Commit | Line | Data |
---|---|---|
d7e09d03 PT |
1 | /* |
2 | * GPL HEADER START | |
3 | * | |
4 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or modify | |
7 | * it under the terms of the GNU General Public License version 2 only, | |
8 | * as published by the Free Software Foundation. | |
9 | * | |
10 | * This program is distributed in the hope that it will be useful, but | |
11 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
13 | * General Public License version 2 for more details (a copy is included | |
14 | * in the LICENSE file that accompanied this code). | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
17 | * version 2 along with this program; If not, see | |
18 | * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf | |
19 | * | |
20 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
21 | * CA 95054 USA or visit www.sun.com if you need additional information or | |
22 | * have any questions. | |
23 | * | |
24 | * GPL HEADER END | |
25 | */ | |
26 | /* | |
27 | * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. | |
28 | * Use is subject to license terms. | |
29 | * | |
30 | * Copyright (c) 2010, 2012, Intel Corporation. | |
31 | */ | |
32 | /* | |
33 | * This file is part of Lustre, http://www.lustre.org/ | |
34 | * Lustre is a trademark of Sun Microsystems, Inc. | |
35 | */ | |
36 | /** | |
37 | * This file contains Asynchronous System Trap (AST) handlers and related | |
38 | * LDLM request-processing routines. | |
39 | * | |
40 | * An AST is a callback issued on a lock when its state is changed. There are | |
41 | * several different types of ASTs (callbacks) registered for each lock: | |
42 | * | |
43 | * - completion AST: when a lock is enqueued by some process, but cannot be | |
44 | * granted immediately due to other conflicting locks on the same resource, | |
45 | * the completion AST is sent to notify the caller when the lock is | |
46 | * eventually granted | |
47 | * | |
48 | * - blocking AST: when a lock is granted to some process, if another process | |
49 | * enqueues a conflicting (blocking) lock on a resource, a blocking AST is | |
50 | * sent to notify the holder(s) of the lock(s) of the conflicting lock | |
51 | * request. The lock holder(s) must release their lock(s) on that resource in | |
52 | * a timely manner or be evicted by the server. | |
53 | * | |
54 | * - glimpse AST: this is used when a process wants information about a lock | |
55 | * (i.e. the lock value block (LVB)) but does not necessarily require holding | |
56 | * the lock. If the resource is locked, the lock holder(s) are sent glimpse | |
57 | * ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL | |
58 | * their lock(s) if they are idle. If the resource is not locked, the server | |
59 | * may grant the lock. | |
60 | */ | |
61 | ||
62 | #define DEBUG_SUBSYSTEM S_LDLM | |
63 | ||
e27db149 GKH |
64 | #include "../include/lustre_dlm.h" |
65 | #include "../include/obd_class.h" | |
66 | #include "../include/obd.h" | |
d7e09d03 PT |
67 | |
68 | #include "ldlm_internal.h" | |
69 | ||
70 | int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT; | |
8cc7b4b9 PT |
71 | module_param(ldlm_enqueue_min, int, 0644); |
72 | MODULE_PARM_DESC(ldlm_enqueue_min, "lock enqueue timeout minimum"); | |
d7e09d03 PT |
73 | |
74 | /* in client side, whether the cached locks will be canceled before replay */ | |
75 | unsigned int ldlm_cancel_unused_locks_before_replay = 1; | |
76 | ||
77 | static void interrupted_completion_wait(void *data) | |
78 | { | |
79 | } | |
80 | ||
81 | struct lock_wait_data { | |
82 | struct ldlm_lock *lwd_lock; | |
83 | __u32 lwd_conn_cnt; | |
84 | }; | |
85 | ||
86 | struct ldlm_async_args { | |
87 | struct lustre_handle lock_handle; | |
88 | }; | |
89 | ||
90 | int ldlm_expired_completion_wait(void *data) | |
91 | { | |
92 | struct lock_wait_data *lwd = data; | |
93 | struct ldlm_lock *lock = lwd->lwd_lock; | |
94 | struct obd_import *imp; | |
95 | struct obd_device *obd; | |
96 | ||
d7e09d03 | 97 | if (lock->l_conn_export == NULL) { |
a649ad1d | 98 | static unsigned long next_dump = 0, last_dump = 0; |
d7e09d03 | 99 | |
d7e09d03 PT |
100 | LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", " |
101 | CFS_DURATION_T"s ago)\n", | |
102 | lock->l_last_activity, | |
7264b8a5 | 103 | cfs_time_sub(get_seconds(), |
d7e09d03 PT |
104 | lock->l_last_activity)); |
105 | LDLM_DEBUG(lock, "lock timed out (enqueued at "CFS_TIME_T", " | |
106 | CFS_DURATION_T"s ago); not entering recovery in " | |
107 | "server code, just going back to sleep", | |
108 | lock->l_last_activity, | |
7264b8a5 | 109 | cfs_time_sub(get_seconds(), |
d7e09d03 PT |
110 | lock->l_last_activity)); |
111 | if (cfs_time_after(cfs_time_current(), next_dump)) { | |
112 | last_dump = next_dump; | |
113 | next_dump = cfs_time_shift(300); | |
114 | ldlm_namespace_dump(D_DLMTRACE, | |
115 | ldlm_lock_to_ns(lock)); | |
116 | if (last_dump == 0) | |
117 | libcfs_debug_dumplog(); | |
118 | } | |
0a3bdb00 | 119 | return 0; |
d7e09d03 PT |
120 | } |
121 | ||
122 | obd = lock->l_conn_export->exp_obd; | |
123 | imp = obd->u.cli.cl_import; | |
124 | ptlrpc_fail_import(imp, lwd->lwd_conn_cnt); | |
125 | LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", " | |
126 | CFS_DURATION_T"s ago), entering recovery for %s@%s", | |
127 | lock->l_last_activity, | |
7264b8a5 | 128 | cfs_time_sub(get_seconds(), lock->l_last_activity), |
d7e09d03 PT |
129 | obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid); |
130 | ||
0a3bdb00 | 131 | return 0; |
d7e09d03 PT |
132 | } |
133 | EXPORT_SYMBOL(ldlm_expired_completion_wait); | |
134 | ||
135 | /* We use the same basis for both server side and client side functions | |
136 | from a single node. */ | |
137 | int ldlm_get_enq_timeout(struct ldlm_lock *lock) | |
138 | { | |
139 | int timeout = at_get(ldlm_lock_to_ns_at(lock)); | |
140 | if (AT_OFF) | |
141 | return obd_timeout / 2; | |
142 | /* Since these are non-updating timeouts, we should be conservative. | |
143 | It would be nice to have some kind of "early reply" mechanism for | |
144 | lock callbacks too... */ | |
145 | timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */ | |
146 | return max(timeout, ldlm_enqueue_min); | |
147 | } | |
148 | EXPORT_SYMBOL(ldlm_get_enq_timeout); | |
149 | ||
150 | /** | |
151 | * Helper function for ldlm_completion_ast(), updating timings when lock is | |
152 | * actually granted. | |
153 | */ | |
154 | static int ldlm_completion_tail(struct ldlm_lock *lock) | |
155 | { | |
156 | long delay; | |
157 | int result; | |
158 | ||
f2145eae | 159 | if (lock->l_flags & (LDLM_FL_DESTROYED | LDLM_FL_FAILED)) { |
d7e09d03 PT |
160 | LDLM_DEBUG(lock, "client-side enqueue: destroyed"); |
161 | result = -EIO; | |
162 | } else { | |
7264b8a5 | 163 | delay = cfs_time_sub(get_seconds(), |
d7e09d03 PT |
164 | lock->l_last_activity); |
165 | LDLM_DEBUG(lock, "client-side enqueue: granted after " | |
166 | CFS_DURATION_T"s", delay); | |
167 | ||
168 | /* Update our time estimate */ | |
169 | at_measured(ldlm_lock_to_ns_at(lock), | |
170 | delay); | |
171 | result = 0; | |
172 | } | |
173 | return result; | |
174 | } | |
175 | ||
176 | /** | |
177 | * Implementation of ->l_completion_ast() for a client, that doesn't wait | |
178 | * until lock is granted. Suitable for locks enqueued through ptlrpcd, of | |
179 | * other threads that cannot block for long. | |
180 | */ | |
181 | int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data) | |
182 | { | |
d7e09d03 PT |
183 | if (flags == LDLM_FL_WAIT_NOREPROC) { |
184 | LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock"); | |
0a3bdb00 | 185 | return 0; |
d7e09d03 PT |
186 | } |
187 | ||
188 | if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED | | |
189 | LDLM_FL_BLOCK_CONV))) { | |
190 | wake_up(&lock->l_waitq); | |
0a3bdb00 | 191 | return ldlm_completion_tail(lock); |
d7e09d03 PT |
192 | } |
193 | ||
194 | LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, " | |
195 | "going forward"); | |
196 | ldlm_reprocess_all(lock->l_resource); | |
0a3bdb00 | 197 | return 0; |
d7e09d03 PT |
198 | } |
199 | EXPORT_SYMBOL(ldlm_completion_ast_async); | |
200 | ||
201 | /** | |
202 | * Generic LDLM "completion" AST. This is called in several cases: | |
203 | * | |
204 | * - when a reply to an ENQUEUE RPC is received from the server | |
205 | * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at | |
206 | * this point (determined by flags); | |
207 | * | |
208 | * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has | |
209 | * been granted; | |
210 | * | |
211 | * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock | |
212 | * gets correct lvb; | |
213 | * | |
214 | * - to force all locks when resource is destroyed (cleanup_resource()); | |
215 | * | |
216 | * - during lock conversion (not used currently). | |
217 | * | |
218 | * If lock is not granted in the first case, this function waits until second | |
219 | * or penultimate cases happen in some other thread. | |
220 | * | |
221 | */ | |
222 | int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) | |
223 | { | |
224 | /* XXX ALLOCATE - 160 bytes */ | |
225 | struct lock_wait_data lwd; | |
226 | struct obd_device *obd; | |
227 | struct obd_import *imp = NULL; | |
228 | struct l_wait_info lwi; | |
229 | __u32 timeout; | |
230 | int rc = 0; | |
d7e09d03 PT |
231 | |
232 | if (flags == LDLM_FL_WAIT_NOREPROC) { | |
233 | LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock"); | |
234 | goto noreproc; | |
235 | } | |
236 | ||
237 | if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED | | |
238 | LDLM_FL_BLOCK_CONV))) { | |
239 | wake_up(&lock->l_waitq); | |
0a3bdb00 | 240 | return 0; |
d7e09d03 PT |
241 | } |
242 | ||
243 | LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, " | |
244 | "sleeping"); | |
245 | ||
246 | noreproc: | |
247 | ||
248 | obd = class_exp2obd(lock->l_conn_export); | |
249 | ||
250 | /* if this is a local lock, then there is no import */ | |
251 | if (obd != NULL) { | |
252 | imp = obd->u.cli.cl_import; | |
253 | } | |
254 | ||
255 | /* Wait a long time for enqueue - server may have to callback a | |
256 | lock from another client. Server will evict the other client if it | |
257 | doesn't respond reasonably, and then give us the lock. */ | |
258 | timeout = ldlm_get_enq_timeout(lock) * 2; | |
259 | ||
260 | lwd.lwd_lock = lock; | |
261 | ||
262 | if (lock->l_flags & LDLM_FL_NO_TIMEOUT) { | |
263 | LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT"); | |
264 | lwi = LWI_INTR(interrupted_completion_wait, &lwd); | |
265 | } else { | |
266 | lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout), | |
267 | ldlm_expired_completion_wait, | |
268 | interrupted_completion_wait, &lwd); | |
269 | } | |
270 | ||
271 | if (imp != NULL) { | |
272 | spin_lock(&imp->imp_lock); | |
273 | lwd.lwd_conn_cnt = imp->imp_conn_cnt; | |
274 | spin_unlock(&imp->imp_lock); | |
275 | } | |
276 | ||
277 | if (ns_is_client(ldlm_lock_to_ns(lock)) && | |
278 | OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST, | |
279 | OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) { | |
280 | lock->l_flags |= LDLM_FL_FAIL_LOC; | |
281 | rc = -EINTR; | |
282 | } else { | |
283 | /* Go to sleep until the lock is granted or cancelled. */ | |
284 | rc = l_wait_event(lock->l_waitq, | |
285 | is_granted_or_cancelled(lock), &lwi); | |
286 | } | |
287 | ||
288 | if (rc) { | |
289 | LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)", | |
290 | rc); | |
0a3bdb00 | 291 | return rc; |
d7e09d03 PT |
292 | } |
293 | ||
0a3bdb00 | 294 | return ldlm_completion_tail(lock); |
d7e09d03 PT |
295 | } |
296 | EXPORT_SYMBOL(ldlm_completion_ast); | |
297 | ||
298 | /** | |
299 | * A helper to build a blocking AST function | |
300 | * | |
301 | * Perform a common operation for blocking ASTs: | |
34ca8748 | 302 | * deferred lock cancellation. |
d7e09d03 PT |
303 | * |
304 | * \param lock the lock blocking or canceling AST was called on | |
305 | * \retval 0 | |
306 | * \see mdt_blocking_ast | |
307 | * \see ldlm_blocking_ast | |
308 | */ | |
309 | int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock) | |
310 | { | |
311 | int do_ast; | |
d7e09d03 PT |
312 | |
313 | lock->l_flags |= LDLM_FL_CBPENDING; | |
314 | do_ast = (!lock->l_readers && !lock->l_writers); | |
315 | unlock_res_and_lock(lock); | |
316 | ||
317 | if (do_ast) { | |
318 | struct lustre_handle lockh; | |
319 | int rc; | |
320 | ||
321 | LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel"); | |
322 | ldlm_lock2handle(lock, &lockh); | |
323 | rc = ldlm_cli_cancel(&lockh, LCF_ASYNC); | |
324 | if (rc < 0) | |
325 | CERROR("ldlm_cli_cancel: %d\n", rc); | |
326 | } else { | |
327 | LDLM_DEBUG(lock, "Lock still has references, will be " | |
328 | "cancelled later"); | |
329 | } | |
0a3bdb00 | 330 | return 0; |
d7e09d03 PT |
331 | } |
332 | EXPORT_SYMBOL(ldlm_blocking_ast_nocheck); | |
333 | ||
334 | /** | |
335 | * Server blocking AST | |
336 | * | |
337 | * ->l_blocking_ast() callback for LDLM locks acquired by server-side | |
338 | * OBDs. | |
339 | * | |
340 | * \param lock the lock which blocks a request or cancelling lock | |
341 | * \param desc unused | |
342 | * \param data unused | |
343 | * \param flag indicates whether this cancelling or blocking callback | |
344 | * \retval 0 | |
345 | * \see ldlm_blocking_ast_nocheck | |
346 | */ | |
347 | int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, | |
348 | void *data, int flag) | |
349 | { | |
d7e09d03 PT |
350 | if (flag == LDLM_CB_CANCELING) { |
351 | /* Don't need to do anything here. */ | |
0a3bdb00 | 352 | return 0; |
d7e09d03 PT |
353 | } |
354 | ||
355 | lock_res_and_lock(lock); | |
356 | /* Get this: if ldlm_blocking_ast is racing with intent_policy, such | |
357 | * that ldlm_blocking_ast is called just before intent_policy method | |
358 | * takes the lr_lock, then by the time we get the lock, we might not | |
359 | * be the correct blocking function anymore. So check, and return | |
360 | * early, if so. */ | |
361 | if (lock->l_blocking_ast != ldlm_blocking_ast) { | |
362 | unlock_res_and_lock(lock); | |
0a3bdb00 | 363 | return 0; |
d7e09d03 | 364 | } |
0a3bdb00 | 365 | return ldlm_blocking_ast_nocheck(lock); |
d7e09d03 PT |
366 | } |
367 | EXPORT_SYMBOL(ldlm_blocking_ast); | |
368 | ||
369 | /** | |
370 | * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See | |
371 | * comment in filter_intent_policy() on why you may need this. | |
372 | */ | |
373 | int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp) | |
374 | { | |
375 | /* | |
376 | * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for | |
377 | * that is rather subtle: with OST-side locking, it may so happen that | |
378 | * _all_ extent locks are held by the OST. If client wants to obtain | |
379 | * current file size it calls ll{,u}_glimpse_size(), and (as locks are | |
380 | * on the server), dummy glimpse callback fires and does | |
381 | * nothing. Client still receives correct file size due to the | |
382 | * following fragment in filter_intent_policy(): | |
383 | * | |
384 | * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB | |
385 | * if (rc != 0 && res->lr_namespace->ns_lvbo && | |
386 | * res->lr_namespace->ns_lvbo->lvbo_update) { | |
387 | * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1); | |
388 | * } | |
389 | * | |
390 | * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and | |
391 | * returns correct file size to the client. | |
392 | */ | |
393 | return -ELDLM_NO_LOCK_DATA; | |
394 | } | |
395 | EXPORT_SYMBOL(ldlm_glimpse_ast); | |
396 | ||
397 | /** | |
398 | * Enqueue a local lock (typically on a server). | |
399 | */ | |
400 | int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, | |
401 | const struct ldlm_res_id *res_id, | |
402 | ldlm_type_t type, ldlm_policy_data_t *policy, | |
403 | ldlm_mode_t mode, __u64 *flags, | |
404 | ldlm_blocking_callback blocking, | |
405 | ldlm_completion_callback completion, | |
406 | ldlm_glimpse_callback glimpse, | |
407 | void *data, __u32 lvb_len, enum lvb_type lvb_type, | |
408 | const __u64 *client_cookie, | |
409 | struct lustre_handle *lockh) | |
410 | { | |
411 | struct ldlm_lock *lock; | |
412 | int err; | |
413 | const struct ldlm_callback_suite cbs = { .lcs_completion = completion, | |
414 | .lcs_blocking = blocking, | |
415 | .lcs_glimpse = glimpse, | |
416 | }; | |
d7e09d03 PT |
417 | |
418 | LASSERT(!(*flags & LDLM_FL_REPLAY)); | |
419 | if (unlikely(ns_is_client(ns))) { | |
420 | CERROR("Trying to enqueue local lock in a shadow namespace\n"); | |
421 | LBUG(); | |
422 | } | |
423 | ||
424 | lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len, | |
425 | lvb_type); | |
426 | if (unlikely(!lock)) | |
427 | GOTO(out_nolock, err = -ENOMEM); | |
428 | ||
429 | ldlm_lock2handle(lock, lockh); | |
430 | ||
431 | /* NB: we don't have any lock now (lock_res_and_lock) | |
432 | * because it's a new lock */ | |
433 | ldlm_lock_addref_internal_nolock(lock, mode); | |
434 | lock->l_flags |= LDLM_FL_LOCAL; | |
435 | if (*flags & LDLM_FL_ATOMIC_CB) | |
436 | lock->l_flags |= LDLM_FL_ATOMIC_CB; | |
437 | ||
438 | if (policy != NULL) | |
439 | lock->l_policy_data = *policy; | |
440 | if (client_cookie != NULL) | |
441 | lock->l_client_cookie = *client_cookie; | |
442 | if (type == LDLM_EXTENT) | |
443 | lock->l_req_extent = policy->l_extent; | |
444 | ||
445 | err = ldlm_lock_enqueue(ns, &lock, policy, flags); | |
446 | if (unlikely(err != ELDLM_OK)) | |
447 | GOTO(out, err); | |
448 | ||
449 | if (policy != NULL) | |
450 | *policy = lock->l_policy_data; | |
451 | ||
452 | if (lock->l_completion_ast) | |
453 | lock->l_completion_ast(lock, *flags, NULL); | |
454 | ||
455 | LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created"); | |
d7e09d03 PT |
456 | out: |
457 | LDLM_LOCK_RELEASE(lock); | |
458 | out_nolock: | |
459 | return err; | |
460 | } | |
461 | EXPORT_SYMBOL(ldlm_cli_enqueue_local); | |
462 | ||
463 | static void failed_lock_cleanup(struct ldlm_namespace *ns, | |
464 | struct ldlm_lock *lock, int mode) | |
465 | { | |
466 | int need_cancel = 0; | |
467 | ||
468 | /* Set a flag to prevent us from sending a CANCEL (bug 407) */ | |
469 | lock_res_and_lock(lock); | |
470 | /* Check that lock is not granted or failed, we might race. */ | |
471 | if ((lock->l_req_mode != lock->l_granted_mode) && | |
472 | !(lock->l_flags & LDLM_FL_FAILED)) { | |
473 | /* Make sure that this lock will not be found by raced | |
474 | * bl_ast and -EINVAL reply is sent to server anyways. | |
475 | * bug 17645 */ | |
476 | lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED | | |
477 | LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING; | |
478 | need_cancel = 1; | |
479 | } | |
480 | unlock_res_and_lock(lock); | |
481 | ||
482 | if (need_cancel) | |
483 | LDLM_DEBUG(lock, | |
484 | "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | " | |
485 | "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING"); | |
486 | else | |
487 | LDLM_DEBUG(lock, "lock was granted or failed in race"); | |
488 | ||
489 | ldlm_lock_decref_internal(lock, mode); | |
490 | ||
491 | /* XXX - HACK because we shouldn't call ldlm_lock_destroy() | |
492 | * from llite/file.c/ll_file_flock(). */ | |
493 | /* This code makes for the fact that we do not have blocking handler on | |
494 | * a client for flock locks. As such this is the place where we must | |
495 | * completely kill failed locks. (interrupted and those that | |
496 | * were waiting to be granted when server evicted us. */ | |
497 | if (lock->l_resource->lr_type == LDLM_FLOCK) { | |
498 | lock_res_and_lock(lock); | |
499 | ldlm_resource_unlink_lock(lock); | |
500 | ldlm_lock_destroy_nolock(lock); | |
501 | unlock_res_and_lock(lock); | |
502 | } | |
503 | } | |
504 | ||
505 | /** | |
506 | * Finishing portion of client lock enqueue code. | |
507 | * | |
508 | * Called after receiving reply from server. | |
509 | */ | |
510 | int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, | |
511 | ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode, | |
512 | __u64 *flags, void *lvb, __u32 lvb_len, | |
513 | struct lustre_handle *lockh,int rc) | |
514 | { | |
515 | struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; | |
516 | int is_replay = *flags & LDLM_FL_REPLAY; | |
517 | struct ldlm_lock *lock; | |
518 | struct ldlm_reply *reply; | |
519 | int cleanup_phase = 1; | |
520 | int size = 0; | |
d7e09d03 PT |
521 | |
522 | lock = ldlm_handle2lock(lockh); | |
523 | /* ldlm_cli_enqueue is holding a reference on this lock. */ | |
524 | if (!lock) { | |
525 | LASSERT(type == LDLM_FLOCK); | |
0a3bdb00 | 526 | return -ENOLCK; |
d7e09d03 PT |
527 | } |
528 | ||
529 | LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len), | |
530 | "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len); | |
531 | ||
532 | if (rc != ELDLM_OK) { | |
533 | LASSERT(!is_replay); | |
534 | LDLM_DEBUG(lock, "client-side enqueue END (%s)", | |
535 | rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED"); | |
536 | ||
537 | if (rc != ELDLM_LOCK_ABORTED) | |
538 | GOTO(cleanup, rc); | |
539 | } | |
540 | ||
541 | /* Before we return, swab the reply */ | |
542 | reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); | |
543 | if (reply == NULL) | |
544 | GOTO(cleanup, rc = -EPROTO); | |
545 | ||
546 | if (lvb_len != 0) { | |
547 | LASSERT(lvb != NULL); | |
548 | ||
549 | size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, | |
550 | RCL_SERVER); | |
551 | if (size < 0) { | |
552 | LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size); | |
553 | GOTO(cleanup, rc = size); | |
554 | } else if (unlikely(size > lvb_len)) { | |
555 | LDLM_ERROR(lock, "Replied LVB is larger than " | |
556 | "expectation, expected = %d, replied = %d", | |
557 | lvb_len, size); | |
558 | GOTO(cleanup, rc = -EINVAL); | |
559 | } | |
560 | } | |
561 | ||
562 | if (rc == ELDLM_LOCK_ABORTED) { | |
563 | if (lvb_len != 0) | |
564 | rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER, | |
565 | lvb, size); | |
566 | GOTO(cleanup, rc = (rc != 0 ? rc : ELDLM_LOCK_ABORTED)); | |
567 | } | |
568 | ||
569 | /* lock enqueued on the server */ | |
570 | cleanup_phase = 0; | |
571 | ||
572 | lock_res_and_lock(lock); | |
573 | /* Key change rehash lock in per-export hash with new key */ | |
574 | if (exp->exp_lock_hash) { | |
575 | /* In the function below, .hs_keycmp resolves to | |
576 | * ldlm_export_lock_keycmp() */ | |
577 | /* coverity[overrun-buffer-val] */ | |
578 | cfs_hash_rehash_key(exp->exp_lock_hash, | |
579 | &lock->l_remote_handle, | |
580 | &reply->lock_handle, | |
581 | &lock->l_exp_hash); | |
582 | } else { | |
583 | lock->l_remote_handle = reply->lock_handle; | |
584 | } | |
585 | ||
586 | *flags = ldlm_flags_from_wire(reply->lock_flags); | |
587 | lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags & | |
588 | LDLM_INHERIT_FLAGS); | |
589 | /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match() | |
590 | * to wait with no timeout as well */ | |
591 | lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags & | |
592 | LDLM_FL_NO_TIMEOUT); | |
593 | unlock_res_and_lock(lock); | |
594 | ||
55f5a824 | 595 | CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: 0x%llx\n", |
d7e09d03 PT |
596 | lock, reply->lock_handle.cookie, *flags); |
597 | ||
598 | /* If enqueue returned a blocked lock but the completion handler has | |
599 | * already run, then it fixed up the resource and we don't need to do it | |
600 | * again. */ | |
601 | if ((*flags) & LDLM_FL_LOCK_CHANGED) { | |
602 | int newmode = reply->lock_desc.l_req_mode; | |
603 | LASSERT(!is_replay); | |
604 | if (newmode && newmode != lock->l_req_mode) { | |
605 | LDLM_DEBUG(lock, "server returned different mode %s", | |
606 | ldlm_lockname[newmode]); | |
607 | lock->l_req_mode = newmode; | |
608 | } | |
609 | ||
6d95e048 AD |
610 | if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name, |
611 | &lock->l_resource->lr_name)) { | |
612 | CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES | |
613 | " instead of "DLDLMRES"\n", | |
614 | PLDLMRES(&reply->lock_desc.l_resource), | |
615 | PLDLMRES(lock->l_resource)); | |
d7e09d03 PT |
616 | |
617 | rc = ldlm_lock_change_resource(ns, lock, | |
618 | &reply->lock_desc.l_resource.lr_name); | |
619 | if (rc || lock->l_resource == NULL) | |
620 | GOTO(cleanup, rc = -ENOMEM); | |
621 | LDLM_DEBUG(lock, "client-side enqueue, new resource"); | |
622 | } | |
623 | if (with_policy) | |
624 | if (!(type == LDLM_IBITS && | |
625 | !(exp_connect_flags(exp) & OBD_CONNECT_IBITS))) | |
626 | /* We assume lock type cannot change on server*/ | |
627 | ldlm_convert_policy_to_local(exp, | |
628 | lock->l_resource->lr_type, | |
629 | &reply->lock_desc.l_policy_data, | |
630 | &lock->l_policy_data); | |
631 | if (type != LDLM_PLAIN) | |
632 | LDLM_DEBUG(lock,"client-side enqueue, new policy data"); | |
633 | } | |
634 | ||
635 | if ((*flags) & LDLM_FL_AST_SENT || | |
636 | /* Cancel extent locks as soon as possible on a liblustre client, | |
637 | * because it cannot handle asynchronous ASTs robustly (see | |
638 | * bug 7311). */ | |
639 | (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) { | |
640 | lock_res_and_lock(lock); | |
641 | lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST; | |
642 | unlock_res_and_lock(lock); | |
643 | LDLM_DEBUG(lock, "enqueue reply includes blocking AST"); | |
644 | } | |
645 | ||
646 | /* If the lock has already been granted by a completion AST, don't | |
647 | * clobber the LVB with an older one. */ | |
648 | if (lvb_len != 0) { | |
649 | /* We must lock or a racing completion might update lvb without | |
650 | * letting us know and we'll clobber the correct value. | |
651 | * Cannot unlock after the check either, a that still leaves | |
652 | * a tiny window for completion to get in */ | |
653 | lock_res_and_lock(lock); | |
654 | if (lock->l_req_mode != lock->l_granted_mode) | |
655 | rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER, | |
656 | lock->l_lvb_data, size); | |
657 | unlock_res_and_lock(lock); | |
658 | if (rc < 0) { | |
659 | cleanup_phase = 1; | |
660 | GOTO(cleanup, rc); | |
661 | } | |
662 | } | |
663 | ||
664 | if (!is_replay) { | |
665 | rc = ldlm_lock_enqueue(ns, &lock, NULL, flags); | |
666 | if (lock->l_completion_ast != NULL) { | |
667 | int err = lock->l_completion_ast(lock, *flags, NULL); | |
668 | if (!rc) | |
669 | rc = err; | |
670 | if (rc) | |
671 | cleanup_phase = 1; | |
672 | } | |
673 | } | |
674 | ||
675 | if (lvb_len && lvb != NULL) { | |
676 | /* Copy the LVB here, and not earlier, because the completion | |
677 | * AST (if any) can override what we got in the reply */ | |
678 | memcpy(lvb, lock->l_lvb_data, lvb_len); | |
679 | } | |
680 | ||
681 | LDLM_DEBUG(lock, "client-side enqueue END"); | |
d7e09d03 PT |
682 | cleanup: |
683 | if (cleanup_phase == 1 && rc) | |
684 | failed_lock_cleanup(ns, lock, mode); | |
685 | /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */ | |
686 | LDLM_LOCK_PUT(lock); | |
687 | LDLM_LOCK_RELEASE(lock); | |
688 | return rc; | |
689 | } | |
690 | EXPORT_SYMBOL(ldlm_cli_enqueue_fini); | |
691 | ||
692 | /** | |
693 | * Estimate number of lock handles that would fit into request of given | |
694 | * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into | |
695 | * a single page on the send/receive side. XXX: 512 should be changed to | |
696 | * more adequate value. | |
697 | */ | |
698 | static inline int ldlm_req_handles_avail(int req_size, int off) | |
699 | { | |
700 | int avail; | |
701 | ||
702 | avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size; | |
703 | if (likely(avail >= 0)) | |
704 | avail /= (int)sizeof(struct lustre_handle); | |
705 | else | |
706 | avail = 0; | |
707 | avail += LDLM_LOCKREQ_HANDLES - off; | |
708 | ||
709 | return avail; | |
710 | } | |
711 | ||
712 | static inline int ldlm_capsule_handles_avail(struct req_capsule *pill, | |
713 | enum req_location loc, | |
714 | int off) | |
715 | { | |
716 | int size = req_capsule_msg_size(pill, loc); | |
717 | return ldlm_req_handles_avail(size, off); | |
718 | } | |
719 | ||
720 | static inline int ldlm_format_handles_avail(struct obd_import *imp, | |
721 | const struct req_format *fmt, | |
722 | enum req_location loc, int off) | |
723 | { | |
724 | int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc); | |
725 | return ldlm_req_handles_avail(size, off); | |
726 | } | |
727 | ||
728 | /** | |
729 | * Cancel LRU locks and pack them into the enqueue request. Pack there the given | |
730 | * \a count locks in \a cancels. | |
731 | * | |
732 | * This is to be called by functions preparing their own requests that | |
733 | * might contain lists of locks to cancel in addition to actual operation | |
734 | * that needs to be performed. | |
735 | */ | |
736 | int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req, | |
737 | int version, int opc, int canceloff, | |
738 | struct list_head *cancels, int count) | |
739 | { | |
740 | struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; | |
741 | struct req_capsule *pill = &req->rq_pill; | |
742 | struct ldlm_request *dlm = NULL; | |
743 | int flags, avail, to_free, pack = 0; | |
744 | LIST_HEAD(head); | |
745 | int rc; | |
d7e09d03 PT |
746 | |
747 | if (cancels == NULL) | |
748 | cancels = &head; | |
749 | if (ns_connect_cancelset(ns)) { | |
750 | /* Estimate the amount of available space in the request. */ | |
751 | req_capsule_filled_sizes(pill, RCL_CLIENT); | |
752 | avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff); | |
753 | ||
754 | flags = ns_connect_lru_resize(ns) ? | |
755 | LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED; | |
756 | to_free = !ns_connect_lru_resize(ns) && | |
757 | opc == LDLM_ENQUEUE ? 1 : 0; | |
758 | ||
759 | /* Cancel LRU locks here _only_ if the server supports | |
760 | * EARLY_CANCEL. Otherwise we have to send extra CANCEL | |
761 | * RPC, which will make us slower. */ | |
762 | if (avail > count) | |
763 | count += ldlm_cancel_lru_local(ns, cancels, to_free, | |
764 | avail - count, 0, flags); | |
765 | if (avail > count) | |
766 | pack = count; | |
767 | else | |
768 | pack = avail; | |
769 | req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT, | |
770 | ldlm_request_bufsize(pack, opc)); | |
771 | } | |
772 | ||
773 | rc = ptlrpc_request_pack(req, version, opc); | |
774 | if (rc) { | |
775 | ldlm_lock_list_put(cancels, l_bl_ast, count); | |
0a3bdb00 | 776 | return rc; |
d7e09d03 PT |
777 | } |
778 | ||
779 | if (ns_connect_cancelset(ns)) { | |
780 | if (canceloff) { | |
781 | dlm = req_capsule_client_get(pill, &RMF_DLM_REQ); | |
782 | LASSERT(dlm); | |
783 | /* Skip first lock handler in ldlm_request_pack(), | |
6e3dd654 | 784 | * this method will increment @lock_count according |
d7e09d03 PT |
785 | * to the lock handle amount actually written to |
786 | * the buffer. */ | |
787 | dlm->lock_count = canceloff; | |
788 | } | |
789 | /* Pack into the request @pack lock handles. */ | |
790 | ldlm_cli_cancel_list(cancels, pack, req, 0); | |
791 | /* Prepare and send separate cancel RPC for others. */ | |
792 | ldlm_cli_cancel_list(cancels, count - pack, NULL, 0); | |
793 | } else { | |
794 | ldlm_lock_list_put(cancels, l_bl_ast, count); | |
795 | } | |
0a3bdb00 | 796 | return 0; |
d7e09d03 PT |
797 | } |
798 | EXPORT_SYMBOL(ldlm_prep_elc_req); | |
799 | ||
800 | int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req, | |
801 | struct list_head *cancels, int count) | |
802 | { | |
803 | return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, | |
804 | LDLM_ENQUEUE_CANCEL_OFF, cancels, count); | |
805 | } | |
806 | EXPORT_SYMBOL(ldlm_prep_enqueue_req); | |
807 | ||
808 | struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len) | |
809 | { | |
810 | struct ptlrpc_request *req; | |
811 | int rc; | |
d7e09d03 PT |
812 | |
813 | req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE); | |
814 | if (req == NULL) | |
0a3bdb00 | 815 | return ERR_PTR(-ENOMEM); |
d7e09d03 PT |
816 | |
817 | rc = ldlm_prep_enqueue_req(exp, req, NULL, 0); | |
818 | if (rc) { | |
819 | ptlrpc_request_free(req); | |
0a3bdb00 | 820 | return ERR_PTR(rc); |
d7e09d03 PT |
821 | } |
822 | ||
823 | req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len); | |
824 | ptlrpc_request_set_replen(req); | |
0a3bdb00 | 825 | return req; |
d7e09d03 PT |
826 | } |
827 | EXPORT_SYMBOL(ldlm_enqueue_pack); | |
828 | ||
829 | /** | |
830 | * Client-side lock enqueue. | |
831 | * | |
832 | * If a request has some specific initialisation it is passed in \a reqp, | |
833 | * otherwise it is created in ldlm_cli_enqueue. | |
834 | * | |
835 | * Supports sync and async requests, pass \a async flag accordingly. If a | |
836 | * request was created in ldlm_cli_enqueue and it is the async request, | |
837 | * pass it to the caller in \a reqp. | |
838 | */ | |
839 | int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, | |
840 | struct ldlm_enqueue_info *einfo, | |
841 | const struct ldlm_res_id *res_id, | |
842 | ldlm_policy_data_t const *policy, __u64 *flags, | |
843 | void *lvb, __u32 lvb_len, enum lvb_type lvb_type, | |
844 | struct lustre_handle *lockh, int async) | |
845 | { | |
846 | struct ldlm_namespace *ns; | |
847 | struct ldlm_lock *lock; | |
848 | struct ldlm_request *body; | |
849 | int is_replay = *flags & LDLM_FL_REPLAY; | |
850 | int req_passed_in = 1; | |
851 | int rc, err; | |
852 | struct ptlrpc_request *req; | |
d7e09d03 PT |
853 | |
854 | LASSERT(exp != NULL); | |
855 | ||
856 | ns = exp->exp_obd->obd_namespace; | |
857 | ||
858 | /* If we're replaying this lock, just check some invariants. | |
859 | * If we're creating a new lock, get everything all setup nice. */ | |
860 | if (is_replay) { | |
861 | lock = ldlm_handle2lock_long(lockh, 0); | |
862 | LASSERT(lock != NULL); | |
863 | LDLM_DEBUG(lock, "client-side enqueue START"); | |
864 | LASSERT(exp == lock->l_conn_export); | |
865 | } else { | |
866 | const struct ldlm_callback_suite cbs = { | |
867 | .lcs_completion = einfo->ei_cb_cp, | |
f2145eae BK |
868 | .lcs_blocking = einfo->ei_cb_bl, |
869 | .lcs_glimpse = einfo->ei_cb_gl | |
d7e09d03 PT |
870 | }; |
871 | lock = ldlm_lock_create(ns, res_id, einfo->ei_type, | |
872 | einfo->ei_mode, &cbs, einfo->ei_cbdata, | |
873 | lvb_len, lvb_type); | |
874 | if (lock == NULL) | |
0a3bdb00 | 875 | return -ENOMEM; |
d7e09d03 PT |
876 | /* for the local lock, add the reference */ |
877 | ldlm_lock_addref_internal(lock, einfo->ei_mode); | |
878 | ldlm_lock2handle(lock, lockh); | |
f261f48a | 879 | if (policy != NULL) |
d7e09d03 | 880 | lock->l_policy_data = *policy; |
d7e09d03 PT |
881 | |
882 | if (einfo->ei_type == LDLM_EXTENT) | |
883 | lock->l_req_extent = policy->l_extent; | |
884 | LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n", | |
885 | *flags); | |
886 | } | |
887 | ||
888 | lock->l_conn_export = exp; | |
889 | lock->l_export = NULL; | |
890 | lock->l_blocking_ast = einfo->ei_cb_bl; | |
d3a8a4e2 | 891 | lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL)); |
d7e09d03 PT |
892 | |
893 | /* lock not sent to server yet */ | |
894 | ||
895 | if (reqp == NULL || *reqp == NULL) { | |
896 | req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), | |
897 | &RQF_LDLM_ENQUEUE, | |
898 | LUSTRE_DLM_VERSION, | |
899 | LDLM_ENQUEUE); | |
900 | if (req == NULL) { | |
901 | failed_lock_cleanup(ns, lock, einfo->ei_mode); | |
902 | LDLM_LOCK_RELEASE(lock); | |
0a3bdb00 | 903 | return -ENOMEM; |
d7e09d03 PT |
904 | } |
905 | req_passed_in = 0; | |
906 | if (reqp) | |
907 | *reqp = req; | |
908 | } else { | |
909 | int len; | |
910 | ||
911 | req = *reqp; | |
912 | len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, | |
913 | RCL_CLIENT); | |
914 | LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n", | |
915 | DLM_LOCKREQ_OFF, len, (int)sizeof(*body)); | |
916 | } | |
917 | ||
918 | /* Dump lock data into the request buffer */ | |
919 | body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); | |
920 | ldlm_lock2desc(lock, &body->lock_desc); | |
921 | body->lock_flags = ldlm_flags_to_wire(*flags); | |
922 | body->lock_handle[0] = *lockh; | |
923 | ||
924 | /* Continue as normal. */ | |
925 | if (!req_passed_in) { | |
926 | if (lvb_len > 0) | |
927 | req_capsule_extend(&req->rq_pill, | |
928 | &RQF_LDLM_ENQUEUE_LVB); | |
929 | req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, | |
930 | lvb_len); | |
931 | ptlrpc_request_set_replen(req); | |
932 | } | |
933 | ||
934 | /* | |
935 | * Liblustre client doesn't get extent locks, except for O_APPEND case | |
936 | * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where | |
937 | * [i_size, OBD_OBJECT_EOF] lock is taken. | |
938 | */ | |
939 | LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT || | |
940 | policy->l_extent.end == OBD_OBJECT_EOF)); | |
941 | ||
942 | if (async) { | |
943 | LASSERT(reqp != NULL); | |
0a3bdb00 | 944 | return 0; |
d7e09d03 PT |
945 | } |
946 | ||
947 | LDLM_DEBUG(lock, "sending request"); | |
948 | ||
949 | rc = ptlrpc_queue_wait(req); | |
950 | ||
951 | err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0, | |
952 | einfo->ei_mode, flags, lvb, lvb_len, | |
953 | lockh, rc); | |
954 | ||
955 | /* If ldlm_cli_enqueue_fini did not find the lock, we need to free | |
956 | * one reference that we took */ | |
957 | if (err == -ENOLCK) | |
958 | LDLM_LOCK_RELEASE(lock); | |
959 | else | |
960 | rc = err; | |
961 | ||
962 | if (!req_passed_in && req != NULL) { | |
963 | ptlrpc_req_finished(req); | |
964 | if (reqp) | |
965 | *reqp = NULL; | |
966 | } | |
967 | ||
0a3bdb00 | 968 | return rc; |
d7e09d03 PT |
969 | } |
970 | EXPORT_SYMBOL(ldlm_cli_enqueue); | |
971 | ||
972 | static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode, | |
973 | __u32 *flags) | |
974 | { | |
975 | struct ldlm_resource *res; | |
976 | int rc; | |
29aaf496 | 977 | |
d7e09d03 PT |
978 | if (ns_is_client(ldlm_lock_to_ns(lock))) { |
979 | CERROR("Trying to cancel local lock\n"); | |
980 | LBUG(); | |
981 | } | |
982 | LDLM_DEBUG(lock, "client-side local convert"); | |
983 | ||
984 | res = ldlm_lock_convert(lock, new_mode, flags); | |
985 | if (res) { | |
986 | ldlm_reprocess_all(res); | |
987 | rc = 0; | |
988 | } else { | |
2d58de78 | 989 | rc = LUSTRE_EDEADLK; |
d7e09d03 PT |
990 | } |
991 | LDLM_DEBUG(lock, "client-side local convert handler END"); | |
992 | LDLM_LOCK_PUT(lock); | |
0a3bdb00 | 993 | return rc; |
d7e09d03 PT |
994 | } |
995 | ||
996 | /* FIXME: one of ldlm_cli_convert or the server side should reject attempted | |
997 | * conversion of locks which are on the waiting or converting queue */ | |
998 | /* Caller of this code is supposed to take care of lock readers/writers | |
999 | accounting */ | |
1000 | int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags) | |
1001 | { | |
1002 | struct ldlm_request *body; | |
1003 | struct ldlm_reply *reply; | |
1004 | struct ldlm_lock *lock; | |
1005 | struct ldlm_resource *res; | |
1006 | struct ptlrpc_request *req; | |
1007 | int rc; | |
d7e09d03 PT |
1008 | |
1009 | lock = ldlm_handle2lock(lockh); | |
1010 | if (!lock) { | |
1011 | LBUG(); | |
0a3bdb00 | 1012 | return -EINVAL; |
d7e09d03 PT |
1013 | } |
1014 | *flags = 0; | |
1015 | ||
1016 | if (lock->l_conn_export == NULL) | |
0a3bdb00 | 1017 | return ldlm_cli_convert_local(lock, new_mode, flags); |
d7e09d03 PT |
1018 | |
1019 | LDLM_DEBUG(lock, "client-side convert"); | |
1020 | ||
1021 | req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export), | |
1022 | &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION, | |
1023 | LDLM_CONVERT); | |
1024 | if (req == NULL) { | |
1025 | LDLM_LOCK_PUT(lock); | |
0a3bdb00 | 1026 | return -ENOMEM; |
d7e09d03 PT |
1027 | } |
1028 | ||
1029 | body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); | |
1030 | body->lock_handle[0] = lock->l_remote_handle; | |
1031 | ||
1032 | body->lock_desc.l_req_mode = new_mode; | |
1033 | body->lock_flags = ldlm_flags_to_wire(*flags); | |
1034 | ||
1035 | ||
1036 | ptlrpc_request_set_replen(req); | |
1037 | rc = ptlrpc_queue_wait(req); | |
1038 | if (rc != ELDLM_OK) | |
1039 | GOTO(out, rc); | |
1040 | ||
1041 | reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); | |
1042 | if (reply == NULL) | |
1043 | GOTO(out, rc = -EPROTO); | |
1044 | ||
1045 | if (req->rq_status) | |
1046 | GOTO(out, rc = req->rq_status); | |
1047 | ||
1048 | res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags); | |
1049 | if (res != NULL) { | |
1050 | ldlm_reprocess_all(res); | |
1051 | /* Go to sleep until the lock is granted. */ | |
1052 | /* FIXME: or cancelled. */ | |
1053 | if (lock->l_completion_ast) { | |
1054 | rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC, | |
1055 | NULL); | |
1056 | if (rc) | |
1057 | GOTO(out, rc); | |
1058 | } | |
1059 | } else { | |
2d58de78 | 1060 | rc = LUSTRE_EDEADLK; |
d7e09d03 | 1061 | } |
d7e09d03 PT |
1062 | out: |
1063 | LDLM_LOCK_PUT(lock); | |
1064 | ptlrpc_req_finished(req); | |
1065 | return rc; | |
1066 | } | |
1067 | EXPORT_SYMBOL(ldlm_cli_convert); | |
1068 | ||
1069 | /** | |
1070 | * Cancel locks locally. | |
1071 | * Returns: | |
1072 | * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server | |
1073 | * \retval LDLM_FL_CANCELING otherwise; | |
1074 | * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC. | |
1075 | */ | |
1076 | static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock) | |
1077 | { | |
1078 | __u64 rc = LDLM_FL_LOCAL_ONLY; | |
d7e09d03 PT |
1079 | |
1080 | if (lock->l_conn_export) { | |
1081 | bool local_only; | |
1082 | ||
1083 | LDLM_DEBUG(lock, "client-side cancel"); | |
1084 | /* Set this flag to prevent others from getting new references*/ | |
1085 | lock_res_and_lock(lock); | |
1086 | lock->l_flags |= LDLM_FL_CBPENDING; | |
1087 | local_only = !!(lock->l_flags & | |
1088 | (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK)); | |
1089 | ldlm_cancel_callback(lock); | |
1090 | rc = (lock->l_flags & LDLM_FL_BL_AST) ? | |
1091 | LDLM_FL_BL_AST : LDLM_FL_CANCELING; | |
1092 | unlock_res_and_lock(lock); | |
1093 | ||
1094 | if (local_only) { | |
1095 | CDEBUG(D_DLMTRACE, "not sending request (at caller's " | |
1096 | "instruction)\n"); | |
1097 | rc = LDLM_FL_LOCAL_ONLY; | |
1098 | } | |
1099 | ldlm_lock_cancel(lock); | |
1100 | } else { | |
1101 | if (ns_is_client(ldlm_lock_to_ns(lock))) { | |
1102 | LDLM_ERROR(lock, "Trying to cancel local lock"); | |
1103 | LBUG(); | |
1104 | } | |
1105 | LDLM_DEBUG(lock, "server-side local cancel"); | |
1106 | ldlm_lock_cancel(lock); | |
1107 | ldlm_reprocess_all(lock->l_resource); | |
1108 | } | |
1109 | ||
0a3bdb00 | 1110 | return rc; |
d7e09d03 PT |
1111 | } |
1112 | ||
1113 | /** | |
1114 | * Pack \a count locks in \a head into ldlm_request buffer of request \a req. | |
1115 | */ | |
1116 | static void ldlm_cancel_pack(struct ptlrpc_request *req, | |
1117 | struct list_head *head, int count) | |
1118 | { | |
1119 | struct ldlm_request *dlm; | |
1120 | struct ldlm_lock *lock; | |
1121 | int max, packed = 0; | |
d7e09d03 PT |
1122 | |
1123 | dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); | |
1124 | LASSERT(dlm != NULL); | |
1125 | ||
1126 | /* Check the room in the request buffer. */ | |
1127 | max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) - | |
1128 | sizeof(struct ldlm_request); | |
1129 | max /= sizeof(struct lustre_handle); | |
1130 | max += LDLM_LOCKREQ_HANDLES; | |
1131 | LASSERT(max >= dlm->lock_count + count); | |
1132 | ||
1133 | /* XXX: it would be better to pack lock handles grouped by resource. | |
1134 | * so that the server cancel would call filter_lvbo_update() less | |
1135 | * frequently. */ | |
1136 | list_for_each_entry(lock, head, l_bl_ast) { | |
1137 | if (!count--) | |
1138 | break; | |
1139 | LASSERT(lock->l_conn_export); | |
1140 | /* Pack the lock handle to the given request buffer. */ | |
1141 | LDLM_DEBUG(lock, "packing"); | |
1142 | dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle; | |
1143 | packed++; | |
1144 | } | |
1145 | CDEBUG(D_DLMTRACE, "%d locks packed\n", packed); | |
d7e09d03 PT |
1146 | } |
1147 | ||
1148 | /** | |
1149 | * Prepare and send a batched cancel RPC. It will include \a count lock | |
1150 | * handles of locks given in \a cancels list. */ | |
1151 | int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels, | |
1152 | int count, ldlm_cancel_flags_t flags) | |
1153 | { | |
1154 | struct ptlrpc_request *req = NULL; | |
1155 | struct obd_import *imp; | |
1156 | int free, sent = 0; | |
1157 | int rc = 0; | |
d7e09d03 PT |
1158 | |
1159 | LASSERT(exp != NULL); | |
1160 | LASSERT(count > 0); | |
1161 | ||
1162 | CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val); | |
1163 | ||
1164 | if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE)) | |
0a3bdb00 | 1165 | return count; |
d7e09d03 PT |
1166 | |
1167 | free = ldlm_format_handles_avail(class_exp2cliimp(exp), | |
1168 | &RQF_LDLM_CANCEL, RCL_CLIENT, 0); | |
1169 | if (count > free) | |
1170 | count = free; | |
1171 | ||
1172 | while (1) { | |
1173 | imp = class_exp2cliimp(exp); | |
1174 | if (imp == NULL || imp->imp_invalid) { | |
1175 | CDEBUG(D_DLMTRACE, | |
1176 | "skipping cancel on invalid import %p\n", imp); | |
0a3bdb00 | 1177 | return count; |
d7e09d03 PT |
1178 | } |
1179 | ||
1180 | req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL); | |
1181 | if (req == NULL) | |
1182 | GOTO(out, rc = -ENOMEM); | |
1183 | ||
1184 | req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT); | |
1185 | req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT, | |
1186 | ldlm_request_bufsize(count, LDLM_CANCEL)); | |
1187 | ||
1188 | rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL); | |
1189 | if (rc) { | |
1190 | ptlrpc_request_free(req); | |
1191 | GOTO(out, rc); | |
1192 | } | |
1193 | ||
1194 | req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL; | |
1195 | req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL; | |
1196 | ptlrpc_at_set_req_timeout(req); | |
1197 | ||
1198 | ldlm_cancel_pack(req, cancels, count); | |
1199 | ||
1200 | ptlrpc_request_set_replen(req); | |
1201 | if (flags & LCF_ASYNC) { | |
1202 | ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1); | |
1203 | sent = count; | |
1204 | GOTO(out, 0); | |
1205 | } else { | |
1206 | rc = ptlrpc_queue_wait(req); | |
1207 | } | |
2d58de78 | 1208 | if (rc == LUSTRE_ESTALE) { |
d7e09d03 PT |
1209 | CDEBUG(D_DLMTRACE, "client/server (nid %s) " |
1210 | "out of sync -- not fatal\n", | |
1211 | libcfs_nid2str(req->rq_import-> | |
1212 | imp_connection->c_peer.nid)); | |
1213 | rc = 0; | |
1214 | } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/ | |
1215 | req->rq_import_generation == imp->imp_generation) { | |
1216 | ptlrpc_req_finished(req); | |
1217 | continue; | |
1218 | } else if (rc != ELDLM_OK) { | |
1219 | /* -ESHUTDOWN is common on umount */ | |
1220 | CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR, | |
1221 | "Got rc %d from cancel RPC: " | |
1222 | "canceling anyway\n", rc); | |
1223 | break; | |
1224 | } | |
1225 | sent = count; | |
1226 | break; | |
1227 | } | |
1228 | ||
1229 | ptlrpc_req_finished(req); | |
d7e09d03 PT |
1230 | out: |
1231 | return sent ? sent : rc; | |
1232 | } | |
1233 | EXPORT_SYMBOL(ldlm_cli_cancel_req); | |
1234 | ||
1235 | static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp) | |
1236 | { | |
1237 | LASSERT(imp != NULL); | |
1238 | return &imp->imp_obd->obd_namespace->ns_pool; | |
1239 | } | |
1240 | ||
1241 | /** | |
1242 | * Update client's OBD pool related fields with new SLV and Limit from \a req. | |
1243 | */ | |
1244 | int ldlm_cli_update_pool(struct ptlrpc_request *req) | |
1245 | { | |
1246 | struct obd_device *obd; | |
1247 | __u64 new_slv; | |
1248 | __u32 new_limit; | |
29aaf496 | 1249 | |
d7e09d03 PT |
1250 | if (unlikely(!req->rq_import || !req->rq_import->imp_obd || |
1251 | !imp_connect_lru_resize(req->rq_import))) | |
1252 | { | |
1253 | /* | |
1254 | * Do nothing for corner cases. | |
1255 | */ | |
0a3bdb00 | 1256 | return 0; |
d7e09d03 PT |
1257 | } |
1258 | ||
1259 | /* In some cases RPC may contain SLV and limit zeroed out. This | |
1260 | * is the case when server does not support LRU resize feature. | |
1261 | * This is also possible in some recovery cases when server-side | |
1262 | * reqs have no reference to the OBD export and thus access to | |
1263 | * server-side namespace is not possible. */ | |
1264 | if (lustre_msg_get_slv(req->rq_repmsg) == 0 || | |
1265 | lustre_msg_get_limit(req->rq_repmsg) == 0) { | |
b0f5aad5 | 1266 | DEBUG_REQ(D_HA, req, "Zero SLV or Limit found (SLV: %llu, Limit: %u)", |
d7e09d03 PT |
1267 | lustre_msg_get_slv(req->rq_repmsg), |
1268 | lustre_msg_get_limit(req->rq_repmsg)); | |
0a3bdb00 | 1269 | return 0; |
d7e09d03 PT |
1270 | } |
1271 | ||
1272 | new_limit = lustre_msg_get_limit(req->rq_repmsg); | |
1273 | new_slv = lustre_msg_get_slv(req->rq_repmsg); | |
1274 | obd = req->rq_import->imp_obd; | |
1275 | ||
1276 | /* Set new SLV and limit in OBD fields to make them accessible | |
1277 | * to the pool thread. We do not access obd_namespace and pool | |
1278 | * directly here as there is no reliable way to make sure that | |
1279 | * they are still alive at cleanup time. Evil races are possible | |
1280 | * which may cause Oops at that time. */ | |
1281 | write_lock(&obd->obd_pool_lock); | |
1282 | obd->obd_pool_slv = new_slv; | |
1283 | obd->obd_pool_limit = new_limit; | |
1284 | write_unlock(&obd->obd_pool_lock); | |
1285 | ||
0a3bdb00 | 1286 | return 0; |
d7e09d03 PT |
1287 | } |
1288 | EXPORT_SYMBOL(ldlm_cli_update_pool); | |
1289 | ||
1290 | /** | |
1291 | * Client side lock cancel. | |
1292 | * | |
1293 | * Lock must not have any readers or writers by this time. | |
1294 | */ | |
1295 | int ldlm_cli_cancel(struct lustre_handle *lockh, | |
1296 | ldlm_cancel_flags_t cancel_flags) | |
1297 | { | |
1298 | struct obd_export *exp; | |
1299 | int avail, flags, count = 1; | |
1300 | __u64 rc = 0; | |
1301 | struct ldlm_namespace *ns; | |
1302 | struct ldlm_lock *lock; | |
1303 | LIST_HEAD(cancels); | |
d7e09d03 PT |
1304 | |
1305 | /* concurrent cancels on the same handle can happen */ | |
1306 | lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING); | |
1307 | if (lock == NULL) { | |
1308 | LDLM_DEBUG_NOLOCK("lock is already being destroyed\n"); | |
0a3bdb00 | 1309 | return 0; |
d7e09d03 PT |
1310 | } |
1311 | ||
1312 | rc = ldlm_cli_cancel_local(lock); | |
d3a8a4e2 | 1313 | if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) { |
d7e09d03 | 1314 | LDLM_LOCK_RELEASE(lock); |
0a3bdb00 | 1315 | return 0; |
d7e09d03 PT |
1316 | } |
1317 | /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL | |
1318 | * RPC which goes to canceld portal, so we can cancel other LRU locks | |
1319 | * here and send them all as one LDLM_CANCEL RPC. */ | |
1320 | LASSERT(list_empty(&lock->l_bl_ast)); | |
1321 | list_add(&lock->l_bl_ast, &cancels); | |
1322 | ||
1323 | exp = lock->l_conn_export; | |
1324 | if (exp_connect_cancelset(exp)) { | |
1325 | avail = ldlm_format_handles_avail(class_exp2cliimp(exp), | |
1326 | &RQF_LDLM_CANCEL, | |
1327 | RCL_CLIENT, 0); | |
1328 | LASSERT(avail > 0); | |
1329 | ||
1330 | ns = ldlm_lock_to_ns(lock); | |
1331 | flags = ns_connect_lru_resize(ns) ? | |
1332 | LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED; | |
1333 | count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1, | |
1334 | LCF_BL_AST, flags); | |
1335 | } | |
1336 | ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags); | |
0a3bdb00 | 1337 | return 0; |
d7e09d03 PT |
1338 | } |
1339 | EXPORT_SYMBOL(ldlm_cli_cancel); | |
1340 | ||
1341 | /** | |
1342 | * Locally cancel up to \a count locks in list \a cancels. | |
1343 | * Return the number of cancelled locks. | |
1344 | */ | |
1345 | int ldlm_cli_cancel_list_local(struct list_head *cancels, int count, | |
1346 | ldlm_cancel_flags_t flags) | |
1347 | { | |
1348 | LIST_HEAD(head); | |
1349 | struct ldlm_lock *lock, *next; | |
1350 | int left = 0, bl_ast = 0; | |
1351 | __u64 rc; | |
1352 | ||
1353 | left = count; | |
1354 | list_for_each_entry_safe(lock, next, cancels, l_bl_ast) { | |
1355 | if (left-- == 0) | |
1356 | break; | |
1357 | ||
1358 | if (flags & LCF_LOCAL) { | |
1359 | rc = LDLM_FL_LOCAL_ONLY; | |
1360 | ldlm_lock_cancel(lock); | |
1361 | } else { | |
1362 | rc = ldlm_cli_cancel_local(lock); | |
1363 | } | |
1364 | /* Until we have compound requests and can send LDLM_CANCEL | |
1365 | * requests batched with generic RPCs, we need to send cancels | |
1366 | * with the LDLM_FL_BL_AST flag in a separate RPC from | |
1367 | * the one being generated now. */ | |
1368 | if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) { | |
1369 | LDLM_DEBUG(lock, "Cancel lock separately"); | |
1370 | list_del_init(&lock->l_bl_ast); | |
1371 | list_add(&lock->l_bl_ast, &head); | |
1372 | bl_ast++; | |
1373 | continue; | |
1374 | } | |
1375 | if (rc == LDLM_FL_LOCAL_ONLY) { | |
1376 | /* CANCEL RPC should not be sent to server. */ | |
1377 | list_del_init(&lock->l_bl_ast); | |
1378 | LDLM_LOCK_RELEASE(lock); | |
1379 | count--; | |
1380 | } | |
1381 | } | |
1382 | if (bl_ast > 0) { | |
1383 | count -= bl_ast; | |
1384 | ldlm_cli_cancel_list(&head, bl_ast, NULL, 0); | |
1385 | } | |
1386 | ||
0a3bdb00 | 1387 | return count; |
d7e09d03 PT |
1388 | } |
1389 | EXPORT_SYMBOL(ldlm_cli_cancel_list_local); | |
1390 | ||
1391 | /** | |
1392 | * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back | |
1393 | * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g. | |
1394 | * readahead requests, ...) | |
1395 | */ | |
1396 | static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, | |
1397 | struct ldlm_lock *lock, | |
1398 | int unused, int added, | |
1399 | int count) | |
1400 | { | |
1401 | ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK; | |
1402 | ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery; | |
1403 | lock_res_and_lock(lock); | |
1404 | ||
1405 | /* don't check added & count since we want to process all locks | |
1406 | * from unused list */ | |
1407 | switch (lock->l_resource->lr_type) { | |
1408 | case LDLM_EXTENT: | |
1409 | case LDLM_IBITS: | |
1410 | if (cb && cb(lock)) | |
1411 | break; | |
1412 | default: | |
1413 | result = LDLM_POLICY_SKIP_LOCK; | |
1414 | lock->l_flags |= LDLM_FL_SKIPPED; | |
1415 | break; | |
1416 | } | |
1417 | ||
1418 | unlock_res_and_lock(lock); | |
0a3bdb00 | 1419 | return result; |
d7e09d03 PT |
1420 | } |
1421 | ||
1422 | /** | |
1423 | * Callback function for LRU-resize policy. Decides whether to keep | |
1424 | * \a lock in LRU for current \a LRU size \a unused, added in current | |
1425 | * scan \a added and number of locks to be preferably canceled \a count. | |
1426 | * | |
1427 | * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning | |
1428 | * | |
1429 | * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU | |
1430 | */ | |
1431 | static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns, | |
1432 | struct ldlm_lock *lock, | |
1433 | int unused, int added, | |
1434 | int count) | |
1435 | { | |
a649ad1d | 1436 | unsigned long cur = cfs_time_current(); |
d7e09d03 PT |
1437 | struct ldlm_pool *pl = &ns->ns_pool; |
1438 | __u64 slv, lvf, lv; | |
a649ad1d | 1439 | unsigned long la; |
d7e09d03 PT |
1440 | |
1441 | /* Stop LRU processing when we reach past @count or have checked all | |
1442 | * locks in LRU. */ | |
1443 | if (count && added >= count) | |
1444 | return LDLM_POLICY_KEEP_LOCK; | |
1445 | ||
1446 | slv = ldlm_pool_get_slv(pl); | |
1447 | lvf = ldlm_pool_get_lvf(pl); | |
1448 | la = cfs_duration_sec(cfs_time_sub(cur, | |
1449 | lock->l_last_used)); | |
1450 | lv = lvf * la * unused; | |
1451 | ||
1452 | /* Inform pool about current CLV to see it via proc. */ | |
1453 | ldlm_pool_set_clv(pl, lv); | |
1454 | ||
1455 | /* Stop when SLV is not yet come from server or lv is smaller than | |
1456 | * it is. */ | |
1457 | return (slv == 0 || lv < slv) ? | |
1458 | LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; | |
1459 | } | |
1460 | ||
1461 | /** | |
1462 | * Callback function for proc used policy. Makes decision whether to keep | |
1463 | * \a lock in LRU for current \a LRU size \a unused, added in current scan \a | |
1464 | * added and number of locks to be preferably canceled \a count. | |
1465 | * | |
1466 | * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning | |
1467 | * | |
1468 | * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU | |
1469 | */ | |
1470 | static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns, | |
1471 | struct ldlm_lock *lock, | |
1472 | int unused, int added, | |
1473 | int count) | |
1474 | { | |
1475 | /* Stop LRU processing when we reach past @count or have checked all | |
1476 | * locks in LRU. */ | |
1477 | return (added >= count) ? | |
1478 | LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; | |
1479 | } | |
1480 | ||
1481 | /** | |
1482 | * Callback function for aged policy. Makes decision whether to keep \a lock in | |
1483 | * LRU for current LRU size \a unused, added in current scan \a added and | |
1484 | * number of locks to be preferably canceled \a count. | |
1485 | * | |
1486 | * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning | |
1487 | * | |
1488 | * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU | |
1489 | */ | |
1490 | static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns, | |
1491 | struct ldlm_lock *lock, | |
1492 | int unused, int added, | |
1493 | int count) | |
1494 | { | |
1495 | /* Stop LRU processing if young lock is found and we reach past count */ | |
1496 | return ((added >= count) && | |
699503bc GKH |
1497 | time_before(cfs_time_current(), |
1498 | cfs_time_add(lock->l_last_used, ns->ns_max_age))) ? | |
d7e09d03 PT |
1499 | LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; |
1500 | } | |
1501 | ||
1502 | /** | |
1503 | * Callback function for default policy. Makes decision whether to keep \a lock | |
1504 | * in LRU for current LRU size \a unused, added in current scan \a added and | |
1505 | * number of locks to be preferably canceled \a count. | |
1506 | * | |
1507 | * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning | |
1508 | * | |
1509 | * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU | |
1510 | */ | |
1511 | static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns, | |
1512 | struct ldlm_lock *lock, | |
1513 | int unused, int added, | |
1514 | int count) | |
1515 | { | |
1516 | /* Stop LRU processing when we reach past count or have checked all | |
1517 | * locks in LRU. */ | |
1518 | return (added >= count) ? | |
1519 | LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; | |
1520 | } | |
1521 | ||
1522 | typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *, | |
1523 | struct ldlm_lock *, int, | |
1524 | int, int); | |
1525 | ||
1526 | static ldlm_cancel_lru_policy_t | |
1527 | ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags) | |
1528 | { | |
1529 | if (flags & LDLM_CANCEL_NO_WAIT) | |
1530 | return ldlm_cancel_no_wait_policy; | |
1531 | ||
1532 | if (ns_connect_lru_resize(ns)) { | |
1533 | if (flags & LDLM_CANCEL_SHRINK) | |
1534 | /* We kill passed number of old locks. */ | |
1535 | return ldlm_cancel_passed_policy; | |
1536 | else if (flags & LDLM_CANCEL_LRUR) | |
1537 | return ldlm_cancel_lrur_policy; | |
1538 | else if (flags & LDLM_CANCEL_PASSED) | |
1539 | return ldlm_cancel_passed_policy; | |
1540 | } else { | |
1541 | if (flags & LDLM_CANCEL_AGED) | |
1542 | return ldlm_cancel_aged_policy; | |
1543 | } | |
1544 | ||
1545 | return ldlm_cancel_default_policy; | |
1546 | } | |
1547 | ||
1548 | /** | |
1549 | * - Free space in LRU for \a count new locks, | |
1550 | * redundant unused locks are canceled locally; | |
1551 | * - also cancel locally unused aged locks; | |
1552 | * - do not cancel more than \a max locks; | |
1553 | * - GET the found locks and add them into the \a cancels list. | |
1554 | * | |
1555 | * A client lock can be added to the l_bl_ast list only when it is | |
1556 | * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing | |
1557 | * CANCEL. There are the following use cases: | |
1558 | * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and | |
1559 | * ldlm_cli_cancel(), which check and set this flag properly. As any | |
1560 | * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed | |
1561 | * later without any special locking. | |
1562 | * | |
1563 | * Calling policies for enabled LRU resize: | |
1564 | * ---------------------------------------- | |
1565 | * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to | |
1566 | * cancel not more than \a count locks; | |
1567 | * | |
1568 | * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at | |
1569 | * the beginning of LRU list); | |
1570 | * | |
1571 | * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to | |
6e3dd654 | 1572 | * memory pressure policy function; |
d7e09d03 PT |
1573 | * |
1574 | * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy". | |
1575 | * | |
1576 | * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible | |
1577 | * (typically before replaying locks) w/o | |
1578 | * sending any RPCs or waiting for any | |
1579 | * outstanding RPC to complete. | |
1580 | */ | |
1581 | static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, struct list_head *cancels, | |
1582 | int count, int max, int flags) | |
1583 | { | |
1584 | ldlm_cancel_lru_policy_t pf; | |
1585 | struct ldlm_lock *lock, *next; | |
1586 | int added = 0, unused, remained; | |
d7e09d03 PT |
1587 | |
1588 | spin_lock(&ns->ns_lock); | |
1589 | unused = ns->ns_nr_unused; | |
1590 | remained = unused; | |
1591 | ||
1592 | if (!ns_connect_lru_resize(ns)) | |
1593 | count += unused - ns->ns_max_unused; | |
1594 | ||
1595 | pf = ldlm_cancel_lru_policy(ns, flags); | |
1596 | LASSERT(pf != NULL); | |
1597 | ||
1598 | while (!list_empty(&ns->ns_unused_list)) { | |
1599 | ldlm_policy_res_t result; | |
1600 | ||
1601 | /* all unused locks */ | |
1602 | if (remained-- <= 0) | |
1603 | break; | |
1604 | ||
1605 | /* For any flags, stop scanning if @max is reached. */ | |
1606 | if (max && added >= max) | |
1607 | break; | |
1608 | ||
1609 | list_for_each_entry_safe(lock, next, &ns->ns_unused_list, | |
1610 | l_lru) { | |
1611 | /* No locks which got blocking requests. */ | |
1612 | LASSERT(!(lock->l_flags & LDLM_FL_BL_AST)); | |
1613 | ||
1614 | if (flags & LDLM_CANCEL_NO_WAIT && | |
1615 | lock->l_flags & LDLM_FL_SKIPPED) | |
1616 | /* already processed */ | |
1617 | continue; | |
1618 | ||
1619 | /* Somebody is already doing CANCEL. No need for this | |
1620 | * lock in LRU, do not traverse it again. */ | |
1621 | if (!(lock->l_flags & LDLM_FL_CANCELING)) | |
1622 | break; | |
1623 | ||
1624 | ldlm_lock_remove_from_lru_nolock(lock); | |
1625 | } | |
1626 | if (&lock->l_lru == &ns->ns_unused_list) | |
1627 | break; | |
1628 | ||
1629 | LDLM_LOCK_GET(lock); | |
1630 | spin_unlock(&ns->ns_lock); | |
6756bb7c | 1631 | lu_ref_add(&lock->l_reference, __func__, current); |
d7e09d03 PT |
1632 | |
1633 | /* Pass the lock through the policy filter and see if it | |
1634 | * should stay in LRU. | |
1635 | * | |
1636 | * Even for shrinker policy we stop scanning if | |
1637 | * we find a lock that should stay in the cache. | |
1638 | * We should take into account lock age anyway | |
1639 | * as a new lock is a valuable resource even if | |
1640 | * it has a low weight. | |
1641 | * | |
1642 | * That is, for shrinker policy we drop only | |
1643 | * old locks, but additionally choose them by | |
1644 | * their weight. Big extent locks will stay in | |
1645 | * the cache. */ | |
1646 | result = pf(ns, lock, unused, added, count); | |
1647 | if (result == LDLM_POLICY_KEEP_LOCK) { | |
1648 | lu_ref_del(&lock->l_reference, | |
6756bb7c | 1649 | __func__, current); |
d7e09d03 PT |
1650 | LDLM_LOCK_RELEASE(lock); |
1651 | spin_lock(&ns->ns_lock); | |
1652 | break; | |
1653 | } | |
1654 | if (result == LDLM_POLICY_SKIP_LOCK) { | |
1655 | lu_ref_del(&lock->l_reference, | |
1656 | __func__, current); | |
1657 | LDLM_LOCK_RELEASE(lock); | |
1658 | spin_lock(&ns->ns_lock); | |
1659 | continue; | |
1660 | } | |
1661 | ||
1662 | lock_res_and_lock(lock); | |
1663 | /* Check flags again under the lock. */ | |
1664 | if ((lock->l_flags & LDLM_FL_CANCELING) || | |
1665 | (ldlm_lock_remove_from_lru(lock) == 0)) { | |
1666 | /* Another thread is removing lock from LRU, or | |
1667 | * somebody is already doing CANCEL, or there | |
1668 | * is a blocking request which will send cancel | |
1669 | * by itself, or the lock is no longer unused. */ | |
1670 | unlock_res_and_lock(lock); | |
1671 | lu_ref_del(&lock->l_reference, | |
6756bb7c | 1672 | __func__, current); |
d7e09d03 PT |
1673 | LDLM_LOCK_RELEASE(lock); |
1674 | spin_lock(&ns->ns_lock); | |
1675 | continue; | |
1676 | } | |
1677 | LASSERT(!lock->l_readers && !lock->l_writers); | |
1678 | ||
1679 | /* If we have chosen to cancel this lock voluntarily, we | |
1680 | * better send cancel notification to server, so that it | |
1681 | * frees appropriate state. This might lead to a race | |
1682 | * where while we are doing cancel here, server is also | |
1683 | * silently cancelling this lock. */ | |
1684 | lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK; | |
1685 | ||
1686 | /* Setting the CBPENDING flag is a little misleading, | |
1687 | * but prevents an important race; namely, once | |
1688 | * CBPENDING is set, the lock can accumulate no more | |
1689 | * readers/writers. Since readers and writers are | |
1690 | * already zero here, ldlm_lock_decref() won't see | |
1691 | * this flag and call l_blocking_ast */ | |
1692 | lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING; | |
1693 | ||
1694 | /* We can't re-add to l_lru as it confuses the | |
1695 | * refcounting in ldlm_lock_remove_from_lru() if an AST | |
1696 | * arrives after we drop lr_lock below. We use l_bl_ast | |
1697 | * and can't use l_pending_chain as it is used both on | |
1698 | * server and client nevertheless bug 5666 says it is | |
1699 | * used only on server */ | |
1700 | LASSERT(list_empty(&lock->l_bl_ast)); | |
1701 | list_add(&lock->l_bl_ast, cancels); | |
1702 | unlock_res_and_lock(lock); | |
6756bb7c | 1703 | lu_ref_del(&lock->l_reference, __func__, current); |
d7e09d03 PT |
1704 | spin_lock(&ns->ns_lock); |
1705 | added++; | |
1706 | unused--; | |
1707 | } | |
1708 | spin_unlock(&ns->ns_lock); | |
0a3bdb00 | 1709 | return added; |
d7e09d03 PT |
1710 | } |
1711 | ||
1712 | int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, | |
1713 | int count, int max, ldlm_cancel_flags_t cancel_flags, | |
1714 | int flags) | |
1715 | { | |
1716 | int added; | |
1717 | added = ldlm_prepare_lru_list(ns, cancels, count, max, flags); | |
1718 | if (added <= 0) | |
1719 | return added; | |
1720 | return ldlm_cli_cancel_list_local(cancels, added, cancel_flags); | |
1721 | } | |
1722 | ||
1723 | /** | |
1724 | * Cancel at least \a nr locks from given namespace LRU. | |
1725 | * | |
1726 | * When called with LCF_ASYNC the blocking callback will be handled | |
1727 | * in a thread and this function will return after the thread has been | |
1728 | * asked to call the callback. When called with LCF_ASYNC the blocking | |
1729 | * callback will be performed in this function. | |
1730 | */ | |
1731 | int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, | |
1732 | ldlm_cancel_flags_t cancel_flags, | |
1733 | int flags) | |
1734 | { | |
1735 | LIST_HEAD(cancels); | |
1736 | int count, rc; | |
d7e09d03 PT |
1737 | |
1738 | /* Just prepare the list of locks, do not actually cancel them yet. | |
1739 | * Locks are cancelled later in a separate thread. */ | |
1740 | count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags); | |
1741 | rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags); | |
1742 | if (rc == 0) | |
0a3bdb00 | 1743 | return count; |
d7e09d03 | 1744 | |
0a3bdb00 | 1745 | return 0; |
d7e09d03 PT |
1746 | } |
1747 | ||
1748 | /** | |
1749 | * Find and cancel locally unused locks found on resource, matched to the | |
1750 | * given policy, mode. GET the found locks and add them into the \a cancels | |
1751 | * list. | |
1752 | */ | |
1753 | int ldlm_cancel_resource_local(struct ldlm_resource *res, | |
1754 | struct list_head *cancels, | |
1755 | ldlm_policy_data_t *policy, | |
875332d4 | 1756 | ldlm_mode_t mode, __u64 lock_flags, |
d7e09d03 PT |
1757 | ldlm_cancel_flags_t cancel_flags, void *opaque) |
1758 | { | |
1759 | struct ldlm_lock *lock; | |
1760 | int count = 0; | |
d7e09d03 PT |
1761 | |
1762 | lock_res(res); | |
1763 | list_for_each_entry(lock, &res->lr_granted, l_res_link) { | |
1764 | if (opaque != NULL && lock->l_ast_data != opaque) { | |
1765 | LDLM_ERROR(lock, "data %p doesn't match opaque %p", | |
1766 | lock->l_ast_data, opaque); | |
1767 | //LBUG(); | |
1768 | continue; | |
1769 | } | |
1770 | ||
1771 | if (lock->l_readers || lock->l_writers) | |
1772 | continue; | |
1773 | ||
1774 | /* If somebody is already doing CANCEL, or blocking AST came, | |
1775 | * skip this lock. */ | |
1776 | if (lock->l_flags & LDLM_FL_BL_AST || | |
1777 | lock->l_flags & LDLM_FL_CANCELING) | |
1778 | continue; | |
1779 | ||
1780 | if (lockmode_compat(lock->l_granted_mode, mode)) | |
1781 | continue; | |
1782 | ||
1783 | /* If policy is given and this is IBITS lock, add to list only | |
1784 | * those locks that match by policy. */ | |
1785 | if (policy && (lock->l_resource->lr_type == LDLM_IBITS) && | |
1786 | !(lock->l_policy_data.l_inodebits.bits & | |
1787 | policy->l_inodebits.bits)) | |
1788 | continue; | |
1789 | ||
1790 | /* See CBPENDING comment in ldlm_cancel_lru */ | |
1791 | lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING | | |
1792 | lock_flags; | |
1793 | ||
1794 | LASSERT(list_empty(&lock->l_bl_ast)); | |
1795 | list_add(&lock->l_bl_ast, cancels); | |
1796 | LDLM_LOCK_GET(lock); | |
1797 | count++; | |
1798 | } | |
1799 | unlock_res(res); | |
1800 | ||
0a3bdb00 | 1801 | return ldlm_cli_cancel_list_local(cancels, count, cancel_flags); |
d7e09d03 PT |
1802 | } |
1803 | EXPORT_SYMBOL(ldlm_cancel_resource_local); | |
1804 | ||
1805 | /** | |
1806 | * Cancel client-side locks from a list and send/prepare cancel RPCs to the | |
1807 | * server. | |
1808 | * If \a req is NULL, send CANCEL request to server with handles of locks | |
1809 | * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests | |
1810 | * separately per lock. | |
1811 | * If \a req is not NULL, put handles of locks in \a cancels into the request | |
1812 | * buffer at the offset \a off. | |
1813 | * Destroy \a cancels at the end. | |
1814 | */ | |
1815 | int ldlm_cli_cancel_list(struct list_head *cancels, int count, | |
1816 | struct ptlrpc_request *req, ldlm_cancel_flags_t flags) | |
1817 | { | |
1818 | struct ldlm_lock *lock; | |
1819 | int res = 0; | |
d7e09d03 PT |
1820 | |
1821 | if (list_empty(cancels) || count == 0) | |
0a3bdb00 | 1822 | return 0; |
d7e09d03 PT |
1823 | |
1824 | /* XXX: requests (both batched and not) could be sent in parallel. | |
1825 | * Usually it is enough to have just 1 RPC, but it is possible that | |
1826 | * there are too many locks to be cancelled in LRU or on a resource. | |
1827 | * It would also speed up the case when the server does not support | |
1828 | * the feature. */ | |
1829 | while (count > 0) { | |
1830 | LASSERT(!list_empty(cancels)); | |
1831 | lock = list_entry(cancels->next, struct ldlm_lock, | |
1832 | l_bl_ast); | |
1833 | LASSERT(lock->l_conn_export); | |
1834 | ||
1835 | if (exp_connect_cancelset(lock->l_conn_export)) { | |
1836 | res = count; | |
1837 | if (req) | |
1838 | ldlm_cancel_pack(req, cancels, count); | |
1839 | else | |
1840 | res = ldlm_cli_cancel_req(lock->l_conn_export, | |
1841 | cancels, count, | |
1842 | flags); | |
1843 | } else { | |
1844 | res = ldlm_cli_cancel_req(lock->l_conn_export, | |
1845 | cancels, 1, flags); | |
1846 | } | |
1847 | ||
1848 | if (res < 0) { | |
1849 | CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR, | |
1850 | "ldlm_cli_cancel_list: %d\n", res); | |
1851 | res = count; | |
1852 | } | |
1853 | ||
1854 | count -= res; | |
1855 | ldlm_lock_list_put(cancels, l_bl_ast, res); | |
1856 | } | |
1857 | LASSERT(count == 0); | |
0a3bdb00 | 1858 | return 0; |
d7e09d03 PT |
1859 | } |
1860 | EXPORT_SYMBOL(ldlm_cli_cancel_list); | |
1861 | ||
1862 | /** | |
1863 | * Cancel all locks on a resource that have 0 readers/writers. | |
1864 | * | |
1865 | * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying | |
1866 | * to notify the server. */ | |
1867 | int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns, | |
1868 | const struct ldlm_res_id *res_id, | |
1869 | ldlm_policy_data_t *policy, | |
1870 | ldlm_mode_t mode, | |
1871 | ldlm_cancel_flags_t flags, | |
1872 | void *opaque) | |
1873 | { | |
1874 | struct ldlm_resource *res; | |
1875 | LIST_HEAD(cancels); | |
1876 | int count; | |
1877 | int rc; | |
d7e09d03 PT |
1878 | |
1879 | res = ldlm_resource_get(ns, NULL, res_id, 0, 0); | |
1880 | if (res == NULL) { | |
1881 | /* This is not a problem. */ | |
b0f5aad5 | 1882 | CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]); |
0a3bdb00 | 1883 | return 0; |
d7e09d03 PT |
1884 | } |
1885 | ||
1886 | LDLM_RESOURCE_ADDREF(res); | |
1887 | count = ldlm_cancel_resource_local(res, &cancels, policy, mode, | |
1888 | 0, flags | LCF_BL_AST, opaque); | |
1889 | rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags); | |
1890 | if (rc != ELDLM_OK) | |
6d95e048 AD |
1891 | CERROR("canceling unused lock "DLDLMRES": rc = %d\n", |
1892 | PLDLMRES(res), rc); | |
d7e09d03 PT |
1893 | |
1894 | LDLM_RESOURCE_DELREF(res); | |
1895 | ldlm_resource_putref(res); | |
0a3bdb00 | 1896 | return 0; |
d7e09d03 PT |
1897 | } |
1898 | EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource); | |
1899 | ||
1900 | struct ldlm_cli_cancel_arg { | |
1901 | int lc_flags; | |
1902 | void *lc_opaque; | |
1903 | }; | |
1904 | ||
6da6eabe | 1905 | static int ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd, |
d7e09d03 PT |
1906 | struct hlist_node *hnode, void *arg) |
1907 | { | |
1908 | struct ldlm_resource *res = cfs_hash_object(hs, hnode); | |
1909 | struct ldlm_cli_cancel_arg *lc = arg; | |
6d95e048 AD |
1910 | |
1911 | ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name, | |
1912 | NULL, LCK_MINMODE, | |
1913 | lc->lc_flags, lc->lc_opaque); | |
d7e09d03 PT |
1914 | /* must return 0 for hash iteration */ |
1915 | return 0; | |
1916 | } | |
1917 | ||
1918 | /** | |
1919 | * Cancel all locks on a namespace (or a specific resource, if given) | |
1920 | * that have 0 readers/writers. | |
1921 | * | |
1922 | * If flags & LCF_LOCAL, throw the locks away without trying | |
1923 | * to notify the server. */ | |
1924 | int ldlm_cli_cancel_unused(struct ldlm_namespace *ns, | |
1925 | const struct ldlm_res_id *res_id, | |
1926 | ldlm_cancel_flags_t flags, void *opaque) | |
1927 | { | |
1928 | struct ldlm_cli_cancel_arg arg = { | |
1929 | .lc_flags = flags, | |
1930 | .lc_opaque = opaque, | |
1931 | }; | |
1932 | ||
d7e09d03 | 1933 | if (ns == NULL) |
0a3bdb00 | 1934 | return ELDLM_OK; |
d7e09d03 PT |
1935 | |
1936 | if (res_id != NULL) { | |
0a3bdb00 | 1937 | return ldlm_cli_cancel_unused_resource(ns, res_id, NULL, |
d7e09d03 | 1938 | LCK_MINMODE, flags, |
0a3bdb00 | 1939 | opaque); |
d7e09d03 PT |
1940 | } else { |
1941 | cfs_hash_for_each_nolock(ns->ns_rs_hash, | |
1942 | ldlm_cli_hash_cancel_unused, &arg); | |
0a3bdb00 | 1943 | return ELDLM_OK; |
d7e09d03 PT |
1944 | } |
1945 | } | |
1946 | EXPORT_SYMBOL(ldlm_cli_cancel_unused); | |
1947 | ||
1948 | /* Lock iterators. */ | |
1949 | ||
1950 | int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter, | |
1951 | void *closure) | |
1952 | { | |
1953 | struct list_head *tmp, *next; | |
1954 | struct ldlm_lock *lock; | |
1955 | int rc = LDLM_ITER_CONTINUE; | |
1956 | ||
d7e09d03 | 1957 | if (!res) |
0a3bdb00 | 1958 | return LDLM_ITER_CONTINUE; |
d7e09d03 PT |
1959 | |
1960 | lock_res(res); | |
1961 | list_for_each_safe(tmp, next, &res->lr_granted) { | |
1962 | lock = list_entry(tmp, struct ldlm_lock, l_res_link); | |
1963 | ||
1964 | if (iter(lock, closure) == LDLM_ITER_STOP) | |
1965 | GOTO(out, rc = LDLM_ITER_STOP); | |
1966 | } | |
1967 | ||
1968 | list_for_each_safe(tmp, next, &res->lr_converting) { | |
1969 | lock = list_entry(tmp, struct ldlm_lock, l_res_link); | |
1970 | ||
1971 | if (iter(lock, closure) == LDLM_ITER_STOP) | |
1972 | GOTO(out, rc = LDLM_ITER_STOP); | |
1973 | } | |
1974 | ||
1975 | list_for_each_safe(tmp, next, &res->lr_waiting) { | |
1976 | lock = list_entry(tmp, struct ldlm_lock, l_res_link); | |
1977 | ||
1978 | if (iter(lock, closure) == LDLM_ITER_STOP) | |
1979 | GOTO(out, rc = LDLM_ITER_STOP); | |
1980 | } | |
1981 | out: | |
1982 | unlock_res(res); | |
0a3bdb00 | 1983 | return rc; |
d7e09d03 PT |
1984 | } |
1985 | EXPORT_SYMBOL(ldlm_resource_foreach); | |
1986 | ||
1987 | struct iter_helper_data { | |
1988 | ldlm_iterator_t iter; | |
1989 | void *closure; | |
1990 | }; | |
1991 | ||
1992 | static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure) | |
1993 | { | |
1994 | struct iter_helper_data *helper = closure; | |
1995 | return helper->iter(lock, helper->closure); | |
1996 | } | |
1997 | ||
6da6eabe | 1998 | static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd, |
d7e09d03 PT |
1999 | struct hlist_node *hnode, void *arg) |
2000 | ||
2001 | { | |
2002 | struct ldlm_resource *res = cfs_hash_object(hs, hnode); | |
2003 | ||
2004 | return ldlm_resource_foreach(res, ldlm_iter_helper, arg) == | |
2005 | LDLM_ITER_STOP; | |
2006 | } | |
2007 | ||
2008 | void ldlm_namespace_foreach(struct ldlm_namespace *ns, | |
2009 | ldlm_iterator_t iter, void *closure) | |
2010 | ||
2011 | { | |
805e517a EG |
2012 | struct iter_helper_data helper = { |
2013 | .iter = iter, | |
2014 | .closure = closure, | |
2015 | }; | |
d7e09d03 PT |
2016 | |
2017 | cfs_hash_for_each_nolock(ns->ns_rs_hash, | |
2018 | ldlm_res_iter_helper, &helper); | |
2019 | ||
2020 | } | |
2021 | EXPORT_SYMBOL(ldlm_namespace_foreach); | |
2022 | ||
2023 | /* non-blocking function to manipulate a lock whose cb_data is being put away. | |
2024 | * return 0: find no resource | |
2025 | * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE. | |
2026 | * < 0: errors | |
2027 | */ | |
2028 | int ldlm_resource_iterate(struct ldlm_namespace *ns, | |
2029 | const struct ldlm_res_id *res_id, | |
2030 | ldlm_iterator_t iter, void *data) | |
2031 | { | |
2032 | struct ldlm_resource *res; | |
2033 | int rc; | |
d7e09d03 PT |
2034 | |
2035 | if (ns == NULL) { | |
2036 | CERROR("must pass in namespace\n"); | |
2037 | LBUG(); | |
2038 | } | |
2039 | ||
2040 | res = ldlm_resource_get(ns, NULL, res_id, 0, 0); | |
2041 | if (res == NULL) | |
0a3bdb00 | 2042 | return 0; |
d7e09d03 PT |
2043 | |
2044 | LDLM_RESOURCE_ADDREF(res); | |
2045 | rc = ldlm_resource_foreach(res, iter, data); | |
2046 | LDLM_RESOURCE_DELREF(res); | |
2047 | ldlm_resource_putref(res); | |
0a3bdb00 | 2048 | return rc; |
d7e09d03 PT |
2049 | } |
2050 | EXPORT_SYMBOL(ldlm_resource_iterate); | |
2051 | ||
2052 | /* Lock replay */ | |
2053 | ||
2054 | static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure) | |
2055 | { | |
2056 | struct list_head *list = closure; | |
2057 | ||
2058 | /* we use l_pending_chain here, because it's unused on clients. */ | |
2059 | LASSERTF(list_empty(&lock->l_pending_chain), | |
2060 | "lock %p next %p prev %p\n", | |
2061 | lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev); | |
2062 | /* bug 9573: don't replay locks left after eviction, or | |
2063 | * bug 17614: locks being actively cancelled. Get a reference | |
6e3dd654 | 2064 | * on a lock so that it does not disappear under us (e.g. due to cancel) |
d7e09d03 PT |
2065 | */ |
2066 | if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) { | |
2067 | list_add(&lock->l_pending_chain, list); | |
2068 | LDLM_LOCK_GET(lock); | |
2069 | } | |
2070 | ||
2071 | return LDLM_ITER_CONTINUE; | |
2072 | } | |
2073 | ||
2074 | static int replay_lock_interpret(const struct lu_env *env, | |
2075 | struct ptlrpc_request *req, | |
2076 | struct ldlm_async_args *aa, int rc) | |
2077 | { | |
2078 | struct ldlm_lock *lock; | |
2079 | struct ldlm_reply *reply; | |
2080 | struct obd_export *exp; | |
2081 | ||
d7e09d03 PT |
2082 | atomic_dec(&req->rq_import->imp_replay_inflight); |
2083 | if (rc != ELDLM_OK) | |
2084 | GOTO(out, rc); | |
2085 | ||
2086 | ||
2087 | reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); | |
2088 | if (reply == NULL) | |
2089 | GOTO(out, rc = -EPROTO); | |
2090 | ||
2091 | lock = ldlm_handle2lock(&aa->lock_handle); | |
2092 | if (!lock) { | |
55f5a824 GKH |
2093 | CERROR("received replay ack for unknown local cookie %#llx" |
2094 | " remote cookie %#llx from server %s id %s\n", | |
d7e09d03 PT |
2095 | aa->lock_handle.cookie, reply->lock_handle.cookie, |
2096 | req->rq_export->exp_client_uuid.uuid, | |
2097 | libcfs_id2str(req->rq_peer)); | |
2098 | GOTO(out, rc = -ESTALE); | |
2099 | } | |
2100 | ||
2101 | /* Key change rehash lock in per-export hash with new key */ | |
2102 | exp = req->rq_export; | |
2103 | if (exp && exp->exp_lock_hash) { | |
2104 | /* In the function below, .hs_keycmp resolves to | |
2105 | * ldlm_export_lock_keycmp() */ | |
2106 | /* coverity[overrun-buffer-val] */ | |
2107 | cfs_hash_rehash_key(exp->exp_lock_hash, | |
2108 | &lock->l_remote_handle, | |
2109 | &reply->lock_handle, | |
2110 | &lock->l_exp_hash); | |
2111 | } else { | |
2112 | lock->l_remote_handle = reply->lock_handle; | |
2113 | } | |
2114 | ||
2115 | LDLM_DEBUG(lock, "replayed lock:"); | |
2116 | ptlrpc_import_recovery_state_machine(req->rq_import); | |
2117 | LDLM_LOCK_PUT(lock); | |
2118 | out: | |
2119 | if (rc != ELDLM_OK) | |
2120 | ptlrpc_connect_import(req->rq_import); | |
2121 | ||
0a3bdb00 | 2122 | return rc; |
d7e09d03 PT |
2123 | } |
2124 | ||
2125 | static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) | |
2126 | { | |
2127 | struct ptlrpc_request *req; | |
2128 | struct ldlm_async_args *aa; | |
2129 | struct ldlm_request *body; | |
2130 | int flags; | |
d7e09d03 PT |
2131 | |
2132 | /* Bug 11974: Do not replay a lock which is actively being canceled */ | |
2133 | if (lock->l_flags & LDLM_FL_CANCELING) { | |
2134 | LDLM_DEBUG(lock, "Not replaying canceled lock:"); | |
0a3bdb00 | 2135 | return 0; |
d7e09d03 PT |
2136 | } |
2137 | ||
2138 | /* If this is reply-less callback lock, we cannot replay it, since | |
2139 | * server might have long dropped it, but notification of that event was | |
2140 | * lost by network. (and server granted conflicting lock already) */ | |
2141 | if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) { | |
2142 | LDLM_DEBUG(lock, "Not replaying reply-less lock:"); | |
2143 | ldlm_lock_cancel(lock); | |
0a3bdb00 | 2144 | return 0; |
d7e09d03 PT |
2145 | } |
2146 | ||
2147 | /* | |
2148 | * If granted mode matches the requested mode, this lock is granted. | |
2149 | * | |
2150 | * If they differ, but we have a granted mode, then we were granted | |
2151 | * one mode and now want another: ergo, converting. | |
2152 | * | |
2153 | * If we haven't been granted anything and are on a resource list, | |
2154 | * then we're blocked/waiting. | |
2155 | * | |
2156 | * If we haven't been granted anything and we're NOT on a resource list, | |
2157 | * then we haven't got a reply yet and don't have a known disposition. | |
2158 | * This happens whenever a lock enqueue is the request that triggers | |
2159 | * recovery. | |
2160 | */ | |
2161 | if (lock->l_granted_mode == lock->l_req_mode) | |
2162 | flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED; | |
2163 | else if (lock->l_granted_mode) | |
2164 | flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV; | |
2165 | else if (!list_empty(&lock->l_res_link)) | |
2166 | flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT; | |
2167 | else | |
2168 | flags = LDLM_FL_REPLAY; | |
2169 | ||
2170 | req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE, | |
2171 | LUSTRE_DLM_VERSION, LDLM_ENQUEUE); | |
2172 | if (req == NULL) | |
0a3bdb00 | 2173 | return -ENOMEM; |
d7e09d03 PT |
2174 | |
2175 | /* We're part of recovery, so don't wait for it. */ | |
2176 | req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS; | |
2177 | ||
2178 | body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); | |
2179 | ldlm_lock2desc(lock, &body->lock_desc); | |
2180 | body->lock_flags = ldlm_flags_to_wire(flags); | |
2181 | ||
2182 | ldlm_lock2handle(lock, &body->lock_handle[0]); | |
2183 | if (lock->l_lvb_len > 0) | |
2184 | req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB); | |
2185 | req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, | |
2186 | lock->l_lvb_len); | |
2187 | ptlrpc_request_set_replen(req); | |
2188 | /* notify the server we've replayed all requests. | |
2189 | * also, we mark the request to be put on a dedicated | |
2190 | * queue to be processed after all request replayes. | |
2191 | * bug 6063 */ | |
2192 | lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE); | |
2193 | ||
2194 | LDLM_DEBUG(lock, "replaying lock:"); | |
2195 | ||
2196 | atomic_inc(&req->rq_import->imp_replay_inflight); | |
2197 | CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args)); | |
2198 | aa = ptlrpc_req_async_args(req); | |
2199 | aa->lock_handle = body->lock_handle[0]; | |
2200 | req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret; | |
2201 | ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1); | |
2202 | ||
0a3bdb00 | 2203 | return 0; |
d7e09d03 PT |
2204 | } |
2205 | ||
2206 | /** | |
2207 | * Cancel as many unused locks as possible before replay. since we are | |
2208 | * in recovery, we can't wait for any outstanding RPCs to send any RPC | |
2209 | * to the server. | |
2210 | * | |
2211 | * Called only in recovery before replaying locks. there is no need to | |
2212 | * replay locks that are unused. since the clients may hold thousands of | |
2213 | * cached unused locks, dropping the unused locks can greatly reduce the | |
2214 | * load on the servers at recovery time. | |
2215 | */ | |
2216 | static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns) | |
2217 | { | |
2218 | int canceled; | |
2219 | LIST_HEAD(cancels); | |
2220 | ||
2221 | CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before" | |
2222 | "replay for namespace %s (%d)\n", | |
2223 | ldlm_ns_name(ns), ns->ns_nr_unused); | |
2224 | ||
2225 | /* We don't need to care whether or not LRU resize is enabled | |
2226 | * because the LDLM_CANCEL_NO_WAIT policy doesn't use the | |
2227 | * count parameter */ | |
2228 | canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0, | |
2229 | LCF_LOCAL, LDLM_CANCEL_NO_WAIT); | |
2230 | ||
2231 | CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n", | |
2232 | canceled, ldlm_ns_name(ns)); | |
2233 | } | |
2234 | ||
2235 | int ldlm_replay_locks(struct obd_import *imp) | |
2236 | { | |
2237 | struct ldlm_namespace *ns = imp->imp_obd->obd_namespace; | |
2238 | LIST_HEAD(list); | |
2239 | struct ldlm_lock *lock, *next; | |
2240 | int rc = 0; | |
2241 | ||
d7e09d03 PT |
2242 | LASSERT(atomic_read(&imp->imp_replay_inflight) == 0); |
2243 | ||
2244 | /* don't replay locks if import failed recovery */ | |
2245 | if (imp->imp_vbr_failed) | |
0a3bdb00 | 2246 | return 0; |
d7e09d03 PT |
2247 | |
2248 | /* ensure this doesn't fall to 0 before all have been queued */ | |
2249 | atomic_inc(&imp->imp_replay_inflight); | |
2250 | ||
2251 | if (ldlm_cancel_unused_locks_before_replay) | |
2252 | ldlm_cancel_unused_locks_for_replay(ns); | |
2253 | ||
2254 | ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list); | |
2255 | ||
2256 | list_for_each_entry_safe(lock, next, &list, l_pending_chain) { | |
2257 | list_del_init(&lock->l_pending_chain); | |
2258 | if (rc) { | |
2259 | LDLM_LOCK_RELEASE(lock); | |
2260 | continue; /* or try to do the rest? */ | |
2261 | } | |
2262 | rc = replay_one_lock(imp, lock); | |
2263 | LDLM_LOCK_RELEASE(lock); | |
2264 | } | |
2265 | ||
2266 | atomic_dec(&imp->imp_replay_inflight); | |
2267 | ||
0a3bdb00 | 2268 | return rc; |
d7e09d03 PT |
2269 | } |
2270 | EXPORT_SYMBOL(ldlm_replay_locks); |