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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
41 #define DEBUG_SUBSYSTEM S_CLASS
43 #include <libcfs/libcfs.h>
44 #include <obd_class.h>
45 #include <obd_support.h>
46 #include <libcfs/list.h>
48 #include <cl_object.h>
49 #include "cl_internal.h"
51 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
55 # define PASSERT(env, page, expr) \
57 if (unlikely(!(expr))) { \
58 CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n"); \
62 #else /* !LIBCFS_DEBUG */
63 # define PASSERT(env, page, exp) \
64 ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
65 #endif /* !LIBCFS_DEBUG */
67 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
68 # define PINVRNT(env, page, expr) \
70 if (unlikely(!(expr))) { \
71 CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n"); \
75 #else /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
76 # define PINVRNT(env, page, exp) \
77 ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
78 #endif /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
80 /* Disable page statistic by default due to huge performance penalty. */
81 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
82 #define CS_PAGE_INC(o, item) \
83 cfs_atomic_inc(&cl_object_site(o)->cs_pages.cs_stats[CS_##item])
84 #define CS_PAGE_DEC(o, item) \
85 cfs_atomic_dec(&cl_object_site(o)->cs_pages.cs_stats[CS_##item])
86 #define CS_PAGESTATE_INC(o, state) \
87 cfs_atomic_inc(&cl_object_site(o)->cs_pages_state[state])
88 #define CS_PAGESTATE_DEC(o, state) \
89 cfs_atomic_dec(&cl_object_site(o)->cs_pages_state[state])
91 #define CS_PAGE_INC(o, item)
92 #define CS_PAGE_DEC(o, item)
93 #define CS_PAGESTATE_INC(o, state)
94 #define CS_PAGESTATE_DEC(o, state)
98 * Internal version of cl_page_top, it should be called if the page is
99 * known to be not freed, says with page referenced, or radix tree lock held,
102 static struct cl_page *cl_page_top_trusted(struct cl_page *page)
104 while (page->cp_parent != NULL)
105 page = page->cp_parent;
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 locking page radix-tree
115 * (cl_object_header::hdr->coh_page_guard), or 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(cfs_atomic_read(&page->cp_ref) > 0);
123 cfs_atomic_inc(&page->cp_ref);
127 * Returns a slice within a page, corresponding to the given layer in the
132 static const struct cl_page_slice *
133 cl_page_at_trusted(const struct cl_page *page,
134 const struct lu_device_type *dtype)
136 const struct cl_page_slice *slice;
139 page = cl_page_top_trusted((struct cl_page *)page);
141 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
142 if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
145 page = page->cp_child;
146 } while (page != NULL);
151 * Returns a page with given index in the given object, or NULL if no page is
152 * found. Acquires a reference on \a page.
154 * Locking: called under cl_object_header::coh_page_guard spin-lock.
156 struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index)
158 struct cl_page *page;
160 LASSERT(spin_is_locked(&hdr->coh_page_guard));
162 page = radix_tree_lookup(&hdr->coh_tree, index);
164 cl_page_get_trust(page);
167 EXPORT_SYMBOL(cl_page_lookup);
170 * Returns a list of pages by a given [start, end] of \a obj.
172 * \param resched If not NULL, then we give up before hogging CPU for too
173 * long and set *resched = 1, in that case caller should implement a retry
176 * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
177 * crucial in the face of [offset, EOF] locks.
179 * Return at least one page in @queue unless there is no covered page.
181 int cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj,
182 struct cl_io *io, pgoff_t start, pgoff_t end,
183 cl_page_gang_cb_t cb, void *cbdata)
185 struct cl_object_header *hdr;
186 struct cl_page *page;
187 struct cl_page **pvec;
188 const struct cl_page_slice *slice;
189 const struct lu_device_type *dtype;
194 int res = CLP_GANG_OKAY;
199 hdr = cl_object_header(obj);
200 pvec = cl_env_info(env)->clt_pvec;
201 dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type;
202 spin_lock(&hdr->coh_page_guard);
203 while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec,
204 idx, CLT_PVEC_SIZE)) > 0) {
205 int end_of_region = 0;
206 idx = pvec[nr - 1]->cp_index + 1;
207 for (i = 0, j = 0; i < nr; ++i) {
211 LASSERT(page->cp_type == CPT_CACHEABLE);
212 if (page->cp_index > end) {
216 if (page->cp_state == CPS_FREEING)
219 slice = cl_page_at_trusted(page, dtype);
221 * Pages for lsm-less file has no underneath sub-page
222 * for osc, in case of ...
224 PASSERT(env, page, slice != NULL);
226 page = slice->cpl_page;
228 * Can safely call cl_page_get_trust() under
229 * radix-tree spin-lock.
231 * XXX not true, because @page is from object another
232 * than @hdr and protected by different tree lock.
234 cl_page_get_trust(page);
235 lu_ref_add_atomic(&page->cp_reference,
236 "gang_lookup", cfs_current());
241 * Here a delicate locking dance is performed. Current thread
242 * holds a reference to a page, but has to own it before it
243 * can be placed into queue. Owning implies waiting, so
244 * radix-tree lock is to be released. After a wait one has to
245 * check that pages weren't truncated (cl_page_own() returns
246 * error in the latter case).
248 spin_unlock(&hdr->coh_page_guard);
251 for (i = 0; i < j; ++i) {
253 if (res == CLP_GANG_OKAY)
254 res = (*cb)(env, io, page, cbdata);
255 lu_ref_del(&page->cp_reference,
256 "gang_lookup", cfs_current());
257 cl_page_put(env, page);
259 if (nr < CLT_PVEC_SIZE || end_of_region)
262 if (res == CLP_GANG_OKAY && need_resched())
263 res = CLP_GANG_RESCHED;
264 if (res != CLP_GANG_OKAY)
267 spin_lock(&hdr->coh_page_guard);
271 spin_unlock(&hdr->coh_page_guard);
274 EXPORT_SYMBOL(cl_page_gang_lookup);
276 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
278 struct cl_object *obj = page->cp_obj;
279 int pagesize = cl_object_header(obj)->coh_page_bufsize;
281 PASSERT(env, page, cfs_list_empty(&page->cp_batch));
282 PASSERT(env, page, page->cp_owner == NULL);
283 PASSERT(env, page, page->cp_req == NULL);
284 PASSERT(env, page, page->cp_parent == NULL);
285 PASSERT(env, page, page->cp_state == CPS_FREEING);
289 while (!cfs_list_empty(&page->cp_layers)) {
290 struct cl_page_slice *slice;
292 slice = cfs_list_entry(page->cp_layers.next,
293 struct cl_page_slice, cpl_linkage);
294 cfs_list_del_init(page->cp_layers.next);
295 slice->cpl_ops->cpo_fini(env, slice);
297 CS_PAGE_DEC(obj, total);
298 CS_PAGESTATE_DEC(obj, page->cp_state);
299 lu_object_ref_del_at(&obj->co_lu, &page->cp_obj_ref, "cl_page", page);
300 cl_object_put(env, obj);
301 lu_ref_fini(&page->cp_reference);
302 OBD_FREE(page, pagesize);
307 * Helper function updating page state. This is the only place in the code
308 * where cl_page::cp_state field is mutated.
310 static inline void cl_page_state_set_trust(struct cl_page *page,
311 enum cl_page_state state)
314 *(enum cl_page_state *)&page->cp_state = state;
317 static struct cl_page *cl_page_alloc(const struct lu_env *env,
318 struct cl_object *o, pgoff_t ind, struct page *vmpage,
319 enum cl_page_type type)
321 struct cl_page *page;
322 struct lu_object_header *head;
325 OBD_ALLOC_GFP(page, cl_object_header(o)->coh_page_bufsize,
329 cfs_atomic_set(&page->cp_ref, 1);
330 if (type == CPT_CACHEABLE) /* for radix tree */
331 cfs_atomic_inc(&page->cp_ref);
334 lu_object_ref_add_at(&o->co_lu, &page->cp_obj_ref, "cl_page",
336 page->cp_index = ind;
337 cl_page_state_set_trust(page, CPS_CACHED);
338 page->cp_type = type;
339 CFS_INIT_LIST_HEAD(&page->cp_layers);
340 CFS_INIT_LIST_HEAD(&page->cp_batch);
341 CFS_INIT_LIST_HEAD(&page->cp_flight);
342 mutex_init(&page->cp_mutex);
343 lu_ref_init(&page->cp_reference);
344 head = o->co_lu.lo_header;
345 cfs_list_for_each_entry(o, &head->loh_layers,
347 if (o->co_ops->coo_page_init != NULL) {
348 result = o->co_ops->coo_page_init(env, o,
351 cl_page_delete0(env, page, 0);
352 cl_page_free(env, page);
353 page = ERR_PTR(result);
359 CS_PAGE_INC(o, total);
360 CS_PAGE_INC(o, create);
361 CS_PAGESTATE_DEC(o, CPS_CACHED);
364 page = ERR_PTR(-ENOMEM);
370 * Returns a cl_page with index \a idx at the object \a o, and associated with
371 * the VM page \a vmpage.
373 * This is the main entry point into the cl_page caching interface. First, a
374 * cache (implemented as a per-object radix tree) is consulted. If page is
375 * found there, it is returned immediately. Otherwise new page is allocated
376 * and returned. In any case, additional reference to page is acquired.
378 * \see cl_object_find(), cl_lock_find()
380 static struct cl_page *cl_page_find0(const struct lu_env *env,
382 pgoff_t idx, struct page *vmpage,
383 enum cl_page_type type,
384 struct cl_page *parent)
386 struct cl_page *page = NULL;
387 struct cl_page *ghost = NULL;
388 struct cl_object_header *hdr;
391 LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
396 hdr = cl_object_header(o);
397 CS_PAGE_INC(o, lookup);
399 CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
400 idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
402 if (type == CPT_CACHEABLE) {
403 /* vmpage lock is used to protect the child/parent
405 KLASSERT(PageLocked(vmpage));
407 * cl_vmpage_page() can be called here without any locks as
409 * - "vmpage" is locked (which prevents ->private from
410 * concurrent updates), and
412 * - "o" cannot be destroyed while current thread holds a
415 page = cl_vmpage_page(vmpage, o);
418 cl_page_vmpage(env, page) == vmpage &&
419 (void *)radix_tree_lookup(&hdr->coh_tree,
428 /* allocate and initialize cl_page */
429 page = cl_page_alloc(env, o, idx, vmpage, type);
433 if (type == CPT_TRANSIENT) {
435 LASSERT(page->cp_parent == NULL);
436 page->cp_parent = parent;
437 parent->cp_child = page;
443 * XXX optimization: use radix_tree_preload() here, and change tree
444 * gfp mask to GFP_KERNEL in cl_object_header_init().
446 spin_lock(&hdr->coh_page_guard);
447 err = radix_tree_insert(&hdr->coh_tree, idx, page);
451 * Noted by Jay: a lock on \a vmpage protects cl_page_find()
452 * from this race, but
454 * 0. it's better to have cl_page interface "locally
455 * consistent" so that its correctness can be reasoned
456 * about without appealing to the (obscure world of) VM
459 * 1. handling this race allows ->coh_tree to remain
460 * consistent even when VM locking is somehow busted,
461 * which is very useful during diagnosing and debugging.
464 CL_PAGE_DEBUG(D_ERROR, env, ghost,
465 "fail to insert into radix tree: %d\n", err);
468 LASSERT(page->cp_parent == NULL);
469 page->cp_parent = parent;
470 parent->cp_child = page;
474 spin_unlock(&hdr->coh_page_guard);
476 if (unlikely(ghost != NULL)) {
477 cl_page_delete0(env, ghost, 0);
478 cl_page_free(env, ghost);
483 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o,
484 pgoff_t idx, struct page *vmpage,
485 enum cl_page_type type)
487 return cl_page_find0(env, o, idx, vmpage, type, NULL);
489 EXPORT_SYMBOL(cl_page_find);
492 struct cl_page *cl_page_find_sub(const struct lu_env *env, struct cl_object *o,
493 pgoff_t idx, struct page *vmpage,
494 struct cl_page *parent)
496 return cl_page_find0(env, o, idx, vmpage, parent->cp_type, parent);
498 EXPORT_SYMBOL(cl_page_find_sub);
500 static inline int cl_page_invariant(const struct cl_page *pg)
502 struct cl_object_header *header;
503 struct cl_page *parent;
504 struct cl_page *child;
508 * Page invariant is protected by a VM lock.
510 LINVRNT(cl_page_is_vmlocked(NULL, pg));
512 header = cl_object_header(pg->cp_obj);
513 parent = pg->cp_parent;
514 child = pg->cp_child;
515 owner = pg->cp_owner;
517 return cl_page_in_use_noref(pg) &&
518 ergo(parent != NULL, parent->cp_child == pg) &&
519 ergo(child != NULL, child->cp_parent == pg) &&
520 ergo(child != NULL, pg->cp_obj != child->cp_obj) &&
521 ergo(parent != NULL, pg->cp_obj != parent->cp_obj) &&
522 ergo(owner != NULL && parent != NULL,
523 parent->cp_owner == pg->cp_owner->ci_parent) &&
524 ergo(owner != NULL && child != NULL,
525 child->cp_owner->ci_parent == owner) &&
527 * Either page is early in initialization (has neither child
528 * nor parent yet), or it is in the object radix tree.
530 ergo(pg->cp_state < CPS_FREEING && pg->cp_type == CPT_CACHEABLE,
531 (void *)radix_tree_lookup(&header->coh_tree,
532 pg->cp_index) == pg ||
533 (child == NULL && parent == NULL));
536 static void cl_page_state_set0(const struct lu_env *env,
537 struct cl_page *page, enum cl_page_state state)
539 enum cl_page_state old;
542 * Matrix of allowed state transitions [old][new], for sanity
545 static const int allowed_transitions[CPS_NR][CPS_NR] = {
548 [CPS_OWNED] = 1, /* io finds existing cached page */
550 [CPS_PAGEOUT] = 1, /* write-out from the cache */
551 [CPS_FREEING] = 1, /* eviction on the memory pressure */
554 [CPS_CACHED] = 1, /* release to the cache */
556 [CPS_PAGEIN] = 1, /* start read immediately */
557 [CPS_PAGEOUT] = 1, /* start write immediately */
558 [CPS_FREEING] = 1, /* lock invalidation or truncate */
561 [CPS_CACHED] = 1, /* io completion */
568 [CPS_CACHED] = 1, /* io completion */
584 old = page->cp_state;
585 PASSERT(env, page, allowed_transitions[old][state]);
586 CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
587 for (; page != NULL; page = page->cp_child) {
588 PASSERT(env, page, page->cp_state == old);
590 equi(state == CPS_OWNED, page->cp_owner != NULL));
592 CS_PAGESTATE_DEC(page->cp_obj, page->cp_state);
593 CS_PAGESTATE_INC(page->cp_obj, state);
594 cl_page_state_set_trust(page, state);
599 static void cl_page_state_set(const struct lu_env *env,
600 struct cl_page *page, enum cl_page_state state)
602 cl_page_state_set0(env, page, state);
606 * Acquires an additional reference to a page.
608 * This can be called only by caller already possessing a reference to \a
611 * \see cl_object_get(), cl_lock_get().
613 void cl_page_get(struct cl_page *page)
616 cl_page_get_trust(page);
619 EXPORT_SYMBOL(cl_page_get);
622 * Releases a reference to a page.
624 * When last reference is released, page is returned to the cache, unless it
625 * is in cl_page_state::CPS_FREEING state, in which case it is immediately
628 * \see cl_object_put(), cl_lock_put().
630 void cl_page_put(const struct lu_env *env, struct cl_page *page)
632 PASSERT(env, page, cfs_atomic_read(&page->cp_ref) > !!page->cp_parent);
635 CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
636 cfs_atomic_read(&page->cp_ref));
638 if (cfs_atomic_dec_and_test(&page->cp_ref)) {
639 LASSERT(page->cp_state == CPS_FREEING);
641 LASSERT(cfs_atomic_read(&page->cp_ref) == 0);
642 PASSERT(env, page, page->cp_owner == NULL);
643 PASSERT(env, page, cfs_list_empty(&page->cp_batch));
645 * Page is no longer reachable by other threads. Tear
648 cl_page_free(env, page);
653 EXPORT_SYMBOL(cl_page_put);
656 * Returns a VM page associated with a given cl_page.
658 struct page *cl_page_vmpage(const struct lu_env *env, struct cl_page *page)
660 const struct cl_page_slice *slice;
663 * Find uppermost layer with ->cpo_vmpage() method, and return its
666 page = cl_page_top(page);
668 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
669 if (slice->cpl_ops->cpo_vmpage != NULL)
670 RETURN(slice->cpl_ops->cpo_vmpage(env, slice));
672 page = page->cp_child;
673 } while (page != NULL);
674 LBUG(); /* ->cpo_vmpage() has to be defined somewhere in the stack */
676 EXPORT_SYMBOL(cl_page_vmpage);
679 * Returns a cl_page associated with a VM page, and given cl_object.
681 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj)
684 struct cl_page *page;
687 KLASSERT(PageLocked(vmpage));
690 * NOTE: absence of races and liveness of data are guaranteed by page
691 * lock on a "vmpage". That works because object destruction has
692 * bottom-to-top pass.
696 * This loop assumes that ->private points to the top-most page. This
697 * can be rectified easily.
699 top = (struct cl_page *)vmpage->private;
703 for (page = top; page != NULL; page = page->cp_child) {
704 if (cl_object_same(page->cp_obj, obj)) {
705 cl_page_get_trust(page);
709 LASSERT(ergo(page, page->cp_type == CPT_CACHEABLE));
712 EXPORT_SYMBOL(cl_vmpage_page);
715 * Returns the top-page for a given page.
717 * \see cl_object_top(), cl_io_top()
719 struct cl_page *cl_page_top(struct cl_page *page)
721 return cl_page_top_trusted(page);
723 EXPORT_SYMBOL(cl_page_top);
725 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
726 const struct lu_device_type *dtype)
728 return cl_page_at_trusted(page, dtype);
730 EXPORT_SYMBOL(cl_page_at);
732 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
734 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...) \
736 const struct lu_env *__env = (_env); \
737 struct cl_page *__page = (_page); \
738 const struct cl_page_slice *__scan; \
740 ptrdiff_t __op = (_op); \
741 int (*__method)_proto; \
744 __page = cl_page_top(__page); \
746 cfs_list_for_each_entry(__scan, &__page->cp_layers, \
748 __method = *(void **)((char *)__scan->cpl_ops + \
750 if (__method != NULL) { \
751 __result = (*__method)(__env, __scan, \
757 __page = __page->cp_child; \
758 } while (__page != NULL && __result == 0); \
764 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...) \
766 const struct lu_env *__env = (_env); \
767 struct cl_page *__page = (_page); \
768 const struct cl_page_slice *__scan; \
769 ptrdiff_t __op = (_op); \
770 void (*__method)_proto; \
772 __page = cl_page_top(__page); \
774 cfs_list_for_each_entry(__scan, &__page->cp_layers, \
776 __method = *(void **)((char *)__scan->cpl_ops + \
778 if (__method != NULL) \
779 (*__method)(__env, __scan, \
782 __page = __page->cp_child; \
783 } while (__page != NULL); \
786 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...) \
788 const struct lu_env *__env = (_env); \
789 struct cl_page *__page = (_page); \
790 const struct cl_page_slice *__scan; \
791 ptrdiff_t __op = (_op); \
792 void (*__method)_proto; \
794 /* get to the bottom page. */ \
795 while (__page->cp_child != NULL) \
796 __page = __page->cp_child; \
798 cfs_list_for_each_entry_reverse(__scan, &__page->cp_layers, \
800 __method = *(void **)((char *)__scan->cpl_ops + \
802 if (__method != NULL) \
803 (*__method)(__env, __scan, \
806 __page = __page->cp_parent; \
807 } while (__page != NULL); \
810 static int cl_page_invoke(const struct lu_env *env,
811 struct cl_io *io, struct cl_page *page, ptrdiff_t op)
814 PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
816 RETURN(CL_PAGE_INVOKE(env, page, op,
817 (const struct lu_env *,
818 const struct cl_page_slice *, struct cl_io *),
822 static void cl_page_invoid(const struct lu_env *env,
823 struct cl_io *io, struct cl_page *page, ptrdiff_t op)
826 PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
828 CL_PAGE_INVOID(env, page, op,
829 (const struct lu_env *,
830 const struct cl_page_slice *, struct cl_io *), io);
834 static void cl_page_owner_clear(struct cl_page *page)
837 for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
838 if (page->cp_owner != NULL) {
839 LASSERT(page->cp_owner->ci_owned_nr > 0);
840 page->cp_owner->ci_owned_nr--;
841 page->cp_owner = NULL;
842 page->cp_task = NULL;
848 static void cl_page_owner_set(struct cl_page *page)
851 for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
852 LASSERT(page->cp_owner != NULL);
853 page->cp_owner->ci_owned_nr++;
858 void cl_page_disown0(const struct lu_env *env,
859 struct cl_io *io, struct cl_page *pg)
861 enum cl_page_state state;
864 state = pg->cp_state;
865 PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
866 PINVRNT(env, pg, cl_page_invariant(pg) || state == CPS_FREEING);
867 cl_page_owner_clear(pg);
869 if (state == CPS_OWNED)
870 cl_page_state_set(env, pg, CPS_CACHED);
872 * Completion call-backs are executed in the bottom-up order, so that
873 * uppermost layer (llite), responsible for VFS/VM interaction runs
874 * last and can release locks safely.
876 CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
877 (const struct lu_env *,
878 const struct cl_page_slice *, struct cl_io *),
884 * returns true, iff page is owned by the given io.
886 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
888 LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
890 RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == io);
892 EXPORT_SYMBOL(cl_page_is_owned);
895 * Try to own a page by IO.
897 * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
898 * into cl_page_state::CPS_OWNED state.
900 * \pre !cl_page_is_owned(pg, io)
901 * \post result == 0 iff cl_page_is_owned(pg, io)
905 * \retval -ve failure, e.g., page was destroyed (and landed in
906 * cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
907 * or, page was owned by another thread, or in IO.
909 * \see cl_page_disown()
910 * \see cl_page_operations::cpo_own()
911 * \see cl_page_own_try()
914 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
915 struct cl_page *pg, int nonblock)
919 PINVRNT(env, pg, !cl_page_is_owned(pg, io));
922 pg = cl_page_top(pg);
925 if (pg->cp_state == CPS_FREEING) {
928 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
929 (const struct lu_env *,
930 const struct cl_page_slice *,
931 struct cl_io *, int),
934 PASSERT(env, pg, pg->cp_owner == NULL);
935 PASSERT(env, pg, pg->cp_req == NULL);
937 pg->cp_task = current;
938 cl_page_owner_set(pg);
939 if (pg->cp_state != CPS_FREEING) {
940 cl_page_state_set(env, pg, CPS_OWNED);
942 cl_page_disown0(env, io, pg);
947 PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
952 * Own a page, might be blocked.
954 * \see cl_page_own0()
956 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
958 return cl_page_own0(env, io, pg, 0);
960 EXPORT_SYMBOL(cl_page_own);
963 * Nonblock version of cl_page_own().
965 * \see cl_page_own0()
967 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
970 return cl_page_own0(env, io, pg, 1);
972 EXPORT_SYMBOL(cl_page_own_try);
976 * Assume page ownership.
978 * Called when page is already locked by the hosting VM.
980 * \pre !cl_page_is_owned(pg, io)
981 * \post cl_page_is_owned(pg, io)
983 * \see cl_page_operations::cpo_assume()
985 void cl_page_assume(const struct lu_env *env,
986 struct cl_io *io, struct cl_page *pg)
988 PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
991 pg = cl_page_top(pg);
994 cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
995 PASSERT(env, pg, pg->cp_owner == NULL);
997 pg->cp_task = current;
998 cl_page_owner_set(pg);
999 cl_page_state_set(env, pg, CPS_OWNED);
1002 EXPORT_SYMBOL(cl_page_assume);
1005 * Releases page ownership without unlocking the page.
1007 * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
1008 * underlying VM page (as VM is supposed to do this itself).
1010 * \pre cl_page_is_owned(pg, io)
1011 * \post !cl_page_is_owned(pg, io)
1013 * \see cl_page_assume()
1015 void cl_page_unassume(const struct lu_env *env,
1016 struct cl_io *io, struct cl_page *pg)
1018 PINVRNT(env, pg, cl_page_is_owned(pg, io));
1019 PINVRNT(env, pg, cl_page_invariant(pg));
1022 pg = cl_page_top(pg);
1024 cl_page_owner_clear(pg);
1025 cl_page_state_set(env, pg, CPS_CACHED);
1026 CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
1027 (const struct lu_env *,
1028 const struct cl_page_slice *, struct cl_io *),
1032 EXPORT_SYMBOL(cl_page_unassume);
1035 * Releases page ownership.
1037 * Moves page into cl_page_state::CPS_CACHED.
1039 * \pre cl_page_is_owned(pg, io)
1040 * \post !cl_page_is_owned(pg, io)
1042 * \see cl_page_own()
1043 * \see cl_page_operations::cpo_disown()
1045 void cl_page_disown(const struct lu_env *env,
1046 struct cl_io *io, struct cl_page *pg)
1048 PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
1049 pg->cp_state == CPS_FREEING);
1052 pg = cl_page_top(pg);
1054 cl_page_disown0(env, io, pg);
1057 EXPORT_SYMBOL(cl_page_disown);
1060 * Called when page is to be removed from the object, e.g., as a result of
1063 * Calls cl_page_operations::cpo_discard() top-to-bottom.
1065 * \pre cl_page_is_owned(pg, io)
1067 * \see cl_page_operations::cpo_discard()
1069 void cl_page_discard(const struct lu_env *env,
1070 struct cl_io *io, struct cl_page *pg)
1072 PINVRNT(env, pg, cl_page_is_owned(pg, io));
1073 PINVRNT(env, pg, cl_page_invariant(pg));
1075 cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
1077 EXPORT_SYMBOL(cl_page_discard);
1080 * Version of cl_page_delete() that can be called for not fully constructed
1081 * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
1082 * path. Doesn't check page invariant.
1084 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
1087 struct cl_page *tmp = pg;
1090 PASSERT(env, pg, pg == cl_page_top(pg));
1091 PASSERT(env, pg, pg->cp_state != CPS_FREEING);
1094 * Severe all ways to obtain new pointers to @pg.
1096 cl_page_owner_clear(pg);
1099 * unexport the page firstly before freeing it so that
1100 * the page content is considered to be invalid.
1101 * We have to do this because a CPS_FREEING cl_page may
1102 * be NOT under the protection of a cl_lock.
1103 * Afterwards, if this page is found by other threads, then this
1104 * page will be forced to reread.
1106 cl_page_export(env, pg, 0);
1107 cl_page_state_set0(env, pg, CPS_FREEING);
1109 CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete),
1110 (const struct lu_env *, const struct cl_page_slice *));
1112 if (tmp->cp_type == CPT_CACHEABLE) {
1114 /* !radix means that @pg is not yet in the radix tree,
1118 for (; tmp != NULL; tmp = tmp->cp_child) {
1120 struct cl_object_header *hdr;
1122 hdr = cl_object_header(tmp->cp_obj);
1123 spin_lock(&hdr->coh_page_guard);
1124 value = radix_tree_delete(&hdr->coh_tree,
1126 PASSERT(env, tmp, value == tmp);
1127 PASSERT(env, tmp, hdr->coh_pages > 0);
1129 spin_unlock(&hdr->coh_page_guard);
1130 cl_page_put(env, tmp);
1138 * Called when a decision is made to throw page out of memory.
1140 * Notifies all layers about page destruction by calling
1141 * cl_page_operations::cpo_delete() method top-to-bottom.
1143 * Moves page into cl_page_state::CPS_FREEING state (this is the only place
1144 * where transition to this state happens).
1146 * Eliminates all venues through which new references to the page can be
1149 * - removes page from the radix trees,
1151 * - breaks linkage from VM page to cl_page.
1153 * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
1154 * drain after some time, at which point page will be recycled.
1156 * \pre pg == cl_page_top(pg)
1157 * \pre VM page is locked
1158 * \post pg->cp_state == CPS_FREEING
1160 * \see cl_page_operations::cpo_delete()
1162 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
1164 PINVRNT(env, pg, cl_page_invariant(pg));
1166 cl_page_delete0(env, pg, 1);
1169 EXPORT_SYMBOL(cl_page_delete);
1172 * Unmaps page from user virtual memory.
1174 * Calls cl_page_operations::cpo_unmap() through all layers top-to-bottom. The
1175 * layer responsible for VM interaction has to unmap page from user space
1178 * \see cl_page_operations::cpo_unmap()
1180 int cl_page_unmap(const struct lu_env *env,
1181 struct cl_io *io, struct cl_page *pg)
1183 PINVRNT(env, pg, cl_page_is_owned(pg, io));
1184 PINVRNT(env, pg, cl_page_invariant(pg));
1186 return cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_unmap));
1188 EXPORT_SYMBOL(cl_page_unmap);
1191 * Marks page up-to-date.
1193 * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
1194 * layer responsible for VM interaction has to mark/clear page as up-to-date
1195 * by the \a uptodate argument.
1197 * \see cl_page_operations::cpo_export()
1199 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
1201 PINVRNT(env, pg, cl_page_invariant(pg));
1202 CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
1203 (const struct lu_env *,
1204 const struct cl_page_slice *, int), uptodate);
1206 EXPORT_SYMBOL(cl_page_export);
1209 * Returns true, iff \a pg is VM locked in a suitable sense by the calling
1212 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
1215 const struct cl_page_slice *slice;
1218 pg = cl_page_top_trusted((struct cl_page *)pg);
1219 slice = container_of(pg->cp_layers.next,
1220 const struct cl_page_slice, cpl_linkage);
1221 PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
1223 * Call ->cpo_is_vmlocked() directly instead of going through
1224 * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
1225 * cl_page_invariant().
1227 result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
1228 PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
1229 RETURN(result == -EBUSY);
1231 EXPORT_SYMBOL(cl_page_is_vmlocked);
1233 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
1236 RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
1239 static void cl_page_io_start(const struct lu_env *env,
1240 struct cl_page *pg, enum cl_req_type crt)
1243 * Page is queued for IO, change its state.
1246 cl_page_owner_clear(pg);
1247 cl_page_state_set(env, pg, cl_req_type_state(crt));
1252 * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
1253 * called top-to-bottom. Every layer either agrees to submit this page (by
1254 * returning 0), or requests to omit this page (by returning -EALREADY). Layer
1255 * handling interactions with the VM also has to inform VM that page is under
1258 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
1259 struct cl_page *pg, enum cl_req_type crt)
1263 PINVRNT(env, pg, cl_page_is_owned(pg, io));
1264 PINVRNT(env, pg, cl_page_invariant(pg));
1265 PINVRNT(env, pg, crt < CRT_NR);
1268 * XXX this has to be called bottom-to-top, so that llite can set up
1269 * PG_writeback without risking other layers deciding to skip this
1274 result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
1276 cl_page_io_start(env, pg, crt);
1278 KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
1280 PageWriteback(cl_page_vmpage(env, pg)))));
1281 CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1284 EXPORT_SYMBOL(cl_page_prep);
1287 * Notify layers about transfer completion.
1289 * Invoked by transfer sub-system (which is a part of osc) to notify layers
1290 * that a transfer, of which this page is a part of has completed.
1292 * Completion call-backs are executed in the bottom-up order, so that
1293 * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1294 * and can release locks safely.
1296 * \pre pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1297 * \post pg->cp_state == CPS_CACHED
1299 * \see cl_page_operations::cpo_completion()
1301 void cl_page_completion(const struct lu_env *env,
1302 struct cl_page *pg, enum cl_req_type crt, int ioret)
1304 struct cl_sync_io *anchor = pg->cp_sync_io;
1306 PASSERT(env, pg, crt < CRT_NR);
1307 /* cl_page::cp_req already cleared by the caller (osc_completion()) */
1308 PASSERT(env, pg, pg->cp_req == NULL);
1309 PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
1312 CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
1313 if (crt == CRT_READ && ioret == 0) {
1314 PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED));
1315 pg->cp_flags |= CPF_READ_COMPLETED;
1318 cl_page_state_set(env, pg, CPS_CACHED);
1321 CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
1322 (const struct lu_env *,
1323 const struct cl_page_slice *, int), ioret);
1325 LASSERT(cl_page_is_vmlocked(env, pg));
1326 LASSERT(pg->cp_sync_io == anchor);
1327 pg->cp_sync_io = NULL;
1330 * As page->cp_obj is pinned by a reference from page->cp_req, it is
1331 * safe to call cl_page_put() without risking object destruction in a
1332 * non-blocking context.
1334 cl_page_put(env, pg);
1337 cl_sync_io_note(anchor, ioret);
1341 EXPORT_SYMBOL(cl_page_completion);
1344 * Notify layers that transfer formation engine decided to yank this page from
1345 * the cache and to make it a part of a transfer.
1347 * \pre pg->cp_state == CPS_CACHED
1348 * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1350 * \see cl_page_operations::cpo_make_ready()
1352 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
1353 enum cl_req_type crt)
1357 PINVRNT(env, pg, crt < CRT_NR);
1362 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
1363 (const struct lu_env *,
1364 const struct cl_page_slice *));
1366 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
1367 cl_page_io_start(env, pg, crt);
1369 CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1372 EXPORT_SYMBOL(cl_page_make_ready);
1375 * Notify layers that high level io decided to place this page into a cache
1376 * for future transfer.
1378 * The layer implementing transfer engine (osc) has to register this page in
1381 * \pre cl_page_is_owned(pg, io)
1382 * \post cl_page_is_owned(pg, io)
1384 * \see cl_page_operations::cpo_cache_add()
1386 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
1387 struct cl_page *pg, enum cl_req_type crt)
1389 const struct cl_page_slice *scan;
1392 PINVRNT(env, pg, crt < CRT_NR);
1393 PINVRNT(env, pg, cl_page_is_owned(pg, io));
1394 PINVRNT(env, pg, cl_page_invariant(pg));
1401 cfs_list_for_each_entry(scan, &pg->cp_layers, cpl_linkage) {
1402 if (scan->cpl_ops->io[crt].cpo_cache_add == NULL)
1405 result = scan->cpl_ops->io[crt].cpo_cache_add(env, scan, io);
1409 CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1412 EXPORT_SYMBOL(cl_page_cache_add);
1415 * Called if a pge is being written back by kernel's intention.
1417 * \pre cl_page_is_owned(pg, io)
1418 * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT)
1420 * \see cl_page_operations::cpo_flush()
1422 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
1427 PINVRNT(env, pg, cl_page_is_owned(pg, io));
1428 PINVRNT(env, pg, cl_page_invariant(pg));
1432 result = cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_flush));
1434 CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result);
1437 EXPORT_SYMBOL(cl_page_flush);
1440 * Checks whether page is protected by any extent lock is at least required
1443 * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1444 * \see cl_page_operations::cpo_is_under_lock()
1446 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1447 struct cl_page *page)
1451 PINVRNT(env, page, cl_page_invariant(page));
1454 rc = CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1455 (const struct lu_env *,
1456 const struct cl_page_slice *, struct cl_io *),
1458 PASSERT(env, page, rc != 0);
1461 EXPORT_SYMBOL(cl_page_is_under_lock);
1463 static int page_prune_cb(const struct lu_env *env, struct cl_io *io,
1464 struct cl_page *page, void *cbdata)
1466 cl_page_own(env, io, page);
1467 cl_page_unmap(env, io, page);
1468 cl_page_discard(env, io, page);
1469 cl_page_disown(env, io, page);
1470 return CLP_GANG_OKAY;
1474 * Purges all cached pages belonging to the object \a obj.
1476 int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj)
1478 struct cl_thread_info *info;
1479 struct cl_object *obj = cl_object_top(clobj);
1484 info = cl_env_info(env);
1488 * initialize the io. This is ugly since we never do IO in this
1489 * function, we just make cl_page_list functions happy. -jay
1492 io->ci_ignore_layout = 1;
1493 result = cl_io_init(env, io, CIT_MISC, obj);
1495 cl_io_fini(env, io);
1496 RETURN(io->ci_result);
1500 result = cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF,
1501 page_prune_cb, NULL);
1502 if (result == CLP_GANG_RESCHED)
1504 } while (result != CLP_GANG_OKAY);
1506 cl_io_fini(env, io);
1509 EXPORT_SYMBOL(cl_pages_prune);
1512 * Tells transfer engine that only part of a page is to be transmitted.
1514 * \see cl_page_operations::cpo_clip()
1516 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1519 PINVRNT(env, pg, cl_page_invariant(pg));
1521 CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1522 CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1523 (const struct lu_env *,
1524 const struct cl_page_slice *,int, int),
1527 EXPORT_SYMBOL(cl_page_clip);
1530 * Prints human readable representation of \a pg to the \a f.
1532 void cl_page_header_print(const struct lu_env *env, void *cookie,
1533 lu_printer_t printer, const struct cl_page *pg)
1535 (*printer)(env, cookie,
1536 "page@%p[%d %p:%lu ^%p_%p %d %d %d %p %p %#x]\n",
1537 pg, cfs_atomic_read(&pg->cp_ref), pg->cp_obj,
1538 pg->cp_index, pg->cp_parent, pg->cp_child,
1539 pg->cp_state, pg->cp_error, pg->cp_type,
1540 pg->cp_owner, pg->cp_req, pg->cp_flags);
1542 EXPORT_SYMBOL(cl_page_header_print);
1545 * Prints human readable representation of \a pg to the \a f.
1547 void cl_page_print(const struct lu_env *env, void *cookie,
1548 lu_printer_t printer, const struct cl_page *pg)
1550 struct cl_page *scan;
1552 for (scan = cl_page_top((struct cl_page *)pg);
1553 scan != NULL; scan = scan->cp_child)
1554 cl_page_header_print(env, cookie, printer, scan);
1555 CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1556 (const struct lu_env *env,
1557 const struct cl_page_slice *slice,
1558 void *cookie, lu_printer_t p), cookie, printer);
1559 (*printer)(env, cookie, "end page@%p\n", pg);
1561 EXPORT_SYMBOL(cl_page_print);
1564 * Cancel a page which is still in a transfer.
1566 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1568 return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1569 (const struct lu_env *,
1570 const struct cl_page_slice *));
1572 EXPORT_SYMBOL(cl_page_cancel);
1575 * Converts a byte offset within object \a obj into a page index.
1577 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1579 return (loff_t)idx << PAGE_CACHE_SHIFT;
1581 EXPORT_SYMBOL(cl_offset);
1584 * Converts a page index into a byte offset within object \a obj.
1586 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1588 return offset >> PAGE_CACHE_SHIFT;
1590 EXPORT_SYMBOL(cl_index);
1592 int cl_page_size(const struct cl_object *obj)
1594 return 1 << PAGE_CACHE_SHIFT;
1596 EXPORT_SYMBOL(cl_page_size);
1599 * Adds page slice to the compound page.
1601 * This is called by cl_object_operations::coo_page_init() methods to add a
1602 * per-layer state to the page. New state is added at the end of
1603 * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1605 * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1607 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1608 struct cl_object *obj,
1609 const struct cl_page_operations *ops)
1612 cfs_list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1613 slice->cpl_obj = obj;
1614 slice->cpl_ops = ops;
1615 slice->cpl_page = page;
1618 EXPORT_SYMBOL(cl_page_slice_add);
1620 int cl_page_init(void)
1625 void cl_page_fini(void)