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/
33 * Author: Nikita Danilov <nikita.danilov@sun.com>
34 * Author: Jinshan Xiong <jinshan.xiong@intel.com>
37 #define DEBUG_SUBSYSTEM S_CLASS
39 #include <linux/list.h>
40 #include <libcfs/libcfs.h>
41 #include <obd_class.h>
42 #include <obd_support.h>
44 #include <cl_object.h>
45 #include "cl_internal.h"
47 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg);
48 static DEFINE_MUTEX(cl_page_kmem_mutex);
51 # define PASSERT(env, page, expr) \
53 if (unlikely(!(expr))) { \
54 CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n"); \
58 #else /* !LIBCFS_DEBUG */
59 # define PASSERT(env, page, exp) \
60 ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
61 #endif /* !LIBCFS_DEBUG */
63 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
64 # define PINVRNT(env, page, expr) \
66 if (unlikely(!(expr))) { \
67 CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n"); \
71 #else /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
72 # define PINVRNT(env, page, exp) \
73 ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
74 #endif /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
76 /* Disable page statistic by default due to huge performance penalty. */
77 static void cs_page_inc(const struct cl_object *obj,
78 enum cache_stats_item item)
80 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
81 atomic_inc(&cl_object_site(obj)->cs_pages.cs_stats[item]);
85 static void cs_page_dec(const struct cl_object *obj,
86 enum cache_stats_item item)
88 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
89 atomic_dec(&cl_object_site(obj)->cs_pages.cs_stats[item]);
93 static void cs_pagestate_inc(const struct cl_object *obj,
94 enum cl_page_state state)
96 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
97 atomic_inc(&cl_object_site(obj)->cs_pages_state[state]);
101 static void cs_pagestate_dec(const struct cl_object *obj,
102 enum cl_page_state state)
104 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
105 atomic_dec(&cl_object_site(obj)->cs_pages_state[state]);
110 * Internal version of cl_page_get().
112 * This function can be used to obtain initial reference to previously
113 * unreferenced cached object. It can be called only if concurrent page
114 * reclamation is somehow prevented, e.g., by keeping a lock on a VM page,
115 * associated with \a page.
117 * Use with care! Not exported.
119 static void cl_page_get_trust(struct cl_page *page)
121 LASSERT(atomic_read(&page->cp_ref) > 0);
122 atomic_inc(&page->cp_ref);
125 static struct cl_page_slice *
126 cl_page_slice_get(const struct cl_page *cl_page, int index)
128 if (index < 0 || index >= cl_page->cp_layer_count)
131 /* To get the cp_layer_offset values fit under 256 bytes, we
132 * use the offset beyond the end of struct cl_page.
134 return (struct cl_page_slice *)((char *)cl_page + sizeof(*cl_page) +
135 cl_page->cp_layer_offset[index]);
138 #define cl_page_slice_for_each(cl_page, slice, i) \
139 for (i = 0, slice = cl_page_slice_get(cl_page, 0); \
140 i < (cl_page)->cp_layer_count; \
141 slice = cl_page_slice_get(cl_page, ++i))
143 #define cl_page_slice_for_each_reverse(cl_page, slice, i) \
144 for (i = (cl_page)->cp_layer_count - 1, \
145 slice = cl_page_slice_get(cl_page, i); i >= 0; \
146 slice = cl_page_slice_get(cl_page, --i))
148 static void __cl_page_free(struct cl_page *cl_page, unsigned short bufsize)
150 int index = cl_page->cp_kmem_index;
153 LASSERT(index < ARRAY_SIZE(cl_page_kmem_array));
154 LASSERT(cl_page_kmem_size_array[index] == bufsize);
155 OBD_SLAB_FREE(cl_page, cl_page_kmem_array[index], bufsize);
157 OBD_FREE(cl_page, bufsize);
161 static void cl_page_free(const struct lu_env *env, struct cl_page *cp,
162 struct pagevec *pvec)
164 struct cl_object *obj = cp->cp_obj;
165 unsigned short bufsize = cl_object_header(obj)->coh_page_bufsize;
169 PASSERT(env, cp, list_empty(&cp->cp_batch));
170 PASSERT(env, cp, cp->cp_owner == NULL);
171 PASSERT(env, cp, cp->cp_state == CPS_FREEING);
173 if (cp->cp_type == CPT_CACHEABLE) {
174 /* vmpage->private was already cleared when page was
175 * moved into CPS_FREEING state. */
176 vmpage = cp->cp_vmpage;
177 LASSERT(vmpage != NULL);
178 LASSERT((struct cl_page *)vmpage->private != cp);
181 if (!pagevec_add(pvec, vmpage))
182 pagevec_release(pvec);
188 cp->cp_layer_count = 0;
189 cs_page_dec(obj, CS_total);
190 cs_pagestate_dec(obj, cp->cp_state);
191 lu_object_ref_del_at(&obj->co_lu, &cp->cp_obj_ref, "cl_page", cp);
192 if (cp->cp_type != CPT_TRANSIENT)
193 cl_object_put(env, obj);
194 lu_ref_fini(&cp->cp_reference);
195 __cl_page_free(cp, bufsize);
199 static struct cl_page *__cl_page_alloc(struct cl_object *o)
202 struct cl_page *cl_page = NULL;
203 unsigned short bufsize = cl_object_header(o)->coh_page_bufsize;
205 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_PAGE_ALLOC))
209 /* the number of entries in cl_page_kmem_array is expected to
210 * only be 2-3 entries, so the lookup overhead should be low.
212 for ( ; i < ARRAY_SIZE(cl_page_kmem_array); i++) {
213 if (smp_load_acquire(&cl_page_kmem_size_array[i])
215 OBD_SLAB_ALLOC_GFP(cl_page, cl_page_kmem_array[i],
218 cl_page->cp_kmem_index = i;
221 if (cl_page_kmem_size_array[i] == 0)
225 if (i < ARRAY_SIZE(cl_page_kmem_array)) {
228 mutex_lock(&cl_page_kmem_mutex);
229 if (cl_page_kmem_size_array[i]) {
230 mutex_unlock(&cl_page_kmem_mutex);
233 snprintf(cache_name, sizeof(cache_name),
234 "cl_page_kmem-%u", bufsize);
235 cl_page_kmem_array[i] =
236 kmem_cache_create(cache_name, bufsize,
238 if (cl_page_kmem_array[i] == NULL) {
239 mutex_unlock(&cl_page_kmem_mutex);
242 smp_store_release(&cl_page_kmem_size_array[i],
244 mutex_unlock(&cl_page_kmem_mutex);
247 OBD_ALLOC_GFP(cl_page, bufsize, GFP_NOFS);
249 cl_page->cp_kmem_index = -1;
255 struct cl_page *cl_page_alloc(const struct lu_env *env, struct cl_object *o,
256 pgoff_t ind, struct page *vmpage,
257 enum cl_page_type type)
259 struct cl_page *cl_page;
260 struct cl_object *head;
264 cl_page = __cl_page_alloc(o);
265 if (cl_page != NULL) {
269 * Please fix cl_page:cp_state/type declaration if
270 * these assertions fail in the future.
272 BUILD_BUG_ON((1 << CP_STATE_BITS) < CPS_NR); /* cp_state */
273 BUILD_BUG_ON((1 << CP_TYPE_BITS) < CPT_NR); /* cp_type */
274 atomic_set(&cl_page->cp_ref, 1);
276 if (type != CPT_TRANSIENT)
278 lu_object_ref_add_at(&o->co_lu, &cl_page->cp_obj_ref,
280 cl_page->cp_vmpage = vmpage;
281 cl_page->cp_state = CPS_CACHED;
282 cl_page->cp_type = type;
283 if (type == CPT_TRANSIENT)
284 /* ref to correct inode will be added
285 * in ll_direct_rw_pages
287 cl_page->cp_inode = NULL;
289 cl_page->cp_inode = page2inode(vmpage);
290 INIT_LIST_HEAD(&cl_page->cp_batch);
291 lu_ref_init(&cl_page->cp_reference);
293 cl_page->cp_page_index = ind;
294 cl_object_for_each(o, head) {
295 if (o->co_ops->coo_page_init != NULL) {
296 result = o->co_ops->coo_page_init(env, o,
299 cl_page_delete0(env, cl_page);
300 cl_page_free(env, cl_page, NULL);
301 cl_page = ERR_PTR(result);
307 cs_page_inc(o, CS_total);
308 cs_page_inc(o, CS_create);
309 cs_pagestate_dec(o, CPS_CACHED);
312 cl_page = ERR_PTR(-ENOMEM);
318 * Returns a cl_page with index \a idx at the object \a o, and associated with
319 * the VM page \a vmpage.
321 * This is the main entry point into the cl_page caching interface. First, a
322 * cache (implemented as a per-object radix tree) is consulted. If page is
323 * found there, it is returned immediately. Otherwise new page is allocated
324 * and returned. In any case, additional reference to page is acquired.
326 * \see cl_object_find(), cl_lock_find()
328 struct cl_page *cl_page_find(const struct lu_env *env,
330 pgoff_t idx, struct page *vmpage,
331 enum cl_page_type type)
333 struct cl_page *page = NULL;
334 struct cl_object_header *hdr;
336 LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
341 hdr = cl_object_header(o);
342 cs_page_inc(o, CS_lookup);
344 CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
345 idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
347 if (type == CPT_CACHEABLE) {
348 /* vmpage lock is used to protect the child/parent
350 LASSERT(PageLocked(vmpage));
352 * cl_vmpage_page() can be called here without any locks as
354 * - "vmpage" is locked (which prevents ->private from
355 * concurrent updates), and
357 * - "o" cannot be destroyed while current thread holds a
360 page = cl_vmpage_page(vmpage, o);
362 cs_page_inc(o, CS_hit);
367 /* allocate and initialize cl_page */
368 page = cl_page_alloc(env, o, idx, vmpage, type);
371 EXPORT_SYMBOL(cl_page_find);
373 static inline int cl_page_invariant(const struct cl_page *pg)
375 return cl_page_in_use_noref(pg);
378 static void cl_page_state_set0(const struct lu_env *env,
379 struct cl_page *cl_page,
380 enum cl_page_state state)
382 enum cl_page_state old;
385 * Matrix of allowed state transitions [old][new], for sanity
388 static const int allowed_transitions[CPS_NR][CPS_NR] = {
391 [CPS_OWNED] = 1, /* io finds existing cached page */
393 [CPS_PAGEOUT] = 1, /* write-out from the cache */
394 [CPS_FREEING] = 1, /* eviction on the memory pressure */
397 [CPS_CACHED] = 1, /* release to the cache */
399 [CPS_PAGEIN] = 1, /* start read immediately */
400 [CPS_PAGEOUT] = 1, /* start write immediately */
401 [CPS_FREEING] = 1, /* lock invalidation or truncate */
404 [CPS_CACHED] = 1, /* io completion */
411 [CPS_CACHED] = 1, /* io completion */
427 old = cl_page->cp_state;
428 PASSERT(env, cl_page, allowed_transitions[old][state]);
429 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d -> %d\n", old, state);
430 PASSERT(env, cl_page, cl_page->cp_state == old);
431 PASSERT(env, cl_page, equi(state == CPS_OWNED,
432 cl_page->cp_owner != NULL));
434 cs_pagestate_dec(cl_page->cp_obj, cl_page->cp_state);
435 cs_pagestate_inc(cl_page->cp_obj, state);
436 cl_page->cp_state = state;
440 static void cl_page_state_set(const struct lu_env *env,
441 struct cl_page *page, enum cl_page_state state)
443 cl_page_state_set0(env, page, state);
447 * Acquires an additional reference to a page.
449 * This can be called only by caller already possessing a reference to \a
452 * \see cl_object_get(), cl_lock_get().
454 void cl_page_get(struct cl_page *page)
457 cl_page_get_trust(page);
460 EXPORT_SYMBOL(cl_page_get);
463 * Releases a reference to a page, use the pagevec to release the pages
464 * in batch if provided.
466 * Users need to do a final pagevec_release() to release any trailing pages.
468 void cl_pagevec_put(const struct lu_env *env, struct cl_page *page,
469 struct pagevec *pvec)
472 CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
473 atomic_read(&page->cp_ref));
475 if (atomic_dec_and_test(&page->cp_ref)) {
476 LASSERT(page->cp_state == CPS_FREEING);
478 LASSERT(atomic_read(&page->cp_ref) == 0);
479 PASSERT(env, page, page->cp_owner == NULL);
480 PASSERT(env, page, list_empty(&page->cp_batch));
482 * Page is no longer reachable by other threads. Tear
485 cl_page_free(env, page, pvec);
490 EXPORT_SYMBOL(cl_pagevec_put);
493 * Releases a reference to a page, wrapper to cl_pagevec_put
495 * When last reference is released, page is returned to the cache, unless it
496 * is in cl_page_state::CPS_FREEING state, in which case it is immediately
499 * \see cl_object_put(), cl_lock_put().
501 void cl_page_put(const struct lu_env *env, struct cl_page *page)
503 cl_pagevec_put(env, page, NULL);
505 EXPORT_SYMBOL(cl_page_put);
508 * Returns a cl_page associated with a VM page, and given cl_object.
510 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj)
512 struct cl_page *page;
515 LASSERT(PageLocked(vmpage));
518 * NOTE: absence of races and liveness of data are guaranteed by page
519 * lock on a "vmpage". That works because object destruction has
520 * bottom-to-top pass.
523 page = (struct cl_page *)vmpage->private;
525 cl_page_get_trust(page);
526 LASSERT(page->cp_type == CPT_CACHEABLE);
530 EXPORT_SYMBOL(cl_vmpage_page);
532 static void cl_page_owner_clear(struct cl_page *page)
535 if (page->cp_owner != NULL) {
536 LASSERT(page->cp_owner->ci_owned_nr > 0);
537 page->cp_owner->ci_owned_nr--;
538 page->cp_owner = NULL;
543 static void cl_page_owner_set(struct cl_page *page)
546 LASSERT(page->cp_owner != NULL);
547 page->cp_owner->ci_owned_nr++;
551 void cl_page_disown0(const struct lu_env *env, struct cl_page *cp)
554 enum cl_page_state state;
557 state = cp->cp_state;
558 PINVRNT(env, cp, state == CPS_OWNED || state == CPS_FREEING);
559 PINVRNT(env, cp, cl_page_invariant(cp) || state == CPS_FREEING);
560 cl_page_owner_clear(cp);
562 if (state == CPS_OWNED)
563 cl_page_state_set(env, cp, CPS_CACHED);
565 if (cp->cp_type == CPT_CACHEABLE) {
566 vmpage = cp->cp_vmpage;
567 LASSERT(vmpage != NULL);
568 LASSERT(PageLocked(vmpage));
576 * returns true, iff page is owned by the given io.
578 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
580 struct cl_io *top = cl_io_top((struct cl_io *)io);
581 LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
583 RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top);
585 EXPORT_SYMBOL(cl_page_is_owned);
588 * Try to own a page by IO.
590 * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
591 * into cl_page_state::CPS_OWNED state.
593 * \pre !cl_page_is_owned(cl_page, io)
594 * \post result == 0 iff cl_page_is_owned(cl_page, io)
598 * \retval -ve failure, e.g., cl_page was destroyed (and landed in
599 * cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
600 * or, page was owned by another thread, or in IO.
602 * \see cl_page_disown()
603 * \see cl_page_own_try()
606 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
607 struct cl_page *cl_page, int nonblock)
609 struct page *vmpage = cl_page->cp_vmpage;
613 PINVRNT(env, cl_page, !cl_page_is_owned(cl_page, io));
615 if (cl_page->cp_state == CPS_FREEING) {
620 LASSERT(vmpage != NULL);
622 if (cl_page->cp_type == CPT_TRANSIENT) {
624 } else if (nonblock) {
625 if (!trylock_page(vmpage)) {
630 if (unlikely(PageWriteback(vmpage))) {
637 wait_on_page_writeback(vmpage);
640 PASSERT(env, cl_page, cl_page->cp_owner == NULL);
641 cl_page->cp_owner = cl_io_top(io);
642 cl_page_owner_set(cl_page);
644 if (cl_page->cp_state == CPS_FREEING) {
645 cl_page_disown0(env, cl_page);
650 cl_page_state_set(env, cl_page, CPS_OWNED);
653 PINVRNT(env, cl_page, ergo(result == 0,
654 cl_page_invariant(cl_page)));
659 * Own a page, might be blocked.
661 * \see cl_page_own0()
663 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
665 return cl_page_own0(env, io, pg, 0);
667 EXPORT_SYMBOL(cl_page_own);
670 * Nonblock version of cl_page_own().
672 * \see cl_page_own0()
674 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
677 return cl_page_own0(env, io, pg, 1);
679 EXPORT_SYMBOL(cl_page_own_try);
683 * Assume page ownership.
685 * Called when page is already locked by the hosting VM.
687 * \pre !cl_page_is_owned(cp, io)
688 * \post cl_page_is_owned(cp, io)
690 void cl_page_assume(const struct lu_env *env,
691 struct cl_io *io, struct cl_page *cp)
696 PINVRNT(env, cp, cl_object_same(cp->cp_obj, io->ci_obj));
698 if (cp->cp_type == CPT_CACHEABLE) {
699 vmpage = cp->cp_vmpage;
700 LASSERT(vmpage != NULL);
701 LASSERT(PageLocked(vmpage));
702 wait_on_page_writeback(vmpage);
705 PASSERT(env, cp, cp->cp_owner == NULL);
706 cp->cp_owner = cl_io_top(io);
707 cl_page_owner_set(cp);
708 cl_page_state_set(env, cp, CPS_OWNED);
711 EXPORT_SYMBOL(cl_page_assume);
714 * Releases page ownership without unlocking the page.
716 * Moves cl_page into cl_page_state::CPS_CACHED without releasing a lock
717 * on the underlying VM page (as VM is supposed to do this itself).
719 * \pre cl_page_is_owned(cp, io)
720 * \post !cl_page_is_owned(cp, io)
722 void cl_page_unassume(const struct lu_env *env,
723 struct cl_io *io, struct cl_page *cp)
728 PINVRNT(env, cp, cl_page_is_owned(cp, io));
729 PINVRNT(env, cp, cl_page_invariant(cp));
731 cl_page_owner_clear(cp);
732 cl_page_state_set(env, cp, CPS_CACHED);
734 if (cp->cp_type == CPT_CACHEABLE) {
735 vmpage = cp->cp_vmpage;
736 LASSERT(vmpage != NULL);
737 LASSERT(PageLocked(vmpage));
742 EXPORT_SYMBOL(cl_page_unassume);
745 * Releases page ownership.
747 * Moves page into cl_page_state::CPS_CACHED.
749 * \pre cl_page_is_owned(pg, io)
750 * \post !cl_page_is_owned(pg, io)
754 void cl_page_disown(const struct lu_env *env,
755 struct cl_io *io, struct cl_page *pg)
757 PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
758 pg->cp_state == CPS_FREEING);
760 cl_page_disown0(env, pg);
762 EXPORT_SYMBOL(cl_page_disown);
765 * Called when cl_page is to be removed from the object, e.g.,
766 * as a result of truncate.
768 * Calls cl_page_operations::cpo_discard() top-to-bottom.
770 * \pre cl_page_is_owned(cl_page, io)
772 * \see cl_page_operations::cpo_discard()
774 void cl_page_discard(const struct lu_env *env,
775 struct cl_io *io, struct cl_page *cp)
778 const struct cl_page_slice *slice;
781 PINVRNT(env, cp, cl_page_is_owned(cp, io));
782 PINVRNT(env, cp, cl_page_invariant(cp));
784 cl_page_slice_for_each(cp, slice, i) {
785 if (slice->cpl_ops->cpo_discard != NULL)
786 (*slice->cpl_ops->cpo_discard)(env, slice, io);
789 if (cp->cp_type == CPT_CACHEABLE) {
790 vmpage = cp->cp_vmpage;
791 LASSERT(vmpage != NULL);
792 LASSERT(PageLocked(vmpage));
793 generic_error_remove_page(vmpage->mapping, vmpage);
795 cl_page_delete(env, cp);
798 EXPORT_SYMBOL(cl_page_discard);
801 * Version of cl_page_delete() that can be called for not fully constructed
802 * cl_pages, e.g. in an error handling cl_page_find()->cl_page_delete0()
803 * path. Doesn't check cl_page invariant.
805 static void cl_page_delete0(const struct lu_env *env, struct cl_page *cp)
808 const struct cl_page_slice *slice;
813 PASSERT(env, cp, cp->cp_state != CPS_FREEING);
816 * Severe all ways to obtain new pointers to @pg.
818 cl_page_owner_clear(cp);
819 cl_page_state_set0(env, cp, CPS_FREEING);
821 cl_page_slice_for_each_reverse(cp, slice, i) {
822 if (slice->cpl_ops->cpo_delete != NULL)
823 (*slice->cpl_ops->cpo_delete)(env, slice);
826 if (cp->cp_type == CPT_CACHEABLE) {
827 vmpage = cp->cp_vmpage;
828 LASSERT(PageLocked(vmpage));
829 LASSERT((struct cl_page *)vmpage->private == cp);
831 /* Drop the reference count held in vvp_page_init */
832 refc = atomic_dec_return(&cp->cp_ref);
833 LASSERTF(refc >= 1, "page = %p, refc = %d\n", cp, refc);
835 ClearPagePrivate(vmpage);
839 * The reference from vmpage to cl_page is removed,
840 * but the reference back is still here. It is removed
841 * later in cl_page_free().
849 * Called when a decision is made to throw page out of memory.
851 * Notifies all layers about page destruction by calling
852 * cl_page_operations::cpo_delete() method top-to-bottom.
854 * Moves page into cl_page_state::CPS_FREEING state (this is the only place
855 * where transition to this state happens).
857 * Eliminates all venues through which new references to the page can be
860 * - removes page from the radix trees,
862 * - breaks linkage from VM page to cl_page.
864 * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
865 * drain after some time, at which point page will be recycled.
867 * \pre VM page is locked
868 * \post pg->cp_state == CPS_FREEING
870 * \see cl_page_operations::cpo_delete()
872 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
874 PINVRNT(env, pg, cl_page_invariant(pg));
876 cl_page_delete0(env, pg);
879 EXPORT_SYMBOL(cl_page_delete);
881 void cl_page_touch(const struct lu_env *env,
882 const struct cl_page *cl_page, size_t to)
884 const struct cl_page_slice *slice;
889 cl_page_slice_for_each(cl_page, slice, i) {
890 if (slice->cpl_ops->cpo_page_touch != NULL)
891 (*slice->cpl_ops->cpo_page_touch)(env, slice, to);
896 EXPORT_SYMBOL(cl_page_touch);
898 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
901 RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
904 static void cl_page_io_start(const struct lu_env *env,
905 struct cl_page *pg, enum cl_req_type crt)
908 * Page is queued for IO, change its state.
911 cl_page_owner_clear(pg);
912 cl_page_state_set(env, pg, cl_req_type_state(crt));
917 * Prepares page for immediate transfer. Return -EALREADY if this page
918 * should be omitted from transfer.
920 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
921 struct cl_page *cp, enum cl_req_type crt)
923 struct page *vmpage = cp->cp_vmpage;
926 PASSERT(env, cp, crt < CRT_NR);
927 PINVRNT(env, cp, cl_page_is_owned(cp, io));
928 PINVRNT(env, cp, cl_page_invariant(cp));
930 if (cp->cp_type == CPT_TRANSIENT) {
932 } else if (crt == CRT_READ) {
933 if (PageUptodate(vmpage))
934 GOTO(out, rc = -EALREADY);
936 LASSERT(PageLocked(vmpage));
937 LASSERT(!PageDirty(vmpage));
939 /* ll_writepage path is not a sync write, so need to
940 * set page writeback flag
942 if (cp->cp_sync_io == NULL)
943 set_page_writeback(vmpage);
946 cl_page_io_start(env, cp, crt);
949 CL_PAGE_HEADER(D_TRACE, env, cp, "%d %d\n", crt, rc);
953 EXPORT_SYMBOL(cl_page_prep);
956 * Notify layers about transfer completion.
958 * Invoked by transfer sub-system (which is a part of osc) to notify layers
959 * that a transfer, of which this page is a part of has completed.
961 * Completion call-backs are executed in the bottom-up order, so that
962 * uppermost layer (llite), responsible for the VFS/VM interaction runs last
963 * and can release locks safely.
965 * \pre cl_page->cp_state == CPS_PAGEIN || cl_page->cp_state == CPS_PAGEOUT
966 * \post cl_page->cl_page_state == CPS_CACHED
968 * \see cl_page_operations::cpo_completion()
970 void cl_page_completion(const struct lu_env *env,
971 struct cl_page *cl_page, enum cl_req_type crt,
974 const struct cl_page_slice *slice;
975 struct cl_sync_io *anchor = cl_page->cp_sync_io;
979 PASSERT(env, cl_page, crt < CRT_NR);
980 PASSERT(env, cl_page, cl_page->cp_state == cl_req_type_state(crt));
982 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", crt, ioret);
983 cl_page_state_set(env, cl_page, CPS_CACHED);
987 cl_page_slice_for_each_reverse(cl_page, slice, i) {
988 if (slice->cpl_ops->io[crt].cpo_completion != NULL)
989 (*slice->cpl_ops->io[crt].cpo_completion)(env, slice,
993 if (anchor != NULL) {
994 LASSERT(cl_page->cp_sync_io == anchor);
995 cl_page->cp_sync_io = NULL;
996 cl_sync_io_note(env, anchor, ioret);
1000 EXPORT_SYMBOL(cl_page_completion);
1003 * Notify layers that transfer formation engine decided to yank this page from
1004 * the cache and to make it a part of a transfer.
1006 * \pre cl_page->cp_state == CPS_CACHED
1007 * \post cl_page->cp_state == CPS_PAGEIN || cl_page->cp_state == CPS_PAGEOUT
1009 int cl_page_make_ready(const struct lu_env *env, struct cl_page *cp,
1010 enum cl_req_type crt)
1012 struct page *vmpage = cp->cp_vmpage;
1016 PASSERT(env, cp, crt == CRT_WRITE);
1018 if (cp->cp_type == CPT_TRANSIENT)
1023 if (clear_page_dirty_for_io(vmpage)) {
1024 LASSERT(cp->cp_state == CPS_CACHED);
1025 /* This actually clears the dirty bit in the
1028 set_page_writeback(vmpage);
1029 CL_PAGE_HEADER(D_PAGE, env, cp, "readied\n");
1031 } else if (cp->cp_state == CPS_PAGEOUT) {
1032 /* is it possible for osc_flush_async_page()
1033 * to already make it ready?
1037 CL_PAGE_DEBUG(D_ERROR, env, cp,
1038 "unexpecting page state %d\n",
1043 unlock_page(vmpage);
1046 PASSERT(env, cp, cp->cp_state == CPS_CACHED);
1047 cl_page_io_start(env, cp, crt);
1050 CL_PAGE_HEADER(D_TRACE, env, cp, "%d %d\n", crt, rc);
1054 EXPORT_SYMBOL(cl_page_make_ready);
1057 * Called if a page is being written back by kernel's intention.
1059 * \pre cl_page_is_owned(cl_page, io)
1060 * \post ergo(result == 0, cl_page->cp_state == CPS_PAGEOUT)
1062 * \see cl_page_operations::cpo_flush()
1064 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
1065 struct cl_page *cl_page)
1067 const struct cl_page_slice *slice;
1072 PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
1073 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
1075 cl_page_slice_for_each(cl_page, slice, i) {
1076 if (slice->cpl_ops->cpo_flush != NULL)
1077 result = (*slice->cpl_ops->cpo_flush)(env, slice, io);
1084 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d\n", result);
1087 EXPORT_SYMBOL(cl_page_flush);
1090 * Tells transfer engine that only part of a page is to be transmitted.
1092 * \see cl_page_operations::cpo_clip()
1094 void cl_page_clip(const struct lu_env *env, struct cl_page *cl_page,
1097 const struct cl_page_slice *slice;
1100 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
1102 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", from, to);
1103 cl_page_slice_for_each(cl_page, slice, i) {
1104 if (slice->cpl_ops->cpo_clip != NULL)
1105 (*slice->cpl_ops->cpo_clip)(env, slice, from, to);
1108 EXPORT_SYMBOL(cl_page_clip);
1111 * Prints human readable representation of \a pg to the \a f.
1113 void cl_page_header_print(const struct lu_env *env, void *cookie,
1114 lu_printer_t printer, const struct cl_page *pg)
1116 (*printer)(env, cookie,
1117 "page@%p[%d %p %d %d %p]\n",
1118 pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1119 pg->cp_state, pg->cp_type,
1122 EXPORT_SYMBOL(cl_page_header_print);
1125 * Prints human readable representation of \a cl_page to the \a f.
1127 void cl_page_print(const struct lu_env *env, void *cookie,
1128 lu_printer_t printer, const struct cl_page *cp)
1130 struct page *vmpage = cp->cp_vmpage;
1131 const struct cl_page_slice *slice;
1135 cl_page_header_print(env, cookie, printer, cp);
1137 (*printer)(env, cookie, "vmpage @%p", vmpage);
1139 if (vmpage != NULL) {
1140 (*printer)(env, cookie, " %lx %d:%d %lx %lu %slru",
1141 (long)vmpage->flags, page_count(vmpage),
1142 page_mapcount(vmpage), vmpage->private,
1144 list_empty(&vmpage->lru) ? "not-" : "");
1147 (*printer)(env, cookie, "\n");
1149 cl_page_slice_for_each(cp, slice, i) {
1150 if (slice->cpl_ops->cpo_print != NULL)
1151 result = (*slice->cpl_ops->cpo_print)(env, slice,
1157 (*printer)(env, cookie, "end page@%p\n", cp);
1159 EXPORT_SYMBOL(cl_page_print);
1162 * Converts a byte offset within object \a obj into a page index.
1164 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1166 return (loff_t)idx << PAGE_SHIFT;
1168 EXPORT_SYMBOL(cl_offset);
1171 * Converts a page index into a byte offset within object \a obj.
1173 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1175 return offset >> PAGE_SHIFT;
1177 EXPORT_SYMBOL(cl_index);
1179 size_t cl_page_size(const struct cl_object *obj)
1181 return 1UL << PAGE_SHIFT;
1183 EXPORT_SYMBOL(cl_page_size);
1186 * Adds page slice to the compound page.
1188 * This is called by cl_object_operations::coo_page_init() methods to add a
1189 * per-layer state to the page. New state is added at the end of
1190 * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1192 * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1194 void cl_page_slice_add(struct cl_page *cl_page, struct cl_page_slice *slice,
1195 struct cl_object *obj,
1196 const struct cl_page_operations *ops)
1198 unsigned int offset = (char *)slice -
1199 ((char *)cl_page + sizeof(*cl_page));
1202 LASSERT(cl_page->cp_layer_count < CP_MAX_LAYER);
1203 LASSERT(offset < (1 << sizeof(cl_page->cp_layer_offset[0]) * 8));
1204 cl_page->cp_layer_offset[cl_page->cp_layer_count++] = offset;
1205 slice->cpl_ops = ops;
1206 slice->cpl_page = cl_page;
1210 EXPORT_SYMBOL(cl_page_slice_add);
1213 * Allocate and initialize cl_cache, called by ll_init_sbi().
1215 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max)
1217 struct cl_client_cache *cache = NULL;
1220 OBD_ALLOC(cache, sizeof(*cache));
1224 /* Initialize cache data */
1225 atomic_set(&cache->ccc_users, 1);
1226 cache->ccc_lru_max = lru_page_max;
1227 atomic_long_set(&cache->ccc_lru_left, lru_page_max);
1228 spin_lock_init(&cache->ccc_lru_lock);
1229 INIT_LIST_HEAD(&cache->ccc_lru);
1231 /* turn unstable check off by default as it impacts performance */
1232 cache->ccc_unstable_check = 0;
1233 atomic_long_set(&cache->ccc_unstable_nr, 0);
1234 init_waitqueue_head(&cache->ccc_unstable_waitq);
1235 mutex_init(&cache->ccc_max_cache_mb_lock);
1239 EXPORT_SYMBOL(cl_cache_init);
1242 * Increase cl_cache refcount
1244 void cl_cache_incref(struct cl_client_cache *cache)
1246 atomic_inc(&cache->ccc_users);
1248 EXPORT_SYMBOL(cl_cache_incref);
1251 * Decrease cl_cache refcount and free the cache if refcount=0.
1252 * Since llite, lov and osc all hold cl_cache refcount,
1253 * the free will not cause race. (LU-6173)
1255 void cl_cache_decref(struct cl_client_cache *cache)
1257 if (atomic_dec_and_test(&cache->ccc_users))
1258 OBD_FREE(cache, sizeof(*cache));
1260 EXPORT_SYMBOL(cl_cache_decref);