X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fobdclass%2Fcl_page.c;h=03ad075187257aa5a55e77c6ac2f3686c36236ed;hp=ba12d8a8f2a8a8de70efc26556d9b2b75f6766df;hb=f625f670afbe954030ff81f0f8522137d6cdd335;hpb=c2ddf78e51e7a674c3bf9e40559c5b7ca2bfe120 diff --git a/lustre/obdclass/cl_page.c b/lustre/obdclass/cl_page.c index ba12d8a..03ad075 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, 2014, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -36,6 +36,7 @@ * Client Lustre Page. * * Author: Nikita Danilov + * Author: Jinshan Xiong */ #define DEBUG_SUBSYSTEM S_CLASS @@ -48,21 +49,7 @@ #include #include "cl_internal.h" -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 - } -}; +static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg); #ifdef LIBCFS_DEBUG # define PASSERT(env, page, expr) \ @@ -77,7 +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))) { \ @@ -85,40 +72,42 @@ 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 */ - -/** - * Internal version of cl_page_top, it should be called with page referenced, - * or cp_lock held. - */ -static struct cl_page *cl_page_top_trusted(struct cl_page *page) -{ - while (page->cp_parent != NULL) - page = page->cp_parent; - return page; -} + ((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) \ + atomic_inc(&cl_object_site(o)->cs_pages.cs_stats[CS_##item]) +#define CS_PAGE_DEC(o, item) \ + atomic_dec(&cl_object_site(o)->cs_pages.cs_stats[CS_##item]) +#define CS_PAGESTATE_INC(o, state) \ + atomic_inc(&cl_object_site(o)->cs_pages_state[state]) +#define CS_PAGESTATE_DEC(o, state) \ + 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_get(). * * This function can be used to obtain initial reference to previously * unreferenced cached object. It can be called only if concurrent page - * reclamation is somehow prevented, e.g., by locking page radix-tree - * (cl_object_header::hdr->coh_page_guard), or by keeping a lock on a VM page, + * reclamation is somehow prevented, e.g., by keeping a lock on a VM page, * associated with \a page. * * Use with care! Not exported. */ 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(atomic_read(&page->cp_ref) > 0); + atomic_inc(&page->cp_ref); } /** @@ -131,183 +120,43 @@ static const struct cl_page_slice * 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); - do { - cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) { - if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype) - RETURN(slice); - } - page = page->cp_child; - } while (page != NULL); - RETURN(NULL); -} - -/** - * Returns a page with given index in the given object, or NULL if no page is - * found. Acquires a reference on \a page. - * - * Locking: called under cl_object_header::coh_page_guard spin-lock. - */ -struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index) -{ - struct cl_page *page; - - LASSERT_SPIN_LOCKED(&hdr->coh_page_guard); - - page = radix_tree_lookup(&hdr->coh_tree, index); - 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 \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. - */ -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; - struct cl_page **pvec; - const struct cl_page_slice *slice; - const struct lu_device_type *dtype; - pgoff_t idx; - unsigned int nr; - unsigned int i; - unsigned int j; - int res = CLP_GANG_OKAY; - int tree_lock = 1; - ENTRY; - - 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); - 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; - - LASSERT(page->cp_type == CPT_CACHEABLE); - if (page->cp_index > end) { - end_of_region = 1; - break; - } - if (page->cp_state == CPS_FREEING) - continue; - - slice = cl_page_at_trusted(page, dtype); - /* - * Pages for lsm-less file has no underneath sub-page - * for osc, in case of ... - */ - PASSERT(env, page, slice != NULL); - - page = slice->cpl_page; - /* - * Can safely call cl_page_get_trust() under - * radix-tree spin-lock. - * - * XXX not true, because @page is from object another - * than @hdr and protected by different tree lock. - */ - cl_page_get_trust(page); - lu_ref_add_atomic(&page->cp_reference, - "gang_lookup", cfs_current()); - pvec[j++] = page; - } - - /* - * Here a delicate locking dance is performed. Current thread - * holds a reference to a page, but has to own it before it - * can be placed into queue. Owning implies waiting, so - * radix-tree lock is to be released. After a wait one has to - * check that pages weren't truncated (cl_page_own() returns - * error in the latter case). - */ - cfs_spin_unlock(&hdr->coh_page_guard); - tree_lock = 0; - - for (i = 0; i < j; ++i) { - page = pvec[i]; - if (res == CLP_GANG_OKAY) - res = (*cb)(env, io, page, cbdata); - lu_ref_del(&page->cp_reference, - "gang_lookup", cfs_current()); - cl_page_put(env, page); - } - if (nr < CLT_PVEC_SIZE || end_of_region) - break; - - if (res == CLP_GANG_OKAY && cfs_need_resched()) - res = CLP_GANG_RESCHED; - if (res != CLP_GANG_OKAY) - break; + const struct cl_page_slice *slice; + ENTRY; - cfs_spin_lock(&hdr->coh_page_guard); - tree_lock = 1; - } - if (tree_lock) - cfs_spin_unlock(&hdr->coh_page_guard); - RETURN(res); + list_for_each_entry(slice, &page->cp_layers, cpl_linkage) { + if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype) + RETURN(slice); + } + RETURN(NULL); } -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); + struct cl_object *obj = page->cp_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); - PASSERT(env, page, page->cp_req == NULL); - PASSERT(env, page, page->cp_parent == NULL); - PASSERT(env, page, page->cp_state == CPS_FREEING); + PASSERT(env, page, list_empty(&page->cp_batch)); + PASSERT(env, page, page->cp_owner == NULL); + PASSERT(env, page, page->cp_req == NULL); + PASSERT(env, page, page->cp_state == CPS_FREEING); - ENTRY; - cfs_might_sleep(); - while (!cfs_list_empty(&page->cp_layers)) { - struct cl_page_slice *slice; - - 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); - } - 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); - cl_object_put(env, obj); - lu_ref_fini(&page->cp_reference); - OBD_SLAB_FREE_PTR(page, cl_page_kmem); - EXIT; + ENTRY; + while (!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); + if (unlikely(slice->cpl_ops->cpo_fini != NULL)) + slice->cpl_ops->cpo_fini(env, slice); + } + 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_FREE(page, pagesize); + EXIT; } /** @@ -321,62 +170,53 @@ 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) +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; + struct cl_page *page; + struct lu_object_header *head; - 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); - -#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, + GFP_NOFS); + if (page != NULL) { + int result = 0; + atomic_set(&page->cp_ref, 1); + 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_vmpage = vmpage; + cl_page_state_set_trust(page, CPS_CACHED); + page->cp_type = type; + INIT_LIST_HEAD(&page->cp_layers); + INIT_LIST_HEAD(&page->cp_batch); + INIT_LIST_HEAD(&page->cp_flight); + 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) { + result = o->co_ops->coo_page_init(env, o, page, + ind); + if (result != 0) { + cl_page_delete0(env, page); + 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); } /** @@ -390,34 +230,28 @@ static int cl_page_alloc(const struct lu_env *env, struct cl_object *o, * * \see cl_object_find(), cl_lock_find() */ -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 *cl_page_find(const struct lu_env *env, + struct cl_object *o, + pgoff_t idx, struct page *vmpage, + enum cl_page_type type) { - 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; + struct cl_page *page = NULL; + struct cl_object_header *hdr; - LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT); - cfs_might_sleep(); + LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT); + might_sleep(); - ENTRY; + ENTRY; - hdr = cl_object_header(o); - cfs_atomic_inc(&site->cs_pages.cs_lookup); + hdr = cl_object_header(o); + 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 @@ -429,134 +263,27 @@ static struct cl_page *cl_page_find0(const struct lu_env *env, * reference on it. */ page = cl_vmpage_page(vmpage, o); - PINVRNT(env, page, - ergo(page != NULL, - cl_page_vmpage(env, page) == vmpage && - (void *)radix_tree_lookup(&hdr->coh_tree, - idx) == page)); - } - - if (page != NULL) { - cfs_atomic_inc(&site->cs_pages.cs_hit); - RETURN(page); + if (page != NULL) { + 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) - 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); - err = radix_tree_insert(&hdr->coh_tree, idx, page); - if (err != 0) { - ghost = page; - /* - * Noted by Jay: a lock on \a vmpage protects cl_page_find() - * from this race, but - * - * 0. it's better to have cl_page interface "locally - * consistent" so that its correctness can be reasoned - * about without appealing to the (obscure world of) VM - * locking. - * - * 1. handling this race allows ->coh_tree to remain - * consistent even when VM locking is somehow busted, - * which is very useful during diagnosing and debugging. - */ - 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++; - } - cfs_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); - } - 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); + page = cl_page_alloc(env, o, idx, vmpage, type); + RETURN(page); } 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; - struct cl_page *parent; - struct cl_page *child; - struct cl_io *owner; - - /* - * Page invariant is protected by a VM lock. - */ - LINVRNT(cl_page_is_vmlocked(NULL, pg)); - - header = cl_object_header(pg->cp_obj); - parent = pg->cp_parent; - child = pg->cp_child; - owner = pg->cp_owner; - - return cfs_atomic_read(&pg->cp_ref) > 0 && - ergo(parent != NULL, parent->cp_child == pg) && - ergo(child != NULL, child->cp_parent == pg) && - ergo(child != NULL, pg->cp_obj != child->cp_obj) && - ergo(parent != NULL, pg->cp_obj != parent->cp_obj) && - ergo(owner != NULL && parent != NULL, - parent->cp_owner == pg->cp_owner->ci_parent) && - ergo(owner != NULL && child != NULL, - child->cp_owner->ci_parent == owner) && - /* - * 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 && pg->cp_type == CPT_CACHEABLE, - (void *)radix_tree_lookup(&header->coh_tree, - pg->cp_index) == pg || - (child == NULL && parent == NULL)); + return cl_page_in_use_noref(pg); } 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 @@ -604,18 +331,13 @@ static void cl_page_state_set0(const struct lu_env *env, old = page->cp_state; PASSERT(env, page, allowed_transitions[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)); - -#ifdef LUSTRE_PAGESTATE_TRACKING - cfs_atomic_dec(&site->cs_pages_state[page->cp_state]); - cfs_atomic_inc(&site->cs_pages_state[state]); -#endif - cl_page_state_set_trust(page, state); - } - EXIT; + PASSERT(env, page, page->cp_state == old); + PASSERT(env, page, equi(state == CPS_OWNED, page->cp_owner != NULL)); + + CS_PAGESTATE_DEC(page->cp_obj, page->cp_state); + CS_PAGESTATE_INC(page->cp_obj, state); + cl_page_state_set_trust(page, state); + EXIT; } static void cl_page_state_set(const struct lu_env *env, @@ -651,118 +373,52 @@ 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); - } + atomic_read(&page->cp_ref)); + + if (atomic_dec_and_test(&page->cp_ref)) { + LASSERT(page->cp_state == CPS_FREEING); + + LASSERT(atomic_read(&page->cp_ref) == 0); + 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; + 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) -{ - const struct cl_page_slice *slice; - - /* - * Find uppermost layer with ->cpo_vmpage() method, and return its - * result. - */ - page = cl_page_top(page); - do { - 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)); - } - page = page->cp_child; - } while (page != NULL); - LBUG(); /* ->cpo_vmpage() has to be defined somewhere in the stack */ -} -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 *top; struct cl_page *page; - ENTRY; - KLASSERT(PageLocked(vmpage)); - - /* - * NOTE: absence of races and liveness of data are guaranteed by page - * lock on a "vmpage". That works because object destruction has - * bottom-to-top pass. - */ - - /* - * This loop assumes that ->private points to the top-most page. This - * can be rectified easily. - */ - top = (struct cl_page *)vmpage->private; - 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); + ENTRY; + KLASSERT(PageLocked(vmpage)); + + /* + * NOTE: absence of races and liveness of data are guaranteed by page + * lock on a "vmpage". That works because object destruction has + * bottom-to-top pass. + */ + + page = (struct cl_page *)vmpage->private; + if (page != NULL) { + cl_page_get_trust(page); + LASSERT(page->cp_type == CPT_CACHEABLE); + } + RETURN(page); } EXPORT_SYMBOL(cl_vmpage_page); -/** - * Returns the top-page for a given page. - * - * \see cl_object_top(), cl_io_top() - */ -struct cl_page *cl_page_top(struct cl_page *page) -{ - return cl_page_top_trusted(page); -} -EXPORT_SYMBOL(cl_page_top); - const struct cl_page_slice *cl_page_at(const struct cl_page *page, const struct lu_device_type *dtype) { @@ -772,80 +428,59 @@ EXPORT_SYMBOL(cl_page_at); #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname) -#define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...) \ -({ \ - const struct lu_env *__env = (_env); \ - struct cl_page *__page = (_page); \ - const struct cl_page_slice *__scan; \ - int __result; \ - ptrdiff_t __op = (_op); \ - int (*__method)_proto; \ - \ - __result = 0; \ - __page = cl_page_top(__page); \ - do { \ - cfs_list_for_each_entry(__scan, &__page->cp_layers, \ - cpl_linkage) { \ - __method = *(void **)((char *)__scan->cpl_ops + \ - __op); \ - if (__method != NULL) { \ - __result = (*__method)(__env, __scan, \ - ## __VA_ARGS__); \ - if (__result != 0) \ - break; \ - } \ - } \ - __page = __page->cp_child; \ - } while (__page != NULL && __result == 0); \ - if (__result > 0) \ - __result = 0; \ - __result; \ +#define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...) \ +({ \ + const struct lu_env *__env = (_env); \ + struct cl_page *__page = (_page); \ + const struct cl_page_slice *__scan; \ + int __result; \ + ptrdiff_t __op = (_op); \ + int (*__method)_proto; \ + \ + __result = 0; \ + list_for_each_entry(__scan, &__page->cp_layers, cpl_linkage) { \ + __method = *(void **)((char *)__scan->cpl_ops + __op); \ + if (__method != NULL) { \ + __result = (*__method)(__env, __scan, ## __VA_ARGS__); \ + if (__result != 0) \ + break; \ + } \ + } \ + if (__result > 0) \ + __result = 0; \ + __result; \ }) -#define CL_PAGE_INVOID(_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; \ - \ - __page = cl_page_top(__page); \ - do { \ - cfs_list_for_each_entry(__scan, &__page->cp_layers, \ - cpl_linkage) { \ - __method = *(void **)((char *)__scan->cpl_ops + \ - __op); \ - if (__method != NULL) \ - (*__method)(__env, __scan, \ - ## __VA_ARGS__); \ - } \ - __page = __page->cp_child; \ - } while (__page != NULL); \ +#define CL_PAGE_INVOID(_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; \ + \ + list_for_each_entry(__scan, &__page->cp_layers, cpl_linkage) { \ + __method = *(void **)((char *)__scan->cpl_ops + __op); \ + if (__method != NULL) \ + (*__method)(__env, __scan, ## __VA_ARGS__); \ + } \ } 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 { \ - 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); \ +#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. */ \ + 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__); \ + } \ } while (0) static int cl_page_invoke(const struct lu_env *env, @@ -874,26 +509,21 @@ static void cl_page_invoid(const struct lu_env *env, static void cl_page_owner_clear(struct cl_page *page) { - ENTRY; - for (page = cl_page_top(page); page != NULL; page = page->cp_child) { - if (page->cp_owner != NULL) { - LASSERT(page->cp_owner->ci_owned_nr > 0); - page->cp_owner->ci_owned_nr--; - page->cp_owner = NULL; - page->cp_task = NULL; - } - } - EXIT; + ENTRY; + if (page->cp_owner != NULL) { + LASSERT(page->cp_owner->ci_owned_nr > 0); + page->cp_owner->ci_owned_nr--; + page->cp_owner = NULL; + } + EXIT; } static void cl_page_owner_set(struct cl_page *page) { - ENTRY; - for (page = cl_page_top(page); page != NULL; page = page->cp_child) { - LASSERT(page->cp_owner != NULL); - page->cp_owner->ci_owned_nr++; - } - EXIT; + ENTRY; + LASSERT(page->cp_owner != NULL); + page->cp_owner->ci_owned_nr++; + EXIT; } void cl_page_disown0(const struct lu_env *env, @@ -904,7 +534,7 @@ void cl_page_disown0(const struct lu_env *env, ENTRY; state = pg->cp_state; PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING); - PINVRNT(env, pg, cl_page_invariant(pg)); + PINVRNT(env, pg, cl_page_invariant(pg) || state == CPS_FREEING); cl_page_owner_clear(pg); if (state == CPS_OWNED) @@ -926,9 +556,10 @@ void cl_page_disown0(const struct lu_env *env, */ int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io) { - LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj)); - ENTRY; - RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == io); + struct cl_io *top = cl_io_top((struct cl_io *)io); + LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj)); + ENTRY; + RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top); } EXPORT_SYMBOL(cl_page_is_owned); @@ -960,7 +591,6 @@ static int cl_page_own0(const struct lu_env *env, struct cl_io *io, PINVRNT(env, pg, !cl_page_is_owned(pg, io)); ENTRY; - pg = cl_page_top(pg); io = cl_io_top(io); if (pg->cp_state == CPS_FREEING) { @@ -974,8 +604,7 @@ static int cl_page_own0(const struct lu_env *env, struct cl_io *io, if (result == 0) { PASSERT(env, pg, pg->cp_owner == NULL); PASSERT(env, pg, pg->cp_req == NULL); - pg->cp_owner = io; - pg->cp_task = current; + pg->cp_owner = cl_io_top(io);; cl_page_owner_set(pg); if (pg->cp_state != CPS_FREEING) { cl_page_state_set(env, pg, CPS_OWNED); @@ -1026,19 +655,17 @@ EXPORT_SYMBOL(cl_page_own_try); void cl_page_assume(const struct lu_env *env, struct cl_io *io, struct cl_page *pg) { - PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj)); + PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj)); - ENTRY; - pg = cl_page_top(pg); - io = cl_io_top(io); + ENTRY; + 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); - cl_page_state_set(env, pg, CPS_OWNED); - EXIT; + cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume)); + PASSERT(env, pg, pg->cp_owner == NULL); + pg->cp_owner = cl_io_top(io); + cl_page_owner_set(pg); + cl_page_state_set(env, pg, CPS_OWNED); + EXIT; } EXPORT_SYMBOL(cl_page_assume); @@ -1060,7 +687,6 @@ void cl_page_unassume(const struct lu_env *env, PINVRNT(env, pg, cl_page_invariant(pg)); ENTRY; - pg = cl_page_top(pg); io = cl_io_top(io); cl_page_owner_clear(pg); cl_page_state_set(env, pg, CPS_CACHED); @@ -1086,13 +712,13 @@ EXPORT_SYMBOL(cl_page_unassume); void cl_page_disown(const struct lu_env *env, struct cl_io *io, struct cl_page *pg) { - PINVRNT(env, pg, cl_page_is_owned(pg, io)); + PINVRNT(env, pg, cl_page_is_owned(pg, io) || + pg->cp_state == CPS_FREEING); - ENTRY; - pg = cl_page_top(pg); - io = cl_io_top(io); - cl_page_disown0(env, io, pg); - EXIT; + ENTRY; + io = cl_io_top(io); + cl_page_disown0(env, io, pg); + EXIT; } EXPORT_SYMBOL(cl_page_disown); @@ -1109,25 +735,22 @@ EXPORT_SYMBOL(cl_page_disown); void cl_page_discard(const struct lu_env *env, struct cl_io *io, struct cl_page *pg) { - PINVRNT(env, pg, cl_page_is_owned(pg, io)); - PINVRNT(env, pg, cl_page_invariant(pg)); + PINVRNT(env, pg, cl_page_is_owned(pg, io)); + PINVRNT(env, pg, cl_page_invariant(pg)); - cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard)); + cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard)); } EXPORT_SYMBOL(cl_page_discard); /** * Version of cl_page_delete() that can be called for not fully constructed - * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0() + * pages, e.g. in an error handling cl_page_find()->cl_page_delete0() * path. Doesn't check page invariant. */ -static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg, - int radix) +static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg) { - struct cl_page *tmp = pg; ENTRY; - PASSERT(env, pg, pg == cl_page_top(pg)); PASSERT(env, pg, pg->cp_state != CPS_FREEING); /* @@ -1135,40 +758,11 @@ static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg, */ cl_page_owner_clear(pg); - /* - * unexport the page firstly before freeing it so that - * the page content is considered to be invalid. - * We have to do this because a CPS_FREEING cl_page may - * be NOT under the protection of a cl_lock. - * Afterwards, if this page is found by other threads, then this - * page will be forced to reread. - */ - cl_page_export(env, pg, 0); cl_page_state_set0(env, pg, CPS_FREEING); - 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); - 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), + CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_delete), (const struct lu_env *, const struct cl_page_slice *)); + EXIT; } @@ -1191,7 +785,6 @@ static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg, * Once page reaches cl_page_state::CPS_FREEING, all remaining references will * drain after some time, at which point page will be recycled. * - * \pre pg == cl_page_top(pg) * \pre VM page is locked * \post pg->cp_state == CPS_FREEING * @@ -1199,33 +792,14 @@ static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg, */ void cl_page_delete(const struct lu_env *env, struct cl_page *pg) { - PINVRNT(env, pg, cl_page_invariant(pg)); - ENTRY; - cl_page_delete0(env, pg, 1); - EXIT; + PINVRNT(env, pg, cl_page_invariant(pg)); + ENTRY; + cl_page_delete0(env, pg); + EXIT; } EXPORT_SYMBOL(cl_page_delete); /** - * Unmaps page from user virtual memory. - * - * Calls cl_page_operations::cpo_unmap() through all layers top-to-bottom. The - * layer responsible for VM interaction has to unmap page from user space - * virtual memory. - * - * \see cl_page_operations::cpo_unmap() - */ -int cl_page_unmap(const struct lu_env *env, - struct cl_io *io, struct cl_page *pg) -{ - PINVRNT(env, pg, cl_page_is_owned(pg, io)); - PINVRNT(env, pg, cl_page_invariant(pg)); - - return cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_unmap)); -} -EXPORT_SYMBOL(cl_page_unmap); - -/** * Marks page up-to-date. * * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The @@ -1253,7 +827,6 @@ int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg) const struct cl_page_slice *slice; ENTRY; - pg = cl_page_top_trusted((struct cl_page *)pg); slice = container_of(pg->cp_layers.next, const struct cl_page_slice, cpl_linkage); PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL); @@ -1307,15 +880,17 @@ int cl_page_prep(const struct lu_env *env, struct cl_io *io, * PG_writeback without risking other layers deciding to skip this * page. */ - result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep)); - if (result == 0) - cl_page_io_start(env, pg, crt); - - 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, "%d %d\n", crt, result); - return result; + 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); + + KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE, + equi(result == 0, + PageWriteback(cl_page_vmpage(pg))))); + CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result); + return result; } EXPORT_SYMBOL(cl_page_prep); @@ -1346,22 +921,27 @@ void cl_page_completion(const struct lu_env *env, ENTRY; 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; - } - 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; - cl_sync_io_note(anchor, ioret); - } - EXIT; + } + /* + * 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 != NULL) + cl_sync_io_note(env, anchor, ioret); + + EXIT; } EXPORT_SYMBOL(cl_page_completion); @@ -1382,6 +962,8 @@ 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 *)); @@ -1395,43 +977,6 @@ int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg, EXPORT_SYMBOL(cl_page_make_ready); /** - * Notify layers that high level io decided to place this page into a cache - * for future transfer. - * - * The layer implementing transfer engine (osc) has to register this page in - * its queues. - * - * \pre cl_page_is_owned(pg, io) - * \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) -{ - 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)); - - 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) @@ -1457,78 +1002,6 @@ int cl_page_flush(const struct lu_env *env, struct cl_io *io, EXPORT_SYMBOL(cl_page_flush); /** - * Checks whether page is protected by any extent lock is at least required - * mode. - * - * \return the same as in cl_page_operations::cpo_is_under_lock() method. - * \see cl_page_operations::cpo_is_under_lock() - */ -int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io, - struct cl_page *page) -{ - int rc; - - PINVRNT(env, page, cl_page_invariant(page)); - - ENTRY; - rc = CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_is_under_lock), - (const struct lu_env *, - const struct cl_page_slice *, struct cl_io *), - io); - PASSERT(env, page, rc != 0); - RETURN(rc); -} -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. - */ -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; - int result; - - ENTRY; - info = cl_env_info(env); - io = &info->clt_io; - - /* - * initialize the io. This is ugly since we never do IO in this - * 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); - } - - 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); -} -EXPORT_SYMBOL(cl_pages_prune); - -/** * Tells transfer engine that only part of a page is to be transmitted. * * \see cl_page_operations::cpo_clip() @@ -1552,12 +1025,11 @@ EXPORT_SYMBOL(cl_page_clip); void cl_page_header_print(const struct lu_env *env, void *cookie, lu_printer_t printer, const struct cl_page *pg) { - (*printer)(env, cookie, - "page@%p[%d %p:%lu ^%p_%p %d %d %d %p %p %#x]\n", - 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); + (*printer)(env, cookie, + "page@%p[%d %p %d %d %d %p %p]\n", + pg, atomic_read(&pg->cp_ref), pg->cp_obj, + pg->cp_state, pg->cp_error, pg->cp_type, + pg->cp_owner, pg->cp_req); } EXPORT_SYMBOL(cl_page_header_print); @@ -1567,16 +1039,12 @@ EXPORT_SYMBOL(cl_page_header_print); void cl_page_print(const struct lu_env *env, void *cookie, lu_printer_t printer, const struct cl_page *pg) { - struct cl_page *scan; - - for (scan = cl_page_top((struct cl_page *)pg); - scan != NULL; scan = scan->cp_child) - cl_page_header_print(env, cookie, printer, scan); - CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print), - (const struct lu_env *env, - const struct cl_page_slice *slice, - void *cookie, lu_printer_t p), cookie, printer); - (*printer)(env, cookie, "end page@%p\n", pg); + cl_page_header_print(env, cookie, printer, pg); + CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print), + (const struct lu_env *env, + const struct cl_page_slice *slice, + void *cookie, lu_printer_t p), cookie, printer); + (*printer)(env, cookie, "end page@%p\n", pg); } EXPORT_SYMBOL(cl_page_print); @@ -1589,17 +1057,13 @@ int cl_page_cancel(const struct lu_env *env, struct cl_page *page) (const struct lu_env *, const struct cl_page_slice *)); } -EXPORT_SYMBOL(cl_page_cancel); /** * Converts a byte offset within object \a obj into a page index. */ 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); @@ -1608,16 +1072,13 @@ 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) +size_t cl_page_size(const struct cl_object *obj) { - return 1 << CFS_PAGE_SHIFT; + return 1UL << PAGE_CACHE_SHIFT; } EXPORT_SYMBOL(cl_page_size); @@ -1631,24 +1092,64 @@ EXPORT_SYMBOL(cl_page_size); * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add() */ void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice, - struct cl_object *obj, - const struct cl_page_operations *ops) + struct cl_object *obj, pgoff_t index, + const struct cl_page_operations *ops) { - ENTRY; - cfs_list_add_tail(&slice->cpl_linkage, &page->cp_layers); - slice->cpl_obj = obj; - slice->cpl_ops = ops; - slice->cpl_page = page; - EXIT; + ENTRY; + list_add_tail(&slice->cpl_linkage, &page->cp_layers); + slice->cpl_obj = obj; + slice->cpl_index = index; + slice->cpl_ops = ops; + slice->cpl_page = page; + EXIT; } EXPORT_SYMBOL(cl_page_slice_add); -int cl_page_init(void) +/** + * Allocate and initialize cl_cache, called by ll_init_sbi(). + */ +struct cl_client_cache *cl_cache_init(unsigned long lru_page_max) +{ + struct cl_client_cache *cache = NULL; + + ENTRY; + OBD_ALLOC(cache, sizeof(*cache)); + if (cache == NULL) + RETURN(NULL); + + /* Initialize cache data */ + atomic_set(&cache->ccc_users, 1); + cache->ccc_lru_max = lru_page_max; + atomic_long_set(&cache->ccc_lru_left, lru_page_max); + spin_lock_init(&cache->ccc_lru_lock); + INIT_LIST_HEAD(&cache->ccc_lru); + + /* turn unstable check off by default as it impacts performance */ + cache->ccc_unstable_check = 0; + atomic_long_set(&cache->ccc_unstable_nr, 0); + init_waitqueue_head(&cache->ccc_unstable_waitq); + + RETURN(cache); +} +EXPORT_SYMBOL(cl_cache_init); + +/** + * Increase cl_cache refcount + */ +void cl_cache_incref(struct cl_client_cache *cache) { - return lu_kmem_init(cl_page_caches); + atomic_inc(&cache->ccc_users); } +EXPORT_SYMBOL(cl_cache_incref); -void cl_page_fini(void) +/** + * Decrease cl_cache refcount and free the cache if refcount=0. + * Since llite, lov and osc all hold cl_cache refcount, + * the free will not cause race. (LU-6173) + */ +void cl_cache_decref(struct cl_client_cache *cache) { - lu_kmem_fini(cl_page_caches); + if (atomic_dec_and_test(&cache->ccc_users)) + OBD_FREE(cache, sizeof(*cache)); } +EXPORT_SYMBOL(cl_cache_decref);