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))
149 * Returns a slice within a cl_page, corresponding to the given layer in the
154 static const struct cl_page_slice *
155 cl_page_at_trusted(const struct cl_page *cl_page,
156 const struct lu_device_type *dtype)
158 const struct cl_page_slice *slice;
163 cl_page_slice_for_each(cl_page, slice, i) {
164 if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
171 static void __cl_page_free(struct cl_page *cl_page, unsigned short bufsize)
173 int index = cl_page->cp_kmem_index;
176 LASSERT(index < ARRAY_SIZE(cl_page_kmem_array));
177 LASSERT(cl_page_kmem_size_array[index] == bufsize);
178 OBD_SLAB_FREE(cl_page, cl_page_kmem_array[index], bufsize);
180 OBD_FREE(cl_page, bufsize);
184 static void cl_page_free(const struct lu_env *env, struct cl_page *cl_page,
185 struct pagevec *pvec)
187 struct cl_object *obj = cl_page->cp_obj;
188 unsigned short bufsize = cl_object_header(obj)->coh_page_bufsize;
189 struct cl_page_slice *slice;
193 PASSERT(env, cl_page, list_empty(&cl_page->cp_batch));
194 PASSERT(env, cl_page, cl_page->cp_owner == NULL);
195 PASSERT(env, cl_page, cl_page->cp_state == CPS_FREEING);
197 cl_page_slice_for_each(cl_page, slice, i) {
198 if (unlikely(slice->cpl_ops->cpo_fini != NULL))
199 slice->cpl_ops->cpo_fini(env, slice, pvec);
201 cl_page->cp_layer_count = 0;
202 cs_page_dec(obj, CS_total);
203 cs_pagestate_dec(obj, cl_page->cp_state);
204 lu_object_ref_del_at(&obj->co_lu, &cl_page->cp_obj_ref,
206 if (cl_page->cp_type != CPT_TRANSIENT)
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;
219 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_PAGE_ALLOC))
223 /* the number of entries in cl_page_kmem_array is expected to
224 * only be 2-3 entries, so the lookup overhead should be low.
226 for ( ; i < ARRAY_SIZE(cl_page_kmem_array); i++) {
227 if (smp_load_acquire(&cl_page_kmem_size_array[i])
229 OBD_SLAB_ALLOC_GFP(cl_page, cl_page_kmem_array[i],
232 cl_page->cp_kmem_index = i;
235 if (cl_page_kmem_size_array[i] == 0)
239 if (i < ARRAY_SIZE(cl_page_kmem_array)) {
242 mutex_lock(&cl_page_kmem_mutex);
243 if (cl_page_kmem_size_array[i]) {
244 mutex_unlock(&cl_page_kmem_mutex);
247 snprintf(cache_name, sizeof(cache_name),
248 "cl_page_kmem-%u", bufsize);
249 cl_page_kmem_array[i] =
250 kmem_cache_create(cache_name, bufsize,
252 if (cl_page_kmem_array[i] == NULL) {
253 mutex_unlock(&cl_page_kmem_mutex);
256 smp_store_release(&cl_page_kmem_size_array[i],
258 mutex_unlock(&cl_page_kmem_mutex);
261 OBD_ALLOC_GFP(cl_page, bufsize, GFP_NOFS);
263 cl_page->cp_kmem_index = -1;
269 struct cl_page *cl_page_alloc(const struct lu_env *env, struct cl_object *o,
270 pgoff_t ind, struct page *vmpage,
271 enum cl_page_type type)
273 struct cl_page *cl_page;
274 struct cl_object *head;
278 cl_page = __cl_page_alloc(o);
279 if (cl_page != NULL) {
283 * Please fix cl_page:cp_state/type declaration if
284 * these assertions fail in the future.
286 BUILD_BUG_ON((1 << CP_STATE_BITS) < CPS_NR); /* cp_state */
287 BUILD_BUG_ON((1 << CP_TYPE_BITS) < CPT_NR); /* cp_type */
288 atomic_set(&cl_page->cp_ref, 1);
290 if (type != CPT_TRANSIENT)
292 lu_object_ref_add_at(&o->co_lu, &cl_page->cp_obj_ref,
294 cl_page->cp_vmpage = vmpage;
295 cl_page->cp_state = CPS_CACHED;
296 cl_page->cp_type = type;
297 cl_page->cp_inode = NULL;
298 INIT_LIST_HEAD(&cl_page->cp_batch);
299 lu_ref_init(&cl_page->cp_reference);
301 cl_object_for_each(o, head) {
302 if (o->co_ops->coo_page_init != NULL) {
303 result = o->co_ops->coo_page_init(env, o,
306 cl_page_delete0(env, cl_page);
307 cl_page_free(env, cl_page, NULL);
308 cl_page = ERR_PTR(result);
314 cs_page_inc(o, CS_total);
315 cs_page_inc(o, CS_create);
316 cs_pagestate_dec(o, CPS_CACHED);
319 cl_page = ERR_PTR(-ENOMEM);
325 * Returns a cl_page with index \a idx at the object \a o, and associated with
326 * the VM page \a vmpage.
328 * This is the main entry point into the cl_page caching interface. First, a
329 * cache (implemented as a per-object radix tree) is consulted. If page is
330 * found there, it is returned immediately. Otherwise new page is allocated
331 * and returned. In any case, additional reference to page is acquired.
333 * \see cl_object_find(), cl_lock_find()
335 struct cl_page *cl_page_find(const struct lu_env *env,
337 pgoff_t idx, struct page *vmpage,
338 enum cl_page_type type)
340 struct cl_page *page = NULL;
341 struct cl_object_header *hdr;
343 LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
348 hdr = cl_object_header(o);
349 cs_page_inc(o, CS_lookup);
351 CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
352 idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
354 if (type == CPT_CACHEABLE) {
355 /* vmpage lock is used to protect the child/parent
357 LASSERT(PageLocked(vmpage));
359 * cl_vmpage_page() can be called here without any locks as
361 * - "vmpage" is locked (which prevents ->private from
362 * concurrent updates), and
364 * - "o" cannot be destroyed while current thread holds a
367 page = cl_vmpage_page(vmpage, o);
369 cs_page_inc(o, CS_hit);
374 /* allocate and initialize cl_page */
375 page = cl_page_alloc(env, o, idx, vmpage, type);
378 EXPORT_SYMBOL(cl_page_find);
380 static inline int cl_page_invariant(const struct cl_page *pg)
382 return cl_page_in_use_noref(pg);
385 static void cl_page_state_set0(const struct lu_env *env,
386 struct cl_page *cl_page,
387 enum cl_page_state state)
389 enum cl_page_state old;
392 * Matrix of allowed state transitions [old][new], for sanity
395 static const int allowed_transitions[CPS_NR][CPS_NR] = {
398 [CPS_OWNED] = 1, /* io finds existing cached page */
400 [CPS_PAGEOUT] = 1, /* write-out from the cache */
401 [CPS_FREEING] = 1, /* eviction on the memory pressure */
404 [CPS_CACHED] = 1, /* release to the cache */
406 [CPS_PAGEIN] = 1, /* start read immediately */
407 [CPS_PAGEOUT] = 1, /* start write immediately */
408 [CPS_FREEING] = 1, /* lock invalidation or truncate */
411 [CPS_CACHED] = 1, /* io completion */
418 [CPS_CACHED] = 1, /* io completion */
434 old = cl_page->cp_state;
435 PASSERT(env, cl_page, allowed_transitions[old][state]);
436 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d -> %d\n", old, state);
437 PASSERT(env, cl_page, cl_page->cp_state == old);
438 PASSERT(env, cl_page, equi(state == CPS_OWNED,
439 cl_page->cp_owner != NULL));
441 cs_pagestate_dec(cl_page->cp_obj, cl_page->cp_state);
442 cs_pagestate_inc(cl_page->cp_obj, state);
443 cl_page->cp_state = state;
447 static void cl_page_state_set(const struct lu_env *env,
448 struct cl_page *page, enum cl_page_state state)
450 cl_page_state_set0(env, page, state);
454 * Acquires an additional reference to a page.
456 * This can be called only by caller already possessing a reference to \a
459 * \see cl_object_get(), cl_lock_get().
461 void cl_page_get(struct cl_page *page)
464 cl_page_get_trust(page);
467 EXPORT_SYMBOL(cl_page_get);
470 * Releases a reference to a page, use the pagevec to release the pages
471 * in batch if provided.
473 * Users need to do a final pagevec_release() to release any trailing pages.
475 void cl_pagevec_put(const struct lu_env *env, struct cl_page *page,
476 struct pagevec *pvec)
479 CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
480 atomic_read(&page->cp_ref));
482 if (atomic_dec_and_test(&page->cp_ref)) {
483 LASSERT(page->cp_state == CPS_FREEING);
485 LASSERT(atomic_read(&page->cp_ref) == 0);
486 PASSERT(env, page, page->cp_owner == NULL);
487 PASSERT(env, page, list_empty(&page->cp_batch));
489 * Page is no longer reachable by other threads. Tear
492 cl_page_free(env, page, pvec);
497 EXPORT_SYMBOL(cl_pagevec_put);
500 * Releases a reference to a page, wrapper to cl_pagevec_put
502 * When last reference is released, page is returned to the cache, unless it
503 * is in cl_page_state::CPS_FREEING state, in which case it is immediately
506 * \see cl_object_put(), cl_lock_put().
508 void cl_page_put(const struct lu_env *env, struct cl_page *page)
510 cl_pagevec_put(env, page, NULL);
512 EXPORT_SYMBOL(cl_page_put);
515 * Returns a cl_page associated with a VM page, and given cl_object.
517 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj)
519 struct cl_page *page;
522 LASSERT(PageLocked(vmpage));
525 * NOTE: absence of races and liveness of data are guaranteed by page
526 * lock on a "vmpage". That works because object destruction has
527 * bottom-to-top pass.
530 page = (struct cl_page *)vmpage->private;
532 cl_page_get_trust(page);
533 LASSERT(page->cp_type == CPT_CACHEABLE);
537 EXPORT_SYMBOL(cl_vmpage_page);
539 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
540 const struct lu_device_type *dtype)
542 return cl_page_at_trusted(page, dtype);
544 EXPORT_SYMBOL(cl_page_at);
546 static void cl_page_owner_clear(struct cl_page *page)
549 if (page->cp_owner != NULL) {
550 LASSERT(page->cp_owner->ci_owned_nr > 0);
551 page->cp_owner->ci_owned_nr--;
552 page->cp_owner = NULL;
557 static void cl_page_owner_set(struct cl_page *page)
560 LASSERT(page->cp_owner != NULL);
561 page->cp_owner->ci_owned_nr++;
565 void cl_page_disown0(const struct lu_env *env,
566 struct cl_io *io, struct cl_page *cl_page)
568 const struct cl_page_slice *slice;
569 enum cl_page_state state;
573 state = cl_page->cp_state;
574 PINVRNT(env, cl_page, state == CPS_OWNED ||
575 state == CPS_FREEING);
576 PINVRNT(env, cl_page, cl_page_invariant(cl_page) ||
577 state == CPS_FREEING);
578 cl_page_owner_clear(cl_page);
580 if (state == CPS_OWNED)
581 cl_page_state_set(env, cl_page, CPS_CACHED);
583 * Completion call-backs are executed in the bottom-up order, so that
584 * uppermost layer (llite), responsible for VFS/VM interaction runs
585 * last and can release locks safely.
587 cl_page_slice_for_each_reverse(cl_page, slice, i) {
588 if (slice->cpl_ops->cpo_disown != NULL)
589 (*slice->cpl_ops->cpo_disown)(env, slice, io);
596 * returns true, iff page is owned by the given io.
598 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
600 struct cl_io *top = cl_io_top((struct cl_io *)io);
601 LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
603 RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top);
605 EXPORT_SYMBOL(cl_page_is_owned);
608 * Try to own a page by IO.
610 * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
611 * into cl_page_state::CPS_OWNED state.
613 * \pre !cl_page_is_owned(cl_page, io)
614 * \post result == 0 iff cl_page_is_owned(cl_page, io)
618 * \retval -ve failure, e.g., cl_page was destroyed (and landed in
619 * cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
620 * or, page was owned by another thread, or in IO.
622 * \see cl_page_disown()
623 * \see cl_page_operations::cpo_own()
624 * \see cl_page_own_try()
627 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
628 struct cl_page *cl_page, int nonblock)
630 const struct cl_page_slice *slice;
635 PINVRNT(env, cl_page, !cl_page_is_owned(cl_page, io));
638 if (cl_page->cp_state == CPS_FREEING) {
643 cl_page_slice_for_each(cl_page, slice, i) {
644 if (slice->cpl_ops->cpo_own)
645 result = (*slice->cpl_ops->cpo_own)(env, slice,
654 PASSERT(env, cl_page, cl_page->cp_owner == NULL);
655 cl_page->cp_owner = cl_io_top(io);
656 cl_page_owner_set(cl_page);
657 if (cl_page->cp_state != CPS_FREEING) {
658 cl_page_state_set(env, cl_page, CPS_OWNED);
660 cl_page_disown0(env, io, cl_page);
666 PINVRNT(env, cl_page, ergo(result == 0,
667 cl_page_invariant(cl_page)));
672 * Own a page, might be blocked.
674 * \see cl_page_own0()
676 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
678 return cl_page_own0(env, io, pg, 0);
680 EXPORT_SYMBOL(cl_page_own);
683 * Nonblock version of cl_page_own().
685 * \see cl_page_own0()
687 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
690 return cl_page_own0(env, io, pg, 1);
692 EXPORT_SYMBOL(cl_page_own_try);
696 * Assume page ownership.
698 * Called when page is already locked by the hosting VM.
700 * \pre !cl_page_is_owned(cl_page, io)
701 * \post cl_page_is_owned(cl_page, io)
703 * \see cl_page_operations::cpo_assume()
705 void cl_page_assume(const struct lu_env *env,
706 struct cl_io *io, struct cl_page *cl_page)
708 const struct cl_page_slice *slice;
713 PINVRNT(env, cl_page,
714 cl_object_same(cl_page->cp_obj, io->ci_obj));
717 cl_page_slice_for_each(cl_page, slice, i) {
718 if (slice->cpl_ops->cpo_assume != NULL)
719 (*slice->cpl_ops->cpo_assume)(env, slice, io);
722 PASSERT(env, cl_page, cl_page->cp_owner == NULL);
723 cl_page->cp_owner = cl_io_top(io);
724 cl_page_owner_set(cl_page);
725 cl_page_state_set(env, cl_page, CPS_OWNED);
728 EXPORT_SYMBOL(cl_page_assume);
731 * Releases page ownership without unlocking the page.
733 * Moves cl_page into cl_page_state::CPS_CACHED without releasing a lock
734 * on the underlying VM page (as VM is supposed to do this itself).
736 * \pre cl_page_is_owned(cl_page, io)
737 * \post !cl_page_is_owned(cl_page, io)
739 * \see cl_page_assume()
741 void cl_page_unassume(const struct lu_env *env,
742 struct cl_io *io, struct cl_page *cl_page)
744 const struct cl_page_slice *slice;
748 PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
749 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
752 cl_page_owner_clear(cl_page);
753 cl_page_state_set(env, cl_page, CPS_CACHED);
755 cl_page_slice_for_each_reverse(cl_page, slice, i) {
756 if (slice->cpl_ops->cpo_unassume != NULL)
757 (*slice->cpl_ops->cpo_unassume)(env, slice, io);
762 EXPORT_SYMBOL(cl_page_unassume);
765 * Releases page ownership.
767 * Moves page into cl_page_state::CPS_CACHED.
769 * \pre cl_page_is_owned(pg, io)
770 * \post !cl_page_is_owned(pg, io)
773 * \see cl_page_operations::cpo_disown()
775 void cl_page_disown(const struct lu_env *env,
776 struct cl_io *io, struct cl_page *pg)
778 PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
779 pg->cp_state == CPS_FREEING);
783 cl_page_disown0(env, io, pg);
786 EXPORT_SYMBOL(cl_page_disown);
789 * Called when cl_page is to be removed from the object, e.g.,
790 * as a result of truncate.
792 * Calls cl_page_operations::cpo_discard() top-to-bottom.
794 * \pre cl_page_is_owned(cl_page, io)
796 * \see cl_page_operations::cpo_discard()
798 void cl_page_discard(const struct lu_env *env,
799 struct cl_io *io, struct cl_page *cl_page)
801 const struct cl_page_slice *slice;
804 PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
805 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
807 cl_page_slice_for_each(cl_page, slice, i) {
808 if (slice->cpl_ops->cpo_discard != NULL)
809 (*slice->cpl_ops->cpo_discard)(env, slice, io);
812 EXPORT_SYMBOL(cl_page_discard);
815 * Version of cl_page_delete() that can be called for not fully constructed
816 * cl_pages, e.g. in an error handling cl_page_find()->cl_page_delete0()
817 * path. Doesn't check cl_page invariant.
819 static void cl_page_delete0(const struct lu_env *env,
820 struct cl_page *cl_page)
822 const struct cl_page_slice *slice;
827 PASSERT(env, cl_page, cl_page->cp_state != CPS_FREEING);
830 * Severe all ways to obtain new pointers to @pg.
832 cl_page_owner_clear(cl_page);
833 cl_page_state_set0(env, cl_page, CPS_FREEING);
835 cl_page_slice_for_each_reverse(cl_page, slice, i) {
836 if (slice->cpl_ops->cpo_delete != NULL)
837 (*slice->cpl_ops->cpo_delete)(env, slice);
844 * Called when a decision is made to throw page out of memory.
846 * Notifies all layers about page destruction by calling
847 * cl_page_operations::cpo_delete() method top-to-bottom.
849 * Moves page into cl_page_state::CPS_FREEING state (this is the only place
850 * where transition to this state happens).
852 * Eliminates all venues through which new references to the page can be
855 * - removes page from the radix trees,
857 * - breaks linkage from VM page to cl_page.
859 * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
860 * drain after some time, at which point page will be recycled.
862 * \pre VM page is locked
863 * \post pg->cp_state == CPS_FREEING
865 * \see cl_page_operations::cpo_delete()
867 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
869 PINVRNT(env, pg, cl_page_invariant(pg));
871 cl_page_delete0(env, pg);
874 EXPORT_SYMBOL(cl_page_delete);
877 * Marks page up-to-date.
879 * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
880 * layer responsible for VM interaction has to mark/clear page as up-to-date
881 * by the \a uptodate argument.
883 * \see cl_page_operations::cpo_export()
885 void cl_page_export(const struct lu_env *env, struct cl_page *cl_page,
888 const struct cl_page_slice *slice;
891 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
893 cl_page_slice_for_each(cl_page, slice, i) {
894 if (slice->cpl_ops->cpo_export != NULL)
895 (*slice->cpl_ops->cpo_export)(env, slice, uptodate);
898 EXPORT_SYMBOL(cl_page_export);
901 * Returns true, if \a page is VM locked in a suitable sense by the calling
904 int cl_page_is_vmlocked(const struct lu_env *env,
905 const struct cl_page *cl_page)
907 const struct cl_page_slice *slice;
911 slice = cl_page_slice_get(cl_page, 0);
912 PASSERT(env, cl_page, slice->cpl_ops->cpo_is_vmlocked != NULL);
914 * Call ->cpo_is_vmlocked() directly instead of going through
915 * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
916 * cl_page_invariant().
918 result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
919 PASSERT(env, cl_page, result == -EBUSY || result == -ENODATA);
921 RETURN(result == -EBUSY);
923 EXPORT_SYMBOL(cl_page_is_vmlocked);
925 void cl_page_touch(const struct lu_env *env,
926 const struct cl_page *cl_page, size_t to)
928 const struct cl_page_slice *slice;
933 cl_page_slice_for_each(cl_page, slice, i) {
934 if (slice->cpl_ops->cpo_page_touch != NULL)
935 (*slice->cpl_ops->cpo_page_touch)(env, slice, to);
940 EXPORT_SYMBOL(cl_page_touch);
942 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
945 RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
948 static void cl_page_io_start(const struct lu_env *env,
949 struct cl_page *pg, enum cl_req_type crt)
952 * Page is queued for IO, change its state.
955 cl_page_owner_clear(pg);
956 cl_page_state_set(env, pg, cl_req_type_state(crt));
961 * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
962 * called top-to-bottom. Every layer either agrees to submit this page (by
963 * returning 0), or requests to omit this page (by returning -EALREADY). Layer
964 * handling interactions with the VM also has to inform VM that page is under
967 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
968 struct cl_page *cl_page, enum cl_req_type crt)
970 const struct cl_page_slice *slice;
974 PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
975 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
976 PINVRNT(env, cl_page, crt < CRT_NR);
979 * this has to be called bottom-to-top, so that llite can set up
980 * PG_writeback without risking other layers deciding to skip this
986 if (cl_page->cp_type != CPT_TRANSIENT) {
987 cl_page_slice_for_each(cl_page, slice, i) {
988 if (slice->cpl_ops->cpo_own)
990 (*slice->cpl_ops->io[crt].cpo_prep)(env,
1000 cl_page_io_start(env, cl_page, crt);
1003 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", crt, result);
1006 EXPORT_SYMBOL(cl_page_prep);
1009 * Notify layers about transfer completion.
1011 * Invoked by transfer sub-system (which is a part of osc) to notify layers
1012 * that a transfer, of which this page is a part of has completed.
1014 * Completion call-backs are executed in the bottom-up order, so that
1015 * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1016 * and can release locks safely.
1018 * \pre cl_page->cp_state == CPS_PAGEIN || cl_page->cp_state == CPS_PAGEOUT
1019 * \post cl_page->cl_page_state == CPS_CACHED
1021 * \see cl_page_operations::cpo_completion()
1023 void cl_page_completion(const struct lu_env *env,
1024 struct cl_page *cl_page, enum cl_req_type crt,
1027 const struct cl_page_slice *slice;
1028 struct cl_sync_io *anchor = cl_page->cp_sync_io;
1032 PASSERT(env, cl_page, crt < CRT_NR);
1033 PASSERT(env, cl_page, cl_page->cp_state == cl_req_type_state(crt));
1035 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", crt, ioret);
1036 cl_page_state_set(env, cl_page, CPS_CACHED);
1040 cl_page_slice_for_each_reverse(cl_page, slice, i) {
1041 if (slice->cpl_ops->io[crt].cpo_completion != NULL)
1042 (*slice->cpl_ops->io[crt].cpo_completion)(env, slice,
1046 if (anchor != NULL) {
1047 LASSERT(cl_page->cp_sync_io == anchor);
1048 cl_page->cp_sync_io = NULL;
1049 cl_sync_io_note(env, anchor, ioret);
1053 EXPORT_SYMBOL(cl_page_completion);
1056 * Notify layers that transfer formation engine decided to yank this page from
1057 * the cache and to make it a part of a transfer.
1059 * \pre cl_page->cp_state == CPS_CACHED
1060 * \post cl_page->cp_state == CPS_PAGEIN || cl_page->cp_state == CPS_PAGEOUT
1062 * \see cl_page_operations::cpo_make_ready()
1064 int cl_page_make_ready(const struct lu_env *env, struct cl_page *cl_page,
1065 enum cl_req_type crt)
1067 const struct cl_page_slice *slice;
1072 PINVRNT(env, cl_page, crt < CRT_NR);
1076 cl_page_slice_for_each(cl_page, slice, i) {
1077 if (slice->cpl_ops->io[crt].cpo_make_ready != NULL)
1078 result = (*slice->cpl_ops->io[crt].cpo_make_ready)(env, slice);
1085 PASSERT(env, cl_page, cl_page->cp_state == CPS_CACHED);
1086 cl_page_io_start(env, cl_page, crt);
1088 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", crt, result);
1092 EXPORT_SYMBOL(cl_page_make_ready);
1095 * Called if a page is being written back by kernel's intention.
1097 * \pre cl_page_is_owned(cl_page, io)
1098 * \post ergo(result == 0, cl_page->cp_state == CPS_PAGEOUT)
1100 * \see cl_page_operations::cpo_flush()
1102 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
1103 struct cl_page *cl_page)
1105 const struct cl_page_slice *slice;
1110 PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
1111 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
1113 cl_page_slice_for_each(cl_page, slice, i) {
1114 if (slice->cpl_ops->cpo_flush != NULL)
1115 result = (*slice->cpl_ops->cpo_flush)(env, slice, io);
1122 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d\n", result);
1125 EXPORT_SYMBOL(cl_page_flush);
1128 * Tells transfer engine that only part of a page is to be transmitted.
1130 * \see cl_page_operations::cpo_clip()
1132 void cl_page_clip(const struct lu_env *env, struct cl_page *cl_page,
1135 const struct cl_page_slice *slice;
1138 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
1140 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", from, to);
1141 cl_page_slice_for_each(cl_page, slice, i) {
1142 if (slice->cpl_ops->cpo_clip != NULL)
1143 (*slice->cpl_ops->cpo_clip)(env, slice, from, to);
1146 EXPORT_SYMBOL(cl_page_clip);
1149 * Prints human readable representation of \a pg to the \a f.
1151 void cl_page_header_print(const struct lu_env *env, void *cookie,
1152 lu_printer_t printer, const struct cl_page *pg)
1154 (*printer)(env, cookie,
1155 "page@%p[%d %p %d %d %p]\n",
1156 pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1157 pg->cp_state, pg->cp_type,
1160 EXPORT_SYMBOL(cl_page_header_print);
1163 * Prints human readable representation of \a cl_page to the \a f.
1165 void cl_page_print(const struct lu_env *env, void *cookie,
1166 lu_printer_t printer, const struct cl_page *cl_page)
1168 const struct cl_page_slice *slice;
1172 cl_page_header_print(env, cookie, printer, cl_page);
1173 cl_page_slice_for_each(cl_page, slice, i) {
1174 if (slice->cpl_ops->cpo_print != NULL)
1175 result = (*slice->cpl_ops->cpo_print)(env, slice,
1180 (*printer)(env, cookie, "end page@%p\n", cl_page);
1182 EXPORT_SYMBOL(cl_page_print);
1185 * Converts a byte offset within object \a obj into a page index.
1187 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1189 return (loff_t)idx << PAGE_SHIFT;
1191 EXPORT_SYMBOL(cl_offset);
1194 * Converts a page index into a byte offset within object \a obj.
1196 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1198 return offset >> PAGE_SHIFT;
1200 EXPORT_SYMBOL(cl_index);
1202 size_t cl_page_size(const struct cl_object *obj)
1204 return 1UL << PAGE_SHIFT;
1206 EXPORT_SYMBOL(cl_page_size);
1209 * Adds page slice to the compound page.
1211 * This is called by cl_object_operations::coo_page_init() methods to add a
1212 * per-layer state to the page. New state is added at the end of
1213 * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1215 * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1217 void cl_page_slice_add(struct cl_page *cl_page, struct cl_page_slice *slice,
1218 struct cl_object *obj,
1219 const struct cl_page_operations *ops)
1221 unsigned int offset = (char *)slice -
1222 ((char *)cl_page + sizeof(*cl_page));
1225 LASSERT(cl_page->cp_layer_count < CP_MAX_LAYER);
1226 LASSERT(offset < (1 << sizeof(cl_page->cp_layer_offset[0]) * 8));
1227 cl_page->cp_layer_offset[cl_page->cp_layer_count++] = offset;
1228 slice->cpl_obj = obj;
1229 slice->cpl_ops = ops;
1230 slice->cpl_page = cl_page;
1234 EXPORT_SYMBOL(cl_page_slice_add);
1237 * Allocate and initialize cl_cache, called by ll_init_sbi().
1239 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max)
1241 struct cl_client_cache *cache = NULL;
1244 OBD_ALLOC(cache, sizeof(*cache));
1248 /* Initialize cache data */
1249 atomic_set(&cache->ccc_users, 1);
1250 cache->ccc_lru_max = lru_page_max;
1251 atomic_long_set(&cache->ccc_lru_left, lru_page_max);
1252 spin_lock_init(&cache->ccc_lru_lock);
1253 INIT_LIST_HEAD(&cache->ccc_lru);
1255 /* turn unstable check off by default as it impacts performance */
1256 cache->ccc_unstable_check = 0;
1257 atomic_long_set(&cache->ccc_unstable_nr, 0);
1258 init_waitqueue_head(&cache->ccc_unstable_waitq);
1259 mutex_init(&cache->ccc_max_cache_mb_lock);
1263 EXPORT_SYMBOL(cl_cache_init);
1266 * Increase cl_cache refcount
1268 void cl_cache_incref(struct cl_client_cache *cache)
1270 atomic_inc(&cache->ccc_users);
1272 EXPORT_SYMBOL(cl_cache_incref);
1275 * Decrease cl_cache refcount and free the cache if refcount=0.
1276 * Since llite, lov and osc all hold cl_cache refcount,
1277 * the free will not cause race. (LU-6173)
1279 void cl_cache_decref(struct cl_client_cache *cache)
1281 if (atomic_dec_and_test(&cache->ccc_users))
1282 OBD_FREE(cache, sizeof(*cache));
1284 EXPORT_SYMBOL(cl_cache_decref);