1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Phil Schwan <phil@clusterfs.com>
8 * This file is part of Lustre, http://www.lustre.org.
10 * Lustre is free software; you can redistribute it and/or
11 * modify it under the terms of version 2 of the GNU General Public
12 * License as published by the Free Software Foundation.
14 * Lustre is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with Lustre; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 #define DEBUG_SUBSYSTEM S_LDLM
26 # include <liblustre.h>
29 #include <linux/lustre_dlm.h>
30 #include <linux/obd_support.h>
31 #include <linux/lustre_lib.h>
33 #include "ldlm_internal.h"
35 /* The purpose of this function is to return:
36 * - the maximum extent
37 * - containing the requested extent
38 * - and not overlapping existing conflicting extents outside the requested one
41 ldlm_extent_internal_policy(struct list_head *queue, struct ldlm_lock *req,
42 struct ldlm_extent *new_ex)
44 struct list_head *tmp;
45 ldlm_mode_t req_mode = req->l_req_mode;
46 __u64 req_start = req->l_req_extent.start;
47 __u64 req_end = req->l_req_extent.end;
51 lockmode_verify(req_mode);
53 list_for_each(tmp, queue) {
54 struct ldlm_lock *lock;
55 struct ldlm_extent *l_extent;
57 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
58 l_extent = &lock->l_policy_data.l_extent;
60 if (new_ex->start == req_start && new_ex->end == req_end) {
65 /* Don't conflict with ourselves */
69 /* Locks are compatible, overlap doesn't matter */
70 /* Until bug 20 is fixed, try to avoid granting overlapping
71 * locks on one client (they take a long time to cancel) */
72 if (lockmode_compat(lock->l_req_mode, req_mode) &&
73 lock->l_export != req->l_export)
76 /* If this is a high-traffic lock, don't grow downwards at all
77 * or grow upwards too much */
80 new_ex->start = req_start;
82 /* If lock doesn't overlap new_ex, skip it. */
83 if (l_extent->end < new_ex->start ||
84 l_extent->start > new_ex->end)
87 /* Locks conflicting in requested extents and we can't satisfy
88 * both locks, so ignore it. Either we will ping-pong this
89 * extent (we would regardless of what extent we granted) or
90 * lock is unused and it shouldn't limit our extent growth. */
91 if (lock->l_req_extent.end >= req_start &&
92 lock->l_req_extent.start <= req_end)
95 /* We grow extents downwards only as far as they don't overlap
96 * with already-granted locks, on the assumtion that clients
97 * will be writing beyond the initial requested end and would
98 * then need to enqueue a new lock beyond previous request.
99 * l_req_extent->end strictly < req_start, checked above. */
100 if (l_extent->start < req_start && new_ex->start != req_start) {
101 if (l_extent->end >= req_start)
102 new_ex->start = req_start;
104 new_ex->start = min(l_extent->end+1, req_start);
107 /* If we need to cancel this lock anyways because our request
108 * overlaps the granted lock, we grow up to its requested
109 * extent start instead of limiting this extent, assuming that
110 * clients are writing forwards and the lock had over grown
111 * its extent downwards before we enqueued our request. */
112 if (l_extent->end > req_end) {
113 if (l_extent->start <= req_end)
114 new_ex->end = max(lock->l_req_extent.start - 1,
117 new_ex->end = max(l_extent->start - 1, req_end);
121 #define LDLM_MAX_GROWN_EXTENT (32 * 1024 * 1024 - 1)
122 if (conflicting > 32 && (req_mode == LCK_PW || req_mode == LCK_CW)) {
123 if (req_end < req_start + LDLM_MAX_GROWN_EXTENT)
124 new_ex->end = min(req_start + LDLM_MAX_GROWN_EXTENT,
130 /* In order to determine the largest possible extent we can grant, we need
131 * to scan all of the queues. */
132 static void ldlm_extent_policy(struct ldlm_resource *res,
133 struct ldlm_lock *lock, int *flags)
135 struct ldlm_extent new_ex = { .start = 0, .end = ~0};
137 ldlm_extent_internal_policy(&res->lr_granted, lock, &new_ex);
138 ldlm_extent_internal_policy(&res->lr_waiting, lock, &new_ex);
140 if (new_ex.start != lock->l_policy_data.l_extent.start ||
141 new_ex.end != lock->l_policy_data.l_extent.end) {
142 *flags |= LDLM_FL_LOCK_CHANGED;
143 lock->l_policy_data.l_extent.start = new_ex.start;
144 lock->l_policy_data.l_extent.end = new_ex.end;
148 /* Determine if the lock is compatible with all locks on the queue.
149 * We stop walking the queue if we hit ourselves so we don't take
150 * conflicting locks enqueued after us into accound, or we'd wait forever.
152 * 0 if the lock is not compatible
153 * 1 if the lock is compatible
154 * 2 if this group lock is compatible and requires no further checking
155 * negative error, such as EWOULDBLOCK for group locks
158 ldlm_extent_compat_queue(struct list_head *queue, struct ldlm_lock *req,
159 int send_cbs, int *flags, ldlm_error_t *err)
161 struct list_head *tmp;
162 struct ldlm_lock *lock;
163 ldlm_mode_t req_mode = req->l_req_mode;
164 __u64 req_start = req->l_req_extent.start;
165 __u64 req_end = req->l_req_extent.end;
169 lockmode_verify(req_mode);
171 list_for_each(tmp, queue) {
172 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
177 /* locks are compatible, overlap doesn't matter */
178 if (lockmode_compat(lock->l_req_mode, req_mode)) {
179 /* non-group locks are compatible, overlap doesn't
181 if (req_mode != LCK_GROUP)
184 /* If we are trying to get a GROUP lock and there is
185 another one of this kind, we need to compare gid */
186 if (req->l_policy_data.l_extent.gid ==
187 lock->l_policy_data.l_extent.gid) {
188 if (lock->l_req_mode == lock->l_granted_mode)
191 /* If we are in nonblocking mode - return
193 if (*flags & LDLM_FL_BLOCK_NOWAIT) {
194 compat = -EWOULDBLOCK;
197 /* If this group lock is compatible with another
198 * group lock on the waiting list, they must be
199 * together in the list, so they can be granted
200 * at the same time. Otherwise the later lock
201 * can get stuck behind another, incompatible,
203 ldlm_resource_insert_lock_after(lock, req);
204 /* Because 'lock' is not granted, we can stop
205 * processing this queue and return immediately.
206 * There is no need to check the rest of the
212 if (lock->l_req_mode == LCK_GROUP) {
213 /* If compared lock is GROUP, then requested is PR/PW/=>
214 * this is not compatible; extent range does not
216 if (*flags & LDLM_FL_BLOCK_NOWAIT) {
217 compat = -EWOULDBLOCK;
220 *flags |= LDLM_FL_NO_TIMEOUT;
222 } else if (lock->l_policy_data.l_extent.end < req_start ||
223 lock->l_policy_data.l_extent.start > req_end) {
224 /* if a non grouplock doesn't overlap skip it */
232 if (lock->l_blocking_ast)
233 ldlm_add_ast_work_item(lock, req, NULL, 0);
238 list_del_init(&req->l_res_link);
239 ldlm_lock_destroy(req);
244 /* If first_enq is 0 (ie, called from ldlm_reprocess_queue):
245 * - blocking ASTs have already been sent
246 * - the caller has already initialized req->lr_tmp
247 * - must call this function with the ns lock held
249 * If first_enq is 1 (ie, called from ldlm_lock_enqueue):
250 * - blocking ASTs have not been sent
251 * - the caller has NOT initialized req->lr_tmp, so we must
252 * - must call this function with the ns lock held once */
253 int ldlm_process_extent_lock(struct ldlm_lock *lock, int *flags, int first_enq,
256 struct ldlm_resource *res = lock->l_resource;
257 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
261 LASSERT(list_empty(&res->lr_converting));
265 /* Careful observers will note that we don't handle -EWOULDBLOCK
266 * here, but it's ok for a non-obvious reason -- compat_queue
267 * can only return -EWOULDBLOCK if (flags & BLOCK_NOWAIT).
268 * flags should always be zero here, and if that ever stops
269 * being true, we want to find out. */
270 LASSERT(*flags == 0);
271 LASSERT(res->lr_tmp != NULL);
272 rc = ldlm_extent_compat_queue(&res->lr_granted, lock, 0, flags,
275 rc = ldlm_extent_compat_queue(&res->lr_waiting, lock, 0,
279 RETURN(LDLM_ITER_STOP);
281 ldlm_resource_unlink_lock(lock);
283 ldlm_extent_policy(res, lock, flags);
284 ldlm_grant_lock(lock, NULL, 0, 1);
285 RETURN(LDLM_ITER_CONTINUE);
289 LASSERT(res->lr_tmp == NULL);
290 res->lr_tmp = &rpc_list;
291 rc = ldlm_extent_compat_queue(&res->lr_granted, lock, 1, flags, err);
293 RETURN(rc); /* lock was destroyed */
297 rc2 = ldlm_extent_compat_queue(&res->lr_waiting, lock, 1, flags, err);
299 RETURN(rc2); /* lock was destroyed */
304 ldlm_extent_policy(res, lock, flags);
305 ldlm_resource_unlink_lock(lock);
306 ldlm_grant_lock(lock, NULL, 0, 0);
308 /* If either of the compat_queue()s returned failure, then we
309 * have ASTs to send and must go onto the waiting list.
311 * bug 2322: we used to unlink and re-add here, which was a
312 * terrible folly -- if we goto restart, we could get
313 * re-ordered! Causes deadlock, because ASTs aren't sent! */
314 if (list_empty(&lock->l_res_link))
315 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
316 l_unlock(&res->lr_namespace->ns_lock);
317 rc = ldlm_run_ast_work(res->lr_namespace, &rpc_list);
318 l_lock(&res->lr_namespace->ns_lock);
320 GOTO(restart, -ERESTART);
321 *flags |= LDLM_FL_BLOCK_GRANTED;
326 /* When a lock is cancelled by a client, the KMS may undergo change if this
327 * is the "highest lock". This function returns the new KMS value.
329 * NB: A lock on [x,y] protects a KMS of up to y + 1 bytes! */
330 __u64 ldlm_extent_shift_kms(struct ldlm_lock *lock, __u64 old_kms)
332 struct ldlm_resource *res = lock->l_resource;
333 struct list_head *tmp;
334 struct ldlm_lock *lck;
338 l_lock(&res->lr_namespace->ns_lock);
340 /* don't let another thread in ldlm_extent_shift_kms race in just after
341 * we finish and take our lock into account in its calculation of the
343 lock->l_flags |= LDLM_FL_KMS_IGNORE;
345 list_for_each(tmp, &res->lr_granted) {
346 lck = list_entry(tmp, struct ldlm_lock, l_res_link);
348 if (lck->l_flags & LDLM_FL_KMS_IGNORE)
351 if (lck->l_policy_data.l_extent.end >= old_kms)
352 GOTO(out, kms = old_kms);
354 /* This extent _has_ to be smaller than old_kms (checked above)
355 * so kms can only ever be smaller or the same as old_kms. */
356 if (lck->l_policy_data.l_extent.end + 1 > kms)
357 kms = lck->l_policy_data.l_extent.end + 1;
359 LASSERTF(kms <= old_kms, "kms "LPU64" old_kms "LPU64"\n", kms, old_kms);
363 l_unlock(&res->lr_namespace->ns_lock);