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 * True, iff \a io is a sendfile().
80 int cl_io_is_sendfile(const struct cl_io *io)
82 return io->ci_type == CIT_READ && io->u.ci_rd.rd_is_sendfile;
84 EXPORT_SYMBOL(cl_io_is_sendfile);
87 * Returns true iff there is an IO ongoing in the given environment.
89 int cl_io_is_going(const struct lu_env *env)
91 return cl_env_info(env)->clt_current_io != NULL;
93 EXPORT_SYMBOL(cl_io_is_going);
96 * cl_io invariant that holds at all times when exported cl_io_*() functions
97 * are entered and left.
99 static int cl_io_invariant(const struct cl_io *io)
106 * io can own pages only when it is ongoing. Sub-io might
107 * still be in CIS_LOCKED state when top-io is in
110 ergo(io->ci_owned_nr > 0, io->ci_state == CIS_IO_GOING ||
111 (io->ci_state == CIS_LOCKED && up != NULL));
115 * Finalize \a io, by calling cl_io_operations::cio_fini() bottom-to-top.
117 void cl_io_fini(const struct lu_env *env, struct cl_io *io)
119 struct cl_io_slice *slice;
120 struct cl_thread_info *info;
122 LINVRNT(cl_io_type_is_valid(io->ci_type));
123 LINVRNT(cl_io_invariant(io));
126 while (!list_empty(&io->ci_layers)) {
127 slice = container_of(io->ci_layers.next, struct cl_io_slice,
129 list_del_init(&slice->cis_linkage);
130 if (slice->cis_iop->op[io->ci_type].cio_fini != NULL)
131 slice->cis_iop->op[io->ci_type].cio_fini(env, slice);
133 * Invalidate slice to catch use after free. This assumes that
134 * slices are allocated within session and can be touched
135 * after ->cio_fini() returns.
137 slice->cis_io = NULL;
139 io->ci_state = CIS_FINI;
140 info = cl_env_info(env);
141 if (info->clt_current_io == io)
142 info->clt_current_io = NULL;
145 EXPORT_SYMBOL(cl_io_fini);
147 static int cl_io_init0(const struct lu_env *env, struct cl_io *io,
148 enum cl_io_type iot, struct cl_object *obj)
150 struct cl_object *scan;
153 LINVRNT(io->ci_state == CIS_ZERO || io->ci_state == CIS_FINI);
154 LINVRNT(cl_io_type_is_valid(iot));
155 LINVRNT(cl_io_invariant(io));
159 CFS_INIT_LIST_HEAD(&io->ci_lockset.cls_todo);
160 CFS_INIT_LIST_HEAD(&io->ci_lockset.cls_curr);
161 CFS_INIT_LIST_HEAD(&io->ci_lockset.cls_done);
162 CFS_INIT_LIST_HEAD(&io->ci_layers);
165 cl_object_for_each(scan, obj) {
166 if (scan->co_ops->coo_io_init != NULL) {
167 result = scan->co_ops->coo_io_init(env, scan, io);
173 io->ci_state = CIS_INIT;
178 * Initialize sub-io, by calling cl_io_operations::cio_init() top-to-bottom.
180 * \pre obj != cl_object_top(obj)
182 int cl_io_sub_init(const struct lu_env *env, struct cl_io *io,
183 enum cl_io_type iot, struct cl_object *obj)
185 struct cl_thread_info *info = cl_env_info(env);
187 LASSERT(obj != cl_object_top(obj));
188 if (info->clt_current_io == NULL)
189 info->clt_current_io = io;
190 return cl_io_init0(env, io, iot, obj);
192 EXPORT_SYMBOL(cl_io_sub_init);
195 * Initialize \a io, by calling cl_io_operations::cio_init() top-to-bottom.
197 * Caller has to call cl_io_fini() after a call to cl_io_init(), no matter
198 * what the latter returned.
200 * \pre obj == cl_object_top(obj)
201 * \pre cl_io_type_is_valid(iot)
202 * \post cl_io_type_is_valid(io->ci_type) && io->ci_type == iot
204 int cl_io_init(const struct lu_env *env, struct cl_io *io,
205 enum cl_io_type iot, struct cl_object *obj)
207 struct cl_thread_info *info = cl_env_info(env);
209 LASSERT(obj == cl_object_top(obj));
210 LASSERT(info->clt_current_io == NULL);
212 info->clt_current_io = io;
213 return cl_io_init0(env, io, iot, obj);
215 EXPORT_SYMBOL(cl_io_init);
218 * Initialize read or write io.
220 * \pre iot == CIT_READ || iot == CIT_WRITE
222 int cl_io_rw_init(const struct lu_env *env, struct cl_io *io,
223 enum cl_io_type iot, loff_t pos, size_t count)
225 LINVRNT(iot == CIT_READ || iot == CIT_WRITE);
226 LINVRNT(io->ci_obj != NULL);
229 LU_OBJECT_HEADER(D_VFSTRACE, env, &io->ci_obj->co_lu,
230 "io range: %i [%llu, %llu) %i %i\n",
231 iot, (__u64)pos, (__u64)(pos + count),
232 io->u.ci_rw.crw_nonblock, io->u.ci_wr.wr_append);
233 io->u.ci_rw.crw_pos = pos;
234 io->u.ci_rw.crw_count = count;
235 RETURN(cl_io_init(env, io, iot, io->ci_obj));
237 EXPORT_SYMBOL(cl_io_rw_init);
239 static inline const struct lu_fid *
240 cl_lock_descr_fid(const struct cl_lock_descr *descr)
242 return lu_object_fid(&descr->cld_obj->co_lu);
245 static int cl_lock_descr_cmp(const struct cl_lock_descr *d0,
246 const struct cl_lock_descr *d1)
248 return lu_fid_cmp(cl_lock_descr_fid(d0), cl_lock_descr_fid(d1)) ?:
249 __diff_normalize(d0->cld_start, d1->cld_start);
253 * Sort locks in lexicographical order of their (fid, start-offset) pairs.
255 static void cl_io_locks_sort(struct cl_io *io)
260 /* hidden treasure: bubble sort for now. */
262 struct cl_io_lock_link *curr;
263 struct cl_io_lock_link *prev;
264 struct cl_io_lock_link *temp;
269 list_for_each_entry_safe(curr, temp, &io->ci_lockset.cls_todo,
272 switch (cl_lock_descr_cmp(&prev->cill_descr,
273 &curr->cill_descr)) {
276 * IMPOSSIBLE: Identical locks are
283 list_move_tail(&curr->cill_linkage,
284 &prev->cill_linkage);
286 continue; /* don't change prev: it's
287 * still "previous" */
288 case -1: /* already in order */
299 * Check whether \a queue contains locks matching \a need.
301 * \retval +ve there is a matching lock in the \a queue
302 * \retval 0 there are no matching locks in the \a queue
304 int cl_queue_match(const struct list_head *queue,
305 const struct cl_lock_descr *need)
307 struct cl_io_lock_link *scan;
310 list_for_each_entry(scan, queue, cill_linkage) {
311 if (cl_lock_descr_match(&scan->cill_descr, need))
316 EXPORT_SYMBOL(cl_queue_match);
318 static int cl_lockset_match(const struct cl_lockset *set,
319 const struct cl_lock_descr *need, int all_queues)
321 return (all_queues ? cl_queue_match(&set->cls_todo, need) : 0) ||
322 cl_queue_match(&set->cls_curr, need) ||
323 cl_queue_match(&set->cls_done, need);
326 static int cl_lockset_lock_one(const struct lu_env *env,
327 struct cl_io *io, struct cl_lockset *set,
328 struct cl_io_lock_link *link)
330 struct cl_lock *lock;
335 lock = cl_lock_request(env, io, &link->cill_descr, link->cill_enq_flags,
338 link->cill_lock = lock;
339 list_move(&link->cill_linkage, &set->cls_curr);
340 if (!(link->cill_enq_flags & CEF_ASYNC)) {
341 result = cl_wait(env, lock);
343 list_move(&link->cill_linkage, &set->cls_done);
347 result = PTR_ERR(lock);
351 static void cl_lock_link_fini(const struct lu_env *env, struct cl_io *io,
352 struct cl_io_lock_link *link)
354 struct cl_lock *lock = link->cill_lock;
357 list_del_init(&link->cill_linkage);
359 cl_lock_release(env, lock, "io", io);
360 link->cill_lock = NULL;
362 if (link->cill_fini != NULL)
363 link->cill_fini(env, link);
367 static int cl_lockset_lock(const struct lu_env *env, struct cl_io *io,
368 struct cl_lockset *set)
370 struct cl_io_lock_link *link;
371 struct cl_io_lock_link *temp;
372 struct cl_lock *lock;
377 list_for_each_entry_safe(link, temp, &set->cls_todo, cill_linkage) {
378 if (!cl_lockset_match(set, &link->cill_descr, 0)) {
379 /* XXX some locking to guarantee that locks aren't
380 * expanded in between. */
381 result = cl_lockset_lock_one(env, io, set, link);
385 cl_lock_link_fini(env, io, link);
388 list_for_each_entry_safe(link, temp,
389 &set->cls_curr, cill_linkage) {
390 lock = link->cill_lock;
391 result = cl_wait(env, lock);
393 list_move(&link->cill_linkage, &set->cls_done);
402 * Takes locks necessary for the current iteration of io.
404 * Calls cl_io_operations::cio_lock() top-to-bottom to collect locks required
405 * by layers for the current iteration. Then sort locks (to avoid dead-locks),
408 int cl_io_lock(const struct lu_env *env, struct cl_io *io)
410 const struct cl_io_slice *scan;
413 LINVRNT(cl_io_is_loopable(io));
414 LINVRNT(io->ci_state == CIS_IT_STARTED);
415 LINVRNT(cl_io_invariant(io));
418 cl_io_for_each(scan, io) {
419 if (scan->cis_iop->op[io->ci_type].cio_lock == NULL)
421 result = scan->cis_iop->op[io->ci_type].cio_lock(env, scan);
426 cl_io_locks_sort(io);
427 result = cl_lockset_lock(env, io, &io->ci_lockset);
430 cl_io_unlock(env, io);
432 io->ci_state = CIS_LOCKED;
435 EXPORT_SYMBOL(cl_io_lock);
438 * Release locks takes by io.
440 void cl_io_unlock(const struct lu_env *env, struct cl_io *io)
442 struct cl_lockset *set;
443 struct cl_io_lock_link *link;
444 struct cl_io_lock_link *temp;
445 const struct cl_io_slice *scan;
447 LASSERT(cl_io_is_loopable(io));
448 LASSERT(CIS_IT_STARTED <= io->ci_state && io->ci_state < CIS_UNLOCKED);
449 LINVRNT(cl_io_invariant(io));
452 set = &io->ci_lockset;
454 list_for_each_entry_safe(link, temp, &set->cls_todo, cill_linkage)
455 cl_lock_link_fini(env, io, link);
457 list_for_each_entry_safe(link, temp, &set->cls_curr, cill_linkage)
458 cl_lock_link_fini(env, io, link);
460 list_for_each_entry_safe(link, temp, &set->cls_done, cill_linkage) {
461 cl_unuse(env, link->cill_lock);
462 cl_lock_link_fini(env, io, link);
464 cl_io_for_each_reverse(scan, io) {
465 if (scan->cis_iop->op[io->ci_type].cio_unlock != NULL)
466 scan->cis_iop->op[io->ci_type].cio_unlock(env, scan);
468 io->ci_state = CIS_UNLOCKED;
469 LASSERT(!cl_env_info(env)->clt_counters[CNL_TOP].ctc_nr_locks_acquired);
472 EXPORT_SYMBOL(cl_io_unlock);
475 * Prepares next iteration of io.
477 * Calls cl_io_operations::cio_iter_init() top-to-bottom. This exists to give
478 * layers a chance to modify io parameters, e.g., so that lov can restrict io
479 * to a single stripe.
481 int cl_io_iter_init(const struct lu_env *env, struct cl_io *io)
483 const struct cl_io_slice *scan;
486 LINVRNT(cl_io_is_loopable(io));
487 LINVRNT(io->ci_state == CIS_INIT || io->ci_state == CIS_IT_ENDED);
488 LINVRNT(cl_io_invariant(io));
492 cl_io_for_each(scan, io) {
493 if (scan->cis_iop->op[io->ci_type].cio_iter_init == NULL)
495 result = scan->cis_iop->op[io->ci_type].cio_iter_init(env,
501 io->ci_state = CIS_IT_STARTED;
504 EXPORT_SYMBOL(cl_io_iter_init);
507 * Finalizes io iteration.
509 * Calls cl_io_operations::cio_iter_fini() bottom-to-top.
511 void cl_io_iter_fini(const struct lu_env *env, struct cl_io *io)
513 const struct cl_io_slice *scan;
515 LINVRNT(cl_io_is_loopable(io));
516 LINVRNT(io->ci_state == CIS_UNLOCKED);
517 LINVRNT(cl_io_invariant(io));
520 cl_io_for_each_reverse(scan, io) {
521 if (scan->cis_iop->op[io->ci_type].cio_iter_fini != NULL)
522 scan->cis_iop->op[io->ci_type].cio_iter_fini(env, scan);
524 io->ci_state = CIS_IT_ENDED;
527 EXPORT_SYMBOL(cl_io_iter_fini);
530 * Records that read or write io progressed \a nob bytes forward.
532 void cl_io_rw_advance(const struct lu_env *env, struct cl_io *io, size_t nob)
534 const struct cl_io_slice *scan;
536 LINVRNT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
538 LINVRNT(cl_io_is_loopable(io));
539 LINVRNT(cl_io_invariant(io));
543 io->u.ci_rw.crw_pos += nob;
544 io->u.ci_rw.crw_count -= nob;
546 /* layers have to be notified. */
547 cl_io_for_each_reverse(scan, io) {
548 if (scan->cis_iop->op[io->ci_type].cio_advance != NULL)
549 scan->cis_iop->op[io->ci_type].cio_advance(env, scan,
554 EXPORT_SYMBOL(cl_io_rw_advance);
557 * Adds a lock to a lockset.
559 int cl_io_lock_add(const struct lu_env *env, struct cl_io *io,
560 struct cl_io_lock_link *link)
565 if (cl_lockset_match(&io->ci_lockset, &link->cill_descr, 1))
568 list_add(&link->cill_linkage, &io->ci_lockset.cls_todo);
573 EXPORT_SYMBOL(cl_io_lock_add);
575 static void cl_free_io_lock_link(const struct lu_env *env,
576 struct cl_io_lock_link *link)
582 * Allocates new lock link, and uses it to add a lock to a lockset.
584 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
585 struct cl_lock_descr *descr, int enqflags)
587 struct cl_io_lock_link *link;
593 link->cill_descr = *descr;
594 link->cill_enq_flags = enqflags;
595 link->cill_fini = cl_free_io_lock_link;
596 result = cl_io_lock_add(env, io, link);
597 if (result) /* lock match */
598 link->cill_fini(env, link);
604 EXPORT_SYMBOL(cl_io_lock_alloc_add);
607 * Starts io by calling cl_io_operations::cio_start() top-to-bottom.
609 int cl_io_start(const struct lu_env *env, struct cl_io *io)
611 const struct cl_io_slice *scan;
614 LINVRNT(cl_io_is_loopable(io));
615 LINVRNT(io->ci_state == CIS_LOCKED);
616 LINVRNT(cl_io_invariant(io));
619 io->ci_state = CIS_IO_GOING;
620 cl_io_for_each(scan, io) {
621 if (scan->cis_iop->op[io->ci_type].cio_start == NULL)
623 result = scan->cis_iop->op[io->ci_type].cio_start(env, scan);
631 EXPORT_SYMBOL(cl_io_start);
634 * Wait until current io iteration is finished by calling
635 * cl_io_operations::cio_end() bottom-to-top.
637 void cl_io_end(const struct lu_env *env, struct cl_io *io)
639 const struct cl_io_slice *scan;
641 LINVRNT(cl_io_is_loopable(io));
642 LINVRNT(io->ci_state == CIS_IO_GOING);
643 LINVRNT(cl_io_invariant(io));
646 cl_io_for_each_reverse(scan, io) {
647 if (scan->cis_iop->op[io->ci_type].cio_end != NULL)
648 scan->cis_iop->op[io->ci_type].cio_end(env, scan);
649 /* TODO: error handling. */
651 io->ci_state = CIS_IO_FINISHED;
654 EXPORT_SYMBOL(cl_io_end);
656 static const struct cl_page_slice *
657 cl_io_slice_page(const struct cl_io_slice *ios, struct cl_page *page)
659 const struct cl_page_slice *slice;
661 slice = cl_page_at(page, ios->cis_obj->co_lu.lo_dev->ld_type);
662 LINVRNT(slice != NULL);
667 * True iff \a page is within \a io range.
669 static int cl_page_in_io(const struct cl_page *page, const struct cl_io *io)
676 idx = page->cp_index;
677 switch (io->ci_type) {
681 * check that [start, end) and [pos, pos + count) extents
684 start = cl_offset(page->cp_obj, idx);
685 end = cl_offset(page->cp_obj, idx + 1);
686 result = io->u.ci_rw.crw_pos < end &&
687 start < io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count;
690 result = io->u.ci_fault.ft_index == idx;
699 * Called by read io, when page has to be read from the server.
701 * \see cl_io_operations::cio_read_page()
703 int cl_io_read_page(const struct lu_env *env, struct cl_io *io,
704 struct cl_page *page)
706 const struct cl_io_slice *scan;
707 struct cl_2queue *queue;
710 LINVRNT(io->ci_type == CIT_READ || io->ci_type == CIT_FAULT);
711 LINVRNT(cl_page_is_owned(page, io));
712 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
713 LINVRNT(cl_page_in_io(page, io));
714 LINVRNT(cl_io_invariant(io));
717 queue = &io->ci_queue;
719 cl_2queue_init(queue);
721 * ->cio_read_page() methods called in the loop below are supposed to
722 * never block waiting for network (the only subtle point is the
723 * creation of new pages for read-ahead that might result in cache
724 * shrinking, but currently only clean pages are shrunk and this
725 * requires no network io).
727 * Should this ever starts blocking, retry loop would be needed for
728 * "parallel io" (see CLO_REPEAT loops in cl_lock.c).
730 cl_io_for_each(scan, io) {
731 if (scan->cis_iop->cio_read_page != NULL) {
732 const struct cl_page_slice *slice;
734 slice = cl_io_slice_page(scan, page);
735 LINVRNT(slice != NULL);
736 result = scan->cis_iop->cio_read_page(env, scan, slice);
742 result = cl_io_submit_rw(env, io, CRT_READ, queue, CRP_NORMAL);
744 * Unlock unsent pages in case of error.
746 cl_page_list_disown(env, io, &queue->c2_qin);
747 cl_2queue_fini(env, queue);
750 EXPORT_SYMBOL(cl_io_read_page);
753 * Called by write io to prepare page to receive data from user buffer.
755 * \see cl_io_operations::cio_prepare_write()
757 int cl_io_prepare_write(const struct lu_env *env, struct cl_io *io,
758 struct cl_page *page, unsigned from, unsigned to)
760 const struct cl_io_slice *scan;
763 LINVRNT(io->ci_type == CIT_WRITE);
764 LINVRNT(cl_page_is_owned(page, io));
765 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
766 LINVRNT(cl_io_invariant(io));
767 LASSERT(cl_page_in_io(page, io));
770 cl_io_for_each_reverse(scan, io) {
771 if (scan->cis_iop->cio_prepare_write != NULL) {
772 const struct cl_page_slice *slice;
774 slice = cl_io_slice_page(scan, page);
775 result = scan->cis_iop->cio_prepare_write(env, scan,
784 EXPORT_SYMBOL(cl_io_prepare_write);
787 * Called by write io after user data were copied into a page.
789 * \see cl_io_operations::cio_commit_write()
791 int cl_io_commit_write(const struct lu_env *env, struct cl_io *io,
792 struct cl_page *page, unsigned from, unsigned to)
794 const struct cl_io_slice *scan;
797 LINVRNT(io->ci_type == CIT_WRITE);
798 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
799 LINVRNT(cl_io_invariant(io));
801 * XXX Uh... not nice. Top level cl_io_commit_write() call (vvp->lov)
802 * already called cl_page_cache_add(), moving page into CPS_CACHED
803 * state. Better (and more general) way of dealing with such situation
806 LASSERT(cl_page_is_owned(page, io) || page->cp_parent != NULL);
807 LASSERT(cl_page_in_io(page, io));
810 cl_io_for_each(scan, io) {
811 if (scan->cis_iop->cio_commit_write != NULL) {
812 const struct cl_page_slice *slice;
814 slice = cl_io_slice_page(scan, page);
815 result = scan->cis_iop->cio_commit_write(env, scan,
822 LINVRNT(result <= 0);
825 EXPORT_SYMBOL(cl_io_commit_write);
828 * Submits a list of pages for immediate io.
830 * After the function gets returned, The submitted pages are moved to
831 * queue->c2_qout queue, and queue->c2_qin contain both the pages don't need
832 * to be submitted, and the pages are errant to submit.
834 * \returns 0 if at least one page was submitted, error code otherwise.
835 * \see cl_io_operations::cio_submit()
837 int cl_io_submit_rw(const struct lu_env *env, struct cl_io *io,
838 enum cl_req_type crt, struct cl_2queue *queue,
839 enum cl_req_priority priority)
841 const struct cl_io_slice *scan;
844 LINVRNT(crt < ARRAY_SIZE(scan->cis_iop->req_op));
847 cl_io_for_each(scan, io) {
848 if (scan->cis_iop->req_op[crt].cio_submit == NULL)
850 result = scan->cis_iop->req_op[crt].cio_submit(env, scan, crt,
856 * If ->cio_submit() failed, no pages were sent.
858 LASSERT(ergo(result != 0, list_empty(&queue->c2_qout.pl_pages)));
861 EXPORT_SYMBOL(cl_io_submit_rw);
864 * Submit a sync_io and wait for the IO to be finished, or error happens.
865 * If @timeout is zero, it means to wait for the IO unconditionally.
867 int cl_io_submit_sync(const struct lu_env *env, struct cl_io *io,
868 enum cl_req_type iot, struct cl_2queue *queue,
869 enum cl_req_priority prio, long timeout)
871 struct cl_sync_io *anchor = &cl_env_info(env)->clt_anchor;
875 LASSERT(prio == CRP_NORMAL || prio == CRP_CANCEL);
877 cl_page_list_for_each(pg, &queue->c2_qin) {
878 LASSERT(pg->cp_sync_io == NULL);
879 pg->cp_sync_io = anchor;
882 cl_sync_io_init(anchor, queue->c2_qin.pl_nr);
883 rc = cl_io_submit_rw(env, io, iot, queue, prio);
886 * If some pages weren't sent for any reason (e.g.,
887 * read found up-to-date pages in the cache, or write found
888 * clean pages), count them as completed to avoid infinite
891 cl_page_list_for_each(pg, &queue->c2_qin) {
892 pg->cp_sync_io = NULL;
893 cl_sync_io_note(anchor, +1);
896 /* wait for the IO to be finished. */
897 rc = cl_sync_io_wait(env, io, &queue->c2_qout,
900 LASSERT(list_empty(&queue->c2_qout.pl_pages));
901 cl_page_list_for_each(pg, &queue->c2_qin)
902 pg->cp_sync_io = NULL;
906 EXPORT_SYMBOL(cl_io_submit_sync);
909 * Cancel an IO which has been submitted by cl_io_submit_rw.
911 int cl_io_cancel(const struct lu_env *env, struct cl_io *io,
912 struct cl_page_list *queue)
914 struct cl_page *page;
917 CERROR("Canceling ongoing page trasmission\n");
918 cl_page_list_for_each(page, queue) {
921 LINVRNT(cl_page_in_io(page, io));
922 rc = cl_page_cancel(env, page);
923 result = result ?: rc;
927 EXPORT_SYMBOL(cl_io_cancel);
932 * Pumps io through iterations calling
934 * - cl_io_iter_init()
944 * - cl_io_iter_fini()
946 * repeatedly until there is no more io to do.
948 int cl_io_loop(const struct lu_env *env, struct cl_io *io)
952 LINVRNT(cl_io_is_loopable(io));
959 result = cl_io_iter_init(env, io);
962 result = cl_io_lock(env, io);
965 * Notify layers that locks has been taken,
968 * - llite: kms, short read;
969 * - llite: generic_file_read();
971 result = cl_io_start(env, io);
973 * Send any remaining pending
976 * - llite: ll_rw_stats_tally.
979 cl_io_unlock(env, io);
980 cl_io_rw_advance(env, io, io->ci_nob - nob);
983 cl_io_iter_fini(env, io);
984 } while (result == 0 && io->ci_continue);
985 RETURN(result < 0 ? result : 0);
987 EXPORT_SYMBOL(cl_io_loop);
990 * Adds io slice to the cl_io.
992 * This is called by cl_object_operations::coo_io_init() methods to add a
993 * per-layer state to the io. New state is added at the end of
994 * cl_io::ci_layers list, that is, it is at the bottom of the stack.
996 * \see cl_lock_slice_add(), cl_req_slice_add(), cl_page_slice_add()
998 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
999 struct cl_object *obj,
1000 const struct cl_io_operations *ops)
1002 struct list_head *linkage = &slice->cis_linkage;
1004 LASSERT((linkage->prev == NULL && linkage->next == NULL) ||
1005 list_empty(linkage));
1008 list_add_tail(linkage, &io->ci_layers);
1010 slice->cis_obj = obj;
1011 slice->cis_iop = ops;
1014 EXPORT_SYMBOL(cl_io_slice_add);
1018 * Initializes page list.
1020 void cl_page_list_init(struct cl_page_list *plist)
1024 CFS_INIT_LIST_HEAD(&plist->pl_pages);
1025 plist->pl_owner = cfs_current();
1028 EXPORT_SYMBOL(cl_page_list_init);
1031 * Adds a page to a page list.
1033 void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page)
1036 /* it would be better to check that page is owned by "current" io, but
1037 * it is not passed here. */
1038 LASSERT(page->cp_owner != NULL);
1039 LINVRNT(plist->pl_owner == cfs_current());
1042 mutex_lock(&page->cp_mutex);
1044 LASSERT(list_empty(&page->cp_batch));
1045 list_add_tail(&page->cp_batch, &plist->pl_pages);
1047 page->cp_queue_ref = lu_ref_add(&page->cp_reference, "queue", plist);
1051 EXPORT_SYMBOL(cl_page_list_add);
1054 * Removes a page from a page list.
1056 void cl_page_list_del(const struct lu_env *env,
1057 struct cl_page_list *plist, struct cl_page *page)
1059 LASSERT(plist->pl_nr > 0);
1060 LINVRNT(plist->pl_owner == cfs_current());
1063 list_del_init(&page->cp_batch);
1065 mutex_unlock(&page->cp_mutex);
1068 lu_ref_del_at(&page->cp_reference, page->cp_queue_ref, "queue", plist);
1069 cl_page_put(env, page);
1072 EXPORT_SYMBOL(cl_page_list_del);
1075 * Moves a page from one page list to another.
1077 void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src,
1078 struct cl_page *page)
1080 LASSERT(src->pl_nr > 0);
1081 LINVRNT(dst->pl_owner == cfs_current());
1082 LINVRNT(src->pl_owner == cfs_current());
1085 list_move_tail(&page->cp_batch, &dst->pl_pages);
1088 lu_ref_set_at(&page->cp_reference,
1089 page->cp_queue_ref, "queue", src, dst);
1092 EXPORT_SYMBOL(cl_page_list_move);
1095 * splice the cl_page_list, just as list head does
1097 void cl_page_list_splice(struct cl_page_list *list, struct cl_page_list *head)
1099 struct cl_page *page;
1100 struct cl_page *tmp;
1102 LINVRNT(list->pl_owner == cfs_current());
1103 LINVRNT(head->pl_owner == cfs_current());
1106 cl_page_list_for_each_safe(page, tmp, list)
1107 cl_page_list_move(head, list, page);
1110 EXPORT_SYMBOL(cl_page_list_splice);
1112 void cl_page_disown0(const struct lu_env *env,
1113 struct cl_io *io, struct cl_page *pg);
1116 * Disowns pages in a queue.
1118 void cl_page_list_disown(const struct lu_env *env,
1119 struct cl_io *io, struct cl_page_list *plist)
1121 struct cl_page *page;
1122 struct cl_page *temp;
1124 LINVRNT(plist->pl_owner == cfs_current());
1127 cl_page_list_for_each_safe(page, temp, plist) {
1128 LASSERT(plist->pl_nr > 0);
1130 list_del_init(&page->cp_batch);
1132 mutex_unlock(&page->cp_mutex);
1136 * cl_page_disown0 rather than usual cl_page_disown() is used,
1137 * because pages are possibly in CPS_FREEING state already due
1138 * to the call to cl_page_list_discard().
1141 * XXX cl_page_disown0() will fail if page is not locked.
1143 cl_page_disown0(env, io, page);
1144 lu_ref_del(&page->cp_reference, "queue", plist);
1145 cl_page_put(env, page);
1149 EXPORT_SYMBOL(cl_page_list_disown);
1152 * Releases pages from queue.
1154 void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist)
1156 struct cl_page *page;
1157 struct cl_page *temp;
1159 LINVRNT(plist->pl_owner == cfs_current());
1162 cl_page_list_for_each_safe(page, temp, plist)
1163 cl_page_list_del(env, plist, page);
1164 LASSERT(plist->pl_nr == 0);
1167 EXPORT_SYMBOL(cl_page_list_fini);
1170 * Owns all pages in a queue.
1172 int cl_page_list_own(const struct lu_env *env,
1173 struct cl_io *io, struct cl_page_list *plist)
1175 struct cl_page *page;
1176 struct cl_page *temp;
1180 LINVRNT(plist->pl_owner == cfs_current());
1184 cl_page_list_for_each_safe(page, temp, plist) {
1185 LASSERT(index <= page->cp_index);
1186 index = page->cp_index;
1187 if (cl_page_own(env, io, page) == 0)
1188 result = result ?: page->cp_error;
1190 cl_page_list_del(env, plist, page);
1194 EXPORT_SYMBOL(cl_page_list_own);
1197 * Assumes all pages in a queue.
1199 void cl_page_list_assume(const struct lu_env *env,
1200 struct cl_io *io, struct cl_page_list *plist)
1202 struct cl_page *page;
1204 LINVRNT(plist->pl_owner == cfs_current());
1206 cl_page_list_for_each(page, plist)
1207 cl_page_assume(env, io, page);
1209 EXPORT_SYMBOL(cl_page_list_assume);
1212 * Discards all pages in a queue.
1214 void cl_page_list_discard(const struct lu_env *env, struct cl_io *io,
1215 struct cl_page_list *plist)
1217 struct cl_page *page;
1219 LINVRNT(plist->pl_owner == cfs_current());
1221 cl_page_list_for_each(page, plist)
1222 cl_page_discard(env, io, page);
1225 EXPORT_SYMBOL(cl_page_list_discard);
1228 * Unmaps all pages in a queue from user virtual memory.
1230 int cl_page_list_unmap(const struct lu_env *env, struct cl_io *io,
1231 struct cl_page_list *plist)
1233 struct cl_page *page;
1236 LINVRNT(plist->pl_owner == cfs_current());
1239 cl_page_list_for_each(page, plist) {
1240 result = cl_page_unmap(env, io, page);
1246 EXPORT_SYMBOL(cl_page_list_unmap);
1249 * Initialize dual page queue.
1251 void cl_2queue_init(struct cl_2queue *queue)
1254 cl_page_list_init(&queue->c2_qin);
1255 cl_page_list_init(&queue->c2_qout);
1258 EXPORT_SYMBOL(cl_2queue_init);
1261 * Add a page to the incoming page list of 2-queue.
1263 void cl_2queue_add(struct cl_2queue *queue, struct cl_page *page)
1266 cl_page_list_add(&queue->c2_qin, page);
1269 EXPORT_SYMBOL(cl_2queue_add);
1272 * Disown pages in both lists of a 2-queue.
1274 void cl_2queue_disown(const struct lu_env *env,
1275 struct cl_io *io, struct cl_2queue *queue)
1278 cl_page_list_disown(env, io, &queue->c2_qin);
1279 cl_page_list_disown(env, io, &queue->c2_qout);
1282 EXPORT_SYMBOL(cl_2queue_disown);
1285 * Discard (truncate) pages in both lists of a 2-queue.
1287 void cl_2queue_discard(const struct lu_env *env,
1288 struct cl_io *io, struct cl_2queue *queue)
1291 cl_page_list_discard(env, io, &queue->c2_qin);
1292 cl_page_list_discard(env, io, &queue->c2_qout);
1295 EXPORT_SYMBOL(cl_2queue_discard);
1298 * Assume to own the pages in cl_2queue
1300 void cl_2queue_assume(const struct lu_env *env,
1301 struct cl_io *io, struct cl_2queue *queue)
1303 cl_page_list_assume(env, io, &queue->c2_qin);
1304 cl_page_list_assume(env, io, &queue->c2_qout);
1306 EXPORT_SYMBOL(cl_2queue_assume);
1309 * Finalize both page lists of a 2-queue.
1311 void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue)
1314 cl_page_list_fini(env, &queue->c2_qout);
1315 cl_page_list_fini(env, &queue->c2_qin);
1318 EXPORT_SYMBOL(cl_2queue_fini);
1321 * Initialize a 2-queue to contain \a page in its incoming page list.
1323 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page)
1326 cl_2queue_init(queue);
1327 cl_2queue_add(queue, page);
1330 EXPORT_SYMBOL(cl_2queue_init_page);
1333 * Returns top-level io.
1335 * \see cl_object_top(), cl_page_top().
1337 struct cl_io *cl_io_top(struct cl_io *io)
1340 while (io->ci_parent != NULL)
1344 EXPORT_SYMBOL(cl_io_top);
1347 * Prints human readable representation of \a io to the \a f.
1349 void cl_io_print(const struct lu_env *env, void *cookie,
1350 lu_printer_t printer, const struct cl_io *io)
1355 * Adds request slice to the compound request.
1357 * This is called by cl_device_operations::cdo_req_init() methods to add a
1358 * per-layer state to the request. New state is added at the end of
1359 * cl_req::crq_layers list, that is, it is at the bottom of the stack.
1361 * \see cl_lock_slice_add(), cl_page_slice_add(), cl_io_slice_add()
1363 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
1364 struct cl_device *dev,
1365 const struct cl_req_operations *ops)
1368 list_add_tail(&slice->crs_linkage, &req->crq_layers);
1369 slice->crs_dev = dev;
1370 slice->crs_ops = ops;
1371 slice->crs_req = req;
1374 EXPORT_SYMBOL(cl_req_slice_add);
1376 static void cl_req_free(const struct lu_env *env, struct cl_req *req)
1380 LASSERT(list_empty(&req->crq_pages));
1381 LASSERT(req->crq_nrpages == 0);
1382 LINVRNT(list_empty(&req->crq_layers));
1383 LINVRNT(equi(req->crq_nrobjs > 0, req->crq_o != NULL));
1386 if (req->crq_o != NULL) {
1387 for (i = 0; i < req->crq_nrobjs; ++i) {
1388 struct cl_object *obj = req->crq_o[i].ro_obj;
1390 lu_object_ref_del_at(&obj->co_lu,
1391 req->crq_o[i].ro_obj_ref,
1393 cl_object_put(env, obj);
1396 OBD_FREE(req->crq_o, req->crq_nrobjs * sizeof req->crq_o[0]);
1402 static int cl_req_init(const struct lu_env *env, struct cl_req *req,
1403 struct cl_page *page)
1405 struct cl_device *dev;
1406 struct cl_page_slice *slice;
1411 page = cl_page_top(page);
1413 list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
1414 dev = lu2cl_dev(slice->cpl_obj->co_lu.lo_dev);
1415 if (dev->cd_ops->cdo_req_init != NULL) {
1416 result = dev->cd_ops->cdo_req_init(env,
1422 page = page->cp_child;
1423 } while (page != NULL && result == 0);
1428 * Invokes per-request transfer completion call-backs
1429 * (cl_req_operations::cro_completion()) bottom-to-top.
1431 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int rc)
1433 struct cl_req_slice *slice;
1437 * for the lack of list_for_each_entry_reverse_safe()...
1439 while (!list_empty(&req->crq_layers)) {
1440 slice = list_entry(req->crq_layers.prev,
1441 struct cl_req_slice, crs_linkage);
1442 list_del_init(&slice->crs_linkage);
1443 if (slice->crs_ops->cro_completion != NULL)
1444 slice->crs_ops->cro_completion(env, slice, rc);
1446 cl_req_free(env, req);
1449 EXPORT_SYMBOL(cl_req_completion);
1452 * Allocates new transfer request.
1454 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
1455 enum cl_req_type crt, int nr_objects)
1459 LINVRNT(nr_objects > 0);
1466 OBD_ALLOC(req->crq_o, nr_objects * sizeof req->crq_o[0]);
1467 if (req->crq_o != NULL) {
1468 req->crq_nrobjs = nr_objects;
1469 req->crq_type = crt;
1470 CFS_INIT_LIST_HEAD(&req->crq_pages);
1471 CFS_INIT_LIST_HEAD(&req->crq_layers);
1472 result = cl_req_init(env, req, page);
1476 cl_req_completion(env, req, result);
1477 req = ERR_PTR(result);
1480 req = ERR_PTR(-ENOMEM);
1483 EXPORT_SYMBOL(cl_req_alloc);
1486 * Adds a page to a request.
1488 void cl_req_page_add(const struct lu_env *env,
1489 struct cl_req *req, struct cl_page *page)
1491 struct cl_object *obj;
1492 struct cl_req_obj *rqo;
1496 page = cl_page_top(page);
1498 LINVRNT(cl_page_is_vmlocked(env, page));
1499 LASSERT(list_empty(&page->cp_flight));
1500 LASSERT(page->cp_req == NULL);
1502 list_add_tail(&page->cp_flight, &req->crq_pages);
1505 obj = cl_object_top(page->cp_obj);
1506 for (i = 0, rqo = req->crq_o; obj != rqo->ro_obj; ++i, ++rqo) {
1507 if (rqo->ro_obj == NULL) {
1510 rqo->ro_obj_ref = lu_object_ref_add(&obj->co_lu,
1515 LASSERT(i < req->crq_nrobjs);
1518 EXPORT_SYMBOL(cl_req_page_add);
1521 * Removes a page from a request.
1523 void cl_req_page_done(const struct lu_env *env, struct cl_page *page)
1525 struct cl_req *req = page->cp_req;
1528 page = cl_page_top(page);
1530 LINVRNT(cl_page_is_vmlocked(env, page));
1531 LASSERT(!list_empty(&page->cp_flight));
1532 LASSERT(req->crq_nrpages > 0);
1534 list_del_init(&page->cp_flight);
1536 page->cp_req = NULL;
1539 EXPORT_SYMBOL(cl_req_page_done);
1542 * Notifies layers that request is about to depart by calling
1543 * cl_req_operations::cro_prep() top-to-bottom.
1545 int cl_req_prep(const struct lu_env *env, struct cl_req *req)
1549 const struct cl_req_slice *slice;
1553 * Check that the caller of cl_req_alloc() didn't lie about the number
1556 for (i = 0; i < req->crq_nrobjs; ++i)
1557 LASSERT(req->crq_o[i].ro_obj != NULL);
1560 list_for_each_entry(slice, &req->crq_layers, crs_linkage) {
1561 if (slice->crs_ops->cro_prep != NULL) {
1562 result = slice->crs_ops->cro_prep(env, slice);
1569 EXPORT_SYMBOL(cl_req_prep);
1572 * Fills in attributes that are passed to server together with transfer. Only
1573 * attributes from \a flags may be touched. This can be called multiple times
1574 * for the same request.
1576 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
1577 struct cl_req_attr *attr, obd_valid flags)
1579 const struct cl_req_slice *slice;
1580 struct cl_page *page;
1583 LASSERT(!list_empty(&req->crq_pages));
1586 /* Take any page to use as a model. */
1587 page = list_entry(req->crq_pages.next, struct cl_page, cp_flight);
1589 for (i = 0; i < req->crq_nrobjs; ++i) {
1590 list_for_each_entry(slice, &req->crq_layers, crs_linkage) {
1591 const struct cl_page_slice *scan;
1592 const struct cl_object *obj;
1594 scan = cl_page_at(page,
1595 slice->crs_dev->cd_lu_dev.ld_type);
1596 LASSERT(scan != NULL);
1597 obj = scan->cpl_obj;
1598 if (slice->crs_ops->cro_attr_set != NULL)
1599 slice->crs_ops->cro_attr_set(env, slice, obj,
1605 EXPORT_SYMBOL(cl_req_attr_set);
1607 /* XXX complete(), init_completion(), and wait_for_completion(), until they are
1608 * implemented in libcfs. */
1610 # include <linux/sched.h>
1611 #else /* __KERNEL__ */
1612 # include <liblustre.h>
1616 * Initialize synchronous io wait anchor, for transfer of \a nrpages pages.
1618 void cl_sync_io_init(struct cl_sync_io *anchor, int nrpages)
1621 cfs_waitq_init(&anchor->csi_waitq);
1622 atomic_set(&anchor->csi_sync_nr, nrpages);
1623 anchor->csi_sync_rc = 0;
1626 EXPORT_SYMBOL(cl_sync_io_init);
1629 * Wait until all transfer completes. Transfer completion routine has to call
1630 * cl_sync_io_note() for every page.
1632 int cl_sync_io_wait(const struct lu_env *env, struct cl_io *io,
1633 struct cl_page_list *queue, struct cl_sync_io *anchor,
1636 struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
1641 LASSERT(timeout >= 0);
1643 rc = l_wait_event(anchor->csi_waitq,
1644 atomic_read(&anchor->csi_sync_nr) == 0,
1649 CERROR("SYNC IO failed with error: %d, try to cancel "
1650 "the remaining page\n", rc);
1652 rc2 = cl_io_cancel(env, io, queue);
1654 lwi = (struct l_wait_info) { 0 };
1655 /* Too bad, some pages are still in IO. */
1656 CERROR("Failed to cancel transfer error: %d, mostly "
1657 "because of they are still being transferred, "
1658 "waiting for %i pages\n",
1659 rc2, atomic_read(&anchor->csi_sync_nr));
1660 (void)l_wait_event(anchor->csi_waitq,
1661 atomic_read(&anchor->csi_sync_nr) == 0,
1665 rc = anchor->csi_sync_rc;
1667 LASSERT(atomic_read(&anchor->csi_sync_nr) == 0);
1668 cl_page_list_assume(env, io, queue);
1669 POISON(anchor, 0x5a, sizeof *anchor);
1672 EXPORT_SYMBOL(cl_sync_io_wait);
1675 * Indicate that transfer of a single page completed.
1677 void cl_sync_io_note(struct cl_sync_io *anchor, int ioret)
1680 if (anchor->csi_sync_rc == 0 && ioret < 0)
1681 anchor->csi_sync_rc = ioret;
1683 * Synchronous IO done without releasing page lock (e.g., as a part of
1684 * ->{prepare,commit}_write(). Completion is used to signal the end of
1687 if (atomic_dec_and_test(&anchor->csi_sync_nr))
1688 cfs_waitq_broadcast(&anchor->csi_waitq);
1691 EXPORT_SYMBOL(cl_sync_io_note);