X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fobdclass%2Fcl_page.c;h=b7dfab1e1df8748896db632a418842f2eea7e4d3;hb=bc71055256db623ba2062f8f299d8b603d89e0d9;hp=926035e5ace1c43d0392a0e0539087ca1f81d25f;hpb=84a795f90412ee575703476728f1522c06de4a2c;p=fs%2Flustre-release.git diff --git a/lustre/obdclass/cl_page.c b/lustre/obdclass/cl_page.c index 926035e..b7dfab1 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. @@ -28,6 +26,8 @@ /* * 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/ @@ -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 { \ @@ -80,7 +64,7 @@ static struct lu_kmem_descr cl_page_caches[] = { ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp)) #endif /* !LIBCFS_DEBUG */ -#ifdef INVARIANT_CHECK +#ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK # define PINVRNT(env, page, expr) \ do { \ if (unlikely(!(expr))) { \ @@ -88,14 +72,32 @@ static struct lu_kmem_descr cl_page_caches[] = { LINVRNT(0); \ } \ } while (0) -#else /* !INVARIANT_CHECK */ +#else /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */ # define PINVRNT(env, page, exp) \ - ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp)) -#endif /* !INVARIANT_CHECK */ + ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp)) +#endif /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_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) { @@ -117,11 +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); } /** @@ -135,13 +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 - struct cl_object_header *ch = cl_object_header(page->cp_obj); - - if (!cfs_atomic_read(&page->cp_ref)) - LASSERT_SPIN_LOCKED(&ch->coh_page_guard); -#endif ENTRY; page = cl_page_top_trusted((struct cl_page *)page); @@ -168,9 +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); @@ -184,11 +175,12 @@ EXPORT_SYMBOL(cl_page_lookup); * * 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 *resched) +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; @@ -199,34 +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; - if (resched != NULL) - *resched = 0; - 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; - 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; idx = pvec[nr - 1]->cp_index + 1; for (i = 0, j = 0; i < nr; ++i) { page = pvec[i]; 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); /* @@ -245,7 +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; } @@ -257,32 +245,38 @@ 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). */ - cfs_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); } - cfs_spin_lock(&hdr->coh_page_guard); - if (nr < CLT_PVEC_SIZE) + if (nr < CLT_PVEC_SIZE || end_of_region) break; - if (resched != NULL && cfs_need_resched()) { - *resched = 1; + + if (res == CLP_GANG_OKAY && cfs_need_resched()) + res = CLP_GANG_RESCHED; + if (res != CLP_GANG_OKAY) break; - } - } - cfs_spin_unlock(&hdr->coh_page_guard); - EXIT; + + 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); @@ -300,15 +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 - lu_object_ref_del_at(&obj->co_lu, page->cp_obj_ref, "cl_page", page); + 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; } @@ -323,62 +314,56 @@ 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); - 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_state_set_trust(page, - CPS_FREEING); - 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); - -#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); + struct cl_page *page; + struct lu_object_header *head; + + ENTRY; + OBD_ALLOC_GFP(page, cl_object_header(o)->coh_page_bufsize, + __GFP_IO); + if (page != NULL) { + int result = 0; + 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); + lu_object_ref_add_at(&o->co_lu, &page->cp_obj_ref, "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); } /** @@ -398,24 +383,26 @@ static struct cl_page *cl_page_find0(const struct lu_env *env, 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); + LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT); cfs_might_sleep(); 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) { + /* vmpage lock is used to protect the child/parent + * relationship */ + KLASSERT(PageLocked(vmpage)); /* * cl_vmpage_page() can be called here without any locks as * @@ -431,25 +418,32 @@ static struct cl_page *cl_page_find0(const struct lu_env *env, cl_page_vmpage(env, page) == vmpage && (void *)radix_tree_lookup(&hdr->coh_tree, idx) == page)); - } else { - cfs_spin_lock(&hdr->coh_page_guard); - page = cl_page_lookup(hdr, idx); - cfs_spin_unlock(&hdr->coh_page_guard); } + 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) { + 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(). */ - 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; @@ -467,27 +461,8 @@ static struct cl_page *cl_page_find0(const struct lu_env *env, * which is very useful during diagnosing and debugging. */ page = ERR_PTR(err); - 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. - */ - cfs_spin_unlock(&hdr->coh_page_guard); - cl_page_put(env, page); - cfs_spin_lock(&hdr->coh_page_guard); - page = ERR_PTR(-EBUSY); - } - } + CL_PAGE_DEBUG(D_ERROR, env, ghost, + "fail to insert into radix tree: %d\n", err); } else { if (parent) { LASSERT(page->cp_parent == NULL); @@ -496,10 +471,9 @@ 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); } @@ -540,7 +514,7 @@ 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) && @@ -553,7 +527,7 @@ 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)); @@ -563,9 +537,6 @@ 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 @@ -618,10 +589,8 @@ 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; @@ -630,7 +599,6 @@ 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); } @@ -644,10 +612,9 @@ 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); @@ -662,52 +629,33 @@ 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, 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)); - hdr = cl_object_header(cl_object_top(page->cp_obj)); - if (cfs_atomic_dec_and_lock(&page->cp_ref, &hdr->coh_page_guard)) { - 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 coh_page_guard, 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 coh_page_guard. So that if it gets here, - * it is the REALLY last reference to this page. - */ - cfs_spin_unlock(&hdr->coh_page_guard); - - 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); + if (cfs_atomic_dec_and_test(&page->cp_ref)) { + LASSERT(page->cp_state == CPS_FREEING); - EXIT; - return; - } - cfs_spin_unlock(&hdr->coh_page_guard); - } + 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); /** * Returns a VM page associated with a given cl_page. */ -cfs_page_t *cl_page_vmpage(const struct lu_env *env, struct cl_page *page) +struct page *cl_page_vmpage(const struct lu_env *env, struct cl_page *page) { const struct cl_page_slice *slice; @@ -730,10 +678,10 @@ EXPORT_SYMBOL(cl_page_vmpage); /** * Returns a cl_page associated with a VM page, and given cl_object. */ -struct cl_page *cl_vmpage_page(cfs_page_t *vmpage, struct cl_object *obj) +struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj) { - struct cl_page *page; - struct cl_object_header *hdr; + struct cl_page *top; + struct cl_page *page; ENTRY; KLASSERT(PageLocked(vmpage)); @@ -748,18 +696,18 @@ 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. */ - hdr = cl_object_header(cl_object_top(obj)); - cfs_spin_lock(&hdr->coh_page_guard); - 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; - } - } - cfs_spin_unlock(&hdr->coh_page_guard); - LASSERT(ergo(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); @@ -975,7 +923,7 @@ 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 *, @@ -992,7 +940,7 @@ 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; } } } @@ -1037,16 +985,14 @@ 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); @@ -1159,27 +1105,31 @@ 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); - 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); - } - 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; } @@ -1318,6 +1268,8 @@ 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); @@ -1354,7 +1306,6 @@ void cl_page_completion(const struct lu_env *env, /* 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, "%d %d\n", crt, ioret); @@ -1364,21 +1315,26 @@ 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); if (anchor) { + 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); - } - /* 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); @@ -1400,6 +1356,8 @@ 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 *)); @@ -1420,32 +1378,64 @@ 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, "%d %d\n", crt, result); - RETURN(result); + 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; + + 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. * @@ -1469,6 +1459,16 @@ 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. */ @@ -1477,13 +1477,10 @@ 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 resched; int result; ENTRY; info = cl_env_info(env); - plist = &info->clt_list; io = &info->clt_io; /* @@ -1491,6 +1488,7 @@ 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); @@ -1498,21 +1496,11 @@ int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj) } do { - cl_page_list_init(plist); - cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF, plist, 0, - &resched); - /* - * 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); - - if (resched) + 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 (resched); + } while (result != CLP_GANG_OKAY); cl_io_fini(env, io); RETURN(result); @@ -1587,10 +1575,7 @@ EXPORT_SYMBOL(cl_page_cancel); */ loff_t cl_offset(const struct cl_object *obj, pgoff_t idx) { - /* - * XXX for now. - */ - return (loff_t)idx << CFS_PAGE_SHIFT; + return (loff_t)idx << PAGE_CACHE_SHIFT; } EXPORT_SYMBOL(cl_offset); @@ -1599,16 +1584,13 @@ EXPORT_SYMBOL(cl_offset); */ pgoff_t cl_index(const struct cl_object *obj, loff_t offset) { - /* - * XXX for now. - */ - return offset >> CFS_PAGE_SHIFT; + return offset >> PAGE_CACHE_SHIFT; } EXPORT_SYMBOL(cl_index); int cl_page_size(const struct cl_object *obj) { - return 1 << CFS_PAGE_SHIFT; + return 1 << PAGE_CACHE_SHIFT; } EXPORT_SYMBOL(cl_page_size); @@ -1636,10 +1618,9 @@ 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); }