even !! 1203 if (epi->event.events & POLLOUT)
1368 ewake 1204 ewake = 1;
1369 break; 1205 break;
1370 case 0: 1206 case 0:
1371 ewake = 1; 1207 ewake = 1;
1372 break; 1208 break;
1373 } 1209 }
1374 } 1210 }
1375 wake_up(&ep->wq); !! 1211 wake_up_locked(&ep->wq);
1376 } 1212 }
1377 if (waitqueue_active(&ep->poll_wait)) 1213 if (waitqueue_active(&ep->poll_wait))
1378 pwake++; 1214 pwake++;
1379 1215
1380 out_unlock: 1216 out_unlock:
1381 read_unlock_irqrestore(&ep->lock, fla !! 1217 spin_unlock_irqrestore(&ep->lock, flags);
1382 1218
1383 /* We have to call this outside the l 1219 /* We have to call this outside the lock */
1384 if (pwake) 1220 if (pwake)
1385 ep_poll_safewake(ep, epi, pol !! 1221 ep_poll_safewake(&ep->poll_wait);
1386 1222
1387 if (!(epi->event.events & EPOLLEXCLUS 1223 if (!(epi->event.events & EPOLLEXCLUSIVE))
1388 ewake = 1; 1224 ewake = 1;
1389 1225
1390 if (pollflags & POLLFREE) { !! 1226 if ((unsigned long)key & POLLFREE) {
1391 /* 1227 /*
1392 * If we race with ep_remove_ 1228 * If we race with ep_remove_wait_queue() it can miss
1393 * ->whead = NULL and do anot 1229 * ->whead = NULL and do another remove_wait_queue() after
1394 * us, so we can't use __remo 1230 * us, so we can't use __remove_wait_queue().
1395 */ 1231 */
1396 list_del_init(&wait->entry); 1232 list_del_init(&wait->entry);
1397 /* 1233 /*
1398 * ->whead != NULL protects u !! 1234 * ->whead != NULL protects us from the race with ep_free()
1399 * ep_clear_and_put() or ep_r !! 1235 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1400 * takes whead->lock held by !! 1236 * held by the caller. Once we nullify it, nothing protects
1401 * nothing protects ep/epi or !! 1237 * ep/epi or even wait.
1402 */ 1238 */
1403 smp_store_release(&ep_pwq_fro 1239 smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1404 } 1240 }
1405 1241
1406 return ewake; 1242 return ewake;
1407 } 1243 }
1408 1244
1409 /* 1245 /*
1410 * This is the callback that is used to add o 1246 * This is the callback that is used to add our wait queue to the
1411 * target file wakeup lists. 1247 * target file wakeup lists.
1412 */ 1248 */
1413 static void ep_ptable_queue_proc(struct file 1249 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1414 poll_table * 1250 poll_table *pt)
1415 { 1251 {
1416 struct ep_pqueue *epq = container_of( !! 1252 struct epitem *epi = ep_item_from_epqueue(pt);
1417 struct epitem *epi = epq->epi; <<
1418 struct eppoll_entry *pwq; 1253 struct eppoll_entry *pwq;
1419 1254
1420 if (unlikely(!epi)) // an earlier !! 1255 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1421 return; !! 1256 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1422 !! 1257 pwq->whead = whead;
1423 pwq = kmem_cache_alloc(pwq_cache, GFP !! 1258 pwq->base = epi;
1424 if (unlikely(!pwq)) { !! 1259 if (epi->event.events & EPOLLEXCLUSIVE)
1425 epq->epi = NULL; !! 1260 add_wait_queue_exclusive(whead, &pwq->wait);
1426 return; !! 1261 else
>> 1262 add_wait_queue(whead, &pwq->wait);
>> 1263 list_add_tail(&pwq->llink, &epi->pwqlist);
>> 1264 epi->nwait++;
>> 1265 } else {
>> 1266 /* We have to signal that an error occurred */
>> 1267 epi->nwait = -1;
1427 } 1268 }
1428 <<
1429 init_waitqueue_func_entry(&pwq->wait, <<
1430 pwq->whead = whead; <<
1431 pwq->base = epi; <<
1432 if (epi->event.events & EPOLLEXCLUSIV <<
1433 add_wait_queue_exclusive(whea <<
1434 else <<
1435 add_wait_queue(whead, &pwq->w <<
1436 pwq->next = epi->pwqlist; <<
1437 epi->pwqlist = pwq; <<
1438 } 1269 }
1439 1270
1440 static void ep_rbtree_insert(struct eventpoll 1271 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1441 { 1272 {
1442 int kcmp; 1273 int kcmp;
1443 struct rb_node **p = &ep->rbr.rb_root 1274 struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;
1444 struct epitem *epic; 1275 struct epitem *epic;
1445 bool leftmost = true; 1276 bool leftmost = true;
1446 1277
1447 while (*p) { 1278 while (*p) {
1448 parent = *p; 1279 parent = *p;
1449 epic = rb_entry(parent, struc 1280 epic = rb_entry(parent, struct epitem, rbn);
1450 kcmp = ep_cmp_ffd(&epi->ffd, 1281 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1451 if (kcmp > 0) { 1282 if (kcmp > 0) {
1452 p = &parent->rb_right 1283 p = &parent->rb_right;
1453 leftmost = false; 1284 leftmost = false;
1454 } else 1285 } else
1455 p = &parent->rb_left; 1286 p = &parent->rb_left;
1456 } 1287 }
1457 rb_link_node(&epi->rbn, parent, p); 1288 rb_link_node(&epi->rbn, parent, p);
1458 rb_insert_color_cached(&epi->rbn, &ep 1289 rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);
1459 } 1290 }
1460 1291
1461 1292
1462 1293
1463 #define PATH_ARR_SIZE 5 1294 #define PATH_ARR_SIZE 5
1464 /* 1295 /*
1465 * These are the number paths of length 1 to 1296 * 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 1297 * from a single file of interest. For example, we allow 1000 paths of length
1467 * 1, to emanate from each file of interest. 1298 * 1, to emanate from each file of interest. This essentially represents the
1468 * potential wakeup paths, which need to be l 1299 * potential wakeup paths, which need to be limited in order to avoid massive
1469 * uncontrolled wakeup storms. The common use 1300 * uncontrolled wakeup storms. The common use case should be a single ep which
1470 * is connected to n file sources. In this ca 1301 * is connected to n file sources. In this case each file source has 1 path
1471 * of length 1. Thus, the numbers below shoul 1302 * of length 1. Thus, the numbers below should be more than sufficient. These
1472 * path limits are enforced during an EPOLL_C 1303 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1473 * and delete can't add additional paths. Pro !! 1304 * and delete can't add additional paths. Protected by the epmutex.
1474 */ 1305 */
1475 static const int path_limits[PATH_ARR_SIZE] = 1306 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1476 static int path_count[PATH_ARR_SIZE]; 1307 static int path_count[PATH_ARR_SIZE];
1477 1308
1478 static int path_count_inc(int nests) 1309 static int path_count_inc(int nests)
1479 { 1310 {
1480 /* Allow an arbitrary number of depth 1311 /* Allow an arbitrary number of depth 1 paths */
1481 if (nests == 0) 1312 if (nests == 0)
1482 return 0; 1313 return 0;
1483 1314
1484 if (++path_count[nests] > path_limits 1315 if (++path_count[nests] > path_limits[nests])
1485 return -1; 1316 return -1;
1486 return 0; 1317 return 0;
1487 } 1318 }
1488 1319
1489 static void path_count_init(void) 1320 static void path_count_init(void)
1490 { 1321 {
1491 int i; 1322 int i;
1492 1323
1493 for (i = 0; i < PATH_ARR_SIZE; i++) 1324 for (i = 0; i < PATH_ARR_SIZE; i++)
1494 path_count[i] = 0; 1325 path_count[i] = 0;
1495 } 1326 }
1496 1327
1497 static int reverse_path_check_proc(struct hli !! 1328 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1498 { 1329 {
1499 int error = 0; 1330 int error = 0;
>> 1331 struct file *file = priv;
>> 1332 struct file *child_file;
1500 struct epitem *epi; 1333 struct epitem *epi;
1501 1334
1502 if (depth > EP_MAX_NESTS) /* too deep <<
1503 return -1; <<
1504 <<
1505 /* CTL_DEL can remove links here, but 1335 /* CTL_DEL can remove links here, but that can't increase our count */
1506 hlist_for_each_entry_rcu(epi, refs, f !! 1336 rcu_read_lock();
1507 struct hlist_head *refs = &ep !! 1337 list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1508 if (hlist_empty(refs)) !! 1338 child_file = epi->ep->file;
1509 error = path_count_in !! 1339 if (is_file_epoll(child_file)) {
1510 else !! 1340 if (list_empty(&child_file->f_ep_links)) {
1511 error = reverse_path_ !! 1341 if (path_count_inc(call_nests)) {
1512 if (error != 0) !! 1342 error = -1;
1513 break; !! 1343 break;
>> 1344 }
>> 1345 } else {
>> 1346 error = ep_call_nested(&poll_loop_ncalls,
>> 1347 EP_MAX_NESTS,
>> 1348 reverse_path_check_proc,
>> 1349 child_file, child_file,
>> 1350 current);
>> 1351 }
>> 1352 if (error != 0)
>> 1353 break;
>> 1354 } else {
>> 1355 printk(KERN_ERR "reverse_path_check_proc: "
>> 1356 "file is not an ep!\n");
>> 1357 }
1514 } 1358 }
>> 1359 rcu_read_unlock();
1515 return error; 1360 return error;
1516 } 1361 }
1517 1362
1518 /** 1363 /**
1519 * reverse_path_check - The tfile_check_list !! 1364 * reverse_path_check - The tfile_check_list is list of file *, which have
1520 * links that are propos 1365 * links that are proposed to be newly added. We need to
1521 * make sure that those 1366 * make sure that those added links don't add too many
1522 * paths such that we wi 1367 * paths such that we will spend all our time waking up
1523 * eventpoll objects. 1368 * eventpoll objects.
1524 * 1369 *
1525 * Return: %zero if the proposed links don't !! 1370 * Returns: Returns zero if the proposed links don't create too many paths,
1526 * %-1 otherwise. !! 1371 * -1 otherwise.
1527 */ 1372 */
1528 static int reverse_path_check(void) 1373 static int reverse_path_check(void)
1529 { 1374 {
1530 struct epitems_head *p; !! 1375 int error = 0;
>> 1376 struct file *current_file;
1531 1377
1532 for (p = tfile_check_list; p != EP_UN !! 1378 /* let's call this for all tfiles */
1533 int error; !! 1379 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1534 path_count_init(); 1380 path_count_init();
1535 rcu_read_lock(); !! 1381 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1536 error = reverse_path_check_pr !! 1382 reverse_path_check_proc, current_file,
1537 rcu_read_unlock(); !! 1383 current_file, current);
1538 if (error) 1384 if (error)
1539 return error; !! 1385 break;
1540 } 1386 }
1541 return 0; !! 1387 return error;
1542 } 1388 }
1543 1389
1544 static int ep_create_wakeup_source(struct epi 1390 static int ep_create_wakeup_source(struct epitem *epi)
1545 { 1391 {
1546 struct name_snapshot n; 1392 struct name_snapshot n;
1547 struct wakeup_source *ws; 1393 struct wakeup_source *ws;
1548 1394
1549 if (!epi->ep->ws) { 1395 if (!epi->ep->ws) {
1550 epi->ep->ws = wakeup_source_r !! 1396 epi->ep->ws = wakeup_source_register("eventpoll");
1551 if (!epi->ep->ws) 1397 if (!epi->ep->ws)
1552 return -ENOMEM; 1398 return -ENOMEM;
1553 } 1399 }
1554 1400
1555 take_dentry_name_snapshot(&n, epi->ff 1401 take_dentry_name_snapshot(&n, epi->ffd.file->f_path.dentry);
1556 ws = wakeup_source_register(NULL, n.n !! 1402 ws = wakeup_source_register(n.name);
1557 release_dentry_name_snapshot(&n); 1403 release_dentry_name_snapshot(&n);
1558 1404
1559 if (!ws) 1405 if (!ws)
1560 return -ENOMEM; 1406 return -ENOMEM;
1561 rcu_assign_pointer(epi->ws, ws); 1407 rcu_assign_pointer(epi->ws, ws);
1562 1408
1563 return 0; 1409 return 0;
1564 } 1410 }
1565 1411
1566 /* rare code path, only used when EPOLL_CTL_M 1412 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1567 static noinline void ep_destroy_wakeup_source 1413 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1568 { 1414 {
1569 struct wakeup_source *ws = ep_wakeup_ 1415 struct wakeup_source *ws = ep_wakeup_source(epi);
1570 1416
1571 RCU_INIT_POINTER(epi->ws, NULL); 1417 RCU_INIT_POINTER(epi->ws, NULL);
1572 1418
1573 /* 1419 /*
1574 * wait for ep_pm_stay_awake_rcu to f 1420 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1575 * used internally by wakeup_source_r 1421 * used internally by wakeup_source_remove, too (called by
1576 * wakeup_source_unregister), so we c 1422 * wakeup_source_unregister), so we cannot use call_rcu
1577 */ 1423 */
1578 synchronize_rcu(); 1424 synchronize_rcu();
1579 wakeup_source_unregister(ws); 1425 wakeup_source_unregister(ws);
1580 } 1426 }
1581 1427
1582 static int attach_epitem(struct file *file, s <<
1583 { <<
1584 struct epitems_head *to_free = NULL; <<
1585 struct hlist_head *head = NULL; <<
1586 struct eventpoll *ep = NULL; <<
1587 <<
1588 if (is_file_epoll(file)) <<
1589 ep = file->private_data; <<
1590 <<
1591 if (ep) { <<
1592 head = &ep->refs; <<
1593 } else if (!READ_ONCE(file->f_ep)) { <<
1594 allocate: <<
1595 to_free = kmem_cache_zalloc(e <<
1596 if (!to_free) <<
1597 return -ENOMEM; <<
1598 head = &to_free->epitems; <<
1599 } <<
1600 spin_lock(&file->f_lock); <<
1601 if (!file->f_ep) { <<
1602 if (unlikely(!head)) { <<
1603 spin_unlock(&file->f_ <<
1604 goto allocate; <<
1605 } <<
1606 file->f_ep = head; <<
1607 to_free = NULL; <<
1608 } <<
1609 hlist_add_head_rcu(&epi->fllink, file <<
1610 spin_unlock(&file->f_lock); <<
1611 free_ephead(to_free); <<
1612 return 0; <<
1613 } <<
1614 <<
1615 /* 1428 /*
1616 * Must be called with "mtx" held. 1429 * Must be called with "mtx" held.
1617 */ 1430 */
1618 static int ep_insert(struct eventpoll *ep, co !! 1431 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1619 struct file *tfile, int 1432 struct file *tfile, int fd, int full_check)
1620 { 1433 {
1621 int error, pwake = 0; !! 1434 int error, revents, pwake = 0;
1622 __poll_t revents; !! 1435 unsigned long flags;
>> 1436 long user_watches;
1623 struct epitem *epi; 1437 struct epitem *epi;
1624 struct ep_pqueue epq; 1438 struct ep_pqueue epq;
1625 struct eventpoll *tep = NULL; <<
1626 <<
1627 if (is_file_epoll(tfile)) <<
1628 tep = tfile->private_data; <<
1629 1439
1630 lockdep_assert_irqs_enabled(); !! 1440 user_watches = atomic_long_read(&ep->user->epoll_watches);
1631 !! 1441 if (unlikely(user_watches >= max_user_watches))
1632 if (unlikely(percpu_counter_compare(& <<
1633 m <<
1634 return -ENOSPC; 1442 return -ENOSPC;
1635 percpu_counter_inc(&ep->user->epoll_w !! 1443 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1636 <<
1637 if (!(epi = kmem_cache_zalloc(epi_cac <<
1638 percpu_counter_dec(&ep->user- <<
1639 return -ENOMEM; 1444 return -ENOMEM;
1640 } <<
1641 1445
1642 /* Item initialization follow here .. 1446 /* Item initialization follow here ... */
1643 INIT_LIST_HEAD(&epi->rdllink); 1447 INIT_LIST_HEAD(&epi->rdllink);
>> 1448 INIT_LIST_HEAD(&epi->fllink);
>> 1449 INIT_LIST_HEAD(&epi->pwqlist);
1644 epi->ep = ep; 1450 epi->ep = ep;
1645 ep_set_ffd(&epi->ffd, tfile, fd); 1451 ep_set_ffd(&epi->ffd, tfile, fd);
1646 epi->event = *event; 1452 epi->event = *event;
>> 1453 epi->nwait = 0;
1647 epi->next = EP_UNACTIVE_PTR; 1454 epi->next = EP_UNACTIVE_PTR;
1648 !! 1455 if (epi->event.events & EPOLLWAKEUP) {
1649 if (tep) !! 1456 error = ep_create_wakeup_source(epi);
1650 mutex_lock_nested(&tep->mtx, !! 1457 if (error)
1651 /* Add the current item to the list o !! 1458 goto error_create_wakeup_source;
1652 if (unlikely(attach_epitem(tfile, epi !! 1459 } else {
1653 if (tep) !! 1460 RCU_INIT_POINTER(epi->ws, NULL);
1654 mutex_unlock(&tep->mt <<
1655 kmem_cache_free(epi_cache, ep <<
1656 percpu_counter_dec(&ep->user- <<
1657 return -ENOMEM; <<
1658 } 1461 }
1659 1462
1660 if (full_check && !tep) !! 1463 /* Add the current item to the list of active epoll hook for this file */
1661 list_file(tfile); !! 1464 spin_lock(&tfile->f_lock);
>> 1465 list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
>> 1466 spin_unlock(&tfile->f_lock);
1662 1467
1663 /* 1468 /*
1664 * Add the current item to the RB tre 1469 * Add the current item to the RB tree. All RB tree operations are
1665 * protected by "mtx", and ep_insert( 1470 * protected by "mtx", and ep_insert() is called with "mtx" held.
1666 */ 1471 */
1667 ep_rbtree_insert(ep, epi); 1472 ep_rbtree_insert(ep, epi);
1668 if (tep) <<
1669 mutex_unlock(&tep->mtx); <<
1670 <<
1671 /* <<
1672 * ep_remove_safe() calls in the late <<
1673 * ep_free() as the ep file itself st <<
1674 */ <<
1675 ep_get(ep); <<
1676 1473
1677 /* now check if we've created too man 1474 /* now check if we've created too many backpaths */
1678 if (unlikely(full_check && reverse_pa !! 1475 error = -EINVAL;
1679 ep_remove_safe(ep, epi); !! 1476 if (full_check && reverse_path_check())
1680 return -EINVAL; !! 1477 goto error_remove_epi;
1681 } <<
1682 <<
1683 if (epi->event.events & EPOLLWAKEUP) <<
1684 error = ep_create_wakeup_sour <<
1685 if (error) { <<
1686 ep_remove_safe(ep, ep <<
1687 return error; <<
1688 } <<
1689 } <<
1690 1478
1691 /* Initialize the poll table using th 1479 /* Initialize the poll table using the queue callback */
1692 epq.epi = epi; 1480 epq.epi = epi;
1693 init_poll_funcptr(&epq.pt, ep_ptable_ 1481 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1694 1482
1695 /* 1483 /*
1696 * Attach the item to the poll hooks 1484 * Attach the item to the poll hooks and get current event bits.
1697 * We can safely use the file* here b 1485 * We can safely use the file* here because its usage count has
1698 * been increased by the caller of th 1486 * been increased by the caller of this function. Note that after
1699 * this operation completes, the poll 1487 * this operation completes, the poll callback can start hitting
1700 * the new item. 1488 * the new item.
1701 */ 1489 */
1702 revents = ep_item_poll(epi, &epq.pt, !! 1490 revents = ep_item_poll(epi, &epq.pt);
1703 1491
1704 /* 1492 /*
1705 * We have to check if something went 1493 * We have to check if something went wrong during the poll wait queue
1706 * install process. Namely an allocat 1494 * install process. Namely an allocation for a wait queue failed due
1707 * high memory pressure. 1495 * high memory pressure.
1708 */ 1496 */
1709 if (unlikely(!epq.epi)) { !! 1497 error = -ENOMEM;
1710 ep_remove_safe(ep, epi); !! 1498 if (epi->nwait < 0)
1711 return -ENOMEM; !! 1499 goto error_unregister;
1712 } <<
1713 1500
1714 /* We have to drop the new item insid 1501 /* We have to drop the new item inside our item list to keep track of it */
1715 write_lock_irq(&ep->lock); !! 1502 spin_lock_irqsave(&ep->lock, flags);
1716 1503
1717 /* record NAPI ID of new item if pres 1504 /* record NAPI ID of new item if present */
1718 ep_set_busy_poll_napi_id(epi); 1505 ep_set_busy_poll_napi_id(epi);
1719 1506
1720 /* If the file is already "ready" we 1507 /* If the file is already "ready" we drop it inside the ready list */
1721 if (revents && !ep_is_linked(epi)) { !! 1508 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
1722 list_add_tail(&epi->rdllink, 1509 list_add_tail(&epi->rdllink, &ep->rdllist);
1723 ep_pm_stay_awake(epi); 1510 ep_pm_stay_awake(epi);
1724 1511
1725 /* Notify waiting tasks that 1512 /* Notify waiting tasks that events are available */
1726 if (waitqueue_active(&ep->wq) 1513 if (waitqueue_active(&ep->wq))
1727 wake_up(&ep->wq); !! 1514 wake_up_locked(&ep->wq);
1728 if (waitqueue_active(&ep->pol 1515 if (waitqueue_active(&ep->poll_wait))
1729 pwake++; 1516 pwake++;
1730 } 1517 }
1731 1518
1732 write_unlock_irq(&ep->lock); !! 1519 spin_unlock_irqrestore(&ep->lock, flags);
>> 1520
>> 1521 atomic_long_inc(&ep->user->epoll_watches);
1733 1522
1734 /* We have to call this outside the l 1523 /* We have to call this outside the lock */
1735 if (pwake) 1524 if (pwake)
1736 ep_poll_safewake(ep, NULL, 0) !! 1525 ep_poll_safewake(&ep->poll_wait);
1737 1526
1738 return 0; 1527 return 0;
>> 1528
>> 1529 error_unregister:
>> 1530 ep_unregister_pollwait(ep, epi);
>> 1531 error_remove_epi:
>> 1532 spin_lock(&tfile->f_lock);
>> 1533 list_del_rcu(&epi->fllink);
>> 1534 spin_unlock(&tfile->f_lock);
>> 1535
>> 1536 rb_erase_cached(&epi->rbn, &ep->rbr);
>> 1537
>> 1538 /*
>> 1539 * We need to do this because an event could have been arrived on some
>> 1540 * allocated wait queue. Note that we don't care about the ep->ovflist
>> 1541 * list, since that is used/cleaned only inside a section bound by "mtx".
>> 1542 * And ep_insert() is called with "mtx" held.
>> 1543 */
>> 1544 spin_lock_irqsave(&ep->lock, flags);
>> 1545 if (ep_is_linked(&epi->rdllink))
>> 1546 list_del_init(&epi->rdllink);
>> 1547 spin_unlock_irqrestore(&ep->lock, flags);
>> 1548
>> 1549 wakeup_source_unregister(ep_wakeup_source(epi));
>> 1550
>> 1551 error_create_wakeup_source:
>> 1552 kmem_cache_free(epi_cache, epi);
>> 1553
>> 1554 return error;
1739 } 1555 }
1740 1556
1741 /* 1557 /*
1742 * Modify the interest event mask by dropping 1558 * Modify the interest event mask by dropping an event if the new mask
1743 * has a match in the current file status. Mu 1559 * has a match in the current file status. Must be called with "mtx" held.
1744 */ 1560 */
1745 static int ep_modify(struct eventpoll *ep, st !! 1561 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1746 const struct epoll_event <<
1747 { 1562 {
1748 int pwake = 0; 1563 int pwake = 0;
>> 1564 unsigned int revents;
1749 poll_table pt; 1565 poll_table pt;
1750 1566
1751 lockdep_assert_irqs_enabled(); <<
1752 <<
1753 init_poll_funcptr(&pt, NULL); 1567 init_poll_funcptr(&pt, NULL);
1754 1568
1755 /* 1569 /*
1756 * Set the new event interest mask be 1570 * Set the new event interest mask before calling f_op->poll();
1757 * otherwise we might miss an event t 1571 * otherwise we might miss an event that happens between the
1758 * f_op->poll() call and the new even 1572 * f_op->poll() call and the new event set registering.
1759 */ 1573 */
1760 epi->event.events = event->events; /* 1574 epi->event.events = event->events; /* need barrier below */
1761 epi->event.data = event->data; /* pro 1575 epi->event.data = event->data; /* protected by mtx */
1762 if (epi->event.events & EPOLLWAKEUP) 1576 if (epi->event.events & EPOLLWAKEUP) {
1763 if (!ep_has_wakeup_source(epi 1577 if (!ep_has_wakeup_source(epi))
1764 ep_create_wakeup_sour 1578 ep_create_wakeup_source(epi);
1765 } else if (ep_has_wakeup_source(epi)) 1579 } else if (ep_has_wakeup_source(epi)) {
1766 ep_destroy_wakeup_source(epi) 1580 ep_destroy_wakeup_source(epi);
1767 } 1581 }
1768 1582
1769 /* 1583 /*
1770 * The following barrier has two effe 1584 * The following barrier has two effects:
1771 * 1585 *
1772 * 1) Flush epi changes above to othe 1586 * 1) Flush epi changes above to other CPUs. This ensures
1773 * we do not miss events from ep_p 1587 * we do not miss events from ep_poll_callback if an
1774 * event occurs immediately after 1588 * event occurs immediately after we call f_op->poll().
1775 * We need this because we did not 1589 * We need this because we did not take ep->lock while
1776 * changing epi above (but ep_poll 1590 * changing epi above (but ep_poll_callback does take
1777 * ep->lock). 1591 * ep->lock).
1778 * 1592 *
1779 * 2) We also need to ensure we do no 1593 * 2) We also need to ensure we do not miss _past_ events
1780 * when calling f_op->poll(). Thi 1594 * when calling f_op->poll(). This barrier also
1781 * pairs with the barrier in wq_ha 1595 * pairs with the barrier in wq_has_sleeper (see
1782 * comments for wq_has_sleeper). 1596 * comments for wq_has_sleeper).
1783 * 1597 *
1784 * This barrier will now guarantee ep 1598 * This barrier will now guarantee ep_poll_callback or f_op->poll
1785 * (or both) will notice the readines 1599 * (or both) will notice the readiness of an item.
1786 */ 1600 */
1787 smp_mb(); 1601 smp_mb();
1788 1602
1789 /* 1603 /*
1790 * Get current event bits. We can saf 1604 * Get current event bits. We can safely use the file* here because
1791 * its usage count has been increased 1605 * its usage count has been increased by the caller of this function.
>> 1606 */
>> 1607 revents = ep_item_poll(epi, &pt);
>> 1608
>> 1609 /*
1792 * If the item is "hot" and it is not 1610 * If the item is "hot" and it is not registered inside the ready
1793 * list, push it inside. 1611 * list, push it inside.
1794 */ 1612 */
1795 if (ep_item_poll(epi, &pt, 1)) { !! 1613 if (revents & event->events) {
1796 write_lock_irq(&ep->lock); !! 1614 spin_lock_irq(&ep->lock);
1797 if (!ep_is_linked(epi)) { !! 1615 if (!ep_is_linked(&epi->rdllink)) {
1798 list_add_tail(&epi->r 1616 list_add_tail(&epi->rdllink, &ep->rdllist);
1799 ep_pm_stay_awake(epi) 1617 ep_pm_stay_awake(epi);
1800 1618
1801 /* Notify waiting tas 1619 /* Notify waiting tasks that events are available */
1802 if (waitqueue_active( 1620 if (waitqueue_active(&ep->wq))
1803 wake_up(&ep-> !! 1621 wake_up_locked(&ep->wq);
1804 if (waitqueue_active( 1622 if (waitqueue_active(&ep->poll_wait))
1805 pwake++; 1623 pwake++;
1806 } 1624 }
1807 write_unlock_irq(&ep->lock); !! 1625 spin_unlock_irq(&ep->lock);
1808 } 1626 }
1809 1627
1810 /* We have to call this outside the l 1628 /* We have to call this outside the lock */
1811 if (pwake) 1629 if (pwake)
1812 ep_poll_safewake(ep, NULL, 0) !! 1630 ep_poll_safewake(&ep->poll_wait);
1813 1631
1814 return 0; 1632 return 0;
1815 } 1633 }
1816 1634
1817 static int ep_send_events(struct eventpoll *e !! 1635 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1818 struct epoll_event !! 1636 void *priv)
1819 { 1637 {
1820 struct epitem *epi, *tmp; !! 1638 struct ep_send_events_data *esed = priv;
1821 LIST_HEAD(txlist); !! 1639 int eventcnt;
>> 1640 unsigned int revents;
>> 1641 struct epitem *epi;
>> 1642 struct epoll_event __user *uevent;
>> 1643 struct wakeup_source *ws;
1822 poll_table pt; 1644 poll_table pt;
1823 int res = 0; <<
1824 <<
1825 /* <<
1826 * Always short-circuit for fatal sig <<
1827 * timely exit without the chance of <<
1828 * fetching repeatedly. <<
1829 */ <<
1830 if (fatal_signal_pending(current)) <<
1831 return -EINTR; <<
1832 1645
1833 init_poll_funcptr(&pt, NULL); 1646 init_poll_funcptr(&pt, NULL);
1834 1647
1835 mutex_lock(&ep->mtx); <<
1836 ep_start_scan(ep, &txlist); <<
1837 <<
1838 /* 1648 /*
1839 * We can loop without lock because w 1649 * We can loop without lock because we are passed a task private list.
1840 * Items cannot vanish during the loo !! 1650 * Items cannot vanish during the loop because ep_scan_ready_list() is
>> 1651 * holding "mtx" during this call.
1841 */ 1652 */
1842 list_for_each_entry_safe(epi, tmp, &t !! 1653 for (eventcnt = 0, uevent = esed->events;
1843 struct wakeup_source *ws; !! 1654 !list_empty(head) && eventcnt < esed->maxevents;) {
1844 __poll_t revents; !! 1655 epi = list_first_entry(head, struct epitem, rdllink);
1845 <<
1846 if (res >= maxevents) <<
1847 break; <<
1848 1656
1849 /* 1657 /*
1850 * Activate ep->ws before dea 1658 * Activate ep->ws before deactivating epi->ws to prevent
1851 * triggering auto-suspend he 1659 * triggering auto-suspend here (in case we reactive epi->ws
1852 * below). 1660 * below).
1853 * 1661 *
1854 * This could be rearranged t 1662 * This could be rearranged to delay the deactivation of epi->ws
1855 * instead, but then epi->ws 1663 * instead, but then epi->ws would temporarily be out of sync
1856 * with ep_is_linked(). 1664 * with ep_is_linked().
1857 */ 1665 */
1858 ws = ep_wakeup_source(epi); 1666 ws = ep_wakeup_source(epi);
1859 if (ws) { 1667 if (ws) {
1860 if (ws->active) 1668 if (ws->active)
1861 __pm_stay_awa 1669 __pm_stay_awake(ep->ws);
1862 __pm_relax(ws); 1670 __pm_relax(ws);
1863 } 1671 }
1864 1672
1865 list_del_init(&epi->rdllink); 1673 list_del_init(&epi->rdllink);
1866 1674
>> 1675 revents = ep_item_poll(epi, &pt);
>> 1676
1867 /* 1677 /*
1868 * If the event mask intersec 1678 * If the event mask intersect the caller-requested one,
1869 * deliver the event to users !! 1679 * deliver the event to userspace. Again, ep_scan_ready_list()
1870 * so no operations coming fr !! 1680 * is holding "mtx", so no operations coming from userspace
>> 1681 * can change the item.
1871 */ 1682 */
1872 revents = ep_item_poll(epi, & !! 1683 if (revents) {
1873 if (!revents) !! 1684 if (__put_user(revents, &uevent->events) ||
1874 continue; !! 1685 __put_user(epi->event.data, &uevent->data)) {
1875 !! 1686 list_add(&epi->rdllink, head);
1876 events = epoll_put_uevent(rev !! 1687 ep_pm_stay_awake(epi);
1877 if (!events) { !! 1688 return eventcnt ? eventcnt : -EFAULT;
1878 list_add(&epi->rdllin !! 1689 }
1879 ep_pm_stay_awake(epi) !! 1690 eventcnt++;
1880 if (!res) !! 1691 uevent++;
1881 res = -EFAULT !! 1692 if (epi->event.events & EPOLLONESHOT)
1882 break; !! 1693 epi->event.events &= EP_PRIVATE_BITS;
1883 } !! 1694 else if (!(epi->event.events & EPOLLET)) {
1884 res++; !! 1695 /*
1885 if (epi->event.events & EPOLL !! 1696 * If this file has been added with Level
1886 epi->event.events &= !! 1697 * Trigger mode, we need to insert back inside
1887 else if (!(epi->event.events !! 1698 * the ready list, so that the next call to
1888 /* !! 1699 * epoll_wait() will check again the events
1889 * If this file has b !! 1700 * availability. At this point, no one can insert
1890 * Trigger mode, we n !! 1701 * into ep->rdllist besides us. The epoll_ctl()
1891 * the ready list, so !! 1702 * callers are locked out by
1892 * epoll_wait() will !! 1703 * ep_scan_ready_list() holding "mtx" and the
1893 * availability. At t !! 1704 * poll callback will queue them in ep->ovflist.
1894 * into ep->rdllist b !! 1705 */
1895 * callers are locked !! 1706 list_add_tail(&epi->rdllink, &ep->rdllist);
1896 * ep_send_events() h !! 1707 ep_pm_stay_awake(epi);
1897 * poll callback will !! 1708 }
1898 */ <<
1899 list_add_tail(&epi->r <<
1900 ep_pm_stay_awake(epi) <<
1901 } 1709 }
1902 } 1710 }
1903 ep_done_scan(ep, &txlist); <<
1904 mutex_unlock(&ep->mtx); <<
1905 1711
1906 return res; !! 1712 return eventcnt;
1907 } 1713 }
1908 1714
1909 static struct timespec64 *ep_timeout_to_times !! 1715 static int ep_send_events(struct eventpoll *ep,
>> 1716 struct epoll_event __user *events, int maxevents)
1910 { 1717 {
1911 struct timespec64 now; !! 1718 struct ep_send_events_data esed;
1912 1719
1913 if (ms < 0) !! 1720 esed.maxevents = maxevents;
1914 return NULL; !! 1721 esed.events = events;
1915 1722
1916 if (!ms) { !! 1723 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1917 to->tv_sec = 0; <<
1918 to->tv_nsec = 0; <<
1919 return to; <<
1920 } <<
1921 <<
1922 to->tv_sec = ms / MSEC_PER_SEC; <<
1923 to->tv_nsec = NSEC_PER_MSEC * (ms % M <<
1924 <<
1925 ktime_get_ts64(&now); <<
1926 *to = timespec64_add_safe(now, *to); <<
1927 return to; <<
1928 } 1724 }
1929 1725
1930 /* !! 1726 static inline struct timespec64 ep_set_mstimeout(long ms)
1931 * autoremove_wake_function, but remove even <<
1932 * know that default_wake_function/ttwu will <<
1933 * woken, and in that case the ep_poll loop w <<
1934 * try to reuse it. <<
1935 */ <<
1936 static int ep_autoremove_wake_function(struct <<
1937 unsign <<
1938 { 1727 {
1939 int ret = default_wake_function(wq_en !! 1728 struct timespec64 now, ts = {
>> 1729 .tv_sec = ms / MSEC_PER_SEC,
>> 1730 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
>> 1731 };
1940 1732
1941 /* !! 1733 ktime_get_ts64(&now);
1942 * Pairs with list_empty_careful in e !! 1734 return timespec64_add_safe(now, ts);
1943 * iterations see the cause of this w <<
1944 */ <<
1945 list_del_init_careful(&wq_entry->entr <<
1946 return ret; <<
1947 } 1735 }
1948 1736
1949 /** 1737 /**
1950 * ep_poll - Retrieves ready events, and deli !! 1738 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1951 * event buffer. 1739 * event buffer.
1952 * 1740 *
1953 * @ep: Pointer to the eventpoll context. 1741 * @ep: Pointer to the eventpoll context.
1954 * @events: Pointer to the userspace buffer w 1742 * @events: Pointer to the userspace buffer where the ready events should be
1955 * stored. 1743 * stored.
1956 * @maxevents: Size (in terms of number of ev 1744 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1957 * @timeout: Maximum timeout for the ready ev 1745 * @timeout: Maximum timeout for the ready events fetch operation, in
1958 * timespec. If the timeout is zero !! 1746 * milliseconds. If the @timeout is zero, the function will not block,
1959 * while if the @timeout ptr is NUL !! 1747 * while if the @timeout is less than zero, the function will block
1960 * until at least one event has bee 1748 * until at least one event has been retrieved (or an error
1961 * occurred). 1749 * occurred).
1962 * 1750 *
1963 * Return: the number of ready events which h !! 1751 * Returns: Returns the number of ready events which have been fetched, or an
1964 * error code, in case of error. 1752 * error code, in case of error.
1965 */ 1753 */
1966 static int ep_poll(struct eventpoll *ep, stru 1754 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1967 int maxevents, struct time !! 1755 int maxevents, long timeout)
1968 { 1756 {
1969 int res, eavail, timed_out = 0; !! 1757 int res = 0, eavail, timed_out = 0;
>> 1758 unsigned long flags;
1970 u64 slack = 0; 1759 u64 slack = 0;
1971 wait_queue_entry_t wait; 1760 wait_queue_entry_t wait;
1972 ktime_t expires, *to = NULL; 1761 ktime_t expires, *to = NULL;
1973 1762
1974 lockdep_assert_irqs_enabled(); !! 1763 if (timeout > 0) {
>> 1764 struct timespec64 end_time = ep_set_mstimeout(timeout);
1975 1765
1976 if (timeout && (timeout->tv_sec | tim !! 1766 slack = select_estimate_accuracy(&end_time);
1977 slack = select_estimate_accur <<
1978 to = &expires; 1767 to = &expires;
1979 *to = timespec64_to_ktime(*ti !! 1768 *to = timespec64_to_ktime(end_time);
1980 } else if (timeout) { !! 1769 } else if (timeout == 0) {
1981 /* 1770 /*
1982 * Avoid the unnecessary trip 1771 * Avoid the unnecessary trip to the wait queue loop, if the
1983 * caller specified a non blo 1772 * caller specified a non blocking operation.
1984 */ 1773 */
1985 timed_out = 1; 1774 timed_out = 1;
>> 1775 spin_lock_irqsave(&ep->lock, flags);
>> 1776 goto check_events;
1986 } 1777 }
1987 1778
1988 /* !! 1779 fetch_events:
1989 * This call is racy: We may or may n <<
1990 * to the ready list under the lock ( <<
1991 * with a non-zero timeout, this thre <<
1992 * lock and will add to the wait queu <<
1993 * timeout, the user by definition sh <<
1994 * recheck again. <<
1995 */ <<
1996 eavail = ep_events_available(ep); <<
1997 1780
1998 while (1) { !! 1781 if (!ep_events_available(ep))
1999 if (eavail) { !! 1782 ep_busy_loop(ep, timed_out);
2000 /* <<
2001 * Try to transfer ev <<
2002 * 0 events and there <<
2003 * trying again in se <<
2004 */ <<
2005 res = ep_send_events( <<
2006 if (res) <<
2007 return res; <<
2008 } <<
2009 <<
2010 if (timed_out) <<
2011 return 0; <<
2012 1783
2013 eavail = ep_busy_loop(ep, tim !! 1784 spin_lock_irqsave(&ep->lock, flags);
2014 if (eavail) <<
2015 continue; <<
2016 <<
2017 if (signal_pending(current)) <<
2018 return -EINTR; <<
2019 1785
>> 1786 if (!ep_events_available(ep)) {
2020 /* 1787 /*
2021 * Internally init_wait() use !! 1788 * Busy poll timed out. Drop NAPI ID for now, we can add
2022 * thus wait entry is removed !! 1789 * it back in when we have moved a socket with a valid NAPI
2023 * wakeup. Why it is importan !! 1790 * ID onto the ready list.
2024 * each new wakeup will hit t <<
2025 * chance to harvest new even <<
2026 * lost. This is also good pe <<
2027 * normal wakeup path no need <<
2028 * explicitly, thus ep->lock <<
2029 * event delivery. <<
2030 * <<
2031 * In fact, we now use an eve <<
2032 * unconditionally removes, b <<
2033 * entry between loop iterati <<
2034 * performance issue if a pro <<
2035 * threads to wake up without <<
2036 */ <<
2037 init_wait(&wait); <<
2038 wait.func = ep_autoremove_wak <<
2039 <<
2040 write_lock_irq(&ep->lock); <<
2041 /* <<
2042 * Barrierless variant, waitq <<
2043 * the same lock on wakeup ep <<
2044 * is safe to avoid an explic <<
2045 */ <<
2046 __set_current_state(TASK_INTE <<
2047 <<
2048 /* <<
2049 * Do the final check under t <<
2050 * plays with two lists (->rd <<
2051 * is always a race when both <<
2052 * period of time although ev <<
2053 * important. <<
2054 */ 1791 */
2055 eavail = ep_events_available( !! 1792 ep_reset_busy_poll_napi_id(ep);
2056 if (!eavail) <<
2057 __add_wait_queue_excl <<
2058 <<
2059 write_unlock_irq(&ep->lock); <<
2060 <<
2061 if (!eavail) <<
2062 timed_out = !schedule <<
2063 <<
2064 __set_current_state(TASK_RUNN <<
2065 1793
2066 /* 1794 /*
2067 * We were woken up, thus go !! 1795 * We don't have any available event to return to the caller.
2068 * If timed out and still on !! 1796 * We need to sleep here, and we will be wake up by
2069 * carefully under lock, belo !! 1797 * ep_poll_callback() when events will become available.
2070 */ 1798 */
2071 eavail = 1; !! 1799 init_waitqueue_entry(&wait, current);
>> 1800 __add_wait_queue_exclusive(&ep->wq, &wait);
2072 1801
2073 if (!list_empty_careful(&wait !! 1802 for (;;) {
2074 write_lock_irq(&ep->l !! 1803 /*
>> 1804 * We don't want to sleep if the ep_poll_callback() sends us
>> 1805 * a wakeup in between. That's why we set the task state
>> 1806 * to TASK_INTERRUPTIBLE before doing the checks.
>> 1807 */
>> 1808 set_current_state(TASK_INTERRUPTIBLE);
2075 /* 1809 /*
2076 * If the thread time !! 1810 * Always short-circuit for fatal signals to allow
2077 * it means that the !! 1811 * threads to make a timely exit without the chance of
2078 * timeout expired be !! 1812 * finding more events available and fetching
2079 * Thus, when wait.en !! 1813 * repeatedly.
2080 * events. <<
2081 */ 1814 */
2082 if (timed_out) !! 1815 if (fatal_signal_pending(current)) {
2083 eavail = list !! 1816 res = -EINTR;
2084 __remove_wait_queue(& !! 1817 break;
2085 write_unlock_irq(&ep- !! 1818 }
>> 1819 if (ep_events_available(ep) || timed_out)
>> 1820 break;
>> 1821 if (signal_pending(current)) {
>> 1822 res = -EINTR;
>> 1823 break;
>> 1824 }
>> 1825
>> 1826 spin_unlock_irqrestore(&ep->lock, flags);
>> 1827 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
>> 1828 timed_out = 1;
>> 1829
>> 1830 spin_lock_irqsave(&ep->lock, flags);
2086 } 1831 }
>> 1832
>> 1833 __remove_wait_queue(&ep->wq, &wait);
>> 1834 __set_current_state(TASK_RUNNING);
2087 } 1835 }
>> 1836 check_events:
>> 1837 /* Is it worth to try to dig for events ? */
>> 1838 eavail = ep_events_available(ep);
>> 1839
>> 1840 spin_unlock_irqrestore(&ep->lock, flags);
>> 1841
>> 1842 /*
>> 1843 * Try to transfer events to user space. In case we get 0 events and
>> 1844 * there's still timeout left over, we go trying again in search of
>> 1845 * more luck.
>> 1846 */
>> 1847 if (!res && eavail &&
>> 1848 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
>> 1849 goto fetch_events;
>> 1850
>> 1851 return res;
2088 } 1852 }
2089 1853
2090 /** 1854 /**
2091 * ep_loop_check_proc - verify that adding an !! 1855 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
2092 * epoll structure does !! 1856 * API, to verify that adding an epoll file inside another
>> 1857 * epoll structure, does not violate the constraints, in
2093 * terms of closed loops 1858 * terms of closed loops, or too deep chains (which can
2094 * result in excessive s 1859 * result in excessive stack usage).
2095 * 1860 *
2096 * @ep: the &struct eventpoll to be currently !! 1861 * @priv: Pointer to the epoll file to be currently checked.
2097 * @depth: Current depth of the path being ch !! 1862 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
>> 1863 * data structure pointer.
>> 1864 * @call_nests: Current dept of the @ep_call_nested() call stack.
2098 * 1865 *
2099 * Return: %zero if adding the epoll @file in !! 1866 * Returns: Returns zero if adding the epoll @file inside current epoll
2100 * structure @ep does not violate th !! 1867 * structure @ep does not violate the constraints, or -1 otherwise.
2101 */ 1868 */
2102 static int ep_loop_check_proc(struct eventpol !! 1869 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
2103 { 1870 {
2104 int error = 0; 1871 int error = 0;
>> 1872 struct file *file = priv;
>> 1873 struct eventpoll *ep = file->private_data;
>> 1874 struct eventpoll *ep_tovisit;
2105 struct rb_node *rbp; 1875 struct rb_node *rbp;
2106 struct epitem *epi; 1876 struct epitem *epi;
2107 1877
2108 mutex_lock_nested(&ep->mtx, depth + 1 !! 1878 mutex_lock_nested(&ep->mtx, call_nests + 1);
2109 ep->gen = loop_check_gen; 1879 ep->gen = loop_check_gen;
2110 for (rbp = rb_first_cached(&ep->rbr); 1880 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
2111 epi = rb_entry(rbp, struct ep 1881 epi = rb_entry(rbp, struct epitem, rbn);
2112 if (unlikely(is_file_epoll(ep 1882 if (unlikely(is_file_epoll(epi->ffd.file))) {
2113 struct eventpoll *ep_ <<
2114 ep_tovisit = epi->ffd 1883 ep_tovisit = epi->ffd.file->private_data;
2115 if (ep_tovisit->gen = 1884 if (ep_tovisit->gen == loop_check_gen)
2116 continue; 1885 continue;
2117 if (ep_tovisit == ins !! 1886 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
2118 error = -1; !! 1887 ep_loop_check_proc, epi->ffd.file,
2119 else !! 1888 ep_tovisit, current);
2120 error = ep_lo <<
2121 if (error != 0) 1889 if (error != 0)
2122 break; 1890 break;
2123 } else { 1891 } else {
2124 /* 1892 /*
2125 * If we've reached a 1893 * If we've reached a file that is not associated with
2126 * an ep, then we nee 1894 * an ep, then we need to check if the newly added
2127 * links are going to 1895 * links are going to add too many wakeup paths. We do
2128 * this by adding it 1896 * this by adding it to the tfile_check_list, if it's
2129 * not already there, 1897 * not already there, and calling reverse_path_check()
2130 * during ep_insert() 1898 * during ep_insert().
2131 */ 1899 */
2132 list_file(epi->ffd.fi !! 1900 if (list_empty(&epi->ffd.file->f_tfile_llink)) {
>> 1901 if (get_file_rcu(epi->ffd.file))
>> 1902 list_add(&epi->ffd.file->f_tfile_llink,
>> 1903 &tfile_check_list);
>> 1904 }
2133 } 1905 }
2134 } 1906 }
2135 mutex_unlock(&ep->mtx); 1907 mutex_unlock(&ep->mtx);
2136 1908
2137 return error; 1909 return error;
2138 } 1910 }
2139 1911
2140 /** 1912 /**
2141 * ep_loop_check - Performs a check to verify !! 1913 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
2142 * into another epoll file (r !! 1914 * another epoll file (represented by @ep) does not create
2143 * closed loops or too deep c 1915 * closed loops or too deep chains.
2144 * 1916 *
2145 * @ep: Pointer to the epoll we are inserting !! 1917 * @ep: Pointer to the epoll private data structure.
2146 * @to: Pointer to the epoll to be inserted. !! 1918 * @file: Pointer to the epoll file to be checked.
2147 * 1919 *
2148 * Return: %zero if adding the epoll @to insi !! 1920 * Returns: Returns zero if adding the epoll @file inside current epoll
2149 * does not violate the constraints, or %-1 o !! 1921 * structure @ep does not violate the constraints, or -1 otherwise.
2150 */ 1922 */
2151 static int ep_loop_check(struct eventpoll *ep !! 1923 static int ep_loop_check(struct eventpoll *ep, struct file *file)
2152 { 1924 {
2153 inserting_into = ep; !! 1925 return ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
2154 return ep_loop_check_proc(to, 0); !! 1926 ep_loop_check_proc, file, ep, current);
2155 } 1927 }
2156 1928
2157 static void clear_tfile_check_list(void) 1929 static void clear_tfile_check_list(void)
2158 { 1930 {
2159 rcu_read_lock(); !! 1931 struct file *file;
2160 while (tfile_check_list != EP_UNACTIV !! 1932
2161 struct epitems_head *head = t !! 1933 /* first clear the tfile_check_list */
2162 tfile_check_list = head->next !! 1934 while (!list_empty(&tfile_check_list)) {
2163 unlist_file(head); !! 1935 file = list_first_entry(&tfile_check_list, struct file,
>> 1936 f_tfile_llink);
>> 1937 list_del_init(&file->f_tfile_llink);
>> 1938 fput(file);
2164 } 1939 }
2165 rcu_read_unlock(); !! 1940 INIT_LIST_HEAD(&tfile_check_list);
2166 } 1941 }
2167 1942
2168 /* 1943 /*
2169 * Open an eventpoll file descriptor. 1944 * Open an eventpoll file descriptor.
2170 */ 1945 */
2171 static int do_epoll_create(int flags) !! 1946 SYSCALL_DEFINE1(epoll_create1, int, flags)
2172 { 1947 {
2173 int error, fd; 1948 int error, fd;
2174 struct eventpoll *ep = NULL; 1949 struct eventpoll *ep = NULL;
2175 struct file *file; 1950 struct file *file;
2176 1951
2177 /* Check the EPOLL_* constant for con 1952 /* Check the EPOLL_* constant for consistency. */
2178 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEX 1953 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
2179 1954
2180 if (flags & ~EPOLL_CLOEXEC) 1955 if (flags & ~EPOLL_CLOEXEC)
2181 return -EINVAL; 1956 return -EINVAL;
2182 /* 1957 /*
2183 * Create the internal data structure 1958 * Create the internal data structure ("struct eventpoll").
2184 */ 1959 */
2185 error = ep_alloc(&ep); 1960 error = ep_alloc(&ep);
2186 if (error < 0) 1961 if (error < 0)
2187 return error; 1962 return error;
2188 /* 1963 /*
2189 * Creates all the items needed to se 1964 * Creates all the items needed to setup an eventpoll file. That is,
2190 * a file structure and a free file d 1965 * a file structure and a free file descriptor.
2191 */ 1966 */
2192 fd = get_unused_fd_flags(O_RDWR | (fl 1967 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
2193 if (fd < 0) { 1968 if (fd < 0) {
2194 error = fd; 1969 error = fd;
2195 goto out_free_ep; 1970 goto out_free_ep;
2196 } 1971 }
2197 file = anon_inode_getfile("[eventpoll 1972 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
2198 O_RDWR | (fl 1973 O_RDWR | (flags & O_CLOEXEC));
2199 if (IS_ERR(file)) { 1974 if (IS_ERR(file)) {
2200 error = PTR_ERR(file); 1975 error = PTR_ERR(file);
2201 goto out_free_fd; 1976 goto out_free_fd;
2202 } 1977 }
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; 1978 ep->file = file;
2209 fd_install(fd, file); 1979 fd_install(fd, file);
2210 return fd; 1980 return fd;
2211 1981
2212 out_free_fd: 1982 out_free_fd:
2213 put_unused_fd(fd); 1983 put_unused_fd(fd);
2214 out_free_ep: 1984 out_free_ep:
2215 ep_clear_and_put(ep); !! 1985 ep_free(ep);
2216 return error; 1986 return error;
2217 } 1987 }
2218 1988
2219 SYSCALL_DEFINE1(epoll_create1, int, flags) <<
2220 { <<
2221 return do_epoll_create(flags); <<
2222 } <<
2223 <<
2224 SYSCALL_DEFINE1(epoll_create, int, size) 1989 SYSCALL_DEFINE1(epoll_create, int, size)
2225 { 1990 {
2226 if (size <= 0) 1991 if (size <= 0)
2227 return -EINVAL; 1992 return -EINVAL;
2228 1993
2229 return do_epoll_create(0); !! 1994 return sys_epoll_create1(0);
2230 } <<
2231 <<
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 <<
2246 bool nonbl <<
2247 { <<
2248 if (!nonblock) { <<
2249 mutex_lock_nested(mutex, dept <<
2250 return 0; <<
2251 } <<
2252 if (mutex_trylock(mutex)) <<
2253 return 0; <<
2254 return -EAGAIN; <<
2255 } 1995 }
2256 1996
2257 int do_epoll_ctl(int epfd, int op, int fd, st !! 1997 /*
2258 bool nonblock) !! 1998 * The following function implements the controller interface for
>> 1999 * the eventpoll file that enables the insertion/removal/change of
>> 2000 * file descriptors inside the interest set.
>> 2001 */
>> 2002 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
>> 2003 struct epoll_event __user *, event)
2259 { 2004 {
2260 int error; 2005 int error;
2261 int full_check = 0; 2006 int full_check = 0;
2262 struct fd f, tf; 2007 struct fd f, tf;
2263 struct eventpoll *ep; 2008 struct eventpoll *ep;
2264 struct epitem *epi; 2009 struct epitem *epi;
>> 2010 struct epoll_event epds;
2265 struct eventpoll *tep = NULL; 2011 struct eventpoll *tep = NULL;
2266 2012
>> 2013 error = -EFAULT;
>> 2014 if (ep_op_has_event(op) &&
>> 2015 copy_from_user(&epds, event, sizeof(struct epoll_event)))
>> 2016 goto error_return;
>> 2017
2267 error = -EBADF; 2018 error = -EBADF;
2268 f = fdget(epfd); 2019 f = fdget(epfd);
2269 if (!f.file) 2020 if (!f.file)
2270 goto error_return; 2021 goto error_return;
2271 2022
2272 /* Get the "struct file *" for the ta 2023 /* Get the "struct file *" for the target file */
2273 tf = fdget(fd); 2024 tf = fdget(fd);
2274 if (!tf.file) 2025 if (!tf.file)
2275 goto error_fput; 2026 goto error_fput;
2276 2027
2277 /* The target file descriptor must su 2028 /* The target file descriptor must support poll */
2278 error = -EPERM; 2029 error = -EPERM;
2279 if (!file_can_poll(tf.file)) !! 2030 if (!tf.file->f_op->poll)
2280 goto error_tgt_fput; 2031 goto error_tgt_fput;
2281 2032
2282 /* Check if EPOLLWAKEUP is allowed */ 2033 /* Check if EPOLLWAKEUP is allowed */
2283 if (ep_op_has_event(op)) 2034 if (ep_op_has_event(op))
2284 ep_take_care_of_epollwakeup(e !! 2035 ep_take_care_of_epollwakeup(&epds);
2285 2036
2286 /* 2037 /*
2287 * We have to check that the file str 2038 * We have to check that the file structure underneath the file descriptor
2288 * the user passed to us _is_ an even 2039 * the user passed to us _is_ an eventpoll file. And also we do not permit
2289 * adding an epoll file descriptor in 2040 * adding an epoll file descriptor inside itself.
2290 */ 2041 */
2291 error = -EINVAL; 2042 error = -EINVAL;
2292 if (f.file == tf.file || !is_file_epo 2043 if (f.file == tf.file || !is_file_epoll(f.file))
2293 goto error_tgt_fput; 2044 goto error_tgt_fput;
2294 2045
2295 /* 2046 /*
2296 * epoll adds to the wakeup queue at 2047 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2297 * so EPOLLEXCLUSIVE is not allowed f 2048 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2298 * Also, we do not currently supporte 2049 * Also, we do not currently supported nested exclusive wakeups.
2299 */ 2050 */
2300 if (ep_op_has_event(op) && (epds->eve !! 2051 if (ep_op_has_event(op) && (epds.events & EPOLLEXCLUSIVE)) {
2301 if (op == EPOLL_CTL_MOD) 2052 if (op == EPOLL_CTL_MOD)
2302 goto error_tgt_fput; 2053 goto error_tgt_fput;
2303 if (op == EPOLL_CTL_ADD && (i 2054 if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2304 (epds->events !! 2055 (epds.events & ~EPOLLEXCLUSIVE_OK_BITS)))
2305 goto error_tgt_fput; 2056 goto error_tgt_fput;
2306 } 2057 }
2307 2058
2308 /* 2059 /*
2309 * At this point it is safe to assume 2060 * At this point it is safe to assume that the "private_data" contains
2310 * our own data structure. 2061 * our own data structure.
2311 */ 2062 */
2312 ep = f.file->private_data; 2063 ep = f.file->private_data;
2313 2064
2314 /* 2065 /*
2315 * When we insert an epoll file descr !! 2066 * When we insert an epoll file descriptor, inside another epoll file
2316 * descriptor, there is the chance of !! 2067 * descriptor, there is the change of creating closed loops, which are
2317 * better be handled here, than in mo 2068 * better be handled here, than in more critical paths. While we are
2318 * checking for loops we also determi 2069 * checking for loops we also determine the list of files reachable
2319 * and hang them on the tfile_check_l 2070 * and hang them on the tfile_check_list, so we can check that we
2320 * haven't created too many possible 2071 * haven't created too many possible wakeup paths.
2321 * 2072 *
2322 * We do not need to take the global 2073 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2323 * the epoll file descriptor is attac 2074 * the epoll file descriptor is attaching directly to a wakeup source,
2324 * unless the epoll file descriptor i 2075 * unless the epoll file descriptor is nested. The purpose of taking the
2325 * 'epnested_mutex' on add is to prev !! 2076 * 'epmutex' on add is to prevent complex toplogies such as loops and
2326 * deep wakeup paths from forming in 2077 * deep wakeup paths from forming in parallel through multiple
2327 * EPOLL_CTL_ADD operations. 2078 * EPOLL_CTL_ADD operations.
2328 */ 2079 */
2329 error = epoll_mutex_lock(&ep->mtx, 0, !! 2080 mutex_lock_nested(&ep->mtx, 0);
2330 if (error) <<
2331 goto error_tgt_fput; <<
2332 if (op == EPOLL_CTL_ADD) { 2081 if (op == EPOLL_CTL_ADD) {
2333 if (READ_ONCE(f.file->f_ep) | !! 2082 if (!list_empty(&f.file->f_ep_links) ||
2334 is_file_epoll(tf.file)) { !! 2083 ep->gen == loop_check_gen ||
2335 mutex_unlock(&ep->mtx !! 2084 is_file_epoll(tf.file)) {
2336 error = epoll_mutex_l <<
2337 if (error) <<
2338 goto error_tg <<
2339 loop_check_gen++; <<
2340 full_check = 1; 2085 full_check = 1;
>> 2086 mutex_unlock(&ep->mtx);
>> 2087 mutex_lock(&epmutex);
2341 if (is_file_epoll(tf. 2088 if (is_file_epoll(tf.file)) {
2342 tep = tf.file <<
2343 error = -ELOO 2089 error = -ELOOP;
2344 if (ep_loop_c !! 2090 if (ep_loop_check(ep, tf.file) != 0)
2345 goto 2091 goto error_tgt_fput;
>> 2092 } else {
>> 2093 get_file(tf.file);
>> 2094 list_add(&tf.file->f_tfile_llink,
>> 2095 &tfile_check_list);
>> 2096 }
>> 2097 mutex_lock_nested(&ep->mtx, 0);
>> 2098 if (is_file_epoll(tf.file)) {
>> 2099 tep = tf.file->private_data;
>> 2100 mutex_lock_nested(&tep->mtx, 1);
2346 } 2101 }
2347 error = epoll_mutex_l <<
2348 if (error) <<
2349 goto error_tg <<
2350 } 2102 }
2351 } 2103 }
2352 2104
2353 /* 2105 /*
2354 * Try to lookup the file inside our !! 2106 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
2355 * above, we can be sure to be able t 2107 * above, we can be sure to be able to use the item looked up by
2356 * ep_find() till we release the mute 2108 * ep_find() till we release the mutex.
2357 */ 2109 */
2358 epi = ep_find(ep, tf.file, fd); 2110 epi = ep_find(ep, tf.file, fd);
2359 2111
2360 error = -EINVAL; 2112 error = -EINVAL;
2361 switch (op) { 2113 switch (op) {
2362 case EPOLL_CTL_ADD: 2114 case EPOLL_CTL_ADD:
2363 if (!epi) { 2115 if (!epi) {
2364 epds->events |= EPOLL !! 2116 epds.events |= POLLERR | POLLHUP;
2365 error = ep_insert(ep, !! 2117 error = ep_insert(ep, &epds, tf.file, fd, full_check);
2366 } else 2118 } else
2367 error = -EEXIST; 2119 error = -EEXIST;
2368 break; 2120 break;
2369 case EPOLL_CTL_DEL: 2121 case EPOLL_CTL_DEL:
2370 if (epi) { !! 2122 if (epi)
2371 /* !! 2123 error = ep_remove(ep, epi);
2372 * The eventpoll itse !! 2124 else
2373 * can't go to zero h <<
2374 */ <<
2375 ep_remove_safe(ep, ep <<
2376 error = 0; <<
2377 } else { <<
2378 error = -ENOENT; 2125 error = -ENOENT;
2379 } <<
2380 break; 2126 break;
2381 case EPOLL_CTL_MOD: 2127 case EPOLL_CTL_MOD:
2382 if (epi) { 2128 if (epi) {
2383 if (!(epi->event.even 2129 if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2384 epds->events !! 2130 epds.events |= POLLERR | POLLHUP;
2385 error = ep_mo !! 2131 error = ep_modify(ep, epi, &epds);
2386 } 2132 }
2387 } else 2133 } else
2388 error = -ENOENT; 2134 error = -ENOENT;
2389 break; 2135 break;
2390 } 2136 }
>> 2137 if (tep != NULL)
>> 2138 mutex_unlock(&tep->mtx);
2391 mutex_unlock(&ep->mtx); 2139 mutex_unlock(&ep->mtx);
2392 2140
2393 error_tgt_fput: 2141 error_tgt_fput:
2394 if (full_check) { 2142 if (full_check) {
2395 clear_tfile_check_list(); 2143 clear_tfile_check_list();
2396 loop_check_gen++; 2144 loop_check_gen++;
2397 mutex_unlock(&epnested_mutex) !! 2145 mutex_unlock(&epmutex);
2398 } 2146 }
2399 2147
2400 fdput(tf); 2148 fdput(tf);
2401 error_fput: 2149 error_fput:
2402 fdput(f); 2150 fdput(f);
2403 error_return: 2151 error_return:
2404 2152
2405 return error; 2153 return error;
2406 } 2154 }
2407 2155
2408 /* 2156 /*
2409 * The following function implements the cont <<
2410 * the eventpoll file that enables the insert <<
2411 * file descriptors inside the interest set. <<
2412 */ <<
2413 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op <<
2414 struct epoll_event __user *, <<
2415 { <<
2416 struct epoll_event epds; <<
2417 <<
2418 if (ep_op_has_event(op) && <<
2419 copy_from_user(&epds, event, size <<
2420 return -EFAULT; <<
2421 <<
2422 return do_epoll_ctl(epfd, op, fd, &ep <<
2423 } <<
2424 <<
2425 /* <<
2426 * Implement the event wait interface for the 2157 * Implement the event wait interface for the eventpoll file. It is the kernel
2427 * part of the user space epoll_wait(2). 2158 * part of the user space epoll_wait(2).
2428 */ 2159 */
2429 static int do_epoll_wait(int epfd, struct epo !! 2160 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2430 int maxevents, struc !! 2161 int, maxevents, int, timeout)
2431 { 2162 {
2432 int error; 2163 int error;
2433 struct fd f; 2164 struct fd f;
2434 struct eventpoll *ep; 2165 struct eventpoll *ep;
2435 2166
2436 /* The maximum number of event must b 2167 /* The maximum number of event must be greater than zero */
2437 if (maxevents <= 0 || maxevents > EP_ 2168 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2438 return -EINVAL; 2169 return -EINVAL;
2439 2170
2440 /* Verify that the area passed by the 2171 /* Verify that the area passed by the user is writeable */
2441 if (!access_ok(events, maxevents * si !! 2172 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
2442 return -EFAULT; 2173 return -EFAULT;
2443 2174
2444 /* Get the "struct file *" for the ev 2175 /* Get the "struct file *" for the eventpoll file */
2445 f = fdget(epfd); 2176 f = fdget(epfd);
2446 if (!f.file) 2177 if (!f.file)
2447 return -EBADF; 2178 return -EBADF;
2448 2179
2449 /* 2180 /*
2450 * We have to check that the file str 2181 * We have to check that the file structure underneath the fd
2451 * the user passed to us _is_ an even 2182 * the user passed to us _is_ an eventpoll file.
2452 */ 2183 */
2453 error = -EINVAL; 2184 error = -EINVAL;
2454 if (!is_file_epoll(f.file)) 2185 if (!is_file_epoll(f.file))
2455 goto error_fput; 2186 goto error_fput;
2456 2187
2457 /* 2188 /*
2458 * At this point it is safe to assume 2189 * At this point it is safe to assume that the "private_data" contains
2459 * our own data structure. 2190 * our own data structure.
2460 */ 2191 */
2461 ep = f.file->private_data; 2192 ep = f.file->private_data;
2462 2193
2463 /* Time to fish for events ... */ 2194 /* Time to fish for events ... */
2464 error = ep_poll(ep, events, maxevents !! 2195 error = ep_poll(ep, events, maxevents, timeout);
2465 2196
2466 error_fput: 2197 error_fput:
2467 fdput(f); 2198 fdput(f);
2468 return error; 2199 return error;
2469 } 2200 }
2470 2201
2471 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct <<
2472 int, maxevents, int, timeout) <<
2473 { <<
2474 struct timespec64 to; <<
2475 <<
2476 return do_epoll_wait(epfd, events, ma <<
2477 ep_timeout_to_ti <<
2478 } <<
2479 <<
2480 /* 2202 /*
2481 * Implement the event wait interface for the 2203 * Implement the event wait interface for the eventpoll file. It is the kernel
2482 * part of the user space epoll_pwait(2). 2204 * part of the user space epoll_pwait(2).
2483 */ 2205 */
2484 static int do_epoll_pwait(int epfd, struct ep !! 2206 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2485 int maxevents, stru !! 2207 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2486 const sigset_t __us !! 2208 size_t, sigsetsize)
2487 { 2209 {
2488 int error; 2210 int error;
>> 2211 sigset_t ksigmask, sigsaved;
2489 2212
2490 /* 2213 /*
2491 * If the caller wants a certain sign 2214 * If the caller wants a certain signal mask to be set during the wait,
2492 * we apply it here. 2215 * we apply it here.
2493 */ 2216 */
2494 error = set_user_sigmask(sigmask, sig !! 2217 if (sigmask) {
2495 if (error) !! 2218 if (sigsetsize != sizeof(sigset_t))
2496 return error; !! 2219 return -EINVAL;
2497 !! 2220 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
2498 error = do_epoll_wait(epfd, events, m !! 2221 return -EFAULT;
2499 !! 2222 sigsaved = current->blocked;
2500 restore_saved_sigmask_unless(error == !! 2223 set_current_blocked(&ksigmask);
2501 !! 2224 }
2502 return error; <<
2503 } <<
2504 <<
2505 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struc <<
2506 int, maxevents, int, timeout, <<
2507 size_t, sigsetsize) <<
2508 { <<
2509 struct timespec64 to; <<
2510 <<
2511 return do_epoll_pwait(epfd, events, m <<
2512 ep_timeout_to_t <<
2513 sigmask, sigset <<
2514 } <<
2515 2225
2516 SYSCALL_DEFINE6(epoll_pwait2, int, epfd, stru !! 2226 error = sys_epoll_wait(epfd, events, maxevents, timeout);
2517 int, maxevents, const struct <<
2518 const sigset_t __user *, sigm <<
2519 { <<
2520 struct timespec64 ts, *to = NULL; <<
2521 2227
2522 if (timeout) { !! 2228 /*
2523 if (get_timespec64(&ts, timeo !! 2229 * If we changed the signal mask, we need to restore the original one.
2524 return -EFAULT; !! 2230 * In case we've got a signal while waiting, we do not restore the
2525 to = &ts; !! 2231 * signal mask yet, and we allow do_signal() to deliver the signal on
2526 if (poll_select_set_timeout(t !! 2232 * the way back to userspace, before the signal mask is restored.
2527 return -EINVAL; !! 2233 */
>> 2234 if (sigmask) {
>> 2235 if (error == -EINTR) {
>> 2236 memcpy(¤t->saved_sigmask, &sigsaved,
>> 2237 sizeof(sigsaved));
>> 2238 set_restore_sigmask();
>> 2239 } else
>> 2240 set_current_blocked(&sigsaved);
2528 } 2241 }
2529 2242
2530 return do_epoll_pwait(epfd, events, m !! 2243 return error;
2531 sigmask, sigset <<
2532 } 2244 }
2533 2245
2534 #ifdef CONFIG_COMPAT 2246 #ifdef CONFIG_COMPAT
2535 static int do_compat_epoll_pwait(int epfd, st !! 2247 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2536 int maxevent !! 2248 struct epoll_event __user *, events,
2537 const compat !! 2249 int, maxevents, int, timeout,
2538 compat_size_ !! 2250 const compat_sigset_t __user *, sigmask,
>> 2251 compat_size_t, sigsetsize)
2539 { 2252 {
2540 long err; 2253 long err;
>> 2254 compat_sigset_t csigmask;
>> 2255 sigset_t ksigmask, sigsaved;
2541 2256
2542 /* 2257 /*
2543 * If the caller wants a certain sign 2258 * If the caller wants a certain signal mask to be set during the wait,
2544 * we apply it here. 2259 * we apply it here.
2545 */ 2260 */
2546 err = set_compat_user_sigmask(sigmask !! 2261 if (sigmask) {
2547 if (err) !! 2262 if (sigsetsize != sizeof(compat_sigset_t))
2548 return err; !! 2263 return -EINVAL;
2549 !! 2264 if (copy_from_user(&csigmask, sigmask, sizeof(csigmask)))
2550 err = do_epoll_wait(epfd, events, max !! 2265 return -EFAULT;
2551 !! 2266 sigset_from_compat(&ksigmask, &csigmask);
2552 restore_saved_sigmask_unless(err == - !! 2267 sigsaved = current->blocked;
2553 !! 2268 set_current_blocked(&ksigmask);
2554 return err; !! 2269 }
2555 } <<
2556 <<
2557 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd <<
2558 struct epoll_event __u <<
2559 int, maxevents, int, t <<
2560 const compat_sigset_t <<
2561 compat_size_t, sigsets <<
2562 { <<
2563 struct timespec64 to; <<
2564 <<
2565 return do_compat_epoll_pwait(epfd, ev <<
2566 ep_timeo <<
2567 sigmask, <<
2568 } <<
2569 2270
2570 COMPAT_SYSCALL_DEFINE6(epoll_pwait2, int, epf !! 2271 err = sys_epoll_wait(epfd, events, maxevents, timeout);
2571 struct epoll_event __u <<
2572 int, maxevents, <<
2573 const struct __kernel_ <<
2574 const compat_sigset_t <<
2575 compat_size_t, sigsets <<
2576 { <<
2577 struct timespec64 ts, *to = NULL; <<
2578 2272
2579 if (timeout) { !! 2273 /*
2580 if (get_timespec64(&ts, timeo !! 2274 * If we changed the signal mask, we need to restore the original one.
2581 return -EFAULT; !! 2275 * In case we've got a signal while waiting, we do not restore the
2582 to = &ts; !! 2276 * signal mask yet, and we allow do_signal() to deliver the signal on
2583 if (poll_select_set_timeout(t !! 2277 * the way back to userspace, before the signal mask is restored.
2584 return -EINVAL; !! 2278 */
>> 2279 if (sigmask) {
>> 2280 if (err == -EINTR) {
>> 2281 memcpy(¤t->saved_sigmask, &sigsaved,
>> 2282 sizeof(sigsaved));
>> 2283 set_restore_sigmask();
>> 2284 } else
>> 2285 set_current_blocked(&sigsaved);
2585 } 2286 }
2586 2287
2587 return do_compat_epoll_pwait(epfd, ev !! 2288 return err;
2588 sigmask, <<
2589 } 2289 }
2590 <<
2591 #endif 2290 #endif
2592 2291
2593 static int __init eventpoll_init(void) 2292 static int __init eventpoll_init(void)
2594 { 2293 {
2595 struct sysinfo si; 2294 struct sysinfo si;
2596 2295
2597 si_meminfo(&si); 2296 si_meminfo(&si);
2598 /* 2297 /*
2599 * Allows top 4% of lomem to be alloc 2298 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2600 */ 2299 */
2601 max_user_watches = (((si.totalram - s 2300 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2602 EP_ITEM_COST; 2301 EP_ITEM_COST;
2603 BUG_ON(max_user_watches < 0); 2302 BUG_ON(max_user_watches < 0);
2604 2303
2605 /* 2304 /*
>> 2305 * Initialize the structure used to perform epoll file descriptor
>> 2306 * inclusion loops checks.
>> 2307 */
>> 2308 ep_nested_calls_init(&poll_loop_ncalls);
>> 2309
>> 2310 /* Initialize the structure used to perform safe poll wait head wake ups */
>> 2311 ep_nested_calls_init(&poll_safewake_ncalls);
>> 2312
>> 2313 /* Initialize the structure used to perform file's f_op->poll() calls */
>> 2314 ep_nested_calls_init(&poll_readywalk_ncalls);
>> 2315
>> 2316 /*
2606 * We can have many thousands of epit 2317 * We can have many thousands of epitems, so prevent this from
2607 * using an extra cache line on 64-bi 2318 * using an extra cache line on 64-bit (and smaller) CPUs
2608 */ 2319 */
2609 BUILD_BUG_ON(sizeof(void *) <= 8 && s 2320 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2610 2321
2611 /* Allocates slab cache used to alloc 2322 /* Allocates slab cache used to allocate "struct epitem" items */
2612 epi_cache = kmem_cache_create("eventp 2323 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2613 0, SLAB_HWCACHE_ALIGN !! 2324 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2614 2325
2615 /* Allocates slab cache used to alloc 2326 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2616 pwq_cache = kmem_cache_create("eventp 2327 pwq_cache = kmem_cache_create("eventpoll_pwq",
2617 sizeof(struct eppoll_entry), !! 2328 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
2618 epoll_sysctls_init(); <<
2619 <<
2620 ephead_cache = kmem_cache_create("ep_ <<
2621 sizeof(struct epitems_head), <<
2622 2329
2623 return 0; 2330 return 0;
2624 } 2331 }
2625 fs_initcall(eventpoll_init); 2332 fs_initcall(eventpoll_init);
2626 2333
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.