X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fobdclass%2Fcl_page.c;h=b34555a1fd65b19c4e50321d3d19969fcf7220b8;hp=bbca0021bc867a5645becac3d274711a67d12343;hb=3bffa4d32bc5b0bc71ba6873e262ddbca436bae1;hpb=c5607338d96aa222319bebe58aa16785f463315e diff --git a/lustre/obdclass/cl_page.c b/lustre/obdclass/cl_page.c index bbca002..b34555a 100644 --- a/lustre/obdclass/cl_page.c +++ b/lustre/obdclass/cl_page.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, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -39,9 +39,6 @@ */ #define DEBUG_SUBSYSTEM S_CLASS -#ifndef EXPORT_SYMTAB -# define EXPORT_SYMTAB -#endif #include #include @@ -54,19 +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 { \ @@ -93,13 +77,30 @@ 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 coh_page_guard 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) { - LASSERT(cl_is_page(page)); while (page->cp_parent != NULL) page = page->cp_parent; return page; @@ -118,12 +119,8 @@ static struct cl_page *cl_page_top_trusted(struct cl_page *page) */ static void cl_page_get_trust(struct cl_page *page) { - LASSERT(cl_is_page(page)); - /* - * Checkless version for trusted users. - */ - if (atomic_inc_return(&page->cp_ref) == 1) - 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); } /** @@ -137,18 +134,11 @@ cl_page_at_trusted(const struct cl_page *page, const struct lu_device_type *dtype) { const struct cl_page_slice *slice; - -#ifdef INVARIANT_CHECK - struct cl_object_header *ch = cl_object_header(page->cp_obj); - - if (!atomic_read(&page->cp_ref)) - LASSERT_SPIN_LOCKED(&ch->coh_page_guard); -#endif ENTRY; page = cl_page_top_trusted((struct cl_page *)page); do { - list_for_each_entry(slice, &page->cp_layers, cpl_linkage) { + cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) { if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype) RETURN(slice); } @@ -170,23 +160,27 @@ 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) { - LASSERT(cl_is_page(page)); + if (page != NULL) cl_page_get_trust(page); - } return page; } EXPORT_SYMBOL(cl_page_lookup); /** - * Returns a list of pages by a given [start, end] of @obj. + * Returns a list of pages by a given [start, end] of \a obj. + * + * \param resched If not NULL, then we give up before hogging CPU for too + * long and set *resched = 1, in that case caller should implement a retry + * logic. * * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely * crucial in the face of [offset, EOF] locks. + * + * Return at least one page in @queue unless there is no covered page. */ -void cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj, - struct cl_io *io, pgoff_t start, pgoff_t end, - struct cl_page_list *queue, int nonblock) +int cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj, + struct cl_io *io, pgoff_t start, pgoff_t end, + cl_page_gang_cb_t cb, void *cbdata) { struct cl_object_header *hdr; struct cl_page *page; @@ -197,33 +191,30 @@ void cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj, unsigned int nr; unsigned int i; unsigned int j; - int (*page_own)(const struct lu_env *env, - struct cl_io *io, - struct cl_page *pg); + int res = CLP_GANG_OKAY; + int tree_lock = 1; ENTRY; - page_own = nonblock ? cl_page_own_try : cl_page_own; - idx = start; hdr = cl_object_header(obj); pvec = cl_env_info(env)->clt_pvec; dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type; - 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; idx = pvec[nr - 1]->cp_index + 1; for (i = 0, j = 0; i < nr; ++i) { page = pvec[i]; - PASSERT(env, page, cl_is_page(page)); pvec[i] = NULL; - if (page->cp_index > end) + + LASSERT(page->cp_type == CPT_CACHEABLE); + if (page->cp_index > end) { + end_of_region = 1; break; + } if (page->cp_state == CPS_FREEING) continue; - if (page->cp_type == CPT_TRANSIENT) { - /* God, we found a transient page!*/ - continue; - } slice = cl_page_at_trusted(page, dtype); /* @@ -231,6 +222,7 @@ void cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj, * for osc, in case of ... */ PASSERT(env, page, slice != NULL); + page = slice->cpl_page; /* * Can safely call cl_page_get_trust() under @@ -241,7 +233,7 @@ void cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj, */ cl_page_get_trust(page); lu_ref_add_atomic(&page->cp_reference, - "page_list", cfs_current()); + "gang_lookup", cfs_current()); pvec[j++] = page; } @@ -253,52 +245,61 @@ void 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). */ - spin_unlock(&hdr->coh_page_guard); + spin_unlock(&hdr->coh_page_guard); + tree_lock = 0; + for (i = 0; i < j; ++i) { page = pvec[i]; - if (page_own(env, io, page) == 0) - cl_page_list_add(queue, page); + if (res == CLP_GANG_OKAY) + res = (*cb)(env, io, page, cbdata); lu_ref_del(&page->cp_reference, - "page_list", cfs_current()); + "gang_lookup", cfs_current()); cl_page_put(env, page); } - spin_lock(&hdr->coh_page_guard); - if (nr < CLT_PVEC_SIZE) + if (nr < CLT_PVEC_SIZE || end_of_region) break; - } - spin_unlock(&hdr->coh_page_guard); - EXIT; + + if (res == CLP_GANG_OKAY && cfs_need_resched()) + res = CLP_GANG_RESCHED; + if (res != CLP_GANG_OKAY) + break; + + 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, cl_is_page(page)); - PASSERT(env, page, list_empty(&page->cp_batch)); + PASSERT(env, page, cfs_list_empty(&page->cp_batch)); PASSERT(env, page, page->cp_owner == NULL); PASSERT(env, page, page->cp_req == NULL); PASSERT(env, page, page->cp_parent == NULL); PASSERT(env, page, page->cp_state == CPS_FREEING); ENTRY; - might_sleep(); - while (!list_empty(&page->cp_layers)) { + cfs_might_sleep(); + while (!cfs_list_empty(&page->cp_layers)) { struct cl_page_slice *slice; - slice = list_entry(page->cp_layers.next, struct cl_page_slice, - cpl_linkage); - list_del_init(page->cp_layers.next); + slice = cfs_list_entry(page->cp_layers.next, + struct cl_page_slice, cpl_linkage); + cfs_list_del_init(page->cp_layers.next); slice->cpl_ops->cpo_fini(env, slice); } - atomic_dec(&site->cs_pages.cs_total); - atomic_dec(&site->cs_pages_state[page->cp_state]); + 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; } @@ -313,58 +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) { - 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); - 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; - 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_state_set_trust(page, - CPS_FREEING); - cl_page_free(env, page); - page = err; - break; - } - } - } - if (err == NULL) { - atomic_inc(&site->cs_pages.cs_busy); - atomic_inc(&site->cs_pages.cs_total); - atomic_inc(&site->cs_pages_state[CPS_CACHED]); - atomic_inc(&site->cs_pages.cs_created); - result = 0; - } - } else - page = ERR_PTR(-ENOMEM); - *out = page; - RETURN(result); + struct cl_page *page; + struct lu_object_header *head; + + 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); } /** @@ -378,28 +376,32 @@ static int cl_page_alloc(const struct lu_env *env, struct cl_object *o, * * \see cl_object_find(), cl_lock_find() */ -struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o, - pgoff_t idx, struct page *vmpage, - enum cl_page_type type) +static struct cl_page *cl_page_find0(const struct lu_env *env, + struct cl_object *o, + pgoff_t idx, struct page *vmpage, + enum cl_page_type type, + struct cl_page *parent) { - struct cl_page *page; + 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; - LINVRNT(type == CPT_CACHEABLE || type == CPT_TRANSIENT); - might_sleep(); + LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT); + cfs_might_sleep(); ENTRY; hdr = cl_object_header(o); - atomic_inc(&site->cs_pages.cs_lookup); + CS_PAGE_INC(o, lookup); - CDEBUG(D_PAGE, "%lu@"DFID" %p %lu %i\n", + 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) { + /* vmpage lock is used to protect the child/parent + * relationship */ + KLASSERT(PageLocked(vmpage)); /* * cl_vmpage_page() can be called here without any locks as * @@ -415,25 +417,32 @@ struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o, cl_page_vmpage(env, page) == vmpage && (void *)radix_tree_lookup(&hdr->coh_tree, idx) == page)); - } else { - spin_lock(&hdr->coh_page_guard); - page = cl_page_lookup(hdr, idx); - spin_unlock(&hdr->coh_page_guard); } + if (page != NULL) { - 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) { + if (parent) { + LASSERT(page->cp_parent == NULL); + page->cp_parent = parent; + parent->cp_child = page; + } + RETURN(page); + } + /* * XXX optimization: use radix_tree_preload() here, and change tree * gfp mask to GFP_KERNEL in cl_object_header_init(). */ - 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; @@ -450,41 +459,43 @@ struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o, * consistent even when VM locking is somehow busted, * which is very useful during diagnosing and debugging. */ - if (err == -EEXIST) { - /* - * XXX in case of a lookup for CPT_TRANSIENT page, - * nothing protects a CPT_CACHEABLE page from being - * concurrently moved into CPS_FREEING state. - */ - page = cl_page_lookup(hdr, idx); - PASSERT(env, page, page != NULL); - if (page->cp_type == CPT_TRANSIENT && - type == CPT_CACHEABLE) { - /* XXX: We should make sure that inode sem - * keeps being held in the lifetime of - * transient pages, so it is impossible to - * have conflicting transient pages. - */ - spin_unlock(&hdr->coh_page_guard); - cl_page_put(env, page); - spin_lock(&hdr->coh_page_guard); - page = ERR_PTR(-EBUSY); - } - } else - page = ERR_PTR(err); - } else + page = ERR_PTR(err); + CL_PAGE_DEBUG(D_ERROR, env, ghost, + "fail to insert into radix tree: %d\n", err); + } else { + if (parent) { + LASSERT(page->cp_parent == NULL); + page->cp_parent = parent; + parent->cp_child = page; + } hdr->coh_pages++; - spin_unlock(&hdr->coh_page_guard); + } + spin_unlock(&hdr->coh_page_guard); if (unlikely(ghost != NULL)) { - atomic_dec(&site->cs_pages.cs_busy); cl_page_delete0(env, ghost, 0); cl_page_free(env, ghost); } RETURN(page); } + +struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o, + pgoff_t idx, struct page *vmpage, + enum cl_page_type type) +{ + return cl_page_find0(env, o, idx, vmpage, type, NULL); +} EXPORT_SYMBOL(cl_page_find); + +struct cl_page *cl_page_find_sub(const struct lu_env *env, struct cl_object *o, + pgoff_t idx, struct page *vmpage, + struct cl_page *parent) +{ + return cl_page_find0(env, o, idx, vmpage, parent->cp_type, parent); +} +EXPORT_SYMBOL(cl_page_find_sub); + static inline int cl_page_invariant(const struct cl_page *pg) { struct cl_object_header *header; @@ -492,7 +503,6 @@ static inline int cl_page_invariant(const struct cl_page *pg) struct cl_page *child; struct cl_io *owner; - LASSERT(cl_is_page(pg)); /* * Page invariant is protected by a VM lock. */ @@ -503,7 +513,7 @@ static inline int cl_page_invariant(const struct cl_page *pg) child = pg->cp_child; owner = pg->cp_owner; - return 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) && @@ -516,7 +526,7 @@ static inline int cl_page_invariant(const struct cl_page *pg) * Either page is early in initialization (has neither child * nor parent yet), or it is in the object radix tree. */ - ergo(pg->cp_state < CPS_FREEING, + ergo(pg->cp_state < CPS_FREEING && pg->cp_type == CPT_CACHEABLE, (void *)radix_tree_lookup(&header->coh_tree, pg->cp_index) == pg || (child == NULL && parent == NULL)); @@ -526,7 +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; - struct cl_site *site = cl_object_site(page->cp_obj); /* * Matrix of allowed state transitions [old][new], for sanity @@ -573,14 +582,14 @@ static void cl_page_state_set0(const struct lu_env *env, ENTRY; old = page->cp_state; PASSERT(env, page, allowed_transitions[old][state]); - CL_PAGE_HEADER(D_TRACE, env, page, "%i -> %i\n", old, state); + CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state); for (; page != NULL; page = page->cp_child) { PASSERT(env, page, page->cp_state == old); PASSERT(env, page, equi(state == CPS_OWNED, page->cp_owner != NULL)); - atomic_dec(&site->cs_pages_state[page->cp_state]); - atomic_inc(&site->cs_pages_state[state]); + CS_PAGESTATE_DEC(page->cp_obj, page->cp_state); + CS_PAGESTATE_INC(page->cp_obj, state); cl_page_state_set_trust(page, state); } EXIT; @@ -589,7 +598,6 @@ static void cl_page_state_set0(const struct lu_env *env, static void cl_page_state_set(const struct lu_env *env, struct cl_page *page, enum cl_page_state state) { - PINVRNT(env, page, cl_page_invariant(page)); cl_page_state_set0(env, page, state); } @@ -603,10 +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); @@ -621,27 +628,26 @@ EXPORT_SYMBOL(cl_page_get); */ void cl_page_put(const struct lu_env *env, struct cl_page *page) { - struct cl_object_header *hdr; - struct cl_site *site = cl_object_site(page->cp_obj); - - PASSERT(env, page, atomic_read(&page->cp_ref) > !!page->cp_parent); + PASSERT(env, page, cfs_atomic_read(&page->cp_ref) > !!page->cp_parent); ENTRY; - CL_PAGE_HEADER(D_TRACE, env, page, "%i\n", atomic_read(&page->cp_ref)); - hdr = cl_object_header(page->cp_obj); - if (atomic_dec_and_test(&page->cp_ref)) { - atomic_dec(&site->cs_pages.cs_busy); - if (page->cp_state == CPS_FREEING) { - PASSERT(env, page, page->cp_owner == NULL); - PASSERT(env, page, list_empty(&page->cp_batch)); - /* - * Page is no longer reachable by other threads. Tear - * it down. - */ - cl_page_free(env, page); - } - } - EXIT; + CL_PAGE_HEADER(D_TRACE, env, page, "%d\n", + cfs_atomic_read(&page->cp_ref)); + + 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; } EXPORT_SYMBOL(cl_page_put); @@ -658,7 +664,7 @@ cfs_page_t *cl_page_vmpage(const struct lu_env *env, struct cl_page *page) */ page = cl_page_top(page); do { - list_for_each_entry(slice, &page->cp_layers, cpl_linkage) { + cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) { if (slice->cpl_ops->cpo_vmpage != NULL) RETURN(slice->cpl_ops->cpo_vmpage(env, slice)); } @@ -673,7 +679,8 @@ EXPORT_SYMBOL(cl_page_vmpage); */ struct cl_page *cl_vmpage_page(cfs_page_t *vmpage, struct cl_object *obj) { - struct cl_page *page; + struct cl_page *top; + struct cl_page *page; ENTRY; KLASSERT(PageLocked(vmpage)); @@ -688,15 +695,18 @@ struct cl_page *cl_vmpage_page(cfs_page_t *vmpage, struct cl_object *obj) * This loop assumes that ->private points to the top-most page. This * can be rectified easily. */ - for (page = (void *)vmpage->private; - page != NULL; page = page->cp_child) { - if (cl_object_same(page->cp_obj, obj)) { - cl_page_get_trust(page); - break; - } - } - LASSERT(ergo(page, cl_is_page(page) && page->cp_type == CPT_CACHEABLE)); - RETURN(page); + top = (struct cl_page *)vmpage->private; + if (top == NULL) + RETURN(NULL); + + 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); @@ -711,20 +721,6 @@ struct cl_page *cl_page_top(struct cl_page *page) } EXPORT_SYMBOL(cl_page_top); -/** - * Returns true if \a addr is an address of an allocated cl_page. Used in - * assertions. This check is optimistically imprecise, i.e., it occasionally - * returns true for the incorrect addresses, but if it returns false, then the - * address is guaranteed to be incorrect. (Should be named cl_pagep().) - * - * \see cl_is_lock() - */ -int cl_is_page(const void *addr) -{ - return cfs_mem_is_in_cache(addr, cl_page_kmem); -} -EXPORT_SYMBOL(cl_is_page); - const struct cl_page_slice *cl_page_at(const struct cl_page *page, const struct lu_device_type *dtype) { @@ -746,8 +742,8 @@ EXPORT_SYMBOL(cl_page_at); __result = 0; \ __page = cl_page_top(__page); \ do { \ - list_for_each_entry(__scan, &__page->cp_layers, \ - cpl_linkage) { \ + cfs_list_for_each_entry(__scan, &__page->cp_layers, \ + cpl_linkage) { \ __method = *(void **)((char *)__scan->cpl_ops + \ __op); \ if (__method != NULL) { \ @@ -774,8 +770,8 @@ do { \ \ __page = cl_page_top(__page); \ do { \ - list_for_each_entry(__scan, &__page->cp_layers, \ - cpl_linkage) { \ + cfs_list_for_each_entry(__scan, &__page->cp_layers, \ + cpl_linkage) { \ __method = *(void **)((char *)__scan->cpl_ops + \ __op); \ if (__method != NULL) \ @@ -786,28 +782,28 @@ do { \ } while (__page != NULL); \ } while (0) -#define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...) \ -do { \ - const struct lu_env *__env = (_env); \ - struct cl_page *__page = (_page); \ - const struct cl_page_slice *__scan; \ - ptrdiff_t __op = (_op); \ - void (*__method)_proto; \ - \ - /* get to the bottom page. */ \ - while (__page->cp_child != NULL) \ - __page = __page->cp_child; \ - do { \ - list_for_each_entry_reverse(__scan, &__page->cp_layers, \ - cpl_linkage) { \ - __method = *(void **)((char *)__scan->cpl_ops + \ - __op); \ - if (__method != NULL) \ - (*__method)(__env, __scan, \ - ## __VA_ARGS__); \ - } \ - __page = __page->cp_parent; \ - } while (__page != NULL); \ +#define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...) \ +do { \ + const struct lu_env *__env = (_env); \ + struct cl_page *__page = (_page); \ + const struct cl_page_slice *__scan; \ + ptrdiff_t __op = (_op); \ + void (*__method)_proto; \ + \ + /* get to the bottom page. */ \ + while (__page->cp_child != NULL) \ + __page = __page->cp_child; \ + do { \ + cfs_list_for_each_entry_reverse(__scan, &__page->cp_layers, \ + cpl_linkage) { \ + __method = *(void **)((char *)__scan->cpl_ops + \ + __op); \ + if (__method != NULL) \ + (*__method)(__env, __scan, \ + ## __VA_ARGS__); \ + } \ + __page = __page->cp_parent; \ + } while (__page != NULL); \ } while (0) static int cl_page_invoke(const struct lu_env *env, @@ -926,7 +922,7 @@ static int cl_page_own0(const struct lu_env *env, struct cl_io *io, io = cl_io_top(io); if (pg->cp_state == CPS_FREEING) { - result = -EAGAIN; + result = -ENOENT; } else { result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own), (const struct lu_env *, @@ -943,7 +939,7 @@ static int cl_page_own0(const struct lu_env *env, struct cl_io *io, cl_page_state_set(env, pg, CPS_OWNED); } else { cl_page_disown0(env, io, pg); - result = -EAGAIN; + result = -ENOENT; } } } @@ -988,16 +984,14 @@ EXPORT_SYMBOL(cl_page_own_try); void cl_page_assume(const struct lu_env *env, struct cl_io *io, struct cl_page *pg) { - PASSERT(env, pg, pg->cp_state < CPS_OWNED); - PASSERT(env, pg, pg->cp_owner == NULL); PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj)); - PINVRNT(env, pg, cl_page_invariant(pg)); ENTRY; pg = cl_page_top(pg); io = cl_io_top(io); cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume)); + PASSERT(env, pg, pg->cp_owner == NULL); pg->cp_owner = io; pg->cp_task = current; cl_page_owner_set(pg); @@ -1110,27 +1104,31 @@ 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); - if (!radix) - /* - * !radix means that @pg is not yet in the radix tree, skip - * removing it. - */ - tmp = pg->cp_child; - for (; tmp != NULL; tmp = tmp->cp_child) { - void *value; - struct cl_object_header *hdr; - - hdr = cl_object_header(tmp->cp_obj); - 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_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, + * skip removing it. + */ + tmp = pg->cp_child; + for (; tmp != NULL; tmp = tmp->cp_child) { + void *value; + struct cl_object_header *hdr; + + hdr = cl_object_header(tmp->cp_obj); + 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); + } + } + EXIT; } @@ -1192,7 +1190,7 @@ EXPORT_SYMBOL(cl_page_unmap); * * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The * layer responsible for VM interaction has to mark/clear page as up-to-date - * by the @uptodate argument. + * by the \a uptodate argument. * * \see cl_page_operations::cpo_export() */ @@ -1276,7 +1274,7 @@ int cl_page_prep(const struct lu_env *env, struct cl_io *io, KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE, equi(result == 0, PageWriteback(cl_page_vmpage(env, pg))))); - CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, result); + CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result); return result; } EXPORT_SYMBOL(cl_page_prep); @@ -1299,14 +1297,15 @@ EXPORT_SYMBOL(cl_page_prep); void cl_page_completion(const struct lu_env *env, struct cl_page *pg, enum cl_req_type crt, int ioret) { + struct cl_sync_io *anchor = pg->cp_sync_io; + PASSERT(env, pg, crt < CRT_NR); /* cl_page::cp_req already cleared by the caller (osc_completion()) */ PASSERT(env, pg, pg->cp_req == NULL); PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt)); - PINVRNT(env, pg, cl_page_invariant(pg)); ENTRY; - CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, ioret); + CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret); if (crt == CRT_READ && ioret == 0) { PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED)); pg->cp_flags |= CPF_READ_COMPLETED; @@ -1316,17 +1315,12 @@ void cl_page_completion(const struct lu_env *env, CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion), (const struct lu_env *, const struct cl_page_slice *, int), ioret); - if (pg->cp_sync_io) { - cl_sync_io_note(pg->cp_sync_io, ioret); + if (anchor) { + LASSERT(cl_page_is_vmlocked(env, pg)); + LASSERT(pg->cp_sync_io == anchor); pg->cp_sync_io = NULL; + cl_sync_io_note(anchor, ioret); } - - /* Don't assert the page writeback bit here because the lustre file - * may be as a backend of swap space. in this case, the page writeback - * is set by VM, and obvious we shouldn't clear it at all. Fortunately - * this type of pages are all TRANSIENT pages. */ - KLASSERT(ergo(pg->cp_type == CPT_CACHEABLE, - !PageWriteback(cl_page_vmpage(env, pg)))); EXIT; } EXPORT_SYMBOL(cl_page_completion); @@ -1355,7 +1349,7 @@ int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg, PASSERT(env, pg, pg->cp_state == CPS_CACHED); cl_page_io_start(env, pg, crt); } - CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, result); + CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result); RETURN(result); } EXPORT_SYMBOL(cl_page_make_ready); @@ -1368,32 +1362,61 @@ EXPORT_SYMBOL(cl_page_make_ready); * its queues. * * \pre cl_page_is_owned(pg, io) - * \post ergo(result == 0, - * pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT) + * \post cl_page_is_owned(pg, io) * * \see cl_page_operations::cpo_cache_add() */ int cl_page_cache_add(const struct lu_env *env, struct cl_io *io, struct cl_page *pg, enum cl_req_type crt) { - int result; + const struct cl_page_slice *scan; + int result = 0; - PINVRNT(env, pg, crt < CRT_NR); - PINVRNT(env, pg, cl_page_is_owned(pg, io)); - PINVRNT(env, pg, cl_page_invariant(pg)); + PINVRNT(env, pg, crt < CRT_NR); + PINVRNT(env, pg, cl_page_is_owned(pg, io)); + PINVRNT(env, pg, cl_page_invariant(pg)); - ENTRY; - result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_cache_add)); - if (result == 0) { - cl_page_owner_clear(pg); - cl_page_state_set(env, pg, CPS_CACHED); - } - CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, result); - RETURN(result); + ENTRY; + + cfs_list_for_each_entry(scan, &pg->cp_layers, cpl_linkage) { + if (scan->cpl_ops->io[crt].cpo_cache_add == NULL) + continue; + + result = scan->cpl_ops->io[crt].cpo_cache_add(env, scan, io); + if (result != 0) + break; + } + CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result); + RETURN(result); } EXPORT_SYMBOL(cl_page_cache_add); /** + * Called if a pge is being written back by kernel's intention. + * + * \pre cl_page_is_owned(pg, io) + * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT) + * + * \see cl_page_operations::cpo_flush() + */ +int cl_page_flush(const struct lu_env *env, struct cl_io *io, + struct cl_page *pg) +{ + int result; + + PINVRNT(env, pg, cl_page_is_owned(pg, io)); + PINVRNT(env, pg, cl_page_invariant(pg)); + + ENTRY; + + result = cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_flush)); + + CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result); + RETURN(result); +} +EXPORT_SYMBOL(cl_page_flush); + +/** * Checks whether page is protected by any extent lock is at least required * mode. * @@ -1417,6 +1440,16 @@ int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io, } EXPORT_SYMBOL(cl_page_is_under_lock); +static int page_prune_cb(const struct lu_env *env, struct cl_io *io, + struct cl_page *page, void *cbdata) +{ + cl_page_own(env, io, page); + cl_page_unmap(env, io, page); + cl_page_discard(env, io, page); + cl_page_disown(env, io, page); + return CLP_GANG_OKAY; +} + /** * Purges all cached pages belonging to the object \a obj. */ @@ -1425,12 +1458,10 @@ int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj) struct cl_thread_info *info; struct cl_object *obj = cl_object_top(clobj); struct cl_io *io; - struct cl_page_list *plist; int result; ENTRY; info = cl_env_info(env); - plist = &info->clt_list; io = &info->clt_io; /* @@ -1438,22 +1469,19 @@ int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj) * function, we just make cl_page_list functions happy. -jay */ io->ci_obj = obj; + io->ci_ignore_layout = 1; result = cl_io_init(env, io, CIT_MISC, obj); if (result != 0) { cl_io_fini(env, io); RETURN(io->ci_result); } - cl_page_list_init(plist); - cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF, plist, 0); - /* - * Since we're purging the pages of an object, we don't care - * the possible outcomes of the following functions. - */ - cl_page_list_unmap(env, io, plist); - cl_page_list_discard(env, io, plist); - cl_page_list_disown(env, io, plist); - cl_page_list_fini(env, plist); + do { + result = cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF, + page_prune_cb, NULL); + if (result == CLP_GANG_RESCHED) + cfs_cond_resched(); + } while (result != CLP_GANG_OKAY); cl_io_fini(env, io); RETURN(result); @@ -1470,7 +1498,7 @@ void cl_page_clip(const struct lu_env *env, struct cl_page *pg, { PINVRNT(env, pg, cl_page_invariant(pg)); - CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", from, to); + CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to); CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip), (const struct lu_env *, const struct cl_page_slice *,int, int), @@ -1486,7 +1514,7 @@ void cl_page_header_print(const struct lu_env *env, void *cookie, { (*printer)(env, cookie, "page@%p[%d %p:%lu ^%p_%p %d %d %d %p %p %#x]\n", - pg, atomic_read(&pg->cp_ref), pg->cp_obj, + pg, cfs_atomic_read(&pg->cp_ref), pg->cp_obj, pg->cp_index, pg->cp_parent, pg->cp_child, pg->cp_state, pg->cp_error, pg->cp_type, pg->cp_owner, pg->cp_req, pg->cp_flags); @@ -1567,7 +1595,7 @@ void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice, const struct cl_page_operations *ops) { ENTRY; - list_add_tail(&slice->cpl_linkage, &page->cp_layers); + cfs_list_add_tail(&slice->cpl_linkage, &page->cp_layers); slice->cpl_obj = obj; slice->cpl_ops = ops; slice->cpl_page = page; @@ -1577,10 +1605,9 @@ 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); }