X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_flock.c;h=238d4bed8775f634fb1837e3e8977dc111cabe35;hp=32837d5518c89a5d0a909677ff57c198d2d924c6;hb=12e311012ae337276dc3e7da3e7ad8d85d11e764;hpb=864669f7853a3325b1433832541d47d71fe0d19d diff --git a/lustre/ldlm/ldlm_flock.c b/lustre/ldlm/ldlm_flock.c index 32837d5..238d4be 100644 --- a/lustre/ldlm/ldlm_flock.c +++ b/lustre/ldlm/ldlm_flock.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,45 +24,46 @@ * GPL HEADER END */ /* - * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. - * Use is subject to license terms. - * * Copyright (c) 2003 Hewlett-Packard Development Company LP. * Developed under the sponsorship of the US Government under * Subcontract No. B514193 + * + * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + * + * Copyright (c) 2010, 2012, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ * Lustre is a trademark of Sun Microsystems, Inc. */ +/** + * This file implements POSIX lock type for Lustre. + * Its policy properties are start and end of extent and PID. + * + * These locks are only done through MDS due to POSIX semantics requiring + * e.g. that locks could be only partially released and as such split into + * two parts, and also that two adjacent locks from the same process may be + * merged into a single wider lock. + * + * Lock modes are mapped like this: + * PR and PW for READ and WRITE locks + * NL to request a releasing of a portion of the lock + * + * These flock locks never timeout. + */ + #define DEBUG_SUBSYSTEM S_LDLM -#ifdef __KERNEL__ #include #include #include #include #include -#else -#include -#include -#endif #include "ldlm_internal.h" -#define l_flock_waitq l_lru - -/** - * Wait queue for Posix lock deadlock detection, added with - * ldlm_lock::l_flock_waitq. - */ -static CFS_LIST_HEAD(ldlm_flock_waitq); -/** - * Lock protecting access to ldlm_flock_waitq. - */ -cfs_spinlock_t ldlm_flock_waitq_lock = CFS_SPIN_LOCK_UNLOCKED; - int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, void *data, int flag); @@ -82,8 +81,8 @@ int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, static inline int ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new) { - return((new->l_policy_data.l_flock.pid == - lock->l_policy_data.l_flock.pid) && + return((new->l_policy_data.l_flock.owner == + lock->l_policy_data.l_flock.owner) && (new->l_export == lock->l_export)); } @@ -96,22 +95,55 @@ ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new) lock->l_policy_data.l_flock.start)); } +static inline void ldlm_flock_blocking_link(struct ldlm_lock *req, + struct ldlm_lock *lock) +{ + /* For server only */ + if (req->l_export == NULL) + return; + + LASSERT(hlist_unhashed(&req->l_exp_flock_hash)); + + req->l_policy_data.l_flock.blocking_owner = + lock->l_policy_data.l_flock.owner; + req->l_policy_data.l_flock.blocking_export = + lock->l_export; + req->l_policy_data.l_flock.blocking_refs = 0; + + cfs_hash_add(req->l_export->exp_flock_hash, + &req->l_policy_data.l_flock.owner, + &req->l_exp_flock_hash); +} + +static inline void ldlm_flock_blocking_unlink(struct ldlm_lock *req) +{ + /* For server only */ + if (req->l_export == NULL) + return; + + check_res_locked(req->l_resource); + if (req->l_export->exp_flock_hash != NULL && + !hlist_unhashed(&req->l_exp_flock_hash)) + cfs_hash_del(req->l_export->exp_flock_hash, + &req->l_policy_data.l_flock.owner, + &req->l_exp_flock_hash); +} + static inline void -ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, int flags) +ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, __u64 flags) { - ENTRY; + ENTRY; - LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: 0x%x)", - mode, flags); + LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: "LPX64")", + mode, flags); - /* Safe to not lock here, since it should be empty anyway */ - LASSERT(cfs_list_empty(&lock->l_flock_waitq)); + /* Safe to not lock here, since it should be empty anyway */ + LASSERT(hlist_unhashed(&lock->l_exp_flock_hash)); - cfs_list_del_init(&lock->l_res_link); - if (flags == LDLM_FL_WAIT_NOREPROC && - !(lock->l_flags & LDLM_FL_FAILED)) { - /* client side - set a flag to prevent sending a CANCEL */ - lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING; + list_del_init(&lock->l_res_link); + if (flags == LDLM_FL_WAIT_NOREPROC) { + /* client side - set a flag to prevent sending a CANCEL */ + lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING; /* when reaching here, it is under lock_res_and_lock(). Thus, need call the nolock version of ldlm_lock_decref_internal*/ @@ -122,45 +154,145 @@ ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, int flags) EXIT; } -static int -ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *blocking_lock) +/** + * POSIX locks deadlock detection code. + * + * Given a new lock \a req and an existing lock \a bl_lock it conflicts + * with, we need to iterate through all blocked POSIX locks for this + * export and see if there is a deadlock condition arising. (i.e. when + * one client holds a lock on something and want a lock on something + * else and at the same time another client has the opposite situation). + */ + +struct ldlm_flock_lookup_cb_data { + __u64 *bl_owner; + struct ldlm_lock *lock; + struct obd_export *exp; +}; + +static int ldlm_flock_lookup_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd, + struct hlist_node *hnode, void *data) { - struct obd_export *req_export = req->l_export; - struct obd_export *blocking_export = blocking_lock->l_export; - pid_t req_pid = req->l_policy_data.l_flock.pid; - pid_t blocking_pid = blocking_lock->l_policy_data.l_flock.pid; - struct ldlm_lock *lock; + struct ldlm_flock_lookup_cb_data *cb_data = data; + struct obd_export *exp = cfs_hash_object(hs, hnode); + struct ldlm_lock *lock; - cfs_spin_lock(&ldlm_flock_waitq_lock); -restart: - cfs_list_for_each_entry(lock, &ldlm_flock_waitq, l_flock_waitq) { - if ((lock->l_policy_data.l_flock.pid != blocking_pid) || - (lock->l_export != blocking_export)) - continue; + lock = cfs_hash_lookup(exp->exp_flock_hash, cb_data->bl_owner); + if (lock == NULL) + return 0; + + /* Stop on first found lock. Same process can't sleep twice */ + cb_data->lock = lock; + cb_data->exp = class_export_get(exp); - blocking_pid = lock->l_policy_data.l_flock.blocking_pid; - blocking_export = (struct obd_export *)(long) - lock->l_policy_data.l_flock.blocking_export; - if (blocking_pid == req_pid && blocking_export == req_export) { - cfs_spin_unlock(&ldlm_flock_waitq_lock); + return 1; +} + +static int +ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *bl_lock) +{ + struct obd_export *req_exp = req->l_export; + struct obd_export *bl_exp = bl_lock->l_export; + __u64 req_owner = req->l_policy_data.l_flock.owner; + __u64 bl_owner = bl_lock->l_policy_data.l_flock.owner; + + /* For server only */ + if (req_exp == NULL) + return 0; + + class_export_get(bl_exp); + while (1) { + struct ldlm_flock_lookup_cb_data cb_data = { + .bl_owner = &bl_owner, + .lock = NULL, + .exp = NULL }; + struct obd_export *bl_exp_new; + struct ldlm_lock *lock = NULL; + struct ldlm_flock *flock; + + if (bl_exp->exp_flock_hash != NULL) { + cfs_hash_for_each_key(bl_exp->exp_obd->obd_nid_hash, + &bl_exp->exp_connection->c_peer.nid, + ldlm_flock_lookup_cb, &cb_data); + lock = cb_data.lock; + } + if (lock == NULL) + break; + + class_export_put(bl_exp); + bl_exp = cb_data.exp; + + LASSERT(req != lock); + flock = &lock->l_policy_data.l_flock; + LASSERT(flock->owner == bl_owner); + bl_owner = flock->blocking_owner; + bl_exp_new = class_export_get(flock->blocking_export); + class_export_put(bl_exp); + + cfs_hash_put(bl_exp->exp_flock_hash, &lock->l_exp_flock_hash); + bl_exp = bl_exp_new; + + if (bl_exp->exp_failed) + break; + + if (bl_owner == req_owner && + (bl_exp->exp_connection->c_peer.nid == + req_exp->exp_connection->c_peer.nid)) { + class_export_put(bl_exp); return 1; } - - goto restart; } - cfs_spin_unlock(&ldlm_flock_waitq_lock); + class_export_put(bl_exp); return 0; } +static void ldlm_flock_cancel_on_deadlock(struct ldlm_lock *lock, + struct list_head *work_list) +{ + CDEBUG(D_INFO, "reprocess deadlock req=%p\n", lock); + + if ((exp_connect_flags(lock->l_export) & + OBD_CONNECT_FLOCK_DEAD) == 0) { + CERROR("deadlock found, but client doesn't " + "support flock canceliation\n"); + } else { + LASSERT(lock->l_completion_ast); + LASSERT(!ldlm_is_ast_sent(lock)); + lock->l_flags |= LDLM_FL_AST_SENT | LDLM_FL_CANCEL_ON_BLOCK | + LDLM_FL_FLOCK_DEADLOCK; + ldlm_flock_blocking_unlink(lock); + ldlm_resource_unlink_lock(lock); + ldlm_add_ast_work_item(lock, NULL, work_list); + } +} + +/** + * Process a granting attempt for flock lock. + * Must be called under ns lock held. + * + * 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. + * + * It is also responsible for splitting a lock if a portion of the lock + * is released. + * + * 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_flock_lock(struct ldlm_lock *req, int *flags, int first_enq, - ldlm_error_t *err, cfs_list_t *work_list) +ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags, int first_enq, + ldlm_error_t *err, struct list_head *work_list) { struct ldlm_resource *res = req->l_resource; struct ldlm_namespace *ns = ldlm_res_to_ns(res); - cfs_list_t *tmp; - cfs_list_t *ownlocks = NULL; + struct list_head *tmp; + struct list_head *ownlocks = NULL; struct ldlm_lock *lock = NULL; struct ldlm_lock *new = req; struct ldlm_lock *new2 = NULL; @@ -172,8 +304,10 @@ ldlm_process_flock_lock(struct ldlm_lock *req, int *flags, int first_enq, const struct ldlm_callback_suite null_cbs = { NULL }; ENTRY; - CDEBUG(D_DLMTRACE, "flags %#x pid %u mode %u start "LPU64" end "LPU64 - "\n", *flags, new->l_policy_data.l_flock.pid, mode, + CDEBUG(D_DLMTRACE, "flags "LPX64" owner "LPU64" pid %u mode %u start " + LPU64" end "LPU64"\n", *flags, + new->l_policy_data.l_flock.owner, + new->l_policy_data.l_flock.pid, mode, req->l_policy_data.l_flock.start, req->l_policy_data.l_flock.end); @@ -192,8 +326,8 @@ reprocess: if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) { /* This loop determines where this processes locks start * in the resource lr_granted list. */ - cfs_list_for_each(tmp, &res->lr_granted) { - lock = cfs_list_entry(tmp, struct ldlm_lock, + list_for_each(tmp, &res->lr_granted) { + lock = list_entry(tmp, struct ldlm_lock, l_res_link); if (ldlm_same_flock_owner(lock, req)) { ownlocks = tmp; @@ -201,12 +335,13 @@ reprocess: } } } else { + int reprocess_failed = 0; lockmode_verify(mode); /* This loop determines if there are existing locks * that conflict with the new lock request. */ - cfs_list_for_each(tmp, &res->lr_granted) { - lock = cfs_list_entry(tmp, struct ldlm_lock, + list_for_each(tmp, &res->lr_granted) { + lock = list_entry(tmp, struct ldlm_lock, l_res_link); if (ldlm_same_flock_owner(lock, req)) { @@ -222,8 +357,15 @@ reprocess: if (!ldlm_flocks_overlap(lock, req)) continue; - if (!first_enq) - RETURN(LDLM_ITER_CONTINUE); + if (!first_enq) { + reprocess_failed = 1; + if (ldlm_flock_deadlock(req, lock)) { + ldlm_flock_cancel_on_deadlock(req, + work_list); + RETURN(LDLM_ITER_CONTINUE); + } + continue; + } if (*flags & LDLM_FL_BLOCK_NOWAIT) { ldlm_flock_destroy(req, mode, *flags); @@ -244,27 +386,23 @@ reprocess: RETURN(LDLM_ITER_STOP); } - if (ldlm_flock_deadlock(req, lock)) { - ldlm_flock_destroy(req, mode, *flags); - *err = -EDEADLK; - RETURN(LDLM_ITER_STOP); - } + /* add lock to blocking list before deadlock + * check to prevent race */ + ldlm_flock_blocking_link(req, lock); - req->l_policy_data.l_flock.blocking_pid = - lock->l_policy_data.l_flock.pid; - req->l_policy_data.l_flock.blocking_export = - (long)(void *)lock->l_export; - - LASSERT(cfs_list_empty(&req->l_flock_waitq)); - cfs_spin_lock(&ldlm_flock_waitq_lock); - cfs_list_add_tail(&req->l_flock_waitq, - &ldlm_flock_waitq); - cfs_spin_unlock(&ldlm_flock_waitq_lock); + if (ldlm_flock_deadlock(req, lock)) { + ldlm_flock_blocking_unlink(req); + ldlm_flock_destroy(req, mode, *flags); + *err = -EDEADLK; + RETURN(LDLM_ITER_STOP); + } ldlm_resource_add_lock(res, &res->lr_waiting, req); *flags |= LDLM_FL_BLOCK_GRANTED; RETURN(LDLM_ITER_STOP); } + if (reprocess_failed) + RETURN(LDLM_ITER_CONTINUE); } if (*flags & LDLM_FL_TEST_LOCK) { @@ -275,10 +413,8 @@ reprocess: } /* In case we had slept on this lock request take it off of the - * deadlock detection waitq. */ - cfs_spin_lock(&ldlm_flock_waitq_lock); - cfs_list_del_init(&req->l_flock_waitq); - cfs_spin_unlock(&ldlm_flock_waitq_lock); + * deadlock detection hash list. */ + ldlm_flock_blocking_unlink(req); /* Scan the locks owned by this process that overlap this request. * We may have to merge or split existing locks. */ @@ -287,7 +423,7 @@ reprocess: ownlocks = &res->lr_granted; list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) { - lock = cfs_list_entry(ownlocks, struct ldlm_lock, l_res_link); + lock = list_entry(ownlocks, struct ldlm_lock, l_res_link); if (!ldlm_same_flock_owner(lock, new)) break; @@ -375,26 +511,28 @@ reprocess: /* XXX - if ldlm_lock_new() can sleep we should * release the lr_lock, allocate the new lock, * and restart processing this lock. */ - if (!new2) { - unlock_res_and_lock(req); - new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK, - lock->l_granted_mode, &null_cbs, - NULL, 0); - lock_res_and_lock(req); - if (!new2) { - ldlm_flock_destroy(req, lock->l_granted_mode, - *flags); - *err = -ENOLCK; - RETURN(LDLM_ITER_STOP); - } - goto reprocess; - } + if (new2 == NULL) { + unlock_res_and_lock(req); + new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK, + lock->l_granted_mode, &null_cbs, + NULL, 0, LVB_T_NONE); + lock_res_and_lock(req); + if (IS_ERR(new2)) { + ldlm_flock_destroy(req, lock->l_granted_mode, + *flags); + *err = PTR_ERR(new2); + RETURN(LDLM_ITER_STOP); + } + goto reprocess; + } splitted = 1; new2->l_granted_mode = lock->l_granted_mode; new2->l_policy_data.l_flock.pid = new->l_policy_data.l_flock.pid; + new2->l_policy_data.l_flock.owner = + new->l_policy_data.l_flock.owner; new2->l_policy_data.l_flock.start = lock->l_policy_data.l_flock.start; new2->l_policy_data.l_flock.end = @@ -405,7 +543,7 @@ reprocess: if (lock->l_export != NULL) { new2->l_export = class_export_lock_get(lock->l_export, new2); if (new2->l_export->exp_lock_hash && - cfs_hlist_unhashed(&new2->l_exp_hash)) + hlist_unhashed(&new2->l_exp_hash)) cfs_hash_add(new2->l_export->exp_lock_hash, &new2->l_remote_handle, &new2->l_exp_hash); @@ -429,12 +567,13 @@ reprocess: /* Add req to the granted queue before calling ldlm_reprocess_all(). */ if (!added) { - cfs_list_del_init(&req->l_res_link); + list_del_init(&req->l_res_link); /* insert new lock before ownlocks in list. */ ldlm_resource_add_lock(res, ownlocks, req); } if (*flags != LDLM_FL_WAIT_NOREPROC) { +#ifdef HAVE_SERVER_SUPPORT if (first_enq) { /* If this is an unlock, reprocess the waitq and * send completions ASTs for locks that can now be @@ -445,30 +584,39 @@ reprocess: * note that ldlm_process_flock_lock() will recurse, * but only once because first_enq will be false from * ldlm_reprocess_queue. */ - if ((mode == LCK_NL) && overlaps) { - CFS_LIST_HEAD(rpc_list); + if ((mode == LCK_NL) && overlaps) { + struct list_head rpc_list; int rc; + + INIT_LIST_HEAD(&rpc_list); restart: ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list); unlock_res_and_lock(req); - rc = ldlm_run_ast_work(&rpc_list, + rc = ldlm_run_ast_work(ns, &rpc_list, LDLM_WORK_CP_AST); lock_res_and_lock(req); - if (rc == -ERESTART) - GOTO(restart, -ERESTART); + if (rc == -ERESTART) + GOTO(restart, rc); } } else { LASSERT(req->l_completion_ast); ldlm_add_ast_work_item(req, NULL, work_list); } +#else /* !HAVE_SERVER_SUPPORT */ + /* The only one possible case for client-side calls flock + * policy function is ldlm_flock_completion_ast inside which + * carries LDLM_FL_WAIT_NOREPROC flag. */ + CERROR("Illegal parameter for client-side-only module.\n"); + LBUG(); +#endif /* HAVE_SERVER_SUPPORT */ } - /* In case we're reprocessing the requested lock we can't destroy - * it until after calling ldlm_ast_work_item() above so that lawi() - * can bump the reference count on req. Otherwise req could be freed - * before the completion AST can be sent. */ + /* In case we're reprocessing the requested lock we can't destroy + * it until after calling ldlm_add_ast_work_item() above so that laawi() + * can bump the reference count on \a req. Otherwise \a req + * could be freed before the completion AST can be sent. */ if (added) ldlm_flock_destroy(req, mode, *flags); @@ -489,19 +637,19 @@ ldlm_flock_interrupted_wait(void *data) lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock; - /* take lock off the deadlock detection waitq. */ - cfs_spin_lock(&ldlm_flock_waitq_lock); - cfs_list_del_init(&lock->l_flock_waitq); - cfs_spin_unlock(&ldlm_flock_waitq_lock); + /* take lock off the deadlock detection hash list. */ + lock_res_and_lock(lock); + ldlm_flock_blocking_unlink(lock); - /* client side - set flag to prevent lock from being put on lru list */ - lock->l_flags |= LDLM_FL_CBPENDING; + /* client side - set flag to prevent lock from being put on LRU list */ + ldlm_set_cbpending(lock); + unlock_res_and_lock(lock); EXIT; } /** - * Flock completion calback function. + * Flock completion callback function. * * \param lock [in,out]: A lock to be handled * \param flags [in]: flags @@ -511,9 +659,9 @@ ldlm_flock_interrupted_wait(void *data) * \retval <0 : failure */ int -ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data) +ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) { - cfs_flock_t *getlk = lock->l_ast_data; + struct file_lock *getlk = lock->l_ast_data; struct obd_device *obd; struct obd_import *imp = NULL; struct ldlm_flock_wait_data fwd; @@ -522,36 +670,30 @@ ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data) int rc = 0; ENTRY; - CDEBUG(D_DLMTRACE, "flags: 0x%x data: %p getlk: %p\n", - flags, data, getlk); - - /* Import invalidation. We need to actually release the lock - * references being held, so that it can go away. No point in - * holding the lock even if app still believes it has it, since - * server already dropped it anyway. Only for granted locks too. */ - if ((lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) == - (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) { - if (lock->l_req_mode == lock->l_granted_mode && - lock->l_granted_mode != LCK_NL && - NULL == data) - ldlm_lock_decref_internal(lock, lock->l_req_mode); - - /* Need to wake up the waiter if we were evicted */ - cfs_waitq_signal(&lock->l_waitq); - RETURN(0); - } - - LASSERT(flags != LDLM_FL_WAIT_NOREPROC); - - if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED | - LDLM_FL_BLOCK_CONV))) { - if (NULL == data) - /* mds granted the lock in the reply */ - goto granted; - /* CP AST RPC: lock get granted, wake it up */ - cfs_waitq_signal(&lock->l_waitq); - RETURN(0); - } + OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT2, 4); + if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT3)) { + lock_res_and_lock(lock); + lock->l_flags |= LDLM_FL_FAIL_LOC; + unlock_res_and_lock(lock); + OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT3, 4); + } + CDEBUG(D_DLMTRACE, "flags: "LPX64" data: %p getlk: %p\n", + flags, data, getlk); + + LASSERT(flags != LDLM_FL_WAIT_NOREPROC); + + if (flags & LDLM_FL_FAILED) + goto granted; + + if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED | + LDLM_FL_BLOCK_CONV))) { + if (NULL == data) + /* mds granted the lock in the reply */ + goto granted; + /* CP AST RPC: lock get granted, wake it up */ + wake_up(&lock->l_waitq); + RETURN(0); + } LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, " "sleeping"); @@ -563,9 +705,9 @@ ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data) imp = obd->u.cli.cl_import; if (NULL != imp) { - cfs_spin_lock(&imp->imp_lock); - fwd.fwd_generation = imp->imp_generation; - cfs_spin_unlock(&imp->imp_lock); + spin_lock(&imp->imp_lock); + fwd.fwd_generation = imp->imp_generation; + spin_unlock(&imp->imp_lock); } lwi = LWI_TIMEOUT_INTR(0, NULL, ldlm_flock_interrupted_wait, &fwd); @@ -582,81 +724,257 @@ ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data) granted: OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10); - if (lock->l_destroyed) { - LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed"); - RETURN(0); - } - - if (lock->l_flags & LDLM_FL_FAILED) { - LDLM_DEBUG(lock, "client-side enqueue waking up: failed"); - RETURN(-EIO); - } + if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT4)) { + lock_res_and_lock(lock); + /* DEADLOCK is always set with CBPENDING */ + lock->l_flags |= LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING; + unlock_res_and_lock(lock); + OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT4, 4); + } + if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT5)) { + lock_res_and_lock(lock); + /* DEADLOCK is always set with CBPENDING */ + lock->l_flags |= LDLM_FL_FAIL_LOC | + LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING; + unlock_res_and_lock(lock); + OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT5, 4); + } + + lock_res_and_lock(lock); + + + /* Protect against race where lock could have been just destroyed + * due to overlap in ldlm_process_flock_lock(). + */ + if (ldlm_is_destroyed(lock)) { + unlock_res_and_lock(lock); + LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed"); + + /* An error is still to be returned, to propagate it up to + * ldlm_cli_enqueue_fini() caller. */ + RETURN(-EIO); + } - if (rc) { - LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)", - rc); - RETURN(rc); - } - - LDLM_DEBUG(lock, "client-side enqueue granted"); - - /* take lock off the deadlock detection waitq. */ - cfs_spin_lock(&ldlm_flock_waitq_lock); - cfs_list_del_init(&lock->l_flock_waitq); - cfs_spin_unlock(&ldlm_flock_waitq_lock); - - lock_res_and_lock(lock); /* ldlm_lock_enqueue() has already placed lock on the granted list. */ - cfs_list_del_init(&lock->l_res_link); - - if (flags & LDLM_FL_TEST_LOCK) { + ldlm_resource_unlink_lock(lock); + + /* Import invalidation. We need to actually release the lock + * references being held, so that it can go away. No point in + * holding the lock even if app still believes it has it, since + * server already dropped it anyway. Only for granted locks too. */ + /* Do the same for DEADLOCK'ed locks. */ + if (ldlm_is_failed(lock) || ldlm_is_flock_deadlock(lock)) { + int mode; + + if (flags & LDLM_FL_TEST_LOCK) + LASSERT(ldlm_is_test_lock(lock)); + + if (ldlm_is_test_lock(lock) || ldlm_is_flock_deadlock(lock)) + mode = flock_type(getlk); + else + mode = lock->l_granted_mode; + + if (ldlm_is_flock_deadlock(lock)) { + LDLM_DEBUG(lock, "client-side enqueue deadlock " + "received"); + rc = -EDEADLK; + } + ldlm_flock_destroy(lock, mode, LDLM_FL_WAIT_NOREPROC); + unlock_res_and_lock(lock); + + /* Need to wake up the waiter if we were evicted */ + wake_up(&lock->l_waitq); + + /* An error is still to be returned, to propagate it up to + * ldlm_cli_enqueue_fini() caller. */ + RETURN(rc ? : -EIO); + } + + LDLM_DEBUG(lock, "client-side enqueue granted"); + + if (flags & LDLM_FL_TEST_LOCK) { /* fcntl(F_GETLK) request */ /* The old mode was saved in getlk->fl_type so that if the mode * in the lock changes we can decref the appropriate refcount.*/ - ldlm_flock_destroy(lock, cfs_flock_type(getlk), - LDLM_FL_WAIT_NOREPROC); - switch (lock->l_granted_mode) { - case LCK_PR: - cfs_flock_set_type(getlk, F_RDLCK); - break; - case LCK_PW: - cfs_flock_set_type(getlk, F_WRLCK); - break; - default: - cfs_flock_set_type(getlk, F_UNLCK); - } - cfs_flock_set_pid(getlk, - (pid_t)lock->l_policy_data.l_flock.pid); - cfs_flock_set_start(getlk, - (loff_t)lock->l_policy_data.l_flock.start); - cfs_flock_set_end(getlk, - (loff_t)lock->l_policy_data.l_flock.end); - } else { - int noreproc = LDLM_FL_WAIT_NOREPROC; - - /* We need to reprocess the lock to do merges or splits - * with existing locks owned by this process. */ - ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL); - } - unlock_res_and_lock(lock); - RETURN(0); + LASSERT(ldlm_is_test_lock(lock)); + ldlm_flock_destroy(lock, flock_type(getlk), + LDLM_FL_WAIT_NOREPROC); + switch (lock->l_granted_mode) { + case LCK_PR: + flock_set_type(getlk, F_RDLCK); + break; + case LCK_PW: + flock_set_type(getlk, F_WRLCK); + break; + default: + flock_set_type(getlk, F_UNLCK); + } + flock_set_pid(getlk, (pid_t)lock->l_policy_data.l_flock.pid); + flock_set_start(getlk, + (loff_t)lock->l_policy_data.l_flock.start); + flock_set_end(getlk, + (loff_t)lock->l_policy_data.l_flock.end); + } else { + __u64 noreproc = LDLM_FL_WAIT_NOREPROC; + + /* We need to reprocess the lock to do merges or splits + * with existing locks owned by this process. */ + ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL); + } + unlock_res_and_lock(lock); + RETURN(rc); } EXPORT_SYMBOL(ldlm_flock_completion_ast); int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, void *data, int flag) { - struct ldlm_namespace *ns; ENTRY; LASSERT(lock); LASSERT(flag == LDLM_CB_CANCELING); - ns = ldlm_lock_to_ns(lock); - - /* take lock off the deadlock detection waitq. */ - cfs_spin_lock(&ldlm_flock_waitq_lock); - cfs_list_del_init(&lock->l_flock_waitq); - cfs_spin_unlock(&ldlm_flock_waitq_lock); + /* take lock off the deadlock detection hash list. */ + lock_res_and_lock(lock); + ldlm_flock_blocking_unlink(lock); + unlock_res_and_lock(lock); RETURN(0); } + +void ldlm_flock_policy_wire18_to_local(const ldlm_wire_policy_data_t *wpolicy, + ldlm_policy_data_t *lpolicy) +{ + memset(lpolicy, 0, sizeof(*lpolicy)); + lpolicy->l_flock.start = wpolicy->l_flock.lfw_start; + lpolicy->l_flock.end = wpolicy->l_flock.lfw_end; + lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid; + /* Compat code, old clients had no idea about owner field and + * relied solely on pid for ownership. Introduced in LU-104, 2.1, + * April 2011 */ + lpolicy->l_flock.owner = wpolicy->l_flock.lfw_pid; +} + + +void ldlm_flock_policy_wire21_to_local(const ldlm_wire_policy_data_t *wpolicy, + ldlm_policy_data_t *lpolicy) +{ + memset(lpolicy, 0, sizeof(*lpolicy)); + lpolicy->l_flock.start = wpolicy->l_flock.lfw_start; + lpolicy->l_flock.end = wpolicy->l_flock.lfw_end; + lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid; + lpolicy->l_flock.owner = wpolicy->l_flock.lfw_owner; +} + +void ldlm_flock_policy_local_to_wire(const ldlm_policy_data_t *lpolicy, + ldlm_wire_policy_data_t *wpolicy) +{ + memset(wpolicy, 0, sizeof(*wpolicy)); + wpolicy->l_flock.lfw_start = lpolicy->l_flock.start; + wpolicy->l_flock.lfw_end = lpolicy->l_flock.end; + wpolicy->l_flock.lfw_pid = lpolicy->l_flock.pid; + wpolicy->l_flock.lfw_owner = lpolicy->l_flock.owner; +} + +/* + * Export handle<->flock hash operations. + */ +static unsigned +ldlm_export_flock_hash(cfs_hash_t *hs, const void *key, unsigned mask) +{ + return cfs_hash_u64_hash(*(__u64 *)key, mask); +} + +static void * +ldlm_export_flock_key(struct hlist_node *hnode) +{ + struct ldlm_lock *lock; + + lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash); + return &lock->l_policy_data.l_flock.owner; +} + +static int +ldlm_export_flock_keycmp(const void *key, struct hlist_node *hnode) +{ + return !memcmp(ldlm_export_flock_key(hnode), key, sizeof(__u64)); +} + +static void * +ldlm_export_flock_object(struct hlist_node *hnode) +{ + return hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash); +} + +static void +ldlm_export_flock_get(cfs_hash_t *hs, struct hlist_node *hnode) +{ + struct ldlm_lock *lock; + struct ldlm_flock *flock; + + lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash); + LDLM_LOCK_GET(lock); + + flock = &lock->l_policy_data.l_flock; + LASSERT(flock->blocking_export != NULL); + class_export_get(flock->blocking_export); + flock->blocking_refs++; +} + +static void +ldlm_export_flock_put(cfs_hash_t *hs, struct hlist_node *hnode) +{ + struct ldlm_lock *lock; + struct ldlm_flock *flock; + + lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash); + LDLM_LOCK_RELEASE(lock); + + flock = &lock->l_policy_data.l_flock; + LASSERT(flock->blocking_export != NULL); + class_export_put(flock->blocking_export); + if (--flock->blocking_refs == 0) { + flock->blocking_owner = 0; + flock->blocking_export = NULL; + } +} + +static cfs_hash_ops_t ldlm_export_flock_ops = { + .hs_hash = ldlm_export_flock_hash, + .hs_key = ldlm_export_flock_key, + .hs_keycmp = ldlm_export_flock_keycmp, + .hs_object = ldlm_export_flock_object, + .hs_get = ldlm_export_flock_get, + .hs_put = ldlm_export_flock_put, + .hs_put_locked = ldlm_export_flock_put, +}; + +int ldlm_init_flock_export(struct obd_export *exp) +{ + if( strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_MDT_NAME) != 0) + RETURN(0); + + exp->exp_flock_hash = + cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid), + HASH_EXP_LOCK_CUR_BITS, + HASH_EXP_LOCK_MAX_BITS, + HASH_EXP_LOCK_BKT_BITS, 0, + CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA, + &ldlm_export_flock_ops, + CFS_HASH_DEFAULT | CFS_HASH_NBLK_CHANGE); + if (!exp->exp_flock_hash) + RETURN(-ENOMEM); + + RETURN(0); +} +EXPORT_SYMBOL(ldlm_init_flock_export); + +void ldlm_destroy_flock_export(struct obd_export *exp) +{ + ENTRY; + if (exp->exp_flock_hash) { + cfs_hash_putref(exp->exp_flock_hash); + exp->exp_flock_hash = NULL; + } + EXIT; +} +EXPORT_SYMBOL(ldlm_destroy_flock_export);