1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved.
30 * Use is subject to license terms.
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 # define EXPORT_SYMTAB
46 #include <libcfs/libcfs.h>
47 #include <obd_class.h>
48 #include <obd_support.h>
49 #include <libcfs/list.h>
51 #include <cl_object.h>
52 #include "cl_internal.h"
54 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
57 static cfs_mem_cache_t *cl_page_kmem = NULL;
59 static struct lu_kmem_descr cl_page_caches[] = {
61 .ckd_cache = &cl_page_kmem,
62 .ckd_name = "cl_page_kmem",
63 .ckd_size = sizeof (struct cl_page)
71 # define PASSERT(env, page, expr) \
73 if (unlikely(!(expr))) { \
74 CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n"); \
78 #else /* !LIBCFS_DEBUG */
79 # define PASSERT(env, page, exp) \
80 ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
81 #endif /* !LIBCFS_DEBUG */
83 #ifdef INVARIANT_CHECK
84 # define PINVRNT(env, page, expr) \
86 if (unlikely(!(expr))) { \
87 CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n"); \
91 #else /* !INVARIANT_CHECK */
92 # define PINVRNT(env, page, exp) \
93 ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
94 #endif /* !INVARIANT_CHECK */
97 * Internal version of cl_page_top, it should be called with page referenced,
98 * or coh_page_guard held.
100 static struct cl_page *cl_page_top_trusted(struct cl_page *page)
102 LASSERT(cl_is_page(page));
103 while (page->cp_parent != NULL)
104 page = page->cp_parent;
109 * Internal version of cl_page_get().
111 * This function can be used to obtain initial reference to previously
112 * unreferenced cached object. It can be called only if concurrent page
113 * reclamation is somehow prevented, e.g., by locking page radix-tree
114 * (cl_object_header::hdr->coh_page_guard), or 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(cl_is_page(page));
123 * Checkless version for trusted users.
125 if (cfs_atomic_inc_return(&page->cp_ref) == 1)
126 cfs_atomic_inc(&cl_object_site(page->cp_obj)->cs_pages.cs_busy);
130 * Returns a slice within a page, corresponding to the given layer in the
135 static const struct cl_page_slice *
136 cl_page_at_trusted(const struct cl_page *page,
137 const struct lu_device_type *dtype)
139 const struct cl_page_slice *slice;
141 #ifdef INVARIANT_CHECK
142 struct cl_object_header *ch = cl_object_header(page->cp_obj);
144 if (!cfs_atomic_read(&page->cp_ref))
145 LASSERT_SPIN_LOCKED(&ch->coh_page_guard);
149 page = cl_page_top_trusted((struct cl_page *)page);
151 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
152 if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
155 page = page->cp_child;
156 } while (page != NULL);
161 * Returns a page with given index in the given object, or NULL if no page is
162 * found. Acquires a reference on \a page.
164 * Locking: called under cl_object_header::coh_page_guard spin-lock.
166 struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index)
168 struct cl_page *page;
170 LASSERT_SPIN_LOCKED(&hdr->coh_page_guard);
172 page = radix_tree_lookup(&hdr->coh_tree, index);
174 LASSERT(cl_is_page(page));
175 cl_page_get_trust(page);
179 EXPORT_SYMBOL(cl_page_lookup);
182 * Returns a list of pages by a given [start, end] of \a obj.
184 * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
185 * crucial in the face of [offset, EOF] locks.
187 void cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj,
188 struct cl_io *io, pgoff_t start, pgoff_t end,
189 struct cl_page_list *queue, int nonblock)
191 struct cl_object_header *hdr;
192 struct cl_page *page;
193 struct cl_page **pvec;
194 const struct cl_page_slice *slice;
195 const struct lu_device_type *dtype;
200 int (*page_own)(const struct lu_env *env,
205 page_own = nonblock ? cl_page_own_try : cl_page_own;
208 hdr = cl_object_header(obj);
209 pvec = cl_env_info(env)->clt_pvec;
210 dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type;
211 cfs_spin_lock(&hdr->coh_page_guard);
212 while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec,
213 idx, CLT_PVEC_SIZE)) > 0) {
214 idx = pvec[nr - 1]->cp_index + 1;
215 for (i = 0, j = 0; i < nr; ++i) {
217 PASSERT(env, page, cl_is_page(page));
219 if (page->cp_index > end)
221 if (page->cp_state == CPS_FREEING)
223 if (page->cp_type == CPT_TRANSIENT) {
224 /* God, we found a transient page!*/
228 slice = cl_page_at_trusted(page, dtype);
230 * Pages for lsm-less file has no underneath sub-page
231 * for osc, in case of ...
233 PASSERT(env, page, slice != NULL);
235 page = slice->cpl_page;
237 * Can safely call cl_page_get_trust() under
238 * radix-tree spin-lock.
240 * XXX not true, because @page is from object another
241 * than @hdr and protected by different tree lock.
243 cl_page_get_trust(page);
244 lu_ref_add_atomic(&page->cp_reference,
245 "page_list", cfs_current());
250 * Here a delicate locking dance is performed. Current thread
251 * holds a reference to a page, but has to own it before it
252 * can be placed into queue. Owning implies waiting, so
253 * radix-tree lock is to be released. After a wait one has to
254 * check that pages weren't truncated (cl_page_own() returns
255 * error in the latter case).
257 cfs_spin_unlock(&hdr->coh_page_guard);
258 for (i = 0; i < j; ++i) {
260 if (page_own(env, io, page) == 0)
261 cl_page_list_add(queue, page);
262 lu_ref_del(&page->cp_reference,
263 "page_list", cfs_current());
264 cl_page_put(env, page);
266 cfs_spin_lock(&hdr->coh_page_guard);
267 if (nr < CLT_PVEC_SIZE)
270 cfs_spin_unlock(&hdr->coh_page_guard);
273 EXPORT_SYMBOL(cl_page_gang_lookup);
275 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
277 struct cl_object *obj = page->cp_obj;
278 struct cl_site *site = cl_object_site(obj);
280 PASSERT(env, page, cl_is_page(page));
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 cfs_atomic_dec(&site->cs_pages.cs_total);
298 cfs_atomic_dec(&site->cs_pages_state[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_SLAB_FREE_PTR(page, cl_page_kmem);
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 int cl_page_alloc(const struct lu_env *env, struct cl_object *o,
318 pgoff_t ind, struct page *vmpage,
319 enum cl_page_type type, struct cl_page **out)
321 struct cl_page *page;
322 struct cl_page *err = NULL;
323 struct lu_object_header *head;
324 struct cl_site *site = cl_object_site(o);
329 OBD_SLAB_ALLOC_PTR_GFP(page, cl_page_kmem, CFS_ALLOC_IO);
331 cfs_atomic_set(&page->cp_ref, 1);
334 page->cp_obj_ref = lu_object_ref_add(&o->co_lu,
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 cfs_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 err = o->co_ops->coo_page_init(env, o,
351 cl_page_state_set_trust(page,
353 cl_page_free(env, page);
360 cfs_atomic_inc(&site->cs_pages.cs_busy);
361 cfs_atomic_inc(&site->cs_pages.cs_total);
362 cfs_atomic_inc(&site->cs_pages_state[CPS_CACHED]);
363 cfs_atomic_inc(&site->cs_pages.cs_created);
367 page = ERR_PTR(-ENOMEM);
373 * Returns a cl_page with index \a idx at the object \a o, and associated with
374 * the VM page \a vmpage.
376 * This is the main entry point into the cl_page caching interface. First, a
377 * cache (implemented as a per-object radix tree) is consulted. If page is
378 * found there, it is returned immediately. Otherwise new page is allocated
379 * and returned. In any case, additional reference to page is acquired.
381 * \see cl_object_find(), cl_lock_find()
383 static struct cl_page *cl_page_find0(const struct lu_env *env,
385 pgoff_t idx, struct page *vmpage,
386 enum cl_page_type type,
387 struct cl_page *parent)
389 struct cl_page *page;
390 struct cl_page *ghost = NULL;
391 struct cl_object_header *hdr;
392 struct cl_site *site = cl_object_site(o);
395 LINVRNT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
400 hdr = cl_object_header(o);
401 cfs_atomic_inc(&site->cs_pages.cs_lookup);
403 CDEBUG(D_PAGE, "%lu@"DFID" %p %lu %i\n",
404 idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
406 if (type == CPT_CACHEABLE) {
408 * cl_vmpage_page() can be called here without any locks as
410 * - "vmpage" is locked (which prevents ->private from
411 * concurrent updates), and
413 * - "o" cannot be destroyed while current thread holds a
416 page = cl_vmpage_page(vmpage, o);
419 cl_page_vmpage(env, page) == vmpage &&
420 (void *)radix_tree_lookup(&hdr->coh_tree,
423 cfs_spin_lock(&hdr->coh_page_guard);
424 page = cl_page_lookup(hdr, idx);
425 cfs_spin_unlock(&hdr->coh_page_guard);
428 cfs_atomic_inc(&site->cs_pages.cs_hit);
432 /* allocate and initialize cl_page */
433 err = cl_page_alloc(env, o, idx, vmpage, type, &page);
437 * XXX optimization: use radix_tree_preload() here, and change tree
438 * gfp mask to GFP_KERNEL in cl_object_header_init().
440 cfs_spin_lock(&hdr->coh_page_guard);
441 err = radix_tree_insert(&hdr->coh_tree, idx, page);
445 * Noted by Jay: a lock on \a vmpage protects cl_page_find()
446 * from this race, but
448 * 0. it's better to have cl_page interface "locally
449 * consistent" so that its correctness can be reasoned
450 * about without appealing to the (obscure world of) VM
453 * 1. handling this race allows ->coh_tree to remain
454 * consistent even when VM locking is somehow busted,
455 * which is very useful during diagnosing and debugging.
458 if (err == -EEXIST) {
460 * XXX in case of a lookup for CPT_TRANSIENT page,
461 * nothing protects a CPT_CACHEABLE page from being
462 * concurrently moved into CPS_FREEING state.
464 page = cl_page_lookup(hdr, idx);
465 PASSERT(env, page, page != NULL);
466 if (page->cp_type == CPT_TRANSIENT &&
467 type == CPT_CACHEABLE) {
468 /* XXX: We should make sure that inode sem
469 * keeps being held in the lifetime of
470 * transient pages, so it is impossible to
471 * have conflicting transient pages.
473 cfs_spin_unlock(&hdr->coh_page_guard);
474 cl_page_put(env, page);
475 cfs_spin_lock(&hdr->coh_page_guard);
476 page = ERR_PTR(-EBUSY);
481 LASSERT(page->cp_parent == NULL);
482 page->cp_parent = parent;
483 parent->cp_child = page;
487 cfs_spin_unlock(&hdr->coh_page_guard);
489 if (unlikely(ghost != NULL)) {
490 cfs_atomic_dec(&site->cs_pages.cs_busy);
491 cl_page_delete0(env, ghost, 0);
492 cl_page_free(env, ghost);
497 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o,
498 pgoff_t idx, struct page *vmpage,
499 enum cl_page_type type)
501 return cl_page_find0(env, o, idx, vmpage, type, NULL);
503 EXPORT_SYMBOL(cl_page_find);
506 struct cl_page *cl_page_find_sub(const struct lu_env *env, struct cl_object *o,
507 pgoff_t idx, struct page *vmpage,
508 struct cl_page *parent)
510 return cl_page_find0(env, o, idx, vmpage, parent->cp_type, parent);
512 EXPORT_SYMBOL(cl_page_find_sub);
514 static inline int cl_page_invariant(const struct cl_page *pg)
516 struct cl_object_header *header;
517 struct cl_page *parent;
518 struct cl_page *child;
521 LASSERT(cl_is_page(pg));
523 * Page invariant is protected by a VM lock.
525 LINVRNT(cl_page_is_vmlocked(NULL, pg));
527 header = cl_object_header(pg->cp_obj);
528 parent = pg->cp_parent;
529 child = pg->cp_child;
530 owner = pg->cp_owner;
532 return cfs_atomic_read(&pg->cp_ref) > 0 &&
533 ergo(parent != NULL, parent->cp_child == pg) &&
534 ergo(child != NULL, child->cp_parent == pg) &&
535 ergo(child != NULL, pg->cp_obj != child->cp_obj) &&
536 ergo(parent != NULL, pg->cp_obj != parent->cp_obj) &&
537 ergo(owner != NULL && parent != NULL,
538 parent->cp_owner == pg->cp_owner->ci_parent) &&
539 ergo(owner != NULL && child != NULL,
540 child->cp_owner->ci_parent == owner) &&
542 * Either page is early in initialization (has neither child
543 * nor parent yet), or it is in the object radix tree.
545 ergo(pg->cp_state < CPS_FREEING,
546 (void *)radix_tree_lookup(&header->coh_tree,
547 pg->cp_index) == pg ||
548 (child == NULL && parent == NULL));
551 static void cl_page_state_set0(const struct lu_env *env,
552 struct cl_page *page, enum cl_page_state state)
554 enum cl_page_state old;
555 struct cl_site *site = cl_object_site(page->cp_obj);
558 * Matrix of allowed state transitions [old][new], for sanity
561 static const int allowed_transitions[CPS_NR][CPS_NR] = {
564 [CPS_OWNED] = 1, /* io finds existing cached page */
566 [CPS_PAGEOUT] = 1, /* write-out from the cache */
567 [CPS_FREEING] = 1, /* eviction on the memory pressure */
570 [CPS_CACHED] = 1, /* release to the cache */
572 [CPS_PAGEIN] = 1, /* start read immediately */
573 [CPS_PAGEOUT] = 1, /* start write immediately */
574 [CPS_FREEING] = 1, /* lock invalidation or truncate */
577 [CPS_CACHED] = 1, /* io completion */
584 [CPS_CACHED] = 1, /* io completion */
600 old = page->cp_state;
601 PASSERT(env, page, allowed_transitions[old][state]);
602 CL_PAGE_HEADER(D_TRACE, env, page, "%i -> %i\n", old, state);
603 for (; page != NULL; page = page->cp_child) {
604 PASSERT(env, page, page->cp_state == old);
606 equi(state == CPS_OWNED, page->cp_owner != NULL));
608 cfs_atomic_dec(&site->cs_pages_state[page->cp_state]);
609 cfs_atomic_inc(&site->cs_pages_state[state]);
610 cl_page_state_set_trust(page, state);
615 static void cl_page_state_set(const struct lu_env *env,
616 struct cl_page *page, enum cl_page_state state)
618 PINVRNT(env, page, cl_page_invariant(page));
619 cl_page_state_set0(env, page, state);
623 * Acquires an additional reference to a page.
625 * This can be called only by caller already possessing a reference to \a
628 * \see cl_object_get(), cl_lock_get().
630 void cl_page_get(struct cl_page *page)
633 LASSERT(page->cp_state != CPS_FREEING);
634 cl_page_get_trust(page);
637 EXPORT_SYMBOL(cl_page_get);
640 * Releases a reference to a page.
642 * When last reference is released, page is returned to the cache, unless it
643 * is in cl_page_state::CPS_FREEING state, in which case it is immediately
646 * \see cl_object_put(), cl_lock_put().
648 void cl_page_put(const struct lu_env *env, struct cl_page *page)
650 struct cl_object_header *hdr;
651 struct cl_site *site = cl_object_site(page->cp_obj);
653 PASSERT(env, page, cfs_atomic_read(&page->cp_ref) > !!page->cp_parent);
656 CL_PAGE_HEADER(D_TRACE, env, page, "%i\n",
657 cfs_atomic_read(&page->cp_ref));
659 hdr = cl_object_header(cl_object_top(page->cp_obj));
660 if (cfs_atomic_dec_and_lock(&page->cp_ref, &hdr->coh_page_guard)) {
661 cfs_atomic_dec(&site->cs_pages.cs_busy);
662 /* We're going to access the page w/o a reference, but it's
663 * ok because we have grabbed the lock coh_page_guard, which
664 * means nobody is able to free this page behind us.
666 if (page->cp_state == CPS_FREEING) {
667 /* We drop the page reference and check the page state
668 * inside the coh_page_guard. So that if it gets here,
669 * it is the REALLY last reference to this page.
671 cfs_spin_unlock(&hdr->coh_page_guard);
673 LASSERT(cfs_atomic_read(&page->cp_ref) == 0);
674 PASSERT(env, page, page->cp_owner == NULL);
675 PASSERT(env, page, cfs_list_empty(&page->cp_batch));
677 * Page is no longer reachable by other threads. Tear
680 cl_page_free(env, page);
685 cfs_spin_unlock(&hdr->coh_page_guard);
690 EXPORT_SYMBOL(cl_page_put);
693 * Returns a VM page associated with a given cl_page.
695 cfs_page_t *cl_page_vmpage(const struct lu_env *env, struct cl_page *page)
697 const struct cl_page_slice *slice;
700 * Find uppermost layer with ->cpo_vmpage() method, and return its
703 page = cl_page_top(page);
705 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
706 if (slice->cpl_ops->cpo_vmpage != NULL)
707 RETURN(slice->cpl_ops->cpo_vmpage(env, slice));
709 page = page->cp_child;
710 } while (page != NULL);
711 LBUG(); /* ->cpo_vmpage() has to be defined somewhere in the stack */
713 EXPORT_SYMBOL(cl_page_vmpage);
716 * Returns a cl_page associated with a VM page, and given cl_object.
718 struct cl_page *cl_vmpage_page(cfs_page_t *vmpage, struct cl_object *obj)
720 struct cl_page *page;
721 struct cl_object_header *hdr;
724 KLASSERT(PageLocked(vmpage));
727 * NOTE: absence of races and liveness of data are guaranteed by page
728 * lock on a "vmpage". That works because object destruction has
729 * bottom-to-top pass.
733 * This loop assumes that ->private points to the top-most page. This
734 * can be rectified easily.
736 hdr = cl_object_header(cl_object_top(obj));
737 cfs_spin_lock(&hdr->coh_page_guard);
738 for (page = (void *)vmpage->private;
739 page != NULL; page = page->cp_child) {
740 if (cl_object_same(page->cp_obj, obj)) {
741 cl_page_get_trust(page);
745 cfs_spin_unlock(&hdr->coh_page_guard);
746 LASSERT(ergo(page, cl_is_page(page) && page->cp_type == CPT_CACHEABLE));
749 EXPORT_SYMBOL(cl_vmpage_page);
752 * Returns the top-page for a given page.
754 * \see cl_object_top(), cl_io_top()
756 struct cl_page *cl_page_top(struct cl_page *page)
758 return cl_page_top_trusted(page);
760 EXPORT_SYMBOL(cl_page_top);
763 * Returns true if \a addr is an address of an allocated cl_page. Used in
764 * assertions. This check is optimistically imprecise, i.e., it occasionally
765 * returns true for the incorrect addresses, but if it returns false, then the
766 * address is guaranteed to be incorrect. (Should be named cl_pagep().)
770 int cl_is_page(const void *addr)
772 return cfs_mem_is_in_cache(addr, cl_page_kmem);
774 EXPORT_SYMBOL(cl_is_page);
776 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
777 const struct lu_device_type *dtype)
779 return cl_page_at_trusted(page, dtype);
781 EXPORT_SYMBOL(cl_page_at);
783 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
785 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...) \
787 const struct lu_env *__env = (_env); \
788 struct cl_page *__page = (_page); \
789 const struct cl_page_slice *__scan; \
791 ptrdiff_t __op = (_op); \
792 int (*__method)_proto; \
795 __page = cl_page_top(__page); \
797 cfs_list_for_each_entry(__scan, &__page->cp_layers, \
799 __method = *(void **)((char *)__scan->cpl_ops + \
801 if (__method != NULL) { \
802 __result = (*__method)(__env, __scan, \
808 __page = __page->cp_child; \
809 } while (__page != NULL && __result == 0); \
815 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...) \
817 const struct lu_env *__env = (_env); \
818 struct cl_page *__page = (_page); \
819 const struct cl_page_slice *__scan; \
820 ptrdiff_t __op = (_op); \
821 void (*__method)_proto; \
823 __page = cl_page_top(__page); \
825 cfs_list_for_each_entry(__scan, &__page->cp_layers, \
827 __method = *(void **)((char *)__scan->cpl_ops + \
829 if (__method != NULL) \
830 (*__method)(__env, __scan, \
833 __page = __page->cp_child; \
834 } while (__page != NULL); \
837 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...) \
839 const struct lu_env *__env = (_env); \
840 struct cl_page *__page = (_page); \
841 const struct cl_page_slice *__scan; \
842 ptrdiff_t __op = (_op); \
843 void (*__method)_proto; \
845 /* get to the bottom page. */ \
846 while (__page->cp_child != NULL) \
847 __page = __page->cp_child; \
849 cfs_list_for_each_entry_reverse(__scan, &__page->cp_layers, \
851 __method = *(void **)((char *)__scan->cpl_ops + \
853 if (__method != NULL) \
854 (*__method)(__env, __scan, \
857 __page = __page->cp_parent; \
858 } while (__page != NULL); \
861 static int cl_page_invoke(const struct lu_env *env,
862 struct cl_io *io, struct cl_page *page, ptrdiff_t op)
865 PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
867 RETURN(CL_PAGE_INVOKE(env, page, op,
868 (const struct lu_env *,
869 const struct cl_page_slice *, struct cl_io *),
873 static void cl_page_invoid(const struct lu_env *env,
874 struct cl_io *io, struct cl_page *page, ptrdiff_t op)
877 PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
879 CL_PAGE_INVOID(env, page, op,
880 (const struct lu_env *,
881 const struct cl_page_slice *, struct cl_io *), io);
885 static void cl_page_owner_clear(struct cl_page *page)
888 for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
889 if (page->cp_owner != NULL) {
890 LASSERT(page->cp_owner->ci_owned_nr > 0);
891 page->cp_owner->ci_owned_nr--;
892 page->cp_owner = NULL;
893 page->cp_task = NULL;
899 static void cl_page_owner_set(struct cl_page *page)
902 for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
903 LASSERT(page->cp_owner != NULL);
904 page->cp_owner->ci_owned_nr++;
909 void cl_page_disown0(const struct lu_env *env,
910 struct cl_io *io, struct cl_page *pg)
912 enum cl_page_state state;
915 state = pg->cp_state;
916 PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
917 PINVRNT(env, pg, cl_page_invariant(pg));
918 cl_page_owner_clear(pg);
920 if (state == CPS_OWNED)
921 cl_page_state_set(env, pg, CPS_CACHED);
923 * Completion call-backs are executed in the bottom-up order, so that
924 * uppermost layer (llite), responsible for VFS/VM interaction runs
925 * last and can release locks safely.
927 CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
928 (const struct lu_env *,
929 const struct cl_page_slice *, struct cl_io *),
935 * returns true, iff page is owned by the given io.
937 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
939 LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
941 RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == io);
943 EXPORT_SYMBOL(cl_page_is_owned);
946 * Try to own a page by IO.
948 * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
949 * into cl_page_state::CPS_OWNED state.
951 * \pre !cl_page_is_owned(pg, io)
952 * \post result == 0 iff cl_page_is_owned(pg, io)
956 * \retval -ve failure, e.g., page was destroyed (and landed in
957 * cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
958 * or, page was owned by another thread, or in IO.
960 * \see cl_page_disown()
961 * \see cl_page_operations::cpo_own()
962 * \see cl_page_own_try()
965 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
966 struct cl_page *pg, int nonblock)
970 PINVRNT(env, pg, !cl_page_is_owned(pg, io));
973 pg = cl_page_top(pg);
976 if (pg->cp_state == CPS_FREEING) {
979 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
980 (const struct lu_env *,
981 const struct cl_page_slice *,
982 struct cl_io *, int),
985 PASSERT(env, pg, pg->cp_owner == NULL);
986 PASSERT(env, pg, pg->cp_req == NULL);
988 pg->cp_task = current;
989 cl_page_owner_set(pg);
990 if (pg->cp_state != CPS_FREEING) {
991 cl_page_state_set(env, pg, CPS_OWNED);
993 cl_page_disown0(env, io, pg);
998 PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
1003 * Own a page, might be blocked.
1005 * \see cl_page_own0()
1007 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
1009 return cl_page_own0(env, io, pg, 0);
1011 EXPORT_SYMBOL(cl_page_own);
1014 * Nonblock version of cl_page_own().
1016 * \see cl_page_own0()
1018 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
1021 return cl_page_own0(env, io, pg, 1);
1023 EXPORT_SYMBOL(cl_page_own_try);
1027 * Assume page ownership.
1029 * Called when page is already locked by the hosting VM.
1031 * \pre !cl_page_is_owned(pg, io)
1032 * \post cl_page_is_owned(pg, io)
1034 * \see cl_page_operations::cpo_assume()
1036 void cl_page_assume(const struct lu_env *env,
1037 struct cl_io *io, struct cl_page *pg)
1039 PASSERT(env, pg, pg->cp_state < CPS_OWNED);
1040 PASSERT(env, pg, pg->cp_owner == NULL);
1041 PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
1042 PINVRNT(env, pg, cl_page_invariant(pg));
1045 pg = cl_page_top(pg);
1048 cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
1050 pg->cp_task = current;
1051 cl_page_owner_set(pg);
1052 cl_page_state_set(env, pg, CPS_OWNED);
1055 EXPORT_SYMBOL(cl_page_assume);
1058 * Releases page ownership without unlocking the page.
1060 * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
1061 * underlying VM page (as VM is supposed to do this itself).
1063 * \pre cl_page_is_owned(pg, io)
1064 * \post !cl_page_is_owned(pg, io)
1066 * \see cl_page_assume()
1068 void cl_page_unassume(const struct lu_env *env,
1069 struct cl_io *io, struct cl_page *pg)
1071 PINVRNT(env, pg, cl_page_is_owned(pg, io));
1072 PINVRNT(env, pg, cl_page_invariant(pg));
1075 pg = cl_page_top(pg);
1077 cl_page_owner_clear(pg);
1078 cl_page_state_set(env, pg, CPS_CACHED);
1079 CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
1080 (const struct lu_env *,
1081 const struct cl_page_slice *, struct cl_io *),
1085 EXPORT_SYMBOL(cl_page_unassume);
1088 * Releases page ownership.
1090 * Moves page into cl_page_state::CPS_CACHED.
1092 * \pre cl_page_is_owned(pg, io)
1093 * \post !cl_page_is_owned(pg, io)
1095 * \see cl_page_own()
1096 * \see cl_page_operations::cpo_disown()
1098 void cl_page_disown(const struct lu_env *env,
1099 struct cl_io *io, struct cl_page *pg)
1101 PINVRNT(env, pg, cl_page_is_owned(pg, io));
1104 pg = cl_page_top(pg);
1106 cl_page_disown0(env, io, pg);
1109 EXPORT_SYMBOL(cl_page_disown);
1112 * Called when page is to be removed from the object, e.g., as a result of
1115 * Calls cl_page_operations::cpo_discard() top-to-bottom.
1117 * \pre cl_page_is_owned(pg, io)
1119 * \see cl_page_operations::cpo_discard()
1121 void cl_page_discard(const struct lu_env *env,
1122 struct cl_io *io, struct cl_page *pg)
1124 PINVRNT(env, pg, cl_page_is_owned(pg, io));
1125 PINVRNT(env, pg, cl_page_invariant(pg));
1127 cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
1129 EXPORT_SYMBOL(cl_page_discard);
1132 * Version of cl_page_delete() that can be called for not fully constructed
1133 * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
1134 * path. Doesn't check page invariant.
1136 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
1139 struct cl_page *tmp = pg;
1142 PASSERT(env, pg, pg == cl_page_top(pg));
1143 PASSERT(env, pg, pg->cp_state != CPS_FREEING);
1146 * Severe all ways to obtain new pointers to @pg.
1148 cl_page_owner_clear(pg);
1151 * unexport the page firstly before freeing it so that
1152 * the page content is considered to be invalid.
1153 * We have to do this because a CPS_FREEING cl_page may
1154 * be NOT under the protection of a cl_lock.
1155 * Afterwards, if this page is found by other threads, then this
1156 * page will be forced to reread.
1158 cl_page_export(env, pg, 0);
1159 cl_page_state_set0(env, pg, CPS_FREEING);
1163 * !radix means that @pg is not yet in the radix tree, skip
1167 for (; tmp != NULL; tmp = tmp->cp_child) {
1169 struct cl_object_header *hdr;
1171 hdr = cl_object_header(tmp->cp_obj);
1172 cfs_spin_lock(&hdr->coh_page_guard);
1173 value = radix_tree_delete(&hdr->coh_tree, tmp->cp_index);
1174 PASSERT(env, tmp, value == tmp);
1175 PASSERT(env, tmp, hdr->coh_pages > 0);
1177 cfs_spin_unlock(&hdr->coh_page_guard);
1180 CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete),
1181 (const struct lu_env *, const struct cl_page_slice *));
1186 * Called when a decision is made to throw page out of memory.
1188 * Notifies all layers about page destruction by calling
1189 * cl_page_operations::cpo_delete() method top-to-bottom.
1191 * Moves page into cl_page_state::CPS_FREEING state (this is the only place
1192 * where transition to this state happens).
1194 * Eliminates all venues through which new references to the page can be
1197 * - removes page from the radix trees,
1199 * - breaks linkage from VM page to cl_page.
1201 * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
1202 * drain after some time, at which point page will be recycled.
1204 * \pre pg == cl_page_top(pg)
1205 * \pre VM page is locked
1206 * \post pg->cp_state == CPS_FREEING
1208 * \see cl_page_operations::cpo_delete()
1210 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
1212 PINVRNT(env, pg, cl_page_invariant(pg));
1214 cl_page_delete0(env, pg, 1);
1217 EXPORT_SYMBOL(cl_page_delete);
1220 * Unmaps page from user virtual memory.
1222 * Calls cl_page_operations::cpo_unmap() through all layers top-to-bottom. The
1223 * layer responsible for VM interaction has to unmap page from user space
1226 * \see cl_page_operations::cpo_unmap()
1228 int cl_page_unmap(const struct lu_env *env,
1229 struct cl_io *io, struct cl_page *pg)
1231 PINVRNT(env, pg, cl_page_is_owned(pg, io));
1232 PINVRNT(env, pg, cl_page_invariant(pg));
1234 return cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_unmap));
1236 EXPORT_SYMBOL(cl_page_unmap);
1239 * Marks page up-to-date.
1241 * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
1242 * layer responsible for VM interaction has to mark/clear page as up-to-date
1243 * by the \a uptodate argument.
1245 * \see cl_page_operations::cpo_export()
1247 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
1249 PINVRNT(env, pg, cl_page_invariant(pg));
1250 CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
1251 (const struct lu_env *,
1252 const struct cl_page_slice *, int), uptodate);
1254 EXPORT_SYMBOL(cl_page_export);
1257 * Returns true, iff \a pg is VM locked in a suitable sense by the calling
1260 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
1263 const struct cl_page_slice *slice;
1266 pg = cl_page_top_trusted((struct cl_page *)pg);
1267 slice = container_of(pg->cp_layers.next,
1268 const struct cl_page_slice, cpl_linkage);
1269 PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
1271 * Call ->cpo_is_vmlocked() directly instead of going through
1272 * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
1273 * cl_page_invariant().
1275 result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
1276 PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
1277 RETURN(result == -EBUSY);
1279 EXPORT_SYMBOL(cl_page_is_vmlocked);
1281 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
1284 RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
1287 static void cl_page_io_start(const struct lu_env *env,
1288 struct cl_page *pg, enum cl_req_type crt)
1291 * Page is queued for IO, change its state.
1294 cl_page_owner_clear(pg);
1295 cl_page_state_set(env, pg, cl_req_type_state(crt));
1300 * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
1301 * called top-to-bottom. Every layer either agrees to submit this page (by
1302 * returning 0), or requests to omit this page (by returning -EALREADY). Layer
1303 * handling interactions with the VM also has to inform VM that page is under
1306 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
1307 struct cl_page *pg, enum cl_req_type crt)
1311 PINVRNT(env, pg, cl_page_is_owned(pg, io));
1312 PINVRNT(env, pg, cl_page_invariant(pg));
1313 PINVRNT(env, pg, crt < CRT_NR);
1316 * XXX this has to be called bottom-to-top, so that llite can set up
1317 * PG_writeback without risking other layers deciding to skip this
1320 result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
1322 cl_page_io_start(env, pg, crt);
1324 KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
1326 PageWriteback(cl_page_vmpage(env, pg)))));
1327 CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, result);
1330 EXPORT_SYMBOL(cl_page_prep);
1333 * Notify layers about transfer completion.
1335 * Invoked by transfer sub-system (which is a part of osc) to notify layers
1336 * that a transfer, of which this page is a part of has completed.
1338 * Completion call-backs are executed in the bottom-up order, so that
1339 * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1340 * and can release locks safely.
1342 * \pre pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1343 * \post pg->cp_state == CPS_CACHED
1345 * \see cl_page_operations::cpo_completion()
1347 void cl_page_completion(const struct lu_env *env,
1348 struct cl_page *pg, enum cl_req_type crt, int ioret)
1350 struct cl_sync_io *anchor = pg->cp_sync_io;
1352 PASSERT(env, pg, crt < CRT_NR);
1353 /* cl_page::cp_req already cleared by the caller (osc_completion()) */
1354 PASSERT(env, pg, pg->cp_req == NULL);
1355 PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
1356 PINVRNT(env, pg, cl_page_invariant(pg));
1359 CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, ioret);
1360 if (crt == CRT_READ && ioret == 0) {
1361 PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED));
1362 pg->cp_flags |= CPF_READ_COMPLETED;
1365 cl_page_state_set(env, pg, CPS_CACHED);
1366 CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
1367 (const struct lu_env *,
1368 const struct cl_page_slice *, int), ioret);
1370 LASSERT(pg->cp_sync_io == anchor);
1371 pg->cp_sync_io = NULL;
1372 cl_sync_io_note(anchor, ioret);
1375 /* Don't assert the page writeback bit here because the lustre file
1376 * may be as a backend of swap space. in this case, the page writeback
1377 * is set by VM, and obvious we shouldn't clear it at all. Fortunately
1378 * this type of pages are all TRANSIENT pages. */
1379 KLASSERT(ergo(pg->cp_type == CPT_CACHEABLE,
1380 !PageWriteback(cl_page_vmpage(env, pg))));
1383 EXPORT_SYMBOL(cl_page_completion);
1386 * Notify layers that transfer formation engine decided to yank this page from
1387 * the cache and to make it a part of a transfer.
1389 * \pre pg->cp_state == CPS_CACHED
1390 * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1392 * \see cl_page_operations::cpo_make_ready()
1394 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
1395 enum cl_req_type crt)
1399 PINVRNT(env, pg, crt < CRT_NR);
1402 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
1403 (const struct lu_env *,
1404 const struct cl_page_slice *));
1406 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
1407 cl_page_io_start(env, pg, crt);
1409 CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, result);
1412 EXPORT_SYMBOL(cl_page_make_ready);
1415 * Notify layers that high level io decided to place this page into a cache
1416 * for future transfer.
1418 * The layer implementing transfer engine (osc) has to register this page in
1421 * \pre cl_page_is_owned(pg, io)
1422 * \post ergo(result == 0,
1423 * pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT)
1425 * \see cl_page_operations::cpo_cache_add()
1427 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
1428 struct cl_page *pg, enum cl_req_type crt)
1432 PINVRNT(env, pg, crt < CRT_NR);
1433 PINVRNT(env, pg, cl_page_is_owned(pg, io));
1434 PINVRNT(env, pg, cl_page_invariant(pg));
1437 result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_cache_add));
1439 cl_page_owner_clear(pg);
1440 cl_page_state_set(env, pg, CPS_CACHED);
1442 CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, result);
1445 EXPORT_SYMBOL(cl_page_cache_add);
1448 * Checks whether page is protected by any extent lock is at least required
1451 * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1452 * \see cl_page_operations::cpo_is_under_lock()
1454 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1455 struct cl_page *page)
1459 PINVRNT(env, page, cl_page_invariant(page));
1462 rc = CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1463 (const struct lu_env *,
1464 const struct cl_page_slice *, struct cl_io *),
1466 PASSERT(env, page, rc != 0);
1469 EXPORT_SYMBOL(cl_page_is_under_lock);
1472 * Purges all cached pages belonging to the object \a obj.
1474 int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj)
1476 struct cl_thread_info *info;
1477 struct cl_object *obj = cl_object_top(clobj);
1479 struct cl_page_list *plist;
1483 info = cl_env_info(env);
1484 plist = &info->clt_list;
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 result = cl_io_init(env, io, CIT_MISC, obj);
1494 cl_io_fini(env, io);
1495 RETURN(io->ci_result);
1498 cl_page_list_init(plist);
1499 cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF, plist, 0);
1501 * Since we're purging the pages of an object, we don't care
1502 * the possible outcomes of the following functions.
1504 cl_page_list_unmap(env, io, plist);
1505 cl_page_list_discard(env, io, plist);
1506 cl_page_list_disown(env, io, plist);
1507 cl_page_list_fini(env, plist);
1509 cl_io_fini(env, io);
1512 EXPORT_SYMBOL(cl_pages_prune);
1515 * Tells transfer engine that only part of a page is to be transmitted.
1517 * \see cl_page_operations::cpo_clip()
1519 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1522 PINVRNT(env, pg, cl_page_invariant(pg));
1524 CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", from, to);
1525 CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1526 (const struct lu_env *,
1527 const struct cl_page_slice *,int, int),
1530 EXPORT_SYMBOL(cl_page_clip);
1533 * Prints human readable representation of \a pg to the \a f.
1535 void cl_page_header_print(const struct lu_env *env, void *cookie,
1536 lu_printer_t printer, const struct cl_page *pg)
1538 (*printer)(env, cookie,
1539 "page@%p[%d %p:%lu ^%p_%p %d %d %d %p %p %#x]\n",
1540 pg, cfs_atomic_read(&pg->cp_ref), pg->cp_obj,
1541 pg->cp_index, pg->cp_parent, pg->cp_child,
1542 pg->cp_state, pg->cp_error, pg->cp_type,
1543 pg->cp_owner, pg->cp_req, pg->cp_flags);
1545 EXPORT_SYMBOL(cl_page_header_print);
1548 * Prints human readable representation of \a pg to the \a f.
1550 void cl_page_print(const struct lu_env *env, void *cookie,
1551 lu_printer_t printer, const struct cl_page *pg)
1553 struct cl_page *scan;
1555 for (scan = cl_page_top((struct cl_page *)pg);
1556 scan != NULL; scan = scan->cp_child)
1557 cl_page_header_print(env, cookie, printer, scan);
1558 CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1559 (const struct lu_env *env,
1560 const struct cl_page_slice *slice,
1561 void *cookie, lu_printer_t p), cookie, printer);
1562 (*printer)(env, cookie, "end page@%p\n", pg);
1564 EXPORT_SYMBOL(cl_page_print);
1567 * Cancel a page which is still in a transfer.
1569 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1571 return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1572 (const struct lu_env *,
1573 const struct cl_page_slice *));
1575 EXPORT_SYMBOL(cl_page_cancel);
1578 * Converts a byte offset within object \a obj into a page index.
1580 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1585 return (loff_t)idx << CFS_PAGE_SHIFT;
1587 EXPORT_SYMBOL(cl_offset);
1590 * Converts a page index into a byte offset within object \a obj.
1592 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1597 return offset >> CFS_PAGE_SHIFT;
1599 EXPORT_SYMBOL(cl_index);
1601 int cl_page_size(const struct cl_object *obj)
1603 return 1 << CFS_PAGE_SHIFT;
1605 EXPORT_SYMBOL(cl_page_size);
1608 * Adds page slice to the compound page.
1610 * This is called by cl_object_operations::coo_page_init() methods to add a
1611 * per-layer state to the page. New state is added at the end of
1612 * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1614 * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1616 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1617 struct cl_object *obj,
1618 const struct cl_page_operations *ops)
1621 cfs_list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1622 slice->cpl_obj = obj;
1623 slice->cpl_ops = ops;
1624 slice->cpl_page = page;
1627 EXPORT_SYMBOL(cl_page_slice_add);
1629 int cl_page_init(void)
1631 return lu_kmem_init(cl_page_caches);
1634 void cl_page_fini(void)
1636 lu_kmem_fini(cl_page_caches);