X-Git-Url: https://git.whamcloud.com/gitweb?a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_extent.c;h=7f299b4a9d78f187da8b680b6fcb09b85f98d054;hb=5516349042147d5553e5861a6f61739771d5f90d;hp=dd04c1c0f306a7ff8a163627706d969cd9211390;hpb=090c677210ee2946d99c71412e4ff762bb300f4f;p=fs%2Flustre-release.git diff --git a/lustre/ldlm/ldlm_extent.c b/lustre/ldlm/ldlm_extent.c index dd04c1c..7f299b4 100644 --- a/lustre/ldlm/ldlm_extent.c +++ b/lustre/ldlm/ldlm_extent.c @@ -5,20 +5,23 @@ * Author: Peter Braam * Author: Phil Schwan * - * This file is part of Lustre, http://www.lustre.org. + * This file is part of the Lustre file system, http://www.lustre.org + * Lustre is a trademark of Cluster File Systems, Inc. * - * Lustre is free software; you can redistribute it and/or - * modify it under the terms of version 2 of the GNU General Public - * License as published by the Free Software Foundation. + * You may have signed or agreed to another license before downloading + * this software. If so, you are bound by the terms and conditions + * of that agreement, and the following does not apply to you. See the + * LICENSE file included with this distribution for more information. * - * Lustre is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. + * If you did not agree to a different license, then this copy of Lustre + * is open source software; you can redistribute it and/or modify it + * under the terms of version 2 of the GNU General Public License as + * published by the Free Software Foundation. * - * You should have received a copy of the GNU General Public License - * along with Lustre; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * In either case, Lustre is distributed in the hope that it will be + * useful, but WITHOUT ANY WARRANTY; without even the implied warranty + * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * license text for more details. */ #define DEBUG_SUBSYSTEM S_LDLM @@ -26,9 +29,9 @@ # include #endif -#include -#include -#include +#include +#include +#include #include "ldlm_internal.h" @@ -45,6 +48,8 @@ ldlm_extent_internal_policy(struct list_head *queue, struct ldlm_lock *req, ldlm_mode_t req_mode = req->l_req_mode; __u64 req_start = req->l_req_extent.start; __u64 req_end = req->l_req_extent.end; + __u64 req_align, mask; + int conflicting = 0; ENTRY; lockmode_verify(req_mode); @@ -56,6 +61,7 @@ ldlm_extent_internal_policy(struct list_head *queue, struct ldlm_lock *req, lock = list_entry(tmp, struct ldlm_lock, l_res_link); l_extent = &lock->l_policy_data.l_extent; + /* We already hit the minimum requested size, search no more */ if (new_ex->start == req_start && new_ex->end == req_end) { EXIT; return; @@ -65,15 +71,24 @@ ldlm_extent_internal_policy(struct list_head *queue, struct ldlm_lock *req, if (req == lock) continue; + /* Locks are compatible, overlap doesn't matter */ + /* Until bug 20 is fixed, try to avoid granting overlapping + * locks on one client (they take a long time to cancel) */ + if (lockmode_compat(lock->l_req_mode, req_mode) && + lock->l_export != req->l_export) + continue; + + /* If this is a high-traffic lock, don't grow downwards at all + * or grow upwards too much */ + ++conflicting; + if (conflicting > 4) + new_ex->start = req_start; + /* If lock doesn't overlap new_ex, skip it. */ if (l_extent->end < new_ex->start || l_extent->start > new_ex->end) continue; - /* Locks are compatible, overlap doesn't matter */ - if (lockmode_compat(lock->l_req_mode, req_mode)) - continue; - /* Locks conflicting in requested extents and we can't satisfy * both locks, so ignore it. Either we will ping-pong this * extent (we would regardless of what extent we granted) or @@ -85,10 +100,10 @@ ldlm_extent_internal_policy(struct list_head *queue, struct ldlm_lock *req, /* We grow extents downwards only as far as they don't overlap * with already-granted locks, on the assumtion that clients * will be writing beyond the initial requested end and would - * then need to enqueue a new lock beyond the previous request. - * We don't grow downwards if there are lots of lockers. */ - if (l_extent->start < req_start) { - if (atomic_read(&req->l_resource->lr_refcount) > 20) + * then need to enqueue a new lock beyond previous request. + * l_req_extent->end strictly < req_start, checked above. */ + if (l_extent->start < req_start && new_ex->start != req_start) { + if (l_extent->end >= req_start) new_ex->start = req_start; else new_ex->start = min(l_extent->end+1, req_start); @@ -107,6 +122,41 @@ ldlm_extent_internal_policy(struct list_head *queue, struct ldlm_lock *req, new_ex->end = max(l_extent->start - 1, req_end); } } + +#define LDLM_MAX_GROWN_EXTENT (32 * 1024 * 1024 - 1) + if (conflicting > 32 && (req_mode == LCK_PW || req_mode == LCK_CW)) { + if (req_end < req_start + LDLM_MAX_GROWN_EXTENT) + new_ex->end = min(req_start + LDLM_MAX_GROWN_EXTENT, + new_ex->end); + } + + if (new_ex->start == 0 && new_ex->end == OBD_OBJECT_EOF) { + EXIT; + return; + } + + /* we need to ensure that the lock extent is properly aligned to what + * the client requested. We align it to the lowest-common denominator + * of the clients requested lock start and end alignment. */ + mask = 0x1000ULL; + req_align = (req_end + 1) | req_start; + if (req_align != 0) { + while ((req_align & mask) == 0) + mask <<= 1; + } + mask -= 1; + /* We can only shrink the lock, not grow it. + * This should never cause lock to be smaller than requested, + * since requested lock was already aligned on these boundaries. */ + new_ex->start = ((new_ex->start - 1) | mask) + 1; + new_ex->end = ((new_ex->end + 1) & ~mask) - 1; + LASSERTF(new_ex->start <= req_start, + "mask "LPX64" grant start "LPU64" req start "LPU64"\n", + mask, new_ex->start, req_start); + LASSERTF(new_ex->end >= req_end, + "mask "LPX64" grant end "LPU64" req end "LPU64"\n", + mask, new_ex->end, req_end); + EXIT; } @@ -115,7 +165,22 @@ ldlm_extent_internal_policy(struct list_head *queue, struct ldlm_lock *req, static void ldlm_extent_policy(struct ldlm_resource *res, struct ldlm_lock *lock, int *flags) { - struct ldlm_extent new_ex = { .start = 0, .end = ~0}; + struct ldlm_extent new_ex = { .start = 0, .end = OBD_OBJECT_EOF }; + + if (lock->l_export == NULL) + /* + * this is local lock taken by server (e.g., as a part of + * OST-side locking, or unlink handling). Expansion doesn't + * make a lot of sense for local locks, because they are + * dropped immediately on operation completion and would only + * conflict with other threads. + */ + return; + + if (lock->l_policy_data.l_extent.start == 0 && + lock->l_policy_data.l_extent.end == OBD_OBJECT_EOF) + /* fast-path whole file locks */ + return; ldlm_extent_internal_policy(&res->lr_granted, lock, &new_ex); ldlm_extent_internal_policy(&res->lr_waiting, lock, &new_ex); @@ -139,7 +204,8 @@ static void ldlm_extent_policy(struct ldlm_resource *res, */ static int ldlm_extent_compat_queue(struct list_head *queue, struct ldlm_lock *req, - int send_cbs, int *flags, ldlm_error_t *err) + int *flags, ldlm_error_t *err, + struct list_head *work_list) { struct list_head *tmp; struct ldlm_lock *lock; @@ -147,6 +213,7 @@ ldlm_extent_compat_queue(struct list_head *queue, struct ldlm_lock *req, __u64 req_start = req->l_req_extent.start; __u64 req_end = req->l_req_extent.end; int compat = 1; + int scan = 0; ENTRY; lockmode_verify(req_mode); @@ -157,21 +224,53 @@ ldlm_extent_compat_queue(struct list_head *queue, struct ldlm_lock *req, if (req == lock) RETURN(compat); + if (unlikely(scan)) { + /* We only get here if we are queuing GROUP lock + and met some incompatible one. The main idea of this + code is to insert GROUP lock past compatible GROUP + lock in the waiting queue or if there is not any, + then in front of first non-GROUP lock */ + if (lock->l_req_mode != LCK_GROUP) { + /* Ok, we hit non-GROUP lock, there should be no + more GROUP locks later on, queue in front of + first non-GROUP lock */ + + ldlm_resource_insert_lock_after(lock, req); + list_del_init(&lock->l_res_link); + ldlm_resource_insert_lock_after(req, lock); + RETURN(0); + } + if (req->l_policy_data.l_extent.gid == + lock->l_policy_data.l_extent.gid) { + /* found it */ + ldlm_resource_insert_lock_after(lock, req); + RETURN(0); + } + continue; + } + /* locks are compatible, overlap doesn't matter */ if (lockmode_compat(lock->l_req_mode, req_mode)) { - /* nonCW locks are compatible, overlap doesn't matter */ - if (req_mode != LCK_CW) + /* non-group locks are compatible, overlap doesn't + matter */ + if (likely(req_mode != LCK_GROUP)) continue; - - /* If we are trying to get a CW lock and there is + + /* If we are trying to get a GROUP lock and there is another one of this kind, we need to compare gid */ if (req->l_policy_data.l_extent.gid == - lock->l_policy_data.l_extent.gid) { + lock->l_policy_data.l_extent.gid) { + /* If existing lock with matched gid is granted, + we grant new one too. */ if (lock->l_req_mode == lock->l_granted_mode) RETURN(2); - /* If we are in nonblocking mode - return - immediately */ + /* Otherwise we are scanning queue of waiting + * locks and it means current request would + * block along with existing lock (that is + * already blocked. + * If we are in nonblocking mode - return + * immediately */ if (*flags & LDLM_FL_BLOCK_NOWAIT) { compat = -EWOULDBLOCK; goto destroylock; @@ -191,9 +290,32 @@ ldlm_extent_compat_queue(struct list_head *queue, struct ldlm_lock *req, } } - if (lock->l_req_mode == LCK_CW) { - /* If compared lock is CW, then requested is PR/PW/ => - * this is not compatible; extent range does not + if (unlikely(req_mode == LCK_GROUP && + (lock->l_req_mode != lock->l_granted_mode))) { + scan = 1; + compat = 0; + if (lock->l_req_mode != LCK_GROUP) { + /* Ok, we hit non-GROUP lock, there should + * be no more GROUP locks later on, queue in + * front of first non-GROUP lock */ + + ldlm_resource_insert_lock_after(lock, req); + list_del_init(&lock->l_res_link); + ldlm_resource_insert_lock_after(req, lock); + RETURN(0); + } + if (req->l_policy_data.l_extent.gid == + lock->l_policy_data.l_extent.gid) { + /* found it */ + ldlm_resource_insert_lock_after(lock, req); + RETURN(0); + } + continue; + } + + if (unlikely(lock->l_req_mode == LCK_GROUP)) { + /* If compared lock is GROUP, then requested is PR/PW/ + * so this is not compatible; extent range does not * matter */ if (*flags & LDLM_FL_BLOCK_NOWAIT) { compat = -EWOULDBLOCK; @@ -203,44 +325,43 @@ ldlm_extent_compat_queue(struct list_head *queue, struct ldlm_lock *req, } } else if (lock->l_policy_data.l_extent.end < req_start || lock->l_policy_data.l_extent.start > req_end) { - /* if a non-CW lock doesn't overlap skip it */ + /* if a non group lock doesn't overlap skip it */ continue; } - if (!send_cbs) + if (!work_list) RETURN(0); compat = 0; if (lock->l_blocking_ast) - ldlm_add_ast_work_item(lock, req, NULL, 0); + ldlm_add_ast_work_item(lock, req, work_list); } - return(compat); + RETURN(compat); destroylock: list_del_init(&req->l_res_link); - ldlm_lock_destroy(req); + ldlm_lock_destroy_nolock(req); *err = compat; RETURN(compat); } /* If first_enq is 0 (ie, called from ldlm_reprocess_queue): * - blocking ASTs have already been sent - * - the caller has already initialized req->lr_tmp * - must call this function with the ns lock held * * If first_enq is 1 (ie, called from ldlm_lock_enqueue): * - blocking ASTs have not been sent - * - the caller has NOT initialized req->lr_tmp, so we must * - must call this function with the ns lock held once */ int ldlm_process_extent_lock(struct ldlm_lock *lock, int *flags, int first_enq, - ldlm_error_t *err) + ldlm_error_t *err, struct list_head *work_list) { struct ldlm_resource *res = lock->l_resource; - struct list_head rpc_list = LIST_HEAD_INIT(rpc_list); + struct list_head rpc_list = CFS_LIST_HEAD_INIT(rpc_list); int rc, rc2; ENTRY; LASSERT(list_empty(&res->lr_converting)); + check_res_locked(res); *err = ELDLM_OK; if (!first_enq) { @@ -250,42 +371,39 @@ int ldlm_process_extent_lock(struct ldlm_lock *lock, int *flags, int first_enq, * flags should always be zero here, and if that ever stops * being true, we want to find out. */ LASSERT(*flags == 0); - LASSERT(res->lr_tmp != NULL); - rc = ldlm_extent_compat_queue(&res->lr_granted, lock, 0, flags, - err); + rc = ldlm_extent_compat_queue(&res->lr_granted, lock, flags, + err, NULL); if (rc == 1) { - rc = ldlm_extent_compat_queue(&res->lr_waiting, lock, 0, - flags, err); + rc = ldlm_extent_compat_queue(&res->lr_waiting, lock, + flags, err, NULL); } if (rc == 0) RETURN(LDLM_ITER_STOP); ldlm_resource_unlink_lock(lock); - ldlm_extent_policy(res, lock, flags); - ldlm_grant_lock(lock, NULL, 0, 1); + if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_EVICT_RACE)) + ldlm_extent_policy(res, lock, flags); + ldlm_grant_lock(lock, work_list); RETURN(LDLM_ITER_CONTINUE); } restart: - LASSERT(res->lr_tmp == NULL); - res->lr_tmp = &rpc_list; - rc = ldlm_extent_compat_queue(&res->lr_granted, lock, 1, flags, err); + rc = ldlm_extent_compat_queue(&res->lr_granted, lock, flags, err, &rpc_list); if (rc < 0) - RETURN(rc); /* lock was destroyed */ + GOTO(out, rc); /* lock was destroyed */ if (rc == 2) goto grant; - rc2 = ldlm_extent_compat_queue(&res->lr_waiting, lock, 1, flags, err); + rc2 = ldlm_extent_compat_queue(&res->lr_waiting, lock, flags, err, &rpc_list); if (rc2 < 0) - RETURN(rc2); /* lock was destroyed */ - res->lr_tmp = NULL; + GOTO(out, rc = rc2); /* lock was destroyed */ if (rc + rc2 == 2) { grant: ldlm_extent_policy(res, lock, flags); ldlm_resource_unlink_lock(lock); - ldlm_grant_lock(lock, NULL, 0, 0); + ldlm_grant_lock(lock, NULL); } else { /* If either of the compat_queue()s returned failure, then we * have ASTs to send and must go onto the waiting list. @@ -295,18 +413,25 @@ int ldlm_process_extent_lock(struct ldlm_lock *lock, int *flags, int first_enq, * re-ordered! Causes deadlock, because ASTs aren't sent! */ if (list_empty(&lock->l_res_link)) ldlm_resource_add_lock(res, &res->lr_waiting, lock); - l_unlock(&res->lr_namespace->ns_lock); - rc = ldlm_run_ast_work(res->lr_namespace, &rpc_list); - l_lock(&res->lr_namespace->ns_lock); + unlock_res(res); + rc = ldlm_run_bl_ast_work(&rpc_list); + lock_res(res); if (rc == -ERESTART) GOTO(restart, -ERESTART); *flags |= LDLM_FL_BLOCK_GRANTED; + /* this way we force client to wait for the lock + * endlessly once the lock is enqueued -bzzz */ + *flags |= LDLM_FL_NO_TIMEOUT; + } - RETURN(0); + rc = 0; +out: + RETURN(rc); } /* When a lock is cancelled by a client, the KMS may undergo change if this * is the "highest lock". This function returns the new KMS value. + * Caller must hold ns_lock already. * * NB: A lock on [x,y] protects a KMS of up to y + 1 bytes! */ __u64 ldlm_extent_shift_kms(struct ldlm_lock *lock, __u64 old_kms) @@ -317,19 +442,26 @@ __u64 ldlm_extent_shift_kms(struct ldlm_lock *lock, __u64 old_kms) __u64 kms = 0; ENTRY; - l_lock(&res->lr_namespace->ns_lock); + /* don't let another thread in ldlm_extent_shift_kms race in + * just after we finish and take our lock into account in its + * calculation of the kms */ + lock->l_flags |= LDLM_FL_KMS_IGNORE; + list_for_each(tmp, &res->lr_granted) { lck = list_entry(tmp, struct ldlm_lock, l_res_link); - if (lock == lck) + if (lck->l_flags & LDLM_FL_KMS_IGNORE) continue; + if (lck->l_policy_data.l_extent.end >= old_kms) - GOTO(out, kms = old_kms); - kms = lck->l_policy_data.l_extent.end + 1; + RETURN(old_kms); + + /* This extent _has_ to be smaller than old_kms (checked above) + * so kms can only ever be smaller or the same as old_kms. */ + if (lck->l_policy_data.l_extent.end + 1 > kms) + kms = lck->l_policy_data.l_extent.end + 1; } + LASSERTF(kms <= old_kms, "kms "LPU64" old_kms "LPU64"\n", kms, old_kms); - GOTO(out, kms); - out: - l_unlock(&res->lr_namespace->ns_lock); - return kms; + RETURN(kms); }