X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_request.c;h=684062f980829a14783c41c8145fc2c037cac809;hp=c5ea6779a44f9aff8120956c5badef174268ac2c;hb=0356990876308f811cc6c61c22946a1cd73e5c23;hpb=929ec628e6fef5609e55d519a1eb9e2cbbf1f1e8 diff --git a/lustre/ldlm/ldlm_request.c b/lustre/ldlm/ldlm_request.c index c5ea677..684062f 100644 --- a/lustre/ldlm/ldlm_request.c +++ b/lustre/ldlm/ldlm_request.c @@ -27,18 +27,39 @@ * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2010, 2011, Whamcloud, Inc. + * Copyright (c) 2010, 2014, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ * Lustre is a trademark of Sun Microsystems, Inc. */ +/** + * This file contains Asynchronous System Trap (AST) handlers and related + * LDLM request-processing routines. + * + * An AST is a callback issued on a lock when its state is changed. There are + * several different types of ASTs (callbacks) registered for each lock: + * + * - completion AST: when a lock is enqueued by some process, but cannot be + * granted immediately due to other conflicting locks on the same resource, + * the completion AST is sent to notify the caller when the lock is + * eventually granted + * + * - blocking AST: when a lock is granted to some process, if another process + * enqueues a conflicting (blocking) lock on a resource, a blocking AST is + * sent to notify the holder(s) of the lock(s) of the conflicting lock + * request. The lock holder(s) must release their lock(s) on that resource in + * a timely manner or be evicted by the server. + * + * - glimpse AST: this is used when a process wants information about a lock + * (i.e. the lock value block (LVB)) but does not necessarily require holding + * the lock. If the resource is locked, the lock holder(s) are sent glimpse + * ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL + * their lock(s) if they are idle. If the resource is not locked, the server + * may grant the lock. + */ #define DEBUG_SUBSYSTEM S_LDLM -#ifndef __KERNEL__ -#include -#include -#endif #include #include @@ -46,8 +67,8 @@ #include "ldlm_internal.h" -int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT; -CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644, +unsigned int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT; +CFS_MODULE_PARM(ldlm_enqueue_min, "i", uint, 0644, "lock enqueue timeout minimum"); /* in client side, whether the cached locks will be canceled before replay */ @@ -77,15 +98,17 @@ int ldlm_expired_completion_wait(void *data) if (lock->l_conn_export == NULL) { static cfs_time_t next_dump = 0, last_dump = 0; - if (ptlrpc_check_suspend()) - RETURN(0); - - LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", " - CFS_DURATION_T"s ago); not entering recovery in " - "server code, just going back to sleep", - lock->l_last_activity, - cfs_time_sub(cfs_time_current_sec(), - lock->l_last_activity)); + LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", " + CFS_DURATION_T"s ago)\n", + lock->l_last_activity, + cfs_time_sub(cfs_time_current_sec(), + lock->l_last_activity)); + LDLM_DEBUG(lock, "lock timed out (enqueued at "CFS_TIME_T", " + CFS_DURATION_T"s ago); not entering recovery in " + "server code, just going back to sleep", + lock->l_last_activity, + cfs_time_sub(cfs_time_current_sec(), + lock->l_last_activity)); if (cfs_time_after(cfs_time_current(), next_dump)) { last_dump = next_dump; next_dump = cfs_time_shift(300); @@ -108,47 +131,58 @@ int ldlm_expired_completion_wait(void *data) RETURN(0); } -EXPORT_SYMBOL(ldlm_expired_completion_wait); + +/** + * Calculate the Completion timeout (covering enqueue, BL AST, data flush, + * lock cancel, and their replies). Used for lock completion timeout on the + * client side. + * + * \param[in] lock lock which is waiting the completion callback + * + * \retval timeout in seconds to wait for the server reply + */ /* We use the same basis for both server side and client side functions from a single node. */ -int ldlm_get_enq_timeout(struct ldlm_lock *lock) +static unsigned int ldlm_cp_timeout(struct ldlm_lock *lock) { - int timeout = at_get(ldlm_lock_to_ns_at(lock)); - if (AT_OFF) - return obd_timeout / 2; - /* Since these are non-updating timeouts, we should be conservative. - It would be nice to have some kind of "early reply" mechanism for - lock callbacks too... */ - timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */ - return max(timeout, ldlm_enqueue_min); + unsigned int timeout; + + if (AT_OFF) + return obd_timeout; + + /* Wait a long time for enqueue - server may have to callback a + * lock from another client. Server will evict the other client if it + * doesn't respond reasonably, and then give us the lock. */ + timeout = at_get(ldlm_lock_to_ns_at(lock)); + return max(3 * timeout, ldlm_enqueue_min); } -EXPORT_SYMBOL(ldlm_get_enq_timeout); /** * Helper function for ldlm_completion_ast(), updating timings when lock is * actually granted. */ -static int ldlm_completion_tail(struct ldlm_lock *lock) +static int ldlm_completion_tail(struct ldlm_lock *lock, void *data) { - long delay; - int result; - - if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) { - LDLM_DEBUG(lock, "client-side enqueue: destroyed"); - result = -EIO; - } else { - delay = cfs_time_sub(cfs_time_current_sec(), - lock->l_last_activity); - LDLM_DEBUG(lock, "client-side enqueue: granted after " - CFS_DURATION_T"s", delay); - - /* Update our time estimate */ - at_measured(ldlm_lock_to_ns_at(lock), - delay); - result = 0; - } - return result; + long delay; + int result = 0; + + if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) { + LDLM_DEBUG(lock, "client-side enqueue: destroyed"); + result = -EIO; + } else if (data == NULL) { + LDLM_DEBUG(lock, "client-side enqueue: granted"); + } else { + /* Take into AT only CP RPC, not immediately granted locks */ + delay = cfs_time_sub(cfs_time_current_sec(), + lock->l_last_activity); + LDLM_DEBUG(lock, "client-side enqueue: granted after " + CFS_DURATION_T"s", delay); + + /* Update our time estimate */ + at_measured(ldlm_lock_to_ns_at(lock), delay); + } + return result; } /** @@ -158,34 +192,33 @@ static int ldlm_completion_tail(struct ldlm_lock *lock) */ int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data) { - ENTRY; + ENTRY; - if (flags == LDLM_FL_WAIT_NOREPROC) { - LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock"); - RETURN(0); - } + if (flags == LDLM_FL_WAIT_NOREPROC) { + LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock"); + RETURN(0); + } - if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED | - LDLM_FL_BLOCK_CONV))) { - cfs_waitq_signal(&lock->l_waitq); - RETURN(ldlm_completion_tail(lock)); - } + if (!(flags & LDLM_FL_BLOCKED_MASK)) { + wake_up(&lock->l_waitq); + RETURN(ldlm_completion_tail(lock, data)); + } - LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, " - "going forward"); - ldlm_reprocess_all(lock->l_resource); - RETURN(0); + LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, " + "going forward"); + ldlm_reprocess_all(lock->l_resource); + RETURN(0); } EXPORT_SYMBOL(ldlm_completion_ast_async); /** - * Client side LDLM "completion" AST. This is called in several cases: + * Generic LDLM "completion" AST. This is called in several cases: * - * - when a reply to an ENQUEUE rpc is received from the server + * - when a reply to an ENQUEUE RPC is received from the server * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at * this point (determined by flags); * - * - when LDLM_CP_CALLBACK rpc comes to client to notify it that lock has + * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has * been granted; * * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock @@ -215,11 +248,10 @@ int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) goto noreproc; } - if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED | - LDLM_FL_BLOCK_CONV))) { - cfs_waitq_signal(&lock->l_waitq); - RETURN(0); - } + if (!(flags & LDLM_FL_BLOCKED_MASK)) { + wake_up(&lock->l_waitq); + RETURN(0); + } LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, " "sleeping"); @@ -233,14 +265,12 @@ noreproc: imp = obd->u.cli.cl_import; } - /* Wait a long time for enqueue - server may have to callback a - lock from another client. Server will evict the other client if it - doesn't respond reasonably, and then give us the lock. */ - timeout = ldlm_get_enq_timeout(lock) * 2; + timeout = ldlm_cp_timeout(lock); - lwd.lwd_lock = lock; + lwd.lwd_lock = lock; + lock->l_last_activity = cfs_time_current_sec(); - if (lock->l_flags & LDLM_FL_NO_TIMEOUT) { + if (ldlm_is_no_timeout(lock)) { LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT"); lwi = LWI_INTR(interrupted_completion_wait, &lwd); } else { @@ -258,7 +288,7 @@ noreproc: if (ns_is_client(ldlm_lock_to_ns(lock)) && OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST, OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) { - lock->l_flags |= LDLM_FL_FAIL_LOC; + ldlm_set_fail_loc(lock); rc = -EINTR; } else { /* Go to sleep until the lock is granted or cancelled. */ @@ -270,19 +300,19 @@ noreproc: LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)", rc); RETURN(rc); - } + } - RETURN(ldlm_completion_tail(lock)); + RETURN(ldlm_completion_tail(lock, data)); } EXPORT_SYMBOL(ldlm_completion_ast); /** - * A helper to build a blocking ast function + * A helper to build a blocking AST function * - * Perform a common operation for blocking asts: + * Perform a common operation for blocking ASTs: * defferred lock cancellation. * - * \param lock the lock blocking or canceling ast was called on + * \param lock the lock blocking or canceling AST was called on * \retval 0 * \see mdt_blocking_ast * \see ldlm_blocking_ast @@ -292,7 +322,7 @@ int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock) int do_ast; ENTRY; - lock->l_flags |= LDLM_FL_CBPENDING; + ldlm_set_cbpending(lock); do_ast = (!lock->l_readers && !lock->l_writers); unlock_res_and_lock(lock); @@ -302,7 +332,7 @@ int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock) LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel"); ldlm_lock2handle(lock, &lockh); - rc = ldlm_cli_cancel(&lockh); + rc = ldlm_cli_cancel(&lockh, LCF_ASYNC); if (rc < 0) CERROR("ldlm_cli_cancel: %d\n", rc); } else { @@ -350,34 +380,43 @@ int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, } EXPORT_SYMBOL(ldlm_blocking_ast); -/* - * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See - * comment in filter_intent_policy() on why you may need this. +/** + * Implements ldlm_lock::l_glimpse_ast for extent locks acquired on the server. + * + * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for that is + * rather subtle: with OST-side locking, it may so happen that _all_ extent + * locks are held by the OST. If client wants to obtain the current file size + * it calls ll_glimpse_size(), and (as all locks are held only on the server), + * this dummy glimpse callback fires and does nothing. The client still + * receives the correct file size due to the following fragment of code in + * ldlm_cb_interpret(): + * + * if (rc == -ELDLM_NO_LOCK_DATA) { + * LDLM_DEBUG(lock, "lost race - client has a lock but no" + * "inode"); + * ldlm_res_lvbo_update(lock->l_resource, NULL, 1); + * } + * + * That is, after the glimpse returns this error, ofd_lvbo_update() is called + * and returns the updated file attributes from the inode to the client. + * + * See also comment in ofd_intent_policy() on why servers must set a non-NULL + * l_glimpse_ast when grabbing DLM locks. Otherwise, the server will assume + * that the object is in the process of being destroyed. + * + * \param[in] lock DLM lock being glimpsed, unused + * \param[in] reqp pointer to ptlrpc_request, unused + * + * \retval -ELDLM_NO_LOCK_DATA to get attributes from disk object */ int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp) { - /* - * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for - * that is rather subtle: with OST-side locking, it may so happen that - * _all_ extent locks are held by the OST. If client wants to obtain - * current file size it calls ll{,u}_glimpse_size(), and (as locks are - * on the server), dummy glimpse callback fires and does - * nothing. Client still receives correct file size due to the - * following fragment in filter_intent_policy(): - * - * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB - * if (rc != 0 && res->lr_namespace->ns_lvbo && - * res->lr_namespace->ns_lvbo->lvbo_update) { - * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1); - * } - * - * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and - * returns correct file size to the client. - */ return -ELDLM_NO_LOCK_DATA; } -EXPORT_SYMBOL(ldlm_glimpse_ast); +/** + * Enqueue a local lock (typically on a server). + */ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, const struct ldlm_res_id *res_id, ldlm_type_t type, ldlm_policy_data_t *policy, @@ -405,24 +444,35 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len, lvb_type); - if (unlikely(!lock)) - GOTO(out_nolock, err = -ENOMEM); + if (IS_ERR(lock)) + GOTO(out_nolock, err = PTR_ERR(lock)); + + err = ldlm_lvbo_init(lock->l_resource); + if (err < 0) { + LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", err); + GOTO(out, err); + } ldlm_lock2handle(lock, lockh); /* NB: we don't have any lock now (lock_res_and_lock) * because it's a new lock */ ldlm_lock_addref_internal_nolock(lock, mode); - lock->l_flags |= LDLM_FL_LOCAL; + ldlm_set_local(lock); if (*flags & LDLM_FL_ATOMIC_CB) - lock->l_flags |= LDLM_FL_ATOMIC_CB; + ldlm_set_atomic_cb(lock); if (policy != NULL) lock->l_policy_data = *policy; if (client_cookie != NULL) lock->l_client_cookie = *client_cookie; - if (type == LDLM_EXTENT) - lock->l_req_extent = policy->l_extent; + if (type == LDLM_EXTENT) { + /* extent lock without policy is a bug */ + if (policy == NULL) + LBUG(); + + lock->l_req_extent = policy->l_extent; + } err = ldlm_lock_enqueue(ns, &lock, policy, flags); if (unlikely(err != ELDLM_OK)) @@ -452,13 +502,13 @@ static void failed_lock_cleanup(struct ldlm_namespace *ns, lock_res_and_lock(lock); /* Check that lock is not granted or failed, we might race. */ if ((lock->l_req_mode != lock->l_granted_mode) && - !(lock->l_flags & LDLM_FL_FAILED)) { - /* Make sure that this lock will not be found by raced - * bl_ast and -EINVAL reply is sent to server anyways. - * bug 17645 */ - lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED | - LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING; - need_cancel = 1; + !ldlm_is_failed(lock)) { + /* Make sure that this lock will not be found by raced + * bl_ast and -EINVAL reply is sent to server anyways. + * b=17645*/ + lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED | + LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING; + need_cancel = 1; } unlock_res_and_lock(lock); @@ -469,22 +519,30 @@ static void failed_lock_cleanup(struct ldlm_namespace *ns, else LDLM_DEBUG(lock, "lock was granted or failed in race"); - ldlm_lock_decref_internal(lock, mode); - - /* XXX - HACK because we shouldn't call ldlm_lock_destroy() - * from llite/file.c/ll_file_flock(). */ - /* This code makes for the fact that we do not have blocking handler on - * a client for flock locks. As such this is the place where we must - * completely kill failed locks. (interrupted and those that - * were waiting to be granted when server evicted us. */ - if (lock->l_resource->lr_type == LDLM_FLOCK) { - lock_res_and_lock(lock); - ldlm_resource_unlink_lock(lock); - ldlm_lock_destroy_nolock(lock); - unlock_res_and_lock(lock); - } + /* XXX - HACK because we shouldn't call ldlm_lock_destroy() + * from llite/file.c/ll_file_flock(). */ + /* This code makes for the fact that we do not have blocking handler on + * a client for flock locks. As such this is the place where we must + * completely kill failed locks. (interrupted and those that + * were waiting to be granted when server evicted us. */ + if (lock->l_resource->lr_type == LDLM_FLOCK) { + lock_res_and_lock(lock); + if (!ldlm_is_destroyed(lock)) { + ldlm_resource_unlink_lock(lock); + ldlm_lock_decref_internal_nolock(lock, mode); + ldlm_lock_destroy_nolock(lock); + } + unlock_res_and_lock(lock); + } else { + ldlm_lock_decref_internal(lock, mode); + } } +/** + * Finishing portion of client lock enqueue code. + * + * Called after receiving reply from server. + */ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode, __u64 *flags, void *lvb, __u32 lvb_len, @@ -495,7 +553,6 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, struct ldlm_lock *lock; struct ldlm_reply *reply; int cleanup_phase = 1; - int size = 0; ENTRY; lock = ldlm_handle2lock(lockh); @@ -522,8 +579,8 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, if (reply == NULL) GOTO(cleanup, rc = -EPROTO); - if (lvb_len != 0) { - LASSERT(lvb != NULL); + if (lvb_len > 0) { + int size = 0; size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER); @@ -536,13 +593,14 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, lvb_len, size); GOTO(cleanup, rc = -EINVAL); } + lvb_len = size; } if (rc == ELDLM_LOCK_ABORTED) { - if (lvb_len != 0) + if (lvb_len > 0 && lvb != NULL) rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER, - lvb, size); - GOTO(cleanup, rc = (rc != 0 ? rc : ELDLM_LOCK_ABORTED)); + lvb, lvb_len); + GOTO(cleanup, rc = rc ? : ELDLM_LOCK_ABORTED); } /* lock enqueued on the server */ @@ -564,73 +622,59 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, *flags = ldlm_flags_from_wire(reply->lock_flags); lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags & - LDLM_INHERIT_FLAGS); - /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match() - * to wait with no timeout as well */ - lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags & - LDLM_FL_NO_TIMEOUT); + LDLM_FL_INHERIT_MASK); unlock_res_and_lock(lock); - CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%llx\n", - lock, reply->lock_handle.cookie, *flags); - - /* If enqueue returned a blocked lock but the completion handler has - * already run, then it fixed up the resource and we don't need to do it - * again. */ - if ((*flags) & LDLM_FL_LOCK_CHANGED) { - int newmode = reply->lock_desc.l_req_mode; - LASSERT(!is_replay); - if (newmode && newmode != lock->l_req_mode) { - LDLM_DEBUG(lock, "server returned different mode %s", - ldlm_lockname[newmode]); - lock->l_req_mode = newmode; - } + CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: "LPX64"\n", + lock, reply->lock_handle.cookie, *flags); + + /* If enqueue returned a blocked lock but the completion handler has + * already run, then it fixed up the resource and we don't need to do it + * again. */ + if ((*flags) & LDLM_FL_LOCK_CHANGED) { + int newmode = reply->lock_desc.l_req_mode; + LASSERT(!is_replay); + if (newmode && newmode != lock->l_req_mode) { + LDLM_DEBUG(lock, "server returned different mode %s", + ldlm_lockname[newmode]); + lock->l_req_mode = newmode; + } - if (memcmp(reply->lock_desc.l_resource.lr_name.name, - lock->l_resource->lr_name.name, - sizeof(struct ldlm_res_id))) { - CDEBUG(D_INFO, "remote intent success, locking " - "(%ld,%ld,%ld) instead of " - "(%ld,%ld,%ld)\n", - (long)reply->lock_desc.l_resource.lr_name.name[0], - (long)reply->lock_desc.l_resource.lr_name.name[1], - (long)reply->lock_desc.l_resource.lr_name.name[2], - (long)lock->l_resource->lr_name.name[0], - (long)lock->l_resource->lr_name.name[1], - (long)lock->l_resource->lr_name.name[2]); - - rc = ldlm_lock_change_resource(ns, lock, - &reply->lock_desc.l_resource.lr_name); - if (rc || lock->l_resource == NULL) - GOTO(cleanup, rc = -ENOMEM); - LDLM_DEBUG(lock, "client-side enqueue, new resource"); - } - if (with_policy) - if (!(type == LDLM_IBITS && !(exp->exp_connect_flags & - OBD_CONNECT_IBITS))) - /* We assume lock type cannot change on server*/ - ldlm_convert_policy_to_local(exp, - lock->l_resource->lr_type, - &reply->lock_desc.l_policy_data, - &lock->l_policy_data); + if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name, + &lock->l_resource->lr_name)) { + CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES + " instead of "DLDLMRES"\n", + PLDLMRES(&reply->lock_desc.l_resource), + PLDLMRES(lock->l_resource)); + + rc = ldlm_lock_change_resource(ns, lock, + &reply->lock_desc.l_resource.lr_name); + if (rc || lock->l_resource == NULL) + GOTO(cleanup, rc = -ENOMEM); + LDLM_DEBUG(lock, "client-side enqueue, new resource"); + } + if (with_policy) + if (!(type == LDLM_IBITS && + !(exp_connect_flags(exp) & OBD_CONNECT_IBITS))) + /* We assume lock type cannot change on server*/ + ldlm_convert_policy_to_local(exp, + lock->l_resource->lr_type, + &reply->lock_desc.l_policy_data, + &lock->l_policy_data); if (type != LDLM_PLAIN) LDLM_DEBUG(lock,"client-side enqueue, new policy data"); } - if ((*flags) & LDLM_FL_AST_SENT || - /* Cancel extent locks as soon as possible on a liblustre client, - * because it cannot handle asynchronous ASTs robustly (see - * bug 7311). */ - (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) { + if ((*flags) & LDLM_FL_AST_SENT) { lock_res_and_lock(lock); - lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST; + lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST; unlock_res_and_lock(lock); LDLM_DEBUG(lock, "enqueue reply includes blocking AST"); } /* If the lock has already been granted by a completion AST, don't * clobber the LVB with an older one. */ - if (lvb_len != 0) { + if (lvb_len > 0) { /* We must lock or a racing completion might update lvb without * letting us know and we'll clobber the correct value. * Cannot unlock after the check either, a that still leaves @@ -638,10 +682,12 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, lock_res_and_lock(lock); if (lock->l_req_mode != lock->l_granted_mode) rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER, - lock->l_lvb_data, size); + lock->l_lvb_data, lvb_len); unlock_res_and_lock(lock); - if (rc < 0) + if (rc < 0) { + cleanup_phase = 1; GOTO(cleanup, rc); + } } if (!is_replay) { @@ -655,11 +701,11 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req, } } - if (lvb_len && lvb != NULL) { - /* Copy the LVB here, and not earlier, because the completion - * AST (if any) can override what we got in the reply */ - memcpy(lvb, lock->l_lvb_data, lvb_len); - } + if (lvb_len > 0 && lvb != NULL) { + /* Copy the LVB here, and not earlier, because the completion + * AST (if any) can override what we got in the reply */ + memcpy(lvb, lock->l_lvb_data, lvb_len); + } LDLM_DEBUG(lock, "client-side enqueue END"); EXIT; @@ -673,14 +719,17 @@ cleanup: } EXPORT_SYMBOL(ldlm_cli_enqueue_fini); -/* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into - * a single page on the send/receive side. XXX: 512 should be changed - * to more adequate value. */ +/** + * Estimate number of lock handles that would fit into request of given + * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into + * a single page on the send/receive side. XXX: 512 should be changed to + * more adequate value. + */ static inline int ldlm_req_handles_avail(int req_size, int off) { int avail; - avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size; + avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size; if (likely(avail >= 0)) avail /= (int)sizeof(struct lustre_handle); else @@ -694,34 +743,40 @@ static inline int ldlm_capsule_handles_avail(struct req_capsule *pill, enum req_location loc, int off) { - int size = req_capsule_msg_size(pill, loc); - return ldlm_req_handles_avail(size, off); + __u32 size = req_capsule_msg_size(pill, loc); + return ldlm_req_handles_avail(size, off); } static inline int ldlm_format_handles_avail(struct obd_import *imp, const struct req_format *fmt, enum req_location loc, int off) { - int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc); - return ldlm_req_handles_avail(size, off); + __u32 size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc); + return ldlm_req_handles_avail(size, off); } -/* Cancel lru locks and pack them into the enqueue request. Pack there the given - * @count locks in @cancels. */ +/** + * Cancel LRU locks and pack them into the enqueue request. Pack there the given + * \a count locks in \a cancels. + * + * This is to be called by functions preparing their own requests that + * might contain lists of locks to cancel in addition to actual operation + * that needs to be performed. + */ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req, - int version, int opc, int canceloff, - cfs_list_t *cancels, int count) -{ - struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; - struct req_capsule *pill = &req->rq_pill; - struct ldlm_request *dlm = NULL; - int flags, avail, to_free, pack = 0; - CFS_LIST_HEAD(head); - int rc; - ENTRY; + int version, int opc, int canceloff, + struct list_head *cancels, int count) + { + struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; + struct req_capsule *pill = &req->rq_pill; + struct ldlm_request *dlm = NULL; + struct list_head head = LIST_HEAD_INIT(head); + int flags, avail, to_free, pack = 0; + int rc; + ENTRY; - if (cancels == NULL) - cancels = &head; + if (cancels == NULL) + cancels = &head; if (ns_connect_cancelset(ns)) { /* Estimate the amount of available space in the request. */ req_capsule_filled_sizes(pill, RCL_CLIENT); @@ -732,9 +787,9 @@ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req, to_free = !ns_connect_lru_resize(ns) && opc == LDLM_ENQUEUE ? 1 : 0; - /* Cancel lru locks here _only_ if the server supports - * EARLY_CANCEL. Otherwise we have to send extra CANCEL - * rpc, what will make us slower. */ + /* Cancel LRU locks here _only_ if the server supports + * EARLY_CANCEL. Otherwise we have to send extra CANCEL + * RPC, which will make us slower. */ if (avail > count) count += ldlm_cancel_lru_local(ns, cancels, to_free, avail - count, 0, flags); @@ -757,14 +812,14 @@ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req, dlm = req_capsule_client_get(pill, &RMF_DLM_REQ); LASSERT(dlm); /* Skip first lock handler in ldlm_request_pack(), - * this method will incrment @lock_count according + * this method will increment @lock_count according * to the lock handle amount actually written to * the buffer. */ dlm->lock_count = canceloff; } /* Pack into the request @pack lock handles. */ ldlm_cli_cancel_list(cancels, pack, req, 0); - /* Prepare and send separate cancel rpc for others. */ + /* Prepare and send separate cancel RPC for others. */ ldlm_cli_cancel_list(cancels, count - pack, NULL, 0); } else { ldlm_lock_list_put(cancels, l_bl_ast, count); @@ -774,19 +829,45 @@ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req, EXPORT_SYMBOL(ldlm_prep_elc_req); int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req, - cfs_list_t *cancels, int count) + struct list_head *cancels, int count) { return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, LDLM_ENQUEUE_CANCEL_OFF, cancels, count); } EXPORT_SYMBOL(ldlm_prep_enqueue_req); -/* If a request has some specific initialisation it is passed in @reqp, +struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len) +{ + struct ptlrpc_request *req; + int rc; + ENTRY; + + req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE); + if (req == NULL) + RETURN(ERR_PTR(-ENOMEM)); + + rc = ldlm_prep_enqueue_req(exp, req, NULL, 0); + if (rc) { + ptlrpc_request_free(req); + RETURN(ERR_PTR(rc)); + } + + req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len); + ptlrpc_request_set_replen(req); + RETURN(req); +} +EXPORT_SYMBOL(ldlm_enqueue_pack); + +/** + * Client-side lock enqueue. + * + * If a request has some specific initialisation it is passed in \a reqp, * otherwise it is created in ldlm_cli_enqueue. * - * Supports sync and async requests, pass @async flag accordingly. If a + * Supports sync and async requests, pass \a async flag accordingly. If a * request was created in ldlm_cli_enqueue and it is the async request, - * pass it to the caller in @reqp. */ + * pass it to the caller in \a reqp. + */ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, struct ldlm_enqueue_info *einfo, const struct ldlm_res_id *res_id, @@ -794,7 +875,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, void *lvb, __u32 lvb_len, enum lvb_type lvb_type, struct lustre_handle *lockh, int async) { - struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; + struct ldlm_namespace *ns; struct ldlm_lock *lock; struct ldlm_request *body; int is_replay = *flags & LDLM_FL_REPLAY; @@ -805,6 +886,8 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, LASSERT(exp != NULL); + ns = exp->exp_obd->obd_namespace; + /* If we're replaying this lock, just check some invariants. * If we're creating a new lock, get everything all setup nice. */ if (is_replay) { @@ -813,45 +896,38 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, LDLM_DEBUG(lock, "client-side enqueue START"); LASSERT(exp == lock->l_conn_export); } else { - const struct ldlm_callback_suite cbs = { - .lcs_completion = einfo->ei_cb_cp, - .lcs_blocking = einfo->ei_cb_bl, - .lcs_glimpse = einfo->ei_cb_gl, - .lcs_weigh = einfo->ei_cb_wg - }; - lock = ldlm_lock_create(ns, res_id, einfo->ei_type, - einfo->ei_mode, &cbs, einfo->ei_cbdata, + const struct ldlm_callback_suite cbs = { + .lcs_completion = einfo->ei_cb_cp, + .lcs_blocking = einfo->ei_cb_bl, + .lcs_glimpse = einfo->ei_cb_gl + }; + lock = ldlm_lock_create(ns, res_id, einfo->ei_type, + einfo->ei_mode, &cbs, einfo->ei_cbdata, lvb_len, lvb_type); - if (lock == NULL) - RETURN(-ENOMEM); + if (IS_ERR(lock)) + RETURN(PTR_ERR(lock)); /* for the local lock, add the reference */ ldlm_lock_addref_internal(lock, einfo->ei_mode); ldlm_lock2handle(lock, lockh); - if (policy != NULL) { - /* INODEBITS_INTEROP: If the server does not support - * inodebits, we will request a plain lock in the - * descriptor (ldlm_lock2desc() below) but use an - * inodebits lock internally with both bits set. - */ - if (einfo->ei_type == LDLM_IBITS && - !(exp->exp_connect_flags & OBD_CONNECT_IBITS)) - lock->l_policy_data.l_inodebits.bits = - MDS_INODELOCK_LOOKUP | - MDS_INODELOCK_UPDATE; - else - lock->l_policy_data = *policy; - } + if (policy != NULL) + lock->l_policy_data = *policy; - if (einfo->ei_type == LDLM_EXTENT) - lock->l_req_extent = policy->l_extent; - LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n", + if (einfo->ei_type == LDLM_EXTENT) { + /* extent lock without policy is a bug */ + if (policy == NULL) + LBUG(); + + lock->l_req_extent = policy->l_extent; + } + LDLM_DEBUG(lock, "client-side enqueue START, flags "LPX64"\n", *flags); - } + } lock->l_conn_export = exp; lock->l_export = NULL; lock->l_blocking_ast = einfo->ei_cb_bl; - lock->l_flags |= (*flags & LDLM_FL_NO_LRU); + lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL)); + lock->l_last_activity = cfs_time_current_sec(); /* lock not sent to server yet */ @@ -894,14 +970,6 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp, ptlrpc_request_set_replen(req); } - /* - * Liblustre client doesn't get extent locks, except for O_APPEND case - * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where - * [i_size, OBD_OBJECT_EOF] lock is taken. - */ - LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT || - policy->l_extent.end == OBD_OBJECT_EOF)); - if (async) { LASSERT(reqp != NULL); RETURN(0); @@ -949,7 +1017,7 @@ static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode, ldlm_reprocess_all(res); rc = 0; } else { - rc = EDEADLOCK; + rc = LUSTRE_EDEADLK; } LDLM_DEBUG(lock, "client-side local convert handler END"); LDLM_LOCK_PUT(lock); @@ -1021,7 +1089,7 @@ int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags) GOTO(out, rc); } } else { - rc = EDEADLOCK; + rc = LUSTRE_EDEADLK; } EXIT; out: @@ -1029,13 +1097,14 @@ int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags) ptlrpc_req_finished(req); return rc; } -EXPORT_SYMBOL(ldlm_cli_convert); -/* Cancel locks locally. +/** + * Cancel locks locally. * Returns: - * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server; - * LDLM_FL_CANCELING otherwise; - * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */ + * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server + * \retval LDLM_FL_CANCELING otherwise; + * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC. + */ static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock) { __u64 rc = LDLM_FL_LOCAL_ONLY; @@ -1047,13 +1116,13 @@ static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock) LDLM_DEBUG(lock, "client-side cancel"); /* Set this flag to prevent others from getting new references*/ lock_res_and_lock(lock); - lock->l_flags |= LDLM_FL_CBPENDING; + ldlm_set_cbpending(lock); local_only = !!(lock->l_flags & (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK)); - ldlm_cancel_callback(lock); - rc = (lock->l_flags & LDLM_FL_BL_AST) ? - LDLM_FL_BL_AST : LDLM_FL_CANCELING; - unlock_res_and_lock(lock); + ldlm_cancel_callback(lock); + rc = (ldlm_is_bl_ast(lock)) ? + LDLM_FL_BL_AST : LDLM_FL_CANCELING; + unlock_res_and_lock(lock); if (local_only) { CDEBUG(D_DLMTRACE, "not sending request (at caller's " @@ -1074,10 +1143,11 @@ static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock) RETURN(rc); } -/* Pack @count locks in @head into ldlm_request buffer at the offset @off, - of the request @req. */ +/** + * Pack \a count locks in \a head into ldlm_request buffer of request \a req. + */ static void ldlm_cancel_pack(struct ptlrpc_request *req, - cfs_list_t *head, int count) + struct list_head *head, int count) { struct ldlm_request *dlm; struct ldlm_lock *lock; @@ -1097,7 +1167,7 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req, /* XXX: it would be better to pack lock handles grouped by resource. * so that the server cancel would call filter_lvbo_update() less * frequently. */ - cfs_list_for_each_entry(lock, head, l_bl_ast) { + list_for_each_entry(lock, head, l_bl_ast) { if (!count--) break; LASSERT(lock->l_conn_export); @@ -1110,9 +1180,10 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req, EXIT; } -/* Prepare and send a batched cancel rpc, it will include count lock handles - * of locks given in @head. */ -int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels, +/** + * Prepare and send a batched cancel RPC. It will include \a count lock + * handles of locks given in \a cancels list. */ +int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels, int count, ldlm_cancel_flags_t flags) { struct ptlrpc_request *req = NULL; @@ -1170,7 +1241,7 @@ int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels, } else { rc = ptlrpc_queue_wait(req); } - if (rc == ESTALE) { + if (rc == LUSTRE_ESTALE) { CDEBUG(D_DLMTRACE, "client/server (nid %s) " "out of sync -- not fatal\n", libcfs_nid2str(req->rq_import-> @@ -1181,10 +1252,11 @@ int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels, ptlrpc_req_finished(req); continue; } else if (rc != ELDLM_OK) { - if (rc != -ESHUTDOWN) - CERROR("Got rc %d from cancel RPC: canceling " - "anyway\n", rc); - break; + /* -ESHUTDOWN is common on umount */ + CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR, + "Got rc %d from cancel RPC: " + "canceling anyway\n", rc); + break; } sent = count; break; @@ -1195,7 +1267,6 @@ int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels, out: return sent ? sent : rc; } -EXPORT_SYMBOL(ldlm_cli_cancel_req); static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp) { @@ -1204,7 +1275,7 @@ static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp) } /** - * Update client's obd pool related fields with new SLV and Limit from \a req. + * Update client's OBD pool related fields with new SLV and Limit from \a req. */ int ldlm_cli_update_pool(struct ptlrpc_request *req) { @@ -1221,13 +1292,11 @@ int ldlm_cli_update_pool(struct ptlrpc_request *req) RETURN(0); } - /* - * In some cases RPC may contain slv and limit zeroed out. This is - * the case when server does not support lru resize feature. This is - * also possible in some recovery cases when server side reqs have no - * ref to obd export and thus access to server side namespace is no - * possible. - */ + /* In some cases RPC may contain SLV and limit zeroed out. This + * is the case when server does not support LRU resize feature. + * This is also possible in some recovery cases when server-side + * reqs have no reference to the OBD export and thus access to + * server-side namespace is not possible. */ if (lustre_msg_get_slv(req->rq_repmsg) == 0 || lustre_msg_get_limit(req->rq_repmsg) == 0) { DEBUG_REQ(D_HA, req, "Zero SLV or Limit found " @@ -1241,13 +1310,11 @@ int ldlm_cli_update_pool(struct ptlrpc_request *req) new_slv = lustre_msg_get_slv(req->rq_repmsg); obd = req->rq_import->imp_obd; - /* - * Set new SLV and Limit to obd fields to make accessible for pool - * thread. We do not access obd_namespace and pool directly here - * as there is no reliable way to make sure that they are still - * alive in cleanup time. Evil races are possible which may cause - * oops in that time. - */ + /* Set new SLV and limit in OBD fields to make them accessible + * to the pool thread. We do not access obd_namespace and pool + * directly here as there is no reliable way to make sure that + * they are still alive at cleanup time. Evil races are possible + * which may cause Oops at that time. */ write_lock(&obd->obd_pool_lock); obd->obd_pool_slv = new_slv; obd->obd_pool_limit = new_limit; @@ -1255,35 +1322,40 @@ int ldlm_cli_update_pool(struct ptlrpc_request *req) RETURN(0); } -EXPORT_SYMBOL(ldlm_cli_update_pool); -int ldlm_cli_cancel(struct lustre_handle *lockh) +/** + * Client side lock cancel. + * + * Lock must not have any readers or writers by this time. + */ +int ldlm_cli_cancel(struct lustre_handle *lockh, + ldlm_cancel_flags_t cancel_flags) { - struct obd_export *exp; + struct obd_export *exp; int avail, flags, count = 1; __u64 rc = 0; - struct ldlm_namespace *ns; - struct ldlm_lock *lock; - CFS_LIST_HEAD(cancels); - ENTRY; + struct ldlm_namespace *ns; + struct ldlm_lock *lock; + struct list_head cancels = LIST_HEAD_INIT(cancels); + ENTRY; /* concurrent cancels on the same handle can happen */ lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING); - if (lock == NULL) { - LDLM_DEBUG_NOLOCK("lock is already being destroyed\n"); - RETURN(0); - } + if (lock == NULL) { + LDLM_DEBUG_NOLOCK("lock is already being destroyed"); + RETURN(0); + } - rc = ldlm_cli_cancel_local(lock); - if (rc == LDLM_FL_LOCAL_ONLY) { - LDLM_LOCK_RELEASE(lock); + rc = ldlm_cli_cancel_local(lock); + if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) { + LDLM_LOCK_RELEASE(lock); RETURN(0); - } - /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL - * rpc which goes to canceld portal, so we can cancel other lru locks - * here and send them all as one LDLM_CANCEL rpc. */ - LASSERT(cfs_list_empty(&lock->l_bl_ast)); - cfs_list_add(&lock->l_bl_ast, &cancels); + } + /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL + * RPC which goes to canceld portal, so we can cancel other LRU locks + * here and send them all as one LDLM_CANCEL RPC. */ + LASSERT(list_empty(&lock->l_bl_ast)); + list_add(&lock->l_bl_ast, &cancels); exp = lock->l_conn_export; if (exp_connect_cancelset(exp)) { @@ -1298,23 +1370,25 @@ int ldlm_cli_cancel(struct lustre_handle *lockh) count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1, LCF_BL_AST, flags); } - ldlm_cli_cancel_list(&cancels, count, NULL, 0); + ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags); RETURN(0); } EXPORT_SYMBOL(ldlm_cli_cancel); -/* XXX until we will have compound requests and can cut cancels from generic rpc - * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */ -int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count, - ldlm_cancel_flags_t flags) +/** + * Locally cancel up to \a count locks in list \a cancels. + * Return the number of cancelled locks. + */ +int ldlm_cli_cancel_list_local(struct list_head *cancels, int count, + ldlm_cancel_flags_t flags) { - CFS_LIST_HEAD(head); - struct ldlm_lock *lock, *next; + struct list_head head = LIST_HEAD_INIT(head); + struct ldlm_lock *lock, *next; int left = 0, bl_ast = 0; __u64 rc; left = count; - cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) { + list_for_each_entry_safe(lock, next, cancels, l_bl_ast) { if (left-- == 0) break; @@ -1324,20 +1398,23 @@ int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count, } else { rc = ldlm_cli_cancel_local(lock); } - if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) { - LDLM_DEBUG(lock, "Cancel lock separately"); - cfs_list_del_init(&lock->l_bl_ast); - cfs_list_add(&lock->l_bl_ast, &head); - bl_ast ++; + /* Until we have compound requests and can send LDLM_CANCEL + * requests batched with generic RPCs, we need to send cancels + * with the LDLM_FL_BL_AST flag in a separate RPC from + * the one being generated now. */ + if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) { + LDLM_DEBUG(lock, "Cancel lock separately"); + list_del_init(&lock->l_bl_ast); + list_add(&lock->l_bl_ast, &head); + bl_ast++; continue; } if (rc == LDLM_FL_LOCAL_ONLY) { /* CANCEL RPC should not be sent to server. */ - cfs_list_del_init(&lock->l_bl_ast); + list_del_init(&lock->l_bl_ast); LDLM_LOCK_RELEASE(lock); count--; } - } if (bl_ast > 0) { count -= bl_ast; @@ -1346,11 +1423,10 @@ int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count, RETURN(count); } -EXPORT_SYMBOL(ldlm_cli_cancel_list_local); /** - * Cancel as many locks as possible w/o sending any rpcs (e.g. to write back - * dirty data, to close a file, ...) or waiting for any rpcs in-flight (e.g. + * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back + * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g. * readahead requests, ...) */ static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, @@ -1358,31 +1434,32 @@ static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, int unused, int added, int count) { - ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK; - ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery; - lock_res_and_lock(lock); - - /* don't check added & count since we want to process all locks - * from unused list */ - switch (lock->l_resource->lr_type) { - case LDLM_EXTENT: - case LDLM_IBITS: - if (cb && cb(lock)) - break; - default: - result = LDLM_POLICY_SKIP_LOCK; - lock->l_flags |= LDLM_FL_SKIPPED; - break; - } + ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK; + + /* don't check added & count since we want to process all locks + * from unused list. + * It's fine to not take lock to access lock->l_resource since + * the lock has already been granted so it won't change. */ + switch (lock->l_resource->lr_type) { + case LDLM_EXTENT: + case LDLM_IBITS: + if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0) + break; + default: + result = LDLM_POLICY_SKIP_LOCK; + lock_res_and_lock(lock); + ldlm_set_skipped(lock); + unlock_res_and_lock(lock); + break; + } - unlock_res_and_lock(lock); - RETURN(result); + RETURN(result); } /** - * Callback function for lru-resize policy. Makes decision whether to keep - * \a lock in LRU for current \a LRU size \a unused, added in current scan - * \a added and number of locks to be preferably canceled \a count. + * Callback function for LRU-resize policy. Decides whether to keep + * \a lock in LRU for current \a LRU size \a unused, added in current + * scan \a added and number of locks to be preferably canceled \a count. * * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning * @@ -1393,41 +1470,43 @@ static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns, int unused, int added, int count) { - cfs_time_t cur = cfs_time_current(); - struct ldlm_pool *pl = &ns->ns_pool; - __u64 slv, lvf, lv; - cfs_time_t la; - - /* - * Stop lru processing when we reached passed @count or checked all - * locks in lru. - */ - if (count && added >= count) - return LDLM_POLICY_KEEP_LOCK; - - slv = ldlm_pool_get_slv(pl); - lvf = ldlm_pool_get_lvf(pl); - la = cfs_duration_sec(cfs_time_sub(cur, - lock->l_last_used)); - - /* - * Stop when slv is not yet come from server or lv is smaller than - * it is. - */ - lv = lvf * la * unused; - - /* - * Inform pool about current CLV to see it via proc. - */ - ldlm_pool_set_clv(pl, lv); - return (slv == 0 || lv < slv) ? - LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; + cfs_time_t cur = cfs_time_current(); + struct ldlm_pool *pl = &ns->ns_pool; + __u64 slv, lvf, lv; + cfs_time_t la; + + /* Stop LRU processing when we reach past @count or have checked all + * locks in LRU. */ + if (count && added >= count) + return LDLM_POLICY_KEEP_LOCK; + + /* Despite of the LV, It doesn't make sense to keep the lock which + * is unused for ns_max_age time. */ + if (cfs_time_after(cfs_time_current(), + cfs_time_add(lock->l_last_used, ns->ns_max_age))) + return LDLM_POLICY_CANCEL_LOCK; + + slv = ldlm_pool_get_slv(pl); + lvf = ldlm_pool_get_lvf(pl); + la = cfs_duration_sec(cfs_time_sub(cur, + lock->l_last_used)); + lv = lvf * la * unused; + + /* Inform pool about current CLV to see it via proc. */ + ldlm_pool_set_clv(pl, lv); + + /* Stop when SLV is not yet come from server or lv is smaller than + * it is. */ + if (slv == 0 || lv < slv) + return LDLM_POLICY_KEEP_LOCK; + + return LDLM_POLICY_CANCEL_LOCK; } /** * Callback function for proc used policy. Makes decision whether to keep - * \a lock in LRU for current \a LRU size \a unused, added in current scan - * \a added and number of locks to be preferably canceled \a count. + * \a lock in LRU for current \a LRU size \a unused, added in current scan \a + * added and number of locks to be preferably canceled \a count. * * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning * @@ -1438,57 +1517,50 @@ static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns, int unused, int added, int count) { - /* - * Stop lru processing when we reached passed @count or checked all - * locks in lru. - */ - return (added >= count) ? - LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; + /* Stop LRU processing when we reach past @count or have checked all + * locks in LRU. */ + return (added >= count) ? + LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; } /** - * Callback function for aged policy. Makes decision whether to keep - * \a lock in LRU for current \a LRU size \a unused, added in current scan - * \a added and number of locks to be preferably canceled \a count. + * Callback function for aged policy. Makes decision whether to keep \a lock in + * LRU for current LRU size \a unused, added in current scan \a added and + * number of locks to be preferably canceled \a count. * * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning * * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU */ static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns, - struct ldlm_lock *lock, - int unused, int added, - int count) + struct ldlm_lock *lock, + int unused, int added, + int count) { - /* - * Stop lru processing if young lock is found and we reached passed - * @count. - */ - return ((added >= count) && - cfs_time_before(cfs_time_current(), - cfs_time_add(lock->l_last_used, - ns->ns_max_age))) ? - LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; + if ((added >= count) && + cfs_time_before(cfs_time_current(), + cfs_time_add(lock->l_last_used, ns->ns_max_age))) + return LDLM_POLICY_KEEP_LOCK; + + return LDLM_POLICY_CANCEL_LOCK; } /** - * Callback function for default policy. Makes decision whether to keep - * \a lock in LRU for current \a LRU size \a unused, added in current scan - * \a added and number of locks to be preferably canceled \a count. + * Callback function for default policy. Makes decision whether to keep \a lock + * in LRU for current LRU size \a unused, added in current scan \a added and + * number of locks to be preferably canceled \a count. * * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning * * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU */ static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns, - struct ldlm_lock *lock, - int unused, int added, - int count) + struct ldlm_lock *lock, + int unused, int added, + int count) { - /* - * Stop lru processing when we reached passed @count or checked all - * locks in lru. - */ + /* Stop LRU processing when we reach past count or have checked all + * locks in LRU. */ return (added >= count) ? LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK; } @@ -1519,39 +1591,42 @@ ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags) return ldlm_cancel_default_policy; } -/* - Free space in lru for @count new locks, +/** + * - Free space in LRU for \a count new locks, * redundant unused locks are canceled locally; * - also cancel locally unused aged locks; - * - do not cancel more than @max locks; - * - GET the found locks and add them into the @cancels list. + * - do not cancel more than \a max locks; + * - GET the found locks and add them into the \a cancels list. * * A client lock can be added to the l_bl_ast list only when it is - * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL. - * There are the following use cases: ldlm_cancel_resource_local(), - * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this - * flag properly. As any attempt to cancel a lock rely on this flag, - * l_bl_ast list is accessed later without any special locking. + * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing + * CANCEL. There are the following use cases: + * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and + * ldlm_cli_cancel(), which check and set this flag properly. As any + * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed + * later without any special locking. * - * Calling policies for enabled lru resize: + * Calling policies for enabled LRU resize: * ---------------------------------------- - * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to - * cancel not more than @count locks; + * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to + * cancel not more than \a count locks; * - * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at - * the beginning of lru list); + * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at + * the beginning of LRU list); * - * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to + * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to * memory pressre policy function; * - * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy". + * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy". * * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible * (typically before replaying locks) w/o - * sending any rpcs or waiting for any - * outstanding rpc to complete. + * sending any RPCs or waiting for any + * outstanding RPC to complete. */ -static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels, - int count, int max, int flags) +static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, + struct list_head *cancels, int count, int max, + int flags) { ldlm_cancel_lru_policy_t pf; struct ldlm_lock *lock, *next; @@ -1568,7 +1643,7 @@ static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels, pf = ldlm_cancel_lru_policy(ns, flags); LASSERT(pf != NULL); - while (!cfs_list_empty(&ns->ns_unused_list)) { + while (!list_empty(&ns->ns_unused_list)) { ldlm_policy_res_t result; /* all unused locks */ @@ -1579,102 +1654,100 @@ static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels, if (max && added >= max) break; - cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list, - l_lru){ + list_for_each_entry_safe(lock, next, &ns->ns_unused_list, + l_lru) { /* No locks which got blocking requests. */ - LASSERT(!(lock->l_flags & LDLM_FL_BL_AST)); + LASSERT(!ldlm_is_bl_ast(lock)); - if (flags & LDLM_CANCEL_NO_WAIT && - lock->l_flags & LDLM_FL_SKIPPED) - /* already processed */ - continue; + if (flags & LDLM_CANCEL_NO_WAIT && + ldlm_is_skipped(lock)) + /* already processed */ + continue; - /* Somebody is already doing CANCEL. No need in this - * lock in lru, do not traverse it again. */ - if (!(lock->l_flags & LDLM_FL_CANCELING)) + /* Somebody is already doing CANCEL. No need for this + * lock in LRU, do not traverse it again. */ + if (!ldlm_is_canceling(lock)) break; - ldlm_lock_remove_from_lru_nolock(lock); - } - if (&lock->l_lru == &ns->ns_unused_list) - break; + ldlm_lock_remove_from_lru_nolock(lock); + } + if (&lock->l_lru == &ns->ns_unused_list) + break; - LDLM_LOCK_GET(lock); + LDLM_LOCK_GET(lock); spin_unlock(&ns->ns_lock); - lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current()); - - /* Pass the lock through the policy filter and see if it - * should stay in lru. - * - * Even for shrinker policy we stop scanning if - * we find a lock that should stay in the cache. - * We should take into account lock age anyway - * as new lock even if it is small of weight is - * valuable resource. - * - * That is, for shrinker policy we drop only - * old locks, but additionally chose them by - * their weight. Big extent locks will stay in - * the cache. */ - result = pf(ns, lock, unused, added, count); - if (result == LDLM_POLICY_KEEP_LOCK) { - lu_ref_del(&lock->l_reference, - __FUNCTION__, cfs_current()); - LDLM_LOCK_RELEASE(lock); + lu_ref_add(&lock->l_reference, __FUNCTION__, current); + + /* Pass the lock through the policy filter and see if it + * should stay in LRU. + * + * Even for shrinker policy we stop scanning if + * we find a lock that should stay in the cache. + * We should take into account lock age anyway + * as a new lock is a valuable resource even if + * it has a low weight. + * + * That is, for shrinker policy we drop only + * old locks, but additionally choose them by + * their weight. Big extent locks will stay in + * the cache. */ + result = pf(ns, lock, unused, added, count); + if (result == LDLM_POLICY_KEEP_LOCK) { + lu_ref_del(&lock->l_reference, + __FUNCTION__, current); + LDLM_LOCK_RELEASE(lock); spin_lock(&ns->ns_lock); break; } if (result == LDLM_POLICY_SKIP_LOCK) { lu_ref_del(&lock->l_reference, - __func__, cfs_current()); + __func__, current); LDLM_LOCK_RELEASE(lock); spin_lock(&ns->ns_lock); - continue; - } + continue; + } - lock_res_and_lock(lock); - /* Check flags again under the lock. */ - if ((lock->l_flags & LDLM_FL_CANCELING) || - (ldlm_lock_remove_from_lru(lock) == 0)) { - /* other thread is removing lock from lru or - * somebody is already doing CANCEL or - * there is a blocking request which will send - * cancel by itseft or the lock is matched - * is already not unused. */ - unlock_res_and_lock(lock); - lu_ref_del(&lock->l_reference, - __FUNCTION__, cfs_current()); - LDLM_LOCK_RELEASE(lock); + lock_res_and_lock(lock); + /* Check flags again under the lock. */ + if (ldlm_is_canceling(lock) || + (ldlm_lock_remove_from_lru(lock) == 0)) { + /* Another thread is removing lock from LRU, or + * somebody is already doing CANCEL, or there + * is a blocking request which will send cancel + * by itself, or the lock is no longer unused. */ + unlock_res_and_lock(lock); + lu_ref_del(&lock->l_reference, __FUNCTION__, current); + LDLM_LOCK_RELEASE(lock); spin_lock(&ns->ns_lock); - continue; - } - LASSERT(!lock->l_readers && !lock->l_writers); - - /* If we have chosen to cancel this lock voluntarily, we - * better send cancel notification to server, so that it - * frees appropriate state. This might lead to a race - * where while we are doing cancel here, server is also - * silently cancelling this lock. */ - lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK; - - /* Setting the CBPENDING flag is a little misleading, - * but prevents an important race; namely, once - * CBPENDING is set, the lock can accumulate no more - * readers/writers. Since readers and writers are - * already zero here, ldlm_lock_decref() won't see - * this flag and call l_blocking_ast */ - lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING; - - /* We can't re-add to l_lru as it confuses the - * refcounting in ldlm_lock_remove_from_lru() if an AST - * arrives after we drop lr_lock below. We use l_bl_ast - * and can't use l_pending_chain as it is used both on - * server and client nevertheless bug 5666 says it is - * used only on server */ - LASSERT(cfs_list_empty(&lock->l_bl_ast)); - cfs_list_add(&lock->l_bl_ast, cancels); - unlock_res_and_lock(lock); - lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current()); + continue; + } + LASSERT(!lock->l_readers && !lock->l_writers); + + /* If we have chosen to cancel this lock voluntarily, we + * better send cancel notification to server, so that it + * frees appropriate state. This might lead to a race + * where while we are doing cancel here, server is also + * silently cancelling this lock. */ + ldlm_clear_cancel_on_block(lock); + + /* Setting the CBPENDING flag is a little misleading, + * but prevents an important race; namely, once + * CBPENDING is set, the lock can accumulate no more + * readers/writers. Since readers and writers are + * already zero here, ldlm_lock_decref() won't see + * this flag and call l_blocking_ast */ + lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING; + + /* We can't re-add to l_lru as it confuses the + * refcounting in ldlm_lock_remove_from_lru() if an AST + * arrives after we drop lr_lock below. We use l_bl_ast + * and can't use l_pending_chain as it is used both on + * server and client nevertheless bug 5666 says it is + * used only on server */ + LASSERT(list_empty(&lock->l_bl_ast)); + list_add(&lock->l_bl_ast, cancels); + unlock_res_and_lock(lock); + lu_ref_del(&lock->l_reference, __FUNCTION__, current); spin_lock(&ns->ns_lock); added++; unused--; @@ -1683,7 +1756,7 @@ static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels, RETURN(added); } -int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels, +int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels, int count, int max, ldlm_cancel_flags_t cancel_flags, int flags) { @@ -1694,45 +1767,49 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels, return ldlm_cli_cancel_list_local(cancels, added, cancel_flags); } -/* when called with LDLM_ASYNC the blocking callback will be handled +/** + * Cancel at least \a nr locks from given namespace LRU. + * + * When called with LCF_ASYNC the blocking callback will be handled * in a thread and this function will return after the thread has been - * asked to call the callback. when called with LDLM_SYNC the blocking - * callback will be performed in this function. */ -int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t mode, - int flags) + * asked to call the callback. When called with LCF_ASYNC the blocking + * callback will be performed in this function. + */ +int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, + ldlm_cancel_flags_t cancel_flags, + int flags) { - CFS_LIST_HEAD(cancels); - int count, rc; - ENTRY; + struct list_head cancels = LIST_HEAD_INIT(cancels); + int count, rc; + ENTRY; -#ifndef __KERNEL__ - mode = LDLM_SYNC; /* force to be sync in user space */ -#endif - /* Just prepare the list of locks, do not actually cancel them yet. - * Locks are cancelled later in a separate thread. */ - count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags); - rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, mode); - if (rc == 0) - RETURN(count); + /* Just prepare the list of locks, do not actually cancel them yet. + * Locks are cancelled later in a separate thread. */ + count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags); + rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags); + if (rc == 0) + RETURN(count); - RETURN(0); + RETURN(0); } -/* Find and cancel locally unused locks found on resource, matched to the - * given policy, mode. GET the found locks and add them into the @cancels - * list. */ +/** + * Find and cancel locally unused locks found on resource, matched to the + * given policy, mode. GET the found locks and add them into the \a cancels + * list. + */ int ldlm_cancel_resource_local(struct ldlm_resource *res, - cfs_list_t *cancels, - ldlm_policy_data_t *policy, - ldlm_mode_t mode, int lock_flags, - ldlm_cancel_flags_t cancel_flags, void *opaque) + struct list_head *cancels, + ldlm_policy_data_t *policy, + ldlm_mode_t mode, __u64 lock_flags, + ldlm_cancel_flags_t cancel_flags, void *opaque) { struct ldlm_lock *lock; int count = 0; ENTRY; lock_res(res); - cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) { + list_for_each_entry(lock, &res->lr_granted, l_res_link) { if (opaque != NULL && lock->l_ast_data != opaque) { LDLM_ERROR(lock, "data %p doesn't match opaque %p", lock->l_ast_data, opaque); @@ -1743,11 +1820,10 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res, if (lock->l_readers || lock->l_writers) continue; - /* If somebody is already doing CANCEL, or blocking ast came, - * skip this lock. */ - if (lock->l_flags & LDLM_FL_BL_AST || - lock->l_flags & LDLM_FL_CANCELING) - continue; + /* If somebody is already doing CANCEL, or blocking AST came, + * skip this lock. */ + if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock)) + continue; if (lockmode_compat(lock->l_granted_mode, mode)) continue; @@ -1759,12 +1835,12 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res, policy->l_inodebits.bits)) continue; - /* See CBPENDING comment in ldlm_cancel_lru */ - lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING | - lock_flags; + /* See CBPENDING comment in ldlm_cancel_lru */ + lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING | + lock_flags; - LASSERT(cfs_list_empty(&lock->l_bl_ast)); - cfs_list_add(&lock->l_bl_ast, cancels); + LASSERT(list_empty(&lock->l_bl_ast)); + list_add(&lock->l_bl_ast, cancels); LDLM_LOCK_GET(lock); count++; } @@ -1774,30 +1850,34 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res, } EXPORT_SYMBOL(ldlm_cancel_resource_local); -/* If @req is NULL, send CANCEL request to server with handles of locks - * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests +/** + * Cancel client-side locks from a list and send/prepare cancel RPCs to the + * server. + * If \a req is NULL, send CANCEL request to server with handles of locks + * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests * separately per lock. - * If @req is not NULL, put handles of locks in @cancels into the request - * buffer at the offset @off. - * Destroy @cancels at the end. */ -int ldlm_cli_cancel_list(cfs_list_t *cancels, int count, + * If \a req is not NULL, put handles of locks in \a cancels into the request + * buffer at the offset \a off. + * Destroy \a cancels at the end. + */ +int ldlm_cli_cancel_list(struct list_head *cancels, int count, struct ptlrpc_request *req, ldlm_cancel_flags_t flags) { struct ldlm_lock *lock; int res = 0; ENTRY; - if (cfs_list_empty(cancels) || count == 0) + if (list_empty(cancels) || count == 0) RETURN(0); /* XXX: requests (both batched and not) could be sent in parallel. * Usually it is enough to have just 1 RPC, but it is possible that - * there are to many locks to be cancelled in LRU or on a resource. + * there are too many locks to be cancelled in LRU or on a resource. * It would also speed up the case when the server does not support * the feature. */ while (count > 0) { - LASSERT(!cfs_list_empty(cancels)); - lock = cfs_list_entry(cancels->next, struct ldlm_lock, + LASSERT(!list_empty(cancels)); + lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast); LASSERT(lock->l_conn_export); @@ -1814,11 +1894,11 @@ int ldlm_cli_cancel_list(cfs_list_t *cancels, int count, cancels, 1, flags); } - if (res < 0) { - if (res != -ESHUTDOWN) - CERROR("ldlm_cli_cancel_list: %d\n", res); - res = count; - } + if (res < 0) { + CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR, + "ldlm_cli_cancel_list: %d\n", res); + res = count; + } count -= res; ldlm_lock_list_put(cancels, l_bl_ast, res); @@ -1828,36 +1908,42 @@ int ldlm_cli_cancel_list(cfs_list_t *cancels, int count, } EXPORT_SYMBOL(ldlm_cli_cancel_list); +/** + * Cancel all locks on a resource that have 0 readers/writers. + * + * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying + * to notify the server. */ int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns, - const struct ldlm_res_id *res_id, - ldlm_policy_data_t *policy, - ldlm_mode_t mode, - ldlm_cancel_flags_t flags, - void *opaque) + const struct ldlm_res_id *res_id, + ldlm_policy_data_t *policy, + ldlm_mode_t mode, + ldlm_cancel_flags_t flags, + void *opaque) { - struct ldlm_resource *res; - CFS_LIST_HEAD(cancels); - int count; - int rc; - ENTRY; - - res = ldlm_resource_get(ns, NULL, res_id, 0, 0); - if (res == NULL) { - /* This is not a problem. */ - CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]); - RETURN(0); - } + struct ldlm_resource *res; + struct list_head cancels = LIST_HEAD_INIT(cancels); + int count; + int rc; + ENTRY; - LDLM_RESOURCE_ADDREF(res); - count = ldlm_cancel_resource_local(res, &cancels, policy, mode, - 0, flags | LCF_BL_AST, opaque); - rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags); - if (rc != ELDLM_OK) - CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc); + res = ldlm_resource_get(ns, NULL, res_id, 0, 0); + if (IS_ERR(res)) { + /* This is not a problem. */ + CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]); + RETURN(0); + } - LDLM_RESOURCE_DELREF(res); - ldlm_resource_putref(res); - RETURN(0); + LDLM_RESOURCE_ADDREF(res); + count = ldlm_cancel_resource_local(res, &cancels, policy, mode, + 0, flags | LCF_BL_AST, opaque); + rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags); + if (rc != ELDLM_OK) + CERROR("canceling unused lock "DLDLMRES": rc = %d\n", + PLDLMRES(res), rc); + + LDLM_RESOURCE_DELREF(res); + ldlm_resource_putref(res); + RETURN(0); } EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource); @@ -1867,24 +1953,20 @@ struct ldlm_cli_cancel_arg { }; static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd, - cfs_hlist_node_t *hnode, void *arg) + struct hlist_node *hnode, void *arg) { - struct ldlm_resource *res = cfs_hash_object(hs, hnode); - struct ldlm_cli_cancel_arg *lc = arg; - int rc; - - rc = ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name, - NULL, LCK_MINMODE, - lc->lc_flags, lc->lc_opaque); - if (rc != 0) { - CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n", - res->lr_name.name[0], rc); - } - /* must return 0 for hash iteration */ - return 0; + struct ldlm_resource *res = cfs_hash_object(hs, hnode); + struct ldlm_cli_cancel_arg *lc = arg; + + ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name, + NULL, LCK_MINMODE, lc->lc_flags, + lc->lc_opaque); + /* must return 0 for hash iteration */ + return 0; } -/* Cancel all locks on a namespace (or a specific resource, if given) +/** + * Cancel all locks on a namespace (or a specific resource, if given) * that have 0 readers/writers. * * If flags & LCF_LOCAL, throw the locks away without trying @@ -1913,14 +1995,13 @@ int ldlm_cli_cancel_unused(struct ldlm_namespace *ns, RETURN(ELDLM_OK); } } -EXPORT_SYMBOL(ldlm_cli_cancel_unused); /* Lock iterators. */ int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter, void *closure) { - cfs_list_t *tmp, *next; + struct list_head *tmp, *next; struct ldlm_lock *lock; int rc = LDLM_ITER_CONTINUE; @@ -1930,22 +2011,22 @@ int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter, RETURN(LDLM_ITER_CONTINUE); lock_res(res); - cfs_list_for_each_safe(tmp, next, &res->lr_granted) { - lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link); + list_for_each_safe(tmp, next, &res->lr_granted) { + lock = list_entry(tmp, struct ldlm_lock, l_res_link); if (iter(lock, closure) == LDLM_ITER_STOP) GOTO(out, rc = LDLM_ITER_STOP); } - cfs_list_for_each_safe(tmp, next, &res->lr_converting) { - lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link); + list_for_each_safe(tmp, next, &res->lr_converting) { + lock = list_entry(tmp, struct ldlm_lock, l_res_link); if (iter(lock, closure) == LDLM_ITER_STOP) GOTO(out, rc = LDLM_ITER_STOP); } - cfs_list_for_each_safe(tmp, next, &res->lr_waiting) { - lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link); + list_for_each_safe(tmp, next, &res->lr_waiting) { + lock = list_entry(tmp, struct ldlm_lock, l_res_link); if (iter(lock, closure) == LDLM_ITER_STOP) GOTO(out, rc = LDLM_ITER_STOP); @@ -1954,7 +2035,6 @@ int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter, unlock_res(res); RETURN(rc); } -EXPORT_SYMBOL(ldlm_resource_foreach); struct iter_helper_data { ldlm_iterator_t iter; @@ -1968,7 +2048,7 @@ static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure) } static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd, - cfs_hlist_node_t *hnode, void *arg) + struct hlist_node *hnode, void *arg) { struct ldlm_resource *res = cfs_hash_object(hs, hnode); @@ -1981,13 +2061,12 @@ void ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter, void *closure) { - struct iter_helper_data helper = { iter: iter, closure: closure }; + struct iter_helper_data helper = { .iter = iter, .closure = closure }; cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_res_iter_helper, &helper); } -EXPORT_SYMBOL(ldlm_namespace_foreach); /* non-blocking function to manipulate a lock whose cb_data is being put away. * return 0: find no resource @@ -1995,27 +2074,24 @@ EXPORT_SYMBOL(ldlm_namespace_foreach); * < 0: errors */ int ldlm_resource_iterate(struct ldlm_namespace *ns, - const struct ldlm_res_id *res_id, - ldlm_iterator_t iter, void *data) + const struct ldlm_res_id *res_id, + ldlm_iterator_t iter, void *data) { - struct ldlm_resource *res; - int rc; - ENTRY; + struct ldlm_resource *res; + int rc; + ENTRY; - if (ns == NULL) { - CERROR("must pass in namespace\n"); - LBUG(); - } + LASSERTF(ns != NULL, "must pass in namespace\n"); - res = ldlm_resource_get(ns, NULL, res_id, 0, 0); - if (res == NULL) - RETURN(0); + res = ldlm_resource_get(ns, NULL, res_id, 0, 0); + if (IS_ERR(res)) + RETURN(0); - LDLM_RESOURCE_ADDREF(res); - rc = ldlm_resource_foreach(res, iter, data); - LDLM_RESOURCE_DELREF(res); - ldlm_resource_putref(res); - RETURN(rc); + LDLM_RESOURCE_ADDREF(res); + rc = ldlm_resource_foreach(res, iter, data); + LDLM_RESOURCE_DELREF(res); + ldlm_resource_putref(res); + RETURN(rc); } EXPORT_SYMBOL(ldlm_resource_iterate); @@ -2023,10 +2099,10 @@ EXPORT_SYMBOL(ldlm_resource_iterate); static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure) { - cfs_list_t *list = closure; + struct list_head *list = closure; /* we use l_pending_chain here, because it's unused on clients. */ - LASSERTF(cfs_list_empty(&lock->l_pending_chain), + LASSERTF(list_empty(&lock->l_pending_chain), "lock %p next %p prev %p\n", lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev); /* bug 9573: don't replay locks left after eviction, or @@ -2034,7 +2110,7 @@ static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure) * on a lock so that it does not disapear under us (e.g. due to cancel) */ if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) { - cfs_list_add(&lock->l_pending_chain, list); + list_add(&lock->l_pending_chain, list); LDLM_LOCK_GET(lock); } @@ -2042,18 +2118,17 @@ static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure) } static int replay_lock_interpret(const struct lu_env *env, - struct ptlrpc_request *req, - struct ldlm_async_args *aa, int rc) + struct ptlrpc_request *req, + struct ldlm_async_args *aa, int rc) { - struct ldlm_lock *lock; - struct ldlm_reply *reply; - struct obd_export *exp; - - ENTRY; - cfs_atomic_dec(&req->rq_import->imp_replay_inflight); - if (rc != ELDLM_OK) - GOTO(out, rc); + struct ldlm_lock *lock; + struct ldlm_reply *reply; + struct obd_export *exp; + ENTRY; + atomic_dec(&req->rq_import->imp_replay_inflight); + if (rc != ELDLM_OK) + GOTO(out, rc); reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); if (reply == NULL) @@ -2103,7 +2178,7 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) /* Bug 11974: Do not replay a lock which is actively being canceled */ - if (lock->l_flags & LDLM_FL_CANCELING) { + if (ldlm_is_canceling(lock)) { LDLM_DEBUG(lock, "Not replaying canceled lock:"); RETURN(0); } @@ -2111,7 +2186,7 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) /* If this is reply-less callback lock, we cannot replay it, since * server might have long dropped it, but notification of that event was * lost by network. (and server granted conflicting lock already) */ - if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) { + if (ldlm_is_cancel_on_block(lock)) { LDLM_DEBUG(lock, "Not replaying reply-less lock:"); ldlm_lock_cancel(lock); RETURN(0); @@ -2135,7 +2210,7 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED; else if (lock->l_granted_mode) flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV; - else if (!cfs_list_empty(&lock->l_res_link)) + else if (!list_empty(&lock->l_res_link)) flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT; else flags = LDLM_FL_REPLAY; @@ -2162,18 +2237,18 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) * also, we mark the request to be put on a dedicated * queue to be processed after all request replayes. * bug 6063 */ - lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE); + lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE); - LDLM_DEBUG(lock, "replaying lock:"); + LDLM_DEBUG(lock, "replaying lock:"); - cfs_atomic_inc(&req->rq_import->imp_replay_inflight); - CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args)); - aa = ptlrpc_req_async_args(req); - aa->lock_handle = body->lock_handle[0]; - req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret; - ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1); + atomic_inc(&req->rq_import->imp_replay_inflight); + CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args)); + aa = ptlrpc_req_async_args(req); + aa->lock_handle = body->lock_handle[0]; + req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret; + ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1); - RETURN(0); + RETURN(0); } /** @@ -2188,58 +2263,57 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock) */ static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns) { - int canceled; - CFS_LIST_HEAD(cancels); + int canceled; + struct list_head cancels = LIST_HEAD_INIT(cancels); - CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before" - "replay for namespace %s (%d)\n", - ldlm_ns_name(ns), ns->ns_nr_unused); + CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before" + "replay for namespace %s (%d)\n", + ldlm_ns_name(ns), ns->ns_nr_unused); - /* We don't need to care whether or not LRU resize is enabled - * because the LDLM_CANCEL_NO_WAIT policy doesn't use the - * count parameter */ - canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0, - LCF_LOCAL, LDLM_CANCEL_NO_WAIT); + /* We don't need to care whether or not LRU resize is enabled + * because the LDLM_CANCEL_NO_WAIT policy doesn't use the + * count parameter */ + canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0, + LCF_LOCAL, LDLM_CANCEL_NO_WAIT); - CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n", - canceled, ldlm_ns_name(ns)); + CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n", + canceled, ldlm_ns_name(ns)); } int ldlm_replay_locks(struct obd_import *imp) { - struct ldlm_namespace *ns = imp->imp_obd->obd_namespace; - CFS_LIST_HEAD(list); - struct ldlm_lock *lock, *next; - int rc = 0; + struct ldlm_namespace *ns = imp->imp_obd->obd_namespace; + struct list_head list = LIST_HEAD_INIT(list); + struct ldlm_lock *lock, *next; + int rc = 0; - ENTRY; + ENTRY; - LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0); + LASSERT(atomic_read(&imp->imp_replay_inflight) == 0); - /* don't replay locks if import failed recovery */ - if (imp->imp_vbr_failed) - RETURN(0); + /* don't replay locks if import failed recovery */ + if (imp->imp_vbr_failed) + RETURN(0); - /* ensure this doesn't fall to 0 before all have been queued */ - cfs_atomic_inc(&imp->imp_replay_inflight); + /* ensure this doesn't fall to 0 before all have been queued */ + atomic_inc(&imp->imp_replay_inflight); - if (ldlm_cancel_unused_locks_before_replay) - ldlm_cancel_unused_locks_for_replay(ns); + if (ldlm_cancel_unused_locks_before_replay) + ldlm_cancel_unused_locks_for_replay(ns); - ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list); + ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list); - cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) { - cfs_list_del_init(&lock->l_pending_chain); - if (rc) { - LDLM_LOCK_RELEASE(lock); - continue; /* or try to do the rest? */ - } - rc = replay_one_lock(imp, lock); - LDLM_LOCK_RELEASE(lock); - } + list_for_each_entry_safe(lock, next, &list, l_pending_chain) { + list_del_init(&lock->l_pending_chain); + if (rc) { + LDLM_LOCK_RELEASE(lock); + continue; /* or try to do the rest? */ + } + rc = replay_one_lock(imp, lock); + LDLM_LOCK_RELEASE(lock); + } - cfs_atomic_dec(&imp->imp_replay_inflight); + atomic_dec(&imp->imp_replay_inflight); - RETURN(rc); + RETURN(rc); } -EXPORT_SYMBOL(ldlm_replay_locks);