X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fobdclass%2Fcl_page.c;h=926035e5ace1c43d0392a0e0539087ca1f81d25f;hb=f4831ab428de8468d8ef25ddb72d3e34b69040b0;hp=e88427b51cfeb617a4b11cd7607c7c96314aa981;hpb=e46545c5af0b582e292b658cf741c47fdde343e9;p=fs%2Flustre-release.git diff --git a/lustre/obdclass/cl_page.c b/lustre/obdclass/cl_page.c index e88427b..926035e 100644 --- a/lustre/obdclass/cl_page.c +++ b/lustre/obdclass/cl_page.c @@ -26,7 +26,7 @@ * 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. */ /* @@ -99,7 +99,6 @@ static struct lu_kmem_descr cl_page_caches[] = { */ 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 +117,11 @@ 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); + if (cfs_atomic_inc_return(&page->cp_ref) == 1) + cfs_atomic_inc(&cl_object_site(page->cp_obj)->cs_pages.cs_busy); } /** @@ -141,14 +139,14 @@ cl_page_at_trusted(const struct cl_page *page, #ifdef INVARIANT_CHECK struct cl_object_header *ch = cl_object_header(page->cp_obj); - if (!atomic_read(&page->cp_ref)) + 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); 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); } @@ -171,7 +169,6 @@ struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index) page = radix_tree_lookup(&hdr->coh_tree, index); if (page != NULL) { - LASSERT(cl_is_page(page)); cl_page_get_trust(page); } return page; @@ -179,14 +176,19 @@ struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index) 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. */ 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) + struct cl_page_list *queue, int nonblock, + int *resched) { struct cl_object_header *hdr; struct cl_page *page; @@ -197,19 +199,25 @@ 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); 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; - spin_lock(&hdr->coh_page_guard); + cfs_spin_lock(&hdr->coh_page_guard); while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec, idx, CLT_PVEC_SIZE)) > 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) break; @@ -226,6 +234,7 @@ 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 @@ -248,20 +257,24 @@ 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); + cfs_spin_unlock(&hdr->coh_page_guard); for (i = 0; i < j; ++i) { page = pvec[i]; - if (cl_page_own(env, io, page) == 0) + if (page_own(env, io, page) == 0) cl_page_list_add(queue, page); lu_ref_del(&page->cp_reference, "page_list", cfs_current()); cl_page_put(env, page); } - spin_lock(&hdr->coh_page_guard); + cfs_spin_lock(&hdr->coh_page_guard); if (nr < CLT_PVEC_SIZE) break; + if (resched != NULL && cfs_need_resched()) { + *resched = 1; + break; + } } - spin_unlock(&hdr->coh_page_guard); + cfs_spin_unlock(&hdr->coh_page_guard); EXIT; } EXPORT_SYMBOL(cl_page_gang_lookup); @@ -271,25 +284,27 @@ 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); - 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]); + 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); @@ -320,9 +335,9 @@ static int cl_page_alloc(const struct lu_env *env, struct cl_object *o, ENTRY; result = +1; - OBD_SLAB_ALLOC_PTR(page, cl_page_kmem); + OBD_SLAB_ALLOC_PTR_GFP(page, cl_page_kmem, CFS_ALLOC_IO); if (page != NULL) { - atomic_set(&page->cp_ref, 1); + 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, @@ -333,10 +348,11 @@ static int cl_page_alloc(const struct lu_env *env, struct cl_object *o, 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); + cfs_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) { + 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); @@ -350,10 +366,13 @@ static int cl_page_alloc(const struct lu_env *env, struct cl_object *o, } } 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); + 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 @@ -373,9 +392,11 @@ 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 *ghost = NULL; @@ -384,14 +405,14 @@ struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o, int err; LINVRNT(type == CPT_CACHEABLE || type == CPT_TRANSIENT); - might_sleep(); + cfs_might_sleep(); ENTRY; hdr = cl_object_header(o); - atomic_inc(&site->cs_pages.cs_lookup); + cfs_atomic_inc(&site->cs_pages.cs_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) { @@ -411,12 +432,12 @@ struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o, (void *)radix_tree_lookup(&hdr->coh_tree, idx) == page)); } else { - spin_lock(&hdr->coh_page_guard); + cfs_spin_lock(&hdr->coh_page_guard); page = cl_page_lookup(hdr, idx); - spin_unlock(&hdr->coh_page_guard); + cfs_spin_unlock(&hdr->coh_page_guard); } if (page != NULL) { - atomic_inc(&site->cs_pages.cs_hit); + cfs_atomic_inc(&site->cs_pages.cs_hit); RETURN(page); } @@ -428,7 +449,7 @@ struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o, * 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); + cfs_spin_lock(&hdr->coh_page_guard); err = radix_tree_insert(&hdr->coh_tree, idx, page); if (err != 0) { ghost = page; @@ -445,6 +466,7 @@ 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. */ + page = ERR_PTR(err); if (err == -EEXIST) { /* * XXX in case of a lookup for CPT_TRANSIENT page, @@ -460,26 +482,47 @@ struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o, * transient pages, so it is impossible to * have conflicting transient pages. */ - spin_unlock(&hdr->coh_page_guard); + cfs_spin_unlock(&hdr->coh_page_guard); cl_page_put(env, page); - spin_lock(&hdr->coh_page_guard); + cfs_spin_lock(&hdr->coh_page_guard); page = ERR_PTR(-EBUSY); } - } else - page = ERR_PTR(err); - } else + } + } 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); + } + cfs_spin_unlock(&hdr->coh_page_guard); if (unlikely(ghost != NULL)) { - atomic_dec(&site->cs_pages.cs_busy); + 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); +} 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; @@ -487,7 +530,6 @@ 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. */ @@ -498,7 +540,7 @@ 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 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) && @@ -521,7 +563,9 @@ 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 @@ -568,14 +612,16 @@ 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]); +#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; @@ -619,23 +665,41 @@ 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); + 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, list_empty(&page->cp_batch)); + 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(&hdr->coh_page_guard); } + EXIT; } EXPORT_SYMBOL(cl_page_put); @@ -653,7 +717,7 @@ 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)); } @@ -669,6 +733,7 @@ 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_object_header *hdr; ENTRY; KLASSERT(PageLocked(vmpage)); @@ -683,6 +748,8 @@ 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)) { @@ -690,7 +757,8 @@ struct cl_page *cl_vmpage_page(cfs_page_t *vmpage, struct cl_object *obj) break; } } - LASSERT(ergo(page, cl_is_page(page) && page->cp_type == CPT_CACHEABLE)); + cfs_spin_unlock(&hdr->coh_page_guard); + LASSERT(ergo(page, page->cp_type == CPT_CACHEABLE)); RETURN(page); } EXPORT_SYMBOL(cl_vmpage_page); @@ -706,20 +774,6 @@ 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) { @@ -741,8 +795,8 @@ 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) { \ @@ -769,8 +823,8 @@ 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) \ @@ -781,28 +835,28 @@ 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, @@ -837,6 +891,7 @@ static void cl_page_owner_clear(struct cl_page *page) LASSERT(page->cp_owner->ci_owned_nr > 0); page->cp_owner->ci_owned_nr--; page->cp_owner = NULL; + page->cp_task = NULL; } } EXIT; @@ -889,7 +944,7 @@ int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io) EXPORT_SYMBOL(cl_page_is_owned); /** - * Owns a page by IO. + * Try to own a page by IO. * * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it * into cl_page_state::CPS_OWNED state. @@ -901,11 +956,15 @@ EXPORT_SYMBOL(cl_page_is_owned); * * \retval -ve failure, e.g., page was destroyed (and landed in * cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED). + * or, page was owned by another thread, or in IO. * * \see cl_page_disown() * \see cl_page_operations::cpo_own() + * \see cl_page_own_try() + * \see cl_page_own */ -int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg) +static int cl_page_own0(const struct lu_env *env, struct cl_io *io, + struct cl_page *pg, int nonblock) { int result; @@ -915,24 +974,57 @@ int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg) pg = cl_page_top(pg); io = cl_io_top(io); - cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_own)); - PASSERT(env, pg, pg->cp_owner == NULL); - PASSERT(env, pg, pg->cp_req == NULL); - pg->cp_owner = io; - cl_page_owner_set(pg); - if (pg->cp_state != CPS_FREEING) { - cl_page_state_set(env, pg, CPS_OWNED); - result = 0; - } else { - cl_page_disown0(env, io, pg); + if (pg->cp_state == CPS_FREEING) { result = -EAGAIN; + } else { + result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own), + (const struct lu_env *, + const struct cl_page_slice *, + struct cl_io *, int), + io, nonblock); + 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; + cl_page_owner_set(pg); + if (pg->cp_state != CPS_FREEING) { + cl_page_state_set(env, pg, CPS_OWNED); + } else { + cl_page_disown0(env, io, pg); + result = -EAGAIN; + } + } } PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg))); RETURN(result); } + +/** + * Own a page, might be blocked. + * + * \see cl_page_own0() + */ +int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg) +{ + return cl_page_own0(env, io, pg, 0); +} EXPORT_SYMBOL(cl_page_own); /** + * Nonblock version of cl_page_own(). + * + * \see cl_page_own0() + */ +int cl_page_own_try(const struct lu_env *env, struct cl_io *io, + struct cl_page *pg) +{ + return cl_page_own0(env, io, pg, 1); +} +EXPORT_SYMBOL(cl_page_own_try); + + +/** * Assume page ownership. * * Called when page is already locked by the hosting VM. @@ -956,6 +1048,7 @@ void cl_page_assume(const struct lu_env *env, cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume)); pg->cp_owner = io; + pg->cp_task = current; cl_page_owner_set(pg); cl_page_state_set(env, pg, CPS_OWNED); EXIT; @@ -1044,35 +1137,49 @@ EXPORT_SYMBOL(cl_page_discard); static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg, int radix) { + struct cl_page *tmp = pg; + ENTRY; + PASSERT(env, pg, pg == cl_page_top(pg)); PASSERT(env, pg, pg->cp_state != CPS_FREEING); - ENTRY; /* * Severe all ways to obtain new pointers to @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); - CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete), - (const struct lu_env *, const struct cl_page_slice *)); + if (!radix) /* * !radix means that @pg is not yet in the radix tree, skip * removing it. */ - pg = pg->cp_child; - for (; pg != NULL; pg = pg->cp_child) { + tmp = pg->cp_child; + for (; tmp != NULL; tmp = tmp->cp_child) { void *value; struct cl_object_header *hdr; - hdr = cl_object_header(pg->cp_obj); - spin_lock(&hdr->coh_page_guard); - value = radix_tree_delete(&hdr->coh_tree, pg->cp_index); - PASSERT(env, pg, value == pg); - PASSERT(env, pg, hdr->coh_pages > 0); + 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--; - spin_unlock(&hdr->coh_page_guard); + 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 *)); EXIT; } @@ -1133,17 +1240,17 @@ EXPORT_SYMBOL(cl_page_unmap); * Marks page up-to-date. * * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The - * layer responsible for VM interaction has to mark page as up-to-date. From - * this moment on, page can be shown to the user space without Lustre being - * notified, hence the name. + * layer responsible for VM interaction has to mark/clear page as up-to-date + * by the \a uptodate argument. * * \see cl_page_operations::cpo_export() */ -void cl_page_export(const struct lu_env *env, struct cl_page *pg) +void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate) { PINVRNT(env, pg, cl_page_invariant(pg)); CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export), - (const struct lu_env *, const struct cl_page_slice *)); + (const struct lu_env *, + const struct cl_page_slice *, int), uptodate); } EXPORT_SYMBOL(cl_page_export); @@ -1218,7 +1325,7 @@ 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); @@ -1241,6 +1348,8 @@ 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); @@ -1248,8 +1357,8 @@ void cl_page_completion(const struct lu_env *env, PINVRNT(env, pg, cl_page_invariant(pg)); ENTRY; - CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, ioret); - if (crt == CRT_READ) { + 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; } @@ -1258,6 +1367,11 @@ 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 (anchor) { + 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 @@ -1293,7 +1407,7 @@ 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); @@ -1326,7 +1440,7 @@ int cl_page_cache_add(const struct lu_env *env, struct cl_io *io, 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); + CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result); RETURN(result); } EXPORT_SYMBOL(cl_page_cache_add); @@ -1364,6 +1478,7 @@ int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj) struct cl_object *obj = cl_object_top(clobj); struct cl_io *io; struct cl_page_list *plist; + int resched; int result; ENTRY; @@ -1382,16 +1497,22 @@ int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj) RETURN(io->ci_result); } - cl_page_list_init(plist); - cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF, plist); - /* - * 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 { + 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) + cfs_cond_resched(); + } while (resched); cl_io_fini(env, io); RETURN(result); @@ -1408,7 +1529,7 @@ 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), @@ -1424,7 +1545,7 @@ 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); @@ -1505,7 +1626,7 @@ 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;