4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * Author: Nikita Danilov <nikita.danilov@sun.com>
35 * Author: Jinshan Xiong <jinshan.xiong@intel.com>
38 #define DEBUG_SUBSYSTEM S_CLASS
40 #include <linux/list.h>
41 #include <libcfs/libcfs.h>
42 #include <obd_class.h>
43 #include <obd_support.h>
45 #include <cl_object.h>
46 #include "cl_internal.h"
48 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg);
49 static DEFINE_MUTEX(cl_page_kmem_mutex);
52 # define PASSERT(env, page, expr) \
54 if (unlikely(!(expr))) { \
55 CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n"); \
59 #else /* !LIBCFS_DEBUG */
60 # define PASSERT(env, page, exp) \
61 ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
62 #endif /* !LIBCFS_DEBUG */
64 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
65 # define PINVRNT(env, page, expr) \
67 if (unlikely(!(expr))) { \
68 CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n"); \
72 #else /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
73 # define PINVRNT(env, page, exp) \
74 ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
75 #endif /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
77 /* Disable page statistic by default due to huge performance penalty. */
78 static void cs_page_inc(const struct cl_object *obj,
79 enum cache_stats_item item)
81 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
82 atomic_inc(&cl_object_site(obj)->cs_pages.cs_stats[item]);
86 static void cs_page_dec(const struct cl_object *obj,
87 enum cache_stats_item item)
89 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
90 atomic_dec(&cl_object_site(obj)->cs_pages.cs_stats[item]);
94 static void cs_pagestate_inc(const struct cl_object *obj,
95 enum cl_page_state state)
97 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
98 atomic_inc(&cl_object_site(obj)->cs_pages_state[state]);
102 static void cs_pagestate_dec(const struct cl_object *obj,
103 enum cl_page_state state)
105 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
106 atomic_dec(&cl_object_site(obj)->cs_pages_state[state]);
111 * Internal version of cl_page_get().
113 * This function can be used to obtain initial reference to previously
114 * unreferenced cached object. It can be called only if concurrent page
115 * reclamation is somehow prevented, e.g., by keeping a lock on a VM page,
116 * associated with \a page.
118 * Use with care! Not exported.
120 static void cl_page_get_trust(struct cl_page *page)
122 LASSERT(atomic_read(&page->cp_ref) > 0);
123 atomic_inc(&page->cp_ref);
126 static struct cl_page_slice *
127 cl_page_slice_get(const struct cl_page *cl_page, int index)
129 if (index < 0 || index >= cl_page->cp_layer_count)
132 /* To get the cp_layer_offset values fit under 256 bytes, we
133 * use the offset beyond the end of struct cl_page.
135 return (struct cl_page_slice *)((char *)cl_page + sizeof(*cl_page) +
136 cl_page->cp_layer_offset[index]);
139 #define cl_page_slice_for_each(cl_page, slice, i) \
140 for (i = 0, slice = cl_page_slice_get(cl_page, 0); \
141 i < (cl_page)->cp_layer_count; \
142 slice = cl_page_slice_get(cl_page, ++i))
144 #define cl_page_slice_for_each_reverse(cl_page, slice, i) \
145 for (i = (cl_page)->cp_layer_count - 1, \
146 slice = cl_page_slice_get(cl_page, i); i >= 0; \
147 slice = cl_page_slice_get(cl_page, --i))
150 * Returns a slice within a cl_page, corresponding to the given layer in the
155 static const struct cl_page_slice *
156 cl_page_at_trusted(const struct cl_page *cl_page,
157 const struct lu_device_type *dtype)
159 const struct cl_page_slice *slice;
164 cl_page_slice_for_each(cl_page, slice, i) {
165 if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
172 static void __cl_page_free(struct cl_page *cl_page, unsigned short bufsize)
174 int index = cl_page->cp_kmem_index;
177 LASSERT(index < ARRAY_SIZE(cl_page_kmem_array));
178 LASSERT(cl_page_kmem_size_array[index] == bufsize);
179 OBD_SLAB_FREE(cl_page, cl_page_kmem_array[index], bufsize);
181 OBD_FREE(cl_page, bufsize);
185 static void cl_page_free(const struct lu_env *env, struct cl_page *cl_page,
186 struct pagevec *pvec)
188 struct cl_object *obj = cl_page->cp_obj;
189 unsigned short bufsize = cl_object_header(obj)->coh_page_bufsize;
190 struct cl_page_slice *slice;
194 PASSERT(env, cl_page, list_empty(&cl_page->cp_batch));
195 PASSERT(env, cl_page, cl_page->cp_owner == NULL);
196 PASSERT(env, cl_page, cl_page->cp_state == CPS_FREEING);
198 cl_page_slice_for_each(cl_page, slice, i) {
199 if (unlikely(slice->cpl_ops->cpo_fini != NULL))
200 slice->cpl_ops->cpo_fini(env, slice, pvec);
202 cl_page->cp_layer_count = 0;
203 cs_page_dec(obj, CS_total);
204 cs_pagestate_dec(obj, cl_page->cp_state);
205 lu_object_ref_del_at(&obj->co_lu, &cl_page->cp_obj_ref,
207 cl_object_put(env, obj);
208 lu_ref_fini(&cl_page->cp_reference);
209 __cl_page_free(cl_page, bufsize);
213 static struct cl_page *__cl_page_alloc(struct cl_object *o)
216 struct cl_page *cl_page = NULL;
217 unsigned short bufsize = cl_object_header(o)->coh_page_bufsize;
220 /* the number of entries in cl_page_kmem_array is expected to
221 * only be 2-3 entries, so the lookup overhead should be low.
223 for ( ; i < ARRAY_SIZE(cl_page_kmem_array); i++) {
224 if (smp_load_acquire(&cl_page_kmem_size_array[i])
226 OBD_SLAB_ALLOC_GFP(cl_page, cl_page_kmem_array[i],
229 cl_page->cp_kmem_index = i;
232 if (cl_page_kmem_size_array[i] == 0)
236 if (i < ARRAY_SIZE(cl_page_kmem_array)) {
239 mutex_lock(&cl_page_kmem_mutex);
240 if (cl_page_kmem_size_array[i]) {
241 mutex_unlock(&cl_page_kmem_mutex);
244 snprintf(cache_name, sizeof(cache_name),
245 "cl_page_kmem-%u", bufsize);
246 cl_page_kmem_array[i] =
247 kmem_cache_create(cache_name, bufsize,
249 if (cl_page_kmem_array[i] == NULL) {
250 mutex_unlock(&cl_page_kmem_mutex);
253 smp_store_release(&cl_page_kmem_size_array[i],
255 mutex_unlock(&cl_page_kmem_mutex);
258 OBD_ALLOC_GFP(cl_page, bufsize, GFP_NOFS);
260 cl_page->cp_kmem_index = -1;
266 struct cl_page *cl_page_alloc(const struct lu_env *env, struct cl_object *o,
267 pgoff_t ind, struct page *vmpage,
268 enum cl_page_type type)
270 struct cl_page *cl_page;
271 struct lu_object_header *head;
275 cl_page = __cl_page_alloc(o);
276 if (cl_page != NULL) {
280 * Please fix cl_page:cp_state/type declaration if
281 * these assertions fail in the future.
283 BUILD_BUG_ON((1 << CP_STATE_BITS) < CPS_NR); /* cp_state */
284 BUILD_BUG_ON((1 << CP_TYPE_BITS) < CPT_NR); /* cp_type */
285 atomic_set(&cl_page->cp_ref, 1);
288 lu_object_ref_add_at(&o->co_lu, &cl_page->cp_obj_ref,
290 cl_page->cp_vmpage = vmpage;
291 cl_page->cp_state = CPS_CACHED;
292 cl_page->cp_type = type;
293 INIT_LIST_HEAD(&cl_page->cp_batch);
294 lu_ref_init(&cl_page->cp_reference);
295 head = o->co_lu.lo_header;
296 list_for_each_entry(o, &head->loh_layers,
298 if (o->co_ops->coo_page_init != NULL) {
299 result = o->co_ops->coo_page_init(env, o,
302 cl_page_delete0(env, cl_page);
303 cl_page_free(env, cl_page, NULL);
304 cl_page = ERR_PTR(result);
310 cs_page_inc(o, CS_total);
311 cs_page_inc(o, CS_create);
312 cs_pagestate_dec(o, CPS_CACHED);
315 cl_page = ERR_PTR(-ENOMEM);
321 * Returns a cl_page with index \a idx at the object \a o, and associated with
322 * the VM page \a vmpage.
324 * This is the main entry point into the cl_page caching interface. First, a
325 * cache (implemented as a per-object radix tree) is consulted. If page is
326 * found there, it is returned immediately. Otherwise new page is allocated
327 * and returned. In any case, additional reference to page is acquired.
329 * \see cl_object_find(), cl_lock_find()
331 struct cl_page *cl_page_find(const struct lu_env *env,
333 pgoff_t idx, struct page *vmpage,
334 enum cl_page_type type)
336 struct cl_page *page = NULL;
337 struct cl_object_header *hdr;
339 LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
344 hdr = cl_object_header(o);
345 cs_page_inc(o, CS_lookup);
347 CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
348 idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
350 if (type == CPT_CACHEABLE) {
351 /* vmpage lock is used to protect the child/parent
353 LASSERT(PageLocked(vmpage));
355 * cl_vmpage_page() can be called here without any locks as
357 * - "vmpage" is locked (which prevents ->private from
358 * concurrent updates), and
360 * - "o" cannot be destroyed while current thread holds a
363 page = cl_vmpage_page(vmpage, o);
365 cs_page_inc(o, CS_hit);
370 /* allocate and initialize cl_page */
371 page = cl_page_alloc(env, o, idx, vmpage, type);
374 EXPORT_SYMBOL(cl_page_find);
376 static inline int cl_page_invariant(const struct cl_page *pg)
378 return cl_page_in_use_noref(pg);
381 static void cl_page_state_set0(const struct lu_env *env,
382 struct cl_page *cl_page,
383 enum cl_page_state state)
385 enum cl_page_state old;
388 * Matrix of allowed state transitions [old][new], for sanity
391 static const int allowed_transitions[CPS_NR][CPS_NR] = {
394 [CPS_OWNED] = 1, /* io finds existing cached page */
396 [CPS_PAGEOUT] = 1, /* write-out from the cache */
397 [CPS_FREEING] = 1, /* eviction on the memory pressure */
400 [CPS_CACHED] = 1, /* release to the cache */
402 [CPS_PAGEIN] = 1, /* start read immediately */
403 [CPS_PAGEOUT] = 1, /* start write immediately */
404 [CPS_FREEING] = 1, /* lock invalidation or truncate */
407 [CPS_CACHED] = 1, /* io completion */
414 [CPS_CACHED] = 1, /* io completion */
430 old = cl_page->cp_state;
431 PASSERT(env, cl_page, allowed_transitions[old][state]);
432 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d -> %d\n", old, state);
433 PASSERT(env, cl_page, cl_page->cp_state == old);
434 PASSERT(env, cl_page, equi(state == CPS_OWNED,
435 cl_page->cp_owner != NULL));
437 cs_pagestate_dec(cl_page->cp_obj, cl_page->cp_state);
438 cs_pagestate_inc(cl_page->cp_obj, state);
439 cl_page->cp_state = state;
443 static void cl_page_state_set(const struct lu_env *env,
444 struct cl_page *page, enum cl_page_state state)
446 cl_page_state_set0(env, page, state);
450 * Acquires an additional reference to a page.
452 * This can be called only by caller already possessing a reference to \a
455 * \see cl_object_get(), cl_lock_get().
457 void cl_page_get(struct cl_page *page)
460 cl_page_get_trust(page);
463 EXPORT_SYMBOL(cl_page_get);
466 * Releases a reference to a page, use the pagevec to release the pages
467 * in batch if provided.
469 * Users need to do a final pagevec_release() to release any trailing pages.
471 void cl_pagevec_put(const struct lu_env *env, struct cl_page *page,
472 struct pagevec *pvec)
475 CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
476 atomic_read(&page->cp_ref));
478 if (atomic_dec_and_test(&page->cp_ref)) {
479 LASSERT(page->cp_state == CPS_FREEING);
481 LASSERT(atomic_read(&page->cp_ref) == 0);
482 PASSERT(env, page, page->cp_owner == NULL);
483 PASSERT(env, page, list_empty(&page->cp_batch));
485 * Page is no longer reachable by other threads. Tear
488 cl_page_free(env, page, pvec);
493 EXPORT_SYMBOL(cl_pagevec_put);
496 * Releases a reference to a page, wrapper to cl_pagevec_put
498 * When last reference is released, page is returned to the cache, unless it
499 * is in cl_page_state::CPS_FREEING state, in which case it is immediately
502 * \see cl_object_put(), cl_lock_put().
504 void cl_page_put(const struct lu_env *env, struct cl_page *page)
506 cl_pagevec_put(env, page, NULL);
508 EXPORT_SYMBOL(cl_page_put);
511 * Returns a cl_page associated with a VM page, and given cl_object.
513 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj)
515 struct cl_page *page;
518 LASSERT(PageLocked(vmpage));
521 * NOTE: absence of races and liveness of data are guaranteed by page
522 * lock on a "vmpage". That works because object destruction has
523 * bottom-to-top pass.
526 page = (struct cl_page *)vmpage->private;
528 cl_page_get_trust(page);
529 LASSERT(page->cp_type == CPT_CACHEABLE);
533 EXPORT_SYMBOL(cl_vmpage_page);
535 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
536 const struct lu_device_type *dtype)
538 return cl_page_at_trusted(page, dtype);
540 EXPORT_SYMBOL(cl_page_at);
542 static void cl_page_owner_clear(struct cl_page *page)
545 if (page->cp_owner != NULL) {
546 LASSERT(page->cp_owner->ci_owned_nr > 0);
547 page->cp_owner->ci_owned_nr--;
548 page->cp_owner = NULL;
553 static void cl_page_owner_set(struct cl_page *page)
556 LASSERT(page->cp_owner != NULL);
557 page->cp_owner->ci_owned_nr++;
561 void cl_page_disown0(const struct lu_env *env,
562 struct cl_io *io, struct cl_page *cl_page)
564 const struct cl_page_slice *slice;
565 enum cl_page_state state;
569 state = cl_page->cp_state;
570 PINVRNT(env, cl_page, state == CPS_OWNED ||
571 state == CPS_FREEING);
572 PINVRNT(env, cl_page, cl_page_invariant(cl_page) ||
573 state == CPS_FREEING);
574 cl_page_owner_clear(cl_page);
576 if (state == CPS_OWNED)
577 cl_page_state_set(env, cl_page, CPS_CACHED);
579 * Completion call-backs are executed in the bottom-up order, so that
580 * uppermost layer (llite), responsible for VFS/VM interaction runs
581 * last and can release locks safely.
583 cl_page_slice_for_each_reverse(cl_page, slice, i) {
584 if (slice->cpl_ops->cpo_disown != NULL)
585 (*slice->cpl_ops->cpo_disown)(env, slice, io);
592 * returns true, iff page is owned by the given io.
594 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
596 struct cl_io *top = cl_io_top((struct cl_io *)io);
597 LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
599 RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top);
601 EXPORT_SYMBOL(cl_page_is_owned);
604 * Try to own a page by IO.
606 * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
607 * into cl_page_state::CPS_OWNED state.
609 * \pre !cl_page_is_owned(cl_page, io)
610 * \post result == 0 iff cl_page_is_owned(cl_page, io)
614 * \retval -ve failure, e.g., cl_page was destroyed (and landed in
615 * cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
616 * or, page was owned by another thread, or in IO.
618 * \see cl_page_disown()
619 * \see cl_page_operations::cpo_own()
620 * \see cl_page_own_try()
623 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
624 struct cl_page *cl_page, int nonblock)
626 const struct cl_page_slice *slice;
631 PINVRNT(env, cl_page, !cl_page_is_owned(cl_page, io));
634 if (cl_page->cp_state == CPS_FREEING) {
639 cl_page_slice_for_each(cl_page, slice, i) {
640 if (slice->cpl_ops->cpo_own)
641 result = (*slice->cpl_ops->cpo_own)(env, slice,
650 PASSERT(env, cl_page, cl_page->cp_owner == NULL);
651 cl_page->cp_owner = cl_io_top(io);
652 cl_page_owner_set(cl_page);
653 if (cl_page->cp_state != CPS_FREEING) {
654 cl_page_state_set(env, cl_page, CPS_OWNED);
656 cl_page_disown0(env, io, cl_page);
662 PINVRNT(env, cl_page, ergo(result == 0,
663 cl_page_invariant(cl_page)));
668 * Own a page, might be blocked.
670 * \see cl_page_own0()
672 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
674 return cl_page_own0(env, io, pg, 0);
676 EXPORT_SYMBOL(cl_page_own);
679 * Nonblock version of cl_page_own().
681 * \see cl_page_own0()
683 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
686 return cl_page_own0(env, io, pg, 1);
688 EXPORT_SYMBOL(cl_page_own_try);
692 * Assume page ownership.
694 * Called when page is already locked by the hosting VM.
696 * \pre !cl_page_is_owned(cl_page, io)
697 * \post cl_page_is_owned(cl_page, io)
699 * \see cl_page_operations::cpo_assume()
701 void cl_page_assume(const struct lu_env *env,
702 struct cl_io *io, struct cl_page *cl_page)
704 const struct cl_page_slice *slice;
709 PINVRNT(env, cl_page,
710 cl_object_same(cl_page->cp_obj, io->ci_obj));
713 cl_page_slice_for_each(cl_page, slice, i) {
714 if (slice->cpl_ops->cpo_assume != NULL)
715 (*slice->cpl_ops->cpo_assume)(env, slice, io);
718 PASSERT(env, cl_page, cl_page->cp_owner == NULL);
719 cl_page->cp_owner = cl_io_top(io);
720 cl_page_owner_set(cl_page);
721 cl_page_state_set(env, cl_page, CPS_OWNED);
724 EXPORT_SYMBOL(cl_page_assume);
727 * Releases page ownership without unlocking the page.
729 * Moves cl_page into cl_page_state::CPS_CACHED without releasing a lock
730 * on the underlying VM page (as VM is supposed to do this itself).
732 * \pre cl_page_is_owned(cl_page, io)
733 * \post !cl_page_is_owned(cl_page, io)
735 * \see cl_page_assume()
737 void cl_page_unassume(const struct lu_env *env,
738 struct cl_io *io, struct cl_page *cl_page)
740 const struct cl_page_slice *slice;
744 PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
745 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
748 cl_page_owner_clear(cl_page);
749 cl_page_state_set(env, cl_page, CPS_CACHED);
751 cl_page_slice_for_each_reverse(cl_page, slice, i) {
752 if (slice->cpl_ops->cpo_unassume != NULL)
753 (*slice->cpl_ops->cpo_unassume)(env, slice, io);
758 EXPORT_SYMBOL(cl_page_unassume);
761 * Releases page ownership.
763 * Moves page into cl_page_state::CPS_CACHED.
765 * \pre cl_page_is_owned(pg, io)
766 * \post !cl_page_is_owned(pg, io)
769 * \see cl_page_operations::cpo_disown()
771 void cl_page_disown(const struct lu_env *env,
772 struct cl_io *io, struct cl_page *pg)
774 PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
775 pg->cp_state == CPS_FREEING);
779 cl_page_disown0(env, io, pg);
782 EXPORT_SYMBOL(cl_page_disown);
785 * Called when cl_page is to be removed from the object, e.g.,
786 * as a result of truncate.
788 * Calls cl_page_operations::cpo_discard() top-to-bottom.
790 * \pre cl_page_is_owned(cl_page, io)
792 * \see cl_page_operations::cpo_discard()
794 void cl_page_discard(const struct lu_env *env,
795 struct cl_io *io, struct cl_page *cl_page)
797 const struct cl_page_slice *slice;
800 PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
801 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
803 cl_page_slice_for_each(cl_page, slice, i) {
804 if (slice->cpl_ops->cpo_discard != NULL)
805 (*slice->cpl_ops->cpo_discard)(env, slice, io);
808 EXPORT_SYMBOL(cl_page_discard);
811 * Version of cl_page_delete() that can be called for not fully constructed
812 * cl_pages, e.g. in an error handling cl_page_find()->cl_page_delete0()
813 * path. Doesn't check cl_page invariant.
815 static void cl_page_delete0(const struct lu_env *env,
816 struct cl_page *cl_page)
818 const struct cl_page_slice *slice;
823 PASSERT(env, cl_page, cl_page->cp_state != CPS_FREEING);
826 * Severe all ways to obtain new pointers to @pg.
828 cl_page_owner_clear(cl_page);
829 cl_page_state_set0(env, cl_page, CPS_FREEING);
831 cl_page_slice_for_each_reverse(cl_page, slice, i) {
832 if (slice->cpl_ops->cpo_delete != NULL)
833 (*slice->cpl_ops->cpo_delete)(env, slice);
840 * Called when a decision is made to throw page out of memory.
842 * Notifies all layers about page destruction by calling
843 * cl_page_operations::cpo_delete() method top-to-bottom.
845 * Moves page into cl_page_state::CPS_FREEING state (this is the only place
846 * where transition to this state happens).
848 * Eliminates all venues through which new references to the page can be
851 * - removes page from the radix trees,
853 * - breaks linkage from VM page to cl_page.
855 * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
856 * drain after some time, at which point page will be recycled.
858 * \pre VM page is locked
859 * \post pg->cp_state == CPS_FREEING
861 * \see cl_page_operations::cpo_delete()
863 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
865 PINVRNT(env, pg, cl_page_invariant(pg));
867 cl_page_delete0(env, pg);
870 EXPORT_SYMBOL(cl_page_delete);
873 * Marks page up-to-date.
875 * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
876 * layer responsible for VM interaction has to mark/clear page as up-to-date
877 * by the \a uptodate argument.
879 * \see cl_page_operations::cpo_export()
881 void cl_page_export(const struct lu_env *env, struct cl_page *cl_page,
884 const struct cl_page_slice *slice;
887 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
889 cl_page_slice_for_each(cl_page, slice, i) {
890 if (slice->cpl_ops->cpo_export != NULL)
891 (*slice->cpl_ops->cpo_export)(env, slice, uptodate);
894 EXPORT_SYMBOL(cl_page_export);
897 * Returns true, if \a page is VM locked in a suitable sense by the calling
900 int cl_page_is_vmlocked(const struct lu_env *env,
901 const struct cl_page *cl_page)
903 const struct cl_page_slice *slice;
907 slice = cl_page_slice_get(cl_page, 0);
908 PASSERT(env, cl_page, slice->cpl_ops->cpo_is_vmlocked != NULL);
910 * Call ->cpo_is_vmlocked() directly instead of going through
911 * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
912 * cl_page_invariant().
914 result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
915 PASSERT(env, cl_page, result == -EBUSY || result == -ENODATA);
917 RETURN(result == -EBUSY);
919 EXPORT_SYMBOL(cl_page_is_vmlocked);
921 void cl_page_touch(const struct lu_env *env,
922 const struct cl_page *cl_page, size_t to)
924 const struct cl_page_slice *slice;
929 cl_page_slice_for_each(cl_page, slice, i) {
930 if (slice->cpl_ops->cpo_page_touch != NULL)
931 (*slice->cpl_ops->cpo_page_touch)(env, slice, to);
936 EXPORT_SYMBOL(cl_page_touch);
938 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
941 RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
944 static void cl_page_io_start(const struct lu_env *env,
945 struct cl_page *pg, enum cl_req_type crt)
948 * Page is queued for IO, change its state.
951 cl_page_owner_clear(pg);
952 cl_page_state_set(env, pg, cl_req_type_state(crt));
957 * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
958 * called top-to-bottom. Every layer either agrees to submit this page (by
959 * returning 0), or requests to omit this page (by returning -EALREADY). Layer
960 * handling interactions with the VM also has to inform VM that page is under
963 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
964 struct cl_page *cl_page, enum cl_req_type crt)
966 const struct cl_page_slice *slice;
970 PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
971 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
972 PINVRNT(env, cl_page, crt < CRT_NR);
975 * this has to be called bottom-to-top, so that llite can set up
976 * PG_writeback without risking other layers deciding to skip this
982 cl_page_slice_for_each(cl_page, slice, i) {
983 if (slice->cpl_ops->cpo_own)
984 result = (*slice->cpl_ops->io[crt].cpo_prep)(env,
993 cl_page_io_start(env, cl_page, crt);
996 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", crt, result);
999 EXPORT_SYMBOL(cl_page_prep);
1002 * Notify layers about transfer completion.
1004 * Invoked by transfer sub-system (which is a part of osc) to notify layers
1005 * that a transfer, of which this page is a part of has completed.
1007 * Completion call-backs are executed in the bottom-up order, so that
1008 * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1009 * and can release locks safely.
1011 * \pre cl_page->cp_state == CPS_PAGEIN || cl_page->cp_state == CPS_PAGEOUT
1012 * \post cl_page->cl_page_state == CPS_CACHED
1014 * \see cl_page_operations::cpo_completion()
1016 void cl_page_completion(const struct lu_env *env,
1017 struct cl_page *cl_page, enum cl_req_type crt,
1020 const struct cl_page_slice *slice;
1021 struct cl_sync_io *anchor = cl_page->cp_sync_io;
1025 PASSERT(env, cl_page, crt < CRT_NR);
1026 PASSERT(env, cl_page, cl_page->cp_state == cl_req_type_state(crt));
1028 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", crt, ioret);
1029 cl_page_state_set(env, cl_page, CPS_CACHED);
1033 cl_page_slice_for_each_reverse(cl_page, slice, i) {
1034 if (slice->cpl_ops->io[crt].cpo_completion != NULL)
1035 (*slice->cpl_ops->io[crt].cpo_completion)(env, slice,
1039 if (anchor != NULL) {
1040 LASSERT(cl_page->cp_sync_io == anchor);
1041 cl_page->cp_sync_io = NULL;
1042 cl_sync_io_note(env, anchor, ioret);
1046 EXPORT_SYMBOL(cl_page_completion);
1049 * Notify layers that transfer formation engine decided to yank this page from
1050 * the cache and to make it a part of a transfer.
1052 * \pre cl_page->cp_state == CPS_CACHED
1053 * \post cl_page->cp_state == CPS_PAGEIN || cl_page->cp_state == CPS_PAGEOUT
1055 * \see cl_page_operations::cpo_make_ready()
1057 int cl_page_make_ready(const struct lu_env *env, struct cl_page *cl_page,
1058 enum cl_req_type crt)
1060 const struct cl_page_slice *slice;
1065 PINVRNT(env, cl_page, crt < CRT_NR);
1069 cl_page_slice_for_each(cl_page, slice, i) {
1070 if (slice->cpl_ops->io[crt].cpo_make_ready != NULL)
1071 result = (*slice->cpl_ops->io[crt].cpo_make_ready)(env, slice);
1078 PASSERT(env, cl_page, cl_page->cp_state == CPS_CACHED);
1079 cl_page_io_start(env, cl_page, crt);
1081 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", crt, result);
1085 EXPORT_SYMBOL(cl_page_make_ready);
1088 * Called if a page is being written back by kernel's intention.
1090 * \pre cl_page_is_owned(cl_page, io)
1091 * \post ergo(result == 0, cl_page->cp_state == CPS_PAGEOUT)
1093 * \see cl_page_operations::cpo_flush()
1095 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
1096 struct cl_page *cl_page)
1098 const struct cl_page_slice *slice;
1103 PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
1104 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
1106 cl_page_slice_for_each(cl_page, slice, i) {
1107 if (slice->cpl_ops->cpo_flush != NULL)
1108 result = (*slice->cpl_ops->cpo_flush)(env, slice, io);
1115 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d\n", result);
1118 EXPORT_SYMBOL(cl_page_flush);
1121 * Tells transfer engine that only part of a page is to be transmitted.
1123 * \see cl_page_operations::cpo_clip()
1125 void cl_page_clip(const struct lu_env *env, struct cl_page *cl_page,
1128 const struct cl_page_slice *slice;
1131 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
1133 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", from, to);
1134 cl_page_slice_for_each(cl_page, slice, i) {
1135 if (slice->cpl_ops->cpo_clip != NULL)
1136 (*slice->cpl_ops->cpo_clip)(env, slice, from, to);
1139 EXPORT_SYMBOL(cl_page_clip);
1142 * Prints human readable representation of \a pg to the \a f.
1144 void cl_page_header_print(const struct lu_env *env, void *cookie,
1145 lu_printer_t printer, const struct cl_page *pg)
1147 (*printer)(env, cookie,
1148 "page@%p[%d %p %d %d %p]\n",
1149 pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1150 pg->cp_state, pg->cp_type,
1153 EXPORT_SYMBOL(cl_page_header_print);
1156 * Prints human readable representation of \a cl_page to the \a f.
1158 void cl_page_print(const struct lu_env *env, void *cookie,
1159 lu_printer_t printer, const struct cl_page *cl_page)
1161 const struct cl_page_slice *slice;
1165 cl_page_header_print(env, cookie, printer, cl_page);
1166 cl_page_slice_for_each(cl_page, slice, i) {
1167 if (slice->cpl_ops->cpo_print != NULL)
1168 result = (*slice->cpl_ops->cpo_print)(env, slice,
1173 (*printer)(env, cookie, "end page@%p\n", cl_page);
1175 EXPORT_SYMBOL(cl_page_print);
1178 * Converts a byte offset within object \a obj into a page index.
1180 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1182 return (loff_t)idx << PAGE_SHIFT;
1184 EXPORT_SYMBOL(cl_offset);
1187 * Converts a page index into a byte offset within object \a obj.
1189 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1191 return offset >> PAGE_SHIFT;
1193 EXPORT_SYMBOL(cl_index);
1195 size_t cl_page_size(const struct cl_object *obj)
1197 return 1UL << PAGE_SHIFT;
1199 EXPORT_SYMBOL(cl_page_size);
1202 * Adds page slice to the compound page.
1204 * This is called by cl_object_operations::coo_page_init() methods to add a
1205 * per-layer state to the page. New state is added at the end of
1206 * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1208 * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1210 void cl_page_slice_add(struct cl_page *cl_page, struct cl_page_slice *slice,
1211 struct cl_object *obj,
1212 const struct cl_page_operations *ops)
1214 unsigned int offset = (char *)slice -
1215 ((char *)cl_page + sizeof(*cl_page));
1218 LASSERT(cl_page->cp_layer_count < CP_MAX_LAYER);
1219 LASSERT(offset < (1 << sizeof(cl_page->cp_layer_offset[0]) * 8));
1220 cl_page->cp_layer_offset[cl_page->cp_layer_count++] = offset;
1221 slice->cpl_obj = obj;
1222 slice->cpl_ops = ops;
1223 slice->cpl_page = cl_page;
1227 EXPORT_SYMBOL(cl_page_slice_add);
1230 * Allocate and initialize cl_cache, called by ll_init_sbi().
1232 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max)
1234 struct cl_client_cache *cache = NULL;
1237 OBD_ALLOC(cache, sizeof(*cache));
1241 /* Initialize cache data */
1242 atomic_set(&cache->ccc_users, 1);
1243 cache->ccc_lru_max = lru_page_max;
1244 atomic_long_set(&cache->ccc_lru_left, lru_page_max);
1245 spin_lock_init(&cache->ccc_lru_lock);
1246 INIT_LIST_HEAD(&cache->ccc_lru);
1248 /* turn unstable check off by default as it impacts performance */
1249 cache->ccc_unstable_check = 0;
1250 atomic_long_set(&cache->ccc_unstable_nr, 0);
1251 init_waitqueue_head(&cache->ccc_unstable_waitq);
1252 mutex_init(&cache->ccc_max_cache_mb_lock);
1256 EXPORT_SYMBOL(cl_cache_init);
1259 * Increase cl_cache refcount
1261 void cl_cache_incref(struct cl_client_cache *cache)
1263 atomic_inc(&cache->ccc_users);
1265 EXPORT_SYMBOL(cl_cache_incref);
1268 * Decrease cl_cache refcount and free the cache if refcount=0.
1269 * Since llite, lov and osc all hold cl_cache refcount,
1270 * the free will not cause race. (LU-6173)
1272 void cl_cache_decref(struct cl_client_cache *cache)
1274 if (atomic_dec_and_test(&cache->ccc_users))
1275 OBD_FREE(cache, sizeof(*cache));
1277 EXPORT_SYMBOL(cl_cache_decref);