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, 2012, Whamcloud, Inc.
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 <obd_class.h>
44 #include <obd_support.h>
45 #include <lustre_fid.h>
46 #include <libcfs/list.h>
47 /* lu_time_global_{init,fini}() */
50 #include <cl_object.h>
51 #include "cl_internal.h"
53 /*****************************************************************************
59 #define cl_io_for_each(slice, io) \
60 cfs_list_for_each_entry((slice), &io->ci_layers, cis_linkage)
61 #define cl_io_for_each_reverse(slice, io) \
62 cfs_list_for_each_entry_reverse((slice), &io->ci_layers, cis_linkage)
64 static inline int cl_io_type_is_valid(enum cl_io_type type)
66 return CIT_READ <= type && type < CIT_OP_NR;
69 static inline int cl_io_is_loopable(const struct cl_io *io)
71 return cl_io_type_is_valid(io->ci_type) && io->ci_type != CIT_MISC;
75 * Returns true iff there is an IO ongoing in the given environment.
77 int cl_io_is_going(const struct lu_env *env)
79 return cl_env_info(env)->clt_current_io != NULL;
81 EXPORT_SYMBOL(cl_io_is_going);
84 * cl_io invariant that holds at all times when exported cl_io_*() functions
85 * are entered and left.
87 static int cl_io_invariant(const struct cl_io *io)
94 * io can own pages only when it is ongoing. Sub-io might
95 * still be in CIS_LOCKED state when top-io is in
98 ergo(io->ci_owned_nr > 0, io->ci_state == CIS_IO_GOING ||
99 (io->ci_state == CIS_LOCKED && up != NULL));
103 * Finalize \a io, by calling cl_io_operations::cio_fini() bottom-to-top.
105 void cl_io_fini(const struct lu_env *env, struct cl_io *io)
107 struct cl_io_slice *slice;
108 struct cl_thread_info *info;
110 LINVRNT(cl_io_type_is_valid(io->ci_type));
111 LINVRNT(cl_io_invariant(io));
114 while (!cfs_list_empty(&io->ci_layers)) {
115 slice = container_of(io->ci_layers.next, struct cl_io_slice,
117 cfs_list_del_init(&slice->cis_linkage);
118 if (slice->cis_iop->op[io->ci_type].cio_fini != NULL)
119 slice->cis_iop->op[io->ci_type].cio_fini(env, slice);
121 * Invalidate slice to catch use after free. This assumes that
122 * slices are allocated within session and can be touched
123 * after ->cio_fini() returns.
125 slice->cis_io = NULL;
127 io->ci_state = CIS_FINI;
128 info = cl_env_info(env);
129 if (info->clt_current_io == io)
130 info->clt_current_io = NULL;
133 EXPORT_SYMBOL(cl_io_fini);
135 static int cl_io_init0(const struct lu_env *env, struct cl_io *io,
136 enum cl_io_type iot, struct cl_object *obj)
138 struct cl_object *scan;
141 LINVRNT(io->ci_state == CIS_ZERO || io->ci_state == CIS_FINI);
142 LINVRNT(cl_io_type_is_valid(iot));
143 LINVRNT(cl_io_invariant(io));
147 CFS_INIT_LIST_HEAD(&io->ci_lockset.cls_todo);
148 CFS_INIT_LIST_HEAD(&io->ci_lockset.cls_curr);
149 CFS_INIT_LIST_HEAD(&io->ci_lockset.cls_done);
150 CFS_INIT_LIST_HEAD(&io->ci_layers);
153 cl_object_for_each(scan, obj) {
154 if (scan->co_ops->coo_io_init != NULL) {
155 result = scan->co_ops->coo_io_init(env, scan, io);
161 io->ci_state = CIS_INIT;
166 * Initialize sub-io, by calling cl_io_operations::cio_init() top-to-bottom.
168 * \pre obj != cl_object_top(obj)
170 int cl_io_sub_init(const struct lu_env *env, struct cl_io *io,
171 enum cl_io_type iot, struct cl_object *obj)
173 struct cl_thread_info *info = cl_env_info(env);
175 LASSERT(obj != cl_object_top(obj));
176 if (info->clt_current_io == NULL)
177 info->clt_current_io = io;
178 return cl_io_init0(env, io, iot, obj);
180 EXPORT_SYMBOL(cl_io_sub_init);
183 * Initialize \a io, by calling cl_io_operations::cio_init() top-to-bottom.
185 * Caller has to call cl_io_fini() after a call to cl_io_init(), no matter
186 * what the latter returned.
188 * \pre obj == cl_object_top(obj)
189 * \pre cl_io_type_is_valid(iot)
190 * \post cl_io_type_is_valid(io->ci_type) && io->ci_type == iot
192 int cl_io_init(const struct lu_env *env, struct cl_io *io,
193 enum cl_io_type iot, struct cl_object *obj)
195 struct cl_thread_info *info = cl_env_info(env);
197 LASSERT(obj == cl_object_top(obj));
198 LASSERT(info->clt_current_io == NULL);
200 info->clt_current_io = io;
201 return cl_io_init0(env, io, iot, obj);
203 EXPORT_SYMBOL(cl_io_init);
206 * Initialize read or write io.
208 * \pre iot == CIT_READ || iot == CIT_WRITE
210 int cl_io_rw_init(const struct lu_env *env, struct cl_io *io,
211 enum cl_io_type iot, loff_t pos, size_t count)
213 LINVRNT(iot == CIT_READ || iot == CIT_WRITE);
214 LINVRNT(io->ci_obj != NULL);
217 LU_OBJECT_HEADER(D_VFSTRACE, env, &io->ci_obj->co_lu,
218 "io range: %u ["LPU64", "LPU64") %u %u\n",
219 iot, (__u64)pos, (__u64)pos + count,
220 io->u.ci_rw.crw_nonblock, io->u.ci_wr.wr_append);
221 io->u.ci_rw.crw_pos = pos;
222 io->u.ci_rw.crw_count = count;
223 RETURN(cl_io_init(env, io, iot, io->ci_obj));
225 EXPORT_SYMBOL(cl_io_rw_init);
227 static inline const struct lu_fid *
228 cl_lock_descr_fid(const struct cl_lock_descr *descr)
230 return lu_object_fid(&descr->cld_obj->co_lu);
233 static int cl_lock_descr_sort(const struct cl_lock_descr *d0,
234 const struct cl_lock_descr *d1)
236 return lu_fid_cmp(cl_lock_descr_fid(d0), cl_lock_descr_fid(d1)) ?:
237 __diff_normalize(d0->cld_start, d1->cld_start);
240 static int cl_lock_descr_cmp(const struct cl_lock_descr *d0,
241 const struct cl_lock_descr *d1)
245 ret = lu_fid_cmp(cl_lock_descr_fid(d0), cl_lock_descr_fid(d1));
248 if (d0->cld_end < d1->cld_start)
250 if (d0->cld_start > d0->cld_end)
255 static void cl_lock_descr_merge(struct cl_lock_descr *d0,
256 const struct cl_lock_descr *d1)
258 d0->cld_start = min(d0->cld_start, d1->cld_start);
259 d0->cld_end = max(d0->cld_end, d1->cld_end);
261 if (d1->cld_mode == CLM_WRITE && d0->cld_mode != CLM_WRITE)
262 d0->cld_mode = CLM_WRITE;
264 if (d1->cld_mode == CLM_GROUP && d0->cld_mode != CLM_GROUP)
265 d0->cld_mode = CLM_GROUP;
269 * Sort locks in lexicographical order of their (fid, start-offset) pairs.
271 static void cl_io_locks_sort(struct cl_io *io)
276 /* hidden treasure: bubble sort for now. */
278 struct cl_io_lock_link *curr;
279 struct cl_io_lock_link *prev;
280 struct cl_io_lock_link *temp;
285 cfs_list_for_each_entry_safe(curr, temp,
286 &io->ci_lockset.cls_todo,
289 switch (cl_lock_descr_sort(&prev->cill_descr,
290 &curr->cill_descr)) {
293 * IMPOSSIBLE: Identical locks are
300 cfs_list_move_tail(&curr->cill_linkage,
301 &prev->cill_linkage);
303 continue; /* don't change prev: it's
304 * still "previous" */
305 case -1: /* already in order */
316 * Check whether \a queue contains locks matching \a need.
318 * \retval +ve there is a matching lock in the \a queue
319 * \retval 0 there are no matching locks in the \a queue
321 int cl_queue_match(const cfs_list_t *queue,
322 const struct cl_lock_descr *need)
324 struct cl_io_lock_link *scan;
327 cfs_list_for_each_entry(scan, queue, cill_linkage) {
328 if (cl_lock_descr_match(&scan->cill_descr, need))
333 EXPORT_SYMBOL(cl_queue_match);
335 static int cl_queue_merge(const cfs_list_t *queue,
336 const struct cl_lock_descr *need)
338 struct cl_io_lock_link *scan;
341 cfs_list_for_each_entry(scan, queue, cill_linkage) {
342 if (cl_lock_descr_cmp(&scan->cill_descr, need))
344 cl_lock_descr_merge(&scan->cill_descr, need);
345 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
346 scan->cill_descr.cld_mode, scan->cill_descr.cld_start,
347 scan->cill_descr.cld_end);
354 static int cl_lockset_match(const struct cl_lockset *set,
355 const struct cl_lock_descr *need)
357 return cl_queue_match(&set->cls_curr, need) ||
358 cl_queue_match(&set->cls_done, need);
361 static int cl_lockset_merge(const struct cl_lockset *set,
362 const struct cl_lock_descr *need)
364 return cl_queue_merge(&set->cls_todo, need) ||
365 cl_lockset_match(set, need);
368 static int cl_lockset_lock_one(const struct lu_env *env,
369 struct cl_io *io, struct cl_lockset *set,
370 struct cl_io_lock_link *link)
372 struct cl_lock *lock;
377 if (io->ci_lockreq == CILR_PEEK) {
378 lock = cl_lock_peek(env, io, &link->cill_descr, "io", io);
380 lock = ERR_PTR(-ENODATA);
382 lock = cl_lock_request(env, io, &link->cill_descr, "io", io);
385 link->cill_lock = lock;
386 cfs_list_move(&link->cill_linkage, &set->cls_curr);
387 if (!(link->cill_descr.cld_enq_flags & CEF_ASYNC)) {
388 result = cl_wait(env, lock);
390 cfs_list_move(&link->cill_linkage,
395 result = PTR_ERR(lock);
399 static void cl_lock_link_fini(const struct lu_env *env, struct cl_io *io,
400 struct cl_io_lock_link *link)
402 struct cl_lock *lock = link->cill_lock;
405 cfs_list_del_init(&link->cill_linkage);
407 cl_lock_release(env, lock, "io", io);
408 link->cill_lock = NULL;
410 if (link->cill_fini != NULL)
411 link->cill_fini(env, link);
415 static int cl_lockset_lock(const struct lu_env *env, struct cl_io *io,
416 struct cl_lockset *set)
418 struct cl_io_lock_link *link;
419 struct cl_io_lock_link *temp;
420 struct cl_lock *lock;
425 cfs_list_for_each_entry_safe(link, temp, &set->cls_todo, cill_linkage) {
426 if (!cl_lockset_match(set, &link->cill_descr)) {
427 /* XXX some locking to guarantee that locks aren't
428 * expanded in between. */
429 result = cl_lockset_lock_one(env, io, set, link);
433 cl_lock_link_fini(env, io, link);
436 cfs_list_for_each_entry_safe(link, temp,
437 &set->cls_curr, cill_linkage) {
438 lock = link->cill_lock;
439 result = cl_wait(env, lock);
441 cfs_list_move(&link->cill_linkage,
451 * Takes locks necessary for the current iteration of io.
453 * Calls cl_io_operations::cio_lock() top-to-bottom to collect locks required
454 * by layers for the current iteration. Then sort locks (to avoid dead-locks),
457 int cl_io_lock(const struct lu_env *env, struct cl_io *io)
459 const struct cl_io_slice *scan;
462 LINVRNT(cl_io_is_loopable(io));
463 LINVRNT(io->ci_state == CIS_IT_STARTED);
464 LINVRNT(cl_io_invariant(io));
467 cl_io_for_each(scan, io) {
468 if (scan->cis_iop->op[io->ci_type].cio_lock == NULL)
470 result = scan->cis_iop->op[io->ci_type].cio_lock(env, scan);
475 cl_io_locks_sort(io);
476 result = cl_lockset_lock(env, io, &io->ci_lockset);
479 cl_io_unlock(env, io);
481 io->ci_state = CIS_LOCKED;
484 EXPORT_SYMBOL(cl_io_lock);
487 * Release locks takes by io.
489 void cl_io_unlock(const struct lu_env *env, struct cl_io *io)
491 struct cl_lockset *set;
492 struct cl_io_lock_link *link;
493 struct cl_io_lock_link *temp;
494 const struct cl_io_slice *scan;
496 LASSERT(cl_io_is_loopable(io));
497 LASSERT(CIS_IT_STARTED <= io->ci_state && io->ci_state < CIS_UNLOCKED);
498 LINVRNT(cl_io_invariant(io));
501 set = &io->ci_lockset;
503 cfs_list_for_each_entry_safe(link, temp, &set->cls_todo, cill_linkage)
504 cl_lock_link_fini(env, io, link);
506 cfs_list_for_each_entry_safe(link, temp, &set->cls_curr, cill_linkage)
507 cl_lock_link_fini(env, io, link);
509 cfs_list_for_each_entry_safe(link, temp, &set->cls_done, cill_linkage) {
510 cl_unuse(env, link->cill_lock);
511 cl_lock_link_fini(env, io, link);
513 cl_io_for_each_reverse(scan, io) {
514 if (scan->cis_iop->op[io->ci_type].cio_unlock != NULL)
515 scan->cis_iop->op[io->ci_type].cio_unlock(env, scan);
517 io->ci_state = CIS_UNLOCKED;
518 LASSERT(!cl_env_info(env)->clt_counters[CNL_TOP].ctc_nr_locks_acquired);
521 EXPORT_SYMBOL(cl_io_unlock);
524 * Prepares next iteration of io.
526 * Calls cl_io_operations::cio_iter_init() top-to-bottom. This exists to give
527 * layers a chance to modify io parameters, e.g., so that lov can restrict io
528 * to a single stripe.
530 int cl_io_iter_init(const struct lu_env *env, struct cl_io *io)
532 const struct cl_io_slice *scan;
535 LINVRNT(cl_io_is_loopable(io));
536 LINVRNT(io->ci_state == CIS_INIT || io->ci_state == CIS_IT_ENDED);
537 LINVRNT(cl_io_invariant(io));
541 cl_io_for_each(scan, io) {
542 if (scan->cis_iop->op[io->ci_type].cio_iter_init == NULL)
544 result = scan->cis_iop->op[io->ci_type].cio_iter_init(env,
550 io->ci_state = CIS_IT_STARTED;
553 EXPORT_SYMBOL(cl_io_iter_init);
556 * Finalizes io iteration.
558 * Calls cl_io_operations::cio_iter_fini() bottom-to-top.
560 void cl_io_iter_fini(const struct lu_env *env, struct cl_io *io)
562 const struct cl_io_slice *scan;
564 LINVRNT(cl_io_is_loopable(io));
565 LINVRNT(io->ci_state == CIS_UNLOCKED);
566 LINVRNT(cl_io_invariant(io));
569 cl_io_for_each_reverse(scan, io) {
570 if (scan->cis_iop->op[io->ci_type].cio_iter_fini != NULL)
571 scan->cis_iop->op[io->ci_type].cio_iter_fini(env, scan);
573 io->ci_state = CIS_IT_ENDED;
576 EXPORT_SYMBOL(cl_io_iter_fini);
579 * Records that read or write io progressed \a nob bytes forward.
581 void cl_io_rw_advance(const struct lu_env *env, struct cl_io *io, size_t nob)
583 const struct cl_io_slice *scan;
585 LINVRNT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
587 LINVRNT(cl_io_is_loopable(io));
588 LINVRNT(cl_io_invariant(io));
592 io->u.ci_rw.crw_pos += nob;
593 io->u.ci_rw.crw_count -= nob;
595 /* layers have to be notified. */
596 cl_io_for_each_reverse(scan, io) {
597 if (scan->cis_iop->op[io->ci_type].cio_advance != NULL)
598 scan->cis_iop->op[io->ci_type].cio_advance(env, scan,
603 EXPORT_SYMBOL(cl_io_rw_advance);
606 * Adds a lock to a lockset.
608 int cl_io_lock_add(const struct lu_env *env, struct cl_io *io,
609 struct cl_io_lock_link *link)
614 if (cl_lockset_merge(&io->ci_lockset, &link->cill_descr))
617 cfs_list_add(&link->cill_linkage, &io->ci_lockset.cls_todo);
622 EXPORT_SYMBOL(cl_io_lock_add);
624 static void cl_free_io_lock_link(const struct lu_env *env,
625 struct cl_io_lock_link *link)
631 * Allocates new lock link, and uses it to add a lock to a lockset.
633 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
634 struct cl_lock_descr *descr)
636 struct cl_io_lock_link *link;
642 link->cill_descr = *descr;
643 link->cill_fini = cl_free_io_lock_link;
644 result = cl_io_lock_add(env, io, link);
645 if (result) /* lock match */
646 link->cill_fini(env, link);
652 EXPORT_SYMBOL(cl_io_lock_alloc_add);
655 * Starts io by calling cl_io_operations::cio_start() top-to-bottom.
657 int cl_io_start(const struct lu_env *env, struct cl_io *io)
659 const struct cl_io_slice *scan;
662 LINVRNT(cl_io_is_loopable(io));
663 LINVRNT(io->ci_state == CIS_LOCKED);
664 LINVRNT(cl_io_invariant(io));
667 io->ci_state = CIS_IO_GOING;
668 cl_io_for_each(scan, io) {
669 if (scan->cis_iop->op[io->ci_type].cio_start == NULL)
671 result = scan->cis_iop->op[io->ci_type].cio_start(env, scan);
679 EXPORT_SYMBOL(cl_io_start);
682 * Wait until current io iteration is finished by calling
683 * cl_io_operations::cio_end() bottom-to-top.
685 void cl_io_end(const struct lu_env *env, struct cl_io *io)
687 const struct cl_io_slice *scan;
689 LINVRNT(cl_io_is_loopable(io));
690 LINVRNT(io->ci_state == CIS_IO_GOING);
691 LINVRNT(cl_io_invariant(io));
694 cl_io_for_each_reverse(scan, io) {
695 if (scan->cis_iop->op[io->ci_type].cio_end != NULL)
696 scan->cis_iop->op[io->ci_type].cio_end(env, scan);
697 /* TODO: error handling. */
699 io->ci_state = CIS_IO_FINISHED;
702 EXPORT_SYMBOL(cl_io_end);
704 static const struct cl_page_slice *
705 cl_io_slice_page(const struct cl_io_slice *ios, struct cl_page *page)
707 const struct cl_page_slice *slice;
709 slice = cl_page_at(page, ios->cis_obj->co_lu.lo_dev->ld_type);
710 LINVRNT(slice != NULL);
715 * True iff \a page is within \a io range.
717 static int cl_page_in_io(const struct cl_page *page, const struct cl_io *io)
724 idx = page->cp_index;
725 switch (io->ci_type) {
729 * check that [start, end) and [pos, pos + count) extents
732 if (!cl_io_is_append(io)) {
733 const struct cl_io_rw_common *crw = &(io->u.ci_rw);
734 start = cl_offset(page->cp_obj, idx);
735 end = cl_offset(page->cp_obj, idx + 1);
736 result = crw->crw_pos < end &&
737 start < crw->crw_pos + crw->crw_count;
741 result = io->u.ci_fault.ft_index == idx;
750 * Called by read io, when page has to be read from the server.
752 * \see cl_io_operations::cio_read_page()
754 int cl_io_read_page(const struct lu_env *env, struct cl_io *io,
755 struct cl_page *page)
757 const struct cl_io_slice *scan;
758 struct cl_2queue *queue;
761 LINVRNT(io->ci_type == CIT_READ || io->ci_type == CIT_FAULT);
762 LINVRNT(cl_page_is_owned(page, io));
763 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
764 LINVRNT(cl_page_in_io(page, io));
765 LINVRNT(cl_io_invariant(io));
768 queue = &io->ci_queue;
770 cl_2queue_init(queue);
772 * ->cio_read_page() methods called in the loop below are supposed to
773 * never block waiting for network (the only subtle point is the
774 * creation of new pages for read-ahead that might result in cache
775 * shrinking, but currently only clean pages are shrunk and this
776 * requires no network io).
778 * Should this ever starts blocking, retry loop would be needed for
779 * "parallel io" (see CLO_REPEAT loops in cl_lock.c).
781 cl_io_for_each(scan, io) {
782 if (scan->cis_iop->cio_read_page != NULL) {
783 const struct cl_page_slice *slice;
785 slice = cl_io_slice_page(scan, page);
786 LINVRNT(slice != NULL);
787 result = scan->cis_iop->cio_read_page(env, scan, slice);
793 result = cl_io_submit_rw(env, io, CRT_READ, queue, CRP_NORMAL);
795 * Unlock unsent pages in case of error.
797 cl_page_list_disown(env, io, &queue->c2_qin);
798 cl_2queue_fini(env, queue);
801 EXPORT_SYMBOL(cl_io_read_page);
804 * Called by write io to prepare page to receive data from user buffer.
806 * \see cl_io_operations::cio_prepare_write()
808 int cl_io_prepare_write(const struct lu_env *env, struct cl_io *io,
809 struct cl_page *page, unsigned from, unsigned to)
811 const struct cl_io_slice *scan;
814 LINVRNT(io->ci_type == CIT_WRITE);
815 LINVRNT(cl_page_is_owned(page, io));
816 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
817 LINVRNT(cl_io_invariant(io));
818 LASSERT(cl_page_in_io(page, io));
821 cl_io_for_each_reverse(scan, io) {
822 if (scan->cis_iop->cio_prepare_write != NULL) {
823 const struct cl_page_slice *slice;
825 slice = cl_io_slice_page(scan, page);
826 result = scan->cis_iop->cio_prepare_write(env, scan,
835 EXPORT_SYMBOL(cl_io_prepare_write);
838 * Called by write io after user data were copied into a page.
840 * \see cl_io_operations::cio_commit_write()
842 int cl_io_commit_write(const struct lu_env *env, struct cl_io *io,
843 struct cl_page *page, unsigned from, unsigned to)
845 const struct cl_io_slice *scan;
848 LINVRNT(io->ci_type == CIT_WRITE);
849 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
850 LINVRNT(cl_io_invariant(io));
852 * XXX Uh... not nice. Top level cl_io_commit_write() call (vvp->lov)
853 * already called cl_page_cache_add(), moving page into CPS_CACHED
854 * state. Better (and more general) way of dealing with such situation
857 LASSERT(cl_page_is_owned(page, io) || page->cp_parent != NULL);
858 LASSERT(cl_page_in_io(page, io));
861 cl_io_for_each(scan, io) {
862 if (scan->cis_iop->cio_commit_write != NULL) {
863 const struct cl_page_slice *slice;
865 slice = cl_io_slice_page(scan, page);
866 result = scan->cis_iop->cio_commit_write(env, scan,
873 LINVRNT(result <= 0);
876 EXPORT_SYMBOL(cl_io_commit_write);
879 * Submits a list of pages for immediate io.
881 * After the function gets returned, The submitted pages are moved to
882 * queue->c2_qout queue, and queue->c2_qin contain both the pages don't need
883 * to be submitted, and the pages are errant to submit.
885 * \returns 0 if at least one page was submitted, error code otherwise.
886 * \see cl_io_operations::cio_submit()
888 int cl_io_submit_rw(const struct lu_env *env, struct cl_io *io,
889 enum cl_req_type crt, struct cl_2queue *queue,
890 enum cl_req_priority priority)
892 const struct cl_io_slice *scan;
895 LINVRNT(crt < ARRAY_SIZE(scan->cis_iop->req_op));
898 cl_io_for_each(scan, io) {
899 if (scan->cis_iop->req_op[crt].cio_submit == NULL)
901 result = scan->cis_iop->req_op[crt].cio_submit(env, scan, crt,
907 * If ->cio_submit() failed, no pages were sent.
909 LASSERT(ergo(result != 0, cfs_list_empty(&queue->c2_qout.pl_pages)));
912 EXPORT_SYMBOL(cl_io_submit_rw);
915 * Submit a sync_io and wait for the IO to be finished, or error happens.
916 * If \a timeout is zero, it means to wait for the IO unconditionally.
918 int cl_io_submit_sync(const struct lu_env *env, struct cl_io *io,
919 enum cl_req_type iot, struct cl_2queue *queue,
920 enum cl_req_priority prio, long timeout)
922 struct cl_sync_io *anchor = &cl_env_info(env)->clt_anchor;
926 LASSERT(prio == CRP_NORMAL || prio == CRP_CANCEL);
928 cl_page_list_for_each(pg, &queue->c2_qin) {
929 LASSERT(pg->cp_sync_io == NULL);
930 pg->cp_sync_io = anchor;
933 cl_sync_io_init(anchor, queue->c2_qin.pl_nr);
934 rc = cl_io_submit_rw(env, io, iot, queue, prio);
937 * If some pages weren't sent for any reason (e.g.,
938 * read found up-to-date pages in the cache, or write found
939 * clean pages), count them as completed to avoid infinite
942 cl_page_list_for_each(pg, &queue->c2_qin) {
943 pg->cp_sync_io = NULL;
944 cl_sync_io_note(anchor, +1);
947 /* wait for the IO to be finished. */
948 rc = cl_sync_io_wait(env, io, &queue->c2_qout,
951 LASSERT(cfs_list_empty(&queue->c2_qout.pl_pages));
952 cl_page_list_for_each(pg, &queue->c2_qin)
953 pg->cp_sync_io = NULL;
957 EXPORT_SYMBOL(cl_io_submit_sync);
960 * Cancel an IO which has been submitted by cl_io_submit_rw.
962 int cl_io_cancel(const struct lu_env *env, struct cl_io *io,
963 struct cl_page_list *queue)
965 struct cl_page *page;
968 CERROR("Canceling ongoing page trasmission\n");
969 cl_page_list_for_each(page, queue) {
972 LINVRNT(cl_page_in_io(page, io));
973 rc = cl_page_cancel(env, page);
974 result = result ?: rc;
978 EXPORT_SYMBOL(cl_io_cancel);
983 * Pumps io through iterations calling
985 * - cl_io_iter_init()
995 * - cl_io_iter_fini()
997 * repeatedly until there is no more io to do.
999 int cl_io_loop(const struct lu_env *env, struct cl_io *io)
1003 LINVRNT(cl_io_is_loopable(io));
1009 io->ci_continue = 0;
1010 result = cl_io_iter_init(env, io);
1013 result = cl_io_lock(env, io);
1016 * Notify layers that locks has been taken,
1017 * and do actual i/o.
1019 * - llite: kms, short read;
1020 * - llite: generic_file_read();
1022 result = cl_io_start(env, io);
1024 * Send any remaining pending
1027 * - llite: ll_rw_stats_tally.
1030 cl_io_unlock(env, io);
1031 cl_io_rw_advance(env, io, io->ci_nob - nob);
1034 cl_io_iter_fini(env, io);
1035 } while (result == 0 && io->ci_continue);
1037 result = io->ci_result;
1038 RETURN(result < 0 ? result : 0);
1040 EXPORT_SYMBOL(cl_io_loop);
1043 * Adds io slice to the cl_io.
1045 * This is called by cl_object_operations::coo_io_init() methods to add a
1046 * per-layer state to the io. New state is added at the end of
1047 * cl_io::ci_layers list, that is, it is at the bottom of the stack.
1049 * \see cl_lock_slice_add(), cl_req_slice_add(), cl_page_slice_add()
1051 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
1052 struct cl_object *obj,
1053 const struct cl_io_operations *ops)
1055 cfs_list_t *linkage = &slice->cis_linkage;
1057 LASSERT((linkage->prev == NULL && linkage->next == NULL) ||
1058 cfs_list_empty(linkage));
1061 cfs_list_add_tail(linkage, &io->ci_layers);
1063 slice->cis_obj = obj;
1064 slice->cis_iop = ops;
1067 EXPORT_SYMBOL(cl_io_slice_add);
1071 * Initializes page list.
1073 void cl_page_list_init(struct cl_page_list *plist)
1077 CFS_INIT_LIST_HEAD(&plist->pl_pages);
1078 plist->pl_owner = cfs_current();
1081 EXPORT_SYMBOL(cl_page_list_init);
1084 * Adds a page to a page list.
1086 void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page)
1089 /* it would be better to check that page is owned by "current" io, but
1090 * it is not passed here. */
1091 LASSERT(page->cp_owner != NULL);
1092 LINVRNT(plist->pl_owner == cfs_current());
1095 cfs_mutex_lock(&page->cp_mutex);
1097 LASSERT(cfs_list_empty(&page->cp_batch));
1098 cfs_list_add_tail(&page->cp_batch, &plist->pl_pages);
1100 page->cp_queue_ref = lu_ref_add(&page->cp_reference, "queue", plist);
1104 EXPORT_SYMBOL(cl_page_list_add);
1107 * Removes a page from a page list.
1109 void cl_page_list_del(const struct lu_env *env,
1110 struct cl_page_list *plist, struct cl_page *page)
1112 LASSERT(plist->pl_nr > 0);
1113 LINVRNT(plist->pl_owner == cfs_current());
1116 cfs_list_del_init(&page->cp_batch);
1118 cfs_mutex_unlock(&page->cp_mutex);
1121 lu_ref_del_at(&page->cp_reference, page->cp_queue_ref, "queue", plist);
1122 cl_page_put(env, page);
1125 EXPORT_SYMBOL(cl_page_list_del);
1128 * Moves a page from one page list to another.
1130 void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src,
1131 struct cl_page *page)
1133 LASSERT(src->pl_nr > 0);
1134 LINVRNT(dst->pl_owner == cfs_current());
1135 LINVRNT(src->pl_owner == cfs_current());
1138 cfs_list_move_tail(&page->cp_batch, &dst->pl_pages);
1141 lu_ref_set_at(&page->cp_reference,
1142 page->cp_queue_ref, "queue", src, dst);
1145 EXPORT_SYMBOL(cl_page_list_move);
1148 * splice the cl_page_list, just as list head does
1150 void cl_page_list_splice(struct cl_page_list *list, struct cl_page_list *head)
1152 struct cl_page *page;
1153 struct cl_page *tmp;
1155 LINVRNT(list->pl_owner == cfs_current());
1156 LINVRNT(head->pl_owner == cfs_current());
1159 cl_page_list_for_each_safe(page, tmp, list)
1160 cl_page_list_move(head, list, page);
1163 EXPORT_SYMBOL(cl_page_list_splice);
1165 void cl_page_disown0(const struct lu_env *env,
1166 struct cl_io *io, struct cl_page *pg);
1169 * Disowns pages in a queue.
1171 void cl_page_list_disown(const struct lu_env *env,
1172 struct cl_io *io, struct cl_page_list *plist)
1174 struct cl_page *page;
1175 struct cl_page *temp;
1177 LINVRNT(plist->pl_owner == cfs_current());
1180 cl_page_list_for_each_safe(page, temp, plist) {
1181 LASSERT(plist->pl_nr > 0);
1183 cfs_list_del_init(&page->cp_batch);
1185 cfs_mutex_unlock(&page->cp_mutex);
1189 * cl_page_disown0 rather than usual cl_page_disown() is used,
1190 * because pages are possibly in CPS_FREEING state already due
1191 * to the call to cl_page_list_discard().
1194 * XXX cl_page_disown0() will fail if page is not locked.
1196 cl_page_disown0(env, io, page);
1197 lu_ref_del(&page->cp_reference, "queue", plist);
1198 cl_page_put(env, page);
1202 EXPORT_SYMBOL(cl_page_list_disown);
1205 * Releases pages from queue.
1207 void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist)
1209 struct cl_page *page;
1210 struct cl_page *temp;
1212 LINVRNT(plist->pl_owner == cfs_current());
1215 cl_page_list_for_each_safe(page, temp, plist)
1216 cl_page_list_del(env, plist, page);
1217 LASSERT(plist->pl_nr == 0);
1220 EXPORT_SYMBOL(cl_page_list_fini);
1223 * Owns all pages in a queue.
1225 int cl_page_list_own(const struct lu_env *env,
1226 struct cl_io *io, struct cl_page_list *plist)
1228 struct cl_page *page;
1229 struct cl_page *temp;
1233 LINVRNT(plist->pl_owner == cfs_current());
1237 cl_page_list_for_each_safe(page, temp, plist) {
1238 LASSERT(index <= page->cp_index);
1239 index = page->cp_index;
1240 if (cl_page_own(env, io, page) == 0)
1241 result = result ?: page->cp_error;
1243 cl_page_list_del(env, plist, page);
1247 EXPORT_SYMBOL(cl_page_list_own);
1250 * Assumes all pages in a queue.
1252 void cl_page_list_assume(const struct lu_env *env,
1253 struct cl_io *io, struct cl_page_list *plist)
1255 struct cl_page *page;
1257 LINVRNT(plist->pl_owner == cfs_current());
1259 cl_page_list_for_each(page, plist)
1260 cl_page_assume(env, io, page);
1262 EXPORT_SYMBOL(cl_page_list_assume);
1265 * Discards all pages in a queue.
1267 void cl_page_list_discard(const struct lu_env *env, struct cl_io *io,
1268 struct cl_page_list *plist)
1270 struct cl_page *page;
1272 LINVRNT(plist->pl_owner == cfs_current());
1274 cl_page_list_for_each(page, plist)
1275 cl_page_discard(env, io, page);
1278 EXPORT_SYMBOL(cl_page_list_discard);
1281 * Unmaps all pages in a queue from user virtual memory.
1283 int cl_page_list_unmap(const struct lu_env *env, struct cl_io *io,
1284 struct cl_page_list *plist)
1286 struct cl_page *page;
1289 LINVRNT(plist->pl_owner == cfs_current());
1292 cl_page_list_for_each(page, plist) {
1293 result = cl_page_unmap(env, io, page);
1299 EXPORT_SYMBOL(cl_page_list_unmap);
1302 * Initialize dual page queue.
1304 void cl_2queue_init(struct cl_2queue *queue)
1307 cl_page_list_init(&queue->c2_qin);
1308 cl_page_list_init(&queue->c2_qout);
1311 EXPORT_SYMBOL(cl_2queue_init);
1314 * Add a page to the incoming page list of 2-queue.
1316 void cl_2queue_add(struct cl_2queue *queue, struct cl_page *page)
1319 cl_page_list_add(&queue->c2_qin, page);
1322 EXPORT_SYMBOL(cl_2queue_add);
1325 * Disown pages in both lists of a 2-queue.
1327 void cl_2queue_disown(const struct lu_env *env,
1328 struct cl_io *io, struct cl_2queue *queue)
1331 cl_page_list_disown(env, io, &queue->c2_qin);
1332 cl_page_list_disown(env, io, &queue->c2_qout);
1335 EXPORT_SYMBOL(cl_2queue_disown);
1338 * Discard (truncate) pages in both lists of a 2-queue.
1340 void cl_2queue_discard(const struct lu_env *env,
1341 struct cl_io *io, struct cl_2queue *queue)
1344 cl_page_list_discard(env, io, &queue->c2_qin);
1345 cl_page_list_discard(env, io, &queue->c2_qout);
1348 EXPORT_SYMBOL(cl_2queue_discard);
1351 * Assume to own the pages in cl_2queue
1353 void cl_2queue_assume(const struct lu_env *env,
1354 struct cl_io *io, struct cl_2queue *queue)
1356 cl_page_list_assume(env, io, &queue->c2_qin);
1357 cl_page_list_assume(env, io, &queue->c2_qout);
1359 EXPORT_SYMBOL(cl_2queue_assume);
1362 * Finalize both page lists of a 2-queue.
1364 void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue)
1367 cl_page_list_fini(env, &queue->c2_qout);
1368 cl_page_list_fini(env, &queue->c2_qin);
1371 EXPORT_SYMBOL(cl_2queue_fini);
1374 * Initialize a 2-queue to contain \a page in its incoming page list.
1376 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page)
1379 cl_2queue_init(queue);
1380 cl_2queue_add(queue, page);
1383 EXPORT_SYMBOL(cl_2queue_init_page);
1386 * Returns top-level io.
1388 * \see cl_object_top(), cl_page_top().
1390 struct cl_io *cl_io_top(struct cl_io *io)
1393 while (io->ci_parent != NULL)
1397 EXPORT_SYMBOL(cl_io_top);
1400 * Prints human readable representation of \a io to the \a f.
1402 void cl_io_print(const struct lu_env *env, void *cookie,
1403 lu_printer_t printer, const struct cl_io *io)
1408 * Adds request slice to the compound request.
1410 * This is called by cl_device_operations::cdo_req_init() methods to add a
1411 * per-layer state to the request. New state is added at the end of
1412 * cl_req::crq_layers list, that is, it is at the bottom of the stack.
1414 * \see cl_lock_slice_add(), cl_page_slice_add(), cl_io_slice_add()
1416 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
1417 struct cl_device *dev,
1418 const struct cl_req_operations *ops)
1421 cfs_list_add_tail(&slice->crs_linkage, &req->crq_layers);
1422 slice->crs_dev = dev;
1423 slice->crs_ops = ops;
1424 slice->crs_req = req;
1427 EXPORT_SYMBOL(cl_req_slice_add);
1429 static void cl_req_free(const struct lu_env *env, struct cl_req *req)
1433 LASSERT(cfs_list_empty(&req->crq_pages));
1434 LASSERT(req->crq_nrpages == 0);
1435 LINVRNT(cfs_list_empty(&req->crq_layers));
1436 LINVRNT(equi(req->crq_nrobjs > 0, req->crq_o != NULL));
1439 if (req->crq_o != NULL) {
1440 for (i = 0; i < req->crq_nrobjs; ++i) {
1441 struct cl_object *obj = req->crq_o[i].ro_obj;
1443 lu_object_ref_del_at(&obj->co_lu,
1444 req->crq_o[i].ro_obj_ref,
1446 cl_object_put(env, obj);
1449 OBD_FREE(req->crq_o, req->crq_nrobjs * sizeof req->crq_o[0]);
1455 static int cl_req_init(const struct lu_env *env, struct cl_req *req,
1456 struct cl_page *page)
1458 struct cl_device *dev;
1459 struct cl_page_slice *slice;
1464 page = cl_page_top(page);
1466 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
1467 dev = lu2cl_dev(slice->cpl_obj->co_lu.lo_dev);
1468 if (dev->cd_ops->cdo_req_init != NULL) {
1469 result = dev->cd_ops->cdo_req_init(env,
1475 page = page->cp_child;
1476 } while (page != NULL && result == 0);
1481 * Invokes per-request transfer completion call-backs
1482 * (cl_req_operations::cro_completion()) bottom-to-top.
1484 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int rc)
1486 struct cl_req_slice *slice;
1490 * for the lack of list_for_each_entry_reverse_safe()...
1492 while (!cfs_list_empty(&req->crq_layers)) {
1493 slice = cfs_list_entry(req->crq_layers.prev,
1494 struct cl_req_slice, crs_linkage);
1495 cfs_list_del_init(&slice->crs_linkage);
1496 if (slice->crs_ops->cro_completion != NULL)
1497 slice->crs_ops->cro_completion(env, slice, rc);
1499 cl_req_free(env, req);
1502 EXPORT_SYMBOL(cl_req_completion);
1505 * Allocates new transfer request.
1507 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
1508 enum cl_req_type crt, int nr_objects)
1512 LINVRNT(nr_objects > 0);
1519 OBD_ALLOC(req->crq_o, nr_objects * sizeof req->crq_o[0]);
1520 if (req->crq_o != NULL) {
1521 req->crq_nrobjs = nr_objects;
1522 req->crq_type = crt;
1523 CFS_INIT_LIST_HEAD(&req->crq_pages);
1524 CFS_INIT_LIST_HEAD(&req->crq_layers);
1525 result = cl_req_init(env, req, page);
1529 cl_req_completion(env, req, result);
1530 req = ERR_PTR(result);
1533 req = ERR_PTR(-ENOMEM);
1536 EXPORT_SYMBOL(cl_req_alloc);
1539 * Adds a page to a request.
1541 void cl_req_page_add(const struct lu_env *env,
1542 struct cl_req *req, struct cl_page *page)
1544 struct cl_object *obj;
1545 struct cl_req_obj *rqo;
1549 page = cl_page_top(page);
1551 LASSERT(cfs_list_empty(&page->cp_flight));
1552 LASSERT(page->cp_req == NULL);
1554 CL_PAGE_DEBUG(D_PAGE, env, page, "req %p, %d, %u\n",
1555 req, req->crq_type, req->crq_nrpages);
1557 cfs_list_add_tail(&page->cp_flight, &req->crq_pages);
1560 obj = cl_object_top(page->cp_obj);
1561 for (i = 0, rqo = req->crq_o; obj != rqo->ro_obj; ++i, ++rqo) {
1562 if (rqo->ro_obj == NULL) {
1565 rqo->ro_obj_ref = lu_object_ref_add(&obj->co_lu,
1570 LASSERT(i < req->crq_nrobjs);
1573 EXPORT_SYMBOL(cl_req_page_add);
1576 * Removes a page from a request.
1578 void cl_req_page_done(const struct lu_env *env, struct cl_page *page)
1580 struct cl_req *req = page->cp_req;
1583 page = cl_page_top(page);
1585 LASSERT(!cfs_list_empty(&page->cp_flight));
1586 LASSERT(req->crq_nrpages > 0);
1588 cfs_list_del_init(&page->cp_flight);
1590 page->cp_req = NULL;
1593 EXPORT_SYMBOL(cl_req_page_done);
1596 * Notifies layers that request is about to depart by calling
1597 * cl_req_operations::cro_prep() top-to-bottom.
1599 int cl_req_prep(const struct lu_env *env, struct cl_req *req)
1603 const struct cl_req_slice *slice;
1607 * Check that the caller of cl_req_alloc() didn't lie about the number
1610 for (i = 0; i < req->crq_nrobjs; ++i)
1611 LASSERT(req->crq_o[i].ro_obj != NULL);
1614 cfs_list_for_each_entry(slice, &req->crq_layers, crs_linkage) {
1615 if (slice->crs_ops->cro_prep != NULL) {
1616 result = slice->crs_ops->cro_prep(env, slice);
1623 EXPORT_SYMBOL(cl_req_prep);
1626 * Fills in attributes that are passed to server together with transfer. Only
1627 * attributes from \a flags may be touched. This can be called multiple times
1628 * for the same request.
1630 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
1631 struct cl_req_attr *attr, obd_valid flags)
1633 const struct cl_req_slice *slice;
1634 struct cl_page *page;
1637 LASSERT(!cfs_list_empty(&req->crq_pages));
1640 /* Take any page to use as a model. */
1641 page = cfs_list_entry(req->crq_pages.next, struct cl_page, cp_flight);
1643 for (i = 0; i < req->crq_nrobjs; ++i) {
1644 cfs_list_for_each_entry(slice, &req->crq_layers, crs_linkage) {
1645 const struct cl_page_slice *scan;
1646 const struct cl_object *obj;
1648 scan = cl_page_at(page,
1649 slice->crs_dev->cd_lu_dev.ld_type);
1650 LASSERT(scan != NULL);
1651 obj = scan->cpl_obj;
1652 if (slice->crs_ops->cro_attr_set != NULL)
1653 slice->crs_ops->cro_attr_set(env, slice, obj,
1659 EXPORT_SYMBOL(cl_req_attr_set);
1661 /* XXX complete(), init_completion(), and wait_for_completion(), until they are
1662 * implemented in libcfs. */
1664 # include <linux/sched.h>
1665 #else /* __KERNEL__ */
1666 # include <liblustre.h>
1670 * Initialize synchronous io wait anchor, for transfer of \a nrpages pages.
1672 void cl_sync_io_init(struct cl_sync_io *anchor, int nrpages)
1675 cfs_waitq_init(&anchor->csi_waitq);
1676 cfs_atomic_set(&anchor->csi_sync_nr, nrpages);
1677 anchor->csi_sync_rc = 0;
1680 EXPORT_SYMBOL(cl_sync_io_init);
1683 * Wait until all transfer completes. Transfer completion routine has to call
1684 * cl_sync_io_note() for every page.
1686 int cl_sync_io_wait(const struct lu_env *env, struct cl_io *io,
1687 struct cl_page_list *queue, struct cl_sync_io *anchor,
1690 struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
1695 LASSERT(timeout >= 0);
1697 rc = l_wait_event(anchor->csi_waitq,
1698 cfs_atomic_read(&anchor->csi_sync_nr) == 0,
1701 CERROR("SYNC IO failed with error: %d, try to cancel "
1702 "%d remaining pages\n",
1703 rc, cfs_atomic_read(&anchor->csi_sync_nr));
1705 (void)cl_io_cancel(env, io, queue);
1707 lwi = (struct l_wait_info) { 0 };
1708 (void)l_wait_event(anchor->csi_waitq,
1709 cfs_atomic_read(&anchor->csi_sync_nr) == 0,
1712 rc = anchor->csi_sync_rc;
1714 LASSERT(cfs_atomic_read(&anchor->csi_sync_nr) == 0);
1715 cl_page_list_assume(env, io, queue);
1716 POISON(anchor, 0x5a, sizeof *anchor);
1719 EXPORT_SYMBOL(cl_sync_io_wait);
1722 * Indicate that transfer of a single page completed.
1724 void cl_sync_io_note(struct cl_sync_io *anchor, int ioret)
1727 if (anchor->csi_sync_rc == 0 && ioret < 0)
1728 anchor->csi_sync_rc = ioret;
1730 * Synchronous IO done without releasing page lock (e.g., as a part of
1731 * ->{prepare,commit}_write(). Completion is used to signal the end of
1734 LASSERT(cfs_atomic_read(&anchor->csi_sync_nr) > 0);
1735 if (cfs_atomic_dec_and_test(&anchor->csi_sync_nr))
1736 cfs_waitq_broadcast(&anchor->csi_waitq);
1739 EXPORT_SYMBOL(cl_sync_io_note);