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 | |
6a5b99a4 | 18 | * http://www.gnu.org/licenses/gpl-2.0.html |
d7e09d03 | 19 | * |
d7e09d03 PT |
20 | * GPL HEADER END |
21 | */ | |
22 | /* | |
23 | * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. | |
24 | * Use is subject to license terms. | |
25 | * | |
1dc563a6 | 26 | * Copyright (c) 2010, 2015, Intel Corporation. |
d7e09d03 PT |
27 | */ |
28 | /* | |
29 | * This file is part of Lustre, http://www.lustre.org/ | |
30 | * Lustre is a trademark of Sun Microsystems, Inc. | |
31 | */ | |
32 | /** | |
33 | * This file contains Asynchronous System Trap (AST) handlers and related | |
34 | * LDLM request-processing routines. | |
35 | * | |
36 | * An AST is a callback issued on a lock when its state is changed. There are | |
37 | * several different types of ASTs (callbacks) registered for each lock: | |
38 | * | |
39 | * - completion AST: when a lock is enqueued by some process, but cannot be | |
40 | * granted immediately due to other conflicting locks on the same resource, | |
41 | * the completion AST is sent to notify the caller when the lock is | |
42 | * eventually granted | |
43 | * | |
44 | * - blocking AST: when a lock is granted to some process, if another process | |
45 | * enqueues a conflicting (blocking) lock on a resource, a blocking AST is | |
46 | * sent to notify the holder(s) of the lock(s) of the conflicting lock | |
47 | * request. The lock holder(s) must release their lock(s) on that resource in | |
48 | * a timely manner or be evicted by the server. | |
49 | * | |
50 | * - glimpse AST: this is used when a process wants information about a lock | |
51 | * (i.e. the lock value block (LVB)) but does not necessarily require holding | |
52 | * the lock. If the resource is locked, the lock holder(s) are sent glimpse | |
53 | * ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL | |
54 | * their lock(s) if they are idle. If the resource is not locked, the server | |
55 | * may grant the lock. | |
56 | */ | |
57 | ||
58 | #define DEBUG_SUBSYSTEM S_LDLM | |
59 | ||
e27db149 GKH |
60 | #include "../include/lustre_dlm.h" |
61 | #include "../include/obd_class.h" | |
62 | #include "../include/obd.h" | |
d7e09d03 PT |
63 | |
64 | #include "ldlm_internal.h" | |
65 | ||
66 | int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT; | |
8cc7b4b9 PT |
67 | module_param(ldlm_enqueue_min, int, 0644); |
68 | MODULE_PARM_DESC(ldlm_enqueue_min, "lock enqueue timeout minimum"); | |
d7e09d03 PT |
69 | |
70 | /* in client side, whether the cached locks will be canceled before replay */ | |
71 | unsigned int ldlm_cancel_unused_locks_before_replay = 1; | |
72 | ||
73 | static void interrupted_completion_wait(void *data) | |
74 | { | |
75 | } | |
76 | ||
77 | struct lock_wait_data { | |
78 | struct ldlm_lock *lwd_lock; | |
79 | __u32 lwd_conn_cnt; | |
80 | }; | |
81 | ||
82 | struct ldlm_async_args { | |
83 | struct lustre_handle lock_handle; | |
84 | }; | |
85 | ||
58c6d133 | 86 | static int ldlm_expired_completion_wait(void *data) |
d7e09d03 PT |
87 | { |
88 | struct lock_wait_data *lwd = data; | |
89 | struct ldlm_lock *lock = lwd->lwd_lock; | |
90 | struct obd_import *imp; | |
91 | struct obd_device *obd; | |
92 | ||
44b53f18 | 93 | if (!lock->l_conn_export) { |
9844f9c9 | 94 | static unsigned long next_dump, last_dump; |
d7e09d03 | 95 | |
bf6d2153 AB |
96 | LCONSOLE_WARN("lock timed out (enqueued at %lld, %llds ago)\n", |
97 | (s64)lock->l_last_activity, | |
98 | (s64)(ktime_get_real_seconds() - | |
99 | lock->l_last_activity)); | |
100 | LDLM_DEBUG(lock, "lock timed out (enqueued at %lld, %llds ago); not entering recovery in server code, just going back to sleep", | |
101 | (s64)lock->l_last_activity, | |
102 | (s64)(ktime_get_real_seconds() - | |
103 | lock->l_last_activity)); | |
d7e09d03 PT |
104 | if (cfs_time_after(cfs_time_current(), next_dump)) { |
105 | last_dump = next_dump; | |
106 | next_dump = cfs_time_shift(300); | |
107 | ldlm_namespace_dump(D_DLMTRACE, | |
108 | ldlm_lock_to_ns(lock)); | |
109 | if (last_dump == 0) | |
110 | libcfs_debug_dumplog(); | |
111 | } | |
0a3bdb00 | 112 | return 0; |
d7e09d03 PT |
113 | } |
114 | ||
115 | obd = lock->l_conn_export->exp_obd; | |
116 | imp = obd->u.cli.cl_import; | |
117 | ptlrpc_fail_import(imp, lwd->lwd_conn_cnt); | |
bf6d2153 AB |
118 | LDLM_ERROR(lock, "lock timed out (enqueued at %lld, %llds ago), entering recovery for %s@%s", |
119 | (s64)lock->l_last_activity, | |
120 | (s64)(ktime_get_real_seconds() - lock->l_last_activity), | |
121 | obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid); | |
d7e09d03 | 122 | |
0a3bdb00 | 123 | return 0; |
d7e09d03 | 124 | } |
d7e09d03 PT |
125 | |
126 | /* We use the same basis for both server side and client side functions | |
6f789a6a OD |
127 | * from a single node. |
128 | */ | |
58c6d133 | 129 | static int ldlm_get_enq_timeout(struct ldlm_lock *lock) |
d7e09d03 PT |
130 | { |
131 | int timeout = at_get(ldlm_lock_to_ns_at(lock)); | |
902f3bb1 | 132 | |
d7e09d03 PT |
133 | if (AT_OFF) |
134 | return obd_timeout / 2; | |
135 | /* Since these are non-updating timeouts, we should be conservative. | |
6f789a6a OD |
136 | * It would be nice to have some kind of "early reply" mechanism for |
137 | * lock callbacks too... | |
138 | */ | |
d7e09d03 PT |
139 | timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */ |
140 | return max(timeout, ldlm_enqueue_min); | |
141 | } | |
d7e09d03 PT |
142 | |
143 | /** | |
144 | * Helper function for ldlm_completion_ast(), updating timings when lock is | |
145 | * actually granted. | |
146 | */ | |
147 | static int ldlm_completion_tail(struct ldlm_lock *lock) | |
148 | { | |
149 | long delay; | |
150 | int result; | |
151 | ||
5a9a80ba | 152 | if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) { |
d7e09d03 PT |
153 | LDLM_DEBUG(lock, "client-side enqueue: destroyed"); |
154 | result = -EIO; | |
155 | } else { | |
bf6d2153 AB |
156 | delay = ktime_get_real_seconds() - lock->l_last_activity; |
157 | LDLM_DEBUG(lock, "client-side enqueue: granted after %lds", | |
158 | delay); | |
d7e09d03 PT |
159 | |
160 | /* Update our time estimate */ | |
161 | at_measured(ldlm_lock_to_ns_at(lock), | |
162 | delay); | |
163 | result = 0; | |
164 | } | |
165 | return result; | |
166 | } | |
167 | ||
168 | /** | |
169 | * Implementation of ->l_completion_ast() for a client, that doesn't wait | |
170 | * until lock is granted. Suitable for locks enqueued through ptlrpcd, of | |
171 | * other threads that cannot block for long. | |
172 | */ | |
173 | int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data) | |
174 | { | |
d7e09d03 PT |
175 | if (flags == LDLM_FL_WAIT_NOREPROC) { |
176 | LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock"); | |
0a3bdb00 | 177 | return 0; |
d7e09d03 PT |
178 | } |
179 | ||
180 | if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED | | |
181 | LDLM_FL_BLOCK_CONV))) { | |
182 | wake_up(&lock->l_waitq); | |
0a3bdb00 | 183 | return ldlm_completion_tail(lock); |
d7e09d03 PT |
184 | } |
185 | ||
2d00bd17 | 186 | LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, going forward"); |
0a3bdb00 | 187 | return 0; |
d7e09d03 PT |
188 | } |
189 | EXPORT_SYMBOL(ldlm_completion_ast_async); | |
190 | ||
191 | /** | |
192 | * Generic LDLM "completion" AST. This is called in several cases: | |
193 | * | |
194 | * - when a reply to an ENQUEUE RPC is received from the server | |
195 | * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at | |
196 | * this point (determined by flags); | |
197 | * | |
198 | * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has | |
199 | * been granted; | |
200 | * | |
201 | * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock | |
202 | * gets correct lvb; | |
203 | * | |
204 | * - to force all locks when resource is destroyed (cleanup_resource()); | |
205 | * | |
206 | * - during lock conversion (not used currently). | |
207 | * | |
208 | * If lock is not granted in the first case, this function waits until second | |
209 | * or penultimate cases happen in some other thread. | |
210 | * | |
211 | */ | |
212 | int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) | |
213 | { | |
214 | /* XXX ALLOCATE - 160 bytes */ | |
215 | struct lock_wait_data lwd; | |
216 | struct obd_device *obd; | |
217 | struct obd_import *imp = NULL; | |
218 | struct l_wait_info lwi; | |
219 | __u32 timeout; | |
220 | int rc = 0; | |
d7e09d03 PT |
221 | |
222 | if (flags == LDLM_FL_WAIT_NOREPROC) { | |
223 | LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock"); | |
224 | goto noreproc; | |
225 | } | |
226 | ||
227 | if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED | | |
228 | LDLM_FL_BLOCK_CONV))) { | |
229 | wake_up(&lock->l_waitq); | |
0a3bdb00 | 230 | return 0; |
d7e09d03 PT |
231 | } |
232 | ||
2d00bd17 | 233 | LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, sleeping"); |
d7e09d03 PT |
234 | |
235 | noreproc: | |
236 | ||
237 | obd = class_exp2obd(lock->l_conn_export); | |
238 | ||
239 | /* if this is a local lock, then there is no import */ | |
44b53f18 | 240 | if (obd) |
d7e09d03 | 241 | imp = obd->u.cli.cl_import; |
d7e09d03 PT |
242 | |
243 | /* Wait a long time for enqueue - server may have to callback a | |
6f789a6a OD |
244 | * lock from another client. Server will evict the other client if it |
245 | * doesn't respond reasonably, and then give us the lock. | |
246 | */ | |
d7e09d03 PT |
247 | timeout = ldlm_get_enq_timeout(lock) * 2; |
248 | ||
249 | lwd.lwd_lock = lock; | |
250 | ||
5a9a80ba | 251 | if (ldlm_is_no_timeout(lock)) { |
d7e09d03 PT |
252 | LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT"); |
253 | lwi = LWI_INTR(interrupted_completion_wait, &lwd); | |
254 | } else { | |
255 | lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout), | |
256 | ldlm_expired_completion_wait, | |
257 | interrupted_completion_wait, &lwd); | |
258 | } | |
259 | ||
44b53f18 | 260 | if (imp) { |
d7e09d03 PT |
261 | spin_lock(&imp->imp_lock); |
262 | lwd.lwd_conn_cnt = imp->imp_conn_cnt; | |
263 | spin_unlock(&imp->imp_lock); | |
264 | } | |
265 | ||
cf739f84 | 266 | if (OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST, |
d7e09d03 | 267 | OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) { |
5a9a80ba | 268 | ldlm_set_fail_loc(lock); |
d7e09d03 PT |
269 | rc = -EINTR; |
270 | } else { | |
271 | /* Go to sleep until the lock is granted or cancelled. */ | |
272 | rc = l_wait_event(lock->l_waitq, | |
273 | is_granted_or_cancelled(lock), &lwi); | |
274 | } | |
275 | ||
276 | if (rc) { | |
277 | LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)", | |
278 | rc); | |
0a3bdb00 | 279 | return rc; |
d7e09d03 PT |
280 | } |
281 | ||
0a3bdb00 | 282 | return ldlm_completion_tail(lock); |
d7e09d03 PT |
283 | } |
284 | EXPORT_SYMBOL(ldlm_completion_ast); | |
285 | ||
d7e09d03 PT |
286 | static void failed_lock_cleanup(struct ldlm_namespace *ns, |
287 | struct ldlm_lock *lock, int mode) | |
288 | { | |
289 | int need_cancel = 0; | |
290 | ||
291 | /* Set a flag to prevent us from sending a CANCEL (bug 407) */ | |
292 | lock_res_and_lock(lock); | |
293 | /* Check that lock is not granted or failed, we might race. */ | |
294 | if ((lock->l_req_mode != lock->l_granted_mode) && | |
5a9a80ba | 295 | !ldlm_is_failed(lock)) { |
d7e09d03 PT |
296 | /* Make sure that this lock will not be found by raced |
297 | * bl_ast and -EINVAL reply is sent to server anyways. | |
6f789a6a OD |
298 | * bug 17645 |
299 | */ | |
d7e09d03 PT |
300 | lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED | |
301 | LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING; | |
302 | need_cancel = 1; | |
303 | } | |
304 | unlock_res_and_lock(lock); | |
305 | ||
306 | if (need_cancel) | |
307 | LDLM_DEBUG(lock, | |
2d00bd17 | 308 | "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING"); |
d7e09d03 PT |
309 | else |
310 | LDLM_DEBUG(lock, "lock was granted or failed in race"); | |
311 | ||
d7e09d03 | 312 | /* XXX - HACK because we shouldn't call ldlm_lock_destroy() |
6f789a6a OD |
313 | * from llite/file.c/ll_file_flock(). |
314 | */ | |
d7e09d03 PT |
315 | /* This code makes for the fact that we do not have blocking handler on |
316 | * a client for flock locks. As such this is the place where we must | |
317 | * completely kill failed locks. (interrupted and those that | |
6f789a6a OD |
318 | * were waiting to be granted when server evicted us. |
319 | */ | |
d7e09d03 PT |
320 | if (lock->l_resource->lr_type == LDLM_FLOCK) { |
321 | lock_res_and_lock(lock); | |
c68c3fa4 VF |
322 | if (!ldlm_is_destroyed(lock)) { |
323 | ldlm_resource_unlink_lock(lock); | |
324 | ldlm_lock_decref_internal_nolock(lock, mode); | |
325 | ldlm_lock_destroy_nolock(lock); | |
326 | } | |
d7e09d03 | 327 | unlock_res_and_lock(lock); |
c68c3fa4 VF |
328 | } else { |
329 | ldlm_lock_decref_internal(lock, mode); | |
d7e09d03 PT |
330 | } |
331 | } | |
332 | ||
333 | /** | |
334 | * Finishing portion of client lock enqueue code. | |
335 | * | |
336 | * Called after receiving reply from server. | |
337 | */ | |
338 | int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, | |
52ee0d20 OD |
339 | enum ldlm_type type, __u8 with_policy, |
340 | enum ldlm_mode mode, | |
d7e09d03 | 341 | __u64 *flags, void *lvb, __u32 lvb_len, |
e8beaf67 | 342 | const struct lustre_handle *lockh, int rc) |
d7e09d03 PT |
343 | { |
344 | struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; | |
345 | int is_replay = *flags & LDLM_FL_REPLAY; | |
346 | struct ldlm_lock *lock; | |
347 | struct ldlm_reply *reply; | |
348 | int cleanup_phase = 1; | |
d7e09d03 PT |
349 | |
350 | lock = ldlm_handle2lock(lockh); | |
351 | /* ldlm_cli_enqueue is holding a reference on this lock. */ | |
352 | if (!lock) { | |
353 | LASSERT(type == LDLM_FLOCK); | |
0a3bdb00 | 354 | return -ENOLCK; |
d7e09d03 PT |
355 | } |
356 | ||
357 | LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len), | |
358 | "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len); | |
359 | ||
360 | if (rc != ELDLM_OK) { | |
361 | LASSERT(!is_replay); | |
362 | LDLM_DEBUG(lock, "client-side enqueue END (%s)", | |
363 | rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED"); | |
364 | ||
365 | if (rc != ELDLM_LOCK_ABORTED) | |
d1c0d446 | 366 | goto cleanup; |
d7e09d03 PT |
367 | } |
368 | ||
369 | /* Before we return, swab the reply */ | |
370 | reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); | |
44b53f18 | 371 | if (!reply) { |
d1c0d446 JL |
372 | rc = -EPROTO; |
373 | goto cleanup; | |
374 | } | |
d7e09d03 | 375 | |
06563b56 JX |
376 | if (lvb_len > 0) { |
377 | int size = 0; | |
d7e09d03 PT |
378 | |
379 | size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, | |
380 | RCL_SERVER); | |
381 | if (size < 0) { | |
382 | LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size); | |
d1c0d446 JL |
383 | rc = size; |
384 | goto cleanup; | |
d7e09d03 | 385 | } else if (unlikely(size > lvb_len)) { |
2d00bd17 | 386 | LDLM_ERROR(lock, "Replied LVB is larger than expectation, expected = %d, replied = %d", |
d7e09d03 | 387 | lvb_len, size); |
d1c0d446 JL |
388 | rc = -EINVAL; |
389 | goto cleanup; | |
d7e09d03 | 390 | } |
06563b56 | 391 | lvb_len = size; |
d7e09d03 PT |
392 | } |
393 | ||
394 | if (rc == ELDLM_LOCK_ABORTED) { | |
06563b56 | 395 | if (lvb_len > 0 && lvb) |
d7e09d03 | 396 | rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER, |
06563b56 | 397 | lvb, lvb_len); |
31664dc6 JL |
398 | if (rc == 0) |
399 | rc = ELDLM_LOCK_ABORTED; | |
d1c0d446 | 400 | goto cleanup; |
d7e09d03 PT |
401 | } |
402 | ||
403 | /* lock enqueued on the server */ | |
404 | cleanup_phase = 0; | |
405 | ||
406 | lock_res_and_lock(lock); | |
407 | /* Key change rehash lock in per-export hash with new key */ | |
408 | if (exp->exp_lock_hash) { | |
409 | /* In the function below, .hs_keycmp resolves to | |
6f789a6a OD |
410 | * ldlm_export_lock_keycmp() |
411 | */ | |
d7e09d03 PT |
412 | /* coverity[overrun-buffer-val] */ |
413 | cfs_hash_rehash_key(exp->exp_lock_hash, | |
414 | &lock->l_remote_handle, | |
415 | &reply->lock_handle, | |
416 | &lock->l_exp_hash); | |
417 | } else { | |
418 | lock->l_remote_handle = reply->lock_handle; | |
419 | } | |
420 | ||
421 | *flags = ldlm_flags_from_wire(reply->lock_flags); | |
422 | lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags & | |
d7696405 | 423 | LDLM_FL_INHERIT_MASK); |
d7e09d03 PT |
424 | unlock_res_and_lock(lock); |
425 | ||
55f5a824 | 426 | CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: 0x%llx\n", |
d7e09d03 PT |
427 | lock, reply->lock_handle.cookie, *flags); |
428 | ||
429 | /* If enqueue returned a blocked lock but the completion handler has | |
430 | * already run, then it fixed up the resource and we don't need to do it | |
6f789a6a OD |
431 | * again. |
432 | */ | |
d7e09d03 PT |
433 | if ((*flags) & LDLM_FL_LOCK_CHANGED) { |
434 | int newmode = reply->lock_desc.l_req_mode; | |
902f3bb1 | 435 | |
d7e09d03 PT |
436 | LASSERT(!is_replay); |
437 | if (newmode && newmode != lock->l_req_mode) { | |
438 | LDLM_DEBUG(lock, "server returned different mode %s", | |
439 | ldlm_lockname[newmode]); | |
440 | lock->l_req_mode = newmode; | |
441 | } | |
442 | ||
6d95e048 AD |
443 | if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name, |
444 | &lock->l_resource->lr_name)) { | |
445 | CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES | |
446 | " instead of "DLDLMRES"\n", | |
447 | PLDLMRES(&reply->lock_desc.l_resource), | |
448 | PLDLMRES(lock->l_resource)); | |
d7e09d03 PT |
449 | |
450 | rc = ldlm_lock_change_resource(ns, lock, | |
451 | &reply->lock_desc.l_resource.lr_name); | |
44b53f18 | 452 | if (rc || !lock->l_resource) { |
d1c0d446 JL |
453 | rc = -ENOMEM; |
454 | goto cleanup; | |
455 | } | |
d7e09d03 PT |
456 | LDLM_DEBUG(lock, "client-side enqueue, new resource"); |
457 | } | |
458 | if (with_policy) | |
459 | if (!(type == LDLM_IBITS && | |
460 | !(exp_connect_flags(exp) & OBD_CONNECT_IBITS))) | |
461 | /* We assume lock type cannot change on server*/ | |
462 | ldlm_convert_policy_to_local(exp, | |
463 | lock->l_resource->lr_type, | |
464 | &reply->lock_desc.l_policy_data, | |
465 | &lock->l_policy_data); | |
466 | if (type != LDLM_PLAIN) | |
1d8cb70c GD |
467 | LDLM_DEBUG(lock, |
468 | "client-side enqueue, new policy data"); | |
d7e09d03 PT |
469 | } |
470 | ||
471 | if ((*flags) & LDLM_FL_AST_SENT || | |
472 | /* Cancel extent locks as soon as possible on a liblustre client, | |
473 | * because it cannot handle asynchronous ASTs robustly (see | |
6f789a6a OD |
474 | * bug 7311). |
475 | */ | |
d7e09d03 PT |
476 | (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) { |
477 | lock_res_and_lock(lock); | |
478 | lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST; | |
479 | unlock_res_and_lock(lock); | |
480 | LDLM_DEBUG(lock, "enqueue reply includes blocking AST"); | |
481 | } | |
482 | ||
483 | /* If the lock has already been granted by a completion AST, don't | |
6f789a6a OD |
484 | * clobber the LVB with an older one. |
485 | */ | |
06563b56 | 486 | if (lvb_len > 0) { |
d7e09d03 PT |
487 | /* We must lock or a racing completion might update lvb without |
488 | * letting us know and we'll clobber the correct value. | |
6f789a6a OD |
489 | * Cannot unlock after the check either, as that still leaves |
490 | * a tiny window for completion to get in | |
491 | */ | |
d7e09d03 PT |
492 | lock_res_and_lock(lock); |
493 | if (lock->l_req_mode != lock->l_granted_mode) | |
494 | rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER, | |
06563b56 | 495 | lock->l_lvb_data, lvb_len); |
d7e09d03 PT |
496 | unlock_res_and_lock(lock); |
497 | if (rc < 0) { | |
498 | cleanup_phase = 1; | |
d1c0d446 | 499 | goto cleanup; |
d7e09d03 PT |
500 | } |
501 | } | |
502 | ||
503 | if (!is_replay) { | |
504 | rc = ldlm_lock_enqueue(ns, &lock, NULL, flags); | |
44b53f18 | 505 | if (lock->l_completion_ast) { |
d7e09d03 | 506 | int err = lock->l_completion_ast(lock, *flags, NULL); |
902f3bb1 | 507 | |
d7e09d03 PT |
508 | if (!rc) |
509 | rc = err; | |
510 | if (rc) | |
511 | cleanup_phase = 1; | |
512 | } | |
513 | } | |
514 | ||
06563b56 | 515 | if (lvb_len > 0 && lvb) { |
d7e09d03 | 516 | /* Copy the LVB here, and not earlier, because the completion |
6f789a6a OD |
517 | * AST (if any) can override what we got in the reply |
518 | */ | |
d7e09d03 PT |
519 | memcpy(lvb, lock->l_lvb_data, lvb_len); |
520 | } | |
521 | ||
522 | LDLM_DEBUG(lock, "client-side enqueue END"); | |
d7e09d03 PT |
523 | cleanup: |
524 | if (cleanup_phase == 1 && rc) | |
525 | failed_lock_cleanup(ns, lock, mode); | |
526 | /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */ | |
527 | LDLM_LOCK_PUT(lock); | |
528 | LDLM_LOCK_RELEASE(lock); | |
529 | return rc; | |
530 | } | |
531 | EXPORT_SYMBOL(ldlm_cli_enqueue_fini); | |
532 | ||
533 | /** | |
534 | * Estimate number of lock handles that would fit into request of given | |
535 | * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into | |
536 | * a single page on the send/receive side. XXX: 512 should be changed to | |
537 | * more adequate value. | |
538 | */ | |
539 | static inline int ldlm_req_handles_avail(int req_size, int off) | |
540 | { | |
541 | int avail; | |
542 | ||
09cbfeaf | 543 | avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512) - req_size; |
d7e09d03 PT |
544 | if (likely(avail >= 0)) |
545 | avail /= (int)sizeof(struct lustre_handle); | |
546 | else | |
547 | avail = 0; | |
548 | avail += LDLM_LOCKREQ_HANDLES - off; | |
549 | ||
550 | return avail; | |
551 | } | |
552 | ||
553 | static inline int ldlm_capsule_handles_avail(struct req_capsule *pill, | |
554 | enum req_location loc, | |
555 | int off) | |
556 | { | |
557 | int size = req_capsule_msg_size(pill, loc); | |
902f3bb1 | 558 | |
d7e09d03 PT |
559 | return ldlm_req_handles_avail(size, off); |
560 | } | |
561 | ||
562 | static inline int ldlm_format_handles_avail(struct obd_import *imp, | |
563 | const struct req_format *fmt, | |
564 | enum req_location loc, int off) | |
565 | { | |
566 | int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc); | |
902f3bb1 | 567 | |
d7e09d03 PT |
568 | return ldlm_req_handles_avail(size, off); |
569 | } | |
570 | ||
571 | /** | |
572 | * Cancel LRU locks and pack them into the enqueue request. Pack there the given | |
573 | * \a count locks in \a cancels. | |
574 | * | |
575 | * This is to be called by functions preparing their own requests that | |
576 | * might contain lists of locks to cancel in addition to actual operation | |
577 | * that needs to be performed. | |
578 | */ | |
579 | int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req, | |
580 | int version, int opc, int canceloff, | |
581 | struct list_head *cancels, int count) | |
582 | { | |
583 | struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; | |
584 | struct req_capsule *pill = &req->rq_pill; | |
585 | struct ldlm_request *dlm = NULL; | |
586 | int flags, avail, to_free, pack = 0; | |
587 | LIST_HEAD(head); | |
588 | int rc; | |
d7e09d03 | 589 | |
44b53f18 | 590 | if (!cancels) |
d7e09d03 PT |
591 | cancels = &head; |
592 | if (ns_connect_cancelset(ns)) { | |
593 | /* Estimate the amount of available space in the request. */ | |
594 | req_capsule_filled_sizes(pill, RCL_CLIENT); | |
595 | avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff); | |
596 | ||
597 | flags = ns_connect_lru_resize(ns) ? | |
2640256e | 598 | LDLM_CANCEL_LRUR_NO_WAIT : LDLM_CANCEL_AGED; |
d7e09d03 PT |
599 | to_free = !ns_connect_lru_resize(ns) && |
600 | opc == LDLM_ENQUEUE ? 1 : 0; | |
601 | ||
602 | /* Cancel LRU locks here _only_ if the server supports | |
603 | * EARLY_CANCEL. Otherwise we have to send extra CANCEL | |
6f789a6a OD |
604 | * RPC, which will make us slower. |
605 | */ | |
d7e09d03 PT |
606 | if (avail > count) |
607 | count += ldlm_cancel_lru_local(ns, cancels, to_free, | |
608 | avail - count, 0, flags); | |
609 | if (avail > count) | |
610 | pack = count; | |
611 | else | |
612 | pack = avail; | |
613 | req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT, | |
614 | ldlm_request_bufsize(pack, opc)); | |
615 | } | |
616 | ||
617 | rc = ptlrpc_request_pack(req, version, opc); | |
618 | if (rc) { | |
619 | ldlm_lock_list_put(cancels, l_bl_ast, count); | |
0a3bdb00 | 620 | return rc; |
d7e09d03 PT |
621 | } |
622 | ||
623 | if (ns_connect_cancelset(ns)) { | |
624 | if (canceloff) { | |
625 | dlm = req_capsule_client_get(pill, &RMF_DLM_REQ); | |
626 | LASSERT(dlm); | |
627 | /* Skip first lock handler in ldlm_request_pack(), | |
6e3dd654 | 628 | * this method will increment @lock_count according |
d7e09d03 | 629 | * to the lock handle amount actually written to |
6f789a6a OD |
630 | * the buffer. |
631 | */ | |
d7e09d03 PT |
632 | dlm->lock_count = canceloff; |
633 | } | |
634 | /* Pack into the request @pack lock handles. */ | |
635 | ldlm_cli_cancel_list(cancels, pack, req, 0); | |
636 | /* Prepare and send separate cancel RPC for others. */ | |
637 | ldlm_cli_cancel_list(cancels, count - pack, NULL, 0); | |
638 | } else { | |
639 | ldlm_lock_list_put(cancels, l_bl_ast, count); | |
640 | } | |
0a3bdb00 | 641 | return 0; |
d7e09d03 PT |
642 | } |
643 | EXPORT_SYMBOL(ldlm_prep_elc_req); | |
644 | ||
645 | int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req, | |
646 | struct list_head *cancels, int count) | |
647 | { | |
648 | return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, | |
649 | LDLM_ENQUEUE_CANCEL_OFF, cancels, count); | |
650 | } | |
651 | EXPORT_SYMBOL(ldlm_prep_enqueue_req); | |
652 | ||
d7e09d03 PT |
653 | /** |
654 | * Client-side lock enqueue. | |
655 | * | |
656 | * If a request has some specific initialisation it is passed in \a reqp, | |
657 | * otherwise it is created in ldlm_cli_enqueue. | |
658 | * | |
659 | * Supports sync and async requests, pass \a async flag accordingly. If a | |
660 | * request was created in ldlm_cli_enqueue and it is the async request, | |
661 | * pass it to the caller in \a reqp. | |
662 | */ | |
663 | int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, | |
664 | struct ldlm_enqueue_info *einfo, | |
665 | const struct ldlm_res_id *res_id, | |
666 | ldlm_policy_data_t const *policy, __u64 *flags, | |
667 | void *lvb, __u32 lvb_len, enum lvb_type lvb_type, | |
668 | struct lustre_handle *lockh, int async) | |
669 | { | |
670 | struct ldlm_namespace *ns; | |
671 | struct ldlm_lock *lock; | |
672 | struct ldlm_request *body; | |
673 | int is_replay = *flags & LDLM_FL_REPLAY; | |
674 | int req_passed_in = 1; | |
675 | int rc, err; | |
676 | struct ptlrpc_request *req; | |
d7e09d03 | 677 | |
d7e09d03 PT |
678 | ns = exp->exp_obd->obd_namespace; |
679 | ||
680 | /* If we're replaying this lock, just check some invariants. | |
6f789a6a OD |
681 | * If we're creating a new lock, get everything all setup nicely. |
682 | */ | |
d7e09d03 PT |
683 | if (is_replay) { |
684 | lock = ldlm_handle2lock_long(lockh, 0); | |
44b53f18 | 685 | LASSERT(lock); |
d7e09d03 PT |
686 | LDLM_DEBUG(lock, "client-side enqueue START"); |
687 | LASSERT(exp == lock->l_conn_export); | |
688 | } else { | |
689 | const struct ldlm_callback_suite cbs = { | |
690 | .lcs_completion = einfo->ei_cb_cp, | |
f2145eae BK |
691 | .lcs_blocking = einfo->ei_cb_bl, |
692 | .lcs_glimpse = einfo->ei_cb_gl | |
d7e09d03 PT |
693 | }; |
694 | lock = ldlm_lock_create(ns, res_id, einfo->ei_type, | |
695 | einfo->ei_mode, &cbs, einfo->ei_cbdata, | |
696 | lvb_len, lvb_type); | |
099d5adf EL |
697 | if (IS_ERR(lock)) |
698 | return PTR_ERR(lock); | |
d7e09d03 PT |
699 | /* for the local lock, add the reference */ |
700 | ldlm_lock_addref_internal(lock, einfo->ei_mode); | |
701 | ldlm_lock2handle(lock, lockh); | |
44b53f18 OD |
702 | if (policy) |
703 | lock->l_policy_data = *policy; | |
d7e09d03 | 704 | |
a739735c SB |
705 | if (einfo->ei_type == LDLM_EXTENT) { |
706 | /* extent lock without policy is a bug */ | |
707 | if (!policy) | |
708 | LBUG(); | |
709 | ||
d7e09d03 | 710 | lock->l_req_extent = policy->l_extent; |
a739735c | 711 | } |
e93876dd | 712 | LDLM_DEBUG(lock, "client-side enqueue START, flags %llx", |
d7e09d03 PT |
713 | *flags); |
714 | } | |
715 | ||
716 | lock->l_conn_export = exp; | |
717 | lock->l_export = NULL; | |
718 | lock->l_blocking_ast = einfo->ei_cb_bl; | |
d3a8a4e2 | 719 | lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL)); |
d7e09d03 PT |
720 | |
721 | /* lock not sent to server yet */ | |
722 | ||
44b53f18 | 723 | if (!reqp || !*reqp) { |
d7e09d03 PT |
724 | req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), |
725 | &RQF_LDLM_ENQUEUE, | |
726 | LUSTRE_DLM_VERSION, | |
727 | LDLM_ENQUEUE); | |
44b53f18 | 728 | if (!req) { |
d7e09d03 PT |
729 | failed_lock_cleanup(ns, lock, einfo->ei_mode); |
730 | LDLM_LOCK_RELEASE(lock); | |
0a3bdb00 | 731 | return -ENOMEM; |
d7e09d03 PT |
732 | } |
733 | req_passed_in = 0; | |
734 | if (reqp) | |
735 | *reqp = req; | |
736 | } else { | |
737 | int len; | |
738 | ||
739 | req = *reqp; | |
740 | len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, | |
741 | RCL_CLIENT); | |
742 | LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n", | |
743 | DLM_LOCKREQ_OFF, len, (int)sizeof(*body)); | |
744 | } | |
745 | ||
746 | /* Dump lock data into the request buffer */ | |
747 | body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); | |
748 | ldlm_lock2desc(lock, &body->lock_desc); | |
749 | body->lock_flags = ldlm_flags_to_wire(*flags); | |
750 | body->lock_handle[0] = *lockh; | |
751 | ||
752 | /* Continue as normal. */ | |
753 | if (!req_passed_in) { | |
754 | if (lvb_len > 0) | |
755 | req_capsule_extend(&req->rq_pill, | |
756 | &RQF_LDLM_ENQUEUE_LVB); | |
757 | req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, | |
758 | lvb_len); | |
759 | ptlrpc_request_set_replen(req); | |
760 | } | |
761 | ||
762 | /* | |
763 | * Liblustre client doesn't get extent locks, except for O_APPEND case | |
764 | * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where | |
765 | * [i_size, OBD_OBJECT_EOF] lock is taken. | |
766 | */ | |
767 | LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT || | |
768 | policy->l_extent.end == OBD_OBJECT_EOF)); | |
769 | ||
770 | if (async) { | |
44b53f18 | 771 | LASSERT(reqp); |
0a3bdb00 | 772 | return 0; |
d7e09d03 PT |
773 | } |
774 | ||
775 | LDLM_DEBUG(lock, "sending request"); | |
776 | ||
777 | rc = ptlrpc_queue_wait(req); | |
778 | ||
779 | err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0, | |
780 | einfo->ei_mode, flags, lvb, lvb_len, | |
781 | lockh, rc); | |
782 | ||
783 | /* If ldlm_cli_enqueue_fini did not find the lock, we need to free | |
6f789a6a OD |
784 | * one reference that we took |
785 | */ | |
d7e09d03 PT |
786 | if (err == -ENOLCK) |
787 | LDLM_LOCK_RELEASE(lock); | |
788 | else | |
789 | rc = err; | |
790 | ||
44b53f18 | 791 | if (!req_passed_in && req) { |
d7e09d03 PT |
792 | ptlrpc_req_finished(req); |
793 | if (reqp) | |
794 | *reqp = NULL; | |
795 | } | |
796 | ||
0a3bdb00 | 797 | return rc; |
d7e09d03 PT |
798 | } |
799 | EXPORT_SYMBOL(ldlm_cli_enqueue); | |
800 | ||
d7e09d03 PT |
801 | /** |
802 | * Cancel locks locally. | |
803 | * Returns: | |
804 | * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server | |
805 | * \retval LDLM_FL_CANCELING otherwise; | |
806 | * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC. | |
807 | */ | |
808 | static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock) | |
809 | { | |
810 | __u64 rc = LDLM_FL_LOCAL_ONLY; | |
d7e09d03 PT |
811 | |
812 | if (lock->l_conn_export) { | |
813 | bool local_only; | |
814 | ||
815 | LDLM_DEBUG(lock, "client-side cancel"); | |
816 | /* Set this flag to prevent others from getting new references*/ | |
817 | lock_res_and_lock(lock); | |
5a9a80ba | 818 | ldlm_set_cbpending(lock); |
d7e09d03 | 819 | local_only = !!(lock->l_flags & |
cd94f231 | 820 | (LDLM_FL_LOCAL_ONLY | LDLM_FL_CANCEL_ON_BLOCK)); |
d7e09d03 | 821 | ldlm_cancel_callback(lock); |
5a9a80ba | 822 | rc = ldlm_is_bl_ast(lock) ? LDLM_FL_BL_AST : LDLM_FL_CANCELING; |
d7e09d03 PT |
823 | unlock_res_and_lock(lock); |
824 | ||
825 | if (local_only) { | |
2d00bd17 | 826 | CDEBUG(D_DLMTRACE, "not sending request (at caller's instruction)\n"); |
d7e09d03 PT |
827 | rc = LDLM_FL_LOCAL_ONLY; |
828 | } | |
829 | ldlm_lock_cancel(lock); | |
830 | } else { | |
cf739f84 OD |
831 | LDLM_ERROR(lock, "Trying to cancel local lock"); |
832 | LBUG(); | |
d7e09d03 PT |
833 | } |
834 | ||
0a3bdb00 | 835 | return rc; |
d7e09d03 PT |
836 | } |
837 | ||
838 | /** | |
839 | * Pack \a count locks in \a head into ldlm_request buffer of request \a req. | |
840 | */ | |
841 | static void ldlm_cancel_pack(struct ptlrpc_request *req, | |
842 | struct list_head *head, int count) | |
843 | { | |
844 | struct ldlm_request *dlm; | |
845 | struct ldlm_lock *lock; | |
846 | int max, packed = 0; | |
d7e09d03 PT |
847 | |
848 | dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); | |
44b53f18 | 849 | LASSERT(dlm); |
d7e09d03 PT |
850 | |
851 | /* Check the room in the request buffer. */ | |
852 | max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) - | |
853 | sizeof(struct ldlm_request); | |
854 | max /= sizeof(struct lustre_handle); | |
855 | max += LDLM_LOCKREQ_HANDLES; | |
856 | LASSERT(max >= dlm->lock_count + count); | |
857 | ||
858 | /* XXX: it would be better to pack lock handles grouped by resource. | |
859 | * so that the server cancel would call filter_lvbo_update() less | |
6f789a6a OD |
860 | * frequently. |
861 | */ | |
d7e09d03 PT |
862 | list_for_each_entry(lock, head, l_bl_ast) { |
863 | if (!count--) | |
864 | break; | |
865 | LASSERT(lock->l_conn_export); | |
866 | /* Pack the lock handle to the given request buffer. */ | |
867 | LDLM_DEBUG(lock, "packing"); | |
868 | dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle; | |
869 | packed++; | |
870 | } | |
871 | CDEBUG(D_DLMTRACE, "%d locks packed\n", packed); | |
d7e09d03 PT |
872 | } |
873 | ||
874 | /** | |
875 | * Prepare and send a batched cancel RPC. It will include \a count lock | |
6f789a6a OD |
876 | * handles of locks given in \a cancels list. |
877 | */ | |
58c6d133 OD |
878 | static int ldlm_cli_cancel_req(struct obd_export *exp, |
879 | struct list_head *cancels, | |
f833ee42 | 880 | int count, enum ldlm_cancel_flags flags) |
d7e09d03 PT |
881 | { |
882 | struct ptlrpc_request *req = NULL; | |
883 | struct obd_import *imp; | |
884 | int free, sent = 0; | |
885 | int rc = 0; | |
d7e09d03 | 886 | |
44b53f18 | 887 | LASSERT(exp); |
d7e09d03 PT |
888 | LASSERT(count > 0); |
889 | ||
890 | CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val); | |
891 | ||
892 | if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE)) | |
0a3bdb00 | 893 | return count; |
d7e09d03 PT |
894 | |
895 | free = ldlm_format_handles_avail(class_exp2cliimp(exp), | |
896 | &RQF_LDLM_CANCEL, RCL_CLIENT, 0); | |
897 | if (count > free) | |
898 | count = free; | |
899 | ||
900 | while (1) { | |
901 | imp = class_exp2cliimp(exp); | |
44b53f18 | 902 | if (!imp || imp->imp_invalid) { |
d7e09d03 PT |
903 | CDEBUG(D_DLMTRACE, |
904 | "skipping cancel on invalid import %p\n", imp); | |
0a3bdb00 | 905 | return count; |
d7e09d03 PT |
906 | } |
907 | ||
908 | req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL); | |
44b53f18 | 909 | if (!req) { |
d1c0d446 JL |
910 | rc = -ENOMEM; |
911 | goto out; | |
912 | } | |
d7e09d03 PT |
913 | |
914 | req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT); | |
915 | req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT, | |
916 | ldlm_request_bufsize(count, LDLM_CANCEL)); | |
917 | ||
918 | rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL); | |
919 | if (rc) { | |
920 | ptlrpc_request_free(req); | |
d1c0d446 | 921 | goto out; |
d7e09d03 PT |
922 | } |
923 | ||
924 | req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL; | |
925 | req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL; | |
926 | ptlrpc_at_set_req_timeout(req); | |
927 | ||
928 | ldlm_cancel_pack(req, cancels, count); | |
929 | ||
930 | ptlrpc_request_set_replen(req); | |
931 | if (flags & LCF_ASYNC) { | |
c5c4c6fa | 932 | ptlrpcd_add_req(req); |
d7e09d03 | 933 | sent = count; |
d1c0d446 | 934 | goto out; |
d7e09d03 | 935 | } |
c5c4c6fa OW |
936 | |
937 | rc = ptlrpc_queue_wait(req); | |
2d58de78 | 938 | if (rc == LUSTRE_ESTALE) { |
2d00bd17 | 939 | CDEBUG(D_DLMTRACE, "client/server (nid %s) out of sync -- not fatal\n", |
d7e09d03 PT |
940 | libcfs_nid2str(req->rq_import-> |
941 | imp_connection->c_peer.nid)); | |
942 | rc = 0; | |
943 | } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/ | |
944 | req->rq_import_generation == imp->imp_generation) { | |
945 | ptlrpc_req_finished(req); | |
946 | continue; | |
947 | } else if (rc != ELDLM_OK) { | |
948 | /* -ESHUTDOWN is common on umount */ | |
949 | CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR, | |
2d00bd17 JP |
950 | "Got rc %d from cancel RPC: canceling anyway\n", |
951 | rc); | |
d7e09d03 PT |
952 | break; |
953 | } | |
954 | sent = count; | |
955 | break; | |
956 | } | |
957 | ||
958 | ptlrpc_req_finished(req); | |
d7e09d03 PT |
959 | out: |
960 | return sent ? sent : rc; | |
961 | } | |
d7e09d03 PT |
962 | |
963 | static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp) | |
964 | { | |
d7e09d03 PT |
965 | return &imp->imp_obd->obd_namespace->ns_pool; |
966 | } | |
967 | ||
968 | /** | |
969 | * Update client's OBD pool related fields with new SLV and Limit from \a req. | |
970 | */ | |
971 | int ldlm_cli_update_pool(struct ptlrpc_request *req) | |
972 | { | |
973 | struct obd_device *obd; | |
974 | __u64 new_slv; | |
975 | __u32 new_limit; | |
29aaf496 | 976 | |
d7e09d03 | 977 | if (unlikely(!req->rq_import || !req->rq_import->imp_obd || |
9d0b2b7a | 978 | !imp_connect_lru_resize(req->rq_import))) { |
d7e09d03 PT |
979 | /* |
980 | * Do nothing for corner cases. | |
981 | */ | |
0a3bdb00 | 982 | return 0; |
d7e09d03 PT |
983 | } |
984 | ||
985 | /* In some cases RPC may contain SLV and limit zeroed out. This | |
986 | * is the case when server does not support LRU resize feature. | |
987 | * This is also possible in some recovery cases when server-side | |
988 | * reqs have no reference to the OBD export and thus access to | |
6f789a6a OD |
989 | * server-side namespace is not possible. |
990 | */ | |
d7e09d03 PT |
991 | if (lustre_msg_get_slv(req->rq_repmsg) == 0 || |
992 | lustre_msg_get_limit(req->rq_repmsg) == 0) { | |
e7ddc48c AR |
993 | DEBUG_REQ(D_HA, req, |
994 | "Zero SLV or Limit found (SLV: %llu, Limit: %u)", | |
d7e09d03 PT |
995 | lustre_msg_get_slv(req->rq_repmsg), |
996 | lustre_msg_get_limit(req->rq_repmsg)); | |
0a3bdb00 | 997 | return 0; |
d7e09d03 PT |
998 | } |
999 | ||
1000 | new_limit = lustre_msg_get_limit(req->rq_repmsg); | |
1001 | new_slv = lustre_msg_get_slv(req->rq_repmsg); | |
1002 | obd = req->rq_import->imp_obd; | |
1003 | ||
1004 | /* Set new SLV and limit in OBD fields to make them accessible | |
1005 | * to the pool thread. We do not access obd_namespace and pool | |
1006 | * directly here as there is no reliable way to make sure that | |
1007 | * they are still alive at cleanup time. Evil races are possible | |
6f789a6a OD |
1008 | * which may cause Oops at that time. |
1009 | */ | |
d7e09d03 PT |
1010 | write_lock(&obd->obd_pool_lock); |
1011 | obd->obd_pool_slv = new_slv; | |
1012 | obd->obd_pool_limit = new_limit; | |
1013 | write_unlock(&obd->obd_pool_lock); | |
1014 | ||
0a3bdb00 | 1015 | return 0; |
d7e09d03 PT |
1016 | } |
1017 | EXPORT_SYMBOL(ldlm_cli_update_pool); | |
1018 | ||
1019 | /** | |
1020 | * Client side lock cancel. | |
1021 | * | |
1022 | * Lock must not have any readers or writers by this time. | |
1023 | */ | |
e8beaf67 | 1024 | int ldlm_cli_cancel(const struct lustre_handle *lockh, |
f833ee42 | 1025 | enum ldlm_cancel_flags cancel_flags) |
d7e09d03 PT |
1026 | { |
1027 | struct obd_export *exp; | |
1028 | int avail, flags, count = 1; | |
1029 | __u64 rc = 0; | |
1030 | struct ldlm_namespace *ns; | |
1031 | struct ldlm_lock *lock; | |
1032 | LIST_HEAD(cancels); | |
d7e09d03 PT |
1033 | |
1034 | /* concurrent cancels on the same handle can happen */ | |
1035 | lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING); | |
44b53f18 | 1036 | if (!lock) { |
19b2056f | 1037 | LDLM_DEBUG_NOLOCK("lock is already being destroyed"); |
0a3bdb00 | 1038 | return 0; |
d7e09d03 PT |
1039 | } |
1040 | ||
1041 | rc = ldlm_cli_cancel_local(lock); | |
d3a8a4e2 | 1042 | if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) { |
d7e09d03 | 1043 | LDLM_LOCK_RELEASE(lock); |
0a3bdb00 | 1044 | return 0; |
d7e09d03 PT |
1045 | } |
1046 | /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL | |
1047 | * RPC which goes to canceld portal, so we can cancel other LRU locks | |
6f789a6a OD |
1048 | * here and send them all as one LDLM_CANCEL RPC. |
1049 | */ | |
d7e09d03 PT |
1050 | LASSERT(list_empty(&lock->l_bl_ast)); |
1051 | list_add(&lock->l_bl_ast, &cancels); | |
1052 | ||
1053 | exp = lock->l_conn_export; | |
1054 | if (exp_connect_cancelset(exp)) { | |
1055 | avail = ldlm_format_handles_avail(class_exp2cliimp(exp), | |
1056 | &RQF_LDLM_CANCEL, | |
1057 | RCL_CLIENT, 0); | |
1058 | LASSERT(avail > 0); | |
1059 | ||
1060 | ns = ldlm_lock_to_ns(lock); | |
1061 | flags = ns_connect_lru_resize(ns) ? | |
1062 | LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED; | |
1063 | count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1, | |
1064 | LCF_BL_AST, flags); | |
1065 | } | |
1066 | ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags); | |
0a3bdb00 | 1067 | return 0; |
d7e09d03 PT |
1068 | } |
1069 | EXPORT_SYMBOL(ldlm_cli_cancel); | |
1070 | ||
1071 | /** | |
1072 | * Locally cancel up to \a count locks in list \a cancels. | |
1073 | * Return the number of cancelled locks. | |
1074 | */ | |
1075 | int ldlm_cli_cancel_list_local(struct list_head *cancels, int count, | |
f833ee42 | 1076 | enum ldlm_cancel_flags flags) |
d7e09d03 PT |
1077 | { |
1078 | LIST_HEAD(head); | |
1079 | struct ldlm_lock *lock, *next; | |
1080 | int left = 0, bl_ast = 0; | |
1081 | __u64 rc; | |
1082 | ||
1083 | left = count; | |
1084 | list_for_each_entry_safe(lock, next, cancels, l_bl_ast) { | |
1085 | if (left-- == 0) | |
1086 | break; | |
1087 | ||
1088 | if (flags & LCF_LOCAL) { | |
1089 | rc = LDLM_FL_LOCAL_ONLY; | |
1090 | ldlm_lock_cancel(lock); | |
1091 | } else { | |
1092 | rc = ldlm_cli_cancel_local(lock); | |
1093 | } | |
1094 | /* Until we have compound requests and can send LDLM_CANCEL | |
1095 | * requests batched with generic RPCs, we need to send cancels | |
1096 | * with the LDLM_FL_BL_AST flag in a separate RPC from | |
6f789a6a OD |
1097 | * the one being generated now. |
1098 | */ | |
d7e09d03 PT |
1099 | if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) { |
1100 | LDLM_DEBUG(lock, "Cancel lock separately"); | |
1101 | list_del_init(&lock->l_bl_ast); | |
1102 | list_add(&lock->l_bl_ast, &head); | |
1103 | bl_ast++; | |
1104 | continue; | |
1105 | } | |
1106 | if (rc == LDLM_FL_LOCAL_ONLY) { | |
1107 | /* CANCEL RPC should not be sent to server. */ | |
1108 | list_del_init(&lock->l_bl_ast); | |
1109 | LDLM_LOCK_RELEASE(lock); | |
1110 | count--; | |
1111 | } | |
1112 | } | |
1113 | if (bl_ast > 0) { | |
1114 | count -= bl_ast; | |
1115 | ldlm_cli_cancel_list(&head, bl_ast, NULL, 0); | |
1116 | } | |
1117 | ||
0a3bdb00 | 1118 | return count; |
d7e09d03 PT |
1119 | } |
1120 | EXPORT_SYMBOL(ldlm_cli_cancel_list_local); | |
1121 | ||
1122 | /** | |
1123 | * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back | |
1124 | * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g. | |
1125 | * readahead requests, ...) | |
1126 | */ | |
7f2d15bb OD |
1127 | static enum ldlm_policy_res |
1128 | ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, struct ldlm_lock *lock, | |
1129 | int unused, int added, int count) | |
d7e09d03 | 1130 | { |
7f2d15bb | 1131 | enum ldlm_policy_res result = LDLM_POLICY_CANCEL_LOCK; |
902f3bb1 | 1132 | |
d7e09d03 | 1133 | /* don't check added & count since we want to process all locks |
f56b355c JX |
1134 | * from unused list. |
1135 | * It's fine to not take lock to access lock->l_resource since | |
1136 | * the lock has already been granted so it won't change. | |
6f789a6a | 1137 | */ |
d7e09d03 | 1138 | switch (lock->l_resource->lr_type) { |
0e1bbbb0 AR |
1139 | case LDLM_EXTENT: |
1140 | case LDLM_IBITS: | |
2640256e | 1141 | if (ns->ns_cancel && ns->ns_cancel(lock) != 0) |
d7e09d03 | 1142 | break; |
0e1bbbb0 AR |
1143 | default: |
1144 | result = LDLM_POLICY_SKIP_LOCK; | |
f56b355c | 1145 | lock_res_and_lock(lock); |
5a9a80ba | 1146 | ldlm_set_skipped(lock); |
f56b355c | 1147 | unlock_res_and_lock(lock); |
0e1bbbb0 | 1148 | break; |
d7e09d03 PT |
1149 | } |
1150 | ||
0a3bdb00 | 1151 | return result; |
d7e09d03 PT |
1152 | } |
1153 | ||
1154 | /** | |
1155 | * Callback function for LRU-resize policy. Decides whether to keep | |
1156 | * \a lock in LRU for current \a LRU size \a unused, added in current | |
1157 | * scan \a added and number of locks to be preferably canceled \a count. | |
1158 | * | |
1159 | * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning | |
1160 | * | |
1161 | * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU | |
1162 | */ | |
7f2d15bb OD |
1163 | static enum ldlm_policy_res ldlm_cancel_lrur_policy(struct ldlm_namespace *ns, |
1164 | struct ldlm_lock *lock, | |
1165 | int unused, int added, | |
1166 | int count) | |
d7e09d03 | 1167 | { |
a649ad1d | 1168 | unsigned long cur = cfs_time_current(); |
d7e09d03 PT |
1169 | struct ldlm_pool *pl = &ns->ns_pool; |
1170 | __u64 slv, lvf, lv; | |
a649ad1d | 1171 | unsigned long la; |
d7e09d03 PT |
1172 | |
1173 | /* Stop LRU processing when we reach past @count or have checked all | |
6f789a6a OD |
1174 | * locks in LRU. |
1175 | */ | |
d7e09d03 PT |
1176 | if (count && added >= count) |
1177 | return LDLM_POLICY_KEEP_LOCK; | |
1178 | ||
1179 | slv = ldlm_pool_get_slv(pl); | |
1180 | lvf = ldlm_pool_get_lvf(pl); | |
24c198e9 | 1181 | la = cfs_duration_sec(cfs_time_sub(cur, lock->l_last_used)); |
d7e09d03 PT |
1182 | lv = lvf * la * unused; |
1183 | ||
700815d4 | 1184 | /* Inform pool about current CLV to see it via debugfs. */ |
d7e09d03 PT |
1185 | ldlm_pool_set_clv(pl, lv); |
1186 | ||
1187 | /* Stop when SLV is not yet come from server or lv is smaller than | |
6f789a6a OD |
1188 | * it is. |
1189 | */ | |
7d443334 JX |
1190 | if (slv == 0 || lv < slv) |
1191 | return LDLM_POLICY_KEEP_LOCK; | |
1192 | ||
1193 | if (ns->ns_cancel && ns->ns_cancel(lock) == 0) | |
1194 | return LDLM_POLICY_KEEP_LOCK; | |
1195 | ||
1196 | return LDLM_POLICY_CANCEL_LOCK; | |
d7e09d03 PT |
1197 | } |
1198 | ||
1199 | /** | |
700815d4 | 1200 | * Callback function for debugfs used policy. Makes decision whether to keep |
d7e09d03 PT |
1201 | * \a lock in LRU for current \a LRU size \a unused, added in current scan \a |
1202 | * added and number of locks to be preferably canceled \a count. | |
1203 | * | |
1204 | * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning | |
1205 | * | |
1206 | * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU | |
1207 | */ | |
7f2d15bb OD |
1208 | static enum ldlm_policy_res ldlm_cancel_passed_policy(struct ldlm_namespace *ns, |
1209 | struct ldlm_lock *lock, | |
1210 | int unused, int added, | |
1211 | int count) | |
d7e09d03 PT |
1212 | { |
1213 | /* Stop LRU processing when we reach past @count or have checked all | |
6f789a6a OD |
1214 | * locks in LRU. |
1215 | */ | |
d7e09d03 PT |
1216 | return (added >= count) ? |
1217 | LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; | |
1218 | } | |
1219 | ||
1220 | /** | |
1221 | * Callback function for aged policy. Makes decision whether to keep \a lock in | |
1222 | * LRU for current LRU size \a unused, added in current scan \a added and | |
1223 | * number of locks to be preferably canceled \a count. | |
1224 | * | |
1225 | * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning | |
1226 | * | |
1227 | * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU | |
1228 | */ | |
7f2d15bb OD |
1229 | static enum ldlm_policy_res ldlm_cancel_aged_policy(struct ldlm_namespace *ns, |
1230 | struct ldlm_lock *lock, | |
1231 | int unused, int added, | |
1232 | int count) | |
d7e09d03 | 1233 | { |
7b2d26b0 NY |
1234 | if ((added >= count) && |
1235 | time_before(cfs_time_current(), | |
7d443334 JX |
1236 | cfs_time_add(lock->l_last_used, ns->ns_max_age))) |
1237 | return LDLM_POLICY_KEEP_LOCK; | |
1238 | ||
1239 | if (ns->ns_cancel && ns->ns_cancel(lock) == 0) | |
1240 | return LDLM_POLICY_KEEP_LOCK; | |
1241 | ||
1242 | return LDLM_POLICY_CANCEL_LOCK; | |
d7e09d03 PT |
1243 | } |
1244 | ||
7f2d15bb | 1245 | static enum ldlm_policy_res |
2640256e VF |
1246 | ldlm_cancel_lrur_no_wait_policy(struct ldlm_namespace *ns, |
1247 | struct ldlm_lock *lock, | |
1248 | int unused, int added, | |
1249 | int count) | |
1250 | { | |
7f2d15bb | 1251 | enum ldlm_policy_res result; |
2640256e VF |
1252 | |
1253 | result = ldlm_cancel_lrur_policy(ns, lock, unused, added, count); | |
1254 | if (result == LDLM_POLICY_KEEP_LOCK) | |
1255 | return result; | |
1256 | ||
1257 | return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count); | |
1258 | } | |
1259 | ||
d7e09d03 PT |
1260 | /** |
1261 | * Callback function for default policy. Makes decision whether to keep \a lock | |
1262 | * in LRU for current LRU size \a unused, added in current scan \a added and | |
1263 | * number of locks to be preferably canceled \a count. | |
1264 | * | |
1265 | * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning | |
1266 | * | |
1267 | * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU | |
1268 | */ | |
7f2d15bb OD |
1269 | static enum ldlm_policy_res |
1270 | ldlm_cancel_default_policy(struct ldlm_namespace *ns, struct ldlm_lock *lock, | |
1271 | int unused, int added, int count) | |
d7e09d03 PT |
1272 | { |
1273 | /* Stop LRU processing when we reach past count or have checked all | |
6f789a6a OD |
1274 | * locks in LRU. |
1275 | */ | |
d7e09d03 PT |
1276 | return (added >= count) ? |
1277 | LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; | |
1278 | } | |
1279 | ||
7f2d15bb OD |
1280 | typedef enum ldlm_policy_res (*ldlm_cancel_lru_policy_t)( |
1281 | struct ldlm_namespace *, | |
d7e09d03 PT |
1282 | struct ldlm_lock *, int, |
1283 | int, int); | |
1284 | ||
1285 | static ldlm_cancel_lru_policy_t | |
1286 | ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags) | |
1287 | { | |
1288 | if (flags & LDLM_CANCEL_NO_WAIT) | |
1289 | return ldlm_cancel_no_wait_policy; | |
1290 | ||
1291 | if (ns_connect_lru_resize(ns)) { | |
1292 | if (flags & LDLM_CANCEL_SHRINK) | |
1293 | /* We kill passed number of old locks. */ | |
1294 | return ldlm_cancel_passed_policy; | |
1295 | else if (flags & LDLM_CANCEL_LRUR) | |
1296 | return ldlm_cancel_lrur_policy; | |
1297 | else if (flags & LDLM_CANCEL_PASSED) | |
1298 | return ldlm_cancel_passed_policy; | |
2640256e VF |
1299 | else if (flags & LDLM_CANCEL_LRUR_NO_WAIT) |
1300 | return ldlm_cancel_lrur_no_wait_policy; | |
d7e09d03 PT |
1301 | } else { |
1302 | if (flags & LDLM_CANCEL_AGED) | |
1303 | return ldlm_cancel_aged_policy; | |
1304 | } | |
1305 | ||
1306 | return ldlm_cancel_default_policy; | |
1307 | } | |
1308 | ||
1309 | /** | |
1310 | * - Free space in LRU for \a count new locks, | |
1311 | * redundant unused locks are canceled locally; | |
1312 | * - also cancel locally unused aged locks; | |
1313 | * - do not cancel more than \a max locks; | |
1314 | * - GET the found locks and add them into the \a cancels list. | |
1315 | * | |
1316 | * A client lock can be added to the l_bl_ast list only when it is | |
1317 | * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing | |
1318 | * CANCEL. There are the following use cases: | |
1319 | * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and | |
1320 | * ldlm_cli_cancel(), which check and set this flag properly. As any | |
1321 | * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed | |
1322 | * later without any special locking. | |
1323 | * | |
1324 | * Calling policies for enabled LRU resize: | |
1325 | * ---------------------------------------- | |
1326 | * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to | |
1327 | * cancel not more than \a count locks; | |
1328 | * | |
1329 | * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at | |
1330 | * the beginning of LRU list); | |
1331 | * | |
1332 | * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to | |
6e3dd654 | 1333 | * memory pressure policy function; |
d7e09d03 PT |
1334 | * |
1335 | * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy". | |
1336 | * | |
1337 | * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible | |
1338 | * (typically before replaying locks) w/o | |
1339 | * sending any RPCs or waiting for any | |
1340 | * outstanding RPC to complete. | |
1341 | */ | |
e7ddc48c AR |
1342 | static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, |
1343 | struct list_head *cancels, int count, int max, | |
1344 | int flags) | |
d7e09d03 PT |
1345 | { |
1346 | ldlm_cancel_lru_policy_t pf; | |
1347 | struct ldlm_lock *lock, *next; | |
1348 | int added = 0, unused, remained; | |
2640256e | 1349 | int no_wait = flags & (LDLM_CANCEL_NO_WAIT | LDLM_CANCEL_LRUR_NO_WAIT); |
d7e09d03 PT |
1350 | |
1351 | spin_lock(&ns->ns_lock); | |
1352 | unused = ns->ns_nr_unused; | |
1353 | remained = unused; | |
1354 | ||
1355 | if (!ns_connect_lru_resize(ns)) | |
1356 | count += unused - ns->ns_max_unused; | |
1357 | ||
1358 | pf = ldlm_cancel_lru_policy(ns, flags); | |
44b53f18 | 1359 | LASSERT(pf); |
d7e09d03 PT |
1360 | |
1361 | while (!list_empty(&ns->ns_unused_list)) { | |
7f2d15bb | 1362 | enum ldlm_policy_res result; |
6cead36d | 1363 | time_t last_use = 0; |
d7e09d03 PT |
1364 | |
1365 | /* all unused locks */ | |
1366 | if (remained-- <= 0) | |
1367 | break; | |
1368 | ||
1369 | /* For any flags, stop scanning if @max is reached. */ | |
1370 | if (max && added >= max) | |
1371 | break; | |
1372 | ||
1373 | list_for_each_entry_safe(lock, next, &ns->ns_unused_list, | |
24c198e9 | 1374 | l_lru) { |
d7e09d03 | 1375 | /* No locks which got blocking requests. */ |
5a9a80ba | 1376 | LASSERT(!ldlm_is_bl_ast(lock)); |
d7e09d03 | 1377 | |
5a9a80ba | 1378 | if (no_wait && ldlm_is_skipped(lock)) |
d7e09d03 PT |
1379 | /* already processed */ |
1380 | continue; | |
1381 | ||
6cead36d VF |
1382 | last_use = lock->l_last_used; |
1383 | if (last_use == cfs_time_current()) | |
1384 | continue; | |
1385 | ||
d7e09d03 | 1386 | /* Somebody is already doing CANCEL. No need for this |
6f789a6a OD |
1387 | * lock in LRU, do not traverse it again. |
1388 | */ | |
5a9a80ba | 1389 | if (!ldlm_is_canceling(lock)) |
d7e09d03 PT |
1390 | break; |
1391 | ||
1392 | ldlm_lock_remove_from_lru_nolock(lock); | |
1393 | } | |
1394 | if (&lock->l_lru == &ns->ns_unused_list) | |
1395 | break; | |
1396 | ||
1397 | LDLM_LOCK_GET(lock); | |
1398 | spin_unlock(&ns->ns_lock); | |
6756bb7c | 1399 | lu_ref_add(&lock->l_reference, __func__, current); |
d7e09d03 PT |
1400 | |
1401 | /* Pass the lock through the policy filter and see if it | |
1402 | * should stay in LRU. | |
1403 | * | |
1404 | * Even for shrinker policy we stop scanning if | |
1405 | * we find a lock that should stay in the cache. | |
1406 | * We should take into account lock age anyway | |
1407 | * as a new lock is a valuable resource even if | |
1408 | * it has a low weight. | |
1409 | * | |
1410 | * That is, for shrinker policy we drop only | |
1411 | * old locks, but additionally choose them by | |
1412 | * their weight. Big extent locks will stay in | |
1413 | * the cache. */ | |
1414 | result = pf(ns, lock, unused, added, count); | |
1415 | if (result == LDLM_POLICY_KEEP_LOCK) { | |
1416 | lu_ref_del(&lock->l_reference, | |
6756bb7c | 1417 | __func__, current); |
d7e09d03 PT |
1418 | LDLM_LOCK_RELEASE(lock); |
1419 | spin_lock(&ns->ns_lock); | |
1420 | break; | |
1421 | } | |
1422 | if (result == LDLM_POLICY_SKIP_LOCK) { | |
1423 | lu_ref_del(&lock->l_reference, | |
1424 | __func__, current); | |
1425 | LDLM_LOCK_RELEASE(lock); | |
1426 | spin_lock(&ns->ns_lock); | |
1427 | continue; | |
1428 | } | |
1429 | ||
1430 | lock_res_and_lock(lock); | |
1431 | /* Check flags again under the lock. */ | |
5a9a80ba | 1432 | if (ldlm_is_canceling(lock) || |
6cead36d | 1433 | (ldlm_lock_remove_from_lru_check(lock, last_use) == 0)) { |
d7e09d03 PT |
1434 | /* Another thread is removing lock from LRU, or |
1435 | * somebody is already doing CANCEL, or there | |
1436 | * is a blocking request which will send cancel | |
6cead36d VF |
1437 | * by itself, or the lock is no longer unused or |
1438 | * the lock has been used since the pf() call and | |
1439 | * pages could be put under it. | |
6f789a6a | 1440 | */ |
d7e09d03 PT |
1441 | unlock_res_and_lock(lock); |
1442 | lu_ref_del(&lock->l_reference, | |
6756bb7c | 1443 | __func__, current); |
d7e09d03 PT |
1444 | LDLM_LOCK_RELEASE(lock); |
1445 | spin_lock(&ns->ns_lock); | |
1446 | continue; | |
1447 | } | |
1448 | LASSERT(!lock->l_readers && !lock->l_writers); | |
1449 | ||
1450 | /* If we have chosen to cancel this lock voluntarily, we | |
1451 | * better send cancel notification to server, so that it | |
1452 | * frees appropriate state. This might lead to a race | |
1453 | * where while we are doing cancel here, server is also | |
6f789a6a OD |
1454 | * silently cancelling this lock. |
1455 | */ | |
5a9a80ba | 1456 | ldlm_clear_cancel_on_block(lock); |
d7e09d03 PT |
1457 | |
1458 | /* Setting the CBPENDING flag is a little misleading, | |
1459 | * but prevents an important race; namely, once | |
1460 | * CBPENDING is set, the lock can accumulate no more | |
1461 | * readers/writers. Since readers and writers are | |
1462 | * already zero here, ldlm_lock_decref() won't see | |
6f789a6a OD |
1463 | * this flag and call l_blocking_ast |
1464 | */ | |
d7e09d03 PT |
1465 | lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING; |
1466 | ||
1467 | /* We can't re-add to l_lru as it confuses the | |
1468 | * refcounting in ldlm_lock_remove_from_lru() if an AST | |
1469 | * arrives after we drop lr_lock below. We use l_bl_ast | |
1470 | * and can't use l_pending_chain as it is used both on | |
1471 | * server and client nevertheless bug 5666 says it is | |
6f789a6a OD |
1472 | * used only on server |
1473 | */ | |
d7e09d03 PT |
1474 | LASSERT(list_empty(&lock->l_bl_ast)); |
1475 | list_add(&lock->l_bl_ast, cancels); | |
1476 | unlock_res_and_lock(lock); | |
6756bb7c | 1477 | lu_ref_del(&lock->l_reference, __func__, current); |
d7e09d03 PT |
1478 | spin_lock(&ns->ns_lock); |
1479 | added++; | |
1480 | unused--; | |
1481 | } | |
1482 | spin_unlock(&ns->ns_lock); | |
0a3bdb00 | 1483 | return added; |
d7e09d03 PT |
1484 | } |
1485 | ||
58c6d133 OD |
1486 | int ldlm_cancel_lru_local(struct ldlm_namespace *ns, |
1487 | struct list_head *cancels, int count, int max, | |
f833ee42 | 1488 | enum ldlm_cancel_flags cancel_flags, int flags) |
d7e09d03 PT |
1489 | { |
1490 | int added; | |
902f3bb1 | 1491 | |
d7e09d03 PT |
1492 | added = ldlm_prepare_lru_list(ns, cancels, count, max, flags); |
1493 | if (added <= 0) | |
1494 | return added; | |
1495 | return ldlm_cli_cancel_list_local(cancels, added, cancel_flags); | |
1496 | } | |
1497 | ||
1498 | /** | |
1499 | * Cancel at least \a nr locks from given namespace LRU. | |
1500 | * | |
1501 | * When called with LCF_ASYNC the blocking callback will be handled | |
1502 | * in a thread and this function will return after the thread has been | |
1503 | * asked to call the callback. When called with LCF_ASYNC the blocking | |
1504 | * callback will be performed in this function. | |
1505 | */ | |
1506 | int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, | |
f833ee42 | 1507 | enum ldlm_cancel_flags cancel_flags, |
d7e09d03 PT |
1508 | int flags) |
1509 | { | |
1510 | LIST_HEAD(cancels); | |
1511 | int count, rc; | |
d7e09d03 PT |
1512 | |
1513 | /* Just prepare the list of locks, do not actually cancel them yet. | |
6f789a6a OD |
1514 | * Locks are cancelled later in a separate thread. |
1515 | */ | |
d7e09d03 PT |
1516 | count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags); |
1517 | rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags); | |
1518 | if (rc == 0) | |
0a3bdb00 | 1519 | return count; |
d7e09d03 | 1520 | |
0a3bdb00 | 1521 | return 0; |
d7e09d03 PT |
1522 | } |
1523 | ||
1524 | /** | |
1525 | * Find and cancel locally unused locks found on resource, matched to the | |
1526 | * given policy, mode. GET the found locks and add them into the \a cancels | |
1527 | * list. | |
1528 | */ | |
1529 | int ldlm_cancel_resource_local(struct ldlm_resource *res, | |
1530 | struct list_head *cancels, | |
1531 | ldlm_policy_data_t *policy, | |
52ee0d20 | 1532 | enum ldlm_mode mode, __u64 lock_flags, |
f833ee42 OD |
1533 | enum ldlm_cancel_flags cancel_flags, |
1534 | void *opaque) | |
d7e09d03 PT |
1535 | { |
1536 | struct ldlm_lock *lock; | |
1537 | int count = 0; | |
d7e09d03 PT |
1538 | |
1539 | lock_res(res); | |
1540 | list_for_each_entry(lock, &res->lr_granted, l_res_link) { | |
44b53f18 | 1541 | if (opaque && lock->l_ast_data != opaque) { |
d7e09d03 PT |
1542 | LDLM_ERROR(lock, "data %p doesn't match opaque %p", |
1543 | lock->l_ast_data, opaque); | |
d7e09d03 PT |
1544 | continue; |
1545 | } | |
1546 | ||
1547 | if (lock->l_readers || lock->l_writers) | |
1548 | continue; | |
1549 | ||
1550 | /* If somebody is already doing CANCEL, or blocking AST came, | |
6f789a6a OD |
1551 | * skip this lock. |
1552 | */ | |
5a9a80ba | 1553 | if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock)) |
d7e09d03 PT |
1554 | continue; |
1555 | ||
1556 | if (lockmode_compat(lock->l_granted_mode, mode)) | |
1557 | continue; | |
1558 | ||
1559 | /* If policy is given and this is IBITS lock, add to list only | |
6f789a6a OD |
1560 | * those locks that match by policy. |
1561 | */ | |
d7e09d03 PT |
1562 | if (policy && (lock->l_resource->lr_type == LDLM_IBITS) && |
1563 | !(lock->l_policy_data.l_inodebits.bits & | |
1564 | policy->l_inodebits.bits)) | |
1565 | continue; | |
1566 | ||
1567 | /* See CBPENDING comment in ldlm_cancel_lru */ | |
1568 | lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING | | |
1569 | lock_flags; | |
1570 | ||
1571 | LASSERT(list_empty(&lock->l_bl_ast)); | |
1572 | list_add(&lock->l_bl_ast, cancels); | |
1573 | LDLM_LOCK_GET(lock); | |
1574 | count++; | |
1575 | } | |
1576 | unlock_res(res); | |
1577 | ||
0a3bdb00 | 1578 | return ldlm_cli_cancel_list_local(cancels, count, cancel_flags); |
d7e09d03 PT |
1579 | } |
1580 | EXPORT_SYMBOL(ldlm_cancel_resource_local); | |
1581 | ||
1582 | /** | |
1583 | * Cancel client-side locks from a list and send/prepare cancel RPCs to the | |
1584 | * server. | |
1585 | * If \a req is NULL, send CANCEL request to server with handles of locks | |
1586 | * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests | |
1587 | * separately per lock. | |
1588 | * If \a req is not NULL, put handles of locks in \a cancels into the request | |
1589 | * buffer at the offset \a off. | |
1590 | * Destroy \a cancels at the end. | |
1591 | */ | |
1592 | int ldlm_cli_cancel_list(struct list_head *cancels, int count, | |
f833ee42 OD |
1593 | struct ptlrpc_request *req, |
1594 | enum ldlm_cancel_flags flags) | |
d7e09d03 PT |
1595 | { |
1596 | struct ldlm_lock *lock; | |
1597 | int res = 0; | |
d7e09d03 PT |
1598 | |
1599 | if (list_empty(cancels) || count == 0) | |
0a3bdb00 | 1600 | return 0; |
d7e09d03 PT |
1601 | |
1602 | /* XXX: requests (both batched and not) could be sent in parallel. | |
1603 | * Usually it is enough to have just 1 RPC, but it is possible that | |
1604 | * there are too many locks to be cancelled in LRU or on a resource. | |
1605 | * It would also speed up the case when the server does not support | |
6f789a6a OD |
1606 | * the feature. |
1607 | */ | |
d7e09d03 PT |
1608 | while (count > 0) { |
1609 | LASSERT(!list_empty(cancels)); | |
24c198e9 | 1610 | lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast); |
d7e09d03 PT |
1611 | LASSERT(lock->l_conn_export); |
1612 | ||
1613 | if (exp_connect_cancelset(lock->l_conn_export)) { | |
1614 | res = count; | |
1615 | if (req) | |
1616 | ldlm_cancel_pack(req, cancels, count); | |
1617 | else | |
1618 | res = ldlm_cli_cancel_req(lock->l_conn_export, | |
1619 | cancels, count, | |
1620 | flags); | |
1621 | } else { | |
1622 | res = ldlm_cli_cancel_req(lock->l_conn_export, | |
1623 | cancels, 1, flags); | |
1624 | } | |
1625 | ||
1626 | if (res < 0) { | |
1627 | CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR, | |
1628 | "ldlm_cli_cancel_list: %d\n", res); | |
1629 | res = count; | |
1630 | } | |
1631 | ||
1632 | count -= res; | |
1633 | ldlm_lock_list_put(cancels, l_bl_ast, res); | |
1634 | } | |
1635 | LASSERT(count == 0); | |
0a3bdb00 | 1636 | return 0; |
d7e09d03 PT |
1637 | } |
1638 | EXPORT_SYMBOL(ldlm_cli_cancel_list); | |
1639 | ||
1640 | /** | |
1641 | * Cancel all locks on a resource that have 0 readers/writers. | |
1642 | * | |
1643 | * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying | |
6f789a6a OD |
1644 | * to notify the server. |
1645 | */ | |
d7e09d03 PT |
1646 | int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns, |
1647 | const struct ldlm_res_id *res_id, | |
1648 | ldlm_policy_data_t *policy, | |
52ee0d20 | 1649 | enum ldlm_mode mode, |
f833ee42 | 1650 | enum ldlm_cancel_flags flags, |
d7e09d03 PT |
1651 | void *opaque) |
1652 | { | |
1653 | struct ldlm_resource *res; | |
1654 | LIST_HEAD(cancels); | |
1655 | int count; | |
1656 | int rc; | |
d7e09d03 PT |
1657 | |
1658 | res = ldlm_resource_get(ns, NULL, res_id, 0, 0); | |
099d5adf | 1659 | if (IS_ERR(res)) { |
d7e09d03 | 1660 | /* This is not a problem. */ |
b0f5aad5 | 1661 | CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]); |
0a3bdb00 | 1662 | return 0; |
d7e09d03 PT |
1663 | } |
1664 | ||
1665 | LDLM_RESOURCE_ADDREF(res); | |
1666 | count = ldlm_cancel_resource_local(res, &cancels, policy, mode, | |
1667 | 0, flags | LCF_BL_AST, opaque); | |
1668 | rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags); | |
1669 | if (rc != ELDLM_OK) | |
6d95e048 AD |
1670 | CERROR("canceling unused lock "DLDLMRES": rc = %d\n", |
1671 | PLDLMRES(res), rc); | |
d7e09d03 PT |
1672 | |
1673 | LDLM_RESOURCE_DELREF(res); | |
1674 | ldlm_resource_putref(res); | |
0a3bdb00 | 1675 | return 0; |
d7e09d03 PT |
1676 | } |
1677 | EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource); | |
1678 | ||
1679 | struct ldlm_cli_cancel_arg { | |
1680 | int lc_flags; | |
1681 | void *lc_opaque; | |
1682 | }; | |
1683 | ||
e7ddc48c AR |
1684 | static int ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, |
1685 | struct cfs_hash_bd *bd, | |
d7e09d03 PT |
1686 | struct hlist_node *hnode, void *arg) |
1687 | { | |
1688 | struct ldlm_resource *res = cfs_hash_object(hs, hnode); | |
1689 | struct ldlm_cli_cancel_arg *lc = arg; | |
6d95e048 AD |
1690 | |
1691 | ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name, | |
1692 | NULL, LCK_MINMODE, | |
1693 | lc->lc_flags, lc->lc_opaque); | |
d7e09d03 PT |
1694 | /* must return 0 for hash iteration */ |
1695 | return 0; | |
1696 | } | |
1697 | ||
1698 | /** | |
1699 | * Cancel all locks on a namespace (or a specific resource, if given) | |
1700 | * that have 0 readers/writers. | |
1701 | * | |
1702 | * If flags & LCF_LOCAL, throw the locks away without trying | |
1703 | * to notify the server. */ | |
1704 | int ldlm_cli_cancel_unused(struct ldlm_namespace *ns, | |
1705 | const struct ldlm_res_id *res_id, | |
f833ee42 | 1706 | enum ldlm_cancel_flags flags, void *opaque) |
d7e09d03 PT |
1707 | { |
1708 | struct ldlm_cli_cancel_arg arg = { | |
1709 | .lc_flags = flags, | |
1710 | .lc_opaque = opaque, | |
1711 | }; | |
1712 | ||
44b53f18 | 1713 | if (!ns) |
0a3bdb00 | 1714 | return ELDLM_OK; |
d7e09d03 | 1715 | |
44b53f18 | 1716 | if (res_id) { |
0a3bdb00 | 1717 | return ldlm_cli_cancel_unused_resource(ns, res_id, NULL, |
d7e09d03 | 1718 | LCK_MINMODE, flags, |
0a3bdb00 | 1719 | opaque); |
d7e09d03 PT |
1720 | } else { |
1721 | cfs_hash_for_each_nolock(ns->ns_rs_hash, | |
1722 | ldlm_cli_hash_cancel_unused, &arg); | |
0a3bdb00 | 1723 | return ELDLM_OK; |
d7e09d03 PT |
1724 | } |
1725 | } | |
1726 | EXPORT_SYMBOL(ldlm_cli_cancel_unused); | |
1727 | ||
1728 | /* Lock iterators. */ | |
1729 | ||
58c6d133 OD |
1730 | static int ldlm_resource_foreach(struct ldlm_resource *res, |
1731 | ldlm_iterator_t iter, void *closure) | |
d7e09d03 PT |
1732 | { |
1733 | struct list_head *tmp, *next; | |
1734 | struct ldlm_lock *lock; | |
1735 | int rc = LDLM_ITER_CONTINUE; | |
1736 | ||
d7e09d03 | 1737 | if (!res) |
0a3bdb00 | 1738 | return LDLM_ITER_CONTINUE; |
d7e09d03 PT |
1739 | |
1740 | lock_res(res); | |
1741 | list_for_each_safe(tmp, next, &res->lr_granted) { | |
1742 | lock = list_entry(tmp, struct ldlm_lock, l_res_link); | |
1743 | ||
d1c0d446 JL |
1744 | if (iter(lock, closure) == LDLM_ITER_STOP) { |
1745 | rc = LDLM_ITER_STOP; | |
1746 | goto out; | |
1747 | } | |
d7e09d03 PT |
1748 | } |
1749 | ||
d7e09d03 PT |
1750 | list_for_each_safe(tmp, next, &res->lr_waiting) { |
1751 | lock = list_entry(tmp, struct ldlm_lock, l_res_link); | |
1752 | ||
d1c0d446 JL |
1753 | if (iter(lock, closure) == LDLM_ITER_STOP) { |
1754 | rc = LDLM_ITER_STOP; | |
1755 | goto out; | |
1756 | } | |
d7e09d03 PT |
1757 | } |
1758 | out: | |
1759 | unlock_res(res); | |
0a3bdb00 | 1760 | return rc; |
d7e09d03 | 1761 | } |
d7e09d03 PT |
1762 | |
1763 | struct iter_helper_data { | |
1764 | ldlm_iterator_t iter; | |
1765 | void *closure; | |
1766 | }; | |
1767 | ||
1768 | static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure) | |
1769 | { | |
1770 | struct iter_helper_data *helper = closure; | |
902f3bb1 | 1771 | |
d7e09d03 PT |
1772 | return helper->iter(lock, helper->closure); |
1773 | } | |
1774 | ||
6da6eabe | 1775 | static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd, |
d7e09d03 PT |
1776 | struct hlist_node *hnode, void *arg) |
1777 | ||
1778 | { | |
1779 | struct ldlm_resource *res = cfs_hash_object(hs, hnode); | |
1780 | ||
1781 | return ldlm_resource_foreach(res, ldlm_iter_helper, arg) == | |
1782 | LDLM_ITER_STOP; | |
1783 | } | |
1784 | ||
58c6d133 OD |
1785 | static void ldlm_namespace_foreach(struct ldlm_namespace *ns, |
1786 | ldlm_iterator_t iter, void *closure) | |
d7e09d03 PT |
1787 | |
1788 | { | |
805e517a EG |
1789 | struct iter_helper_data helper = { |
1790 | .iter = iter, | |
1791 | .closure = closure, | |
1792 | }; | |
d7e09d03 PT |
1793 | |
1794 | cfs_hash_for_each_nolock(ns->ns_rs_hash, | |
1795 | ldlm_res_iter_helper, &helper); | |
d7e09d03 | 1796 | } |
d7e09d03 PT |
1797 | |
1798 | /* non-blocking function to manipulate a lock whose cb_data is being put away. | |
1799 | * return 0: find no resource | |
1800 | * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE. | |
1801 | * < 0: errors | |
1802 | */ | |
1803 | int ldlm_resource_iterate(struct ldlm_namespace *ns, | |
1804 | const struct ldlm_res_id *res_id, | |
1805 | ldlm_iterator_t iter, void *data) | |
1806 | { | |
1807 | struct ldlm_resource *res; | |
1808 | int rc; | |
d7e09d03 | 1809 | |
099d5adf | 1810 | LASSERTF(ns, "must pass in namespace\n"); |
d7e09d03 PT |
1811 | |
1812 | res = ldlm_resource_get(ns, NULL, res_id, 0, 0); | |
099d5adf | 1813 | if (IS_ERR(res)) |
0a3bdb00 | 1814 | return 0; |
d7e09d03 PT |
1815 | |
1816 | LDLM_RESOURCE_ADDREF(res); | |
1817 | rc = ldlm_resource_foreach(res, iter, data); | |
1818 | LDLM_RESOURCE_DELREF(res); | |
1819 | ldlm_resource_putref(res); | |
0a3bdb00 | 1820 | return rc; |
d7e09d03 PT |
1821 | } |
1822 | EXPORT_SYMBOL(ldlm_resource_iterate); | |
1823 | ||
1824 | /* Lock replay */ | |
1825 | ||
1826 | static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure) | |
1827 | { | |
1828 | struct list_head *list = closure; | |
1829 | ||
1830 | /* we use l_pending_chain here, because it's unused on clients. */ | |
1831 | LASSERTF(list_empty(&lock->l_pending_chain), | |
1832 | "lock %p next %p prev %p\n", | |
43ee4160 AR |
1833 | lock, &lock->l_pending_chain.next, |
1834 | &lock->l_pending_chain.prev); | |
d7e09d03 PT |
1835 | /* bug 9573: don't replay locks left after eviction, or |
1836 | * bug 17614: locks being actively cancelled. Get a reference | |
6e3dd654 | 1837 | * on a lock so that it does not disappear under us (e.g. due to cancel) |
d7e09d03 | 1838 | */ |
cd94f231 | 1839 | if (!(lock->l_flags & (LDLM_FL_FAILED | LDLM_FL_CANCELING))) { |
d7e09d03 PT |
1840 | list_add(&lock->l_pending_chain, list); |
1841 | LDLM_LOCK_GET(lock); | |
1842 | } | |
1843 | ||
1844 | return LDLM_ITER_CONTINUE; | |
1845 | } | |
1846 | ||
1847 | static int replay_lock_interpret(const struct lu_env *env, | |
1848 | struct ptlrpc_request *req, | |
1849 | struct ldlm_async_args *aa, int rc) | |
1850 | { | |
1851 | struct ldlm_lock *lock; | |
1852 | struct ldlm_reply *reply; | |
1853 | struct obd_export *exp; | |
1854 | ||
d7e09d03 PT |
1855 | atomic_dec(&req->rq_import->imp_replay_inflight); |
1856 | if (rc != ELDLM_OK) | |
d1c0d446 | 1857 | goto out; |
d7e09d03 | 1858 | |
d7e09d03 | 1859 | reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); |
44b53f18 | 1860 | if (!reply) { |
d1c0d446 JL |
1861 | rc = -EPROTO; |
1862 | goto out; | |
1863 | } | |
d7e09d03 PT |
1864 | |
1865 | lock = ldlm_handle2lock(&aa->lock_handle); | |
1866 | if (!lock) { | |
2d00bd17 | 1867 | CERROR("received replay ack for unknown local cookie %#llx remote cookie %#llx from server %s id %s\n", |
d7e09d03 PT |
1868 | aa->lock_handle.cookie, reply->lock_handle.cookie, |
1869 | req->rq_export->exp_client_uuid.uuid, | |
1870 | libcfs_id2str(req->rq_peer)); | |
d1c0d446 JL |
1871 | rc = -ESTALE; |
1872 | goto out; | |
d7e09d03 PT |
1873 | } |
1874 | ||
1875 | /* Key change rehash lock in per-export hash with new key */ | |
1876 | exp = req->rq_export; | |
1877 | if (exp && exp->exp_lock_hash) { | |
1878 | /* In the function below, .hs_keycmp resolves to | |
6f789a6a OD |
1879 | * ldlm_export_lock_keycmp() |
1880 | */ | |
d7e09d03 PT |
1881 | /* coverity[overrun-buffer-val] */ |
1882 | cfs_hash_rehash_key(exp->exp_lock_hash, | |
1883 | &lock->l_remote_handle, | |
1884 | &reply->lock_handle, | |
1885 | &lock->l_exp_hash); | |
1886 | } else { | |
1887 | lock->l_remote_handle = reply->lock_handle; | |
1888 | } | |
1889 | ||
1890 | LDLM_DEBUG(lock, "replayed lock:"); | |
1891 | ptlrpc_import_recovery_state_machine(req->rq_import); | |
1892 | LDLM_LOCK_PUT(lock); | |
1893 | out: | |
1894 | if (rc != ELDLM_OK) | |
1895 | ptlrpc_connect_import(req->rq_import); | |
1896 | ||
0a3bdb00 | 1897 | return rc; |
d7e09d03 PT |
1898 | } |
1899 | ||
1900 | static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) | |
1901 | { | |
1902 | struct ptlrpc_request *req; | |
1903 | struct ldlm_async_args *aa; | |
1904 | struct ldlm_request *body; | |
1905 | int flags; | |
d7e09d03 PT |
1906 | |
1907 | /* Bug 11974: Do not replay a lock which is actively being canceled */ | |
5a9a80ba | 1908 | if (ldlm_is_canceling(lock)) { |
d7e09d03 | 1909 | LDLM_DEBUG(lock, "Not replaying canceled lock:"); |
0a3bdb00 | 1910 | return 0; |
d7e09d03 PT |
1911 | } |
1912 | ||
1913 | /* If this is reply-less callback lock, we cannot replay it, since | |
1914 | * server might have long dropped it, but notification of that event was | |
6f789a6a OD |
1915 | * lost by network. (and server granted conflicting lock already) |
1916 | */ | |
5a9a80ba | 1917 | if (ldlm_is_cancel_on_block(lock)) { |
d7e09d03 PT |
1918 | LDLM_DEBUG(lock, "Not replaying reply-less lock:"); |
1919 | ldlm_lock_cancel(lock); | |
0a3bdb00 | 1920 | return 0; |
d7e09d03 PT |
1921 | } |
1922 | ||
1923 | /* | |
1924 | * If granted mode matches the requested mode, this lock is granted. | |
1925 | * | |
1926 | * If they differ, but we have a granted mode, then we were granted | |
1927 | * one mode and now want another: ergo, converting. | |
1928 | * | |
1929 | * If we haven't been granted anything and are on a resource list, | |
1930 | * then we're blocked/waiting. | |
1931 | * | |
1932 | * If we haven't been granted anything and we're NOT on a resource list, | |
1933 | * then we haven't got a reply yet and don't have a known disposition. | |
1934 | * This happens whenever a lock enqueue is the request that triggers | |
1935 | * recovery. | |
1936 | */ | |
1937 | if (lock->l_granted_mode == lock->l_req_mode) | |
1938 | flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED; | |
1939 | else if (lock->l_granted_mode) | |
1940 | flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV; | |
1941 | else if (!list_empty(&lock->l_res_link)) | |
1942 | flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT; | |
1943 | else | |
1944 | flags = LDLM_FL_REPLAY; | |
1945 | ||
1946 | req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE, | |
1947 | LUSTRE_DLM_VERSION, LDLM_ENQUEUE); | |
44b53f18 | 1948 | if (!req) |
0a3bdb00 | 1949 | return -ENOMEM; |
d7e09d03 PT |
1950 | |
1951 | /* We're part of recovery, so don't wait for it. */ | |
1952 | req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS; | |
1953 | ||
1954 | body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); | |
1955 | ldlm_lock2desc(lock, &body->lock_desc); | |
1956 | body->lock_flags = ldlm_flags_to_wire(flags); | |
1957 | ||
1958 | ldlm_lock2handle(lock, &body->lock_handle[0]); | |
1959 | if (lock->l_lvb_len > 0) | |
1960 | req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB); | |
1961 | req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, | |
1962 | lock->l_lvb_len); | |
1963 | ptlrpc_request_set_replen(req); | |
1964 | /* notify the server we've replayed all requests. | |
1965 | * also, we mark the request to be put on a dedicated | |
1966 | * queue to be processed after all request replayes. | |
6f789a6a OD |
1967 | * bug 6063 |
1968 | */ | |
d7e09d03 PT |
1969 | lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE); |
1970 | ||
1971 | LDLM_DEBUG(lock, "replaying lock:"); | |
1972 | ||
1973 | atomic_inc(&req->rq_import->imp_replay_inflight); | |
1974 | CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args)); | |
1975 | aa = ptlrpc_req_async_args(req); | |
1976 | aa->lock_handle = body->lock_handle[0]; | |
1977 | req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret; | |
c5c4c6fa | 1978 | ptlrpcd_add_req(req); |
d7e09d03 | 1979 | |
0a3bdb00 | 1980 | return 0; |
d7e09d03 PT |
1981 | } |
1982 | ||
1983 | /** | |
1984 | * Cancel as many unused locks as possible before replay. since we are | |
1985 | * in recovery, we can't wait for any outstanding RPCs to send any RPC | |
1986 | * to the server. | |
1987 | * | |
1988 | * Called only in recovery before replaying locks. there is no need to | |
1989 | * replay locks that are unused. since the clients may hold thousands of | |
1990 | * cached unused locks, dropping the unused locks can greatly reduce the | |
1991 | * load on the servers at recovery time. | |
1992 | */ | |
1993 | static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns) | |
1994 | { | |
1995 | int canceled; | |
1996 | LIST_HEAD(cancels); | |
1997 | ||
2d00bd17 JP |
1998 | CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before replay for namespace %s (%d)\n", |
1999 | ldlm_ns_name(ns), ns->ns_nr_unused); | |
d7e09d03 PT |
2000 | |
2001 | /* We don't need to care whether or not LRU resize is enabled | |
2002 | * because the LDLM_CANCEL_NO_WAIT policy doesn't use the | |
6f789a6a OD |
2003 | * count parameter |
2004 | */ | |
d7e09d03 PT |
2005 | canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0, |
2006 | LCF_LOCAL, LDLM_CANCEL_NO_WAIT); | |
2007 | ||
2008 | CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n", | |
24c198e9 | 2009 | canceled, ldlm_ns_name(ns)); |
d7e09d03 PT |
2010 | } |
2011 | ||
2012 | int ldlm_replay_locks(struct obd_import *imp) | |
2013 | { | |
2014 | struct ldlm_namespace *ns = imp->imp_obd->obd_namespace; | |
2015 | LIST_HEAD(list); | |
2016 | struct ldlm_lock *lock, *next; | |
2017 | int rc = 0; | |
2018 | ||
d7e09d03 PT |
2019 | LASSERT(atomic_read(&imp->imp_replay_inflight) == 0); |
2020 | ||
2021 | /* don't replay locks if import failed recovery */ | |
2022 | if (imp->imp_vbr_failed) | |
0a3bdb00 | 2023 | return 0; |
d7e09d03 PT |
2024 | |
2025 | /* ensure this doesn't fall to 0 before all have been queued */ | |
2026 | atomic_inc(&imp->imp_replay_inflight); | |
2027 | ||
2028 | if (ldlm_cancel_unused_locks_before_replay) | |
2029 | ldlm_cancel_unused_locks_for_replay(ns); | |
2030 | ||
2031 | ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list); | |
2032 | ||
2033 | list_for_each_entry_safe(lock, next, &list, l_pending_chain) { | |
2034 | list_del_init(&lock->l_pending_chain); | |
2035 | if (rc) { | |
2036 | LDLM_LOCK_RELEASE(lock); | |
2037 | continue; /* or try to do the rest? */ | |
2038 | } | |
2039 | rc = replay_one_lock(imp, lock); | |
2040 | LDLM_LOCK_RELEASE(lock); | |
2041 | } | |
2042 | ||
2043 | atomic_dec(&imp->imp_replay_inflight); | |
2044 | ||
0a3bdb00 | 2045 | return rc; |
d7e09d03 PT |
2046 | } |
2047 | EXPORT_SYMBOL(ldlm_replay_locks); |