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 cl_object_put(env, obj);
207 lu_ref_fini(&cl_page->cp_reference);
208 __cl_page_free(cl_page, bufsize);
212 static struct cl_page *__cl_page_alloc(struct cl_object *o)
215 struct cl_page *cl_page = NULL;
216 unsigned short bufsize = cl_object_header(o)->coh_page_bufsize;
218 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_PAGE_ALLOC))
222 /* the number of entries in cl_page_kmem_array is expected to
223 * only be 2-3 entries, so the lookup overhead should be low.
225 for ( ; i < ARRAY_SIZE(cl_page_kmem_array); i++) {
226 if (smp_load_acquire(&cl_page_kmem_size_array[i])
228 OBD_SLAB_ALLOC_GFP(cl_page, cl_page_kmem_array[i],
231 cl_page->cp_kmem_index = i;
234 if (cl_page_kmem_size_array[i] == 0)
238 if (i < ARRAY_SIZE(cl_page_kmem_array)) {
241 mutex_lock(&cl_page_kmem_mutex);
242 if (cl_page_kmem_size_array[i]) {
243 mutex_unlock(&cl_page_kmem_mutex);
246 snprintf(cache_name, sizeof(cache_name),
247 "cl_page_kmem-%u", bufsize);
248 cl_page_kmem_array[i] =
249 kmem_cache_create(cache_name, bufsize,
251 if (cl_page_kmem_array[i] == NULL) {
252 mutex_unlock(&cl_page_kmem_mutex);
255 smp_store_release(&cl_page_kmem_size_array[i],
257 mutex_unlock(&cl_page_kmem_mutex);
260 OBD_ALLOC_GFP(cl_page, bufsize, GFP_NOFS);
262 cl_page->cp_kmem_index = -1;
268 struct cl_page *cl_page_alloc(const struct lu_env *env, struct cl_object *o,
269 pgoff_t ind, struct page *vmpage,
270 enum cl_page_type type)
272 struct cl_page *cl_page;
273 struct cl_object *head;
277 cl_page = __cl_page_alloc(o);
278 if (cl_page != NULL) {
282 * Please fix cl_page:cp_state/type declaration if
283 * these assertions fail in the future.
285 BUILD_BUG_ON((1 << CP_STATE_BITS) < CPS_NR); /* cp_state */
286 BUILD_BUG_ON((1 << CP_TYPE_BITS) < CPT_NR); /* cp_type */
287 atomic_set(&cl_page->cp_ref, 1);
290 lu_object_ref_add_at(&o->co_lu, &cl_page->cp_obj_ref,
292 cl_page->cp_vmpage = vmpage;
293 cl_page->cp_state = CPS_CACHED;
294 cl_page->cp_type = type;
295 cl_page->cp_inode = NULL;
296 INIT_LIST_HEAD(&cl_page->cp_batch);
297 lu_ref_init(&cl_page->cp_reference);
299 cl_object_for_each(o, head) {
300 if (o->co_ops->coo_page_init != NULL) {
301 result = o->co_ops->coo_page_init(env, o,
304 cl_page_delete0(env, cl_page);
305 cl_page_free(env, cl_page, NULL);
306 cl_page = ERR_PTR(result);
312 cs_page_inc(o, CS_total);
313 cs_page_inc(o, CS_create);
314 cs_pagestate_dec(o, CPS_CACHED);
317 cl_page = ERR_PTR(-ENOMEM);
323 * Returns a cl_page with index \a idx at the object \a o, and associated with
324 * the VM page \a vmpage.
326 * This is the main entry point into the cl_page caching interface. First, a
327 * cache (implemented as a per-object radix tree) is consulted. If page is
328 * found there, it is returned immediately. Otherwise new page is allocated
329 * and returned. In any case, additional reference to page is acquired.
331 * \see cl_object_find(), cl_lock_find()
333 struct cl_page *cl_page_find(const struct lu_env *env,
335 pgoff_t idx, struct page *vmpage,
336 enum cl_page_type type)
338 struct cl_page *page = NULL;
339 struct cl_object_header *hdr;
341 LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
346 hdr = cl_object_header(o);
347 cs_page_inc(o, CS_lookup);
349 CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
350 idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
352 if (type == CPT_CACHEABLE) {
353 /* vmpage lock is used to protect the child/parent
355 LASSERT(PageLocked(vmpage));
357 * cl_vmpage_page() can be called here without any locks as
359 * - "vmpage" is locked (which prevents ->private from
360 * concurrent updates), and
362 * - "o" cannot be destroyed while current thread holds a
365 page = cl_vmpage_page(vmpage, o);
367 cs_page_inc(o, CS_hit);
372 /* allocate and initialize cl_page */
373 page = cl_page_alloc(env, o, idx, vmpage, type);
376 EXPORT_SYMBOL(cl_page_find);
378 static inline int cl_page_invariant(const struct cl_page *pg)
380 return cl_page_in_use_noref(pg);
383 static void cl_page_state_set0(const struct lu_env *env,
384 struct cl_page *cl_page,
385 enum cl_page_state state)
387 enum cl_page_state old;
390 * Matrix of allowed state transitions [old][new], for sanity
393 static const int allowed_transitions[CPS_NR][CPS_NR] = {
396 [CPS_OWNED] = 1, /* io finds existing cached page */
398 [CPS_PAGEOUT] = 1, /* write-out from the cache */
399 [CPS_FREEING] = 1, /* eviction on the memory pressure */
402 [CPS_CACHED] = 1, /* release to the cache */
404 [CPS_PAGEIN] = 1, /* start read immediately */
405 [CPS_PAGEOUT] = 1, /* start write immediately */
406 [CPS_FREEING] = 1, /* lock invalidation or truncate */
409 [CPS_CACHED] = 1, /* io completion */
416 [CPS_CACHED] = 1, /* io completion */
432 old = cl_page->cp_state;
433 PASSERT(env, cl_page, allowed_transitions[old][state]);
434 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d -> %d\n", old, state);
435 PASSERT(env, cl_page, cl_page->cp_state == old);
436 PASSERT(env, cl_page, equi(state == CPS_OWNED,
437 cl_page->cp_owner != NULL));
439 cs_pagestate_dec(cl_page->cp_obj, cl_page->cp_state);
440 cs_pagestate_inc(cl_page->cp_obj, state);
441 cl_page->cp_state = state;
445 static void cl_page_state_set(const struct lu_env *env,
446 struct cl_page *page, enum cl_page_state state)
448 cl_page_state_set0(env, page, state);
452 * Acquires an additional reference to a page.
454 * This can be called only by caller already possessing a reference to \a
457 * \see cl_object_get(), cl_lock_get().
459 void cl_page_get(struct cl_page *page)
462 cl_page_get_trust(page);
465 EXPORT_SYMBOL(cl_page_get);
468 * Releases a reference to a page, use the pagevec to release the pages
469 * in batch if provided.
471 * Users need to do a final pagevec_release() to release any trailing pages.
473 void cl_pagevec_put(const struct lu_env *env, struct cl_page *page,
474 struct pagevec *pvec)
477 CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
478 atomic_read(&page->cp_ref));
480 if (atomic_dec_and_test(&page->cp_ref)) {
481 LASSERT(page->cp_state == CPS_FREEING);
483 LASSERT(atomic_read(&page->cp_ref) == 0);
484 PASSERT(env, page, page->cp_owner == NULL);
485 PASSERT(env, page, list_empty(&page->cp_batch));
487 * Page is no longer reachable by other threads. Tear
490 cl_page_free(env, page, pvec);
495 EXPORT_SYMBOL(cl_pagevec_put);
498 * Releases a reference to a page, wrapper to cl_pagevec_put
500 * When last reference is released, page is returned to the cache, unless it
501 * is in cl_page_state::CPS_FREEING state, in which case it is immediately
504 * \see cl_object_put(), cl_lock_put().
506 void cl_page_put(const struct lu_env *env, struct cl_page *page)
508 cl_pagevec_put(env, page, NULL);
510 EXPORT_SYMBOL(cl_page_put);
513 * Returns a cl_page associated with a VM page, and given cl_object.
515 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj)
517 struct cl_page *page;
520 LASSERT(PageLocked(vmpage));
523 * NOTE: absence of races and liveness of data are guaranteed by page
524 * lock on a "vmpage". That works because object destruction has
525 * bottom-to-top pass.
528 page = (struct cl_page *)vmpage->private;
530 cl_page_get_trust(page);
531 LASSERT(page->cp_type == CPT_CACHEABLE);
535 EXPORT_SYMBOL(cl_vmpage_page);
537 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
538 const struct lu_device_type *dtype)
540 return cl_page_at_trusted(page, dtype);
542 EXPORT_SYMBOL(cl_page_at);
544 static void cl_page_owner_clear(struct cl_page *page)
547 if (page->cp_owner != NULL) {
548 LASSERT(page->cp_owner->ci_owned_nr > 0);
549 page->cp_owner->ci_owned_nr--;
550 page->cp_owner = NULL;
555 static void cl_page_owner_set(struct cl_page *page)
558 LASSERT(page->cp_owner != NULL);
559 page->cp_owner->ci_owned_nr++;
563 void cl_page_disown0(const struct lu_env *env,
564 struct cl_io *io, struct cl_page *cl_page)
566 const struct cl_page_slice *slice;
567 enum cl_page_state state;
571 state = cl_page->cp_state;
572 PINVRNT(env, cl_page, state == CPS_OWNED ||
573 state == CPS_FREEING);
574 PINVRNT(env, cl_page, cl_page_invariant(cl_page) ||
575 state == CPS_FREEING);
576 cl_page_owner_clear(cl_page);
578 if (state == CPS_OWNED)
579 cl_page_state_set(env, cl_page, CPS_CACHED);
581 * Completion call-backs are executed in the bottom-up order, so that
582 * uppermost layer (llite), responsible for VFS/VM interaction runs
583 * last and can release locks safely.
585 cl_page_slice_for_each_reverse(cl_page, slice, i) {
586 if (slice->cpl_ops->cpo_disown != NULL)
587 (*slice->cpl_ops->cpo_disown)(env, slice, io);
594 * returns true, iff page is owned by the given io.
596 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
598 struct cl_io *top = cl_io_top((struct cl_io *)io);
599 LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
601 RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top);
603 EXPORT_SYMBOL(cl_page_is_owned);
606 * Try to own a page by IO.
608 * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
609 * into cl_page_state::CPS_OWNED state.
611 * \pre !cl_page_is_owned(cl_page, io)
612 * \post result == 0 iff cl_page_is_owned(cl_page, io)
616 * \retval -ve failure, e.g., cl_page was destroyed (and landed in
617 * cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
618 * or, page was owned by another thread, or in IO.
620 * \see cl_page_disown()
621 * \see cl_page_operations::cpo_own()
622 * \see cl_page_own_try()
625 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
626 struct cl_page *cl_page, int nonblock)
628 const struct cl_page_slice *slice;
633 PINVRNT(env, cl_page, !cl_page_is_owned(cl_page, io));
636 if (cl_page->cp_state == CPS_FREEING) {
641 cl_page_slice_for_each(cl_page, slice, i) {
642 if (slice->cpl_ops->cpo_own)
643 result = (*slice->cpl_ops->cpo_own)(env, slice,
652 PASSERT(env, cl_page, cl_page->cp_owner == NULL);
653 cl_page->cp_owner = cl_io_top(io);
654 cl_page_owner_set(cl_page);
655 if (cl_page->cp_state != CPS_FREEING) {
656 cl_page_state_set(env, cl_page, CPS_OWNED);
658 cl_page_disown0(env, io, cl_page);
664 PINVRNT(env, cl_page, ergo(result == 0,
665 cl_page_invariant(cl_page)));
670 * Own a page, might be blocked.
672 * \see cl_page_own0()
674 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
676 return cl_page_own0(env, io, pg, 0);
678 EXPORT_SYMBOL(cl_page_own);
681 * Nonblock version of cl_page_own().
683 * \see cl_page_own0()
685 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
688 return cl_page_own0(env, io, pg, 1);
690 EXPORT_SYMBOL(cl_page_own_try);
694 * Assume page ownership.
696 * Called when page is already locked by the hosting VM.
698 * \pre !cl_page_is_owned(cl_page, io)
699 * \post cl_page_is_owned(cl_page, io)
701 * \see cl_page_operations::cpo_assume()
703 void cl_page_assume(const struct lu_env *env,
704 struct cl_io *io, struct cl_page *cl_page)
706 const struct cl_page_slice *slice;
711 PINVRNT(env, cl_page,
712 cl_object_same(cl_page->cp_obj, io->ci_obj));
715 cl_page_slice_for_each(cl_page, slice, i) {
716 if (slice->cpl_ops->cpo_assume != NULL)
717 (*slice->cpl_ops->cpo_assume)(env, slice, io);
720 PASSERT(env, cl_page, cl_page->cp_owner == NULL);
721 cl_page->cp_owner = cl_io_top(io);
722 cl_page_owner_set(cl_page);
723 cl_page_state_set(env, cl_page, CPS_OWNED);
726 EXPORT_SYMBOL(cl_page_assume);
729 * Releases page ownership without unlocking the page.
731 * Moves cl_page into cl_page_state::CPS_CACHED without releasing a lock
732 * on the underlying VM page (as VM is supposed to do this itself).
734 * \pre cl_page_is_owned(cl_page, io)
735 * \post !cl_page_is_owned(cl_page, io)
737 * \see cl_page_assume()
739 void cl_page_unassume(const struct lu_env *env,
740 struct cl_io *io, struct cl_page *cl_page)
742 const struct cl_page_slice *slice;
746 PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
747 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
750 cl_page_owner_clear(cl_page);
751 cl_page_state_set(env, cl_page, CPS_CACHED);
753 cl_page_slice_for_each_reverse(cl_page, slice, i) {
754 if (slice->cpl_ops->cpo_unassume != NULL)
755 (*slice->cpl_ops->cpo_unassume)(env, slice, io);
760 EXPORT_SYMBOL(cl_page_unassume);
763 * Releases page ownership.
765 * Moves page into cl_page_state::CPS_CACHED.
767 * \pre cl_page_is_owned(pg, io)
768 * \post !cl_page_is_owned(pg, io)
771 * \see cl_page_operations::cpo_disown()
773 void cl_page_disown(const struct lu_env *env,
774 struct cl_io *io, struct cl_page *pg)
776 PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
777 pg->cp_state == CPS_FREEING);
781 cl_page_disown0(env, io, pg);
784 EXPORT_SYMBOL(cl_page_disown);
787 * Called when cl_page is to be removed from the object, e.g.,
788 * as a result of truncate.
790 * Calls cl_page_operations::cpo_discard() top-to-bottom.
792 * \pre cl_page_is_owned(cl_page, io)
794 * \see cl_page_operations::cpo_discard()
796 void cl_page_discard(const struct lu_env *env,
797 struct cl_io *io, struct cl_page *cl_page)
799 const struct cl_page_slice *slice;
802 PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
803 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
805 cl_page_slice_for_each(cl_page, slice, i) {
806 if (slice->cpl_ops->cpo_discard != NULL)
807 (*slice->cpl_ops->cpo_discard)(env, slice, io);
810 EXPORT_SYMBOL(cl_page_discard);
813 * Version of cl_page_delete() that can be called for not fully constructed
814 * cl_pages, e.g. in an error handling cl_page_find()->cl_page_delete0()
815 * path. Doesn't check cl_page invariant.
817 static void cl_page_delete0(const struct lu_env *env,
818 struct cl_page *cl_page)
820 const struct cl_page_slice *slice;
825 PASSERT(env, cl_page, cl_page->cp_state != CPS_FREEING);
828 * Severe all ways to obtain new pointers to @pg.
830 cl_page_owner_clear(cl_page);
831 cl_page_state_set0(env, cl_page, CPS_FREEING);
833 cl_page_slice_for_each_reverse(cl_page, slice, i) {
834 if (slice->cpl_ops->cpo_delete != NULL)
835 (*slice->cpl_ops->cpo_delete)(env, slice);
842 * Called when a decision is made to throw page out of memory.
844 * Notifies all layers about page destruction by calling
845 * cl_page_operations::cpo_delete() method top-to-bottom.
847 * Moves page into cl_page_state::CPS_FREEING state (this is the only place
848 * where transition to this state happens).
850 * Eliminates all venues through which new references to the page can be
853 * - removes page from the radix trees,
855 * - breaks linkage from VM page to cl_page.
857 * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
858 * drain after some time, at which point page will be recycled.
860 * \pre VM page is locked
861 * \post pg->cp_state == CPS_FREEING
863 * \see cl_page_operations::cpo_delete()
865 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
867 PINVRNT(env, pg, cl_page_invariant(pg));
869 cl_page_delete0(env, pg);
872 EXPORT_SYMBOL(cl_page_delete);
875 * Marks page up-to-date.
877 * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
878 * layer responsible for VM interaction has to mark/clear page as up-to-date
879 * by the \a uptodate argument.
881 * \see cl_page_operations::cpo_export()
883 void cl_page_export(const struct lu_env *env, struct cl_page *cl_page,
886 const struct cl_page_slice *slice;
889 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
891 cl_page_slice_for_each(cl_page, slice, i) {
892 if (slice->cpl_ops->cpo_export != NULL)
893 (*slice->cpl_ops->cpo_export)(env, slice, uptodate);
896 EXPORT_SYMBOL(cl_page_export);
899 * Returns true, if \a page is VM locked in a suitable sense by the calling
902 int cl_page_is_vmlocked(const struct lu_env *env,
903 const struct cl_page *cl_page)
905 const struct cl_page_slice *slice;
909 slice = cl_page_slice_get(cl_page, 0);
910 PASSERT(env, cl_page, slice->cpl_ops->cpo_is_vmlocked != NULL);
912 * Call ->cpo_is_vmlocked() directly instead of going through
913 * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
914 * cl_page_invariant().
916 result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
917 PASSERT(env, cl_page, result == -EBUSY || result == -ENODATA);
919 RETURN(result == -EBUSY);
921 EXPORT_SYMBOL(cl_page_is_vmlocked);
923 void cl_page_touch(const struct lu_env *env,
924 const struct cl_page *cl_page, size_t to)
926 const struct cl_page_slice *slice;
931 cl_page_slice_for_each(cl_page, slice, i) {
932 if (slice->cpl_ops->cpo_page_touch != NULL)
933 (*slice->cpl_ops->cpo_page_touch)(env, slice, to);
938 EXPORT_SYMBOL(cl_page_touch);
940 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
943 RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
946 static void cl_page_io_start(const struct lu_env *env,
947 struct cl_page *pg, enum cl_req_type crt)
950 * Page is queued for IO, change its state.
953 cl_page_owner_clear(pg);
954 cl_page_state_set(env, pg, cl_req_type_state(crt));
959 * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
960 * called top-to-bottom. Every layer either agrees to submit this page (by
961 * returning 0), or requests to omit this page (by returning -EALREADY). Layer
962 * handling interactions with the VM also has to inform VM that page is under
965 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
966 struct cl_page *cl_page, enum cl_req_type crt)
968 const struct cl_page_slice *slice;
972 PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
973 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
974 PINVRNT(env, cl_page, crt < CRT_NR);
977 * this has to be called bottom-to-top, so that llite can set up
978 * PG_writeback without risking other layers deciding to skip this
984 cl_page_slice_for_each(cl_page, slice, i) {
985 if (slice->cpl_ops->cpo_own)
986 result = (*slice->cpl_ops->io[crt].cpo_prep)(env,
995 cl_page_io_start(env, cl_page, crt);
998 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", crt, result);
1001 EXPORT_SYMBOL(cl_page_prep);
1004 * Notify layers about transfer completion.
1006 * Invoked by transfer sub-system (which is a part of osc) to notify layers
1007 * that a transfer, of which this page is a part of has completed.
1009 * Completion call-backs are executed in the bottom-up order, so that
1010 * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1011 * and can release locks safely.
1013 * \pre cl_page->cp_state == CPS_PAGEIN || cl_page->cp_state == CPS_PAGEOUT
1014 * \post cl_page->cl_page_state == CPS_CACHED
1016 * \see cl_page_operations::cpo_completion()
1018 void cl_page_completion(const struct lu_env *env,
1019 struct cl_page *cl_page, enum cl_req_type crt,
1022 const struct cl_page_slice *slice;
1023 struct cl_sync_io *anchor = cl_page->cp_sync_io;
1027 PASSERT(env, cl_page, crt < CRT_NR);
1028 PASSERT(env, cl_page, cl_page->cp_state == cl_req_type_state(crt));
1030 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", crt, ioret);
1031 cl_page_state_set(env, cl_page, CPS_CACHED);
1035 cl_page_slice_for_each_reverse(cl_page, slice, i) {
1036 if (slice->cpl_ops->io[crt].cpo_completion != NULL)
1037 (*slice->cpl_ops->io[crt].cpo_completion)(env, slice,
1041 if (anchor != NULL) {
1042 LASSERT(cl_page->cp_sync_io == anchor);
1043 cl_page->cp_sync_io = NULL;
1044 cl_sync_io_note(env, anchor, ioret);
1048 EXPORT_SYMBOL(cl_page_completion);
1051 * Notify layers that transfer formation engine decided to yank this page from
1052 * the cache and to make it a part of a transfer.
1054 * \pre cl_page->cp_state == CPS_CACHED
1055 * \post cl_page->cp_state == CPS_PAGEIN || cl_page->cp_state == CPS_PAGEOUT
1057 * \see cl_page_operations::cpo_make_ready()
1059 int cl_page_make_ready(const struct lu_env *env, struct cl_page *cl_page,
1060 enum cl_req_type crt)
1062 const struct cl_page_slice *slice;
1067 PINVRNT(env, cl_page, crt < CRT_NR);
1071 cl_page_slice_for_each(cl_page, slice, i) {
1072 if (slice->cpl_ops->io[crt].cpo_make_ready != NULL)
1073 result = (*slice->cpl_ops->io[crt].cpo_make_ready)(env, slice);
1080 PASSERT(env, cl_page, cl_page->cp_state == CPS_CACHED);
1081 cl_page_io_start(env, cl_page, crt);
1083 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", crt, result);
1087 EXPORT_SYMBOL(cl_page_make_ready);
1090 * Called if a page is being written back by kernel's intention.
1092 * \pre cl_page_is_owned(cl_page, io)
1093 * \post ergo(result == 0, cl_page->cp_state == CPS_PAGEOUT)
1095 * \see cl_page_operations::cpo_flush()
1097 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
1098 struct cl_page *cl_page)
1100 const struct cl_page_slice *slice;
1105 PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
1106 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
1108 cl_page_slice_for_each(cl_page, slice, i) {
1109 if (slice->cpl_ops->cpo_flush != NULL)
1110 result = (*slice->cpl_ops->cpo_flush)(env, slice, io);
1117 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d\n", result);
1120 EXPORT_SYMBOL(cl_page_flush);
1123 * Tells transfer engine that only part of a page is to be transmitted.
1125 * \see cl_page_operations::cpo_clip()
1127 void cl_page_clip(const struct lu_env *env, struct cl_page *cl_page,
1130 const struct cl_page_slice *slice;
1133 PINVRNT(env, cl_page, cl_page_invariant(cl_page));
1135 CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", from, to);
1136 cl_page_slice_for_each(cl_page, slice, i) {
1137 if (slice->cpl_ops->cpo_clip != NULL)
1138 (*slice->cpl_ops->cpo_clip)(env, slice, from, to);
1141 EXPORT_SYMBOL(cl_page_clip);
1144 * Prints human readable representation of \a pg to the \a f.
1146 void cl_page_header_print(const struct lu_env *env, void *cookie,
1147 lu_printer_t printer, const struct cl_page *pg)
1149 (*printer)(env, cookie,
1150 "page@%p[%d %p %d %d %p]\n",
1151 pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1152 pg->cp_state, pg->cp_type,
1155 EXPORT_SYMBOL(cl_page_header_print);
1158 * Prints human readable representation of \a cl_page to the \a f.
1160 void cl_page_print(const struct lu_env *env, void *cookie,
1161 lu_printer_t printer, const struct cl_page *cl_page)
1163 const struct cl_page_slice *slice;
1167 cl_page_header_print(env, cookie, printer, cl_page);
1168 cl_page_slice_for_each(cl_page, slice, i) {
1169 if (slice->cpl_ops->cpo_print != NULL)
1170 result = (*slice->cpl_ops->cpo_print)(env, slice,
1175 (*printer)(env, cookie, "end page@%p\n", cl_page);
1177 EXPORT_SYMBOL(cl_page_print);
1180 * Converts a byte offset within object \a obj into a page index.
1182 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1184 return (loff_t)idx << PAGE_SHIFT;
1186 EXPORT_SYMBOL(cl_offset);
1189 * Converts a page index into a byte offset within object \a obj.
1191 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1193 return offset >> PAGE_SHIFT;
1195 EXPORT_SYMBOL(cl_index);
1197 size_t cl_page_size(const struct cl_object *obj)
1199 return 1UL << PAGE_SHIFT;
1201 EXPORT_SYMBOL(cl_page_size);
1204 * Adds page slice to the compound page.
1206 * This is called by cl_object_operations::coo_page_init() methods to add a
1207 * per-layer state to the page. New state is added at the end of
1208 * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1210 * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1212 void cl_page_slice_add(struct cl_page *cl_page, struct cl_page_slice *slice,
1213 struct cl_object *obj,
1214 const struct cl_page_operations *ops)
1216 unsigned int offset = (char *)slice -
1217 ((char *)cl_page + sizeof(*cl_page));
1220 LASSERT(cl_page->cp_layer_count < CP_MAX_LAYER);
1221 LASSERT(offset < (1 << sizeof(cl_page->cp_layer_offset[0]) * 8));
1222 cl_page->cp_layer_offset[cl_page->cp_layer_count++] = offset;
1223 slice->cpl_obj = obj;
1224 slice->cpl_ops = ops;
1225 slice->cpl_page = cl_page;
1229 EXPORT_SYMBOL(cl_page_slice_add);
1232 * Allocate and initialize cl_cache, called by ll_init_sbi().
1234 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max)
1236 struct cl_client_cache *cache = NULL;
1239 OBD_ALLOC(cache, sizeof(*cache));
1243 /* Initialize cache data */
1244 atomic_set(&cache->ccc_users, 1);
1245 cache->ccc_lru_max = lru_page_max;
1246 atomic_long_set(&cache->ccc_lru_left, lru_page_max);
1247 spin_lock_init(&cache->ccc_lru_lock);
1248 INIT_LIST_HEAD(&cache->ccc_lru);
1250 /* turn unstable check off by default as it impacts performance */
1251 cache->ccc_unstable_check = 0;
1252 atomic_long_set(&cache->ccc_unstable_nr, 0);
1253 init_waitqueue_head(&cache->ccc_unstable_waitq);
1254 mutex_init(&cache->ccc_max_cache_mb_lock);
1258 EXPORT_SYMBOL(cl_cache_init);
1261 * Increase cl_cache refcount
1263 void cl_cache_incref(struct cl_client_cache *cache)
1265 atomic_inc(&cache->ccc_users);
1267 EXPORT_SYMBOL(cl_cache_incref);
1270 * Decrease cl_cache refcount and free the cache if refcount=0.
1271 * Since llite, lov and osc all hold cl_cache refcount,
1272 * the free will not cause race. (LU-6173)
1274 void cl_cache_decref(struct cl_client_cache *cache)
1276 if (atomic_dec_and_test(&cache->ccc_users))
1277 OBD_FREE(cache, sizeof(*cache));
1279 EXPORT_SYMBOL(cl_cache_decref);