X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fosc%2Fosc_lock.c;h=ffcfdf767d402166690ef43eb162b774c86e7878;hb=1ba7eb520d6f50908985b45cb1aecad83bdf9c63;hp=d14acdcc5af5cb9283b339ef5662ce1e598573f2;hpb=3d3e8e27d6bfbfcf6970f7aa8942caf7a362c3db;p=fs%2Flustre-release.git diff --git a/lustre/osc/osc_lock.c b/lustre/osc/osc_lock.c index d14acdc..ffcfdf7 100644 --- a/lustre/osc/osc_lock.c +++ b/lustre/osc/osc_lock.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. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved. + * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2012, Whamcloud, Inc. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -38,8 +38,6 @@ * Author: Nikita Danilov */ -/** \addtogroup osc osc @{ */ - #define DEBUG_SUBSYSTEM S_OSC #ifdef __KERNEL__ @@ -52,6 +50,12 @@ #include "osc_cl_internal.h" +/** \addtogroup osc + * @{ + */ + +#define _PAGEREF_MAGIC (-10000000) + /***************************************************************************** * * Type conversions. @@ -133,10 +137,10 @@ static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck) { struct ldlm_lock *dlmlock; - spin_lock(&osc_ast_guard); + cfs_spin_lock(&osc_ast_guard); dlmlock = olck->ols_lock; if (dlmlock == NULL) { - spin_unlock(&osc_ast_guard); + cfs_spin_unlock(&osc_ast_guard); return; } @@ -145,56 +149,79 @@ static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck) * call to osc_lock_detach() */ dlmlock->l_ast_data = NULL; olck->ols_handle.cookie = 0ULL; - spin_unlock(&osc_ast_guard); + cfs_spin_unlock(&osc_ast_guard); lock_res_and_lock(dlmlock); if (dlmlock->l_granted_mode == dlmlock->l_req_mode) { struct cl_object *obj = olck->ols_cl.cls_obj; struct cl_attr *attr = &osc_env_info(env)->oti_attr; - __u64 old_kms = cl2osc(obj)->oo_oinfo->loi_kms; + __u64 old_kms; + cl_object_attr_lock(obj); + /* Must get the value under the lock to avoid possible races. */ + old_kms = cl2osc(obj)->oo_oinfo->loi_kms; /* Update the kms. Need to loop all granted locks. * Not a problem for the client */ attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms); - unlock_res_and_lock(dlmlock); - cl_object_attr_lock(obj); cl_object_attr_set(env, obj, attr, CAT_KMS); cl_object_attr_unlock(obj); - } else - unlock_res_and_lock(dlmlock); + } + unlock_res_and_lock(dlmlock); /* release a reference taken in osc_lock_upcall0(). */ + LASSERT(olck->ols_has_ref); lu_ref_del(&dlmlock->l_reference, "osc_lock", olck); LDLM_LOCK_RELEASE(dlmlock); + olck->ols_has_ref = 0; +} + +static int osc_lock_unhold(struct osc_lock *ols) +{ + int result = 0; + + if (ols->ols_hold) { + ols->ols_hold = 0; + result = osc_cancel_base(&ols->ols_handle, + ols->ols_einfo.ei_mode); + } + return result; } static int osc_lock_unuse(const struct lu_env *env, const struct cl_lock_slice *slice) { struct osc_lock *ols = cl2osc_lock(slice); - int result; - LASSERT(ols->ols_state == OLS_GRANTED || - ols->ols_state == OLS_UPCALL_RECEIVED); LINVRNT(osc_lock_invariant(ols)); - if (ols->ols_glimpse) { - LASSERT(ols->ols_hold == 0); + switch (ols->ols_state) { + case OLS_NEW: + LASSERT(!ols->ols_hold); + LASSERT(ols->ols_agl); return 0; + case OLS_UPCALL_RECEIVED: + osc_lock_unhold(ols); + case OLS_ENQUEUED: + LASSERT(!ols->ols_hold); + osc_lock_detach(env, ols); + ols->ols_state = OLS_NEW; + return 0; + case OLS_GRANTED: + LASSERT(!ols->ols_glimpse); + LASSERT(ols->ols_hold); + /* + * Move lock into OLS_RELEASED state before calling + * osc_cancel_base() so that possible synchronous cancellation + * (that always happens e.g., for liblustre) sees that lock is + * released. + */ + ols->ols_state = OLS_RELEASED; + return osc_lock_unhold(ols); + default: + CERROR("Impossible state: %d\n", ols->ols_state); + LBUG(); } - LASSERT(ols->ols_hold); - - /* - * Move lock into OLS_RELEASED state before calling osc_cancel_base() - * so that possible synchronous cancellation (that always happens - * e.g., for liblustre) sees that lock is released. - */ - ols->ols_state = OLS_RELEASED; - ols->ols_hold = 0; - result = osc_cancel_base(&ols->ols_handle, ols->ols_einfo.ei_mode); - ols->ols_has_ref = 0; - return result; } static void osc_lock_fini(const struct lu_env *env, @@ -209,9 +236,10 @@ static void osc_lock_fini(const struct lu_env *env, * to the lock), before reply from a server was received. In this case * lock is destroyed immediately after upcall. */ - if (ols->ols_hold) - osc_lock_unuse(env, slice); + osc_lock_unhold(ols); LASSERT(ols->ols_lock == NULL); + LASSERT(cfs_atomic_read(&ols->ols_pageref) == 0 || + cfs_atomic_read(&ols->ols_pageref) == _PAGEREF_MAGIC); OBD_SLAB_FREE_PTR(ols, osc_lock_kmem); } @@ -229,9 +257,9 @@ void osc_lock_build_res(const struct lu_env *env, const struct osc_object *obj, } else { /* * In reality, where ost server expects ->lsm_object_id and - * ->lsm_object_gr in rename. + * ->lsm_object_seq in rename. */ - osc_build_res_name(obj->oo_oinfo->loi_id, obj->oo_oinfo->loi_gr, + osc_build_res_name(obj->oo_oinfo->loi_id, obj->oo_oinfo->loi_seq, resname); } } @@ -265,14 +293,14 @@ static int osc_enq2ldlm_flags(__u32 enqflags) * Global spin-lock protecting consistency of ldlm_lock::l_ast_data * pointers. Initialized in osc_init(). */ -spinlock_t osc_ast_guard; +cfs_spinlock_t osc_ast_guard; static struct osc_lock *osc_ast_data_get(struct ldlm_lock *dlm_lock) { struct osc_lock *olck; lock_res_and_lock(dlm_lock); - spin_lock(&osc_ast_guard); + cfs_spin_lock(&osc_ast_guard); olck = dlm_lock->l_ast_data; if (olck != NULL) { struct cl_lock *lock = olck->ols_cl.cls_lock; @@ -292,7 +320,7 @@ static struct osc_lock *osc_ast_data_get(struct ldlm_lock *dlm_lock) } else olck = NULL; } - spin_unlock(&osc_ast_guard); + cfs_spin_unlock(&osc_ast_guard); unlock_res_and_lock(dlm_lock); return olck; } @@ -327,10 +355,8 @@ static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck, ENTRY; - if (!(olck->ols_flags & LDLM_FL_LVB_READY)) { - EXIT; - return; - } + if (!(olck->ols_flags & LDLM_FL_LVB_READY)) + RETURN_EXIT; lvb = &olck->ols_lvb; obj = olck->ols_cl.cls_obj; @@ -397,7 +423,7 @@ static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck, LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode); ENTRY; - if (olck->ols_state != OLS_GRANTED) { + if (olck->ols_state < OLS_GRANTED) { lock = olck->ols_cl.cls_lock; ext = &dlmlock->l_policy_data.l_extent; descr = &osc_env_info(env)->oti_descr; @@ -440,11 +466,11 @@ static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck) LASSERT(dlmlock != NULL); lock_res_and_lock(dlmlock); - spin_lock(&osc_ast_guard); + cfs_spin_lock(&osc_ast_guard); LASSERT(dlmlock->l_ast_data == olck); LASSERT(olck->ols_lock == NULL); olck->ols_lock = dlmlock; - spin_unlock(&osc_ast_guard); + cfs_spin_unlock(&osc_ast_guard); /* * Lock might be not yet granted. In this case, completion ast @@ -460,11 +486,12 @@ static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck) * this. */ ldlm_lock_addref(&olck->ols_handle, olck->ols_einfo.ei_mode); - olck->ols_hold = olck->ols_has_ref = 1; + olck->ols_hold = 1; /* lock reference taken by ldlm_handle2lock_long() is owned by * osc_lock and released in osc_lock_detach() */ lu_ref_add(&dlmlock->l_reference, "osc_lock", olck); + olck->ols_has_ref = 1; } /** @@ -494,7 +521,7 @@ static int osc_lock_upcall(void *cookie, int errcode) } else if (olck->ols_state == OLS_CANCELLED) { rc = -EIO; } else { - CERROR("Impossible state: %i\n", olck->ols_state); + CERROR("Impossible state: %d\n", olck->ols_state); LBUG(); } if (rc) { @@ -503,11 +530,12 @@ static int osc_lock_upcall(void *cookie, int errcode) dlmlock = ldlm_handle2lock(&olck->ols_handle); if (dlmlock != NULL) { lock_res_and_lock(dlmlock); - spin_lock(&osc_ast_guard); + cfs_spin_lock(&osc_ast_guard); LASSERT(olck->ols_lock == NULL); dlmlock->l_ast_data = NULL; olck->ols_handle.cookie = 0ULL; - spin_unlock(&osc_ast_guard); + cfs_spin_unlock(&osc_ast_guard); + ldlm_lock_fail_match_locked(dlmlock); unlock_res_and_lock(dlmlock); LDLM_LOCK_PUT(dlmlock); } @@ -536,22 +564,42 @@ static int osc_lock_upcall(void *cookie, int errcode) rc = 0; } - if (rc == 0) - /* on error, lock was signaled by cl_lock_error() */ + if (rc == 0) { + /* For AGL case, the RPC sponsor may exits the cl_lock + * processing without wait() called before related OSC + * lock upcall(). So update the lock status according + * to the enqueue result inside AGL upcall(). */ + if (olck->ols_agl) { + lock->cll_flags |= CLF_FROM_UPCALL; + cl_wait_try(env, lock); + lock->cll_flags &= ~CLF_FROM_UPCALL; + if (!olck->ols_glimpse) + olck->ols_agl = 0; + } cl_lock_signal(env, lock); - else + /* del user for lock upcall cookie */ + cl_unuse_try(env, lock); + } else { + /* del user for lock upcall cookie */ + cl_lock_user_del(env, lock); cl_lock_error(env, lock, rc); + } - cl_lock_mutex_put(env, lock); - - /* release cookie reference, acquired by osc_lock_enqueue() */ - lu_ref_del(&lock->cll_reference, "upcall", lock); - cl_lock_put(env, lock); - cl_env_nested_put(&nest, env); - } else - /* should never happen, similar to osc_ldlm_blocking_ast(). */ - LBUG(); - RETURN(errcode); + /* release cookie reference, acquired by osc_lock_enqueue() */ + cl_lock_hold_release(env, lock, "upcall", lock); + cl_lock_mutex_put(env, lock); + + lu_ref_del(&lock->cll_reference, "upcall", lock); + /* This maybe the last reference, so must be called after + * cl_lock_mutex_put(). */ + cl_lock_put(env, lock); + + cl_env_nested_put(&nest, env); + } else { + /* should never happen, similar to osc_ldlm_blocking_ast(). */ + LBUG(); + } + RETURN(errcode); } /** @@ -567,12 +615,11 @@ static void osc_lock_blocking(const struct lu_env *env, CLASSERT(OLS_BLOCKED < OLS_CANCELLED); LASSERT(!osc_lock_is_lockless(olck)); - if (olck->ols_hold) - /* - * Lock might be still addref-ed here, if e.g., blocking ast - * is sent for a failed lock. - */ - osc_lock_unuse(env, &olck->ols_cl); + /* + * Lock might be still addref-ed here, if e.g., blocking ast + * is sent for a failed lock. + */ + osc_lock_unhold(olck); if (blocking && olck->ols_state < OLS_BLOCKED) /* @@ -769,7 +816,7 @@ static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock, LASSERT(dlmlock->l_lvb_data != NULL); lock_res_and_lock(dlmlock); olck->ols_lvb = *(struct ost_lvb *)dlmlock->l_lvb_data; - if (olck->ols_lock == NULL) + if (olck->ols_lock == NULL) { /* * upcall (osc_lock_upcall()) hasn't yet been * called. Do nothing now, upcall will bind @@ -779,12 +826,17 @@ static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock, * and ldlm_lock are always bound when * osc_lock is in OLS_GRANTED state. */ - ; - else if (dlmlock->l_granted_mode != LCK_MINMODE) + } else if (dlmlock->l_granted_mode == + dlmlock->l_req_mode) { osc_lock_granted(env, olck, dlmlock, dlmrc); + } unlock_res_and_lock(dlmlock); - if (dlmrc != 0) + + if (dlmrc != 0) { + CL_LOCK_DEBUG(D_ERROR, env, lock, + "dlmlock returned %d\n", dlmrc); cl_lock_error(env, lock, dlmrc); + } cl_lock_mutex_put(env, lock); osc_ast_data_put(env, olck); result = 0; @@ -812,8 +864,7 @@ static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data) env = cl_env_nested_get(&nest); if (!IS_ERR(env)) { - /* - * osc_ast_data_get() has to go after environment is + /* osc_ast_data_get() has to go after environment is * allocated, because osc_ast_data() acquires a * reference to a lock, and it can only be released in * environment. @@ -821,7 +872,12 @@ static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data) olck = osc_ast_data_get(dlmlock); if (olck != NULL) { lock = olck->ols_cl.cls_lock; - cl_lock_mutex_get(env, lock); + /* Do not grab the mutex of cl_lock for glimpse. + * See LU-1274 for details. + * BTW, it's okay for cl_lock to be cancelled during + * this period because server can handle this race. + * See ldlm_server_glimpse_ast() for details. + * cl_lock_mutex_get(env, lock); */ cap = &req->rq_pill; req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK); req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER, @@ -832,7 +888,6 @@ static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data) obj = lock->cll_descr.cld_obj; result = cl_object_glimpse(env, obj, lvb); } - cl_lock_mutex_put(env, lock); osc_ast_data_put(env, olck); } else { /* @@ -874,7 +929,7 @@ static unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock) unsigned long weight; ENTRY; - might_sleep(); + cfs_might_sleep(); /* * osc_ldlm_weigh_ast has a complex context since it might be called * because of lock canceling, or from user's input. We have to make @@ -934,86 +989,6 @@ static void osc_lock_build_einfo(const struct lu_env *env, einfo->ei_cbdata = lock; /* value to be put into ->l_ast_data */ } -static int osc_lock_delete0(struct cl_lock *conflict) -{ - struct cl_env_nest nest; - struct lu_env *env; - int rc = 0; - - env = cl_env_nested_get(&nest); - if (!IS_ERR(env)) { - cl_lock_delete(env, conflict); - cl_env_nested_put(&nest, env); - } else - rc = PTR_ERR(env); - return rc; -} -/** - * Cancels \a conflict lock and waits until it reached CLS_FREEING state. This - * is called as a part of enqueuing to cancel conflicting locks early. - * - * \retval 0: success, \a conflict was cancelled and destroyed. - * - * \retval CLO_REPEAT: \a conflict was cancelled, but \a lock mutex was - * released in the process. Repeat enqueing. - * - * \retval -EWOULDBLOCK: \a conflict cannot be cancelled immediately, and - * either \a lock is non-blocking, or current thread - * holds other locks, that prevent it from waiting - * for cancel to complete. - * - * \retval -ve: other error, including -EINTR. - * - */ -static int osc_lock_cancel_wait(const struct lu_env *env, struct cl_lock *lock, - struct cl_lock *conflict, int canwait) -{ - int rc; - - LASSERT(cl_lock_is_mutexed(lock)); - LASSERT(cl_lock_is_mutexed(conflict)); - - rc = 0; - if (conflict->cll_state != CLS_FREEING) { - cl_lock_cancel(env, conflict); - rc = osc_lock_delete0(conflict); - if (rc) - return rc; - if (conflict->cll_flags & (CLF_CANCELPEND|CLF_DOOMED)) { - rc = -EWOULDBLOCK; - if (cl_lock_nr_mutexed(env) > 2) - /* - * If mutices of locks other than @lock and - * @scan are held by the current thread, it - * cannot wait on @scan state change in a - * dead-lock safe matter, so simply skip early - * cancellation in this case. - * - * This means that early cancellation doesn't - * work when there is even slight mutex - * contention, as top-lock's mutex is usually - * held at this time. - */ - ; - else if (canwait) { - /* Waiting for @scan to be destroyed */ - cl_lock_mutex_put(env, lock); - do { - rc = cl_lock_state_wait(env, conflict); - } while (!rc && - conflict->cll_state < CLS_FREEING); - /* mutex was released, repeat enqueue. */ - rc = rc ?: CLO_REPEAT; - cl_lock_mutex_get(env, lock); - } - } - LASSERT(ergo(!rc, conflict->cll_state == CLS_FREEING)); - CDEBUG(D_INFO, "lock %p was %s freed now, rc (%d)\n", - conflict, rc ? "not":"", rc); - } - return rc; -} - /** * Determine if the lock should be converted into a lockless lock. * @@ -1052,14 +1027,14 @@ static void osc_lock_to_lockless(const struct lu_env *env, io->ci_lockreq == CILR_NEVER); ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data; - ols->ols_locklessable = (io->ci_type != CIT_TRUNC) && + ols->ols_locklessable = (io->ci_type != CIT_SETATTR) && (io->ci_lockreq == CILR_MAYBE) && (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK); if (io->ci_lockreq == CILR_NEVER || /* lockless IO */ (ols->ols_locklessable && osc_object_is_contended(oob)) || /* lockless truncate */ - (io->ci_type == CIT_TRUNC && + (cl_io_is_trunc(io) && (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) && osd->od_lockless_truncate)) { ols->ols_locklessable = 1; @@ -1069,6 +1044,21 @@ static void osc_lock_to_lockless(const struct lu_env *env, LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols))); } +static int osc_lock_compatible(const struct osc_lock *qing, + const struct osc_lock *qed) +{ + enum cl_lock_mode qing_mode; + enum cl_lock_mode qed_mode; + + qing_mode = qing->ols_cl.cls_lock->cll_descr.cld_mode; + if (qed->ols_glimpse && + (qed->ols_state >= OLS_UPCALL_RECEIVED || qing_mode == CLM_READ)) + return 1; + + qed_mode = qed->ols_cl.cls_lock->cll_descr.cld_mode; + return ((qing_mode == CLM_READ) && (qed_mode == CLM_READ)); +} + /** * Cancel all conflicting locks and wait for them to be destroyed. * @@ -1086,36 +1076,31 @@ static int osc_lock_enqueue_wait(const struct lu_env *env, struct cl_lock *lock = olck->ols_cl.cls_lock; struct cl_lock_descr *descr = &lock->cll_descr; struct cl_object_header *hdr = cl_object_header(descr->cld_obj); - struct cl_lock_closure *closure = &osc_env_info(env)->oti_closure; struct cl_lock *scan; - struct cl_lock *temp; + struct cl_lock *conflict= NULL; int lockless = osc_lock_is_lockless(olck); int rc = 0; - int canwait; - int stop; ENTRY; LASSERT(cl_lock_is_mutexed(lock)); - LASSERT(lock->cll_state == CLS_QUEUING); - /* - * XXX This function could be sped up if we had asynchronous - * cancellation. - */ + /* make it enqueue anyway for glimpse lock, because we actually + * don't need to cancel any conflicting locks. */ + if (olck->ols_glimpse) + return 0; + + cfs_spin_lock(&hdr->coh_lock_guard); + cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) { + struct cl_lock_descr *cld = &scan->cll_descr; + const struct osc_lock *scan_ols; - canwait = - !(olck->ols_flags & LDLM_FL_BLOCK_NOWAIT) && - cl_lock_nr_mutexed(env) == 1; - cl_lock_closure_init(env, closure, lock, canwait); - spin_lock(&hdr->coh_lock_guard); - list_for_each_entry_safe(scan, temp, &hdr->coh_locks, cll_linkage) { if (scan == lock) - continue; + break; if (scan->cll_state < CLS_QUEUING || scan->cll_state == CLS_FREEING || - scan->cll_descr.cld_start > descr->cld_end || - scan->cll_descr.cld_end < descr->cld_start) + cld->cld_start > descr->cld_end || + cld->cld_end < descr->cld_start) continue; /* overlapped and living locks. */ @@ -1127,118 +1112,52 @@ static int osc_lock_enqueue_wait(const struct lu_env *env, continue; } - /* A tricky case for lockless pages: - * We need to cancel the compatible locks if we're enqueuing + scan_ols = osc_lock_at(scan); + + /* We need to cancel the compatible locks if we're enqueuing * a lockless lock, for example: * imagine that client has PR lock on [0, 1000], and thread T0 * is doing lockless IO in [500, 1500] region. Concurrent * thread T1 can see lockless data in [500, 1000], which is - * wrong, because these data are possibly stale. - */ - if (!lockless && cl_lock_compatible(scan, lock)) + * wrong, because these data are possibly stale. */ + if (!lockless && osc_lock_compatible(olck, scan_ols)) continue; - /* Now @scan is conflicting with @lock, this means current - * thread have to sleep for @scan being destroyed. */ cl_lock_get_trust(scan); - if (&temp->cll_linkage != &hdr->coh_locks) - cl_lock_get_trust(temp); - spin_unlock(&hdr->coh_lock_guard); - lu_ref_add(&scan->cll_reference, "cancel-wait", lock); - - LASSERT(list_empty(&closure->clc_list)); - rc = cl_lock_closure_build(env, scan, closure); - if (rc == 0) { - rc = osc_lock_cancel_wait(env, lock, scan, canwait); - cl_lock_disclosure(env, closure); - if (rc == -EWOULDBLOCK) - rc = 0; - } - if (rc == CLO_REPEAT && !canwait) - /* cannot wait... no early cancellation. */ - rc = 0; - - lu_ref_del(&scan->cll_reference, "cancel-wait", lock); - cl_lock_put(env, scan); - spin_lock(&hdr->coh_lock_guard); - /* - * Lock list could have been modified, while spin-lock was - * released. Check that it is safe to continue. - */ - stop = list_empty(&temp->cll_linkage); - if (&temp->cll_linkage != &hdr->coh_locks) - cl_lock_put(env, temp); - if (stop || rc != 0) - break; + conflict = scan; + break; } - spin_unlock(&hdr->coh_lock_guard); - cl_lock_closure_fini(closure); - RETURN(rc); -} - -/** - * Deadlock avoidance for osc_lock_enqueue(). Consider following scenario: - * - * - Thread0: obtains PR:[0, 10]. Lock is busy. - * - * - Thread1: enqueues PW:[5, 50]. Blocking ast is sent to - * PR:[0, 10], but cancellation of busy lock is postponed. - * - * - Thread0: enqueue PR:[30, 40]. Lock is locally matched to - * PW:[5, 50], and thread0 waits for the lock completion never - * releasing PR:[0, 10]---deadlock. - * - * The second PR lock can be glimpse (it is to deal with that situation that - * ll_glimpse_size() has second argument, preventing local match of - * not-yet-granted locks, see bug 10295). Similar situation is possible in the - * case of memory mapped user level buffer. - * - * To prevent this we can detect a situation when current "thread" or "io" - * already holds a lock on this object and either add LDLM_FL_BLOCK_GRANTED to - * the ols->ols_flags, or prevent local match with PW locks. - */ -static int osc_deadlock_is_possible(const struct lu_env *env, - struct cl_lock *lock) -{ - struct cl_object *obj; - struct cl_object_header *head; - struct cl_lock *scan; - struct osc_io *oio; - - int result; - - ENTRY; + cfs_spin_unlock(&hdr->coh_lock_guard); - LASSERT(cl_lock_is_mutexed(lock)); - - oio = osc_env_io(env); - obj = lock->cll_descr.cld_obj; - head = cl_object_header(obj); - - result = 0; - spin_lock(&head->coh_lock_guard); - list_for_each_entry(scan, &head->coh_locks, cll_linkage) { - if (scan != lock) { - struct osc_lock *oscan; - - oscan = osc_lock_at(scan); - LASSERT(oscan != NULL); - if (oscan->ols_owner == oio) { - result = 1; - break; - } + if (conflict) { + if (lock->cll_descr.cld_mode == CLM_GROUP) { + /* we want a group lock but a previous lock request + * conflicts, we do not wait but return 0 so the + * request is send to the server + */ + CDEBUG(D_DLMTRACE, "group lock %p is conflicted " + "with %p, no wait, send to server\n", + lock, conflict); + cl_lock_put(env, conflict); + rc = 0; + } else { + CDEBUG(D_DLMTRACE, "lock %p is conflicted with %p, " + "will wait\n", + lock, conflict); + LASSERT(lock->cll_conflict == NULL); + lu_ref_add(&conflict->cll_reference, "cancel-wait", + lock); + lock->cll_conflict = conflict; + rc = CLO_WAIT; } } - spin_unlock(&head->coh_lock_guard); - RETURN(result); + RETURN(rc); } /** * Implementation of cl_lock_operations::clo_enqueue() method for osc * layer. This initiates ldlm enqueue: * - * - checks for possible dead-lock conditions (osc_deadlock_is_possible()); - * * - cancels conflicting locks early (osc_lock_enqueue_wait()); * * - calls osc_enqueue_base() to do actual enqueue. @@ -1255,38 +1174,43 @@ static int osc_lock_enqueue(const struct lu_env *env, { struct osc_lock *ols = cl2osc_lock(slice); struct cl_lock *lock = ols->ols_cl.cls_lock; - struct osc_object *obj = cl2osc(slice->cls_obj); - struct osc_thread_info *info = osc_env_info(env); - struct ldlm_res_id *resname = &info->oti_resname; - ldlm_policy_data_t *policy = &info->oti_policy; - struct ldlm_enqueue_info *einfo = &ols->ols_einfo; int result; ENTRY; LASSERT(cl_lock_is_mutexed(lock)); - LASSERT(lock->cll_state == CLS_QUEUING); - LASSERT(ols->ols_state == OLS_NEW); + LASSERTF(ols->ols_state == OLS_NEW, + "Impossible state: %d\n", ols->ols_state); - osc_lock_build_res(env, obj, resname); - osc_lock_build_policy(env, lock, policy); ols->ols_flags = osc_enq2ldlm_flags(enqflags); - if (osc_deadlock_is_possible(env, lock)) - ols->ols_flags |= LDLM_FL_BLOCK_GRANTED; + if (enqflags & CEF_AGL) { + ols->ols_flags |= LDLM_FL_BLOCK_NOWAIT; + ols->ols_agl = 1; + } else { + ols->ols_agl = 0; + } if (ols->ols_flags & LDLM_FL_HAS_INTENT) ols->ols_glimpse = 1; + if (!osc_lock_is_lockless(ols) && !(enqflags & CEF_MUST)) + /* try to convert this lock to a lockless lock */ + osc_lock_to_lockless(env, ols, (enqflags & CEF_NEVER)); result = osc_lock_enqueue_wait(env, ols); if (result == 0) { - if (!(enqflags & CEF_MUST)) - /* try to convert this lock to a lockless lock */ - osc_lock_to_lockless(env, ols, (enqflags & CEF_NEVER)); if (!osc_lock_is_lockless(ols)) { + struct osc_object *obj = cl2osc(slice->cls_obj); + struct osc_thread_info *info = osc_env_info(env); + struct ldlm_res_id *resname = &info->oti_resname; + ldlm_policy_data_t *policy = &info->oti_policy; + struct ldlm_enqueue_info *einfo = &ols->ols_einfo; + if (ols->ols_locklessable) ols->ols_flags |= LDLM_FL_DENY_ON_CONTENTION; - /* a reference for lock, passed as an upcall cookie */ - cl_lock_get(lock); - lu_ref_add(&lock->cll_reference, "upcall", lock); + /* lock will be passed as upcall cookie, + * hold ref to prevent to be released. */ + cl_lock_hold_add(env, lock, "upcall", lock); + /* a user for lock also */ + cl_lock_user_add(env, lock); ols->ols_state = OLS_ENQUEUED; /* @@ -1294,20 +1218,26 @@ static int osc_lock_enqueue(const struct lu_env *env, * ldlm_lock_match(LDLM_FL_LVB_READY) waits for * LDLM_CP_CALLBACK. */ + osc_lock_build_res(env, obj, resname); + osc_lock_build_policy(env, lock, policy); result = osc_enqueue_base(osc_export(obj), resname, &ols->ols_flags, policy, &ols->ols_lvb, obj->oo_oinfo->loi_kms_valid, osc_lock_upcall, ols, einfo, &ols->ols_handle, - PTLRPCD_SET, 1); + PTLRPCD_SET, 1, ols->ols_agl); if (result != 0) { - lu_ref_del(&lock->cll_reference, - "upcall", lock); - cl_lock_put(env, lock); + cl_lock_user_del(env, lock); + cl_lock_unhold(env, lock, "upcall", lock); + if (unlikely(result == -ECANCELED)) { + ols->ols_state = OLS_NEW; + result = 0; + } } } else { ols->ols_state = OLS_GRANTED; + ols->ols_owner = osc_env_io(env); } } LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols))); @@ -1321,8 +1251,34 @@ static int osc_lock_wait(const struct lu_env *env, struct cl_lock *lock = olck->ols_cl.cls_lock; LINVRNT(osc_lock_invariant(olck)); - if (olck->ols_glimpse && olck->ols_state >= OLS_UPCALL_RECEIVED) - return 0; + + if (olck->ols_glimpse && olck->ols_state >= OLS_UPCALL_RECEIVED) { + if (olck->ols_flags & LDLM_FL_LVB_READY) { + return 0; + } else if (olck->ols_agl) { + if (lock->cll_flags & CLF_FROM_UPCALL) + /* It is from enqueue RPC reply upcall for + * updating state. Do not re-enqueue. */ + return -ENAVAIL; + else + olck->ols_state = OLS_NEW; + } else { + LASSERT(lock->cll_error); + return lock->cll_error; + } + } + + if (olck->ols_state == OLS_NEW) { + int rc; + + LASSERT(olck->ols_agl); + + rc = osc_lock_enqueue(env, slice, NULL, CEF_ASYNC | CEF_MUST); + if (rc != 0) + return rc; + else + return CLO_REENQUEUED; + } LASSERT(equi(olck->ols_state >= OLS_UPCALL_RECEIVED && lock->cll_error == 0, olck->ols_lock != NULL)); @@ -1341,13 +1297,14 @@ static int osc_lock_use(const struct lu_env *env, int rc; LASSERT(!olck->ols_hold); + /* * Atomically check for LDLM_FL_CBPENDING and addref a lock if this * flag is not set. This protects us from a concurrent blocking ast. */ rc = ldlm_lock_addref_try(&olck->ols_handle, olck->ols_einfo.ei_mode); if (rc == 0) { - olck->ols_hold = olck->ols_has_ref = 1; + olck->ols_hold = 1; olck->ols_state = OLS_GRANTED; } else { struct cl_lock *lock; @@ -1359,7 +1316,7 @@ static int osc_lock_use(const struct lu_env *env, * cl_lock mutex. */ lock = slice->cls_lock; - LASSERT(lock->cll_state == CLS_CACHED); + LASSERT(lock->cll_state == CLS_INTRANSIT); LASSERT(lock->cll_users > 0); /* set a flag for osc_dlm_blocking_ast0() to signal the * lock.*/ @@ -1371,14 +1328,32 @@ static int osc_lock_use(const struct lu_env *env, static int osc_lock_flush(struct osc_lock *ols, int discard) { - struct cl_lock *lock = ols->ols_cl.cls_lock; - struct cl_env_nest nest; - struct lu_env *env; - int result = 0; + struct cl_lock *lock = ols->ols_cl.cls_lock; + struct cl_env_nest nest; + struct lu_env *env; + int result = 0; + ENTRY; + + env = cl_env_nested_get(&nest); + if (!IS_ERR(env)) { + struct osc_object *obj = cl2osc(ols->ols_cl.cls_obj); + struct cl_lock_descr *descr = &lock->cll_descr; + int rc = 0; + + if (descr->cld_mode >= CLM_WRITE) { + result = osc_cache_writeback_range(env, obj, + descr->cld_start, descr->cld_end, + 1, discard); + CDEBUG(D_DLMTRACE, "write out %d pages for lock %p.\n", + result, lock); + if (result > 0) + result = 0; + } + + rc = cl_lock_discard_pages(env, lock); + if (result == 0 && rc < 0) + result = rc; - env = cl_env_nested_get(&nest); - if (!IS_ERR(env)) { - result = cl_lock_page_out(env, lock, discard); cl_env_nested_put(&nest, env); } else result = PTR_ERR(env); @@ -1386,7 +1361,7 @@ static int osc_lock_flush(struct osc_lock *ols, int discard) ols->ols_flush = 1; LINVRNT(!osc_lock_has_pages(ols)); } - return result; + RETURN(result); } /** @@ -1418,10 +1393,9 @@ static void osc_lock_cancel(const struct lu_env *env, if (dlmlock != NULL) { int do_cancel; - discard = dlmlock->l_flags & LDLM_FL_DISCARD_DATA; + discard = !!(dlmlock->l_flags & LDLM_FL_DISCARD_DATA); result = osc_lock_flush(olck, discard); - if (olck->ols_hold) - osc_lock_unuse(env, slice); + osc_lock_unhold(olck); lock_res_and_lock(dlmlock); /* Now that we're the only user of dlm read/write reference, @@ -1440,14 +1414,31 @@ static void osc_lock_cancel(const struct lu_env *env, lock, result); } olck->ols_state = OLS_CANCELLED; + olck->ols_flags &= ~LDLM_FL_LVB_READY; osc_lock_detach(env, olck); } -void cl_lock_page_list_fixup(const struct lu_env *env, - struct cl_io *io, struct cl_lock *lock, - struct cl_page_list *queue); - #ifdef INVARIANT_CHECK +static int check_cb(const struct lu_env *env, struct cl_io *io, + struct cl_page *page, void *cbdata) +{ + struct cl_lock *lock = cbdata; + + if (lock->cll_descr.cld_mode == CLM_READ) { + struct cl_lock *tmp; + tmp = cl_lock_at_page(env, lock->cll_descr.cld_obj, + page, lock, 1, 0); + if (tmp != NULL) { + cl_lock_put(env, tmp); + return CLP_GANG_OKAY; + } + } + + CL_LOCK_DEBUG(D_ERROR, env, lock, "still has pages\n"); + CL_PAGE_DEBUG(D_ERROR, env, page, "\n"); + return CLP_GANG_ABORT; +} + /** * Returns true iff there are pages under \a olck not protected by other * locks. @@ -1458,44 +1449,40 @@ static int osc_lock_has_pages(struct osc_lock *olck) struct cl_lock_descr *descr; struct cl_object *obj; struct osc_object *oob; - struct cl_page_list *plist; - struct cl_page *page; struct cl_env_nest nest; struct cl_io *io; struct lu_env *env; int result; env = cl_env_nested_get(&nest); - if (!IS_ERR(env)) { - obj = olck->ols_cl.cls_obj; - oob = cl2osc(obj); - io = &oob->oo_debug_io; - lock = olck->ols_cl.cls_lock; - descr = &lock->cll_descr; - plist = &osc_env_info(env)->oti_plist; - cl_page_list_init(plist); - - mutex_lock(&oob->oo_debug_mutex); - - io->ci_obj = cl_object_top(obj); - cl_io_init(env, io, CIT_MISC, io->ci_obj); - cl_page_gang_lookup(env, obj, io, - descr->cld_start, descr->cld_end, plist); - cl_lock_page_list_fixup(env, io, lock, plist); - if (plist->pl_nr > 0) { - CL_LOCK_DEBUG(D_ERROR, env, lock, "still has pages\n"); - cl_page_list_for_each(page, plist) - CL_PAGE_DEBUG(D_ERROR, env, page, "\n"); - } - result = plist->pl_nr > 0; - cl_page_list_disown(env, io, plist); - cl_page_list_fini(env, plist); - cl_io_fini(env, io); - mutex_unlock(&oob->oo_debug_mutex); - cl_env_nested_put(&nest, env); - } else - result = 0; - return result; + if (IS_ERR(env)) + return 0; + + obj = olck->ols_cl.cls_obj; + oob = cl2osc(obj); + io = &oob->oo_debug_io; + lock = olck->ols_cl.cls_lock; + descr = &lock->cll_descr; + + cfs_mutex_lock(&oob->oo_debug_mutex); + + io->ci_obj = cl_object_top(obj); + io->ci_ignore_layout = 1; + cl_io_init(env, io, CIT_MISC, io->ci_obj); + do { + result = cl_page_gang_lookup(env, obj, io, + descr->cld_start, descr->cld_end, + check_cb, (void *)lock); + if (result == CLP_GANG_ABORT) + break; + if (result == CLP_GANG_RESCHED) + cfs_cond_resched(); + } while (result != CLP_GANG_OKAY); + cl_io_fini(env, io); + cfs_mutex_unlock(&oob->oo_debug_mutex); + cl_env_nested_put(&nest, env); + + return (result == CLP_GANG_ABORT); } #else static int osc_lock_has_pages(struct osc_lock *olck) @@ -1519,8 +1506,7 @@ static void osc_lock_delete(const struct lu_env *env, LINVRNT(osc_lock_invariant(olck)); LINVRNT(!osc_lock_has_pages(olck)); - if (olck->ols_hold) - osc_lock_unuse(env, slice); + osc_lock_unhold(olck); osc_lock_detach(env, olck); } @@ -1539,13 +1525,14 @@ static void osc_lock_state(const struct lu_env *env, enum cl_lock_state state) { struct osc_lock *lock = cl2osc_lock(slice); - struct osc_io *oio = osc_env_io(env); /* * XXX multiple io contexts can use the lock at the same time. */ LINVRNT(osc_lock_invariant(lock)); if (state == CLS_HELD && slice->cls_lock->cll_state != CLS_HELD) { + struct osc_io *oio = osc_env_io(env); + LASSERT(lock->ols_owner == NULL); lock->ols_owner = oio; } else if (state != CLS_HELD) @@ -1560,13 +1547,60 @@ static int osc_lock_print(const struct lu_env *env, void *cookie, /* * XXX print ldlm lock and einfo properly. */ - (*p)(env, cookie, "%p %08x "LPU64" %d %p ", + (*p)(env, cookie, "%p %08x "LPX64" %d %p ", lock->ols_lock, lock->ols_flags, lock->ols_handle.cookie, lock->ols_state, lock->ols_owner); osc_lvb_print(env, cookie, p, &lock->ols_lvb); return 0; } +static int osc_lock_fits_into(const struct lu_env *env, + const struct cl_lock_slice *slice, + const struct cl_lock_descr *need, + const struct cl_io *io) +{ + struct osc_lock *ols = cl2osc_lock(slice); + + if (need->cld_enq_flags & CEF_NEVER) + return 0; + + if (ols->ols_state >= OLS_CANCELLED) + return 0; + + if (need->cld_mode == CLM_PHANTOM) { + if (ols->ols_agl) + return !(ols->ols_state > OLS_RELEASED); + + /* + * Note: the QUEUED lock can't be matched here, otherwise + * it might cause the deadlocks. + * In read_process, + * P1: enqueued read lock, create sublock1 + * P2: enqueued write lock, create sublock2(conflicted + * with sublock1). + * P1: Grant read lock. + * P1: enqueued glimpse lock(with holding sublock1_read), + * matched with sublock2, waiting sublock2 to be granted. + * But sublock2 can not be granted, because P1 + * will not release sublock1. Bang! + */ + if (ols->ols_state < OLS_GRANTED || + ols->ols_state > OLS_RELEASED) + return 0; + } else if (need->cld_enq_flags & CEF_MUST) { + /* + * If the lock hasn't ever enqueued, it can't be matched + * because enqueue process brings in many information + * which can be used to determine things such as lockless, + * CEF_MUST, etc. + */ + if (ols->ols_state < OLS_UPCALL_RECEIVED && + ols->ols_locklessable) + return 0; + } + return 1; +} + static const struct cl_lock_operations osc_lock_ops = { .clo_fini = osc_lock_fini, .clo_enqueue = osc_lock_enqueue, @@ -1577,17 +1611,10 @@ static const struct cl_lock_operations osc_lock_ops = { .clo_state = osc_lock_state, .clo_cancel = osc_lock_cancel, .clo_weigh = osc_lock_weigh, - .clo_print = osc_lock_print + .clo_print = osc_lock_print, + .clo_fits_into = osc_lock_fits_into, }; -static int osc_lock_lockless_enqueue(const struct lu_env *env, - const struct cl_lock_slice *slice, - struct cl_io *unused, __u32 enqflags) -{ - LBUG(); - return 0; -} - static int osc_lock_lockless_unuse(const struct lu_env *env, const struct cl_lock_slice *slice) { @@ -1632,19 +1659,19 @@ static void osc_lock_lockless_state(const struct lu_env *env, enum cl_lock_state state) { struct osc_lock *lock = cl2osc_lock(slice); - struct osc_io *oio = osc_env_io(env); LINVRNT(osc_lock_invariant(lock)); if (state == CLS_HELD) { - LASSERT(lock->ols_owner == NULL); + struct osc_io *oio = osc_env_io(env); + + LASSERT(ergo(lock->ols_owner, lock->ols_owner == oio)); lock->ols_owner = oio; /* set the io to be lockless if this lock is for io's * host object */ if (cl_object_same(oio->oi_cl.cis_obj, slice->cls_obj)) oio->oi_lockless = 1; - } else - lock->ols_owner = NULL; + } } static int osc_lock_lockless_fits_into(const struct lu_env *env, @@ -1652,12 +1679,18 @@ static int osc_lock_lockless_fits_into(const struct lu_env *env, const struct cl_lock_descr *need, const struct cl_io *io) { - return 0; + struct osc_lock *lock = cl2osc_lock(slice); + + if (!(need->cld_enq_flags & CEF_NEVER)) + return 0; + + /* lockless lock should only be used by its owning io. b22147 */ + return (lock->ols_owner == osc_env_io(env)); } static const struct cl_lock_operations osc_lock_lockless_ops = { .clo_fini = osc_lock_fini, - .clo_enqueue = osc_lock_lockless_enqueue, + .clo_enqueue = osc_lock_enqueue, .clo_wait = osc_lock_lockless_wait, .clo_unuse = osc_lock_lockless_unuse, .clo_state = osc_lock_lockless_state, @@ -1676,6 +1709,7 @@ int osc_lock_init(const struct lu_env *env, OBD_SLAB_ALLOC_PTR_GFP(clk, osc_lock_kmem, CFS_ALLOC_IO); if (clk != NULL) { osc_lock_build_einfo(env, lock, clk, &clk->ols_einfo); + cfs_atomic_set(&clk->ols_pageref, 0); clk->ols_state = OLS_NEW; cl_lock_slice_add(lock, &clk->ols_cl, obj, &osc_lock_ops); result = 0; @@ -1684,5 +1718,26 @@ int osc_lock_init(const struct lu_env *env, return result; } +int osc_dlm_lock_pageref(struct ldlm_lock *dlm) +{ + struct osc_lock *olock; + int rc = 0; + + cfs_spin_lock(&osc_ast_guard); + olock = dlm->l_ast_data; + /* + * there's a very rare race with osc_page_addref_lock(), but that + * doesn't matter because in the worst case we don't cancel a lock + * which we actually can, that's no harm. + */ + if (olock != NULL && + cfs_atomic_add_return(_PAGEREF_MAGIC, + &olock->ols_pageref) != _PAGEREF_MAGIC) { + cfs_atomic_sub(_PAGEREF_MAGIC, &olock->ols_pageref); + rc = 1; + } + cfs_spin_unlock(&osc_ast_guard); + return rc; +} /** @} osc */