X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_inodebits.c;h=22daed502957d19618be467edafd321d74179497;hp=636527b566b81d85fe5c35cb8043f56c61662c8a;hb=3cce65712d94cffe8f1626545845b95b88aef672;hpb=d10200a80770f0029d1d665af954187b9ad883df diff --git a/lustre/ldlm/ldlm_inodebits.c b/lustre/ldlm/ldlm_inodebits.c index 636527b..22daed5 100644 --- a/lustre/ldlm/ldlm_inodebits.c +++ b/lustre/ldlm/ldlm_inodebits.c @@ -23,7 +23,7 @@ * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2016, Intel Corporation. + * Copyright (c) 2011, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -57,6 +57,89 @@ #include "ldlm_internal.h" #ifdef HAVE_SERVER_SUPPORT + +/** + * It should iterate through all waiting locks on a given resource queue and + * attempt to grant them. An optimization is to check only heads waitintg + * locks for each inodebit type. + * + * Must be called with resource lock held. + */ +int ldlm_reprocess_inodebits_queue(struct ldlm_resource *res, + struct list_head *queue, + struct list_head *work_list, + enum ldlm_process_intention intention, + struct ldlm_lock *hint) +{ + __u64 flags; + int rc = LDLM_ITER_CONTINUE; + enum ldlm_error err; + struct list_head bl_ast_list = LIST_HEAD_INIT(bl_ast_list); + struct ldlm_ibits_queues *queues = res->lr_ibits_queues; + int i; + + ENTRY; + + check_res_locked(res); + + LASSERT(res->lr_type == LDLM_IBITS); + LASSERT(intention == LDLM_PROCESS_RESCAN || + intention == LDLM_PROCESS_RECOVERY); + + if (intention == LDLM_PROCESS_RECOVERY) + return ldlm_reprocess_queue(res, queue, work_list, intention, + NULL); + +restart: + CDEBUG(D_DLMTRACE, "--- Reprocess resource "DLDLMRES" (%p)\n", + PLDLMRES(res), res); + + for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) { + struct list_head rpc_list = LIST_HEAD_INIT(rpc_list); + struct list_head *head = &queues->liq_waiting[i]; + struct ldlm_lock *pending; + struct ldlm_ibits_node *node; + + if (list_empty(head)) + continue; + if (hint && !(hint->l_policy_data.l_inodebits.bits & (1 << i))) + continue; + + node = list_entry(head->next, struct ldlm_ibits_node, + lin_link[i]); + + pending = node->lock; + LDLM_DEBUG(pending, "Reprocessing lock from queue %d", i); + + flags = 0; + rc = ldlm_process_inodebits_lock(pending, &flags, intention, + &err, &rpc_list); + if (ldlm_is_granted(pending)) { + list_splice(&rpc_list, work_list); + /* Try to grant more locks from current queue */ + i--; + } else { + list_splice(&rpc_list, &bl_ast_list); + } + } + + if (!list_empty(&bl_ast_list)) { + unlock_res(res); + + rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list, + LDLM_WORK_BL_AST); + + lock_res(res); + if (rc == -ERESTART) + GOTO(restart, rc); + } + + if (!list_empty(&bl_ast_list)) + ldlm_discard_bl_list(&bl_ast_list); + + RETURN(rc); +} + /** * Determine if the lock is compatible with all locks on the queue. * @@ -79,12 +162,18 @@ ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req, struct list_head *tmp; struct ldlm_lock *lock; __u64 req_bits = req->l_policy_data.l_inodebits.bits; + __u64 *try_bits = &req->l_policy_data.l_inodebits.try_bits; int compat = 1; + ENTRY; - /* There is no sense in lock with no bits set, I think. - * Also, such a lock would be compatible with any other bit lock */ - LASSERT(req_bits != 0); + /* There is no sense in lock with no bits set. Also such a lock + * would be compatible with any other bit lock. + * Meanwhile that can be true if there were just try_bits and all + * are failed, so just exit gracefully and let the caller to care. + */ + if ((req_bits | *try_bits) == 0) + RETURN(0); list_for_each(tmp, queue) { struct list_head *mode_tail; @@ -99,11 +188,10 @@ ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req, /* last lock in mode group */ LASSERT(lock->l_sl_mode.prev != NULL); - mode_tail = &list_entry(lock->l_sl_mode.prev, - struct ldlm_lock, + mode_tail = &list_entry(lock->l_sl_mode.prev, struct ldlm_lock, l_sl_mode)->l_res_link; - /* if reqest lock is not COS_INCOMPAT and COS is disabled, + /* if request lock is not COS_INCOMPAT and COS is disabled, * they are compatible, IOW this request is from a local * transaction on a DNE system. */ if (lock->l_req_mode == LCK_COS && !ldlm_is_cos_incompat(req) && @@ -125,8 +213,24 @@ ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req, /* Advance loop cursor to last lock in policy group. */ tmp = &list_entry(lock->l_sl_policy.prev, - struct ldlm_lock, - l_sl_policy)->l_res_link; + struct ldlm_lock, + l_sl_policy)->l_res_link; + + /* New lock's try_bits are filtered out by ibits + * of all locks in both granted and waiting queues. + */ + *try_bits &= ~(lock->l_policy_data.l_inodebits.bits | + lock->l_policy_data.l_inodebits.try_bits); + + if ((req_bits | *try_bits) == 0) + RETURN(0); + + /* The new lock ibits is more preferable than try_bits + * of waiting locks so drop conflicting try_bits in + * the waiting queue. + * Notice that try_bits of granted locks must be zero. + */ + lock->l_policy_data.l_inodebits.try_bits &= ~req_bits; /* Locks with overlapping bits conflict. */ if (lock->l_policy_data.l_inodebits.bits & req_bits) { @@ -138,6 +242,7 @@ ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req, ldlm_is_cos_enabled(req) && lock->l_client_cookie == req->l_client_cookie) goto not_conflicting; + /* Found a conflicting policy group. */ if (!work_list) RETURN(0); @@ -146,22 +251,21 @@ ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req, /* Add locks of the policy group to @work_list * as blocking locks for @req */ - if (lock->l_blocking_ast) - ldlm_add_ast_work_item(lock, req, - work_list); - head = &lock->l_sl_policy; + if (lock->l_blocking_ast) + ldlm_add_ast_work_item(lock, req, + work_list); + head = &lock->l_sl_policy; list_for_each_entry(lock, head, l_sl_policy) - if (lock->l_blocking_ast) - ldlm_add_ast_work_item(lock, req, - work_list); - } - not_conflicting: - if (tmp == mode_tail) - break; - - tmp = tmp->next; - lock = list_entry(tmp, struct ldlm_lock, - l_res_link); + if (lock->l_blocking_ast) + ldlm_add_ast_work_item(lock, + req, work_list); + } +not_conflicting: + if (tmp == mode_tail) + break; + + tmp = tmp->next; + lock = list_entry(tmp, struct ldlm_lock, l_res_link); } /* Loop over policy groups within one mode group. */ } /* Loop over mode groups within @queue. */ @@ -175,88 +279,102 @@ ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req, * This function looks for any conflicts for \a lock in the granted or * waiting queues. The lock is granted if no conflicts are found in * either queue. - * - * If \a first_enq is 0 (ie, called from ldlm_reprocess_queue): - * - blocking ASTs have already been sent - * - * If \a first_enq is 1 (ie, called from ldlm_lock_enqueue): - * - blocking ASTs have not been sent yet, so list of conflicting locks - * would be collected and ASTs sent. */ int ldlm_process_inodebits_lock(struct ldlm_lock *lock, __u64 *flags, - int first_enq, enum ldlm_error *err, + enum ldlm_process_intention intention, + enum ldlm_error *err, struct list_head *work_list) { struct ldlm_resource *res = lock->l_resource; - struct list_head rpc_list; + struct list_head *grant_work = intention == LDLM_PROCESS_ENQUEUE ? + NULL : work_list; int rc; + ENTRY; - LASSERT(lock->l_granted_mode != lock->l_req_mode); - LASSERT(list_empty(&res->lr_converting)); - INIT_LIST_HEAD(&rpc_list); + LASSERT(!ldlm_is_granted(lock)); check_res_locked(res); - /* (*flags & LDLM_FL_BLOCK_NOWAIT) is for layout lock right now. */ - if (!first_enq || (*flags & LDLM_FL_BLOCK_NOWAIT)) { - *err = ELDLM_LOCK_ABORTED; - if (*flags & LDLM_FL_BLOCK_NOWAIT) - *err = ELDLM_LOCK_WOULDBLOCK; + if (intention == LDLM_PROCESS_RESCAN) { + struct list_head *bl_list; - rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock, NULL); - if (!rc) - RETURN(LDLM_ITER_STOP); - rc = ldlm_inodebits_compat_queue(&res->lr_waiting, lock, NULL); - if (!rc) - RETURN(LDLM_ITER_STOP); + if (*flags & LDLM_FL_BLOCK_NOWAIT) { + bl_list = NULL; + *err = ELDLM_LOCK_WOULDBLOCK; + } else { + bl_list = work_list; + *err = ELDLM_LOCK_ABORTED; + } - ldlm_resource_unlink_lock(lock); - ldlm_grant_lock(lock, work_list); + LASSERT(lock->l_policy_data.l_inodebits.bits != 0); + + /* It is possible that some of granted locks was not canceled + * but converted and is kept in granted queue. So there is + * a window where lock with 'ast_sent' might become granted + * again. Meanwhile a new lock may appear in that window and + * conflicts with the converted lock so the following scenario + * is possible: + * + * 1) lock1 conflicts with lock2 + * 2) bl_ast was sent for lock2 + * 3) lock3 comes and conflicts with lock2 too + * 4) no bl_ast sent because lock2->l_bl_ast_sent is 1 + * 5) lock2 was converted for lock1 but not for lock3 + * 6) lock1 granted, lock3 still is waiting for lock2, but + * there will never be another bl_ast for that + * + * To avoid this scenario the work_list is used below to collect + * any blocked locks from granted queue during every reprocess + * and bl_ast will be sent if needed. + */ + rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock, + bl_list); + if (!rc) + RETURN(LDLM_ITER_STOP); + rc = ldlm_inodebits_compat_queue(&res->lr_waiting, lock, NULL); + if (!rc) + RETURN(LDLM_ITER_STOP); + + /* grant also try_bits if any */ + if (lock->l_policy_data.l_inodebits.try_bits != 0) { + lock->l_policy_data.l_inodebits.bits |= + lock->l_policy_data.l_inodebits.try_bits; + lock->l_policy_data.l_inodebits.try_bits = 0; + *flags |= LDLM_FL_LOCK_CHANGED; + } + ldlm_resource_unlink_lock(lock); + ldlm_grant_lock(lock, grant_work); *err = ELDLM_OK; RETURN(LDLM_ITER_CONTINUE); } - restart: - rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock, &rpc_list); - rc += ldlm_inodebits_compat_queue(&res->lr_waiting, lock, &rpc_list); - - if (rc != 2) { - /* If either of the compat_queue()s returned 0, then we - * have ASTs to send and must go onto the waiting list. - * - * bug 2322: we used to unlink and re-add here, which was a - * terrible folly -- if we goto restart, we could get - * 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); - unlock_res(res); - rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list, - LDLM_WORK_BL_AST); - lock_res(res); - if (rc == -ERESTART) { - /* We were granted while waiting, nothing left to do */ - if (lock->l_granted_mode == lock->l_req_mode) - GOTO(out, rc = 0); - /* Lock was destroyed while we were waiting, abort */ - if (ldlm_is_destroyed(lock)) - GOTO(out, rc = -EAGAIN); - - /* Otherwise try again */ - GOTO(restart, rc); - } - *flags |= LDLM_FL_BLOCK_GRANTED; - } else { - ldlm_resource_unlink_lock(lock); - ldlm_grant_lock(lock, NULL); - } + rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock, work_list); + rc += ldlm_inodebits_compat_queue(&res->lr_waiting, lock, work_list); - rc = 0; -out: - *err = rc; - LASSERT(list_empty(&rpc_list)); + if (rc != 2) { + /* if there were only bits to try and all are conflicting */ + if ((lock->l_policy_data.l_inodebits.bits | + lock->l_policy_data.l_inodebits.try_bits) == 0) { + *err = ELDLM_LOCK_WOULDBLOCK; + } else { + *err = ELDLM_OK; + } + } else { + /* grant also all remaining try_bits */ + if (lock->l_policy_data.l_inodebits.try_bits != 0) { + lock->l_policy_data.l_inodebits.bits |= + lock->l_policy_data.l_inodebits.try_bits; + lock->l_policy_data.l_inodebits.try_bits = 0; + *flags |= LDLM_FL_LOCK_CHANGED; + } + LASSERT(lock->l_policy_data.l_inodebits.bits); + ldlm_resource_unlink_lock(lock); + ldlm_grant_lock(lock, grant_work); + *err = ELDLM_OK; + } - RETURN(rc); + RETURN(LDLM_ITER_CONTINUE); } #endif /* HAVE_SERVER_SUPPORT */ @@ -264,6 +382,7 @@ void ldlm_ibits_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy, union ldlm_policy_data *lpolicy) { lpolicy->l_inodebits.bits = wpolicy->l_inodebits.bits; + lpolicy->l_inodebits.try_bits = wpolicy->l_inodebits.try_bits; } void ldlm_ibits_policy_local_to_wire(const union ldlm_policy_data *lpolicy, @@ -271,4 +390,175 @@ void ldlm_ibits_policy_local_to_wire(const union ldlm_policy_data *lpolicy, { memset(wpolicy, 0, sizeof(*wpolicy)); wpolicy->l_inodebits.bits = lpolicy->l_inodebits.bits; + wpolicy->l_inodebits.try_bits = lpolicy->l_inodebits.try_bits; +} + +/** + * Attempt to convert already granted IBITS lock with several bits set to + * a lock with less bits (downgrade). + * + * Such lock conversion is used to keep lock with non-blocking bits instead of + * cancelling it, introduced for better support of DoM files. + */ +int ldlm_inodebits_drop(struct ldlm_lock *lock, __u64 to_drop) +{ + ENTRY; + + check_res_locked(lock->l_resource); + + /* Just return if there are no conflicting bits */ + if ((lock->l_policy_data.l_inodebits.bits & to_drop) == 0) { + LDLM_WARN(lock, "try to drop unset bits %#llx/%#llx", + lock->l_policy_data.l_inodebits.bits, to_drop); + /* nothing to do */ + RETURN(0); + } + + /* remove lock from a skiplist and put in the new place + * according with new inodebits */ + ldlm_resource_unlink_lock(lock); + lock->l_policy_data.l_inodebits.bits &= ~to_drop; + ldlm_grant_lock_with_skiplist(lock); + RETURN(0); +} +EXPORT_SYMBOL(ldlm_inodebits_drop); + +/* convert single lock */ +int ldlm_cli_dropbits(struct ldlm_lock *lock, __u64 drop_bits) +{ + struct lustre_handle lockh; + __u32 flags = 0; + int rc; + + ENTRY; + + LASSERT(drop_bits); + LASSERT(!lock->l_readers && !lock->l_writers); + + LDLM_DEBUG(lock, "client lock convert START"); + + ldlm_lock2handle(lock, &lockh); + lock_res_and_lock(lock); + /* check if all bits are blocked */ + if (!(lock->l_policy_data.l_inodebits.bits & ~drop_bits)) { + unlock_res_and_lock(lock); + /* return error to continue with cancel */ + GOTO(exit, rc = -EINVAL); + } + + /* check if no common bits, consider this as successful convert */ + if (!(lock->l_policy_data.l_inodebits.bits & drop_bits)) { + unlock_res_and_lock(lock); + GOTO(exit, rc = 0); + } + + /* check if there is race with cancel */ + if (ldlm_is_canceling(lock) || ldlm_is_cancel(lock)) { + unlock_res_and_lock(lock); + GOTO(exit, rc = -EINVAL); + } + + /* clear cbpending flag early, it is safe to match lock right after + * client convert because it is downgrade always. + */ + ldlm_clear_cbpending(lock); + ldlm_clear_bl_ast(lock); + + /* If lock is being converted already, check drop bits first */ + if (ldlm_is_converting(lock)) { + /* raced lock convert, lock inodebits are remaining bits + * so check if they are conflicting with new convert or not. + */ + if (!(lock->l_policy_data.l_inodebits.bits & drop_bits)) { + unlock_res_and_lock(lock); + GOTO(exit, rc = 0); + } + /* Otherwise drop new conflicting bits in new convert */ + } + ldlm_set_converting(lock); + /* from all bits of blocking lock leave only conflicting */ + drop_bits &= lock->l_policy_data.l_inodebits.bits; + /* save them in cancel_bits, so l_blocking_ast will know + * which bits from the current lock were dropped. */ + lock->l_policy_data.l_inodebits.cancel_bits = drop_bits; + /* Finally clear these bits in lock ibits */ + ldlm_inodebits_drop(lock, drop_bits); + unlock_res_and_lock(lock); + /* Finally call cancel callback for remaining bits only. + * It is important to have converting flag during that + * so blocking_ast callback can distinguish convert from + * cancels. + */ + if (lock->l_blocking_ast) + lock->l_blocking_ast(lock, NULL, lock->l_ast_data, + LDLM_CB_CANCELING); + + /* now notify server about convert */ + rc = ldlm_cli_convert(lock, &flags); + if (rc) { + lock_res_and_lock(lock); + if (ldlm_is_converting(lock)) { + ldlm_clear_converting(lock); + ldlm_set_cbpending(lock); + ldlm_set_bl_ast(lock); + } + unlock_res_and_lock(lock); + GOTO(exit, rc); + } + EXIT; +exit: + LDLM_DEBUG(lock, "client lock convert END"); + return rc; +} + + +int ldlm_inodebits_alloc_lock(struct ldlm_lock *lock) +{ + if (ldlm_is_ns_srv(lock)) { + int i; + + OBD_SLAB_ALLOC_PTR(lock->l_ibits_node, ldlm_inodebits_slab); + if (lock->l_ibits_node == NULL) + return -ENOMEM; + for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) + INIT_LIST_HEAD(&lock->l_ibits_node->lin_link[i]); + lock->l_ibits_node->lock = lock; + } else { + lock->l_ibits_node = NULL; + } + return 0; +} + +void ldlm_inodebits_add_lock(struct ldlm_resource *res, struct list_head *head, + struct ldlm_lock *lock) +{ + int i; + + if (!ldlm_is_ns_srv(lock)) + return; + + if (head == &res->lr_waiting) { + for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) { + if (lock->l_policy_data.l_inodebits.bits & (1 << i)) + list_add_tail(&lock->l_ibits_node->lin_link[i], + &res->lr_ibits_queues->liq_waiting[i]); + } + } else if (head == &res->lr_granted && lock->l_ibits_node != NULL) { + for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) + LASSERT(list_empty(&lock->l_ibits_node->lin_link[i])); + OBD_SLAB_FREE_PTR(lock->l_ibits_node, ldlm_inodebits_slab); + lock->l_ibits_node = NULL; + } +} + +void ldlm_inodebits_unlink_lock(struct ldlm_lock *lock) +{ + int i; + + ldlm_unlink_lock_skiplist(lock); + if (!ldlm_is_ns_srv(lock)) + return; + + for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) + list_del_init(&lock->l_ibits_node->lin_link[i]); }