X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Flov%2Flov_lock.c;h=547bdca89eb543103046706ccd7e990f4a7b4643;hb=db46ab8d95b1c6e040039c65acb0c30641421659;hp=14ecd6850290ccf59e608bf74176cf1437f096fb;hpb=fbf5870b9848929d352460f1f005b79c0b5ccc5a;p=fs%2Flustre-release.git diff --git a/lustre/lov/lov_lock.c b/lustre/lov/lov_lock.c index 14ecd68..547bdca 100644 --- a/lustre/lov/lov_lock.c +++ b/lustre/lov/lov_lock.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved. + * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -42,17 +42,64 @@ #include "lov_cl_internal.h" -/** \addtogroup lov lov @{ */ +/** \addtogroup lov + * @{ + */ static struct cl_lock_closure *lov_closure_get(const struct lu_env *env, struct cl_lock *parent); +static int lov_lock_unuse(const struct lu_env *env, + const struct cl_lock_slice *slice); /***************************************************************************** * * Lov lock operations. * */ +static struct lov_sublock_env *lov_sublock_env_get(const struct lu_env *env, + struct cl_lock *parent, + struct lov_lock_sub *lls) +{ + struct lov_sublock_env *subenv; + struct lov_io *lio = lov_env_io(env); + struct cl_io *io = lio->lis_cl.cis_io; + struct lov_io_sub *sub; + + subenv = &lov_env_session(env)->ls_subenv; + + /* + * FIXME: We tend to use the subio's env & io to call the sublock + * lock operations because osc lock sometimes stores some control + * variables in thread's IO infomation(Now only lockless information). + * However, if the lock's host(object) is different from the object + * for current IO, we have no way to get the subenv and subio because + * they are not initialized at all. As a temp fix, in this case, + * we still borrow the parent's env to call sublock operations. + */ + if (!io || !cl_object_same(io->ci_obj, parent->cll_descr.cld_obj)) { + subenv->lse_env = env; + subenv->lse_io = io; + subenv->lse_sub = NULL; + } else { + sub = lov_sub_get(env, lio, lls->sub_stripe); + if (!IS_ERR(sub)) { + subenv->lse_env = sub->sub_env; + subenv->lse_io = sub->sub_io; + subenv->lse_sub = sub; + } else { + subenv = (void*)sub; + } + } + return subenv; +} + +static void lov_sublock_env_put(struct lov_sublock_env *subenv) +{ + if (subenv && subenv->lse_sub) + lov_sub_put(subenv->lse_sub); +} + static void lov_sublock_adopt(const struct lu_env *env, struct lov_lock *lck, struct cl_lock *sublock, int idx, struct lov_lock_link *link) @@ -75,7 +122,7 @@ static void lov_sublock_adopt(const struct lu_env *env, struct lov_lock *lck, lck->lls_sub[idx].sub_lock = lsl; lck->lls_nr_filled++; LASSERT(lck->lls_nr_filled <= lck->lls_nr); - list_add_tail(&link->lll_list, &lsl->lss_parents); + cfs_list_add_tail(&link->lll_list, &lsl->lss_parents); link->lll_idx = idx; link->lll_super = lck; cl_lock_get(parent); @@ -100,17 +147,32 @@ static struct cl_lock *lov_sublock_alloc(const struct lu_env *env, LASSERT(idx < lck->lls_nr); ENTRY; - OBD_SLAB_ALLOC_PTR(link, lov_lock_link_kmem); + OBD_SLAB_ALLOC_PTR_GFP(link, lov_lock_link_kmem, __GFP_IO); if (link != NULL) { - struct lov_lock_sub *sub; + struct lov_sublock_env *subenv; + struct lov_lock_sub *lls; struct cl_lock_descr *descr; parent = lck->lls_cl.cls_lock; - sub = &lck->lls_sub[idx]; - descr = &sub->sub_descr; + lls = &lck->lls_sub[idx]; + descr = &lls->sub_got; + + subenv = lov_sublock_env_get(env, parent, lls); + if (!IS_ERR(subenv)) { + /* CAVEAT: Don't try to add a field in lov_lock_sub + * to remember the subio. This is because lock is able + * to be cached, but this is not true for IO. This + * further means a sublock might be referenced in + * different io context. -jay */ + + sublock = cl_lock_hold(subenv->lse_env, subenv->lse_io, + descr, "lov-parent", parent); + lov_sublock_env_put(subenv); + } else { + /* error occurs. */ + sublock = (void*)subenv; + } - /* XXX maybe sub-io? */ - sublock = cl_lock_hold(env, io, descr, "lov-parent", parent); if (!IS_ERR(sublock)) *out = link; else @@ -122,28 +184,64 @@ static struct cl_lock *lov_sublock_alloc(const struct lu_env *env, static void lov_sublock_unlock(const struct lu_env *env, struct lovsub_lock *lsl, - struct cl_lock_closure *closure) + struct cl_lock_closure *closure, + struct lov_sublock_env *subenv) { ENTRY; + lov_sublock_env_put(subenv); lsl->lss_active = NULL; cl_lock_disclosure(env, closure); EXIT; } -static int lov_sublock_lock(const struct lu_env *env, struct lovsub_lock *lsl, - struct cl_lock_closure *closure) +static int lov_sublock_lock(const struct lu_env *env, + struct lov_lock *lck, + struct lov_lock_sub *lls, + struct cl_lock_closure *closure, + struct lov_sublock_env **lsep) { - struct cl_lock *child; - int result; + struct lovsub_lock *sublock; + struct cl_lock *child; + int result = 0; + ENTRY; - LASSERT(list_empty(&closure->clc_list)); + LASSERT(cfs_list_empty(&closure->clc_list)); - ENTRY; - child = lsl->lss_cl.cls_lock; + sublock = lls->sub_lock; + child = sublock->lss_cl.cls_lock; result = cl_lock_closure_build(env, child, closure); if (result == 0) { + struct cl_lock *parent = closure->clc_origin; + LASSERT(cl_lock_is_mutexed(child)); - lsl->lss_active = closure->clc_origin; + sublock->lss_active = parent; + + if (unlikely((child->cll_state == CLS_FREEING) || + (child->cll_flags & CLF_CANCELLED))) { + struct lov_lock_link *link; + /* + * we could race with lock deletion which temporarily + * put the lock in freeing state, bug 19080. + */ + LASSERT(!(lls->sub_flags & LSF_HELD)); + + link = lov_lock_link_find(env, lck, sublock); + LASSERT(link != NULL); + lov_lock_unlink(env, link, sublock); + lov_sublock_unlock(env, sublock, closure, NULL); + lck->lls_cancel_race = 1; + result = CLO_REPEAT; + } else if (lsep) { + struct lov_sublock_env *subenv; + subenv = lov_sublock_env_get(env, parent, lls); + if (IS_ERR(subenv)) { + lov_sublock_unlock(env, sublock, + closure, NULL); + result = PTR_ERR(subenv); + } else { + *lsep = subenv; + } + } } RETURN(result); } @@ -169,12 +267,14 @@ static int lov_subresult(int result, int rc) int result_rank; int rc_rank; - LASSERT(result <= 0 || result == CLO_REPEAT || result == CLO_WAIT); - LASSERT(rc <= 0 || rc == CLO_REPEAT || rc == CLO_WAIT); - CLASSERT(CLO_WAIT < CLO_REPEAT); - ENTRY; + LASSERTF(result <= 0 || result == CLO_REPEAT || result == CLO_WAIT, + "result = %d", result); + LASSERTF(rc <= 0 || rc == CLO_REPEAT || rc == CLO_WAIT, + "rc = %d\n", rc); + CLASSERT(CLO_WAIT < CLO_REPEAT); + /* calculate ranks in the ordering above */ result_rank = result < 0 ? 1 + CLO_REPEAT : result; rc_rank = rc < 0 ? 1 + CLO_REPEAT : rc; @@ -197,10 +297,7 @@ static int lov_lock_sub_init(const struct lu_env *env, { int result = 0; int i; - int j; int nr; - int stripe; - int start_stripe; obd_off start; obd_off end; obd_off file_start; @@ -216,19 +313,17 @@ static int lov_lock_sub_init(const struct lu_env *env, file_start = cl_offset(lov2cl(loo), parent->cll_descr.cld_start); file_end = cl_offset(lov2cl(loo), parent->cll_descr.cld_end + 1) - 1; - start_stripe = lov_stripe_number(r0->lo_lsm, file_start); for (i = 0, nr = 0; i < r0->lo_nr; i++) { /* * XXX for wide striping smarter algorithm is desirable, * breaking out of the loop, early. */ - stripe = (start_stripe + i) % r0->lo_nr; - if (lov_stripe_intersects(r0->lo_lsm, stripe, + if (lov_stripe_intersects(loo->lo_lsm, i, file_start, file_end, &start, &end)) nr++; } LASSERT(nr > 0); - OBD_ALLOC(lck->lls_sub, nr * sizeof lck->lls_sub[0]); + OBD_ALLOC_LARGE(lck->lls_sub, nr * sizeof lck->lls_sub[0]); if (lck->lls_sub == NULL) RETURN(-ENOMEM); @@ -240,21 +335,23 @@ static int lov_lock_sub_init(const struct lu_env *env, * create sub-locks. At this moment, no other thread can access * top-lock. */ - for (j = 0, nr = 0; j < i; ++j) { - stripe = (start_stripe + j) % r0->lo_nr; - if (lov_stripe_intersects(r0->lo_lsm, stripe, + for (i = 0, nr = 0; i < r0->lo_nr; ++i) { + if (lov_stripe_intersects(loo->lo_lsm, i, file_start, file_end, &start, &end)) { struct cl_lock_descr *descr; descr = &lck->lls_sub[nr].sub_descr; LASSERT(descr->cld_obj == NULL); - descr->cld_obj = lovsub2cl(r0->lo_sub[stripe]); + descr->cld_obj = lovsub2cl(r0->lo_sub[i]); descr->cld_start = cl_index(descr->cld_obj, start); descr->cld_end = cl_index(descr->cld_obj, end); descr->cld_mode = parent->cll_descr.cld_mode; + descr->cld_gid = parent->cll_descr.cld_gid; + descr->cld_enq_flags = parent->cll_descr.cld_enq_flags; + /* XXX has no effect */ lck->lls_sub[nr].sub_got = *descr; - lck->lls_sub[nr].sub_stripe = stripe; + lck->lls_sub[nr].sub_stripe = i; nr++; } } @@ -273,6 +370,7 @@ static int lov_lock_sub_init(const struct lu_env *env, result = PTR_ERR(sublock); break; } + cl_lock_get_trust(sublock); cl_lock_mutex_get(env, sublock); cl_lock_mutex_get(env, parent); /* @@ -284,11 +382,13 @@ static int lov_lock_sub_init(const struct lu_env *env, lov_sublock_adopt(env, lck, sublock, i, link); cl_lock_mutex_put(env, parent); } else { + OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem); cl_lock_mutex_put(env, parent); cl_lock_unhold(env, sublock, "lov-parent", parent); } cl_lock_mutex_put(env, sublock); + cl_lock_put(env, sublock); } } /* @@ -308,7 +408,7 @@ static int lov_sublock_release(const struct lu_env *env, struct lov_lock *lck, ENTRY; if (lck->lls_sub[i].sub_flags & LSF_HELD) { - struct cl_lock *sublock; + struct cl_lock *sublock; int dying; LASSERT(lck->lls_sub[i].sub_lock != NULL); @@ -324,6 +424,7 @@ static int lov_sublock_release(const struct lu_env *env, struct lov_lock *lck, * while sub-lock is being paged out. */ dying = (sublock->cll_descr.cld_mode == CLM_PHANTOM || + sublock->cll_descr.cld_mode == CLM_GROUP || (sublock->cll_flags & (CLF_CANCELPEND|CLF_DOOMED))) && sublock->cll_holds == 1; if (dying) @@ -386,12 +487,29 @@ static void lov_lock_fini(const struct lu_env *env, * a reference on its parent. */ LASSERT(lck->lls_sub[i].sub_lock == NULL); - OBD_FREE(lck->lls_sub, lck->lls_nr * sizeof lck->lls_sub[0]); + OBD_FREE_LARGE(lck->lls_sub, + lck->lls_nr * sizeof lck->lls_sub[0]); } OBD_SLAB_FREE_PTR(lck, lov_lock_kmem); EXIT; } +static int lov_lock_enqueue_wait(const struct lu_env *env, + struct lov_lock *lck, + struct cl_lock *sublock) +{ + struct cl_lock *lock = lck->lls_cl.cls_lock; + int result; + ENTRY; + + LASSERT(cl_lock_is_mutexed(lock)); + + cl_lock_mutex_put(env, lock); + result = cl_lock_enqueue_wait(env, sublock, 0); + cl_lock_mutex_get(env, lock); + RETURN(result ?: CLO_REPEAT); +} + /** * Tries to advance a state machine of a given sub-lock toward enqueuing of * the top-lock. @@ -404,21 +522,24 @@ static int lov_lock_enqueue_one(const struct lu_env *env, struct lov_lock *lck, struct cl_io *io, __u32 enqflags, int last) { int result; - ENTRY; + /* first, try to enqueue a sub-lock ... */ result = cl_enqueue_try(env, sublock, io, enqflags); - if (sublock->cll_state == CLS_ENQUEUED) - /* if it is enqueued, try to `wait' on it---maybe it's already - * granted */ - result = cl_wait_try(env, sublock); + if ((sublock->cll_state == CLS_ENQUEUED) && !(enqflags & CEF_AGL)) { + /* if it is enqueued, try to `wait' on it---maybe it's already + * granted */ + result = cl_wait_try(env, sublock); + if (result == CLO_REENQUEUED) + result = CLO_WAIT; + } /* * If CEF_ASYNC flag is set, then all sub-locks can be enqueued in * parallel, otherwise---enqueue has to wait until sub-lock is granted * before proceeding to the next one. */ - if (result == CLO_WAIT && sublock->cll_state <= CLS_HELD && - enqflags & CEF_ASYNC && !last) + if ((result == CLO_WAIT) && (sublock->cll_state <= CLS_HELD) && + (enqflags & CEF_ASYNC) && (!last || (enqflags & CEF_AGL))) result = 0; RETURN(result); } @@ -441,10 +562,12 @@ static int lov_sublock_fill(const struct lu_env *env, struct cl_lock *parent, cl_lock_mutex_get(env, parent); if (!IS_ERR(sublock)) { + cl_lock_get_trust(sublock); if (parent->cll_state == CLS_QUEUING && - lck->lls_sub[idx].sub_lock == NULL) + lck->lls_sub[idx].sub_lock == NULL) { lov_sublock_adopt(env, lck, sublock, idx, link); - else { + } else { + OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem); /* other thread allocated sub-lock, or enqueue is no * longer going on */ cl_lock_mutex_put(env, parent); @@ -452,6 +575,7 @@ static int lov_sublock_fill(const struct lu_env *env, struct cl_lock *parent, cl_lock_mutex_get(env, parent); } cl_lock_mutex_put(env, sublock); + cl_lock_put(env, sublock); result = CLO_REPEAT; } else result = PTR_ERR(sublock); @@ -480,8 +604,10 @@ static int lov_lock_enqueue(const struct lu_env *env, for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) { int rc; - struct lovsub_lock *sub; - struct cl_lock *sublock; + struct lovsub_lock *sub; + struct lov_lock_sub *lls; + struct cl_lock *sublock; + struct lov_sublock_env *subenv; if (lock->cll_state != CLS_QUEUING) { /* @@ -493,7 +619,8 @@ static int lov_lock_enqueue(const struct lu_env *env, break; } - sub = lck->lls_sub[i].sub_lock; + lls = &lck->lls_sub[i]; + sub = lls->sub_lock; /* * Sub-lock might have been canceled, while top-lock was * cached. @@ -505,23 +632,51 @@ static int lov_lock_enqueue(const struct lu_env *env, break; } sublock = sub->lss_cl.cls_lock; - rc = lov_sublock_lock(env, sub, closure); + rc = lov_sublock_lock(env, lck, lls, closure, &subenv); if (rc == 0) { lov_sublock_hold(env, lck, i); - rc = lov_lock_enqueue_one(env, lck, sublock, io, - enqflags, + rc = lov_lock_enqueue_one(subenv->lse_env, lck, sublock, + subenv->lse_io, enqflags, i == lck->lls_nr - 1); minstate = min(minstate, sublock->cll_state); - /* - * Don't hold a sub-lock in CLS_CACHED state, see - * description for lov_lock::lls_sub. - */ - if (sublock->cll_state > CLS_HELD) - rc = lov_sublock_release(env, lck, i, 1, rc); - lov_sublock_unlock(env, sub, closure); + if (rc == CLO_WAIT) { + switch (sublock->cll_state) { + case CLS_QUEUING: + /* take recursive mutex, the lock is + * released in lov_lock_enqueue_wait. + */ + cl_lock_mutex_get(env, sublock); + lov_sublock_unlock(env, sub, closure, + subenv); + rc = lov_lock_enqueue_wait(env, lck, + sublock); + break; + case CLS_CACHED: + cl_lock_get(sublock); + /* take recursive mutex of sublock */ + cl_lock_mutex_get(env, sublock); + /* need to release all locks in closure + * otherwise it may deadlock. LU-2683.*/ + lov_sublock_unlock(env, sub, closure, + subenv); + /* sublock and parent are held. */ + rc = lov_sublock_release(env, lck, i, + 1, rc); + cl_lock_mutex_put(env, sublock); + cl_lock_put(env, sublock); + break; + default: + lov_sublock_unlock(env, sub, closure, + subenv); + break; + } + } else { + LASSERT(sublock->cll_conflict == NULL); + lov_sublock_unlock(env, sub, closure, subenv); + } } result = lov_subresult(result, rc); - if (result < 0) + if (result != 0) break; } cl_lock_closure_fini(closure); @@ -540,72 +695,149 @@ static int lov_lock_unuse(const struct lu_env *env, for (result = 0, i = 0; i < lck->lls_nr; ++i) { int rc; - struct lovsub_lock *sub; - struct cl_lock *sublock; + struct lovsub_lock *sub; + struct cl_lock *sublock; + struct lov_lock_sub *lls; + struct lov_sublock_env *subenv; /* top-lock state cannot change concurrently, because single * thread (one that released the last hold) carries unlocking * to the completion. */ - LASSERT(slice->cls_lock->cll_state == CLS_UNLOCKING); - sub = lck->lls_sub[i].sub_lock; + LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT); + lls = &lck->lls_sub[i]; + sub = lls->sub_lock; if (sub == NULL) continue; sublock = sub->lss_cl.cls_lock; - rc = lov_sublock_lock(env, sub, closure); + rc = lov_sublock_lock(env, lck, lls, closure, &subenv); if (rc == 0) { - if (lck->lls_sub[i].sub_flags & LSF_HELD) { - LASSERT(sublock->cll_state == CLS_HELD); - rc = cl_unuse_try(env, sublock); - if (rc != CLO_WAIT) - rc = lov_sublock_release(env, lck, - i, 0, rc); + if (lls->sub_flags & LSF_HELD) { + LASSERT(sublock->cll_state == CLS_HELD || + sublock->cll_state == CLS_ENQUEUED); + rc = cl_unuse_try(subenv->lse_env, sublock); + rc = lov_sublock_release(env, lck, i, 0, rc); } - lov_sublock_unlock(env, sub, closure); + lov_sublock_unlock(env, sub, closure, subenv); } result = lov_subresult(result, rc); - if (result < 0) - break; } - if (result == 0 && lck->lls_unuse_race) { - lck->lls_unuse_race = 0; + + if (result == 0 && lck->lls_cancel_race) { + lck->lls_cancel_race = 0; result = -ESTALE; } cl_lock_closure_fini(closure); RETURN(result); } + +static void lov_lock_cancel(const struct lu_env *env, + const struct cl_lock_slice *slice) +{ + struct lov_lock *lck = cl2lov_lock(slice); + struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock); + int i; + int result; + + ENTRY; + + for (result = 0, i = 0; i < lck->lls_nr; ++i) { + int rc; + struct lovsub_lock *sub; + struct cl_lock *sublock; + struct lov_lock_sub *lls; + struct lov_sublock_env *subenv; + + /* top-lock state cannot change concurrently, because single + * thread (one that released the last hold) carries unlocking + * to the completion. */ + lls = &lck->lls_sub[i]; + sub = lls->sub_lock; + if (sub == NULL) + continue; + + sublock = sub->lss_cl.cls_lock; + rc = lov_sublock_lock(env, lck, lls, closure, &subenv); + if (rc == 0) { + if (!(lls->sub_flags & LSF_HELD)) { + lov_sublock_unlock(env, sub, closure, subenv); + continue; + } + + switch(sublock->cll_state) { + case CLS_HELD: + rc = cl_unuse_try(subenv->lse_env, sublock); + lov_sublock_release(env, lck, i, 0, 0); + break; + default: + lov_sublock_release(env, lck, i, 1, 0); + break; + } + lov_sublock_unlock(env, sub, closure, subenv); + } + + if (rc == CLO_REPEAT) { + --i; + continue; + } + + result = lov_subresult(result, rc); + } + + if (result) + CL_LOCK_DEBUG(D_ERROR, env, slice->cls_lock, + "lov_lock_cancel fails with %d.\n", result); + + cl_lock_closure_fini(closure); +} + static int lov_lock_wait(const struct lu_env *env, const struct cl_lock_slice *slice) { struct lov_lock *lck = cl2lov_lock(slice); struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock); enum cl_lock_state minstate; + int reenqueued; int result; int i; ENTRY; - for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) { +again: + for (result = 0, minstate = CLS_FREEING, i = 0, reenqueued = 0; + i < lck->lls_nr; ++i) { int rc; - struct lovsub_lock *sub; - struct cl_lock *sublock; + struct lovsub_lock *sub; + struct cl_lock *sublock; + struct lov_lock_sub *lls; + struct lov_sublock_env *subenv; - sub = lck->lls_sub[i].sub_lock; + lls = &lck->lls_sub[i]; + sub = lls->sub_lock; LASSERT(sub != NULL); sublock = sub->lss_cl.cls_lock; - rc = lov_sublock_lock(env, sub, closure); + rc = lov_sublock_lock(env, lck, lls, closure, &subenv); if (rc == 0) { LASSERT(sublock->cll_state >= CLS_ENQUEUED); if (sublock->cll_state < CLS_HELD) rc = cl_wait_try(env, sublock); + minstate = min(minstate, sublock->cll_state); - lov_sublock_unlock(env, sub, closure); + lov_sublock_unlock(env, sub, closure, subenv); + } + if (rc == CLO_REENQUEUED) { + reenqueued++; + rc = 0; } result = lov_subresult(result, rc); - if (result < 0) + if (result != 0) break; } + /* Each sublock only can be reenqueued once, so will not loop for + * ever. */ + if (result == 0 && reenqueued != 0) + goto again; cl_lock_closure_fini(closure); RETURN(result ?: minstate >= CLS_HELD ? 0 : CLO_WAIT); } @@ -618,44 +850,63 @@ static int lov_lock_use(const struct lu_env *env, int result; int i; - LASSERT(slice->cls_lock->cll_state == CLS_CACHED); + LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT); ENTRY; for (result = 0, i = 0; i < lck->lls_nr; ++i) { int rc; - struct lovsub_lock *sub; - struct cl_lock *sublock; + struct lovsub_lock *sub; + struct cl_lock *sublock; + struct lov_lock_sub *lls; + struct lov_sublock_env *subenv; - if (slice->cls_lock->cll_state != CLS_CACHED) { - /* see comment in lov_lock_enqueue(). */ - LASSERT(i > 0 && result != 0); + LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT); + + lls = &lck->lls_sub[i]; + sub = lls->sub_lock; + if (sub == NULL) { + /* + * Sub-lock might have been canceled, while top-lock was + * cached. + */ + result = -ESTALE; break; } - /* - * if a sub-lock was destroyed while top-lock was in - * CLS_CACHED state, top-lock would have been moved into - * CLS_NEW state, so all sub-locks have to be in place. - */ - sub = lck->lls_sub[i].sub_lock; - LASSERT(sub != NULL); + sublock = sub->lss_cl.cls_lock; - rc = lov_sublock_lock(env, sub, closure); + rc = lov_sublock_lock(env, lck, lls, closure, &subenv); if (rc == 0) { LASSERT(sublock->cll_state != CLS_FREEING); lov_sublock_hold(env, lck, i); if (sublock->cll_state == CLS_CACHED) { - rc = cl_use_try(env, sublock); + rc = cl_use_try(subenv->lse_env, sublock, 0); if (rc != 0) rc = lov_sublock_release(env, lck, i, 1, rc); - } else - rc = 0; - lov_sublock_unlock(env, sub, closure); + } else if (sublock->cll_state == CLS_NEW) { + /* Sub-lock might have been canceled, while + * top-lock was cached. */ + result = -ESTALE; + lov_sublock_release(env, lck, i, 1, result); + } + lov_sublock_unlock(env, sub, closure, subenv); } result = lov_subresult(result, rc); - if (result < 0) + if (result != 0) break; } + + if (lck->lls_cancel_race) { + /* + * If there is unlocking happened at the same time, then + * sublock_lock state should be FREEING, and lov_sublock_lock + * should return CLO_REPEAT. In this case, it should return + * ESTALE, and up layer should reset the lock state to be NEW. + */ + lck->lls_cancel_race = 0; + LASSERT(result != 0); + result = -ESTALE; + } cl_lock_closure_fini(closure); RETURN(result); } @@ -684,7 +935,7 @@ static int lock_lock_multi_match() if (sub->sub_lock == NULL) continue; subobj = sub->sub_descr.cld_obj; - if (!lov_stripe_intersects(r0->lo_lsm, sub->sub_stripe, + if (!lov_stripe_intersects(loo->lo_lsm, sub->sub_stripe, fstart, fend, &start, &end)) continue; subneed->cld_start = cl_index(subobj, start); @@ -699,19 +950,49 @@ static int lock_lock_multi_match() } #endif -static int lov_is_same_stripe(struct lov_object *lov, int stripe, - const struct cl_lock_descr *descr) +/** + * Check if the extent region \a descr is covered by \a child against the + * specific \a stripe. + */ +static int lov_lock_stripe_is_matching(const struct lu_env *env, + struct lov_object *lov, int stripe, + const struct cl_lock_descr *child, + const struct cl_lock_descr *descr) { - struct lov_stripe_md *lsm = lov_r0(lov)->lo_lsm; + struct lov_stripe_md *lsm = lov->lo_lsm; obd_off start; obd_off end; + int result; + if (lov_r0(lov)->lo_nr == 1) + return cl_lock_ext_match(child, descr); + + /* + * For a multi-stripes object: + * - make sure the descr only covers child's stripe, and + * - check if extent is matching. + */ start = cl_offset(&lov->lo_cl, descr->cld_start); end = cl_offset(&lov->lo_cl, descr->cld_end + 1) - 1; - return - end - start <= lsm->lsm_stripe_size && - stripe == lov_stripe_number(lsm, start) && - stripe == lov_stripe_number(lsm, end); + result = end - start <= lsm->lsm_stripe_size && + stripe == lov_stripe_number(lsm, start) && + stripe == lov_stripe_number(lsm, end); + if (result) { + struct cl_lock_descr *subd = &lov_env_info(env)->lti_ldescr; + obd_off sub_start; + obd_off sub_end; + + subd->cld_obj = NULL; /* don't need sub object at all */ + subd->cld_mode = descr->cld_mode; + subd->cld_gid = descr->cld_gid; + result = lov_stripe_intersects(lsm, stripe, start, end, + &sub_start, &sub_end); + LASSERT(result); + subd->cld_start = cl_index(child->cld_obj, sub_start); + subd->cld_end = cl_index(child->cld_obj, sub_end); + result = cl_lock_ext_match(child, subd); + } + return result; } /** @@ -737,21 +1018,26 @@ static int lov_lock_fits_into(const struct lu_env *env, ENTRY; - if (lov->lls_nr == 1) { + /* for top lock, it's necessary to match enq flags otherwise it will + * run into problem if a sublock is missing and reenqueue. */ + if (need->cld_enq_flags != lov->lls_orig.cld_enq_flags) + return 0; + + if (lov->lls_ever_canceled) + return 0; + + if (need->cld_mode == CLM_GROUP) /* - * If a lock is on a single stripe, it's enough to check that - * @need lock matches actually granted stripe lock, and... + * always allow to match group lock. */ - result = cl_lock_ext_match(&lov->lls_sub[0].sub_got, need); - if (result && lov_r0(obj)->lo_nr > 1) - /* - * ... @need is on the same stripe, if multiple - * stripes are possible at all for this object. - */ - result = lov_is_same_stripe(cl2lov(slice->cls_obj), - lov->lls_sub[0].sub_stripe, - need); - } else if (io->ci_type != CIT_TRUNC && io->ci_type != CIT_MISC && + result = cl_lock_ext_match(&lov->lls_orig, need); + else if (lov->lls_nr == 1) { + struct cl_lock_descr *got = &lov->lls_sub[0].sub_got; + result = lov_lock_stripe_is_matching(env, + cl2lov(slice->cls_obj), + lov->lls_sub[0].sub_stripe, + got, need); + } else if (io->ci_type != CIT_SETATTR && io->ci_type != CIT_MISC && !cl_io_is_append(io) && need->cld_mode != CLM_PHANTOM) /* * Multi-stripe locks are only suitable for `quick' IO and for @@ -768,7 +1054,7 @@ static int lov_lock_fits_into(const struct lu_env *env, * match against original lock extent. */ result = cl_lock_ext_match(&lov->lls_orig, need); - CDEBUG(D_DLMTRACE, DDESCR"/"DDESCR" %i %i/%i: %i\n", + CDEBUG(D_DLMTRACE, DDESCR"/"DDESCR" %d %d/%d: %d\n", PDESCR(&lov->lls_orig), PDESCR(&lov->lls_sub[0].sub_got), lov->lls_sub[0].sub_stripe, lov->lls_nr, lov_r0(obj)->lo_nr, result); @@ -785,7 +1071,7 @@ void lov_lock_unlink(const struct lu_env *env, LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock)); ENTRY; - list_del_init(&link->lll_list); + cfs_list_del_init(&link->lll_list); LASSERT(lck->lls_sub[link->lll_idx].sub_lock == sub); /* yank this sub-lock from parent's array */ lck->lls_sub[link->lll_idx].sub_lock = NULL; @@ -806,7 +1092,7 @@ struct lov_lock_link *lov_lock_link_find(const struct lu_env *env, LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock)); ENTRY; - list_for_each_entry(scan, &sub->lss_parents, lll_list) { + cfs_list_for_each_entry(scan, &sub->lss_parents, lll_list) { if (scan->lll_super == lck) RETURN(scan); } @@ -832,42 +1118,40 @@ static void lov_lock_delete(const struct lu_env *env, { struct lov_lock *lck = cl2lov_lock(slice); struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock); - int i; + struct lov_lock_link *link; + int rc; + int i; LASSERT(slice->cls_lock->cll_state == CLS_FREEING); ENTRY; for (i = 0; i < lck->lls_nr; ++i) { - struct lovsub_lock *lsl; - struct cl_lock *sublock; - int rc; + struct lov_lock_sub *lls = &lck->lls_sub[i]; + struct lovsub_lock *lsl = lls->sub_lock; - lsl = lck->lls_sub[i].sub_lock; - if (lsl == NULL) + if (lsl == NULL) /* already removed */ continue; - sublock = lsl->lss_cl.cls_lock; - rc = lov_sublock_lock(env, lsl, closure); - if (rc == 0) { - if (lck->lls_sub[i].sub_flags & LSF_HELD) - lov_sublock_release(env, lck, i, 1, 0); - if (sublock->cll_state < CLS_FREEING) { - struct lov_lock_link *link; - - link = lov_lock_link_find(env, lck, lsl); - LASSERT(link != NULL); - lov_lock_unlink(env, link, lsl); - LASSERT(lck->lls_sub[i].sub_lock == NULL); - } - lov_sublock_unlock(env, lsl, closure); - } else if (rc == CLO_REPEAT) { - --i; /* repeat with this lock */ - } else { - CL_LOCK_DEBUG(D_ERROR, env, sublock, - "Cannot get sub-lock for delete: %i\n", - rc); + rc = lov_sublock_lock(env, lck, lls, closure, NULL); + if (rc == CLO_REPEAT) { + --i; + continue; } + + LASSERT(rc == 0); + LASSERT(lsl->lss_cl.cls_lock->cll_state < CLS_FREEING); + + if (lls->sub_flags & LSF_HELD) + lov_sublock_release(env, lck, i, 1, 0); + + link = lov_lock_link_find(env, lck, lsl); + LASSERT(link != NULL); + lov_lock_unlink(env, link, lsl); + LASSERT(lck->lls_sub[i].sub_lock == NULL); + + lov_sublock_unlock(env, lsl, closure, NULL); } + cl_lock_closure_fini(closure); EXIT; } @@ -899,6 +1183,7 @@ static const struct cl_lock_operations lov_lock_ops = { .clo_wait = lov_lock_wait, .clo_use = lov_lock_use, .clo_unuse = lov_lock_unuse, + .clo_cancel = lov_lock_cancel, .clo_fits_into = lov_lock_fits_into, .clo_delete = lov_lock_delete, .clo_print = lov_lock_print @@ -911,7 +1196,7 @@ int lov_lock_init_raid0(const struct lu_env *env, struct cl_object *obj, int result; ENTRY; - OBD_SLAB_ALLOC_PTR(lck, lov_lock_kmem); + OBD_SLAB_ALLOC_PTR_GFP(lck, lov_lock_kmem, __GFP_IO); if (lck != NULL) { cl_lock_slice_add(lock, &lck->lls_cl, obj, &lov_lock_ops); result = lov_lock_sub_init(env, lck, io); @@ -920,13 +1205,49 @@ int lov_lock_init_raid0(const struct lu_env *env, struct cl_object *obj, RETURN(result); } +static void lov_empty_lock_fini(const struct lu_env *env, + struct cl_lock_slice *slice) +{ + struct lov_lock *lck = cl2lov_lock(slice); + OBD_SLAB_FREE_PTR(lck, lov_lock_kmem); +} + +static int lov_empty_lock_print(const struct lu_env *env, void *cookie, + lu_printer_t p, const struct cl_lock_slice *slice) +{ + (*p)(env, cookie, "empty\n"); + return 0; +} + +/* XXX: more methods will be added later. */ +static const struct cl_lock_operations lov_empty_lock_ops = { + .clo_fini = lov_empty_lock_fini, + .clo_print = lov_empty_lock_print +}; + +int lov_lock_init_empty(const struct lu_env *env, struct cl_object *obj, + struct cl_lock *lock, const struct cl_io *io) +{ + struct lov_lock *lck; + int result = -ENOMEM; + + ENTRY; + OBD_SLAB_ALLOC_PTR_GFP(lck, lov_lock_kmem, __GFP_IO); + if (lck != NULL) { + cl_lock_slice_add(lock, &lck->lls_cl, obj, &lov_empty_lock_ops); + lck->lls_orig = lock->cll_descr; + result = 0; + } + RETURN(result); +} + static struct cl_lock_closure *lov_closure_get(const struct lu_env *env, struct cl_lock *parent) { struct cl_lock_closure *closure; closure = &lov_env_info(env)->lti_closure; - LINVRNT(list_empty(&closure->clc_list)); + LASSERT(cfs_list_empty(&closure->clc_list)); cl_lock_closure_init(env, closure, parent, 1); return closure; }