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, link->cill_enq_flags,
329 link->cill_lock = lock;
330 list_move(&link->cill_linkage, &set->cls_curr);
331 if (!(link->cill_enq_flags & CEF_ASYNC)) {
332 result = cl_wait(env, lock);
334 list_move(&link->cill_linkage, &set->cls_done);
338 result = PTR_ERR(lock);
342 static void cl_lock_link_fini(const struct lu_env *env, struct cl_io *io,
343 struct cl_io_lock_link *link)
345 struct cl_lock *lock = link->cill_lock;
348 list_del_init(&link->cill_linkage);
350 cl_lock_release(env, lock, "io", io);
351 link->cill_lock = NULL;
353 if (link->cill_fini != NULL)
354 link->cill_fini(env, link);
358 static int cl_lockset_lock(const struct lu_env *env, struct cl_io *io,
359 struct cl_lockset *set)
361 struct cl_io_lock_link *link;
362 struct cl_io_lock_link *temp;
363 struct cl_lock *lock;
368 list_for_each_entry_safe(link, temp, &set->cls_todo, cill_linkage) {
369 if (!cl_lockset_match(set, &link->cill_descr, 0)) {
370 /* XXX some locking to guarantee that locks aren't
371 * expanded in between. */
372 result = cl_lockset_lock_one(env, io, set, link);
376 cl_lock_link_fini(env, io, link);
379 list_for_each_entry_safe(link, temp,
380 &set->cls_curr, cill_linkage) {
381 lock = link->cill_lock;
382 result = cl_wait(env, lock);
384 list_move(&link->cill_linkage, &set->cls_done);
393 * Takes locks necessary for the current iteration of io.
395 * Calls cl_io_operations::cio_lock() top-to-bottom to collect locks required
396 * by layers for the current iteration. Then sort locks (to avoid dead-locks),
399 int cl_io_lock(const struct lu_env *env, struct cl_io *io)
401 const struct cl_io_slice *scan;
404 LINVRNT(cl_io_is_loopable(io));
405 LINVRNT(io->ci_state == CIS_IT_STARTED);
406 LINVRNT(cl_io_invariant(io));
409 cl_io_for_each(scan, io) {
410 if (scan->cis_iop->op[io->ci_type].cio_lock == NULL)
412 result = scan->cis_iop->op[io->ci_type].cio_lock(env, scan);
417 cl_io_locks_sort(io);
418 result = cl_lockset_lock(env, io, &io->ci_lockset);
421 cl_io_unlock(env, io);
423 io->ci_state = CIS_LOCKED;
426 EXPORT_SYMBOL(cl_io_lock);
429 * Release locks takes by io.
431 void cl_io_unlock(const struct lu_env *env, struct cl_io *io)
433 struct cl_lockset *set;
434 struct cl_io_lock_link *link;
435 struct cl_io_lock_link *temp;
436 const struct cl_io_slice *scan;
438 LASSERT(cl_io_is_loopable(io));
439 LASSERT(CIS_IT_STARTED <= io->ci_state && io->ci_state < CIS_UNLOCKED);
440 LINVRNT(cl_io_invariant(io));
443 set = &io->ci_lockset;
445 list_for_each_entry_safe(link, temp, &set->cls_todo, cill_linkage)
446 cl_lock_link_fini(env, io, link);
448 list_for_each_entry_safe(link, temp, &set->cls_curr, cill_linkage)
449 cl_lock_link_fini(env, io, link);
451 list_for_each_entry_safe(link, temp, &set->cls_done, cill_linkage) {
452 cl_unuse(env, link->cill_lock);
453 cl_lock_link_fini(env, io, link);
455 cl_io_for_each_reverse(scan, io) {
456 if (scan->cis_iop->op[io->ci_type].cio_unlock != NULL)
457 scan->cis_iop->op[io->ci_type].cio_unlock(env, scan);
459 io->ci_state = CIS_UNLOCKED;
460 LASSERT(!cl_env_info(env)->clt_counters[CNL_TOP].ctc_nr_locks_acquired);
463 EXPORT_SYMBOL(cl_io_unlock);
466 * Prepares next iteration of io.
468 * Calls cl_io_operations::cio_iter_init() top-to-bottom. This exists to give
469 * layers a chance to modify io parameters, e.g., so that lov can restrict io
470 * to a single stripe.
472 int cl_io_iter_init(const struct lu_env *env, struct cl_io *io)
474 const struct cl_io_slice *scan;
477 LINVRNT(cl_io_is_loopable(io));
478 LINVRNT(io->ci_state == CIS_INIT || io->ci_state == CIS_IT_ENDED);
479 LINVRNT(cl_io_invariant(io));
483 cl_io_for_each(scan, io) {
484 if (scan->cis_iop->op[io->ci_type].cio_iter_init == NULL)
486 result = scan->cis_iop->op[io->ci_type].cio_iter_init(env,
492 io->ci_state = CIS_IT_STARTED;
495 EXPORT_SYMBOL(cl_io_iter_init);
498 * Finalizes io iteration.
500 * Calls cl_io_operations::cio_iter_fini() bottom-to-top.
502 void cl_io_iter_fini(const struct lu_env *env, struct cl_io *io)
504 const struct cl_io_slice *scan;
506 LINVRNT(cl_io_is_loopable(io));
507 LINVRNT(io->ci_state == CIS_UNLOCKED);
508 LINVRNT(cl_io_invariant(io));
511 cl_io_for_each_reverse(scan, io) {
512 if (scan->cis_iop->op[io->ci_type].cio_iter_fini != NULL)
513 scan->cis_iop->op[io->ci_type].cio_iter_fini(env, scan);
515 io->ci_state = CIS_IT_ENDED;
518 EXPORT_SYMBOL(cl_io_iter_fini);
521 * Records that read or write io progressed \a nob bytes forward.
523 void cl_io_rw_advance(const struct lu_env *env, struct cl_io *io, size_t nob)
525 const struct cl_io_slice *scan;
527 LINVRNT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
529 LINVRNT(cl_io_is_loopable(io));
530 LINVRNT(cl_io_invariant(io));
534 io->u.ci_rw.crw_pos += nob;
535 io->u.ci_rw.crw_count -= nob;
537 /* layers have to be notified. */
538 cl_io_for_each_reverse(scan, io) {
539 if (scan->cis_iop->op[io->ci_type].cio_advance != NULL)
540 scan->cis_iop->op[io->ci_type].cio_advance(env, scan,
545 EXPORT_SYMBOL(cl_io_rw_advance);
548 * Adds a lock to a lockset.
550 int cl_io_lock_add(const struct lu_env *env, struct cl_io *io,
551 struct cl_io_lock_link *link)
556 if (cl_lockset_match(&io->ci_lockset, &link->cill_descr, 1))
559 list_add(&link->cill_linkage, &io->ci_lockset.cls_todo);
564 EXPORT_SYMBOL(cl_io_lock_add);
566 static void cl_free_io_lock_link(const struct lu_env *env,
567 struct cl_io_lock_link *link)
573 * Allocates new lock link, and uses it to add a lock to a lockset.
575 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
576 struct cl_lock_descr *descr, int enqflags)
578 struct cl_io_lock_link *link;
584 link->cill_descr = *descr;
585 link->cill_enq_flags = enqflags;
586 link->cill_fini = cl_free_io_lock_link;
587 result = cl_io_lock_add(env, io, link);
588 if (result) /* lock match */
589 link->cill_fini(env, link);
595 EXPORT_SYMBOL(cl_io_lock_alloc_add);
598 * Starts io by calling cl_io_operations::cio_start() top-to-bottom.
600 int cl_io_start(const struct lu_env *env, struct cl_io *io)
602 const struct cl_io_slice *scan;
605 LINVRNT(cl_io_is_loopable(io));
606 LINVRNT(io->ci_state == CIS_LOCKED);
607 LINVRNT(cl_io_invariant(io));
610 io->ci_state = CIS_IO_GOING;
611 cl_io_for_each(scan, io) {
612 if (scan->cis_iop->op[io->ci_type].cio_start == NULL)
614 result = scan->cis_iop->op[io->ci_type].cio_start(env, scan);
622 EXPORT_SYMBOL(cl_io_start);
625 * Wait until current io iteration is finished by calling
626 * cl_io_operations::cio_end() bottom-to-top.
628 void cl_io_end(const struct lu_env *env, struct cl_io *io)
630 const struct cl_io_slice *scan;
632 LINVRNT(cl_io_is_loopable(io));
633 LINVRNT(io->ci_state == CIS_IO_GOING);
634 LINVRNT(cl_io_invariant(io));
637 cl_io_for_each_reverse(scan, io) {
638 if (scan->cis_iop->op[io->ci_type].cio_end != NULL)
639 scan->cis_iop->op[io->ci_type].cio_end(env, scan);
640 /* TODO: error handling. */
642 io->ci_state = CIS_IO_FINISHED;
645 EXPORT_SYMBOL(cl_io_end);
647 static const struct cl_page_slice *
648 cl_io_slice_page(const struct cl_io_slice *ios, struct cl_page *page)
650 const struct cl_page_slice *slice;
652 slice = cl_page_at(page, ios->cis_obj->co_lu.lo_dev->ld_type);
653 LINVRNT(slice != NULL);
658 * True iff \a page is within \a io range.
660 static int cl_page_in_io(const struct cl_page *page, const struct cl_io *io)
667 idx = page->cp_index;
668 switch (io->ci_type) {
672 * check that [start, end) and [pos, pos + count) extents
675 start = cl_offset(page->cp_obj, idx);
676 end = cl_offset(page->cp_obj, idx + 1);
677 result = io->u.ci_rw.crw_pos < end &&
678 start < io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count;
681 result = io->u.ci_fault.ft_index == idx;
690 * Called by read io, when page has to be read from the server.
692 * \see cl_io_operations::cio_read_page()
694 int cl_io_read_page(const struct lu_env *env, struct cl_io *io,
695 struct cl_page *page)
697 const struct cl_io_slice *scan;
698 struct cl_2queue *queue;
701 LINVRNT(io->ci_type == CIT_READ || io->ci_type == CIT_FAULT);
702 LINVRNT(cl_page_is_owned(page, io));
703 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
704 LINVRNT(cl_page_in_io(page, io));
705 LINVRNT(cl_io_invariant(io));
708 queue = &io->ci_queue;
710 cl_2queue_init(queue);
712 * ->cio_read_page() methods called in the loop below are supposed to
713 * never block waiting for network (the only subtle point is the
714 * creation of new pages for read-ahead that might result in cache
715 * shrinking, but currently only clean pages are shrunk and this
716 * requires no network io).
718 * Should this ever starts blocking, retry loop would be needed for
719 * "parallel io" (see CLO_REPEAT loops in cl_lock.c).
721 cl_io_for_each(scan, io) {
722 if (scan->cis_iop->cio_read_page != NULL) {
723 const struct cl_page_slice *slice;
725 slice = cl_io_slice_page(scan, page);
726 LINVRNT(slice != NULL);
727 result = scan->cis_iop->cio_read_page(env, scan, slice);
733 result = cl_io_submit_rw(env, io, CRT_READ, queue, CRP_NORMAL);
735 * Unlock unsent pages in case of error.
737 cl_page_list_disown(env, io, &queue->c2_qin);
738 cl_2queue_fini(env, queue);
741 EXPORT_SYMBOL(cl_io_read_page);
744 * Called by write io to prepare page to receive data from user buffer.
746 * \see cl_io_operations::cio_prepare_write()
748 int cl_io_prepare_write(const struct lu_env *env, struct cl_io *io,
749 struct cl_page *page, unsigned from, unsigned to)
751 const struct cl_io_slice *scan;
754 LINVRNT(io->ci_type == CIT_WRITE);
755 LINVRNT(cl_page_is_owned(page, io));
756 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
757 LINVRNT(cl_io_invariant(io));
758 LASSERT(cl_page_in_io(page, io));
761 cl_io_for_each_reverse(scan, io) {
762 if (scan->cis_iop->cio_prepare_write != NULL) {
763 const struct cl_page_slice *slice;
765 slice = cl_io_slice_page(scan, page);
766 result = scan->cis_iop->cio_prepare_write(env, scan,
775 EXPORT_SYMBOL(cl_io_prepare_write);
778 * Called by write io after user data were copied into a page.
780 * \see cl_io_operations::cio_commit_write()
782 int cl_io_commit_write(const struct lu_env *env, struct cl_io *io,
783 struct cl_page *page, unsigned from, unsigned to)
785 const struct cl_io_slice *scan;
788 LINVRNT(io->ci_type == CIT_WRITE);
789 LINVRNT(io->ci_state == CIS_IO_GOING || io->ci_state == CIS_LOCKED);
790 LINVRNT(cl_io_invariant(io));
792 * XXX Uh... not nice. Top level cl_io_commit_write() call (vvp->lov)
793 * already called cl_page_cache_add(), moving page into CPS_CACHED
794 * state. Better (and more general) way of dealing with such situation
797 LASSERT(cl_page_is_owned(page, io) || page->cp_parent != NULL);
798 LASSERT(cl_page_in_io(page, io));
801 cl_io_for_each(scan, io) {
802 if (scan->cis_iop->cio_commit_write != NULL) {
803 const struct cl_page_slice *slice;
805 slice = cl_io_slice_page(scan, page);
806 result = scan->cis_iop->cio_commit_write(env, scan,
813 LINVRNT(result <= 0);
816 EXPORT_SYMBOL(cl_io_commit_write);
819 * Submits a list of pages for immediate io.
821 * After the function gets returned, The submitted pages are moved to
822 * queue->c2_qout queue, and queue->c2_qin contain both the pages don't need
823 * to be submitted, and the pages are errant to submit.
825 * \returns 0 if at least one page was submitted, error code otherwise.
826 * \see cl_io_operations::cio_submit()
828 int cl_io_submit_rw(const struct lu_env *env, struct cl_io *io,
829 enum cl_req_type crt, struct cl_2queue *queue,
830 enum cl_req_priority priority)
832 const struct cl_io_slice *scan;
835 LINVRNT(crt < ARRAY_SIZE(scan->cis_iop->req_op));
838 cl_io_for_each(scan, io) {
839 if (scan->cis_iop->req_op[crt].cio_submit == NULL)
841 result = scan->cis_iop->req_op[crt].cio_submit(env, scan, crt,
847 * If ->cio_submit() failed, no pages were sent.
849 LASSERT(ergo(result != 0, list_empty(&queue->c2_qout.pl_pages)));
852 EXPORT_SYMBOL(cl_io_submit_rw);
855 * Submit a sync_io and wait for the IO to be finished, or error happens.
856 * If \a timeout is zero, it means to wait for the IO unconditionally.
858 int cl_io_submit_sync(const struct lu_env *env, struct cl_io *io,
859 enum cl_req_type iot, struct cl_2queue *queue,
860 enum cl_req_priority prio, long timeout)
862 struct cl_sync_io *anchor = &cl_env_info(env)->clt_anchor;
866 LASSERT(prio == CRP_NORMAL || prio == CRP_CANCEL);
868 cl_page_list_for_each(pg, &queue->c2_qin) {
869 LASSERT(pg->cp_sync_io == NULL);
870 pg->cp_sync_io = anchor;
873 cl_sync_io_init(anchor, queue->c2_qin.pl_nr);
874 rc = cl_io_submit_rw(env, io, iot, queue, prio);
877 * If some pages weren't sent for any reason (e.g.,
878 * read found up-to-date pages in the cache, or write found
879 * clean pages), count them as completed to avoid infinite
882 cl_page_list_for_each(pg, &queue->c2_qin) {
883 pg->cp_sync_io = NULL;
884 cl_sync_io_note(anchor, +1);
887 /* wait for the IO to be finished. */
888 rc = cl_sync_io_wait(env, io, &queue->c2_qout,
891 LASSERT(list_empty(&queue->c2_qout.pl_pages));
892 cl_page_list_for_each(pg, &queue->c2_qin)
893 pg->cp_sync_io = NULL;
897 EXPORT_SYMBOL(cl_io_submit_sync);
900 * Cancel an IO which has been submitted by cl_io_submit_rw.
902 int cl_io_cancel(const struct lu_env *env, struct cl_io *io,
903 struct cl_page_list *queue)
905 struct cl_page *page;
908 CERROR("Canceling ongoing page trasmission\n");
909 cl_page_list_for_each(page, queue) {
912 LINVRNT(cl_page_in_io(page, io));
913 rc = cl_page_cancel(env, page);
914 result = result ?: rc;
918 EXPORT_SYMBOL(cl_io_cancel);
923 * Pumps io through iterations calling
925 * - cl_io_iter_init()
935 * - cl_io_iter_fini()
937 * repeatedly until there is no more io to do.
939 int cl_io_loop(const struct lu_env *env, struct cl_io *io)
943 LINVRNT(cl_io_is_loopable(io));
950 result = cl_io_iter_init(env, io);
953 result = cl_io_lock(env, io);
956 * Notify layers that locks has been taken,
959 * - llite: kms, short read;
960 * - llite: generic_file_read();
962 result = cl_io_start(env, io);
964 * Send any remaining pending
967 * - llite: ll_rw_stats_tally.
970 cl_io_unlock(env, io);
971 cl_io_rw_advance(env, io, io->ci_nob - nob);
974 cl_io_iter_fini(env, io);
975 } while (result == 0 && io->ci_continue);
976 RETURN(result < 0 ? result : 0);
978 EXPORT_SYMBOL(cl_io_loop);
981 * Adds io slice to the cl_io.
983 * This is called by cl_object_operations::coo_io_init() methods to add a
984 * per-layer state to the io. New state is added at the end of
985 * cl_io::ci_layers list, that is, it is at the bottom of the stack.
987 * \see cl_lock_slice_add(), cl_req_slice_add(), cl_page_slice_add()
989 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
990 struct cl_object *obj,
991 const struct cl_io_operations *ops)
993 struct list_head *linkage = &slice->cis_linkage;
995 LASSERT((linkage->prev == NULL && linkage->next == NULL) ||
996 list_empty(linkage));
999 list_add_tail(linkage, &io->ci_layers);
1001 slice->cis_obj = obj;
1002 slice->cis_iop = ops;
1005 EXPORT_SYMBOL(cl_io_slice_add);
1009 * Initializes page list.
1011 void cl_page_list_init(struct cl_page_list *plist)
1015 CFS_INIT_LIST_HEAD(&plist->pl_pages);
1016 plist->pl_owner = cfs_current();
1019 EXPORT_SYMBOL(cl_page_list_init);
1022 * Adds a page to a page list.
1024 void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page)
1027 /* it would be better to check that page is owned by "current" io, but
1028 * it is not passed here. */
1029 LASSERT(page->cp_owner != NULL);
1030 LINVRNT(plist->pl_owner == cfs_current());
1033 mutex_lock(&page->cp_mutex);
1035 LASSERT(list_empty(&page->cp_batch));
1036 list_add_tail(&page->cp_batch, &plist->pl_pages);
1038 page->cp_queue_ref = lu_ref_add(&page->cp_reference, "queue", plist);
1042 EXPORT_SYMBOL(cl_page_list_add);
1045 * Removes a page from a page list.
1047 void cl_page_list_del(const struct lu_env *env,
1048 struct cl_page_list *plist, struct cl_page *page)
1050 LASSERT(plist->pl_nr > 0);
1051 LINVRNT(plist->pl_owner == cfs_current());
1054 list_del_init(&page->cp_batch);
1056 mutex_unlock(&page->cp_mutex);
1059 lu_ref_del_at(&page->cp_reference, page->cp_queue_ref, "queue", plist);
1060 cl_page_put(env, page);
1063 EXPORT_SYMBOL(cl_page_list_del);
1066 * Moves a page from one page list to another.
1068 void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src,
1069 struct cl_page *page)
1071 LASSERT(src->pl_nr > 0);
1072 LINVRNT(dst->pl_owner == cfs_current());
1073 LINVRNT(src->pl_owner == cfs_current());
1076 list_move_tail(&page->cp_batch, &dst->pl_pages);
1079 lu_ref_set_at(&page->cp_reference,
1080 page->cp_queue_ref, "queue", src, dst);
1083 EXPORT_SYMBOL(cl_page_list_move);
1086 * splice the cl_page_list, just as list head does
1088 void cl_page_list_splice(struct cl_page_list *list, struct cl_page_list *head)
1090 struct cl_page *page;
1091 struct cl_page *tmp;
1093 LINVRNT(list->pl_owner == cfs_current());
1094 LINVRNT(head->pl_owner == cfs_current());
1097 cl_page_list_for_each_safe(page, tmp, list)
1098 cl_page_list_move(head, list, page);
1101 EXPORT_SYMBOL(cl_page_list_splice);
1103 void cl_page_disown0(const struct lu_env *env,
1104 struct cl_io *io, struct cl_page *pg);
1107 * Disowns pages in a queue.
1109 void cl_page_list_disown(const struct lu_env *env,
1110 struct cl_io *io, struct cl_page_list *plist)
1112 struct cl_page *page;
1113 struct cl_page *temp;
1115 LINVRNT(plist->pl_owner == cfs_current());
1118 cl_page_list_for_each_safe(page, temp, plist) {
1119 LASSERT(plist->pl_nr > 0);
1121 list_del_init(&page->cp_batch);
1123 mutex_unlock(&page->cp_mutex);
1127 * cl_page_disown0 rather than usual cl_page_disown() is used,
1128 * because pages are possibly in CPS_FREEING state already due
1129 * to the call to cl_page_list_discard().
1132 * XXX cl_page_disown0() will fail if page is not locked.
1134 cl_page_disown0(env, io, page);
1135 lu_ref_del(&page->cp_reference, "queue", plist);
1136 cl_page_put(env, page);
1140 EXPORT_SYMBOL(cl_page_list_disown);
1143 * Releases pages from queue.
1145 void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist)
1147 struct cl_page *page;
1148 struct cl_page *temp;
1150 LINVRNT(plist->pl_owner == cfs_current());
1153 cl_page_list_for_each_safe(page, temp, plist)
1154 cl_page_list_del(env, plist, page);
1155 LASSERT(plist->pl_nr == 0);
1158 EXPORT_SYMBOL(cl_page_list_fini);
1161 * Owns all pages in a queue.
1163 int cl_page_list_own(const struct lu_env *env,
1164 struct cl_io *io, struct cl_page_list *plist)
1166 struct cl_page *page;
1167 struct cl_page *temp;
1171 LINVRNT(plist->pl_owner == cfs_current());
1175 cl_page_list_for_each_safe(page, temp, plist) {
1176 LASSERT(index <= page->cp_index);
1177 index = page->cp_index;
1178 if (cl_page_own(env, io, page) == 0)
1179 result = result ?: page->cp_error;
1181 cl_page_list_del(env, plist, page);
1185 EXPORT_SYMBOL(cl_page_list_own);
1188 * Assumes all pages in a queue.
1190 void cl_page_list_assume(const struct lu_env *env,
1191 struct cl_io *io, struct cl_page_list *plist)
1193 struct cl_page *page;
1195 LINVRNT(plist->pl_owner == cfs_current());
1197 cl_page_list_for_each(page, plist)
1198 cl_page_assume(env, io, page);
1200 EXPORT_SYMBOL(cl_page_list_assume);
1203 * Discards all pages in a queue.
1205 void cl_page_list_discard(const struct lu_env *env, struct cl_io *io,
1206 struct cl_page_list *plist)
1208 struct cl_page *page;
1210 LINVRNT(plist->pl_owner == cfs_current());
1212 cl_page_list_for_each(page, plist)
1213 cl_page_discard(env, io, page);
1216 EXPORT_SYMBOL(cl_page_list_discard);
1219 * Unmaps all pages in a queue from user virtual memory.
1221 int cl_page_list_unmap(const struct lu_env *env, struct cl_io *io,
1222 struct cl_page_list *plist)
1224 struct cl_page *page;
1227 LINVRNT(plist->pl_owner == cfs_current());
1230 cl_page_list_for_each(page, plist) {
1231 result = cl_page_unmap(env, io, page);
1237 EXPORT_SYMBOL(cl_page_list_unmap);
1240 * Initialize dual page queue.
1242 void cl_2queue_init(struct cl_2queue *queue)
1245 cl_page_list_init(&queue->c2_qin);
1246 cl_page_list_init(&queue->c2_qout);
1249 EXPORT_SYMBOL(cl_2queue_init);
1252 * Add a page to the incoming page list of 2-queue.
1254 void cl_2queue_add(struct cl_2queue *queue, struct cl_page *page)
1257 cl_page_list_add(&queue->c2_qin, page);
1260 EXPORT_SYMBOL(cl_2queue_add);
1263 * Disown pages in both lists of a 2-queue.
1265 void cl_2queue_disown(const struct lu_env *env,
1266 struct cl_io *io, struct cl_2queue *queue)
1269 cl_page_list_disown(env, io, &queue->c2_qin);
1270 cl_page_list_disown(env, io, &queue->c2_qout);
1273 EXPORT_SYMBOL(cl_2queue_disown);
1276 * Discard (truncate) pages in both lists of a 2-queue.
1278 void cl_2queue_discard(const struct lu_env *env,
1279 struct cl_io *io, struct cl_2queue *queue)
1282 cl_page_list_discard(env, io, &queue->c2_qin);
1283 cl_page_list_discard(env, io, &queue->c2_qout);
1286 EXPORT_SYMBOL(cl_2queue_discard);
1289 * Assume to own the pages in cl_2queue
1291 void cl_2queue_assume(const struct lu_env *env,
1292 struct cl_io *io, struct cl_2queue *queue)
1294 cl_page_list_assume(env, io, &queue->c2_qin);
1295 cl_page_list_assume(env, io, &queue->c2_qout);
1297 EXPORT_SYMBOL(cl_2queue_assume);
1300 * Finalize both page lists of a 2-queue.
1302 void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue)
1305 cl_page_list_fini(env, &queue->c2_qout);
1306 cl_page_list_fini(env, &queue->c2_qin);
1309 EXPORT_SYMBOL(cl_2queue_fini);
1312 * Initialize a 2-queue to contain \a page in its incoming page list.
1314 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page)
1317 cl_2queue_init(queue);
1318 cl_2queue_add(queue, page);
1321 EXPORT_SYMBOL(cl_2queue_init_page);
1324 * Returns top-level io.
1326 * \see cl_object_top(), cl_page_top().
1328 struct cl_io *cl_io_top(struct cl_io *io)
1331 while (io->ci_parent != NULL)
1335 EXPORT_SYMBOL(cl_io_top);
1338 * Prints human readable representation of \a io to the \a f.
1340 void cl_io_print(const struct lu_env *env, void *cookie,
1341 lu_printer_t printer, const struct cl_io *io)
1346 * Adds request slice to the compound request.
1348 * This is called by cl_device_operations::cdo_req_init() methods to add a
1349 * per-layer state to the request. New state is added at the end of
1350 * cl_req::crq_layers list, that is, it is at the bottom of the stack.
1352 * \see cl_lock_slice_add(), cl_page_slice_add(), cl_io_slice_add()
1354 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
1355 struct cl_device *dev,
1356 const struct cl_req_operations *ops)
1359 list_add_tail(&slice->crs_linkage, &req->crq_layers);
1360 slice->crs_dev = dev;
1361 slice->crs_ops = ops;
1362 slice->crs_req = req;
1365 EXPORT_SYMBOL(cl_req_slice_add);
1367 static void cl_req_free(const struct lu_env *env, struct cl_req *req)
1371 LASSERT(list_empty(&req->crq_pages));
1372 LASSERT(req->crq_nrpages == 0);
1373 LINVRNT(list_empty(&req->crq_layers));
1374 LINVRNT(equi(req->crq_nrobjs > 0, req->crq_o != NULL));
1377 if (req->crq_o != NULL) {
1378 for (i = 0; i < req->crq_nrobjs; ++i) {
1379 struct cl_object *obj = req->crq_o[i].ro_obj;
1381 lu_object_ref_del_at(&obj->co_lu,
1382 req->crq_o[i].ro_obj_ref,
1384 cl_object_put(env, obj);
1387 OBD_FREE(req->crq_o, req->crq_nrobjs * sizeof req->crq_o[0]);
1393 static int cl_req_init(const struct lu_env *env, struct cl_req *req,
1394 struct cl_page *page)
1396 struct cl_device *dev;
1397 struct cl_page_slice *slice;
1402 page = cl_page_top(page);
1404 list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
1405 dev = lu2cl_dev(slice->cpl_obj->co_lu.lo_dev);
1406 if (dev->cd_ops->cdo_req_init != NULL) {
1407 result = dev->cd_ops->cdo_req_init(env,
1413 page = page->cp_child;
1414 } while (page != NULL && result == 0);
1419 * Invokes per-request transfer completion call-backs
1420 * (cl_req_operations::cro_completion()) bottom-to-top.
1422 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int rc)
1424 struct cl_req_slice *slice;
1428 * for the lack of list_for_each_entry_reverse_safe()...
1430 while (!list_empty(&req->crq_layers)) {
1431 slice = list_entry(req->crq_layers.prev,
1432 struct cl_req_slice, crs_linkage);
1433 list_del_init(&slice->crs_linkage);
1434 if (slice->crs_ops->cro_completion != NULL)
1435 slice->crs_ops->cro_completion(env, slice, rc);
1437 cl_req_free(env, req);
1440 EXPORT_SYMBOL(cl_req_completion);
1443 * Allocates new transfer request.
1445 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
1446 enum cl_req_type crt, int nr_objects)
1450 LINVRNT(nr_objects > 0);
1457 OBD_ALLOC(req->crq_o, nr_objects * sizeof req->crq_o[0]);
1458 if (req->crq_o != NULL) {
1459 req->crq_nrobjs = nr_objects;
1460 req->crq_type = crt;
1461 CFS_INIT_LIST_HEAD(&req->crq_pages);
1462 CFS_INIT_LIST_HEAD(&req->crq_layers);
1463 result = cl_req_init(env, req, page);
1467 cl_req_completion(env, req, result);
1468 req = ERR_PTR(result);
1471 req = ERR_PTR(-ENOMEM);
1474 EXPORT_SYMBOL(cl_req_alloc);
1477 * Adds a page to a request.
1479 void cl_req_page_add(const struct lu_env *env,
1480 struct cl_req *req, struct cl_page *page)
1482 struct cl_object *obj;
1483 struct cl_req_obj *rqo;
1487 page = cl_page_top(page);
1489 LINVRNT(cl_page_is_vmlocked(env, page));
1490 LASSERT(list_empty(&page->cp_flight));
1491 LASSERT(page->cp_req == NULL);
1493 list_add_tail(&page->cp_flight, &req->crq_pages);
1496 obj = cl_object_top(page->cp_obj);
1497 for (i = 0, rqo = req->crq_o; obj != rqo->ro_obj; ++i, ++rqo) {
1498 if (rqo->ro_obj == NULL) {
1501 rqo->ro_obj_ref = lu_object_ref_add(&obj->co_lu,
1506 LASSERT(i < req->crq_nrobjs);
1509 EXPORT_SYMBOL(cl_req_page_add);
1512 * Removes a page from a request.
1514 void cl_req_page_done(const struct lu_env *env, struct cl_page *page)
1516 struct cl_req *req = page->cp_req;
1519 page = cl_page_top(page);
1521 LINVRNT(cl_page_is_vmlocked(env, page));
1522 LASSERT(!list_empty(&page->cp_flight));
1523 LASSERT(req->crq_nrpages > 0);
1525 list_del_init(&page->cp_flight);
1527 page->cp_req = NULL;
1530 EXPORT_SYMBOL(cl_req_page_done);
1533 * Notifies layers that request is about to depart by calling
1534 * cl_req_operations::cro_prep() top-to-bottom.
1536 int cl_req_prep(const struct lu_env *env, struct cl_req *req)
1540 const struct cl_req_slice *slice;
1544 * Check that the caller of cl_req_alloc() didn't lie about the number
1547 for (i = 0; i < req->crq_nrobjs; ++i)
1548 LASSERT(req->crq_o[i].ro_obj != NULL);
1551 list_for_each_entry(slice, &req->crq_layers, crs_linkage) {
1552 if (slice->crs_ops->cro_prep != NULL) {
1553 result = slice->crs_ops->cro_prep(env, slice);
1560 EXPORT_SYMBOL(cl_req_prep);
1563 * Fills in attributes that are passed to server together with transfer. Only
1564 * attributes from \a flags may be touched. This can be called multiple times
1565 * for the same request.
1567 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
1568 struct cl_req_attr *attr, obd_valid flags)
1570 const struct cl_req_slice *slice;
1571 struct cl_page *page;
1574 LASSERT(!list_empty(&req->crq_pages));
1577 /* Take any page to use as a model. */
1578 page = list_entry(req->crq_pages.next, struct cl_page, cp_flight);
1580 for (i = 0; i < req->crq_nrobjs; ++i) {
1581 list_for_each_entry(slice, &req->crq_layers, crs_linkage) {
1582 const struct cl_page_slice *scan;
1583 const struct cl_object *obj;
1585 scan = cl_page_at(page,
1586 slice->crs_dev->cd_lu_dev.ld_type);
1587 LASSERT(scan != NULL);
1588 obj = scan->cpl_obj;
1589 if (slice->crs_ops->cro_attr_set != NULL)
1590 slice->crs_ops->cro_attr_set(env, slice, obj,
1596 EXPORT_SYMBOL(cl_req_attr_set);
1598 /* XXX complete(), init_completion(), and wait_for_completion(), until they are
1599 * implemented in libcfs. */
1601 # include <linux/sched.h>
1602 #else /* __KERNEL__ */
1603 # include <liblustre.h>
1607 * Initialize synchronous io wait anchor, for transfer of \a nrpages pages.
1609 void cl_sync_io_init(struct cl_sync_io *anchor, int nrpages)
1612 cfs_waitq_init(&anchor->csi_waitq);
1613 atomic_set(&anchor->csi_sync_nr, nrpages);
1614 anchor->csi_sync_rc = 0;
1617 EXPORT_SYMBOL(cl_sync_io_init);
1620 * Wait until all transfer completes. Transfer completion routine has to call
1621 * cl_sync_io_note() for every page.
1623 int cl_sync_io_wait(const struct lu_env *env, struct cl_io *io,
1624 struct cl_page_list *queue, struct cl_sync_io *anchor,
1627 struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
1632 LASSERT(timeout >= 0);
1634 rc = l_wait_event(anchor->csi_waitq,
1635 atomic_read(&anchor->csi_sync_nr) == 0,
1640 CERROR("SYNC IO failed with error: %d, try to cancel "
1641 "the remaining page\n", rc);
1643 rc2 = cl_io_cancel(env, io, queue);
1645 lwi = (struct l_wait_info) { 0 };
1646 /* Too bad, some pages are still in IO. */
1647 CERROR("Failed to cancel transfer error: %d, mostly "
1648 "because of they are still being transferred, "
1649 "waiting for %i pages\n",
1650 rc2, atomic_read(&anchor->csi_sync_nr));
1651 (void)l_wait_event(anchor->csi_waitq,
1652 atomic_read(&anchor->csi_sync_nr) == 0,
1656 rc = anchor->csi_sync_rc;
1658 LASSERT(atomic_read(&anchor->csi_sync_nr) == 0);
1659 cl_page_list_assume(env, io, queue);
1660 POISON(anchor, 0x5a, sizeof *anchor);
1663 EXPORT_SYMBOL(cl_sync_io_wait);
1666 * Indicate that transfer of a single page completed.
1668 void cl_sync_io_note(struct cl_sync_io *anchor, int ioret)
1671 if (anchor->csi_sync_rc == 0 && ioret < 0)
1672 anchor->csi_sync_rc = ioret;
1674 * Synchronous IO done without releasing page lock (e.g., as a part of
1675 * ->{prepare,commit}_write(). Completion is used to signal the end of
1678 if (atomic_dec_and_test(&anchor->csi_sync_nr))
1679 cfs_waitq_broadcast(&anchor->csi_waitq);
1682 EXPORT_SYMBOL(cl_sync_io_note);