diff options
Diffstat (limited to 'drivers/staging/lustre/lustre/osc/osc_lock.c')
| -rw-r--r-- | drivers/staging/lustre/lustre/osc/osc_lock.c | 1663 |
1 files changed, 1663 insertions, 0 deletions
diff --git a/drivers/staging/lustre/lustre/osc/osc_lock.c b/drivers/staging/lustre/lustre/osc/osc_lock.c new file mode 100644 index 000000000000..640bc3d34709 --- /dev/null +++ b/drivers/staging/lustre/lustre/osc/osc_lock.c @@ -0,0 +1,1663 @@ +/* + * GPL HEADER START + * + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 only, + * as published by the Free Software Foundation. + * + * This program 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 version 2 for more details (a copy is included + * in the LICENSE file that accompanied this code). + * + * You should have received a copy of the GNU General Public License + * version 2 along with this program; If not, see + * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf + * + * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, + * CA 95054 USA or visit www.sun.com if you need additional information or + * have any questions. + * + * GPL HEADER END + */ +/* + * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + * + * Copyright (c) 2011, 2012, Intel Corporation. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. + * + * Implementation of cl_lock for OSC layer. + * + * Author: Nikita Danilov <nikita.danilov@sun.com> + */ + +#define DEBUG_SUBSYSTEM S_OSC + +# include <linux/libcfs/libcfs.h> +/* fid_build_reg_res_name() */ +#include <lustre_fid.h> + +#include "osc_cl_internal.h" + +/** \addtogroup osc + * @{ + */ + +#define _PAGEREF_MAGIC (-10000000) + +/***************************************************************************** + * + * Type conversions. + * + */ + +static const struct cl_lock_operations osc_lock_ops; +static const struct cl_lock_operations osc_lock_lockless_ops; +static void osc_lock_to_lockless(const struct lu_env *env, + struct osc_lock *ols, int force); +static int osc_lock_has_pages(struct osc_lock *olck); + +int osc_lock_is_lockless(const struct osc_lock *olck) +{ + return (olck->ols_cl.cls_ops == &osc_lock_lockless_ops); +} + +/** + * Returns a weak pointer to the ldlm lock identified by a handle. Returned + * pointer cannot be dereferenced, as lock is not protected from concurrent + * reclaim. This function is a helper for osc_lock_invariant(). + */ +static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle) +{ + struct ldlm_lock *lock; + + lock = ldlm_handle2lock(handle); + if (lock != NULL) + LDLM_LOCK_PUT(lock); + return lock; +} + +/** + * Invariant that has to be true all of the time. + */ +static int osc_lock_invariant(struct osc_lock *ols) +{ + struct ldlm_lock *lock = osc_handle_ptr(&ols->ols_handle); + struct ldlm_lock *olock = ols->ols_lock; + int handle_used = lustre_handle_is_used(&ols->ols_handle); + + return + ergo(osc_lock_is_lockless(ols), + ols->ols_locklessable && ols->ols_lock == NULL) || + (ergo(olock != NULL, handle_used) && + ergo(olock != NULL, + olock->l_handle.h_cookie == ols->ols_handle.cookie) && + /* + * Check that ->ols_handle and ->ols_lock are consistent, but + * take into account that they are set at the different time. + */ + ergo(handle_used, + ergo(lock != NULL && olock != NULL, lock == olock) && + ergo(lock == NULL, olock == NULL)) && + ergo(ols->ols_state == OLS_CANCELLED, + olock == NULL && !handle_used) && + /* + * DLM lock is destroyed only after we have seen cancellation + * ast. + */ + ergo(olock != NULL && ols->ols_state < OLS_CANCELLED, + !olock->l_destroyed) && + ergo(ols->ols_state == OLS_GRANTED, + olock != NULL && + olock->l_req_mode == olock->l_granted_mode && + ols->ols_hold)); +} + +/***************************************************************************** + * + * Lock operations. + * + */ + +/** + * Breaks a link between osc_lock and dlm_lock. + */ +static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck) +{ + struct ldlm_lock *dlmlock; + + spin_lock(&osc_ast_guard); + dlmlock = olck->ols_lock; + if (dlmlock == NULL) { + spin_unlock(&osc_ast_guard); + return; + } + + olck->ols_lock = NULL; + /* wb(); --- for all who checks (ols->ols_lock != NULL) before + * call to osc_lock_detach() */ + dlmlock->l_ast_data = NULL; + olck->ols_handle.cookie = 0ULL; + 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; + + 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); + + cl_object_attr_set(env, obj, attr, CAT_KMS); + cl_object_attr_unlock(obj); + } + 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); + + LINVRNT(osc_lock_invariant(ols)); + + 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(); + } +} + +static void osc_lock_fini(const struct lu_env *env, + struct cl_lock_slice *slice) +{ + struct osc_lock *ols = cl2osc_lock(slice); + + LINVRNT(osc_lock_invariant(ols)); + /* + * ->ols_hold can still be true at this point if, for example, a + * thread that requested a lock was killed (and released a reference + * to the lock), before reply from a server was received. In this case + * lock is destroyed immediately after upcall. + */ + osc_lock_unhold(ols); + LASSERT(ols->ols_lock == NULL); + LASSERT(atomic_read(&ols->ols_pageref) == 0 || + atomic_read(&ols->ols_pageref) == _PAGEREF_MAGIC); + + OBD_SLAB_FREE_PTR(ols, osc_lock_kmem); +} + +static void osc_lock_build_policy(const struct lu_env *env, + const struct cl_lock *lock, + ldlm_policy_data_t *policy) +{ + const struct cl_lock_descr *d = &lock->cll_descr; + + osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end); + policy->l_extent.gid = d->cld_gid; +} + +static __u64 osc_enq2ldlm_flags(__u32 enqflags) +{ + __u64 result = 0; + + LASSERT((enqflags & ~CEF_MASK) == 0); + + if (enqflags & CEF_NONBLOCK) + result |= LDLM_FL_BLOCK_NOWAIT; + if (enqflags & CEF_ASYNC) + result |= LDLM_FL_HAS_INTENT; + if (enqflags & CEF_DISCARD_DATA) + result |= LDLM_AST_DISCARD_DATA; + return result; +} + +/** + * Global spin-lock protecting consistency of ldlm_lock::l_ast_data + * pointers. Initialized in osc_init(). + */ +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); + olck = dlm_lock->l_ast_data; + if (olck != NULL) { + struct cl_lock *lock = olck->ols_cl.cls_lock; + /* + * If osc_lock holds a reference on ldlm lock, return it even + * when cl_lock is in CLS_FREEING state. This way + * + * osc_ast_data_get(dlmlock) == NULL + * + * guarantees that all osc references on dlmlock were + * released. osc_dlm_blocking_ast0() relies on that. + */ + if (lock->cll_state < CLS_FREEING || olck->ols_has_ref) { + cl_lock_get_trust(lock); + lu_ref_add_atomic(&lock->cll_reference, + "ast", current); + } else + olck = NULL; + } + spin_unlock(&osc_ast_guard); + unlock_res_and_lock(dlm_lock); + return olck; +} + +static void osc_ast_data_put(const struct lu_env *env, struct osc_lock *olck) +{ + struct cl_lock *lock; + + lock = olck->ols_cl.cls_lock; + lu_ref_del(&lock->cll_reference, "ast", current); + cl_lock_put(env, lock); +} + +/** + * Updates object attributes from a lock value block (lvb) received together + * with the DLM lock reply from the server. Copy of osc_update_enqueue() + * logic. + * + * This can be optimized to not update attributes when lock is a result of a + * local match. + * + * Called under lock and resource spin-locks. + */ +static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck, + int rc) +{ + struct ost_lvb *lvb; + struct cl_object *obj; + struct lov_oinfo *oinfo; + struct cl_attr *attr; + unsigned valid; + + ENTRY; + + if (!(olck->ols_flags & LDLM_FL_LVB_READY)) + RETURN_EXIT; + + lvb = &olck->ols_lvb; + obj = olck->ols_cl.cls_obj; + oinfo = cl2osc(obj)->oo_oinfo; + attr = &osc_env_info(env)->oti_attr; + valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE; + cl_lvb2attr(attr, lvb); + + cl_object_attr_lock(obj); + if (rc == 0) { + struct ldlm_lock *dlmlock; + __u64 size; + + dlmlock = olck->ols_lock; + LASSERT(dlmlock != NULL); + + /* re-grab LVB from a dlm lock under DLM spin-locks. */ + *lvb = *(struct ost_lvb *)dlmlock->l_lvb_data; + size = lvb->lvb_size; + /* Extend KMS up to the end of this lock and no further + * A lock on [x,y] means a KMS of up to y + 1 bytes! */ + if (size > dlmlock->l_policy_data.l_extent.end) + size = dlmlock->l_policy_data.l_extent.end + 1; + if (size >= oinfo->loi_kms) { + LDLM_DEBUG(dlmlock, "lock acquired, setting rss="LPU64 + ", kms="LPU64, lvb->lvb_size, size); + valid |= CAT_KMS; + attr->cat_kms = size; + } else { + LDLM_DEBUG(dlmlock, "lock acquired, setting rss=" + LPU64"; leaving kms="LPU64", end="LPU64, + lvb->lvb_size, oinfo->loi_kms, + dlmlock->l_policy_data.l_extent.end); + } + ldlm_lock_allow_match_locked(dlmlock); + } else if (rc == -ENAVAIL && olck->ols_glimpse) { + CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving" + " kms="LPU64"\n", lvb->lvb_size, oinfo->loi_kms); + } else + valid = 0; + + if (valid != 0) + cl_object_attr_set(env, obj, attr, valid); + + cl_object_attr_unlock(obj); + + EXIT; +} + +/** + * Called when a lock is granted, from an upcall (when server returned a + * granted lock), or from completion AST, when server returned a blocked lock. + * + * Called under lock and resource spin-locks, that are released temporarily + * here. + */ +static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck, + struct ldlm_lock *dlmlock, int rc) +{ + struct ldlm_extent *ext; + struct cl_lock *lock; + struct cl_lock_descr *descr; + + LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode); + + ENTRY; + 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; + descr->cld_obj = lock->cll_descr.cld_obj; + + /* XXX check that ->l_granted_mode is valid. */ + descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode); + descr->cld_start = cl_index(descr->cld_obj, ext->start); + descr->cld_end = cl_index(descr->cld_obj, ext->end); + descr->cld_gid = ext->gid; + /* + * tell upper layers the extent of the lock that was actually + * granted + */ + olck->ols_state = OLS_GRANTED; + osc_lock_lvb_update(env, olck, rc); + + /* release DLM spin-locks to allow cl_lock_{modify,signal}() + * to take a semaphore on a parent lock. This is safe, because + * spin-locks are needed to protect consistency of + * dlmlock->l_*_mode and LVB, and we have finished processing + * them. */ + unlock_res_and_lock(dlmlock); + cl_lock_modify(env, lock, descr); + cl_lock_signal(env, lock); + LINVRNT(osc_lock_invariant(olck)); + lock_res_and_lock(dlmlock); + } + EXIT; +} + +static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck) + +{ + struct ldlm_lock *dlmlock; + + ENTRY; + + dlmlock = ldlm_handle2lock_long(&olck->ols_handle, 0); + LASSERT(dlmlock != NULL); + + lock_res_and_lock(dlmlock); + 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); + + /* + * Lock might be not yet granted. In this case, completion ast + * (osc_ldlm_completion_ast()) comes later and finishes lock + * granting. + */ + if (dlmlock->l_granted_mode == dlmlock->l_req_mode) + osc_lock_granted(env, olck, dlmlock, 0); + unlock_res_and_lock(dlmlock); + + /* + * osc_enqueue_interpret() decrefs asynchronous locks, counter + * this. + */ + ldlm_lock_addref(&olck->ols_handle, olck->ols_einfo.ei_mode); + 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; +} + +/** + * Lock upcall function that is executed either when a reply to ENQUEUE rpc is + * received from a server, or after osc_enqueue_base() matched a local DLM + * lock. + */ +static int osc_lock_upcall(void *cookie, int errcode) +{ + struct osc_lock *olck = cookie; + struct cl_lock_slice *slice = &olck->ols_cl; + struct cl_lock *lock = slice->cls_lock; + struct lu_env *env; + struct cl_env_nest nest; + + ENTRY; + env = cl_env_nested_get(&nest); + if (!IS_ERR(env)) { + int rc; + + cl_lock_mutex_get(env, lock); + + LASSERT(lock->cll_state >= CLS_QUEUING); + if (olck->ols_state == OLS_ENQUEUED) { + olck->ols_state = OLS_UPCALL_RECEIVED; + rc = ldlm_error2errno(errcode); + } else if (olck->ols_state == OLS_CANCELLED) { + rc = -EIO; + } else { + CERROR("Impossible state: %d\n", olck->ols_state); + LBUG(); + } + if (rc) { + struct ldlm_lock *dlmlock; + + dlmlock = ldlm_handle2lock(&olck->ols_handle); + if (dlmlock != NULL) { + lock_res_and_lock(dlmlock); + 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); + ldlm_lock_fail_match_locked(dlmlock); + unlock_res_and_lock(dlmlock); + LDLM_LOCK_PUT(dlmlock); + } + } else { + if (olck->ols_glimpse) + olck->ols_glimpse = 0; + osc_lock_upcall0(env, olck); + } + + /* Error handling, some errors are tolerable. */ + if (olck->ols_locklessable && rc == -EUSERS) { + /* This is a tolerable error, turn this lock into + * lockless lock. + */ + osc_object_set_contended(cl2osc(slice->cls_obj)); + LASSERT(slice->cls_ops == &osc_lock_ops); + + /* Change this lock to ldlmlock-less lock. */ + osc_lock_to_lockless(env, olck, 1); + olck->ols_state = OLS_GRANTED; + rc = 0; + } else if (olck->ols_glimpse && rc == -ENAVAIL) { + osc_lock_lvb_update(env, olck, rc); + cl_lock_delete(env, lock); + /* Hide the error. */ + rc = 0; + } + + 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); + /* 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); + } + + /* 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); +} + +/** + * Core of osc_dlm_blocking_ast() logic. + */ +static void osc_lock_blocking(const struct lu_env *env, + struct ldlm_lock *dlmlock, + struct osc_lock *olck, int blocking) +{ + struct cl_lock *lock = olck->ols_cl.cls_lock; + + LASSERT(olck->ols_lock == dlmlock); + CLASSERT(OLS_BLOCKED < OLS_CANCELLED); + LASSERT(!osc_lock_is_lockless(olck)); + + /* + * 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) + /* + * Move osc_lock into OLS_BLOCKED before canceling the lock, + * because it recursively re-enters osc_lock_blocking(), with + * the state set to OLS_CANCELLED. + */ + olck->ols_state = OLS_BLOCKED; + /* + * cancel and destroy lock at least once no matter how blocking ast is + * entered (see comment above osc_ldlm_blocking_ast() for use + * cases). cl_lock_cancel() and cl_lock_delete() are idempotent. + */ + cl_lock_cancel(env, lock); + cl_lock_delete(env, lock); +} + +/** + * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock + * and ldlm_lock caches. + */ +static int osc_dlm_blocking_ast0(const struct lu_env *env, + struct ldlm_lock *dlmlock, + void *data, int flag) +{ + struct osc_lock *olck; + struct cl_lock *lock; + int result; + int cancel; + + LASSERT(flag == LDLM_CB_BLOCKING || flag == LDLM_CB_CANCELING); + + cancel = 0; + olck = osc_ast_data_get(dlmlock); + if (olck != NULL) { + lock = olck->ols_cl.cls_lock; + cl_lock_mutex_get(env, lock); + LINVRNT(osc_lock_invariant(olck)); + if (olck->ols_ast_wait) { + /* wake up osc_lock_use() */ + cl_lock_signal(env, lock); + olck->ols_ast_wait = 0; + } + /* + * Lock might have been canceled while this thread was + * sleeping for lock mutex, but olck is pinned in memory. + */ + if (olck == dlmlock->l_ast_data) { + /* + * NOTE: DLM sends blocking AST's for failed locks + * (that are still in pre-OLS_GRANTED state) + * too, and they have to be canceled otherwise + * DLM lock is never destroyed and stuck in + * the memory. + * + * Alternatively, ldlm_cli_cancel() can be + * called here directly for osc_locks with + * ols_state < OLS_GRANTED to maintain an + * invariant that ->clo_cancel() is only called + * for locks that were granted. + */ + LASSERT(data == olck); + osc_lock_blocking(env, dlmlock, + olck, flag == LDLM_CB_BLOCKING); + } else + cancel = 1; + cl_lock_mutex_put(env, lock); + osc_ast_data_put(env, olck); + } else + /* + * DLM lock exists, but there is no cl_lock attached to it. + * This is a `normal' race. cl_object and its cl_lock's can be + * removed by memory pressure, together with all pages. + */ + cancel = (flag == LDLM_CB_BLOCKING); + + if (cancel) { + struct lustre_handle *lockh; + + lockh = &osc_env_info(env)->oti_handle; + ldlm_lock2handle(dlmlock, lockh); + result = ldlm_cli_cancel(lockh, LCF_ASYNC); + } else + result = 0; + return result; +} + +/** + * Blocking ast invoked by ldlm when dlm lock is either blocking progress of + * some other lock, or is canceled. This function is installed as a + * ldlm_lock::l_blocking_ast() for client extent locks. + * + * Control flow is tricky, because ldlm uses the same call-back + * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's. + * + * \param dlmlock lock for which ast occurred. + * + * \param new description of a conflicting lock in case of blocking ast. + * + * \param data value of dlmlock->l_ast_data + * + * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish + * cancellation and blocking ast's. + * + * Possible use cases: + * + * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel + * lock due to lock lru pressure, or explicit user request to purge + * locks. + * + * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify + * us that dlmlock conflicts with another lock that some client is + * enqueing. Lock is canceled. + * + * - cl_lock_cancel() is called. osc_lock_cancel() calls + * ldlm_cli_cancel() that calls + * + * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) + * + * recursively entering osc_ldlm_blocking_ast(). + * + * - client cancels lock voluntary (e.g., as a part of early cancellation): + * + * cl_lock_cancel()-> + * osc_lock_cancel()-> + * ldlm_cli_cancel()-> + * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) + * + */ +static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock, + struct ldlm_lock_desc *new, void *data, + int flag) +{ + struct lu_env *env; + struct cl_env_nest nest; + int result; + + /* + * This can be called in the context of outer IO, e.g., + * + * cl_enqueue()->... + * ->osc_enqueue_base()->... + * ->ldlm_prep_elc_req()->... + * ->ldlm_cancel_callback()->... + * ->osc_ldlm_blocking_ast() + * + * new environment has to be created to not corrupt outer context. + */ + env = cl_env_nested_get(&nest); + if (!IS_ERR(env)) { + result = osc_dlm_blocking_ast0(env, dlmlock, data, flag); + cl_env_nested_put(&nest, env); + } else { + result = PTR_ERR(env); + /* + * XXX This should never happen, as cl_lock is + * stuck. Pre-allocated environment a la vvp_inode_fini_env + * should be used. + */ + LBUG(); + } + if (result != 0) { + if (result == -ENODATA) + result = 0; + else + CERROR("BAST failed: %d\n", result); + } + return result; +} + +static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock, + __u64 flags, void *data) +{ + struct cl_env_nest nest; + struct lu_env *env; + struct osc_lock *olck; + struct cl_lock *lock; + int result; + int dlmrc; + + /* first, do dlm part of the work */ + dlmrc = ldlm_completion_ast_async(dlmlock, flags, data); + /* then, notify cl_lock */ + env = cl_env_nested_get(&nest); + if (!IS_ERR(env)) { + olck = osc_ast_data_get(dlmlock); + if (olck != NULL) { + lock = olck->ols_cl.cls_lock; + cl_lock_mutex_get(env, lock); + /* + * ldlm_handle_cp_callback() copied LVB from request + * to lock->l_lvb_data, store it in osc_lock. + */ + 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) { + /* + * upcall (osc_lock_upcall()) hasn't yet been + * called. Do nothing now, upcall will bind + * olck to dlmlock and signal the waiters. + * + * This maintains an invariant that osc_lock + * and ldlm_lock are always bound when + * osc_lock is in OLS_GRANTED state. + */ + } 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) { + 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; + } else + result = -ELDLM_NO_LOCK_DATA; + cl_env_nested_put(&nest, env); + } else + result = PTR_ERR(env); + return dlmrc ?: result; +} + +static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data) +{ + struct ptlrpc_request *req = data; + struct osc_lock *olck; + struct cl_lock *lock; + struct cl_object *obj; + struct cl_env_nest nest; + struct lu_env *env; + struct ost_lvb *lvb; + struct req_capsule *cap; + int result; + + LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK); + + env = cl_env_nested_get(&nest); + if (!IS_ERR(env)) { + /* 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. + */ + olck = osc_ast_data_get(dlmlock); + if (olck != NULL) { + lock = olck->ols_cl.cls_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, + sizeof *lvb); + result = req_capsule_server_pack(cap); + if (result == 0) { + lvb = req_capsule_server_get(cap, &RMF_DLM_LVB); + obj = lock->cll_descr.cld_obj; + result = cl_object_glimpse(env, obj, lvb); + } + if (!exp_connect_lvb_type(req->rq_export)) + req_capsule_shrink(&req->rq_pill, + &RMF_DLM_LVB, + sizeof(struct ost_lvb_v1), + RCL_SERVER); + osc_ast_data_put(env, olck); + } else { + /* + * These errors are normal races, so we don't want to + * fill the console with messages by calling + * ptlrpc_error() + */ + lustre_pack_reply(req, 1, NULL, NULL); + result = -ELDLM_NO_LOCK_DATA; + } + cl_env_nested_put(&nest, env); + } else + result = PTR_ERR(env); + req->rq_status = result; + return result; +} + +static unsigned long osc_lock_weigh(const struct lu_env *env, + const struct cl_lock_slice *slice) +{ + /* + * don't need to grab coh_page_guard since we don't care the exact # + * of pages.. + */ + return cl_object_header(slice->cls_obj)->coh_pages; +} + +/** + * Get the weight of dlm lock for early cancellation. + * + * XXX: it should return the pages covered by this \a dlmlock. + */ +static unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock) +{ + struct cl_env_nest nest; + struct lu_env *env; + struct osc_lock *lock; + struct cl_lock *cll; + unsigned long weight; + ENTRY; + + 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 + * a new environment for it. Probably it is implementation safe to use + * the upper context because cl_lock_put don't modify environment + * variables. But in case of .. + */ + env = cl_env_nested_get(&nest); + if (IS_ERR(env)) + /* Mostly because lack of memory, tend to eliminate this lock*/ + RETURN(0); + + LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT); + lock = osc_ast_data_get(dlmlock); + if (lock == NULL) { + /* cl_lock was destroyed because of memory pressure. + * It is much reasonable to assign this type of lock + * a lower cost. + */ + GOTO(out, weight = 0); + } + + cll = lock->ols_cl.cls_lock; + cl_lock_mutex_get(env, cll); + weight = cl_lock_weigh(env, cll); + cl_lock_mutex_put(env, cll); + osc_ast_data_put(env, lock); + EXIT; + +out: + cl_env_nested_put(&nest, env); + return weight; +} + +static void osc_lock_build_einfo(const struct lu_env *env, + const struct cl_lock *clock, + struct osc_lock *lock, + struct ldlm_enqueue_info *einfo) +{ + enum cl_lock_mode mode; + + mode = clock->cll_descr.cld_mode; + if (mode == CLM_PHANTOM) + /* + * For now, enqueue all glimpse locks in read mode. In the + * future, client might choose to enqueue LCK_PW lock for + * glimpse on a file opened for write. + */ + mode = CLM_READ; + + einfo->ei_type = LDLM_EXTENT; + einfo->ei_mode = osc_cl_lock2ldlm(mode); + einfo->ei_cb_bl = osc_ldlm_blocking_ast; + einfo->ei_cb_cp = osc_ldlm_completion_ast; + einfo->ei_cb_gl = osc_ldlm_glimpse_ast; + einfo->ei_cb_wg = osc_ldlm_weigh_ast; + einfo->ei_cbdata = lock; /* value to be put into ->l_ast_data */ +} + +/** + * Determine if the lock should be converted into a lockless lock. + * + * Steps to check: + * - if the lock has an explicite requirment for a non-lockless lock; + * - if the io lock request type ci_lockreq; + * - send the enqueue rpc to ost to make the further decision; + * - special treat to truncate lockless lock + * + * Additional policy can be implemented here, e.g., never do lockless-io + * for large extents. + */ +static void osc_lock_to_lockless(const struct lu_env *env, + struct osc_lock *ols, int force) +{ + struct cl_lock_slice *slice = &ols->ols_cl; + + LASSERT(ols->ols_state == OLS_NEW || + ols->ols_state == OLS_UPCALL_RECEIVED); + + if (force) { + ols->ols_locklessable = 1; + slice->cls_ops = &osc_lock_lockless_ops; + } else { + struct osc_io *oio = osc_env_io(env); + struct cl_io *io = oio->oi_cl.cis_io; + struct cl_object *obj = slice->cls_obj; + struct osc_object *oob = cl2osc(obj); + const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev); + struct obd_connect_data *ocd; + + LASSERT(io->ci_lockreq == CILR_MANDATORY || + io->ci_lockreq == CILR_MAYBE || + io->ci_lockreq == CILR_NEVER); + + ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data; + 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 */ + (cl_io_is_trunc(io) && + (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) && + osd->od_lockless_truncate)) { + ols->ols_locklessable = 1; + slice->cls_ops = &osc_lock_lockless_ops; + } + } + 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. + * + * This function is used for two purposes: + * + * - early cancel all conflicting locks before starting IO, and + * + * - guarantee that pages added to the page cache by lockless IO are never + * covered by locks other than lockless IO lock, and, hence, are not + * visible to other threads. + */ +static int osc_lock_enqueue_wait(const struct lu_env *env, + const struct osc_lock *olck) +{ + 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 *scan; + struct cl_lock *conflict= NULL; + int lockless = osc_lock_is_lockless(olck); + int rc = 0; + ENTRY; + + LASSERT(cl_lock_is_mutexed(lock)); + + /* make it enqueue anyway for glimpse lock, because we actually + * don't need to cancel any conflicting locks. */ + if (olck->ols_glimpse) + return 0; + + spin_lock(&hdr->coh_lock_guard); + list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) { + struct cl_lock_descr *cld = &scan->cll_descr; + const struct osc_lock *scan_ols; + + if (scan == lock) + break; + + if (scan->cll_state < CLS_QUEUING || + scan->cll_state == CLS_FREEING || + cld->cld_start > descr->cld_end || + cld->cld_end < descr->cld_start) + continue; + + /* overlapped and living locks. */ + + /* We're not supposed to give up group lock. */ + if (scan->cll_descr.cld_mode == CLM_GROUP) { + LASSERT(descr->cld_mode != CLM_GROUP || + descr->cld_gid != scan->cll_descr.cld_gid); + continue; + } + + 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 && osc_lock_compatible(olck, scan_ols)) + continue; + + cl_lock_get_trust(scan); + conflict = scan; + break; + } + spin_unlock(&hdr->coh_lock_guard); + + 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; + } + } + RETURN(rc); +} + +/** + * Implementation of cl_lock_operations::clo_enqueue() method for osc + * layer. This initiates ldlm enqueue: + * + * - cancels conflicting locks early (osc_lock_enqueue_wait()); + * + * - calls osc_enqueue_base() to do actual enqueue. + * + * osc_enqueue_base() is supplied with an upcall function that is executed + * when lock is received either after a local cached ldlm lock is matched, or + * when a reply from the server is received. + * + * This function does not wait for the network communication to complete. + */ +static int osc_lock_enqueue(const struct lu_env *env, + const struct cl_lock_slice *slice, + struct cl_io *unused, __u32 enqflags) +{ + struct osc_lock *ols = cl2osc_lock(slice); + struct cl_lock *lock = ols->ols_cl.cls_lock; + int result; + ENTRY; + + LASSERT(cl_lock_is_mutexed(lock)); + LASSERTF(ols->ols_state == OLS_NEW, + "Impossible state: %d\n", ols->ols_state); + + LASSERTF(ergo(ols->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ), + "lock = %p, ols = %p\n", lock, ols); + + result = osc_lock_enqueue_wait(env, ols); + if (result == 0) { + 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; + + /* 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; + + /* + * XXX: this is possible blocking point as + * ldlm_lock_match(LDLM_FL_LVB_READY) waits for + * LDLM_CP_CALLBACK. + */ + ostid_build_res_name(&obj->oo_oinfo->loi_oi, 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, ols->ols_agl); + if (result != 0) { + 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))); + RETURN(result); +} + +static int osc_lock_wait(const struct lu_env *env, + const struct cl_lock_slice *slice) +{ + struct osc_lock *olck = cl2osc_lock(slice); + struct cl_lock *lock = olck->ols_cl.cls_lock; + + LINVRNT(osc_lock_invariant(olck)); + + 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); + olck->ols_agl = 0; + 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)); + + return lock->cll_error ?: olck->ols_state >= OLS_GRANTED ? 0 : CLO_WAIT; +} + +/** + * An implementation of cl_lock_operations::clo_use() method that pins cached + * lock. + */ +static int osc_lock_use(const struct lu_env *env, + const struct cl_lock_slice *slice) +{ + struct osc_lock *olck = cl2osc_lock(slice); + 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 = 1; + olck->ols_state = OLS_GRANTED; + } else { + struct cl_lock *lock; + + /* + * Lock is being cancelled somewhere within + * ldlm_handle_bl_callback(): LDLM_FL_CBPENDING is already + * set, but osc_ldlm_blocking_ast() hasn't yet acquired + * cl_lock mutex. + */ + lock = slice->cls_lock; + LASSERT(lock->cll_state == CLS_INTRANSIT); + LASSERT(lock->cll_users > 0); + /* set a flag for osc_dlm_blocking_ast0() to signal the + * lock.*/ + olck->ols_ast_wait = 1; + rc = CLO_WAIT; + } + return rc; +} + +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; + 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); + LDLM_DEBUG(ols->ols_lock, + "lock %p: %d pages were %s.\n", lock, result, + discard ? "discarded" : "written"); + if (result > 0) + result = 0; + } + + rc = cl_lock_discard_pages(env, lock); + if (result == 0 && rc < 0) + result = rc; + + cl_env_nested_put(&nest, env); + } else + result = PTR_ERR(env); + if (result == 0) { + ols->ols_flush = 1; + LINVRNT(!osc_lock_has_pages(ols)); + } + RETURN(result); +} + +/** + * Implements cl_lock_operations::clo_cancel() method for osc layer. This is + * called (as part of cl_lock_cancel()) when lock is canceled either voluntary + * (LRU pressure, early cancellation, umount, etc.) or due to the conflict + * with some other lock some where in the cluster. This function does the + * following: + * + * - invalidates all pages protected by this lock (after sending dirty + * ones to the server, as necessary); + * + * - decref's underlying ldlm lock; + * + * - cancels ldlm lock (ldlm_cli_cancel()). + */ +static void osc_lock_cancel(const struct lu_env *env, + const struct cl_lock_slice *slice) +{ + struct cl_lock *lock = slice->cls_lock; + struct osc_lock *olck = cl2osc_lock(slice); + struct ldlm_lock *dlmlock = olck->ols_lock; + int result = 0; + int discard; + + LASSERT(cl_lock_is_mutexed(lock)); + LINVRNT(osc_lock_invariant(olck)); + + if (dlmlock != NULL) { + int do_cancel; + + discard = !!(dlmlock->l_flags & LDLM_FL_DISCARD_DATA); + if (olck->ols_state >= OLS_GRANTED) + result = osc_lock_flush(olck, discard); + osc_lock_unhold(olck); + + lock_res_and_lock(dlmlock); + /* Now that we're the only user of dlm read/write reference, + * mostly the ->l_readers + ->l_writers should be zero. + * However, there is a corner case. + * See bug 18829 for details.*/ + do_cancel = (dlmlock->l_readers == 0 && + dlmlock->l_writers == 0); + dlmlock->l_flags |= LDLM_FL_CBPENDING; + unlock_res_and_lock(dlmlock); + if (do_cancel) + result = ldlm_cli_cancel(&olck->ols_handle, LCF_ASYNC); + if (result < 0) + CL_LOCK_DEBUG(D_ERROR, env, lock, + "lock %p cancel failure with error(%d)\n", + lock, result); + } + olck->ols_state = OLS_CANCELLED; + olck->ols_flags &= ~LDLM_FL_LVB_READY; + osc_lock_detach(env, olck); +} + +static int osc_lock_has_pages(struct osc_lock *olck) +{ + return 0; +} + +static void osc_lock_delete(const struct lu_env *env, + const struct cl_lock_slice *slice) +{ + struct osc_lock *olck; + + olck = cl2osc_lock(slice); + if (olck->ols_glimpse) { + LASSERT(!olck->ols_hold); + LASSERT(!olck->ols_lock); + return; + } + + LINVRNT(osc_lock_invariant(olck)); + LINVRNT(!osc_lock_has_pages(olck)); + + osc_lock_unhold(olck); + osc_lock_detach(env, olck); +} + +/** + * Implements cl_lock_operations::clo_state() method for osc layer. + * + * Maintains osc_lock::ols_owner field. + * + * This assumes that lock always enters CLS_HELD (from some other state) in + * the same IO context as one that requested the lock. This should not be a + * problem, because context is by definition shared by all activity pertaining + * to the same high-level IO. + */ +static void osc_lock_state(const struct lu_env *env, + const struct cl_lock_slice *slice, + enum cl_lock_state state) +{ + struct osc_lock *lock = cl2osc_lock(slice); + + /* + * 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) + lock->ols_owner = NULL; +} + +static int osc_lock_print(const struct lu_env *env, void *cookie, + lu_printer_t p, const struct cl_lock_slice *slice) +{ + struct osc_lock *lock = cl2osc_lock(slice); + + /* + * XXX print ldlm lock and einfo properly. + */ + (*p)(env, cookie, "%p %#16llx "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, + .clo_wait = osc_lock_wait, + .clo_unuse = osc_lock_unuse, + .clo_use = osc_lock_use, + .clo_delete = osc_lock_delete, + .clo_state = osc_lock_state, + .clo_cancel = osc_lock_cancel, + .clo_weigh = osc_lock_weigh, + .clo_print = osc_lock_print, + .clo_fits_into = osc_lock_fits_into, +}; + +static int osc_lock_lockless_unuse(const struct lu_env *env, + const struct cl_lock_slice *slice) +{ + struct osc_lock *ols = cl2osc_lock(slice); + struct cl_lock *lock = slice->cls_lock; + + LASSERT(ols->ols_state == OLS_GRANTED); + LINVRNT(osc_lock_invariant(ols)); + + cl_lock_cancel(env, lock); + cl_lock_delete(env, lock); + return 0; +} + +static void osc_lock_lockless_cancel(const struct lu_env *env, + const struct cl_lock_slice *slice) +{ + struct osc_lock *ols = cl2osc_lock(slice); + int result; + + result = osc_lock_flush(ols, 0); + if (result) + CERROR("Pages for lockless lock %p were not purged(%d)\n", + ols, result); + ols->ols_state = OLS_CANCELLED; +} + +static int osc_lock_lockless_wait(const struct lu_env *env, + const struct cl_lock_slice *slice) +{ + struct osc_lock *olck = cl2osc_lock(slice); + struct cl_lock *lock = olck->ols_cl.cls_lock; + + LINVRNT(osc_lock_invariant(olck)); + LASSERT(olck->ols_state >= OLS_UPCALL_RECEIVED); + + return lock->cll_error; +} + +static void osc_lock_lockless_state(const struct lu_env *env, + const struct cl_lock_slice *slice, + enum cl_lock_state state) +{ + struct osc_lock *lock = cl2osc_lock(slice); + + LINVRNT(osc_lock_invariant(lock)); + if (state == CLS_HELD) { + 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; + } +} + +static int osc_lock_lockless_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 *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_enqueue, + .clo_wait = osc_lock_lockless_wait, + .clo_unuse = osc_lock_lockless_unuse, + .clo_state = osc_lock_lockless_state, + .clo_fits_into = osc_lock_lockless_fits_into, + .clo_cancel = osc_lock_lockless_cancel, + .clo_print = osc_lock_print +}; + +int osc_lock_init(const struct lu_env *env, + struct cl_object *obj, struct cl_lock *lock, + const struct cl_io *unused) +{ + struct osc_lock *clk; + int result; + + OBD_SLAB_ALLOC_PTR_GFP(clk, osc_lock_kmem, __GFP_IO); + if (clk != NULL) { + __u32 enqflags = lock->cll_descr.cld_enq_flags; + + osc_lock_build_einfo(env, lock, clk, &clk->ols_einfo); + atomic_set(&clk->ols_pageref, 0); + clk->ols_state = OLS_NEW; + + clk->ols_flags = osc_enq2ldlm_flags(enqflags); + clk->ols_agl = !!(enqflags & CEF_AGL); + if (clk->ols_agl) + clk->ols_flags |= LDLM_FL_BLOCK_NOWAIT; + if (clk->ols_flags & LDLM_FL_HAS_INTENT) + clk->ols_glimpse = 1; + + cl_lock_slice_add(lock, &clk->ols_cl, obj, &osc_lock_ops); + + if (!(enqflags & CEF_MUST)) + /* try to convert this lock to a lockless lock */ + osc_lock_to_lockless(env, clk, (enqflags & CEF_NEVER)); + if (clk->ols_locklessable && !(enqflags & CEF_DISCARD_DATA)) + clk->ols_flags |= LDLM_FL_DENY_ON_CONTENTION; + + LDLM_DEBUG_NOLOCK("lock %p, osc lock %p, flags %llx\n", + lock, clk, clk->ols_flags); + + result = 0; + } else + result = -ENOMEM; + return result; +} + +int osc_dlm_lock_pageref(struct ldlm_lock *dlm) +{ + struct osc_lock *olock; + int rc = 0; + + 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 && + atomic_add_return(_PAGEREF_MAGIC, + &olock->ols_pageref) != _PAGEREF_MAGIC) { + atomic_sub(_PAGEREF_MAGIC, &olock->ols_pageref); + rc = 1; + } + spin_unlock(&osc_ast_guard); + return rc; +} + +/** @} osc */ |