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