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 (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011, Intel Corporation.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
40 * Author: Nikita Danilov <nikita.danilov@sun.com>
43 #define DEBUG_SUBSYSTEM S_CLASS
45 # define EXPORT_SYMTAB
48 #include <obd_class.h>
49 #include <obd_support.h>
50 #include <lustre_fid.h>
51 #include <libcfs/list.h>
52 /* lu_time_global_{init,fini}() */
55 #include <cl_object.h>
56 #include "cl_internal.h"
58 /*****************************************************************************
64 #define cl_io_for_each(slice, io) \
65 cfs_list_for_each_entry((slice), &io->ci_layers, cis_linkage)
66 #define cl_io_for_each_reverse(slice, io) \
67 cfs_list_for_each_entry_reverse((slice), &io->ci_layers, cis_linkage)
69 static inline int cl_io_type_is_valid(enum cl_io_type type)
71 return CIT_READ <= type && type < CIT_OP_NR;
74 static inline int cl_io_is_loopable(const struct cl_io *io)
76 return cl_io_type_is_valid(io->ci_type) && io->ci_type != CIT_MISC;
80 * Returns true iff there is an IO ongoing in the given environment.
82 int cl_io_is_going(const struct lu_env *env)
84 return cl_env_info(env)->clt_current_io != NULL;
86 EXPORT_SYMBOL(cl_io_is_going);
89 * cl_io invariant that holds at all times when exported cl_io_*() functions
90 * are entered and left.
92 static int cl_io_invariant(const struct cl_io *io)
99 * io can own pages only when it is ongoing. Sub-io might
100 * still be in CIS_LOCKED state when top-io is in
103 ergo(io->ci_owned_nr > 0, io->ci_state == CIS_IO_GOING ||
104 (io->ci_state == CIS_LOCKED && up != NULL));
108 * Finalize \a io, by calling cl_io_operations::cio_fini() bottom-to-top.
110 void cl_io_fini(const struct lu_env *env, struct cl_io *io)
112 struct cl_io_slice *slice;
113 struct cl_thread_info *info;
115 LINVRNT(cl_io_type_is_valid(io->ci_type));
116 LINVRNT(cl_io_invariant(io));
119 while (!cfs_list_empty(&io->ci_layers)) {
120 slice = container_of(io->ci_layers.next, struct cl_io_slice,
122 cfs_list_del_init(&slice->cis_linkage);
123 if (slice->cis_iop->op[io->ci_type].cio_fini != NULL)
124 slice->cis_iop->op[io->ci_type].cio_fini(env, slice);
126 * Invalidate slice to catch use after free. This assumes that
127 * slices are allocated within session and can be touched
128 * after ->cio_fini() returns.
130 slice->cis_io = NULL;
132 io->ci_state = CIS_FINI;
133 info = cl_env_info(env);
134 if (info->clt_current_io == io)
135 info->clt_current_io = NULL;
138 EXPORT_SYMBOL(cl_io_fini);
140 static int cl_io_init0(const struct lu_env *env, struct cl_io *io,
141 enum cl_io_type iot, struct cl_object *obj)
143 struct cl_object *scan;
146 LINVRNT(io->ci_state == CIS_ZERO || io->ci_state == CIS_FINI);
147 LINVRNT(cl_io_type_is_valid(iot));
148 LINVRNT(cl_io_invariant(io));
152 CFS_INIT_LIST_HEAD(&io->ci_lockset.cls_todo);
153 CFS_INIT_LIST_HEAD(&io->ci_lockset.cls_curr);
154 CFS_INIT_LIST_HEAD(&io->ci_lockset.cls_done);
155 CFS_INIT_LIST_HEAD(&io->ci_layers);
158 cl_object_for_each(scan, obj) {
159 if (scan->co_ops->coo_io_init != NULL) {
160 result = scan->co_ops->coo_io_init(env, scan, io);
166 io->ci_state = CIS_INIT;
171 * Initialize sub-io, by calling cl_io_operations::cio_init() top-to-bottom.
173 * \pre obj != cl_object_top(obj)
175 int cl_io_sub_init(const struct lu_env *env, struct cl_io *io,
176 enum cl_io_type iot, struct cl_object *obj)
178 struct cl_thread_info *info = cl_env_info(env);
180 LASSERT(obj != cl_object_top(obj));
181 if (info->clt_current_io == NULL)
182 info->clt_current_io = io;
183 return cl_io_init0(env, io, iot, obj);
185 EXPORT_SYMBOL(cl_io_sub_init);
188 * Initialize \a io, by calling cl_io_operations::cio_init() top-to-bottom.
190 * Caller has to call cl_io_fini() after a call to cl_io_init(), no matter
191 * what the latter returned.
193 * \pre obj == cl_object_top(obj)
194 * \pre cl_io_type_is_valid(iot)
195 * \post cl_io_type_is_valid(io->ci_type) && io->ci_type == iot
197 int cl_io_init(const struct lu_env *env, struct cl_io *io,
198 enum cl_io_type iot, struct cl_object *obj)
200 struct cl_thread_info *info = cl_env_info(env);
202 LASSERT(obj == cl_object_top(obj));
203 LASSERT(info->clt_current_io == NULL);
205 info->clt_current_io = io;
206 return cl_io_init0(env, io, iot, obj);
208 EXPORT_SYMBOL(cl_io_init);
211 * Initialize read or write io.
213 * \pre iot == CIT_READ || iot == CIT_WRITE
215 int cl_io_rw_init(const struct lu_env *env, struct cl_io *io,
216 enum cl_io_type iot, loff_t pos, size_t count)
218 LINVRNT(iot == CIT_READ || iot == CIT_WRITE);
219 LINVRNT(io->ci_obj != NULL);
222 LU_OBJECT_HEADER(D_VFSTRACE, env, &io->ci_obj->co_lu,
223 "io range: %u ["LPU64", "LPU64") %u %u\n",
224 iot, (__u64)pos, (__u64)pos + count,
225 io->u.ci_rw.crw_nonblock, io->u.ci_wr.wr_append);
226 io->u.ci_rw.crw_pos = pos;
227 io->u.ci_rw.crw_count = count;
228 RETURN(cl_io_init(env, io, iot, io->ci_obj));
230 EXPORT_SYMBOL(cl_io_rw_init);
232 static inline const struct lu_fid *
233 cl_lock_descr_fid(const struct cl_lock_descr *descr)
235 return lu_object_fid(&descr->cld_obj->co_lu);
238 static int cl_lock_descr_sort(const struct cl_lock_descr *d0,
239 const struct cl_lock_descr *d1)
241 return lu_fid_cmp(cl_lock_descr_fid(d0), cl_lock_descr_fid(d1)) ?:
242 __diff_normalize(d0->cld_start, d1->cld_start);
245 static int cl_lock_descr_cmp(const struct cl_lock_descr *d0,
246 const struct cl_lock_descr *d1)
250 ret = lu_fid_cmp(cl_lock_descr_fid(d0), cl_lock_descr_fid(d1));
253 if (d0->cld_end < d1->cld_start)
255 if (d0->cld_start > d0->cld_end)
260 static void cl_lock_descr_merge(struct cl_lock_descr *d0,
261 const struct cl_lock_descr *d1)
263 d0->cld_start = min(d0->cld_start, d1->cld_start);
264 d0->cld_end = max(d0->cld_end, d1->cld_end);
266 if (d1->cld_mode == CLM_WRITE && d0->cld_mode != CLM_WRITE)
267 d0->cld_mode = CLM_WRITE;
269 if (d1->cld_mode == CLM_GROUP && d0->cld_mode != CLM_GROUP)
270 d0->cld_mode = CLM_GROUP;
274 * Sort locks in lexicographical order of their (fid, start-offset) pairs.
276 static void cl_io_locks_sort(struct cl_io *io)
281 /* hidden treasure: bubble sort for now. */
283 struct cl_io_lock_link *curr;
284 struct cl_io_lock_link *prev;
285 struct cl_io_lock_link *temp;
290 cfs_list_for_each_entry_safe(curr, temp,
291 &io->ci_lockset.cls_todo,
294 switch (cl_lock_descr_sort(&prev->cill_descr,
295 &curr->cill_descr)) {
298 * IMPOSSIBLE: Identical locks are
305 cfs_list_move_tail(&curr->cill_linkage,
306 &prev->cill_linkage);
308 continue; /* don't change prev: it's
309 * still "previous" */
310 case -1: /* already in order */
321 * Check whether \a queue contains locks matching \a need.
323 * \retval +ve there is a matching lock in the \a queue
324 * \retval 0 there are no matching locks in the \a queue
326 int cl_queue_match(const cfs_list_t *queue,
327 const struct cl_lock_descr *need)
329 struct cl_io_lock_link *scan;
332 cfs_list_for_each_entry(scan, queue, cill_linkage) {
333 if (cl_lock_descr_match(&scan->cill_descr, need))
338 EXPORT_SYMBOL(cl_queue_match);
340 static int cl_queue_merge(const cfs_list_t *queue,
341 const struct cl_lock_descr *need)
343 struct cl_io_lock_link *scan;
346 cfs_list_for_each_entry(scan, queue, cill_linkage) {
347 if (cl_lock_descr_cmp(&scan->cill_descr, need))
349 cl_lock_descr_merge(&scan->cill_descr, need);
350 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
351 scan->cill_descr.cld_mode, scan->cill_descr.cld_start,
352 scan->cill_descr.cld_end);
359 static int cl_lockset_match(const struct cl_lockset *set,
360 const struct cl_lock_descr *need)
362 return cl_queue_match(&set->cls_curr, need) ||
363 cl_queue_match(&set->cls_done, need);
366 static int cl_lockset_merge(const struct cl_lockset *set,
367 const struct cl_lock_descr *need)
369 return cl_queue_merge(&set->cls_todo, need) ||
370 cl_lockset_match(set, need);
373 static int cl_lockset_lock_one(const struct lu_env *env,
374 struct cl_io *io, struct cl_lockset *set,
375 struct cl_io_lock_link *link)
377 struct cl_lock *lock;
382 lock = cl_lock_request(env, io, &link->cill_descr, "io", io);
384 link->cill_lock = lock;
385 cfs_list_move(&link->cill_linkage, &set->cls_curr);
386 if (!(link->cill_descr.cld_enq_flags & CEF_ASYNC)) {
387 result = cl_wait(env, lock);
389 cfs_list_move(&link->cill_linkage,
394 result = PTR_ERR(lock);
398 static void cl_lock_link_fini(const struct lu_env *env, struct cl_io *io,
399 struct cl_io_lock_link *link)
401 struct cl_lock *lock = link->cill_lock;
404 cfs_list_del_init(&link->cill_linkage);
406 cl_lock_release(env, lock, "io", io);
407 link->cill_lock = NULL;
409 if (link->cill_fini != NULL)
410 link->cill_fini(env, link);
414 static int cl_lockset_lock(const struct lu_env *env, struct cl_io *io,
415 struct cl_lockset *set)
417 struct cl_io_lock_link *link;
418 struct cl_io_lock_link *temp;
419 struct cl_lock *lock;
424 cfs_list_for_each_entry_safe(link, temp, &set->cls_todo, cill_linkage) {
425 if (!cl_lockset_match(set, &link->cill_descr)) {
426 /* XXX some locking to guarantee that locks aren't
427 * expanded in between. */
428 result = cl_lockset_lock_one(env, io, set, link);
432 cl_lock_link_fini(env, io, link);
435 cfs_list_for_each_entry_safe(link, temp,
436 &set->cls_curr, cill_linkage) {
437 lock = link->cill_lock;
438 result = cl_wait(env, lock);
440 cfs_list_move(&link->cill_linkage,
450 * Takes locks necessary for the current iteration of io.
452 * Calls cl_io_operations::cio_lock() top-to-bottom to collect locks required
453 * by layers for the current iteration. Then sort locks (to avoid dead-locks),
456 int cl_io_lock(const struct lu_env *env, struct cl_io *io)
458 const struct cl_io_slice *scan;
461 LINVRNT(cl_io_is_loopable(io));
462 LINVRNT(io->ci_state == CIS_IT_STARTED);
463 LINVRNT(cl_io_invariant(io));
466 cl_io_for_each(scan, io) {
467 if (scan->cis_iop->op[io->ci_type].cio_lock == NULL)
469 result = scan->cis_iop->op[io->ci_type].cio_lock(env, scan);
474 cl_io_locks_sort(io);
475 result = cl_lockset_lock(env, io, &io->ci_lockset);
478 cl_io_unlock(env, io);
480 io->ci_state = CIS_LOCKED;
483 EXPORT_SYMBOL(cl_io_lock);
486 * Release locks takes by io.
488 void cl_io_unlock(const struct lu_env *env, struct cl_io *io)
490 struct cl_lockset *set;
491 struct cl_io_lock_link *link;
492 struct cl_io_lock_link *temp;
493 const struct cl_io_slice *scan;
495 LASSERT(cl_io_is_loopable(io));
496 LASSERT(CIS_IT_STARTED <= io->ci_state && io->ci_state < CIS_UNLOCKED);
497 LINVRNT(cl_io_invariant(io));
500 set = &io->ci_lockset;
502 cfs_list_for_each_entry_safe(link, temp, &set->cls_todo, cill_linkage)
503 cl_lock_link_fini(env, io, link);
505 cfs_list_for_each_entry_safe(link, temp, &set->cls_curr, cill_linkage)
506 cl_lock_link_fini(env, io, link);
508 cfs_list_for_each_entry_safe(link, temp, &set->cls_done, cill_linkage) {
509 cl_unuse(env, link->cill_lock);
510 cl_lock_link_fini(env, io, link);
512 cl_io_for_each_reverse(scan, io) {
513 if (scan->cis_iop->op[io->ci_type].cio_unlock != NULL)
514 scan->cis_iop->op[io->ci_type].cio_unlock(env, scan);
516 io->ci_state = CIS_UNLOCKED;
517 LASSERT(!cl_env_info(env)->clt_counters[CNL_TOP].ctc_nr_locks_acquired);
520 EXPORT_SYMBOL(cl_io_unlock);
523 * Prepares next iteration of io.
525 * Calls cl_io_operations::cio_iter_init() top-to-bottom. This exists to give
526 * layers a chance to modify io parameters, e.g., so that lov can restrict io
527 * to a single stripe.
529 int cl_io_iter_init(const struct lu_env *env, struct cl_io *io)
531 const struct cl_io_slice *scan;
534 LINVRNT(cl_io_is_loopable(io));
535 LINVRNT(io->ci_state == CIS_INIT || io->ci_state == CIS_IT_ENDED);
536 LINVRNT(cl_io_invariant(io));
540 cl_io_for_each(scan, io) {
541 if (scan->cis_iop->op[io->ci_type].cio_iter_init == NULL)
543 result = scan->cis_iop->op[io->ci_type].cio_iter_init(env,
549 io->ci_state = CIS_IT_STARTED;
552 EXPORT_SYMBOL(cl_io_iter_init);
555 * Finalizes io iteration.
557 * Calls cl_io_operations::cio_iter_fini() bottom-to-top.
559 void cl_io_iter_fini(const struct lu_env *env, struct cl_io *io)
561 const struct cl_io_slice *scan;
563 LINVRNT(cl_io_is_loopable(io));
564 LINVRNT(io->ci_state == CIS_UNLOCKED);
565 LINVRNT(cl_io_invariant(io));
568 cl_io_for_each_reverse(scan, io) {
569 if (scan->cis_iop->op[io->ci_type].cio_iter_fini != NULL)
570 scan->cis_iop->op[io->ci_type].cio_iter_fini(env, scan);
572 io->ci_state = CIS_IT_ENDED;
575 EXPORT_SYMBOL(cl_io_iter_fini);
578 * Records that read or write io progressed \a nob bytes forward.
580 void cl_io_rw_advance(const struct lu_env *env, struct cl_io *io, size_t nob)
582 const struct cl_io_slice *scan;
584 LINVRNT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
586 LINVRNT(cl_io_is_loopable(io));
587 LINVRNT(cl_io_invariant(io));
591 io->u.ci_rw.crw_pos += nob;
592 io->u.ci_rw.crw_count -= nob;
594 /* layers have to be notified. */
595 cl_io_for_each_reverse(scan, io) {
596 if (scan->cis_iop->op[io->ci_type].cio_advance != NULL)
597 scan->cis_iop->op[io->ci_type].cio_advance(env, scan,
602 EXPORT_SYMBOL(cl_io_rw_advance);
605 * Adds a lock to a lockset.
607 int cl_io_lock_add(const struct lu_env *env, struct cl_io *io,
608 struct cl_io_lock_link *link)
613 if (cl_lockset_merge(&io->ci_lockset, &link->cill_descr))
616 cfs_list_add(&link->cill_linkage, &io->ci_lockset.cls_todo);
621 EXPORT_SYMBOL(cl_io_lock_add);
623 static void cl_free_io_lock_link(const struct lu_env *env,
624 struct cl_io_lock_link *link)
630 * Allocates new lock link, and uses it to add a lock to a lockset.
632 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
633 struct cl_lock_descr *descr)
635 struct cl_io_lock_link *link;
641 link->cill_descr = *descr;
642 link->cill_fini = cl_free_io_lock_link;
643 result = cl_io_lock_add(env, io, link);
644 if (result) /* lock match */
645 link->cill_fini(env, link);
651 EXPORT_SYMBOL(cl_io_lock_alloc_add);
654 * Starts io by calling cl_io_operations::cio_start() top-to-bottom.
656 int cl_io_start(const struct lu_env *env, struct cl_io *io)
658 const struct cl_io_slice *scan;
661 LINVRNT(cl_io_is_loopable(io));
662 LINVRNT(io->ci_state == CIS_LOCKED);
663 LINVRNT(cl_io_invariant(io));
666 io->ci_state = CIS_IO_GOING;
667 cl_io_for_each(scan, io) {
668 if (scan->cis_iop->op[io->ci_type].cio_start == NULL)
670 result = scan->cis_iop->op[io->ci_type].cio_start(env, scan);
678 EXPORT_SYMBOL(cl_io_start);
681 * Wait until current io iteration is finished by calling
682 * cl_io_operations::cio_end() bottom-to-top.
684 void cl_io_end(const struct lu_env *env, struct cl_io *io)
686 const struct cl_io_slice *scan;
688 LINVRNT(cl_io_is_loopable(io));
689 LINVRNT(io->ci_state == CIS_IO_GOING);
690 LINVRNT(cl_io_invariant(io));
693 cl_io_for_each_reverse(scan, io) {
694 if (scan->cis_iop->op[io->ci_type].cio_end != NULL)
695 scan->cis_iop->op[io->ci_type].cio_end(env, scan);
696 /* TODO: error handling. */
698 io->ci_state = CIS_IO_FINISHED;
701 EXPORT_SYMBOL(cl_io_end);
703 static const struct cl_page_slice *
704 cl_io_slice_page(const struct cl_io_slice *ios, struct cl_page *page)
706 const struct cl_page_slice *slice;
708 slice = cl_page_at(page, ios->cis_obj->co_lu.lo_dev->ld_type);
709 LINVRNT(slice != NULL);
714 * True iff \a page is within \a io range.
716 static int cl_page_in_io(const struct cl_page *page, const struct cl_io *io)
723 idx = page->cp_index;
724 switch (io->ci_type) {
728 * check that [start, end) and [pos, pos + count) extents
731 if (!cl_io_is_append(io)) {
732 const struct cl_io_rw_common *crw = &(io->u.ci_rw);
733 start = cl_offset(page->cp_obj, idx);
734 end = cl_offset(page->cp_obj, idx + 1);
735 result = crw->crw_pos < end &&
736 start < crw->crw_pos + crw->crw_count;
740 result = io->u.ci_fault.ft_index == idx;
749 * Called by read io, when page has to be read from the server.
751 * \see cl_io_operations::cio_read_page()
753 int cl_io_read_page(const struct lu_env *env, struct cl_io *io,
754 struct cl_page *page)
756 const struct cl_io_slice *scan;
757 struct cl_2queue *queue;
760 LINVRNT(io->ci_type == CIT_READ || io->ci_type == CIT_FAULT);
761 LINVRNT(cl_page_is_owned(page, io));
762 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
763 LINVRNT(cl_page_in_io(page, io));
764 LINVRNT(cl_io_invariant(io));
767 queue = &io->ci_queue;
769 cl_2queue_init(queue);
771 * ->cio_read_page() methods called in the loop below are supposed to
772 * never block waiting for network (the only subtle point is the
773 * creation of new pages for read-ahead that might result in cache
774 * shrinking, but currently only clean pages are shrunk and this
775 * requires no network io).
777 * Should this ever starts blocking, retry loop would be needed for
778 * "parallel io" (see CLO_REPEAT loops in cl_lock.c).
780 cl_io_for_each(scan, io) {
781 if (scan->cis_iop->cio_read_page != NULL) {
782 const struct cl_page_slice *slice;
784 slice = cl_io_slice_page(scan, page);
785 LINVRNT(slice != NULL);
786 result = scan->cis_iop->cio_read_page(env, scan, slice);
792 result = cl_io_submit_rw(env, io, CRT_READ, queue, CRP_NORMAL);
794 * Unlock unsent pages in case of error.
796 cl_page_list_disown(env, io, &queue->c2_qin);
797 cl_2queue_fini(env, queue);
800 EXPORT_SYMBOL(cl_io_read_page);
803 * Called by write io to prepare page to receive data from user buffer.
805 * \see cl_io_operations::cio_prepare_write()
807 int cl_io_prepare_write(const struct lu_env *env, struct cl_io *io,
808 struct cl_page *page, unsigned from, unsigned to)
810 const struct cl_io_slice *scan;
813 LINVRNT(io->ci_type == CIT_WRITE);
814 LINVRNT(cl_page_is_owned(page, io));
815 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
816 LINVRNT(cl_io_invariant(io));
817 LASSERT(cl_page_in_io(page, io));
820 cl_io_for_each_reverse(scan, io) {
821 if (scan->cis_iop->cio_prepare_write != NULL) {
822 const struct cl_page_slice *slice;
824 slice = cl_io_slice_page(scan, page);
825 result = scan->cis_iop->cio_prepare_write(env, scan,
834 EXPORT_SYMBOL(cl_io_prepare_write);
837 * Called by write io after user data were copied into a page.
839 * \see cl_io_operations::cio_commit_write()
841 int cl_io_commit_write(const struct lu_env *env, struct cl_io *io,
842 struct cl_page *page, unsigned from, unsigned to)
844 const struct cl_io_slice *scan;
847 LINVRNT(io->ci_type == CIT_WRITE);
848 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
849 LINVRNT(cl_io_invariant(io));
851 * XXX Uh... not nice. Top level cl_io_commit_write() call (vvp->lov)
852 * already called cl_page_cache_add(), moving page into CPS_CACHED
853 * state. Better (and more general) way of dealing with such situation
856 LASSERT(cl_page_is_owned(page, io) || page->cp_parent != NULL);
857 LASSERT(cl_page_in_io(page, io));
860 cl_io_for_each(scan, io) {
861 if (scan->cis_iop->cio_commit_write != NULL) {
862 const struct cl_page_slice *slice;
864 slice = cl_io_slice_page(scan, page);
865 result = scan->cis_iop->cio_commit_write(env, scan,
872 LINVRNT(result <= 0);
875 EXPORT_SYMBOL(cl_io_commit_write);
878 * Submits a list of pages for immediate io.
880 * After the function gets returned, The submitted pages are moved to
881 * queue->c2_qout queue, and queue->c2_qin contain both the pages don't need
882 * to be submitted, and the pages are errant to submit.
884 * \returns 0 if at least one page was submitted, error code otherwise.
885 * \see cl_io_operations::cio_submit()
887 int cl_io_submit_rw(const struct lu_env *env, struct cl_io *io,
888 enum cl_req_type crt, struct cl_2queue *queue,
889 enum cl_req_priority priority)
891 const struct cl_io_slice *scan;
894 LINVRNT(crt < ARRAY_SIZE(scan->cis_iop->req_op));
897 cl_io_for_each(scan, io) {
898 if (scan->cis_iop->req_op[crt].cio_submit == NULL)
900 result = scan->cis_iop->req_op[crt].cio_submit(env, scan, crt,
906 * If ->cio_submit() failed, no pages were sent.
908 LASSERT(ergo(result != 0, cfs_list_empty(&queue->c2_qout.pl_pages)));
911 EXPORT_SYMBOL(cl_io_submit_rw);
914 * Submit a sync_io and wait for the IO to be finished, or error happens.
915 * If \a timeout is zero, it means to wait for the IO unconditionally.
917 int cl_io_submit_sync(const struct lu_env *env, struct cl_io *io,
918 enum cl_req_type iot, struct cl_2queue *queue,
919 enum cl_req_priority prio, long timeout)
921 struct cl_sync_io *anchor = &cl_env_info(env)->clt_anchor;
925 LASSERT(prio == CRP_NORMAL || prio == CRP_CANCEL);
927 cl_page_list_for_each(pg, &queue->c2_qin) {
928 LASSERT(pg->cp_sync_io == NULL);
929 pg->cp_sync_io = anchor;
932 cl_sync_io_init(anchor, queue->c2_qin.pl_nr);
933 rc = cl_io_submit_rw(env, io, iot, queue, prio);
936 * If some pages weren't sent for any reason (e.g.,
937 * read found up-to-date pages in the cache, or write found
938 * clean pages), count them as completed to avoid infinite
941 cl_page_list_for_each(pg, &queue->c2_qin) {
942 pg->cp_sync_io = NULL;
943 cl_sync_io_note(anchor, +1);
946 /* wait for the IO to be finished. */
947 rc = cl_sync_io_wait(env, io, &queue->c2_qout,
950 LASSERT(cfs_list_empty(&queue->c2_qout.pl_pages));
951 cl_page_list_for_each(pg, &queue->c2_qin)
952 pg->cp_sync_io = NULL;
956 EXPORT_SYMBOL(cl_io_submit_sync);
959 * Cancel an IO which has been submitted by cl_io_submit_rw.
961 int cl_io_cancel(const struct lu_env *env, struct cl_io *io,
962 struct cl_page_list *queue)
964 struct cl_page *page;
967 CERROR("Canceling ongoing page trasmission\n");
968 cl_page_list_for_each(page, queue) {
971 LINVRNT(cl_page_in_io(page, io));
972 rc = cl_page_cancel(env, page);
973 result = result ?: rc;
977 EXPORT_SYMBOL(cl_io_cancel);
982 * Pumps io through iterations calling
984 * - cl_io_iter_init()
994 * - cl_io_iter_fini()
996 * repeatedly until there is no more io to do.
998 int cl_io_loop(const struct lu_env *env, struct cl_io *io)
1002 LINVRNT(cl_io_is_loopable(io));
1008 io->ci_continue = 0;
1009 result = cl_io_iter_init(env, io);
1012 result = cl_io_lock(env, io);
1015 * Notify layers that locks has been taken,
1016 * and do actual i/o.
1018 * - llite: kms, short read;
1019 * - llite: generic_file_read();
1021 result = cl_io_start(env, io);
1023 * Send any remaining pending
1026 * - llite: ll_rw_stats_tally.
1029 cl_io_unlock(env, io);
1030 cl_io_rw_advance(env, io, io->ci_nob - nob);
1033 cl_io_iter_fini(env, io);
1034 } while (result == 0 && io->ci_continue);
1035 RETURN(result < 0 ? result : 0);
1037 EXPORT_SYMBOL(cl_io_loop);
1040 * Adds io slice to the cl_io.
1042 * This is called by cl_object_operations::coo_io_init() methods to add a
1043 * per-layer state to the io. New state is added at the end of
1044 * cl_io::ci_layers list, that is, it is at the bottom of the stack.
1046 * \see cl_lock_slice_add(), cl_req_slice_add(), cl_page_slice_add()
1048 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
1049 struct cl_object *obj,
1050 const struct cl_io_operations *ops)
1052 cfs_list_t *linkage = &slice->cis_linkage;
1054 LASSERT((linkage->prev == NULL && linkage->next == NULL) ||
1055 cfs_list_empty(linkage));
1058 cfs_list_add_tail(linkage, &io->ci_layers);
1060 slice->cis_obj = obj;
1061 slice->cis_iop = ops;
1064 EXPORT_SYMBOL(cl_io_slice_add);
1068 * Initializes page list.
1070 void cl_page_list_init(struct cl_page_list *plist)
1074 CFS_INIT_LIST_HEAD(&plist->pl_pages);
1075 plist->pl_owner = cfs_current();
1078 EXPORT_SYMBOL(cl_page_list_init);
1081 * Adds a page to a page list.
1083 void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page)
1086 /* it would be better to check that page is owned by "current" io, but
1087 * it is not passed here. */
1088 LASSERT(page->cp_owner != NULL);
1089 LINVRNT(plist->pl_owner == cfs_current());
1092 cfs_mutex_lock(&page->cp_mutex);
1094 LASSERT(cfs_list_empty(&page->cp_batch));
1095 cfs_list_add_tail(&page->cp_batch, &plist->pl_pages);
1097 page->cp_queue_ref = lu_ref_add(&page->cp_reference, "queue", plist);
1101 EXPORT_SYMBOL(cl_page_list_add);
1104 * Removes a page from a page list.
1106 void cl_page_list_del(const struct lu_env *env,
1107 struct cl_page_list *plist, struct cl_page *page)
1109 LASSERT(plist->pl_nr > 0);
1110 LINVRNT(plist->pl_owner == cfs_current());
1113 cfs_list_del_init(&page->cp_batch);
1115 cfs_mutex_unlock(&page->cp_mutex);
1118 lu_ref_del_at(&page->cp_reference, page->cp_queue_ref, "queue", plist);
1119 cl_page_put(env, page);
1122 EXPORT_SYMBOL(cl_page_list_del);
1125 * Moves a page from one page list to another.
1127 void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src,
1128 struct cl_page *page)
1130 LASSERT(src->pl_nr > 0);
1131 LINVRNT(dst->pl_owner == cfs_current());
1132 LINVRNT(src->pl_owner == cfs_current());
1135 cfs_list_move_tail(&page->cp_batch, &dst->pl_pages);
1138 lu_ref_set_at(&page->cp_reference,
1139 page->cp_queue_ref, "queue", src, dst);
1142 EXPORT_SYMBOL(cl_page_list_move);
1145 * splice the cl_page_list, just as list head does
1147 void cl_page_list_splice(struct cl_page_list *list, struct cl_page_list *head)
1149 struct cl_page *page;
1150 struct cl_page *tmp;
1152 LINVRNT(list->pl_owner == cfs_current());
1153 LINVRNT(head->pl_owner == cfs_current());
1156 cl_page_list_for_each_safe(page, tmp, list)
1157 cl_page_list_move(head, list, page);
1160 EXPORT_SYMBOL(cl_page_list_splice);
1162 void cl_page_disown0(const struct lu_env *env,
1163 struct cl_io *io, struct cl_page *pg);
1166 * Disowns pages in a queue.
1168 void cl_page_list_disown(const struct lu_env *env,
1169 struct cl_io *io, struct cl_page_list *plist)
1171 struct cl_page *page;
1172 struct cl_page *temp;
1174 LINVRNT(plist->pl_owner == cfs_current());
1177 cl_page_list_for_each_safe(page, temp, plist) {
1178 LASSERT(plist->pl_nr > 0);
1180 cfs_list_del_init(&page->cp_batch);
1182 cfs_mutex_unlock(&page->cp_mutex);
1186 * cl_page_disown0 rather than usual cl_page_disown() is used,
1187 * because pages are possibly in CPS_FREEING state already due
1188 * to the call to cl_page_list_discard().
1191 * XXX cl_page_disown0() will fail if page is not locked.
1193 cl_page_disown0(env, io, page);
1194 lu_ref_del(&page->cp_reference, "queue", plist);
1195 cl_page_put(env, page);
1199 EXPORT_SYMBOL(cl_page_list_disown);
1202 * Releases pages from queue.
1204 void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist)
1206 struct cl_page *page;
1207 struct cl_page *temp;
1209 LINVRNT(plist->pl_owner == cfs_current());
1212 cl_page_list_for_each_safe(page, temp, plist)
1213 cl_page_list_del(env, plist, page);
1214 LASSERT(plist->pl_nr == 0);
1217 EXPORT_SYMBOL(cl_page_list_fini);
1220 * Owns all pages in a queue.
1222 int cl_page_list_own(const struct lu_env *env,
1223 struct cl_io *io, struct cl_page_list *plist)
1225 struct cl_page *page;
1226 struct cl_page *temp;
1230 LINVRNT(plist->pl_owner == cfs_current());
1234 cl_page_list_for_each_safe(page, temp, plist) {
1235 LASSERT(index <= page->cp_index);
1236 index = page->cp_index;
1237 if (cl_page_own(env, io, page) == 0)
1238 result = result ?: page->cp_error;
1240 cl_page_list_del(env, plist, page);
1244 EXPORT_SYMBOL(cl_page_list_own);
1247 * Assumes all pages in a queue.
1249 void cl_page_list_assume(const struct lu_env *env,
1250 struct cl_io *io, struct cl_page_list *plist)
1252 struct cl_page *page;
1254 LINVRNT(plist->pl_owner == cfs_current());
1256 cl_page_list_for_each(page, plist)
1257 cl_page_assume(env, io, page);
1259 EXPORT_SYMBOL(cl_page_list_assume);
1262 * Discards all pages in a queue.
1264 void cl_page_list_discard(const struct lu_env *env, struct cl_io *io,
1265 struct cl_page_list *plist)
1267 struct cl_page *page;
1269 LINVRNT(plist->pl_owner == cfs_current());
1271 cl_page_list_for_each(page, plist)
1272 cl_page_discard(env, io, page);
1275 EXPORT_SYMBOL(cl_page_list_discard);
1278 * Unmaps all pages in a queue from user virtual memory.
1280 int cl_page_list_unmap(const struct lu_env *env, struct cl_io *io,
1281 struct cl_page_list *plist)
1283 struct cl_page *page;
1286 LINVRNT(plist->pl_owner == cfs_current());
1289 cl_page_list_for_each(page, plist) {
1290 result = cl_page_unmap(env, io, page);
1296 EXPORT_SYMBOL(cl_page_list_unmap);
1299 * Initialize dual page queue.
1301 void cl_2queue_init(struct cl_2queue *queue)
1304 cl_page_list_init(&queue->c2_qin);
1305 cl_page_list_init(&queue->c2_qout);
1308 EXPORT_SYMBOL(cl_2queue_init);
1311 * Add a page to the incoming page list of 2-queue.
1313 void cl_2queue_add(struct cl_2queue *queue, struct cl_page *page)
1316 cl_page_list_add(&queue->c2_qin, page);
1319 EXPORT_SYMBOL(cl_2queue_add);
1322 * Disown pages in both lists of a 2-queue.
1324 void cl_2queue_disown(const struct lu_env *env,
1325 struct cl_io *io, struct cl_2queue *queue)
1328 cl_page_list_disown(env, io, &queue->c2_qin);
1329 cl_page_list_disown(env, io, &queue->c2_qout);
1332 EXPORT_SYMBOL(cl_2queue_disown);
1335 * Discard (truncate) pages in both lists of a 2-queue.
1337 void cl_2queue_discard(const struct lu_env *env,
1338 struct cl_io *io, struct cl_2queue *queue)
1341 cl_page_list_discard(env, io, &queue->c2_qin);
1342 cl_page_list_discard(env, io, &queue->c2_qout);
1345 EXPORT_SYMBOL(cl_2queue_discard);
1348 * Assume to own the pages in cl_2queue
1350 void cl_2queue_assume(const struct lu_env *env,
1351 struct cl_io *io, struct cl_2queue *queue)
1353 cl_page_list_assume(env, io, &queue->c2_qin);
1354 cl_page_list_assume(env, io, &queue->c2_qout);
1356 EXPORT_SYMBOL(cl_2queue_assume);
1359 * Finalize both page lists of a 2-queue.
1361 void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue)
1364 cl_page_list_fini(env, &queue->c2_qout);
1365 cl_page_list_fini(env, &queue->c2_qin);
1368 EXPORT_SYMBOL(cl_2queue_fini);
1371 * Initialize a 2-queue to contain \a page in its incoming page list.
1373 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page)
1376 cl_2queue_init(queue);
1377 cl_2queue_add(queue, page);
1380 EXPORT_SYMBOL(cl_2queue_init_page);
1383 * Returns top-level io.
1385 * \see cl_object_top(), cl_page_top().
1387 struct cl_io *cl_io_top(struct cl_io *io)
1390 while (io->ci_parent != NULL)
1394 EXPORT_SYMBOL(cl_io_top);
1397 * Prints human readable representation of \a io to the \a f.
1399 void cl_io_print(const struct lu_env *env, void *cookie,
1400 lu_printer_t printer, const struct cl_io *io)
1405 * Adds request slice to the compound request.
1407 * This is called by cl_device_operations::cdo_req_init() methods to add a
1408 * per-layer state to the request. New state is added at the end of
1409 * cl_req::crq_layers list, that is, it is at the bottom of the stack.
1411 * \see cl_lock_slice_add(), cl_page_slice_add(), cl_io_slice_add()
1413 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
1414 struct cl_device *dev,
1415 const struct cl_req_operations *ops)
1418 cfs_list_add_tail(&slice->crs_linkage, &req->crq_layers);
1419 slice->crs_dev = dev;
1420 slice->crs_ops = ops;
1421 slice->crs_req = req;
1424 EXPORT_SYMBOL(cl_req_slice_add);
1426 static void cl_req_free(const struct lu_env *env, struct cl_req *req)
1430 LASSERT(cfs_list_empty(&req->crq_pages));
1431 LASSERT(req->crq_nrpages == 0);
1432 LINVRNT(cfs_list_empty(&req->crq_layers));
1433 LINVRNT(equi(req->crq_nrobjs > 0, req->crq_o != NULL));
1436 if (req->crq_o != NULL) {
1437 for (i = 0; i < req->crq_nrobjs; ++i) {
1438 struct cl_object *obj = req->crq_o[i].ro_obj;
1440 lu_object_ref_del_at(&obj->co_lu,
1441 req->crq_o[i].ro_obj_ref,
1443 cl_object_put(env, obj);
1446 OBD_FREE(req->crq_o, req->crq_nrobjs * sizeof req->crq_o[0]);
1452 static int cl_req_init(const struct lu_env *env, struct cl_req *req,
1453 struct cl_page *page)
1455 struct cl_device *dev;
1456 struct cl_page_slice *slice;
1461 page = cl_page_top(page);
1463 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
1464 dev = lu2cl_dev(slice->cpl_obj->co_lu.lo_dev);
1465 if (dev->cd_ops->cdo_req_init != NULL) {
1466 result = dev->cd_ops->cdo_req_init(env,
1472 page = page->cp_child;
1473 } while (page != NULL && result == 0);
1478 * Invokes per-request transfer completion call-backs
1479 * (cl_req_operations::cro_completion()) bottom-to-top.
1481 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int rc)
1483 struct cl_req_slice *slice;
1487 * for the lack of list_for_each_entry_reverse_safe()...
1489 while (!cfs_list_empty(&req->crq_layers)) {
1490 slice = cfs_list_entry(req->crq_layers.prev,
1491 struct cl_req_slice, crs_linkage);
1492 cfs_list_del_init(&slice->crs_linkage);
1493 if (slice->crs_ops->cro_completion != NULL)
1494 slice->crs_ops->cro_completion(env, slice, rc);
1496 cl_req_free(env, req);
1499 EXPORT_SYMBOL(cl_req_completion);
1502 * Allocates new transfer request.
1504 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
1505 enum cl_req_type crt, int nr_objects)
1509 LINVRNT(nr_objects > 0);
1516 OBD_ALLOC(req->crq_o, nr_objects * sizeof req->crq_o[0]);
1517 if (req->crq_o != NULL) {
1518 req->crq_nrobjs = nr_objects;
1519 req->crq_type = crt;
1520 CFS_INIT_LIST_HEAD(&req->crq_pages);
1521 CFS_INIT_LIST_HEAD(&req->crq_layers);
1522 result = cl_req_init(env, req, page);
1526 cl_req_completion(env, req, result);
1527 req = ERR_PTR(result);
1530 req = ERR_PTR(-ENOMEM);
1533 EXPORT_SYMBOL(cl_req_alloc);
1536 * Adds a page to a request.
1538 void cl_req_page_add(const struct lu_env *env,
1539 struct cl_req *req, struct cl_page *page)
1541 struct cl_object *obj;
1542 struct cl_req_obj *rqo;
1546 page = cl_page_top(page);
1548 LINVRNT(cl_page_is_vmlocked(env, page));
1549 LASSERT(cfs_list_empty(&page->cp_flight));
1550 LASSERT(page->cp_req == NULL);
1552 CL_PAGE_DEBUG(D_PAGE, env, page, "req %p, %d, %u\n",
1553 req, req->crq_type, req->crq_nrpages);
1555 cfs_list_add_tail(&page->cp_flight, &req->crq_pages);
1558 obj = cl_object_top(page->cp_obj);
1559 for (i = 0, rqo = req->crq_o; obj != rqo->ro_obj; ++i, ++rqo) {
1560 if (rqo->ro_obj == NULL) {
1563 rqo->ro_obj_ref = lu_object_ref_add(&obj->co_lu,
1568 LASSERT(i < req->crq_nrobjs);
1571 EXPORT_SYMBOL(cl_req_page_add);
1574 * Removes a page from a request.
1576 void cl_req_page_done(const struct lu_env *env, struct cl_page *page)
1578 struct cl_req *req = page->cp_req;
1581 page = cl_page_top(page);
1583 LINVRNT(cl_page_is_vmlocked(env, page));
1584 LASSERT(!cfs_list_empty(&page->cp_flight));
1585 LASSERT(req->crq_nrpages > 0);
1587 cfs_list_del_init(&page->cp_flight);
1589 page->cp_req = NULL;
1592 EXPORT_SYMBOL(cl_req_page_done);
1595 * Notifies layers that request is about to depart by calling
1596 * cl_req_operations::cro_prep() top-to-bottom.
1598 int cl_req_prep(const struct lu_env *env, struct cl_req *req)
1602 const struct cl_req_slice *slice;
1606 * Check that the caller of cl_req_alloc() didn't lie about the number
1609 for (i = 0; i < req->crq_nrobjs; ++i)
1610 LASSERT(req->crq_o[i].ro_obj != NULL);
1613 cfs_list_for_each_entry(slice, &req->crq_layers, crs_linkage) {
1614 if (slice->crs_ops->cro_prep != NULL) {
1615 result = slice->crs_ops->cro_prep(env, slice);
1622 EXPORT_SYMBOL(cl_req_prep);
1625 * Fills in attributes that are passed to server together with transfer. Only
1626 * attributes from \a flags may be touched. This can be called multiple times
1627 * for the same request.
1629 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
1630 struct cl_req_attr *attr, obd_valid flags)
1632 const struct cl_req_slice *slice;
1633 struct cl_page *page;
1636 LASSERT(!cfs_list_empty(&req->crq_pages));
1639 /* Take any page to use as a model. */
1640 page = cfs_list_entry(req->crq_pages.next, struct cl_page, cp_flight);
1642 for (i = 0; i < req->crq_nrobjs; ++i) {
1643 cfs_list_for_each_entry(slice, &req->crq_layers, crs_linkage) {
1644 const struct cl_page_slice *scan;
1645 const struct cl_object *obj;
1647 scan = cl_page_at(page,
1648 slice->crs_dev->cd_lu_dev.ld_type);
1649 LASSERT(scan != NULL);
1650 obj = scan->cpl_obj;
1651 if (slice->crs_ops->cro_attr_set != NULL)
1652 slice->crs_ops->cro_attr_set(env, slice, obj,
1658 EXPORT_SYMBOL(cl_req_attr_set);
1660 /* XXX complete(), init_completion(), and wait_for_completion(), until they are
1661 * implemented in libcfs. */
1663 # include <linux/sched.h>
1664 #else /* __KERNEL__ */
1665 # include <liblustre.h>
1669 * Initialize synchronous io wait anchor, for transfer of \a nrpages pages.
1671 void cl_sync_io_init(struct cl_sync_io *anchor, int nrpages)
1674 cfs_waitq_init(&anchor->csi_waitq);
1675 cfs_atomic_set(&anchor->csi_sync_nr, nrpages);
1676 anchor->csi_sync_rc = 0;
1679 EXPORT_SYMBOL(cl_sync_io_init);
1682 * Wait until all transfer completes. Transfer completion routine has to call
1683 * cl_sync_io_note() for every page.
1685 int cl_sync_io_wait(const struct lu_env *env, struct cl_io *io,
1686 struct cl_page_list *queue, struct cl_sync_io *anchor,
1689 struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
1694 LASSERT(timeout >= 0);
1696 rc = l_wait_event(anchor->csi_waitq,
1697 cfs_atomic_read(&anchor->csi_sync_nr) == 0,
1700 CERROR("SYNC IO failed with error: %d, try to cancel "
1701 "%d remaining pages\n",
1702 rc, cfs_atomic_read(&anchor->csi_sync_nr));
1704 (void)cl_io_cancel(env, io, queue);
1706 lwi = (struct l_wait_info) { 0 };
1707 (void)l_wait_event(anchor->csi_waitq,
1708 cfs_atomic_read(&anchor->csi_sync_nr) == 0,
1711 rc = anchor->csi_sync_rc;
1713 LASSERT(cfs_atomic_read(&anchor->csi_sync_nr) == 0);
1714 cl_page_list_assume(env, io, queue);
1715 POISON(anchor, 0x5a, sizeof *anchor);
1718 EXPORT_SYMBOL(cl_sync_io_wait);
1721 * Indicate that transfer of a single page completed.
1723 void cl_sync_io_note(struct cl_sync_io *anchor, int ioret)
1726 if (anchor->csi_sync_rc == 0 && ioret < 0)
1727 anchor->csi_sync_rc = ioret;
1729 * Synchronous IO done without releasing page lock (e.g., as a part of
1730 * ->{prepare,commit}_write(). Completion is used to signal the end of
1733 LASSERT(cfs_atomic_read(&anchor->csi_sync_nr) > 0);
1734 if (cfs_atomic_dec_and_test(&anchor->csi_sync_nr))
1735 cfs_waitq_broadcast(&anchor->csi_waitq);
1738 EXPORT_SYMBOL(cl_sync_io_note);