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_matching_downcall(struct orangefs_kernel_op_s *); |
20 | ||
1182fca3 MM |
21 | /* |
22 | * What we do in this function is to walk the list of operations that are | |
23 | * present in the request queue and mark them as purged. | |
24 | * NOTE: This is called from the device close after client-core has | |
25 | * guaranteed that no new operations could appear on the list since the | |
26 | * client-core is anyway going to exit. | |
27 | */ | |
28 | void purge_waiting_ops(void) | |
29 | { | |
8bb8aefd | 30 | struct orangefs_kernel_op_s *op; |
1182fca3 | 31 | |
8bb8aefd YL |
32 | spin_lock(&orangefs_request_list_lock); |
33 | list_for_each_entry(op, &orangefs_request_list, list) { | |
1182fca3 MM |
34 | gossip_debug(GOSSIP_WAIT_DEBUG, |
35 | "pvfs2-client-core: purging op tag %llu %s\n", | |
36 | llu(op->tag), | |
37 | get_opname_string(op)); | |
1182fca3 | 38 | set_op_state_purged(op); |
1182fca3 | 39 | } |
8bb8aefd | 40 | spin_unlock(&orangefs_request_list_lock); |
1182fca3 MM |
41 | } |
42 | ||
fc916da5 | 43 | static inline void |
78699e29 | 44 | __add_op_to_request_list(struct orangefs_kernel_op_s *op) |
fc916da5 | 45 | { |
fc916da5 AV |
46 | spin_lock(&op->lock); |
47 | set_op_state_waiting(op); | |
48 | list_add_tail(&op->list, &orangefs_request_list); | |
fc916da5 AV |
49 | spin_unlock(&op->lock); |
50 | wake_up_interruptible(&orangefs_request_list_waitq); | |
51 | } | |
52 | ||
78699e29 AV |
53 | static inline void |
54 | add_op_to_request_list(struct orangefs_kernel_op_s *op) | |
55 | { | |
56 | spin_lock(&orangefs_request_list_lock); | |
57 | __add_op_to_request_list(op); | |
58 | spin_unlock(&orangefs_request_list_lock); | |
59 | } | |
60 | ||
fc916da5 AV |
61 | static inline |
62 | void add_priority_op_to_request_list(struct orangefs_kernel_op_s *op) | |
63 | { | |
64 | spin_lock(&orangefs_request_list_lock); | |
65 | spin_lock(&op->lock); | |
66 | set_op_state_waiting(op); | |
67 | ||
68 | list_add(&op->list, &orangefs_request_list); | |
69 | spin_unlock(&orangefs_request_list_lock); | |
70 | spin_unlock(&op->lock); | |
71 | wake_up_interruptible(&orangefs_request_list_waitq); | |
72 | } | |
73 | ||
1182fca3 | 74 | /* |
8bb8aefd | 75 | * submits a ORANGEFS operation and waits for it to complete |
1182fca3 MM |
76 | * |
77 | * Note op->downcall.status will contain the status of the operation (in | |
78 | * errno format), whether provided by pvfs2-client or a result of failure to | |
79 | * service the operation. If the caller wishes to distinguish, then | |
80 | * op->state can be checked to see if it was serviced or not. | |
81 | * | |
82 | * Returns contents of op->downcall.status for convenience | |
83 | */ | |
8bb8aefd | 84 | int service_operation(struct orangefs_kernel_op_s *op, |
1182fca3 MM |
85 | const char *op_name, |
86 | int flags) | |
87 | { | |
88 | /* flags to modify behavior */ | |
89 | sigset_t orig_sigset; | |
90 | int ret = 0; | |
91 | ||
ce6c414e | 92 | DEFINE_WAIT(wait_entry); |
1182fca3 MM |
93 | |
94 | op->upcall.tgid = current->tgid; | |
95 | op->upcall.pid = current->pid; | |
96 | ||
97 | retry_servicing: | |
98 | op->downcall.status = 0; | |
99 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
8bb8aefd | 100 | "orangefs: service_operation: %s %p\n", |
1182fca3 MM |
101 | op_name, |
102 | op); | |
103 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
8bb8aefd | 104 | "orangefs: operation posted by process: %s, pid: %i\n", |
1182fca3 MM |
105 | current->comm, |
106 | current->pid); | |
107 | ||
108 | /* mask out signals if this operation is not to be interrupted */ | |
8bb8aefd | 109 | if (!(flags & ORANGEFS_OP_INTERRUPTIBLE)) |
c146c0b8 | 110 | orangefs_block_signals(&orig_sigset); |
1182fca3 | 111 | |
8bb8aefd | 112 | if (!(flags & ORANGEFS_OP_NO_SEMAPHORE)) { |
1182fca3 MM |
113 | ret = mutex_lock_interruptible(&request_mutex); |
114 | /* | |
115 | * check to see if we were interrupted while waiting for | |
116 | * semaphore | |
117 | */ | |
118 | if (ret < 0) { | |
8bb8aefd | 119 | if (!(flags & ORANGEFS_OP_INTERRUPTIBLE)) |
c146c0b8 | 120 | orangefs_set_signals(&orig_sigset); |
1182fca3 MM |
121 | op->downcall.status = ret; |
122 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
8bb8aefd | 123 | "orangefs: service_operation interrupted.\n"); |
1182fca3 MM |
124 | return ret; |
125 | } | |
126 | } | |
127 | ||
128 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
129 | "%s:About to call is_daemon_in_service().\n", | |
130 | __func__); | |
131 | ||
132 | if (is_daemon_in_service() < 0) { | |
133 | /* | |
134 | * By incrementing the per-operation attempt counter, we | |
135 | * directly go into the timeout logic while waiting for | |
136 | * the matching downcall to be read | |
137 | */ | |
138 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
139 | "%s:client core is NOT in service(%d).\n", | |
140 | __func__, | |
141 | is_daemon_in_service()); | |
142 | op->attempts++; | |
143 | } | |
144 | ||
145 | /* queue up the operation */ | |
8bb8aefd | 146 | if (flags & ORANGEFS_OP_PRIORITY) { |
1182fca3 MM |
147 | add_priority_op_to_request_list(op); |
148 | } else { | |
149 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
150 | "%s:About to call add_op_to_request_list().\n", | |
151 | __func__); | |
152 | add_op_to_request_list(op); | |
153 | } | |
154 | ||
8bb8aefd | 155 | if (!(flags & ORANGEFS_OP_NO_SEMAPHORE)) |
1182fca3 MM |
156 | mutex_unlock(&request_mutex); |
157 | ||
158 | /* | |
159 | * If we are asked to service an asynchronous operation from | |
160 | * VFS perspective, we are done. | |
161 | */ | |
8bb8aefd | 162 | if (flags & ORANGEFS_OP_ASYNC) |
1182fca3 MM |
163 | return 0; |
164 | ||
78699e29 | 165 | ret = wait_for_matching_downcall(op); |
1182fca3 MM |
166 | |
167 | if (ret < 0) { | |
168 | /* failed to get matching downcall */ | |
169 | if (ret == -ETIMEDOUT) { | |
8bb8aefd | 170 | gossip_err("orangefs: %s -- wait timed out; aborting attempt.\n", |
1182fca3 MM |
171 | op_name); |
172 | } | |
173 | op->downcall.status = ret; | |
174 | } else { | |
175 | /* got matching downcall; make sure status is in errno format */ | |
176 | op->downcall.status = | |
8bb8aefd | 177 | orangefs_normalize_to_errno(op->downcall.status); |
1182fca3 MM |
178 | ret = op->downcall.status; |
179 | } | |
180 | ||
8bb8aefd | 181 | if (!(flags & ORANGEFS_OP_INTERRUPTIBLE)) |
c146c0b8 | 182 | orangefs_set_signals(&orig_sigset); |
1182fca3 MM |
183 | |
184 | BUG_ON(ret != op->downcall.status); | |
185 | /* retry if operation has not been serviced and if requested */ | |
186 | if (!op_state_serviced(op) && op->downcall.status == -EAGAIN) { | |
187 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
8bb8aefd | 188 | "orangefs: tag %llu (%s)" |
1182fca3 MM |
189 | " -- operation to be retried (%d attempt)\n", |
190 | llu(op->tag), | |
191 | op_name, | |
192 | op->attempts + 1); | |
193 | ||
194 | if (!op->uses_shared_memory) | |
195 | /* | |
196 | * this operation doesn't use the shared memory | |
197 | * system | |
198 | */ | |
199 | goto retry_servicing; | |
200 | ||
201 | /* op uses shared memory */ | |
7d221485 | 202 | if (orangefs_get_bufmap_init() == 0) { |
6ebcc3fc | 203 | WARN_ON(1); |
1182fca3 MM |
204 | /* |
205 | * This operation uses the shared memory system AND | |
206 | * the system is not yet ready. This situation occurs | |
207 | * when the client-core is restarted AND there were | |
208 | * operations waiting to be processed or were already | |
209 | * in process. | |
210 | */ | |
211 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
212 | "uses_shared_memory is true.\n"); | |
213 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
214 | "Client core in-service status(%d).\n", | |
215 | is_daemon_in_service()); | |
216 | gossip_debug(GOSSIP_WAIT_DEBUG, "bufmap_init:%d.\n", | |
7d221485 | 217 | orangefs_get_bufmap_init()); |
1182fca3 MM |
218 | gossip_debug(GOSSIP_WAIT_DEBUG, |
219 | "operation's status is 0x%0x.\n", | |
220 | op->op_state); | |
221 | ||
222 | /* | |
223 | * let process sleep for a few seconds so shared | |
224 | * memory system can be initialized. | |
225 | */ | |
ce6c414e MM |
226 | prepare_to_wait(&orangefs_bufmap_init_waitq, |
227 | &wait_entry, | |
228 | TASK_INTERRUPTIBLE); | |
1182fca3 | 229 | |
1182fca3 | 230 | /* |
8bb8aefd | 231 | * Wait for orangefs_bufmap_initialize() to wake me up |
1182fca3 MM |
232 | * within the allotted time. |
233 | */ | |
727cbfea AV |
234 | ret = schedule_timeout( |
235 | ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS * HZ); | |
1182fca3 MM |
236 | |
237 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
238 | "Value returned from schedule_timeout:" | |
239 | "%d.\n", | |
240 | ret); | |
241 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
242 | "Is shared memory available? (%d).\n", | |
7d221485 | 243 | orangefs_get_bufmap_init()); |
1182fca3 | 244 | |
ce6c414e | 245 | finish_wait(&orangefs_bufmap_init_waitq, &wait_entry); |
1182fca3 | 246 | |
7d221485 | 247 | if (orangefs_get_bufmap_init() == 0) { |
1182fca3 MM |
248 | gossip_err("%s:The shared memory system has not started in %d seconds after the client core restarted. Aborting user's request(%s).\n", |
249 | __func__, | |
8bb8aefd | 250 | ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS, |
1182fca3 MM |
251 | get_opname_string(op)); |
252 | return -EIO; | |
253 | } | |
254 | ||
255 | /* | |
256 | * Return to the calling function and re-populate a | |
257 | * shared memory buffer. | |
258 | */ | |
259 | return -EAGAIN; | |
260 | } | |
261 | } | |
262 | ||
263 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
8bb8aefd | 264 | "orangefs: service_operation %s returning: %d for %p.\n", |
1182fca3 MM |
265 | op_name, |
266 | ret, | |
267 | op); | |
268 | return ret; | |
269 | } | |
270 | ||
78699e29 AV |
271 | bool orangefs_cancel_op_in_progress(struct orangefs_kernel_op_s *op) |
272 | { | |
273 | u64 tag = op->tag; | |
274 | if (!op_state_in_progress(op)) | |
275 | return false; | |
276 | ||
277 | op->slot_to_free = op->upcall.req.io.buf_index; | |
278 | memset(&op->upcall, 0, sizeof(op->upcall)); | |
279 | memset(&op->downcall, 0, sizeof(op->downcall)); | |
280 | op->upcall.type = ORANGEFS_VFS_OP_CANCEL; | |
281 | op->upcall.req.cancel.op_tag = tag; | |
282 | op->downcall.type = ORANGEFS_VFS_OP_INVALID; | |
283 | op->downcall.status = -1; | |
284 | orangefs_new_tag(op); | |
285 | ||
286 | spin_lock(&orangefs_request_list_lock); | |
287 | /* orangefs_request_list_lock is enough of a barrier here */ | |
288 | if (!__is_daemon_in_service()) { | |
289 | spin_unlock(&orangefs_request_list_lock); | |
290 | return false; | |
291 | } | |
292 | __add_op_to_request_list(op); | |
293 | spin_unlock(&orangefs_request_list_lock); | |
294 | ||
295 | gossip_debug(GOSSIP_UTILS_DEBUG, | |
296 | "Attempting ORANGEFS operation cancellation of tag %llu\n", | |
297 | llu(tag)); | |
298 | return true; | |
299 | } | |
300 | ||
e07db0a2 | 301 | static void orangefs_clean_up_interrupted_operation(struct orangefs_kernel_op_s *op) |
1182fca3 MM |
302 | { |
303 | /* | |
304 | * handle interrupted cases depending on what state we were in when | |
305 | * the interruption is detected. there is a coarse grained lock | |
306 | * across the operation. | |
307 | * | |
eab9b389 | 308 | * Called with op->lock held. |
1182fca3 | 309 | */ |
ed42fe05 | 310 | op->op_state |= OP_VFS_STATE_GIVEN_UP; |
1182fca3 MM |
311 | |
312 | if (op_state_waiting(op)) { | |
313 | /* | |
314 | * upcall hasn't been read; remove op from upcall request | |
315 | * list. | |
316 | */ | |
317 | spin_unlock(&op->lock); | |
ed42fe05 AV |
318 | spin_lock(&orangefs_request_list_lock); |
319 | list_del(&op->list); | |
320 | spin_unlock(&orangefs_request_list_lock); | |
1182fca3 MM |
321 | gossip_debug(GOSSIP_WAIT_DEBUG, |
322 | "Interrupted: Removed op %p from request_list\n", | |
323 | op); | |
324 | } else if (op_state_in_progress(op)) { | |
325 | /* op must be removed from the in progress htable */ | |
326 | spin_unlock(&op->lock); | |
327 | spin_lock(&htable_ops_in_progress_lock); | |
328 | list_del(&op->list); | |
329 | spin_unlock(&htable_ops_in_progress_lock); | |
330 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
331 | "Interrupted: Removed op %p" | |
332 | " from htable_ops_in_progress\n", | |
333 | op); | |
334 | } else if (!op_state_serviced(op)) { | |
335 | spin_unlock(&op->lock); | |
336 | gossip_err("interrupted operation is in a weird state 0x%x\n", | |
337 | op->op_state); | |
84d02150 MM |
338 | } else { |
339 | /* | |
340 | * It is not intended for execution to flow here, | |
341 | * but having this unlock here makes sparse happy. | |
342 | */ | |
343 | gossip_err("%s: can't get here.\n", __func__); | |
344 | spin_unlock(&op->lock); | |
1182fca3 MM |
345 | } |
346 | } | |
347 | ||
348 | /* | |
349 | * sleeps on waitqueue waiting for matching downcall. | |
350 | * if client-core finishes servicing, then we are good to go. | |
351 | * else if client-core exits, we get woken up here, and retry with a timeout | |
352 | * | |
353 | * Post when this call returns to the caller, the specified op will no | |
354 | * longer be on any list or htable. | |
355 | * | |
356 | * Returns 0 on success and -errno on failure | |
357 | * Errors are: | |
358 | * EAGAIN in case we want the caller to requeue and try again.. | |
359 | * EINTR/EIO/ETIMEDOUT indicating we are done trying to service this | |
360 | * operation since client-core seems to be exiting too often | |
361 | * or if we were interrupted. | |
362 | */ | |
b7ae37b0 | 363 | static int wait_for_matching_downcall(struct orangefs_kernel_op_s *op) |
1182fca3 MM |
364 | { |
365 | int ret = -EINVAL; | |
ce6c414e | 366 | DEFINE_WAIT(wait_entry); |
1182fca3 MM |
367 | |
368 | while (1) { | |
1182fca3 | 369 | spin_lock(&op->lock); |
ce6c414e | 370 | prepare_to_wait(&op->waitq, &wait_entry, TASK_INTERRUPTIBLE); |
1182fca3 MM |
371 | if (op_state_serviced(op)) { |
372 | spin_unlock(&op->lock); | |
373 | ret = 0; | |
374 | break; | |
375 | } | |
1182fca3 | 376 | |
70c6ea26 AV |
377 | if (unlikely(signal_pending(current))) { |
378 | gossip_debug(GOSSIP_WAIT_DEBUG, | |
379 | "*** %s:" | |
380 | " operation interrupted by a signal (tag " | |
381 | "%llu, op %p)\n", | |
382 | __func__, | |
383 | llu(op->tag), | |
384 | op); | |
385 | orangefs_clean_up_interrupted_operation(op); | |
386 | ret = -EINTR; | |
387 | break; | |
388 | } | |
389 | ||
390 | /* | |
391 | * if this was our first attempt and client-core | |
392 | * has not purged our operation, we are happy to | |
393 | * simply wait | |
394 | */ | |
70c6ea26 AV |
395 | if (op->attempts == 0 && !op_state_purged(op)) { |
396 | spin_unlock(&op->lock); | |
397 | schedule(); | |
398 | } else { | |
399 | spin_unlock(&op->lock); | |
1182fca3 | 400 | /* |
70c6ea26 AV |
401 | * subsequent attempts, we retry exactly once |
402 | * with timeouts | |
1182fca3 | 403 | */ |
727cbfea | 404 | if (!schedule_timeout(op_timeout_secs * HZ)) { |
1182fca3 MM |
405 | gossip_debug(GOSSIP_WAIT_DEBUG, |
406 | "*** %s:" | |
70c6ea26 AV |
407 | " operation timed out (tag" |
408 | " %llu, %p, att %d)\n", | |
1182fca3 MM |
409 | __func__, |
410 | llu(op->tag), | |
411 | op, | |
412 | op->attempts); | |
70c6ea26 | 413 | ret = -ETIMEDOUT; |
eab9b389 | 414 | spin_lock(&op->lock); |
8bb8aefd | 415 | orangefs_clean_up_interrupted_operation(op); |
1182fca3 MM |
416 | break; |
417 | } | |
70c6ea26 AV |
418 | } |
419 | spin_lock(&op->lock); | |
420 | op->attempts++; | |
421 | /* | |
422 | * if the operation was purged in the meantime, it | |
423 | * is better to requeue it afresh but ensure that | |
424 | * we have not been purged repeatedly. This could | |
425 | * happen if client-core crashes when an op | |
426 | * is being serviced, so we requeue the op, client | |
427 | * core crashes again so we requeue the op, client | |
428 | * core starts, and so on... | |
429 | */ | |
430 | if (op_state_purged(op)) { | |
431 | ret = (op->attempts < ORANGEFS_PURGE_RETRY_COUNT) ? | |
432 | -EAGAIN : | |
433 | -EIO; | |
70c6ea26 AV |
434 | gossip_debug(GOSSIP_WAIT_DEBUG, |
435 | "*** %s:" | |
436 | " operation purged (tag " | |
437 | "%llu, %p, att %d)\n", | |
438 | __func__, | |
439 | llu(op->tag), | |
440 | op, | |
441 | op->attempts); | |
442 | orangefs_clean_up_interrupted_operation(op); | |
443 | break; | |
1182fca3 | 444 | } |
70c6ea26 | 445 | spin_unlock(&op->lock); |
1182fca3 MM |
446 | } |
447 | ||
1182fca3 | 448 | spin_lock(&op->lock); |
ce6c414e | 449 | finish_wait(&op->waitq, &wait_entry); |
1182fca3 MM |
450 | spin_unlock(&op->lock); |
451 | ||
452 | return ret; | |
453 | } |