X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Flov%2Flov_lock.c;h=40192cecadaa4381abd1d7956da518c8fef7ba3d;hp=157ee7d69161e244363ee8fad62419d167d329c6;hb=1364ab2c166d69bc857a729f3eff2c965db847c9;hpb=b15f3875f46eec3c5186fe6b84cf2cda7ceb8518 diff --git a/lustre/lov/lov_lock.c b/lustre/lov/lov_lock.c index 157ee7d..40192ce 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, 2012, Whamcloud, Inc. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -49,6 +49,8 @@ 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. @@ -75,12 +77,11 @@ static struct lov_sublock_env *lov_sublock_env_get(const struct lu_env *env, * 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 (!cl_object_same(io->ci_obj, parent->cll_descr.cld_obj)) { + 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 { - LASSERT(io != NULL); sub = lov_sub_get(env, lio, lls->sub_stripe); if (!IS_ERR(sub)) { subenv->lse_env = sub->sub_env; @@ -121,7 +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); @@ -204,7 +205,7 @@ static int lov_sublock_lock(const struct lu_env *env, int result = 0; ENTRY; - LASSERT(list_empty(&closure->clc_list)); + LASSERT(cfs_list_empty(&closure->clc_list)); sublock = lls->sub_lock; child = sublock->lss_cl.cls_lock; @@ -215,7 +216,8 @@ static int lov_sublock_lock(const struct lu_env *env, LASSERT(cl_lock_is_mutexed(child)); sublock->lss_active = parent; - if (unlikely(child->cll_state == CLS_FREEING)) { + 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 @@ -227,6 +229,7 @@ static int lov_sublock_lock(const struct lu_env *env, 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; @@ -264,12 +267,12 @@ static int lov_subresult(int result, int rc) int result_rank; int rc_rank; + ENTRY; + LASSERT(result <= 0 || result == CLO_REPEAT || result == CLO_WAIT); LASSERT(rc <= 0 || rc == CLO_REPEAT || rc == CLO_WAIT); CLASSERT(CLO_WAIT < CLO_REPEAT); - ENTRY; - /* calculate ranks in the ordering above */ result_rank = result < 0 ? 1 + CLO_REPEAT : result; rc_rank = rc < 0 ? 1 + CLO_REPEAT : rc; @@ -318,7 +321,7 @@ static int lov_lock_sub_init(const struct lu_env *env, 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); @@ -343,6 +346,7 @@ static int lov_lock_sub_init(const struct lu_env *env, 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 = i; @@ -364,6 +368,7 @@ 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); /* @@ -381,6 +386,7 @@ static int lov_lock_sub_init(const struct lu_env *env, "lov-parent", parent); } cl_lock_mutex_put(env, sublock); + cl_lock_put(env, sublock); } } /* @@ -479,12 +485,29 @@ 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. @@ -501,7 +524,7 @@ static int lov_lock_enqueue_one(const struct lu_env *env, struct lov_lock *lck, /* first, try to enqueue a sub-lock ... */ result = cl_enqueue_try(env, sublock, io, enqflags); - if (sublock->cll_state == CLS_ENQUEUED) + 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); @@ -510,8 +533,8 @@ static int lov_lock_enqueue_one(const struct lu_env *env, struct lov_lock *lck, * 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); } @@ -534,10 +557,11 @@ 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 */ @@ -546,6 +570,7 @@ 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); @@ -609,13 +634,30 @@ static int lov_lock_enqueue(const struct lu_env *env, 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, subenv); + 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: + rc = lov_sublock_release(env, lck, i, + 1, rc); + 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) @@ -645,7 +687,7 @@ static int lov_lock_unuse(const struct lu_env *env, /* 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); + LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT); lls = &lck->lls_sub[i]; sub = lls->sub_lock; if (sub == NULL) @@ -654,39 +696,101 @@ static int lov_lock_unuse(const struct lu_env *env, sublock = sub->lss_cl.cls_lock; 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); + 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); - if (rc != CLO_WAIT) - rc = lov_sublock_release(env, lck, - i, 0, rc); + rc = lov_sublock_release(env, lck, i, 0, rc); } 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; @@ -706,10 +810,18 @@ static int lov_lock_wait(const struct lu_env *env, minstate = min(minstate, sublock->cll_state); 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); } @@ -722,7 +834,7 @@ 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) { @@ -732,37 +844,53 @@ static int lov_lock_use(const struct lu_env *env, 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); - 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. - */ + LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT); + lls = &lck->lls_sub[i]; sub = lls->sub_lock; - LASSERT(sub != NULL); + if (sub == NULL) { + /* + * Sub-lock might have been canceled, while top-lock was + * cached. + */ + result = -ESTALE; + break; + } + sublock = sub->lss_cl.cls_lock; 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(subenv->lse_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; + } 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); } @@ -885,7 +1013,7 @@ static int lov_lock_fits_into(const struct lu_env *env, cl2lov(slice->cls_obj), lov->lls_sub[0].sub_stripe, got, need); - } else if (io->ci_type != CIT_TRUNC && io->ci_type != CIT_MISC && + } 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 @@ -902,7 +1030,7 @@ 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); @@ -919,7 +1047,7 @@ 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; @@ -940,7 +1068,7 @@ 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); } @@ -966,44 +1094,40 @@ 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 lov_lock_sub *lls; - 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; - lls = &lck->lls_sub[i]; - lsl = lls->sub_lock; - if (lsl == NULL) + if (lsl == NULL) /* already removed */ continue; - sublock = lsl->lss_cl.cls_lock; rc = lov_sublock_lock(env, lck, lls, closure, NULL); - 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, NULL); - } 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); + 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; } @@ -1035,6 +1159,7 @@ 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 @@ -1062,7 +1187,7 @@ static struct cl_lock_closure *lov_closure_get(const struct lu_env *env, struct cl_lock_closure *closure; closure = &lov_env_info(env)->lti_closure; - LASSERT(list_empty(&closure->clc_list)); + LASSERT(cfs_list_empty(&closure->clc_list)); cl_lock_closure_init(env, closure, parent, 1); return closure; }