Commit | Line | Data |
---|---|---|
1182fca3 MM |
1 | /* |
2 | * (C) 2001 Clemson University and The University of Chicago | |
3 | * (C) 2011 Omnibond Systems | |
4 | * | |
5 | * Changes by Acxiom Corporation to implement generic service_operation() | |
6 | * function, Copyright Acxiom Corporation, 2005. | |
7 | * | |
8 | * See COPYING in top-level directory. | |
9 | */ | |
10 | ||
11 | /* | |
12 | * In-kernel waitqueue operations. | |
13 | */ | |
14 | ||
15 | #include "protocol.h" | |
575e9461 MM |
16 | #include "orangefs-kernel.h" |
17 | #include "orangefs-bufmap.h" | |
1182fca3 | 18 | |
ade3d781 AV |
19 | static int wait_for_cancellation_downcall(struct orangefs_kernel_op_s *); |
20 | static int wait_for_matching_downcall(struct orangefs_kernel_op_s *); | |
21 | ||
1182fca3 MM |
22 | /* |
23 | * What we do in this function is to walk the list of operations that are | |
24 | * present in the request queue and mark them as purged. | |
25 | * NOTE: This is called from the device close after client-core has | |
26 | * guaranteed that no new operations could appear on the list since the | |
27 | * client-core is anyway going to exit. | |
28 | */ | |
29 | void purge_waiting_ops(void) | |
30 | { | |
8bb8aefd | 31 | struct orangefs_kernel_op_s *op; |
1182fca3 | 32 | |
8bb8aefd YL |
33 | spin_lock(&orangefs_request_list_lock); |
34 | list_for_each_entry(op, &orangefs_request_list, list) { | |
1182fca3 MM |
35 | gossip_debug(GOSSIP_WAIT_DEBUG, |
36 | "pvfs2-client-core: purging op tag %llu %s\n", | |
37 | llu(op->tag), | |
38 | get_opname_string(op)); | |
39 | spin_lock(&op->lock); | |
40 | set_op_state_purged(op); | |
41 | spin_unlock(&op->lock); | |
1182fca3 | 42 | } |
8bb8aefd | 43 | spin_unlock(&orangefs_request_list_lock); |
1182fca3 MM |
44 | } |
45 | ||
fc916da5 AV |
46 | static inline void |
47 | add_op_to_request_list(struct orangefs_kernel_op_s *op) | |
48 | { | |
49 | spin_lock(&orangefs_request_list_lock); | |
50 | spin_lock(&op->lock); | |
51 | set_op_state_waiting(op); | |
52 | list_add_tail(&op->list, &orangefs_request_list); | |
53 | spin_unlock(&orangefs_request_list_lock); | |
54 | spin_unlock(&op->lock); | |
55 | wake_up_interruptible(&orangefs_request_list_waitq); | |
56 | } | |
57 | ||
58 | static inline | |
59 | void add_priority_op_to_request_list(struct orangefs_kernel_op_s *op) | |
60 | { | |
61 | spin_lock(&orangefs_request_list_lock); | |
62 | spin_lock(&op->lock); | |
63 | set_op_state_waiting(op); | |
64 | ||
65 | list_add(&op->list, &orangefs_request_list); | |
66 | spin_unlock(&orangefs_request_list_lock); | |
67 | spin_unlock(&op->lock); | |
68 | wake_up_interruptible(&orangefs_request_list_waitq); | |
69 | } | |
70 | ||
1182fca3 | 71 | /* |
8bb8aefd | 72 | * submits a ORANGEFS operation and waits for it to complete |
1182fca3 MM |
73 | * |
74 | * Note op->downcall.status will contain the status of the operation (in | |
75 | * errno format), whether provided by pvfs2-client or a result of failure to | |
76 | * service the operation. If the caller wishes to distinguish, then | |
77 | * op->state can be checked to see if it was serviced or not. | |
78 | * | |
79 | * Returns contents of op->downcall.status for convenience | |
80 | */ | |
8bb8aefd | 81 | int service_operation(struct orangefs_kernel_op_s *op, |
1182fca3 MM |
82 | const char *op_name, |
83 | int flags) | |
84 | { | |
85 | /* flags to modify behavior */ | |
86 | sigset_t orig_sigset; | |
87 | int ret = 0; | |
88 | ||
ce6c414e | 89 | DEFINE_WAIT(wait_entry); |
1182fca3 MM |
90 | |
91 | op->upcall.tgid = current->tgid; | |
92 | op->upcall.pid = current->pid; | |
93 | ||
94 | retry_servicing: | |
95 | op->downcall.status = 0; | |
96 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
8bb8aefd | 97 | "orangefs: service_operation: %s %p\n", |
1182fca3 MM |
98 | op_name, |
99 | op); | |
100 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
8bb8aefd | 101 | "orangefs: operation posted by process: %s, pid: %i\n", |
1182fca3 MM |
102 | current->comm, |
103 | current->pid); | |
104 | ||
105 | /* mask out signals if this operation is not to be interrupted */ | |
8bb8aefd | 106 | if (!(flags & ORANGEFS_OP_INTERRUPTIBLE)) |
c146c0b8 | 107 | orangefs_block_signals(&orig_sigset); |
1182fca3 | 108 | |
8bb8aefd | 109 | if (!(flags & ORANGEFS_OP_NO_SEMAPHORE)) { |
1182fca3 MM |
110 | ret = mutex_lock_interruptible(&request_mutex); |
111 | /* | |
112 | * check to see if we were interrupted while waiting for | |
113 | * semaphore | |
114 | */ | |
115 | if (ret < 0) { | |
8bb8aefd | 116 | if (!(flags & ORANGEFS_OP_INTERRUPTIBLE)) |
c146c0b8 | 117 | orangefs_set_signals(&orig_sigset); |
1182fca3 MM |
118 | op->downcall.status = ret; |
119 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
8bb8aefd | 120 | "orangefs: service_operation interrupted.\n"); |
1182fca3 MM |
121 | return ret; |
122 | } | |
123 | } | |
124 | ||
125 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
126 | "%s:About to call is_daemon_in_service().\n", | |
127 | __func__); | |
128 | ||
129 | if (is_daemon_in_service() < 0) { | |
130 | /* | |
131 | * By incrementing the per-operation attempt counter, we | |
132 | * directly go into the timeout logic while waiting for | |
133 | * the matching downcall to be read | |
134 | */ | |
135 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
136 | "%s:client core is NOT in service(%d).\n", | |
137 | __func__, | |
138 | is_daemon_in_service()); | |
139 | op->attempts++; | |
140 | } | |
141 | ||
142 | /* queue up the operation */ | |
8bb8aefd | 143 | if (flags & ORANGEFS_OP_PRIORITY) { |
1182fca3 MM |
144 | add_priority_op_to_request_list(op); |
145 | } else { | |
146 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
147 | "%s:About to call add_op_to_request_list().\n", | |
148 | __func__); | |
149 | add_op_to_request_list(op); | |
150 | } | |
151 | ||
8bb8aefd | 152 | if (!(flags & ORANGEFS_OP_NO_SEMAPHORE)) |
1182fca3 MM |
153 | mutex_unlock(&request_mutex); |
154 | ||
155 | /* | |
156 | * If we are asked to service an asynchronous operation from | |
157 | * VFS perspective, we are done. | |
158 | */ | |
8bb8aefd | 159 | if (flags & ORANGEFS_OP_ASYNC) |
1182fca3 MM |
160 | return 0; |
161 | ||
8bb8aefd | 162 | if (flags & ORANGEFS_OP_CANCELLATION) { |
1182fca3 MM |
163 | gossip_debug(GOSSIP_WAIT_DEBUG, |
164 | "%s:" | |
165 | "About to call wait_for_cancellation_downcall.\n", | |
166 | __func__); | |
167 | ret = wait_for_cancellation_downcall(op); | |
168 | } else { | |
169 | ret = wait_for_matching_downcall(op); | |
170 | } | |
171 | ||
172 | if (ret < 0) { | |
173 | /* failed to get matching downcall */ | |
174 | if (ret == -ETIMEDOUT) { | |
8bb8aefd | 175 | gossip_err("orangefs: %s -- wait timed out; aborting attempt.\n", |
1182fca3 MM |
176 | op_name); |
177 | } | |
178 | op->downcall.status = ret; | |
179 | } else { | |
180 | /* got matching downcall; make sure status is in errno format */ | |
181 | op->downcall.status = | |
8bb8aefd | 182 | orangefs_normalize_to_errno(op->downcall.status); |
1182fca3 MM |
183 | ret = op->downcall.status; |
184 | } | |
185 | ||
8bb8aefd | 186 | if (!(flags & ORANGEFS_OP_INTERRUPTIBLE)) |
c146c0b8 | 187 | orangefs_set_signals(&orig_sigset); |
1182fca3 MM |
188 | |
189 | BUG_ON(ret != op->downcall.status); | |
190 | /* retry if operation has not been serviced and if requested */ | |
191 | if (!op_state_serviced(op) && op->downcall.status == -EAGAIN) { | |
192 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
8bb8aefd | 193 | "orangefs: tag %llu (%s)" |
1182fca3 MM |
194 | " -- operation to be retried (%d attempt)\n", |
195 | llu(op->tag), | |
196 | op_name, | |
197 | op->attempts + 1); | |
198 | ||
199 | if (!op->uses_shared_memory) | |
200 | /* | |
201 | * this operation doesn't use the shared memory | |
202 | * system | |
203 | */ | |
204 | goto retry_servicing; | |
205 | ||
206 | /* op uses shared memory */ | |
7d221485 | 207 | if (orangefs_get_bufmap_init() == 0) { |
1182fca3 MM |
208 | /* |
209 | * This operation uses the shared memory system AND | |
210 | * the system is not yet ready. This situation occurs | |
211 | * when the client-core is restarted AND there were | |
212 | * operations waiting to be processed or were already | |
213 | * in process. | |
214 | */ | |
215 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
216 | "uses_shared_memory is true.\n"); | |
217 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
218 | "Client core in-service status(%d).\n", | |
219 | is_daemon_in_service()); | |
220 | gossip_debug(GOSSIP_WAIT_DEBUG, "bufmap_init:%d.\n", | |
7d221485 | 221 | orangefs_get_bufmap_init()); |
1182fca3 MM |
222 | gossip_debug(GOSSIP_WAIT_DEBUG, |
223 | "operation's status is 0x%0x.\n", | |
224 | op->op_state); | |
225 | ||
226 | /* | |
227 | * let process sleep for a few seconds so shared | |
228 | * memory system can be initialized. | |
229 | */ | |
ce6c414e MM |
230 | prepare_to_wait(&orangefs_bufmap_init_waitq, |
231 | &wait_entry, | |
232 | TASK_INTERRUPTIBLE); | |
1182fca3 | 233 | |
1182fca3 | 234 | /* |
8bb8aefd | 235 | * Wait for orangefs_bufmap_initialize() to wake me up |
1182fca3 MM |
236 | * within the allotted time. |
237 | */ | |
238 | ret = schedule_timeout(MSECS_TO_JIFFIES | |
8bb8aefd | 239 | (1000 * ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS)); |
1182fca3 MM |
240 | |
241 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
242 | "Value returned from schedule_timeout:" | |
243 | "%d.\n", | |
244 | ret); | |
245 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
246 | "Is shared memory available? (%d).\n", | |
7d221485 | 247 | orangefs_get_bufmap_init()); |
1182fca3 | 248 | |
ce6c414e | 249 | finish_wait(&orangefs_bufmap_init_waitq, &wait_entry); |
1182fca3 | 250 | |
7d221485 | 251 | if (orangefs_get_bufmap_init() == 0) { |
1182fca3 MM |
252 | gossip_err("%s:The shared memory system has not started in %d seconds after the client core restarted. Aborting user's request(%s).\n", |
253 | __func__, | |
8bb8aefd | 254 | ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS, |
1182fca3 MM |
255 | get_opname_string(op)); |
256 | return -EIO; | |
257 | } | |
258 | ||
259 | /* | |
260 | * Return to the calling function and re-populate a | |
261 | * shared memory buffer. | |
262 | */ | |
263 | return -EAGAIN; | |
264 | } | |
265 | } | |
266 | ||
267 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
8bb8aefd | 268 | "orangefs: service_operation %s returning: %d for %p.\n", |
1182fca3 MM |
269 | op_name, |
270 | ret, | |
271 | op); | |
272 | return ret; | |
273 | } | |
274 | ||
e07db0a2 | 275 | static void orangefs_clean_up_interrupted_operation(struct orangefs_kernel_op_s *op) |
1182fca3 MM |
276 | { |
277 | /* | |
278 | * handle interrupted cases depending on what state we were in when | |
279 | * the interruption is detected. there is a coarse grained lock | |
280 | * across the operation. | |
281 | * | |
282 | * NOTE: be sure not to reverse lock ordering by locking an op lock | |
283 | * while holding the request_list lock. Here, we first lock the op | |
284 | * and then lock the appropriate list. | |
285 | */ | |
286 | if (!op) { | |
287 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
288 | "%s: op is null, ignoring\n", | |
289 | __func__); | |
290 | return; | |
291 | } | |
292 | ||
293 | /* | |
294 | * one more sanity check, make sure it's in one of the possible states | |
295 | * or don't try to cancel it | |
296 | */ | |
297 | if (!(op_state_waiting(op) || | |
298 | op_state_in_progress(op) || | |
299 | op_state_serviced(op) || | |
300 | op_state_purged(op))) { | |
301 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
302 | "%s: op %p not in a valid state (%0x), " | |
303 | "ignoring\n", | |
304 | __func__, | |
305 | op, | |
306 | op->op_state); | |
307 | return; | |
308 | } | |
309 | ||
310 | spin_lock(&op->lock); | |
ed42fe05 | 311 | op->op_state |= OP_VFS_STATE_GIVEN_UP; |
1182fca3 MM |
312 | |
313 | if (op_state_waiting(op)) { | |
314 | /* | |
315 | * upcall hasn't been read; remove op from upcall request | |
316 | * list. | |
317 | */ | |
318 | spin_unlock(&op->lock); | |
ed42fe05 AV |
319 | spin_lock(&orangefs_request_list_lock); |
320 | list_del(&op->list); | |
321 | spin_unlock(&orangefs_request_list_lock); | |
1182fca3 MM |
322 | gossip_debug(GOSSIP_WAIT_DEBUG, |
323 | "Interrupted: Removed op %p from request_list\n", | |
324 | op); | |
325 | } else if (op_state_in_progress(op)) { | |
326 | /* op must be removed from the in progress htable */ | |
327 | spin_unlock(&op->lock); | |
328 | spin_lock(&htable_ops_in_progress_lock); | |
329 | list_del(&op->list); | |
330 | spin_unlock(&htable_ops_in_progress_lock); | |
331 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
332 | "Interrupted: Removed op %p" | |
333 | " from htable_ops_in_progress\n", | |
334 | op); | |
335 | } else if (!op_state_serviced(op)) { | |
336 | spin_unlock(&op->lock); | |
337 | gossip_err("interrupted operation is in a weird state 0x%x\n", | |
338 | op->op_state); | |
84d02150 MM |
339 | } else { |
340 | /* | |
341 | * It is not intended for execution to flow here, | |
342 | * but having this unlock here makes sparse happy. | |
343 | */ | |
344 | gossip_err("%s: can't get here.\n", __func__); | |
345 | spin_unlock(&op->lock); | |
1182fca3 MM |
346 | } |
347 | } | |
348 | ||
349 | /* | |
350 | * sleeps on waitqueue waiting for matching downcall. | |
351 | * if client-core finishes servicing, then we are good to go. | |
352 | * else if client-core exits, we get woken up here, and retry with a timeout | |
353 | * | |
354 | * Post when this call returns to the caller, the specified op will no | |
355 | * longer be on any list or htable. | |
356 | * | |
357 | * Returns 0 on success and -errno on failure | |
358 | * Errors are: | |
359 | * EAGAIN in case we want the caller to requeue and try again.. | |
360 | * EINTR/EIO/ETIMEDOUT indicating we are done trying to service this | |
361 | * operation since client-core seems to be exiting too often | |
362 | * or if we were interrupted. | |
363 | */ | |
b7ae37b0 | 364 | static int wait_for_matching_downcall(struct orangefs_kernel_op_s *op) |
1182fca3 MM |
365 | { |
366 | int ret = -EINVAL; | |
ce6c414e | 367 | DEFINE_WAIT(wait_entry); |
1182fca3 MM |
368 | |
369 | while (1) { | |
1182fca3 | 370 | spin_lock(&op->lock); |
ce6c414e | 371 | prepare_to_wait(&op->waitq, &wait_entry, TASK_INTERRUPTIBLE); |
1182fca3 MM |
372 | if (op_state_serviced(op)) { |
373 | spin_unlock(&op->lock); | |
374 | ret = 0; | |
375 | break; | |
376 | } | |
377 | spin_unlock(&op->lock); | |
378 | ||
379 | if (!signal_pending(current)) { | |
380 | /* | |
381 | * if this was our first attempt and client-core | |
382 | * has not purged our operation, we are happy to | |
383 | * simply wait | |
384 | */ | |
385 | spin_lock(&op->lock); | |
386 | if (op->attempts == 0 && !op_state_purged(op)) { | |
387 | spin_unlock(&op->lock); | |
388 | schedule(); | |
389 | } else { | |
390 | spin_unlock(&op->lock); | |
391 | /* | |
392 | * subsequent attempts, we retry exactly once | |
393 | * with timeouts | |
394 | */ | |
395 | if (!schedule_timeout(MSECS_TO_JIFFIES | |
396 | (1000 * op_timeout_secs))) { | |
397 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
398 | "*** %s:" | |
399 | " operation timed out (tag" | |
400 | " %llu, %p, att %d)\n", | |
401 | __func__, | |
402 | llu(op->tag), | |
403 | op, | |
404 | op->attempts); | |
405 | ret = -ETIMEDOUT; | |
8bb8aefd | 406 | orangefs_clean_up_interrupted_operation |
1182fca3 MM |
407 | (op); |
408 | break; | |
409 | } | |
410 | } | |
411 | spin_lock(&op->lock); | |
412 | op->attempts++; | |
413 | /* | |
414 | * if the operation was purged in the meantime, it | |
415 | * is better to requeue it afresh but ensure that | |
416 | * we have not been purged repeatedly. This could | |
417 | * happen if client-core crashes when an op | |
418 | * is being serviced, so we requeue the op, client | |
419 | * core crashes again so we requeue the op, client | |
420 | * core starts, and so on... | |
421 | */ | |
422 | if (op_state_purged(op)) { | |
8bb8aefd | 423 | ret = (op->attempts < ORANGEFS_PURGE_RETRY_COUNT) ? |
1182fca3 MM |
424 | -EAGAIN : |
425 | -EIO; | |
426 | spin_unlock(&op->lock); | |
427 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
428 | "*** %s:" | |
429 | " operation purged (tag " | |
430 | "%llu, %p, att %d)\n", | |
431 | __func__, | |
432 | llu(op->tag), | |
433 | op, | |
434 | op->attempts); | |
8bb8aefd | 435 | orangefs_clean_up_interrupted_operation(op); |
1182fca3 MM |
436 | break; |
437 | } | |
438 | spin_unlock(&op->lock); | |
439 | continue; | |
440 | } | |
441 | ||
442 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
443 | "*** %s:" | |
444 | " operation interrupted by a signal (tag " | |
445 | "%llu, op %p)\n", | |
446 | __func__, | |
447 | llu(op->tag), | |
448 | op); | |
8bb8aefd | 449 | orangefs_clean_up_interrupted_operation(op); |
1182fca3 MM |
450 | ret = -EINTR; |
451 | break; | |
452 | } | |
453 | ||
1182fca3 | 454 | spin_lock(&op->lock); |
ce6c414e | 455 | finish_wait(&op->waitq, &wait_entry); |
1182fca3 MM |
456 | spin_unlock(&op->lock); |
457 | ||
458 | return ret; | |
459 | } | |
460 | ||
461 | /* | |
462 | * similar to wait_for_matching_downcall(), but used in the special case | |
463 | * of I/O cancellations. | |
464 | * | |
465 | * Note we need a special wait function because if this is called we already | |
466 | * know that a signal is pending in current and need to service the | |
467 | * cancellation upcall anyway. the only way to exit this is to either | |
468 | * timeout or have the cancellation be serviced properly. | |
469 | */ | |
b7ae37b0 | 470 | static int wait_for_cancellation_downcall(struct orangefs_kernel_op_s *op) |
1182fca3 MM |
471 | { |
472 | int ret = -EINVAL; | |
ce6c414e | 473 | DEFINE_WAIT(wait_entry); |
1182fca3 MM |
474 | |
475 | while (1) { | |
1182fca3 | 476 | spin_lock(&op->lock); |
ce6c414e | 477 | prepare_to_wait(&op->waitq, &wait_entry, TASK_INTERRUPTIBLE); |
1182fca3 MM |
478 | if (op_state_serviced(op)) { |
479 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
480 | "%s:op-state is SERVICED.\n", | |
481 | __func__); | |
482 | spin_unlock(&op->lock); | |
483 | ret = 0; | |
484 | break; | |
485 | } | |
486 | spin_unlock(&op->lock); | |
487 | ||
488 | if (signal_pending(current)) { | |
489 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
490 | "%s:operation interrupted by a signal (tag" | |
491 | " %llu, op %p)\n", | |
492 | __func__, | |
493 | llu(op->tag), | |
494 | op); | |
8bb8aefd | 495 | orangefs_clean_up_interrupted_operation(op); |
1182fca3 MM |
496 | ret = -EINTR; |
497 | break; | |
498 | } | |
499 | ||
500 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
501 | "%s:About to call schedule_timeout.\n", | |
502 | __func__); | |
503 | ret = | |
504 | schedule_timeout(MSECS_TO_JIFFIES(1000 * op_timeout_secs)); | |
505 | ||
506 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
507 | "%s:Value returned from schedule_timeout(%d).\n", | |
508 | __func__, | |
509 | ret); | |
510 | if (!ret) { | |
511 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
512 | "%s:*** operation timed out: %p\n", | |
513 | __func__, | |
514 | op); | |
8bb8aefd | 515 | orangefs_clean_up_interrupted_operation(op); |
1182fca3 MM |
516 | ret = -ETIMEDOUT; |
517 | break; | |
518 | } | |
519 | ||
520 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
521 | "%s:Breaking out of loop, regardless of value returned by schedule_timeout.\n", | |
522 | __func__); | |
523 | ret = -ETIMEDOUT; | |
524 | break; | |
525 | } | |
526 | ||
1182fca3 | 527 | spin_lock(&op->lock); |
ce6c414e | 528 | finish_wait(&op->waitq, &wait_entry); |
1182fca3 MM |
529 | spin_unlock(&op->lock); |
530 | ||
531 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
532 | "%s:returning ret(%d)\n", | |
533 | __func__, | |
534 | ret); | |
535 | ||
536 | return ret; | |
537 | } |