X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fobdclass%2Fcl_page.c;h=a9d81467552357659114e5dc5baad51b35c6f159;hb=59c3911deb8357c4b6aad68d1b1fc81a8ad61b3f;hp=ea160ebe3de6d1f2715f9e8464ac356120d8299e;hpb=60e052d06f7f82e13fc551b96d1ca7585ac505df;p=fs%2Flustre-release.git diff --git a/lustre/obdclass/cl_page.c b/lustre/obdclass/cl_page.c index ea160eb..a9d8146 100644 --- a/lustre/obdclass/cl_page.c +++ b/lustre/obdclass/cl_page.c @@ -27,7 +27,7 @@ * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2012, Whamcloud, Inc. + * Copyright (c) 2011, 2012, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -51,19 +51,6 @@ static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg, int radix); -static cfs_mem_cache_t *cl_page_kmem = NULL; - -static struct lu_kmem_descr cl_page_caches[] = { - { - .ckd_cache = &cl_page_kmem, - .ckd_name = "cl_page_kmem", - .ckd_size = sizeof (struct cl_page) - }, - { - .ckd_cache = NULL - } -}; - #ifdef LIBCFS_DEBUG # define PASSERT(env, page, expr) \ do { \ @@ -90,9 +77,27 @@ static struct lu_kmem_descr cl_page_caches[] = { ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp)) #endif /* !INVARIANT_CHECK */ +/* Disable page statistic by default due to huge performance penalty. */ +#ifdef CONFIG_DEBUG_PAGESTATE_TRACKING +#define CS_PAGE_INC(o, item) \ + cfs_atomic_inc(&cl_object_site(o)->cs_pages.cs_stats[CS_##item]) +#define CS_PAGE_DEC(o, item) \ + cfs_atomic_dec(&cl_object_site(o)->cs_pages.cs_stats[CS_##item]) +#define CS_PAGESTATE_INC(o, state) \ + cfs_atomic_inc(&cl_object_site(o)->cs_pages_state[state]) +#define CS_PAGESTATE_DEC(o, state) \ + cfs_atomic_dec(&cl_object_site(o)->cs_pages_state[state]) +#else +#define CS_PAGE_INC(o, item) +#define CS_PAGE_DEC(o, item) +#define CS_PAGESTATE_INC(o, state) +#define CS_PAGESTATE_DEC(o, state) +#endif + /** - * Internal version of cl_page_top, it should be called with page referenced, - * or cp_lock held. + * Internal version of cl_page_top, it should be called if the page is + * known to be not freed, says with page referenced, or radix tree lock held, + * or page owned. */ static struct cl_page *cl_page_top_trusted(struct cl_page *page) { @@ -114,11 +119,8 @@ static struct cl_page *cl_page_top_trusted(struct cl_page *page) */ static void cl_page_get_trust(struct cl_page *page) { - /* - * Checkless version for trusted users. - */ - if (cfs_atomic_inc_return(&page->cp_ref) == 1) - cfs_atomic_inc(&cl_object_site(page->cp_obj)->cs_pages.cs_busy); + LASSERT(cfs_atomic_read(&page->cp_ref) > 0); + cfs_atomic_inc(&page->cp_ref); } /** @@ -132,11 +134,6 @@ cl_page_at_trusted(const struct cl_page *page, const struct lu_device_type *dtype) { const struct cl_page_slice *slice; - -#ifdef INVARIANT_CHECK - if (!cfs_atomic_read(&page->cp_ref)) - LASSERT_SPIN_LOCKED(&page->cp_lock); -#endif ENTRY; page = cl_page_top_trusted((struct cl_page *)page); @@ -163,9 +160,8 @@ struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index) LASSERT_SPIN_LOCKED(&hdr->coh_page_guard); page = radix_tree_lookup(&hdr->coh_tree, index); - if (page != NULL) { + if (page != NULL) cl_page_get_trust(page); - } return page; } EXPORT_SYMBOL(cl_page_lookup); @@ -203,7 +199,7 @@ int cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj, hdr = cl_object_header(obj); pvec = cl_env_info(env)->clt_pvec; dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type; - cfs_spin_lock(&hdr->coh_page_guard); + spin_lock(&hdr->coh_page_guard); while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec, idx, CLT_PVEC_SIZE)) > 0) { int end_of_region = 0; @@ -249,7 +245,7 @@ int cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj, * check that pages weren't truncated (cl_page_own() returns * error in the latter case). */ - cfs_spin_unlock(&hdr->coh_page_guard); + spin_unlock(&hdr->coh_page_guard); tree_lock = 0; for (i = 0; i < j; ++i) { @@ -268,19 +264,19 @@ int cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj, if (res != CLP_GANG_OKAY) break; - cfs_spin_lock(&hdr->coh_page_guard); - tree_lock = 1; - } - if (tree_lock) - cfs_spin_unlock(&hdr->coh_page_guard); - RETURN(res); + spin_lock(&hdr->coh_page_guard); + tree_lock = 1; + } + if (tree_lock) + spin_unlock(&hdr->coh_page_guard); + RETURN(res); } EXPORT_SYMBOL(cl_page_gang_lookup); static void cl_page_free(const struct lu_env *env, struct cl_page *page) { struct cl_object *obj = page->cp_obj; - struct cl_site *site = cl_object_site(obj); + int pagesize = cl_object_header(obj)->coh_page_bufsize; PASSERT(env, page, cfs_list_empty(&page->cp_batch)); PASSERT(env, page, page->cp_owner == NULL); @@ -298,15 +294,12 @@ static void cl_page_free(const struct lu_env *env, struct cl_page *page) cfs_list_del_init(page->cp_layers.next); slice->cpl_ops->cpo_fini(env, slice); } - cfs_atomic_dec(&site->cs_pages.cs_total); - -#ifdef LUSTRE_PAGESTATE_TRACKING - cfs_atomic_dec(&site->cs_pages_state[page->cp_state]); -#endif + CS_PAGE_DEC(obj, total); + CS_PAGESTATE_DEC(obj, page->cp_state); lu_object_ref_del_at(&obj->co_lu, page->cp_obj_ref, "cl_page", page); cl_object_put(env, obj); lu_ref_fini(&page->cp_reference); - OBD_SLAB_FREE_PTR(page, cl_page_kmem); + OBD_FREE(page, pagesize); EXIT; } @@ -321,62 +314,55 @@ static inline void cl_page_state_set_trust(struct cl_page *page, *(enum cl_page_state *)&page->cp_state = state; } -static int cl_page_alloc(const struct lu_env *env, struct cl_object *o, - pgoff_t ind, struct page *vmpage, - enum cl_page_type type, struct cl_page **out) +static struct cl_page *cl_page_alloc(const struct lu_env *env, + struct cl_object *o, pgoff_t ind, struct page *vmpage, + enum cl_page_type type) { - struct cl_page *page; - struct cl_page *err = NULL; - struct lu_object_header *head; - struct cl_site *site = cl_object_site(o); - int result; - - ENTRY; - result = +1; - OBD_SLAB_ALLOC_PTR_GFP(page, cl_page_kmem, CFS_ALLOC_IO); - if (page != NULL) { - cfs_atomic_set(&page->cp_ref, 1); - page->cp_obj = o; - cl_object_get(o); - page->cp_obj_ref = lu_object_ref_add(&o->co_lu, - "cl_page", page); - page->cp_index = ind; - cl_page_state_set_trust(page, CPS_CACHED); - cfs_spin_lock_init(&page->cp_lock); - page->cp_type = type; - CFS_INIT_LIST_HEAD(&page->cp_layers); - CFS_INIT_LIST_HEAD(&page->cp_batch); - CFS_INIT_LIST_HEAD(&page->cp_flight); - cfs_mutex_init(&page->cp_mutex); - lu_ref_init(&page->cp_reference); - head = o->co_lu.lo_header; - cfs_list_for_each_entry(o, &head->loh_layers, - co_lu.lo_linkage) { - if (o->co_ops->coo_page_init != NULL) { - err = o->co_ops->coo_page_init(env, o, - page, vmpage); - if (err != NULL) { - cl_page_delete0(env, page, 0); - cl_page_free(env, page); - page = err; - break; - } - } - } - if (err == NULL) { - cfs_atomic_inc(&site->cs_pages.cs_busy); - cfs_atomic_inc(&site->cs_pages.cs_total); + struct cl_page *page; + struct lu_object_header *head; -#ifdef LUSTRE_PAGESTATE_TRACKING - cfs_atomic_inc(&site->cs_pages_state[CPS_CACHED]); -#endif - cfs_atomic_inc(&site->cs_pages.cs_created); - result = 0; - } - } else - page = ERR_PTR(-ENOMEM); - *out = page; - RETURN(result); + ENTRY; + OBD_ALLOC_GFP(page, cl_object_header(o)->coh_page_bufsize, + CFS_ALLOC_IO); + if (page != NULL) { + int result; + cfs_atomic_set(&page->cp_ref, 1); + if (type == CPT_CACHEABLE) /* for radix tree */ + cfs_atomic_inc(&page->cp_ref); + page->cp_obj = o; + cl_object_get(o); + page->cp_obj_ref = lu_object_ref_add(&o->co_lu, "cl_page",page); + page->cp_index = ind; + cl_page_state_set_trust(page, CPS_CACHED); + page->cp_type = type; + CFS_INIT_LIST_HEAD(&page->cp_layers); + CFS_INIT_LIST_HEAD(&page->cp_batch); + CFS_INIT_LIST_HEAD(&page->cp_flight); + mutex_init(&page->cp_mutex); + lu_ref_init(&page->cp_reference); + head = o->co_lu.lo_header; + cfs_list_for_each_entry(o, &head->loh_layers, + co_lu.lo_linkage) { + if (o->co_ops->coo_page_init != NULL) { + result = o->co_ops->coo_page_init(env, o, + page, vmpage); + if (result != 0) { + cl_page_delete0(env, page, 0); + cl_page_free(env, page); + page = ERR_PTR(result); + break; + } + } + } + if (result == 0) { + CS_PAGE_INC(o, total); + CS_PAGE_INC(o, create); + CS_PAGESTATE_DEC(o, CPS_CACHED); + } + } else { + page = ERR_PTR(-ENOMEM); + } + RETURN(page); } /** @@ -399,7 +385,6 @@ static struct cl_page *cl_page_find0(const struct lu_env *env, struct cl_page *page = NULL; struct cl_page *ghost = NULL; struct cl_object_header *hdr; - struct cl_site *site = cl_object_site(o); int err; LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT); @@ -408,16 +393,14 @@ static struct cl_page *cl_page_find0(const struct lu_env *env, ENTRY; hdr = cl_object_header(o); - cfs_atomic_inc(&site->cs_pages.cs_lookup); + CS_PAGE_INC(o, lookup); CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n", idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type); /* fast path. */ if (type == CPT_CACHEABLE) { - /* cl_page::cp_lock is used to protect the page state and - * refcount, but need an external lock to protect the - * child/parent relationship, so vmpage lock must be held for - * this purpose. */ + /* vmpage lock is used to protect the child/parent + * relationship */ KLASSERT(PageLocked(vmpage)); /* * cl_vmpage_page() can be called here without any locks as @@ -437,13 +420,13 @@ static struct cl_page *cl_page_find0(const struct lu_env *env, } if (page != NULL) { - cfs_atomic_inc(&site->cs_pages.cs_hit); + CS_PAGE_INC(o, hit); RETURN(page); } /* allocate and initialize cl_page */ - err = cl_page_alloc(env, o, idx, vmpage, type, &page); - if (err != 0) + page = cl_page_alloc(env, o, idx, vmpage, type); + if (IS_ERR(page)) RETURN(page); if (type == CPT_TRANSIENT) { @@ -459,7 +442,7 @@ static struct cl_page *cl_page_find0(const struct lu_env *env, * XXX optimization: use radix_tree_preload() here, and change tree * gfp mask to GFP_KERNEL in cl_object_header_init(). */ - cfs_spin_lock(&hdr->coh_page_guard); + spin_lock(&hdr->coh_page_guard); err = radix_tree_insert(&hdr->coh_tree, idx, page); if (err != 0) { ghost = page; @@ -487,10 +470,9 @@ static struct cl_page *cl_page_find0(const struct lu_env *env, } hdr->coh_pages++; } - cfs_spin_unlock(&hdr->coh_page_guard); + spin_unlock(&hdr->coh_page_guard); if (unlikely(ghost != NULL)) { - cfs_atomic_dec(&site->cs_pages.cs_busy); cl_page_delete0(env, ghost, 0); cl_page_free(env, ghost); } @@ -531,7 +513,7 @@ static inline int cl_page_invariant(const struct cl_page *pg) child = pg->cp_child; owner = pg->cp_owner; - return cfs_atomic_read(&pg->cp_ref) > 0 && + return cl_page_in_use(pg) && ergo(parent != NULL, parent->cp_child == pg) && ergo(child != NULL, child->cp_parent == pg) && ergo(child != NULL, pg->cp_obj != child->cp_obj) && @@ -554,9 +536,6 @@ static void cl_page_state_set0(const struct lu_env *env, struct cl_page *page, enum cl_page_state state) { enum cl_page_state old; -#ifdef LUSTRE_PAGESTATE_TRACKING - struct cl_site *site = cl_object_site(page->cp_obj); -#endif /* * Matrix of allowed state transitions [old][new], for sanity @@ -609,10 +588,8 @@ static void cl_page_state_set0(const struct lu_env *env, PASSERT(env, page, equi(state == CPS_OWNED, page->cp_owner != NULL)); -#ifdef LUSTRE_PAGESTATE_TRACKING - cfs_atomic_dec(&site->cs_pages_state[page->cp_state]); - cfs_atomic_inc(&site->cs_pages_state[state]); -#endif + CS_PAGESTATE_DEC(page->cp_obj, page->cp_state); + CS_PAGESTATE_INC(page->cp_obj, state); cl_page_state_set_trust(page, state); } EXIT; @@ -634,10 +611,9 @@ static void cl_page_state_set(const struct lu_env *env, */ void cl_page_get(struct cl_page *page) { - ENTRY; - LASSERT(page->cp_state != CPS_FREEING); - cl_page_get_trust(page); - EXIT; + ENTRY; + cl_page_get_trust(page); + EXIT; } EXPORT_SYMBOL(cl_page_get); @@ -652,43 +628,26 @@ EXPORT_SYMBOL(cl_page_get); */ void cl_page_put(const struct lu_env *env, struct cl_page *page) { - struct cl_site *site = cl_object_site(page->cp_obj); - PASSERT(env, page, cfs_atomic_read(&page->cp_ref) > !!page->cp_parent); ENTRY; CL_PAGE_HEADER(D_TRACE, env, page, "%d\n", cfs_atomic_read(&page->cp_ref)); - if (cfs_atomic_dec_and_lock(&page->cp_ref, &page->cp_lock)) { - cfs_atomic_dec(&site->cs_pages.cs_busy); - /* We're going to access the page w/o a reference, but it's - * ok because we have grabbed the lock cp_lock, which - * means nobody is able to free this page behind us. - */ - if (page->cp_state == CPS_FREEING) { - /* We drop the page reference and check the page state - * inside the cp_lock. So that if it gets here, - * it is the REALLY last reference to this page. - */ - cfs_spin_unlock(&page->cp_lock); - - LASSERT(cfs_atomic_read(&page->cp_ref) == 0); - PASSERT(env, page, page->cp_owner == NULL); - PASSERT(env, page, cfs_list_empty(&page->cp_batch)); - /* - * Page is no longer reachable by other threads. Tear - * it down. - */ - cl_page_free(env, page); - - EXIT; - return; - } - cfs_spin_unlock(&page->cp_lock); - } + if (cfs_atomic_dec_and_test(&page->cp_ref)) { + LASSERT(page->cp_state == CPS_FREEING); + + LASSERT(cfs_atomic_read(&page->cp_ref) == 0); + PASSERT(env, page, page->cp_owner == NULL); + PASSERT(env, page, cfs_list_empty(&page->cp_batch)); + /* + * Page is no longer reachable by other threads. Tear + * it down. + */ + cl_page_free(env, page); + } - EXIT; + EXIT; } EXPORT_SYMBOL(cl_page_put); @@ -740,16 +699,14 @@ struct cl_page *cl_vmpage_page(cfs_page_t *vmpage, struct cl_object *obj) if (top == NULL) RETURN(NULL); - cfs_spin_lock(&top->cp_lock); - for (page = top; page != NULL; page = page->cp_child) { - if (cl_object_same(page->cp_obj, obj)) { - cl_page_get_trust(page); - break; - } - } - cfs_spin_unlock(&top->cp_lock); - LASSERT(ergo(page, page->cp_type == CPT_CACHEABLE)); - RETURN(page); + for (page = top; page != NULL; page = page->cp_child) { + if (cl_object_same(page->cp_obj, obj)) { + cl_page_get_trust(page); + break; + } + } + LASSERT(ergo(page, page->cp_type == CPT_CACHEABLE)); + RETURN(page); } EXPORT_SYMBOL(cl_vmpage_page); @@ -1147,6 +1104,9 @@ static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg, cl_page_export(env, pg, 0); cl_page_state_set0(env, pg, CPS_FREEING); + CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete), + (const struct lu_env *, const struct cl_page_slice *)); + if (tmp->cp_type == CPT_CACHEABLE) { if (!radix) /* !radix means that @pg is not yet in the radix tree, @@ -1158,18 +1118,17 @@ static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg, struct cl_object_header *hdr; hdr = cl_object_header(tmp->cp_obj); - cfs_spin_lock(&hdr->coh_page_guard); - value = radix_tree_delete(&hdr->coh_tree, - tmp->cp_index); - PASSERT(env, tmp, value == tmp); - PASSERT(env, tmp, hdr->coh_pages > 0); - hdr->coh_pages--; - cfs_spin_unlock(&hdr->coh_page_guard); + spin_lock(&hdr->coh_page_guard); + value = radix_tree_delete(&hdr->coh_tree, + tmp->cp_index); + PASSERT(env, tmp, value == tmp); + PASSERT(env, tmp, hdr->coh_pages > 0); + hdr->coh_pages--; + spin_unlock(&hdr->coh_page_guard); + cl_page_put(env, tmp); } } - CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete), - (const struct lu_env *, const struct cl_page_slice *)); EXIT; } @@ -1308,6 +1267,8 @@ int cl_page_prep(const struct lu_env *env, struct cl_io *io, * PG_writeback without risking other layers deciding to skip this * page. */ + if (crt >= CRT_NR) + return -EINVAL; result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep)); if (result == 0) cl_page_io_start(env, pg, crt); @@ -1353,6 +1314,8 @@ void cl_page_completion(const struct lu_env *env, } cl_page_state_set(env, pg, CPS_CACHED); + if (crt >= CRT_NR) + return; CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion), (const struct lu_env *, const struct cl_page_slice *, int), ioret); @@ -1360,8 +1323,17 @@ void cl_page_completion(const struct lu_env *env, LASSERT(cl_page_is_vmlocked(env, pg)); LASSERT(pg->cp_sync_io == anchor); pg->cp_sync_io = NULL; + } + /* + * As page->cp_obj is pinned by a reference from page->cp_req, it is + * safe to call cl_page_put() without risking object destruction in a + * non-blocking context. + */ + cl_page_put(env, pg); + + if (anchor) cl_sync_io_note(anchor, ioret); - } + EXIT; } EXPORT_SYMBOL(cl_page_completion); @@ -1383,6 +1355,8 @@ int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg, PINVRNT(env, pg, crt < CRT_NR); ENTRY; + if (crt >= CRT_NR) + RETURN(-EINVAL); result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready), (const struct lu_env *, const struct cl_page_slice *)); @@ -1419,6 +1393,9 @@ int cl_page_cache_add(const struct lu_env *env, struct cl_io *io, ENTRY; + if (crt >= CRT_NR) + RETURN(-EINVAL); + cfs_list_for_each_entry(scan, &pg->cp_layers, cpl_linkage) { if (scan->cpl_ops->io[crt].cpo_cache_add == NULL) continue; @@ -1646,10 +1623,9 @@ EXPORT_SYMBOL(cl_page_slice_add); int cl_page_init(void) { - return lu_kmem_init(cl_page_caches); + return 0; } void cl_page_fini(void) { - lu_kmem_fini(cl_page_caches); }