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 <obd_class.h>
47 #include <obd_support.h>
48 #include <lustre_fid.h>
49 #include <libcfs/list.h>
50 /* lu_time_global_{init,fini}() */
53 #include <cl_object.h>
54 #include "cl_internal.h"
56 /*****************************************************************************
62 #define cl_io_for_each(slice, io) \
63 list_for_each_entry((slice), &io->ci_layers, cis_linkage)
64 #define cl_io_for_each_reverse(slice, io) \
65 list_for_each_entry_reverse((slice), &io->ci_layers, cis_linkage)
67 static inline int cl_io_type_is_valid(enum cl_io_type type)
69 return CIT_READ <= type && type < CIT_OP_NR;
72 static inline int cl_io_is_loopable(const struct cl_io *io)
74 return cl_io_type_is_valid(io->ci_type) && io->ci_type != CIT_MISC;
78 * Returns true iff there is an IO ongoing in the given environment.
80 int cl_io_is_going(const struct lu_env *env)
82 return cl_env_info(env)->clt_current_io != NULL;
84 EXPORT_SYMBOL(cl_io_is_going);
87 * cl_io invariant that holds at all times when exported cl_io_*() functions
88 * are entered and left.
90 static int cl_io_invariant(const struct cl_io *io)
97 * io can own pages only when it is ongoing. Sub-io might
98 * still be in CIS_LOCKED state when top-io is in
101 ergo(io->ci_owned_nr > 0, io->ci_state == CIS_IO_GOING ||
102 (io->ci_state == CIS_LOCKED && up != NULL));
106 * Finalize \a io, by calling cl_io_operations::cio_fini() bottom-to-top.
108 void cl_io_fini(const struct lu_env *env, struct cl_io *io)
110 struct cl_io_slice *slice;
111 struct cl_thread_info *info;
113 LINVRNT(cl_io_type_is_valid(io->ci_type));
114 LINVRNT(cl_io_invariant(io));
117 while (!list_empty(&io->ci_layers)) {
118 slice = container_of(io->ci_layers.next, struct cl_io_slice,
120 list_del_init(&slice->cis_linkage);
121 if (slice->cis_iop->op[io->ci_type].cio_fini != NULL)
122 slice->cis_iop->op[io->ci_type].cio_fini(env, slice);
124 * Invalidate slice to catch use after free. This assumes that
125 * slices are allocated within session and can be touched
126 * after ->cio_fini() returns.
128 slice->cis_io = NULL;
130 io->ci_state = CIS_FINI;
131 info = cl_env_info(env);
132 if (info->clt_current_io == io)
133 info->clt_current_io = NULL;
136 EXPORT_SYMBOL(cl_io_fini);
138 static int cl_io_init0(const struct lu_env *env, struct cl_io *io,
139 enum cl_io_type iot, struct cl_object *obj)
141 struct cl_object *scan;
144 LINVRNT(io->ci_state == CIS_ZERO || io->ci_state == CIS_FINI);
145 LINVRNT(cl_io_type_is_valid(iot));
146 LINVRNT(cl_io_invariant(io));
150 CFS_INIT_LIST_HEAD(&io->ci_lockset.cls_todo);
151 CFS_INIT_LIST_HEAD(&io->ci_lockset.cls_curr);
152 CFS_INIT_LIST_HEAD(&io->ci_lockset.cls_done);
153 CFS_INIT_LIST_HEAD(&io->ci_layers);
156 cl_object_for_each(scan, obj) {
157 if (scan->co_ops->coo_io_init != NULL) {
158 result = scan->co_ops->coo_io_init(env, scan, io);
164 io->ci_state = CIS_INIT;
169 * Initialize sub-io, by calling cl_io_operations::cio_init() top-to-bottom.
171 * \pre obj != cl_object_top(obj)
173 int cl_io_sub_init(const struct lu_env *env, struct cl_io *io,
174 enum cl_io_type iot, struct cl_object *obj)
176 struct cl_thread_info *info = cl_env_info(env);
178 LASSERT(obj != cl_object_top(obj));
179 if (info->clt_current_io == NULL)
180 info->clt_current_io = io;
181 return cl_io_init0(env, io, iot, obj);
183 EXPORT_SYMBOL(cl_io_sub_init);
186 * Initialize \a io, by calling cl_io_operations::cio_init() top-to-bottom.
188 * Caller has to call cl_io_fini() after a call to cl_io_init(), no matter
189 * what the latter returned.
191 * \pre obj == cl_object_top(obj)
192 * \pre cl_io_type_is_valid(iot)
193 * \post cl_io_type_is_valid(io->ci_type) && io->ci_type == iot
195 int cl_io_init(const struct lu_env *env, struct cl_io *io,
196 enum cl_io_type iot, struct cl_object *obj)
198 struct cl_thread_info *info = cl_env_info(env);
200 LASSERT(obj == cl_object_top(obj));
201 LASSERT(info->clt_current_io == NULL);
203 info->clt_current_io = io;
204 return cl_io_init0(env, io, iot, obj);
206 EXPORT_SYMBOL(cl_io_init);
209 * Initialize read or write io.
211 * \pre iot == CIT_READ || iot == CIT_WRITE
213 int cl_io_rw_init(const struct lu_env *env, struct cl_io *io,
214 enum cl_io_type iot, loff_t pos, size_t count)
216 LINVRNT(iot == CIT_READ || iot == CIT_WRITE);
217 LINVRNT(io->ci_obj != NULL);
220 LU_OBJECT_HEADER(D_VFSTRACE, env, &io->ci_obj->co_lu,
221 "io range: %u ["LPU64", "LPU64") %u %u\n",
222 iot, (__u64)pos, (__u64)pos + count,
223 io->u.ci_rw.crw_nonblock, io->u.ci_wr.wr_append);
224 io->u.ci_rw.crw_pos = pos;
225 io->u.ci_rw.crw_count = count;
226 RETURN(cl_io_init(env, io, iot, io->ci_obj));
228 EXPORT_SYMBOL(cl_io_rw_init);
230 static inline const struct lu_fid *
231 cl_lock_descr_fid(const struct cl_lock_descr *descr)
233 return lu_object_fid(&descr->cld_obj->co_lu);
236 static int cl_lock_descr_cmp(const struct cl_lock_descr *d0,
237 const struct cl_lock_descr *d1)
239 return lu_fid_cmp(cl_lock_descr_fid(d0), cl_lock_descr_fid(d1)) ?:
240 __diff_normalize(d0->cld_start, d1->cld_start);
244 * Sort locks in lexicographical order of their (fid, start-offset) pairs.
246 static void cl_io_locks_sort(struct cl_io *io)
251 /* hidden treasure: bubble sort for now. */
253 struct cl_io_lock_link *curr;
254 struct cl_io_lock_link *prev;
255 struct cl_io_lock_link *temp;
260 list_for_each_entry_safe(curr, temp, &io->ci_lockset.cls_todo,
263 switch (cl_lock_descr_cmp(&prev->cill_descr,
264 &curr->cill_descr)) {
267 * IMPOSSIBLE: Identical locks are
274 list_move_tail(&curr->cill_linkage,
275 &prev->cill_linkage);
277 continue; /* don't change prev: it's
278 * still "previous" */
279 case -1: /* already in order */
290 * Check whether \a queue contains locks matching \a need.
292 * \retval +ve there is a matching lock in the \a queue
293 * \retval 0 there are no matching locks in the \a queue
295 int cl_queue_match(const struct list_head *queue,
296 const struct cl_lock_descr *need)
298 struct cl_io_lock_link *scan;
301 list_for_each_entry(scan, queue, cill_linkage) {
302 if (cl_lock_descr_match(&scan->cill_descr, need))
307 EXPORT_SYMBOL(cl_queue_match);
309 static int cl_lockset_match(const struct cl_lockset *set,
310 const struct cl_lock_descr *need, int all_queues)
312 return (all_queues ? cl_queue_match(&set->cls_todo, need) : 0) ||
313 cl_queue_match(&set->cls_curr, need) ||
314 cl_queue_match(&set->cls_done, need);
317 static int cl_lockset_lock_one(const struct lu_env *env,
318 struct cl_io *io, struct cl_lockset *set,
319 struct cl_io_lock_link *link)
321 struct cl_lock *lock;
326 lock = cl_lock_request(env, io, &link->cill_descr, "io", io);
328 link->cill_lock = lock;
329 list_move(&link->cill_linkage, &set->cls_curr);
330 if (!(link->cill_descr.cld_enq_flags & CEF_ASYNC)) {
331 result = cl_wait(env, lock);
333 list_move(&link->cill_linkage, &set->cls_done);
337 result = PTR_ERR(lock);
341 static void cl_lock_link_fini(const struct lu_env *env, struct cl_io *io,
342 struct cl_io_lock_link *link)
344 struct cl_lock *lock = link->cill_lock;
347 list_del_init(&link->cill_linkage);
349 cl_lock_release(env, lock, "io", io);
350 link->cill_lock = NULL;
352 if (link->cill_fini != NULL)
353 link->cill_fini(env, link);
357 static int cl_lockset_lock(const struct lu_env *env, struct cl_io *io,
358 struct cl_lockset *set)
360 struct cl_io_lock_link *link;
361 struct cl_io_lock_link *temp;
362 struct cl_lock *lock;
367 list_for_each_entry_safe(link, temp, &set->cls_todo, cill_linkage) {
368 if (!cl_lockset_match(set, &link->cill_descr, 0)) {
369 /* XXX some locking to guarantee that locks aren't
370 * expanded in between. */
371 result = cl_lockset_lock_one(env, io, set, link);
375 cl_lock_link_fini(env, io, link);
378 list_for_each_entry_safe(link, temp,
379 &set->cls_curr, cill_linkage) {
380 lock = link->cill_lock;
381 result = cl_wait(env, lock);
383 list_move(&link->cill_linkage, &set->cls_done);
392 * Takes locks necessary for the current iteration of io.
394 * Calls cl_io_operations::cio_lock() top-to-bottom to collect locks required
395 * by layers for the current iteration. Then sort locks (to avoid dead-locks),
398 int cl_io_lock(const struct lu_env *env, struct cl_io *io)
400 const struct cl_io_slice *scan;
403 LINVRNT(cl_io_is_loopable(io));
404 LINVRNT(io->ci_state == CIS_IT_STARTED);
405 LINVRNT(cl_io_invariant(io));
408 cl_io_for_each(scan, io) {
409 if (scan->cis_iop->op[io->ci_type].cio_lock == NULL)
411 result = scan->cis_iop->op[io->ci_type].cio_lock(env, scan);
416 cl_io_locks_sort(io);
417 result = cl_lockset_lock(env, io, &io->ci_lockset);
420 cl_io_unlock(env, io);
422 io->ci_state = CIS_LOCKED;
425 EXPORT_SYMBOL(cl_io_lock);
428 * Release locks takes by io.
430 void cl_io_unlock(const struct lu_env *env, struct cl_io *io)
432 struct cl_lockset *set;
433 struct cl_io_lock_link *link;
434 struct cl_io_lock_link *temp;
435 const struct cl_io_slice *scan;
437 LASSERT(cl_io_is_loopable(io));
438 LASSERT(CIS_IT_STARTED <= io->ci_state && io->ci_state < CIS_UNLOCKED);
439 LINVRNT(cl_io_invariant(io));
442 set = &io->ci_lockset;
444 list_for_each_entry_safe(link, temp, &set->cls_todo, cill_linkage)
445 cl_lock_link_fini(env, io, link);
447 list_for_each_entry_safe(link, temp, &set->cls_curr, cill_linkage)
448 cl_lock_link_fini(env, io, link);
450 list_for_each_entry_safe(link, temp, &set->cls_done, cill_linkage) {
451 cl_unuse(env, link->cill_lock);
452 cl_lock_link_fini(env, io, link);
454 cl_io_for_each_reverse(scan, io) {
455 if (scan->cis_iop->op[io->ci_type].cio_unlock != NULL)
456 scan->cis_iop->op[io->ci_type].cio_unlock(env, scan);
458 io->ci_state = CIS_UNLOCKED;
459 LASSERT(!cl_env_info(env)->clt_counters[CNL_TOP].ctc_nr_locks_acquired);
462 EXPORT_SYMBOL(cl_io_unlock);
465 * Prepares next iteration of io.
467 * Calls cl_io_operations::cio_iter_init() top-to-bottom. This exists to give
468 * layers a chance to modify io parameters, e.g., so that lov can restrict io
469 * to a single stripe.
471 int cl_io_iter_init(const struct lu_env *env, struct cl_io *io)
473 const struct cl_io_slice *scan;
476 LINVRNT(cl_io_is_loopable(io));
477 LINVRNT(io->ci_state == CIS_INIT || io->ci_state == CIS_IT_ENDED);
478 LINVRNT(cl_io_invariant(io));
482 cl_io_for_each(scan, io) {
483 if (scan->cis_iop->op[io->ci_type].cio_iter_init == NULL)
485 result = scan->cis_iop->op[io->ci_type].cio_iter_init(env,
491 io->ci_state = CIS_IT_STARTED;
494 EXPORT_SYMBOL(cl_io_iter_init);
497 * Finalizes io iteration.
499 * Calls cl_io_operations::cio_iter_fini() bottom-to-top.
501 void cl_io_iter_fini(const struct lu_env *env, struct cl_io *io)
503 const struct cl_io_slice *scan;
505 LINVRNT(cl_io_is_loopable(io));
506 LINVRNT(io->ci_state == CIS_UNLOCKED);
507 LINVRNT(cl_io_invariant(io));
510 cl_io_for_each_reverse(scan, io) {
511 if (scan->cis_iop->op[io->ci_type].cio_iter_fini != NULL)
512 scan->cis_iop->op[io->ci_type].cio_iter_fini(env, scan);
514 io->ci_state = CIS_IT_ENDED;
517 EXPORT_SYMBOL(cl_io_iter_fini);
520 * Records that read or write io progressed \a nob bytes forward.
522 void cl_io_rw_advance(const struct lu_env *env, struct cl_io *io, size_t nob)
524 const struct cl_io_slice *scan;
526 LINVRNT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
528 LINVRNT(cl_io_is_loopable(io));
529 LINVRNT(cl_io_invariant(io));
533 io->u.ci_rw.crw_pos += nob;
534 io->u.ci_rw.crw_count -= nob;
536 /* layers have to be notified. */
537 cl_io_for_each_reverse(scan, io) {
538 if (scan->cis_iop->op[io->ci_type].cio_advance != NULL)
539 scan->cis_iop->op[io->ci_type].cio_advance(env, scan,
544 EXPORT_SYMBOL(cl_io_rw_advance);
547 * Adds a lock to a lockset.
549 int cl_io_lock_add(const struct lu_env *env, struct cl_io *io,
550 struct cl_io_lock_link *link)
555 if (cl_lockset_match(&io->ci_lockset, &link->cill_descr, 1))
558 list_add(&link->cill_linkage, &io->ci_lockset.cls_todo);
563 EXPORT_SYMBOL(cl_io_lock_add);
565 static void cl_free_io_lock_link(const struct lu_env *env,
566 struct cl_io_lock_link *link)
572 * Allocates new lock link, and uses it to add a lock to a lockset.
574 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
575 struct cl_lock_descr *descr)
577 struct cl_io_lock_link *link;
583 link->cill_descr = *descr;
584 link->cill_fini = cl_free_io_lock_link;
585 result = cl_io_lock_add(env, io, link);
586 if (result) /* lock match */
587 link->cill_fini(env, link);
593 EXPORT_SYMBOL(cl_io_lock_alloc_add);
596 * Starts io by calling cl_io_operations::cio_start() top-to-bottom.
598 int cl_io_start(const struct lu_env *env, struct cl_io *io)
600 const struct cl_io_slice *scan;
603 LINVRNT(cl_io_is_loopable(io));
604 LINVRNT(io->ci_state == CIS_LOCKED);
605 LINVRNT(cl_io_invariant(io));
608 io->ci_state = CIS_IO_GOING;
609 cl_io_for_each(scan, io) {
610 if (scan->cis_iop->op[io->ci_type].cio_start == NULL)
612 result = scan->cis_iop->op[io->ci_type].cio_start(env, scan);
620 EXPORT_SYMBOL(cl_io_start);
623 * Wait until current io iteration is finished by calling
624 * cl_io_operations::cio_end() bottom-to-top.
626 void cl_io_end(const struct lu_env *env, struct cl_io *io)
628 const struct cl_io_slice *scan;
630 LINVRNT(cl_io_is_loopable(io));
631 LINVRNT(io->ci_state == CIS_IO_GOING);
632 LINVRNT(cl_io_invariant(io));
635 cl_io_for_each_reverse(scan, io) {
636 if (scan->cis_iop->op[io->ci_type].cio_end != NULL)
637 scan->cis_iop->op[io->ci_type].cio_end(env, scan);
638 /* TODO: error handling. */
640 io->ci_state = CIS_IO_FINISHED;
643 EXPORT_SYMBOL(cl_io_end);
645 static const struct cl_page_slice *
646 cl_io_slice_page(const struct cl_io_slice *ios, struct cl_page *page)
648 const struct cl_page_slice *slice;
650 slice = cl_page_at(page, ios->cis_obj->co_lu.lo_dev->ld_type);
651 LINVRNT(slice != NULL);
656 * True iff \a page is within \a io range.
658 static int cl_page_in_io(const struct cl_page *page, const struct cl_io *io)
665 idx = page->cp_index;
666 switch (io->ci_type) {
670 * check that [start, end) and [pos, pos + count) extents
673 start = cl_offset(page->cp_obj, idx);
674 end = cl_offset(page->cp_obj, idx + 1);
675 result = io->u.ci_rw.crw_pos < end &&
676 start < io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count;
679 result = io->u.ci_fault.ft_index == idx;
688 * Called by read io, when page has to be read from the server.
690 * \see cl_io_operations::cio_read_page()
692 int cl_io_read_page(const struct lu_env *env, struct cl_io *io,
693 struct cl_page *page)
695 const struct cl_io_slice *scan;
696 struct cl_2queue *queue;
699 LINVRNT(io->ci_type == CIT_READ || io->ci_type == CIT_FAULT);
700 LINVRNT(cl_page_is_owned(page, io));
701 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
702 LINVRNT(cl_page_in_io(page, io));
703 LINVRNT(cl_io_invariant(io));
706 queue = &io->ci_queue;
708 cl_2queue_init(queue);
710 * ->cio_read_page() methods called in the loop below are supposed to
711 * never block waiting for network (the only subtle point is the
712 * creation of new pages for read-ahead that might result in cache
713 * shrinking, but currently only clean pages are shrunk and this
714 * requires no network io).
716 * Should this ever starts blocking, retry loop would be needed for
717 * "parallel io" (see CLO_REPEAT loops in cl_lock.c).
719 cl_io_for_each(scan, io) {
720 if (scan->cis_iop->cio_read_page != NULL) {
721 const struct cl_page_slice *slice;
723 slice = cl_io_slice_page(scan, page);
724 LINVRNT(slice != NULL);
725 result = scan->cis_iop->cio_read_page(env, scan, slice);
731 result = cl_io_submit_rw(env, io, CRT_READ, queue, CRP_NORMAL);
733 * Unlock unsent pages in case of error.
735 cl_page_list_disown(env, io, &queue->c2_qin);
736 cl_2queue_fini(env, queue);
739 EXPORT_SYMBOL(cl_io_read_page);
742 * Called by write io to prepare page to receive data from user buffer.
744 * \see cl_io_operations::cio_prepare_write()
746 int cl_io_prepare_write(const struct lu_env *env, struct cl_io *io,
747 struct cl_page *page, unsigned from, unsigned to)
749 const struct cl_io_slice *scan;
752 LINVRNT(io->ci_type == CIT_WRITE);
753 LINVRNT(cl_page_is_owned(page, io));
754 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
755 LINVRNT(cl_io_invariant(io));
756 LASSERT(cl_page_in_io(page, io));
759 cl_io_for_each_reverse(scan, io) {
760 if (scan->cis_iop->cio_prepare_write != NULL) {
761 const struct cl_page_slice *slice;
763 slice = cl_io_slice_page(scan, page);
764 result = scan->cis_iop->cio_prepare_write(env, scan,
773 EXPORT_SYMBOL(cl_io_prepare_write);
776 * Called by write io after user data were copied into a page.
778 * \see cl_io_operations::cio_commit_write()
780 int cl_io_commit_write(const struct lu_env *env, struct cl_io *io,
781 struct cl_page *page, unsigned from, unsigned to)
783 const struct cl_io_slice *scan;
786 LINVRNT(io->ci_type == CIT_WRITE);
787 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
788 LINVRNT(cl_io_invariant(io));
790 * XXX Uh... not nice. Top level cl_io_commit_write() call (vvp->lov)
791 * already called cl_page_cache_add(), moving page into CPS_CACHED
792 * state. Better (and more general) way of dealing with such situation
795 LASSERT(cl_page_is_owned(page, io) || page->cp_parent != NULL);
796 LASSERT(cl_page_in_io(page, io));
799 cl_io_for_each(scan, io) {
800 if (scan->cis_iop->cio_commit_write != NULL) {
801 const struct cl_page_slice *slice;
803 slice = cl_io_slice_page(scan, page);
804 result = scan->cis_iop->cio_commit_write(env, scan,
811 LINVRNT(result <= 0);
814 EXPORT_SYMBOL(cl_io_commit_write);
817 * Submits a list of pages for immediate io.
819 * After the function gets returned, The submitted pages are moved to
820 * queue->c2_qout queue, and queue->c2_qin contain both the pages don't need
821 * to be submitted, and the pages are errant to submit.
823 * \returns 0 if at least one page was submitted, error code otherwise.
824 * \see cl_io_operations::cio_submit()
826 int cl_io_submit_rw(const struct lu_env *env, struct cl_io *io,
827 enum cl_req_type crt, struct cl_2queue *queue,
828 enum cl_req_priority priority)
830 const struct cl_io_slice *scan;
833 LINVRNT(crt < ARRAY_SIZE(scan->cis_iop->req_op));
836 cl_io_for_each(scan, io) {
837 if (scan->cis_iop->req_op[crt].cio_submit == NULL)
839 result = scan->cis_iop->req_op[crt].cio_submit(env, scan, crt,
845 * If ->cio_submit() failed, no pages were sent.
847 LASSERT(ergo(result != 0, list_empty(&queue->c2_qout.pl_pages)));
850 EXPORT_SYMBOL(cl_io_submit_rw);
853 * Submit a sync_io and wait for the IO to be finished, or error happens.
854 * If \a timeout is zero, it means to wait for the IO unconditionally.
856 int cl_io_submit_sync(const struct lu_env *env, struct cl_io *io,
857 enum cl_req_type iot, struct cl_2queue *queue,
858 enum cl_req_priority prio, long timeout)
860 struct cl_sync_io *anchor = &cl_env_info(env)->clt_anchor;
864 LASSERT(prio == CRP_NORMAL || prio == CRP_CANCEL);
866 cl_page_list_for_each(pg, &queue->c2_qin) {
867 LASSERT(pg->cp_sync_io == NULL);
868 pg->cp_sync_io = anchor;
871 cl_sync_io_init(anchor, queue->c2_qin.pl_nr);
872 rc = cl_io_submit_rw(env, io, iot, queue, prio);
875 * If some pages weren't sent for any reason (e.g.,
876 * read found up-to-date pages in the cache, or write found
877 * clean pages), count them as completed to avoid infinite
880 cl_page_list_for_each(pg, &queue->c2_qin) {
881 pg->cp_sync_io = NULL;
882 cl_sync_io_note(anchor, +1);
885 /* wait for the IO to be finished. */
886 rc = cl_sync_io_wait(env, io, &queue->c2_qout,
889 LASSERT(list_empty(&queue->c2_qout.pl_pages));
890 cl_page_list_for_each(pg, &queue->c2_qin)
891 pg->cp_sync_io = NULL;
895 EXPORT_SYMBOL(cl_io_submit_sync);
898 * Cancel an IO which has been submitted by cl_io_submit_rw.
900 int cl_io_cancel(const struct lu_env *env, struct cl_io *io,
901 struct cl_page_list *queue)
903 struct cl_page *page;
906 CERROR("Canceling ongoing page trasmission\n");
907 cl_page_list_for_each(page, queue) {
910 LINVRNT(cl_page_in_io(page, io));
911 rc = cl_page_cancel(env, page);
912 result = result ?: rc;
916 EXPORT_SYMBOL(cl_io_cancel);
921 * Pumps io through iterations calling
923 * - cl_io_iter_init()
933 * - cl_io_iter_fini()
935 * repeatedly until there is no more io to do.
937 int cl_io_loop(const struct lu_env *env, struct cl_io *io)
941 LINVRNT(cl_io_is_loopable(io));
948 result = cl_io_iter_init(env, io);
951 result = cl_io_lock(env, io);
954 * Notify layers that locks has been taken,
957 * - llite: kms, short read;
958 * - llite: generic_file_read();
960 result = cl_io_start(env, io);
962 * Send any remaining pending
965 * - llite: ll_rw_stats_tally.
968 cl_io_unlock(env, io);
969 cl_io_rw_advance(env, io, io->ci_nob - nob);
972 cl_io_iter_fini(env, io);
973 } while (result == 0 && io->ci_continue);
974 RETURN(result < 0 ? result : 0);
976 EXPORT_SYMBOL(cl_io_loop);
979 * Adds io slice to the cl_io.
981 * This is called by cl_object_operations::coo_io_init() methods to add a
982 * per-layer state to the io. New state is added at the end of
983 * cl_io::ci_layers list, that is, it is at the bottom of the stack.
985 * \see cl_lock_slice_add(), cl_req_slice_add(), cl_page_slice_add()
987 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
988 struct cl_object *obj,
989 const struct cl_io_operations *ops)
991 struct list_head *linkage = &slice->cis_linkage;
993 LASSERT((linkage->prev == NULL && linkage->next == NULL) ||
994 list_empty(linkage));
997 list_add_tail(linkage, &io->ci_layers);
999 slice->cis_obj = obj;
1000 slice->cis_iop = ops;
1003 EXPORT_SYMBOL(cl_io_slice_add);
1007 * Initializes page list.
1009 void cl_page_list_init(struct cl_page_list *plist)
1013 CFS_INIT_LIST_HEAD(&plist->pl_pages);
1014 plist->pl_owner = cfs_current();
1017 EXPORT_SYMBOL(cl_page_list_init);
1020 * Adds a page to a page list.
1022 void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page)
1025 /* it would be better to check that page is owned by "current" io, but
1026 * it is not passed here. */
1027 LASSERT(page->cp_owner != NULL);
1028 LINVRNT(plist->pl_owner == cfs_current());
1031 mutex_lock(&page->cp_mutex);
1033 LASSERT(list_empty(&page->cp_batch));
1034 list_add_tail(&page->cp_batch, &plist->pl_pages);
1036 page->cp_queue_ref = lu_ref_add(&page->cp_reference, "queue", plist);
1040 EXPORT_SYMBOL(cl_page_list_add);
1043 * Removes a page from a page list.
1045 void cl_page_list_del(const struct lu_env *env,
1046 struct cl_page_list *plist, struct cl_page *page)
1048 LASSERT(plist->pl_nr > 0);
1049 LINVRNT(plist->pl_owner == cfs_current());
1052 list_del_init(&page->cp_batch);
1054 mutex_unlock(&page->cp_mutex);
1057 lu_ref_del_at(&page->cp_reference, page->cp_queue_ref, "queue", plist);
1058 cl_page_put(env, page);
1061 EXPORT_SYMBOL(cl_page_list_del);
1064 * Moves a page from one page list to another.
1066 void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src,
1067 struct cl_page *page)
1069 LASSERT(src->pl_nr > 0);
1070 LINVRNT(dst->pl_owner == cfs_current());
1071 LINVRNT(src->pl_owner == cfs_current());
1074 list_move_tail(&page->cp_batch, &dst->pl_pages);
1077 lu_ref_set_at(&page->cp_reference,
1078 page->cp_queue_ref, "queue", src, dst);
1081 EXPORT_SYMBOL(cl_page_list_move);
1084 * splice the cl_page_list, just as list head does
1086 void cl_page_list_splice(struct cl_page_list *list, struct cl_page_list *head)
1088 struct cl_page *page;
1089 struct cl_page *tmp;
1091 LINVRNT(list->pl_owner == cfs_current());
1092 LINVRNT(head->pl_owner == cfs_current());
1095 cl_page_list_for_each_safe(page, tmp, list)
1096 cl_page_list_move(head, list, page);
1099 EXPORT_SYMBOL(cl_page_list_splice);
1101 void cl_page_disown0(const struct lu_env *env,
1102 struct cl_io *io, struct cl_page *pg);
1105 * Disowns pages in a queue.
1107 void cl_page_list_disown(const struct lu_env *env,
1108 struct cl_io *io, struct cl_page_list *plist)
1110 struct cl_page *page;
1111 struct cl_page *temp;
1113 LINVRNT(plist->pl_owner == cfs_current());
1116 cl_page_list_for_each_safe(page, temp, plist) {
1117 LASSERT(plist->pl_nr > 0);
1119 list_del_init(&page->cp_batch);
1121 mutex_unlock(&page->cp_mutex);
1125 * cl_page_disown0 rather than usual cl_page_disown() is used,
1126 * because pages are possibly in CPS_FREEING state already due
1127 * to the call to cl_page_list_discard().
1130 * XXX cl_page_disown0() will fail if page is not locked.
1132 cl_page_disown0(env, io, page);
1133 lu_ref_del(&page->cp_reference, "queue", plist);
1134 cl_page_put(env, page);
1138 EXPORT_SYMBOL(cl_page_list_disown);
1141 * Releases pages from queue.
1143 void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist)
1145 struct cl_page *page;
1146 struct cl_page *temp;
1148 LINVRNT(plist->pl_owner == cfs_current());
1151 cl_page_list_for_each_safe(page, temp, plist)
1152 cl_page_list_del(env, plist, page);
1153 LASSERT(plist->pl_nr == 0);
1156 EXPORT_SYMBOL(cl_page_list_fini);
1159 * Owns all pages in a queue.
1161 int cl_page_list_own(const struct lu_env *env,
1162 struct cl_io *io, struct cl_page_list *plist)
1164 struct cl_page *page;
1165 struct cl_page *temp;
1169 LINVRNT(plist->pl_owner == cfs_current());
1173 cl_page_list_for_each_safe(page, temp, plist) {
1174 LASSERT(index <= page->cp_index);
1175 index = page->cp_index;
1176 if (cl_page_own(env, io, page) == 0)
1177 result = result ?: page->cp_error;
1179 cl_page_list_del(env, plist, page);
1183 EXPORT_SYMBOL(cl_page_list_own);
1186 * Assumes all pages in a queue.
1188 void cl_page_list_assume(const struct lu_env *env,
1189 struct cl_io *io, struct cl_page_list *plist)
1191 struct cl_page *page;
1193 LINVRNT(plist->pl_owner == cfs_current());
1195 cl_page_list_for_each(page, plist)
1196 cl_page_assume(env, io, page);
1198 EXPORT_SYMBOL(cl_page_list_assume);
1201 * Discards all pages in a queue.
1203 void cl_page_list_discard(const struct lu_env *env, struct cl_io *io,
1204 struct cl_page_list *plist)
1206 struct cl_page *page;
1208 LINVRNT(plist->pl_owner == cfs_current());
1210 cl_page_list_for_each(page, plist)
1211 cl_page_discard(env, io, page);
1214 EXPORT_SYMBOL(cl_page_list_discard);
1217 * Unmaps all pages in a queue from user virtual memory.
1219 int cl_page_list_unmap(const struct lu_env *env, struct cl_io *io,
1220 struct cl_page_list *plist)
1222 struct cl_page *page;
1225 LINVRNT(plist->pl_owner == cfs_current());
1228 cl_page_list_for_each(page, plist) {
1229 result = cl_page_unmap(env, io, page);
1235 EXPORT_SYMBOL(cl_page_list_unmap);
1238 * Initialize dual page queue.
1240 void cl_2queue_init(struct cl_2queue *queue)
1243 cl_page_list_init(&queue->c2_qin);
1244 cl_page_list_init(&queue->c2_qout);
1247 EXPORT_SYMBOL(cl_2queue_init);
1250 * Add a page to the incoming page list of 2-queue.
1252 void cl_2queue_add(struct cl_2queue *queue, struct cl_page *page)
1255 cl_page_list_add(&queue->c2_qin, page);
1258 EXPORT_SYMBOL(cl_2queue_add);
1261 * Disown pages in both lists of a 2-queue.
1263 void cl_2queue_disown(const struct lu_env *env,
1264 struct cl_io *io, struct cl_2queue *queue)
1267 cl_page_list_disown(env, io, &queue->c2_qin);
1268 cl_page_list_disown(env, io, &queue->c2_qout);
1271 EXPORT_SYMBOL(cl_2queue_disown);
1274 * Discard (truncate) pages in both lists of a 2-queue.
1276 void cl_2queue_discard(const struct lu_env *env,
1277 struct cl_io *io, struct cl_2queue *queue)
1280 cl_page_list_discard(env, io, &queue->c2_qin);
1281 cl_page_list_discard(env, io, &queue->c2_qout);
1284 EXPORT_SYMBOL(cl_2queue_discard);
1287 * Assume to own the pages in cl_2queue
1289 void cl_2queue_assume(const struct lu_env *env,
1290 struct cl_io *io, struct cl_2queue *queue)
1292 cl_page_list_assume(env, io, &queue->c2_qin);
1293 cl_page_list_assume(env, io, &queue->c2_qout);
1295 EXPORT_SYMBOL(cl_2queue_assume);
1298 * Finalize both page lists of a 2-queue.
1300 void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue)
1303 cl_page_list_fini(env, &queue->c2_qout);
1304 cl_page_list_fini(env, &queue->c2_qin);
1307 EXPORT_SYMBOL(cl_2queue_fini);
1310 * Initialize a 2-queue to contain \a page in its incoming page list.
1312 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page)
1315 cl_2queue_init(queue);
1316 cl_2queue_add(queue, page);
1319 EXPORT_SYMBOL(cl_2queue_init_page);
1322 * Returns top-level io.
1324 * \see cl_object_top(), cl_page_top().
1326 struct cl_io *cl_io_top(struct cl_io *io)
1329 while (io->ci_parent != NULL)
1333 EXPORT_SYMBOL(cl_io_top);
1336 * Prints human readable representation of \a io to the \a f.
1338 void cl_io_print(const struct lu_env *env, void *cookie,
1339 lu_printer_t printer, const struct cl_io *io)
1344 * Adds request slice to the compound request.
1346 * This is called by cl_device_operations::cdo_req_init() methods to add a
1347 * per-layer state to the request. New state is added at the end of
1348 * cl_req::crq_layers list, that is, it is at the bottom of the stack.
1350 * \see cl_lock_slice_add(), cl_page_slice_add(), cl_io_slice_add()
1352 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
1353 struct cl_device *dev,
1354 const struct cl_req_operations *ops)
1357 list_add_tail(&slice->crs_linkage, &req->crq_layers);
1358 slice->crs_dev = dev;
1359 slice->crs_ops = ops;
1360 slice->crs_req = req;
1363 EXPORT_SYMBOL(cl_req_slice_add);
1365 static void cl_req_free(const struct lu_env *env, struct cl_req *req)
1369 LASSERT(list_empty(&req->crq_pages));
1370 LASSERT(req->crq_nrpages == 0);
1371 LINVRNT(list_empty(&req->crq_layers));
1372 LINVRNT(equi(req->crq_nrobjs > 0, req->crq_o != NULL));
1375 if (req->crq_o != NULL) {
1376 for (i = 0; i < req->crq_nrobjs; ++i) {
1377 struct cl_object *obj = req->crq_o[i].ro_obj;
1379 lu_object_ref_del_at(&obj->co_lu,
1380 req->crq_o[i].ro_obj_ref,
1382 cl_object_put(env, obj);
1385 OBD_FREE(req->crq_o, req->crq_nrobjs * sizeof req->crq_o[0]);
1391 static int cl_req_init(const struct lu_env *env, struct cl_req *req,
1392 struct cl_page *page)
1394 struct cl_device *dev;
1395 struct cl_page_slice *slice;
1400 page = cl_page_top(page);
1402 list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
1403 dev = lu2cl_dev(slice->cpl_obj->co_lu.lo_dev);
1404 if (dev->cd_ops->cdo_req_init != NULL) {
1405 result = dev->cd_ops->cdo_req_init(env,
1411 page = page->cp_child;
1412 } while (page != NULL && result == 0);
1417 * Invokes per-request transfer completion call-backs
1418 * (cl_req_operations::cro_completion()) bottom-to-top.
1420 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int rc)
1422 struct cl_req_slice *slice;
1426 * for the lack of list_for_each_entry_reverse_safe()...
1428 while (!list_empty(&req->crq_layers)) {
1429 slice = list_entry(req->crq_layers.prev,
1430 struct cl_req_slice, crs_linkage);
1431 list_del_init(&slice->crs_linkage);
1432 if (slice->crs_ops->cro_completion != NULL)
1433 slice->crs_ops->cro_completion(env, slice, rc);
1435 cl_req_free(env, req);
1438 EXPORT_SYMBOL(cl_req_completion);
1441 * Allocates new transfer request.
1443 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
1444 enum cl_req_type crt, int nr_objects)
1448 LINVRNT(nr_objects > 0);
1455 OBD_ALLOC(req->crq_o, nr_objects * sizeof req->crq_o[0]);
1456 if (req->crq_o != NULL) {
1457 req->crq_nrobjs = nr_objects;
1458 req->crq_type = crt;
1459 CFS_INIT_LIST_HEAD(&req->crq_pages);
1460 CFS_INIT_LIST_HEAD(&req->crq_layers);
1461 result = cl_req_init(env, req, page);
1465 cl_req_completion(env, req, result);
1466 req = ERR_PTR(result);
1469 req = ERR_PTR(-ENOMEM);
1472 EXPORT_SYMBOL(cl_req_alloc);
1475 * Adds a page to a request.
1477 void cl_req_page_add(const struct lu_env *env,
1478 struct cl_req *req, struct cl_page *page)
1480 struct cl_object *obj;
1481 struct cl_req_obj *rqo;
1485 page = cl_page_top(page);
1487 LINVRNT(cl_page_is_vmlocked(env, page));
1488 LASSERT(list_empty(&page->cp_flight));
1489 LASSERT(page->cp_req == NULL);
1491 list_add_tail(&page->cp_flight, &req->crq_pages);
1494 obj = cl_object_top(page->cp_obj);
1495 for (i = 0, rqo = req->crq_o; obj != rqo->ro_obj; ++i, ++rqo) {
1496 if (rqo->ro_obj == NULL) {
1499 rqo->ro_obj_ref = lu_object_ref_add(&obj->co_lu,
1504 LASSERT(i < req->crq_nrobjs);
1507 EXPORT_SYMBOL(cl_req_page_add);
1510 * Removes a page from a request.
1512 void cl_req_page_done(const struct lu_env *env, struct cl_page *page)
1514 struct cl_req *req = page->cp_req;
1517 page = cl_page_top(page);
1519 LINVRNT(cl_page_is_vmlocked(env, page));
1520 LASSERT(!list_empty(&page->cp_flight));
1521 LASSERT(req->crq_nrpages > 0);
1523 list_del_init(&page->cp_flight);
1525 page->cp_req = NULL;
1528 EXPORT_SYMBOL(cl_req_page_done);
1531 * Notifies layers that request is about to depart by calling
1532 * cl_req_operations::cro_prep() top-to-bottom.
1534 int cl_req_prep(const struct lu_env *env, struct cl_req *req)
1538 const struct cl_req_slice *slice;
1542 * Check that the caller of cl_req_alloc() didn't lie about the number
1545 for (i = 0; i < req->crq_nrobjs; ++i)
1546 LASSERT(req->crq_o[i].ro_obj != NULL);
1549 list_for_each_entry(slice, &req->crq_layers, crs_linkage) {
1550 if (slice->crs_ops->cro_prep != NULL) {
1551 result = slice->crs_ops->cro_prep(env, slice);
1558 EXPORT_SYMBOL(cl_req_prep);
1561 * Fills in attributes that are passed to server together with transfer. Only
1562 * attributes from \a flags may be touched. This can be called multiple times
1563 * for the same request.
1565 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
1566 struct cl_req_attr *attr, obd_valid flags)
1568 const struct cl_req_slice *slice;
1569 struct cl_page *page;
1572 LASSERT(!list_empty(&req->crq_pages));
1575 /* Take any page to use as a model. */
1576 page = list_entry(req->crq_pages.next, struct cl_page, cp_flight);
1578 for (i = 0; i < req->crq_nrobjs; ++i) {
1579 list_for_each_entry(slice, &req->crq_layers, crs_linkage) {
1580 const struct cl_page_slice *scan;
1581 const struct cl_object *obj;
1583 scan = cl_page_at(page,
1584 slice->crs_dev->cd_lu_dev.ld_type);
1585 LASSERT(scan != NULL);
1586 obj = scan->cpl_obj;
1587 if (slice->crs_ops->cro_attr_set != NULL)
1588 slice->crs_ops->cro_attr_set(env, slice, obj,
1594 EXPORT_SYMBOL(cl_req_attr_set);
1596 /* XXX complete(), init_completion(), and wait_for_completion(), until they are
1597 * implemented in libcfs. */
1599 # include <linux/sched.h>
1600 #else /* __KERNEL__ */
1601 # include <liblustre.h>
1605 * Initialize synchronous io wait anchor, for transfer of \a nrpages pages.
1607 void cl_sync_io_init(struct cl_sync_io *anchor, int nrpages)
1610 cfs_waitq_init(&anchor->csi_waitq);
1611 atomic_set(&anchor->csi_sync_nr, nrpages);
1612 anchor->csi_sync_rc = 0;
1615 EXPORT_SYMBOL(cl_sync_io_init);
1618 * Wait until all transfer completes. Transfer completion routine has to call
1619 * cl_sync_io_note() for every page.
1621 int cl_sync_io_wait(const struct lu_env *env, struct cl_io *io,
1622 struct cl_page_list *queue, struct cl_sync_io *anchor,
1625 struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
1630 LASSERT(timeout >= 0);
1632 rc = l_wait_event(anchor->csi_waitq,
1633 atomic_read(&anchor->csi_sync_nr) == 0,
1636 CERROR("SYNC IO failed with error: %d, try to cancel "
1637 "%d remaining pages\n",
1638 rc, atomic_read(&anchor->csi_sync_nr));
1640 (void)cl_io_cancel(env, io, queue);
1642 lwi = (struct l_wait_info) { 0 };
1643 (void)l_wait_event(anchor->csi_waitq,
1644 atomic_read(&anchor->csi_sync_nr) == 0,
1647 rc = anchor->csi_sync_rc;
1649 LASSERT(atomic_read(&anchor->csi_sync_nr) == 0);
1650 cl_page_list_assume(env, io, queue);
1651 POISON(anchor, 0x5a, sizeof *anchor);
1654 EXPORT_SYMBOL(cl_sync_io_wait);
1657 * Indicate that transfer of a single page completed.
1659 void cl_sync_io_note(struct cl_sync_io *anchor, int ioret)
1662 if (anchor->csi_sync_rc == 0 && ioret < 0)
1663 anchor->csi_sync_rc = ioret;
1665 * Synchronous IO done without releasing page lock (e.g., as a part of
1666 * ->{prepare,commit}_write(). Completion is used to signal the end of
1669 LASSERT(atomic_read(&anchor->csi_sync_nr) > 0);
1670 if (atomic_dec_and_test(&anchor->csi_sync_nr))
1671 cfs_waitq_broadcast(&anchor->csi_waitq);
1674 EXPORT_SYMBOL(cl_sync_io_note);