4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Implementation of cl_io for LOV layer.
34 * Author: Nikita Danilov <nikita.danilov@sun.com>
35 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
38 #define DEBUG_SUBSYSTEM S_LOV
40 #include "lov_cl_internal.h"
46 static inline struct lov_io_sub *lov_sub_alloc(struct lov_io *lio, int index)
48 struct lov_io_sub *sub;
50 if (lio->lis_nr_subios == 0) {
51 LASSERT(lio->lis_single_subio_index == -1);
52 sub = &lio->lis_single_subio;
53 lio->lis_single_subio_index = index;
54 memset(sub, 0, sizeof(*sub));
60 INIT_LIST_HEAD(&sub->sub_list);
61 INIT_LIST_HEAD(&sub->sub_linkage);
62 sub->sub_subio_index = index;
68 static inline void lov_sub_free(struct lov_io *lio, struct lov_io_sub *sub)
70 if (sub->sub_subio_index == lio->lis_single_subio_index) {
71 LASSERT(sub == &lio->lis_single_subio);
72 lio->lis_single_subio_index = -1;
78 static void lov_io_sub_fini(const struct lu_env *env, struct lov_io *lio,
79 struct lov_io_sub *sub)
83 cl_io_fini(sub->sub_env, &sub->sub_io);
85 if (sub->sub_env && !IS_ERR(sub->sub_env)) {
86 cl_env_put(sub->sub_env, &sub->sub_refcheck);
93 is_index_within_mirror(struct lov_object *lov, int index, int mirror_index)
95 struct lov_layout_composite *comp = &lov->u.composite;
96 struct lov_mirror_entry *lre = &comp->lo_mirrors[mirror_index];
98 return (index >= lre->lre_start && index <= lre->lre_end);
101 static int lov_io_sub_init(const struct lu_env *env, struct lov_io *lio,
102 struct lov_io_sub *sub)
104 struct lov_object *lov = lio->lis_object;
105 struct cl_io *sub_io;
106 struct cl_object *sub_obj;
107 struct cl_io *io = lio->lis_cl.cis_io;
108 int index = lov_comp_entry(sub->sub_subio_index);
109 int stripe = lov_comp_stripe(sub->sub_subio_index);
111 LASSERT(sub->sub_env == NULL);
114 if (unlikely(!lov_r0(lov, index)->lo_sub ||
115 !lov_r0(lov, index)->lo_sub[stripe]))
118 LASSERTF(is_index_within_mirror(lov, index, lio->lis_mirror_index),
119 DFID "iot = %d, index = %d, mirror = %d\n",
120 PFID(lu_object_fid(lov2lu(lov))), io->ci_type, index,
121 lio->lis_mirror_index);
123 /* obtain new environment */
124 sub->sub_env = cl_env_get(&sub->sub_refcheck);
125 if (IS_ERR(sub->sub_env)) {
126 result = PTR_ERR(sub->sub_env);
130 sub_obj = lovsub2cl(lov_r0(lov, index)->lo_sub[stripe]);
131 sub_io = &sub->sub_io;
133 sub_io->ci_obj = sub_obj;
134 sub_io->ci_result = 0;
136 sub_io->ci_parent = io;
137 sub_io->ci_lockreq = io->ci_lockreq;
138 sub_io->ci_type = io->ci_type;
139 sub_io->ci_no_srvlock = io->ci_no_srvlock;
140 sub_io->ci_noatime = io->ci_noatime;
141 sub_io->ci_async_readahead = io->ci_async_readahead;
142 sub_io->ci_lock_no_expand = io->ci_lock_no_expand;
143 sub_io->ci_ndelay = io->ci_ndelay;
144 sub_io->ci_layout_version = io->ci_layout_version;
145 sub_io->ci_tried_all_mirrors = io->ci_tried_all_mirrors;
147 result = cl_io_sub_init(sub->sub_env, sub_io, io->ci_type, sub_obj);
150 lov_io_sub_fini(env, lio, sub);
155 struct lov_io_sub *lov_sub_get(const struct lu_env *env,
156 struct lov_io *lio, int index)
158 struct lov_io_sub *sub;
163 list_for_each_entry(sub, &lio->lis_subios, sub_list) {
164 if (sub->sub_subio_index == index) {
171 sub = lov_sub_alloc(lio, index);
173 GOTO(out, rc = -ENOMEM);
175 rc = lov_io_sub_init(env, lio, sub);
177 lov_sub_free(lio, sub);
181 list_add_tail(&sub->sub_list, &lio->lis_subios);
182 lio->lis_nr_subios++;
188 sub->sub_io.ci_noquota = lio->lis_cl.cis_io->ci_noquota;
192 /*****************************************************************************
197 static int lov_io_subio_init(const struct lu_env *env, struct lov_io *lio,
202 LASSERT(lio->lis_object != NULL);
204 INIT_LIST_HEAD(&lio->lis_subios);
205 lio->lis_single_subio_index = -1;
206 lio->lis_nr_subios = 0;
212 * Decide if it will need write intent RPC
214 static int lov_io_mirror_write_intent(struct lov_io *lio,
215 struct lov_object *obj, struct cl_io *io)
217 struct lov_layout_composite *comp = &obj->u.composite;
218 struct lu_extent *ext = &io->ci_write_intent;
219 struct lov_mirror_entry *lre;
220 struct lov_mirror_entry *primary;
221 struct lov_layout_entry *lle;
225 *ext = (typeof(*ext)) { lio->lis_pos, lio->lis_endpos };
226 io->ci_need_write_intent = 0;
228 if (!(io->ci_type == CIT_WRITE || cl_io_is_trunc(io) ||
229 cl_io_is_mkwrite(io)))
233 * FLR: check if it needs to send a write intent RPC to server.
234 * Writing to sync_pending file needs write intent RPC to change
235 * the file state back to write_pending, so that the layout version
236 * can be increased when the state changes to sync_pending at a later
237 * time. Otherwise there exists a chance that an evicted client may
238 * dirty the file data while resync client is working on it.
239 * Designated I/O is allowed for resync workload.
241 if (lov_flr_state(obj) == LCM_FL_RDONLY ||
242 (lov_flr_state(obj) == LCM_FL_SYNC_PENDING &&
243 io->ci_designated_mirror == 0)) {
244 io->ci_need_write_intent = 1;
248 LASSERT((lov_flr_state(obj) == LCM_FL_WRITE_PENDING));
249 LASSERT(comp->lo_preferred_mirror >= 0);
252 * need to iterate all components to see if there are
253 * multiple components covering the writing component
255 primary = &comp->lo_mirrors[comp->lo_preferred_mirror];
256 LASSERT(!primary->lre_stale);
257 lov_foreach_mirror_layout_entry(obj, lle, primary) {
258 LASSERT(lle->lle_valid);
259 if (!lu_extent_is_overlapped(ext, lle->lle_extent))
262 ext->e_start = min(ext->e_start, lle->lle_extent->e_start);
263 ext->e_end = max(ext->e_end, lle->lle_extent->e_end);
267 CERROR(DFID ": cannot find any valid components covering "
268 "file extent "DEXT", mirror: %d\n",
269 PFID(lu_object_fid(lov2lu(obj))), PEXT(ext),
270 primary->lre_mirror_id);
275 lov_foreach_mirror_entry(obj, lre) {
279 lov_foreach_mirror_layout_entry(obj, lle, lre) {
283 if (lu_extent_is_overlapped(ext, lle->lle_extent)) {
290 CDEBUG(D_VFSTRACE, DFID "there are %zd components to be staled to "
291 "modify file extent "DEXT", iot: %d\n",
292 PFID(lu_object_fid(lov2lu(obj))), count, PEXT(ext), io->ci_type);
294 io->ci_need_write_intent = count > 0;
299 static int lov_io_mirror_init(struct lov_io *lio, struct lov_object *obj,
302 struct lov_layout_composite *comp = &obj->u.composite;
308 if (!lov_is_flr(obj)) {
309 LASSERT(comp->lo_preferred_mirror == 0);
310 lio->lis_mirror_index = comp->lo_preferred_mirror;
315 /* transfer the layout version for verification */
316 if (io->ci_layout_version == 0)
317 io->ci_layout_version = obj->lo_lsm->lsm_layout_gen;
319 /* find the corresponding mirror for designated mirror IO */
320 if (io->ci_designated_mirror > 0) {
321 struct lov_mirror_entry *entry;
323 LASSERT(!io->ci_ndelay);
325 CDEBUG(D_LAYOUT, "designated I/O mirror state: %d\n",
328 if ((cl_io_is_trunc(io) || io->ci_type == CIT_WRITE) &&
329 (io->ci_layout_version != obj->lo_lsm->lsm_layout_gen)) {
331 * For resync I/O, the ci_layout_version was the layout
332 * version when resync starts. If it doesn't match the
333 * current object layout version, it means the layout
339 io->ci_layout_version |= LU_LAYOUT_RESYNC;
342 lio->lis_mirror_index = -1;
343 lov_foreach_mirror_entry(obj, entry) {
344 if (entry->lre_mirror_id ==
345 io->ci_designated_mirror) {
346 lio->lis_mirror_index = index;
353 RETURN(lio->lis_mirror_index < 0 ? -EINVAL : 0);
356 result = lov_io_mirror_write_intent(lio, obj, io);
360 if (io->ci_need_write_intent) {
361 CDEBUG(D_VFSTRACE, DFID " need write intent for [%llu, %llu)\n",
362 PFID(lu_object_fid(lov2lu(obj))),
363 lio->lis_pos, lio->lis_endpos);
365 if (cl_io_is_trunc(io)) {
367 * for truncate, we uses [size, EOF) to judge whether
368 * a write intent needs to be send, but we need to
369 * restore the write extent to [0, size).
371 io->ci_write_intent.e_start = 0;
372 io->ci_write_intent.e_end =
373 io->u.ci_setattr.sa_attr.lvb_size;
375 /* stop cl_io_init() loop */
379 if (io->ci_ndelay_tried == 0 || /* first time to try */
380 /* reset the mirror index if layout has changed */
381 lio->lis_mirror_layout_gen != obj->lo_lsm->lsm_layout_gen) {
382 lio->lis_mirror_layout_gen = obj->lo_lsm->lsm_layout_gen;
383 index = lio->lis_mirror_index = comp->lo_preferred_mirror;
385 index = lio->lis_mirror_index;
388 /* move mirror index to the next one */
389 index = (index + 1) % comp->lo_mirror_count;
392 for (i = 0; i < comp->lo_mirror_count; i++) {
393 struct lu_extent ext = { .e_start = lio->lis_pos,
394 .e_end = lio->lis_pos + 1 };
395 struct lov_mirror_entry *lre;
396 struct lov_layout_entry *lle;
399 lre = &comp->lo_mirrors[(index + i) % comp->lo_mirror_count];
403 lov_foreach_mirror_layout_entry(obj, lle, lre) {
407 if (lu_extent_is_overlapped(&ext, lle->lle_extent)) {
411 } /* each component of the mirror */
413 index = (index + i) % comp->lo_mirror_count;
418 if (i == comp->lo_mirror_count) {
419 CERROR(DFID": failed to find a component covering "
420 "I/O region at %llu\n",
421 PFID(lu_object_fid(lov2lu(obj))), lio->lis_pos);
423 dump_lsm(D_ERROR, obj->lo_lsm);
428 CDEBUG(D_VFSTRACE, DFID ": flr state: %d, move mirror from %d to %d, "
429 "have retried: %d, mirror count: %d\n",
430 PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj),
431 lio->lis_mirror_index, index, io->ci_ndelay_tried,
432 comp->lo_mirror_count);
434 lio->lis_mirror_index = index;
437 * FLR: if all mirrors have been tried once, most likely the network
438 * of this client has been partitioned. We should relinquish CPU for
439 * a while before trying again.
441 if (io->ci_ndelay && io->ci_ndelay_tried > 0 &&
442 (io->ci_ndelay_tried % comp->lo_mirror_count == 0)) {
443 schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
444 if (signal_pending(current))
448 * we'd set ci_tried_all_mirrors to turn off fast mirror
449 * switching for read after we've tried all mirrors several
452 io->ci_tried_all_mirrors = io->ci_ndelay_tried %
453 (comp->lo_mirror_count * 4) == 0;
455 ++io->ci_ndelay_tried;
457 CDEBUG(D_VFSTRACE, "use %sdelayed RPC state for this IO\n",
458 io->ci_ndelay ? "non-" : "");
463 static int lov_io_slice_init(struct lov_io *lio,
464 struct lov_object *obj, struct cl_io *io)
471 lio->lis_object = obj;
473 LASSERT(obj->lo_lsm != NULL);
475 switch (io->ci_type) {
478 lio->lis_pos = io->u.ci_rw.crw_pos;
479 lio->lis_endpos = io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count;
480 lio->lis_io_endpos = lio->lis_endpos;
481 if (cl_io_is_append(io)) {
482 LASSERT(io->ci_type == CIT_WRITE);
485 * If there is LOV EA hole, then we may cannot locate
486 * the current file-tail exactly.
488 if (unlikely(obj->lo_lsm->lsm_entries[0]->lsme_pattern &
490 GOTO(out, result = -EIO);
493 lio->lis_endpos = OBD_OBJECT_EOF;
498 if (cl_io_is_trunc(io))
499 lio->lis_pos = io->u.ci_setattr.sa_attr.lvb_size;
502 lio->lis_endpos = OBD_OBJECT_EOF;
505 case CIT_DATA_VERSION:
507 lio->lis_endpos = OBD_OBJECT_EOF;
511 pgoff_t index = io->u.ci_fault.ft_index;
513 lio->lis_pos = cl_offset(io->ci_obj, index);
514 lio->lis_endpos = cl_offset(io->ci_obj, index + 1);
519 lio->lis_pos = io->u.ci_fsync.fi_start;
520 lio->lis_endpos = io->u.ci_fsync.fi_end;
525 lio->lis_pos = io->u.ci_ladvise.li_start;
526 lio->lis_endpos = io->u.ci_ladvise.li_end;
532 lio->lis_endpos = OBD_OBJECT_EOF;
534 if (lov_flr_state(obj) == LCM_FL_RDONLY &&
535 !OBD_FAIL_CHECK(OBD_FAIL_FLR_GLIMPSE_IMMUTABLE))
536 /* SoM is accurate, no need glimpse */
537 GOTO(out, result = 1);
542 lio->lis_endpos = OBD_OBJECT_EOF;
549 result = lov_io_mirror_init(lio, obj, io);
553 /* check if it needs to instantiate layout */
554 if (!(io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io) ||
555 (cl_io_is_trunc(io) && io->u.ci_setattr.sa_attr.lvb_size > 0)))
556 GOTO(out, result = 0);
559 * for truncate, it only needs to instantiate the components
560 * before the truncated size.
562 if (cl_io_is_trunc(io)) {
563 io->ci_write_intent.e_start = 0;
564 /* for writes, e_end is endpos, the location of the file
565 * pointer after the write is completed, so it is not accessed.
566 * For truncate, 'end' is the size, and *is* acccessed.
567 * In other words, writes are [start, end), but truncate is
568 * [start, size], where both are included. So add 1 to the
569 * size when creating the write intent to account for this.
571 io->ci_write_intent.e_end =
572 io->u.ci_setattr.sa_attr.lvb_size + 1;
574 io->ci_write_intent.e_start = lio->lis_pos;
575 io->ci_write_intent.e_end = lio->lis_endpos;
579 lov_foreach_io_layout(index, lio, &io->ci_write_intent) {
580 if (!lsm_entry_inited(obj->lo_lsm, index)) {
581 io->ci_need_write_intent = 1;
586 if (io->ci_need_write_intent && io->ci_designated_mirror > 0) {
588 * REINT_SYNC RPC has already tried to instantiate all of the
589 * components involved, obviously it didn't succeed. Skip this
590 * mirror for now. The server won't be able to figure out
591 * which mirror it should instantiate components
593 CERROR(DFID": trying to instantiate components for designated "
594 "I/O, file state: %d\n",
595 PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj));
597 io->ci_need_write_intent = 0;
598 GOTO(out, result = -EIO);
601 if (io->ci_need_write_intent)
602 GOTO(out, result = 1);
610 static void lov_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
612 struct lov_io *lio = cl2lov_io(env, ios);
613 struct lov_object *lov = cl2lov(ios->cis_obj);
617 LASSERT(list_empty(&lio->lis_active));
619 while (!list_empty(&lio->lis_subios)) {
620 struct lov_io_sub *sub = list_entry(lio->lis_subios.next,
624 list_del_init(&sub->sub_list);
625 lio->lis_nr_subios--;
627 lov_io_sub_fini(env, lio, sub);
628 lov_sub_free(lio, sub);
630 LASSERT(lio->lis_nr_subios == 0);
632 LASSERT(atomic_read(&lov->lo_active_ios) > 0);
633 if (atomic_dec_and_test(&lov->lo_active_ios))
634 wake_up_all(&lov->lo_waitq);
638 static void lov_io_sub_inherit(struct lov_io_sub *sub, struct lov_io *lio,
639 loff_t start, loff_t end)
641 struct cl_io *io = &sub->sub_io;
642 struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
643 struct cl_io *parent = lio->lis_cl.cis_io;
644 int index = lov_comp_entry(sub->sub_subio_index);
645 int stripe = lov_comp_stripe(sub->sub_subio_index);
647 switch (io->ci_type) {
649 io->u.ci_setattr.sa_attr = parent->u.ci_setattr.sa_attr;
650 io->u.ci_setattr.sa_attr_flags =
651 parent->u.ci_setattr.sa_attr_flags;
652 io->u.ci_setattr.sa_avalid = parent->u.ci_setattr.sa_avalid;
653 io->u.ci_setattr.sa_xvalid = parent->u.ci_setattr.sa_xvalid;
654 io->u.ci_setattr.sa_stripe_index = stripe;
655 io->u.ci_setattr.sa_parent_fid =
656 parent->u.ci_setattr.sa_parent_fid;
657 if (cl_io_is_trunc(io)) {
658 loff_t new_size = parent->u.ci_setattr.sa_attr.lvb_size;
660 new_size = lov_size_to_stripe(lsm, index, new_size,
662 io->u.ci_setattr.sa_attr.lvb_size = new_size;
664 lov_lsm2layout(lsm, lsm->lsm_entries[index],
665 &io->u.ci_setattr.sa_layout);
668 case CIT_DATA_VERSION: {
669 io->u.ci_data_version.dv_data_version = 0;
670 io->u.ci_data_version.dv_flags =
671 parent->u.ci_data_version.dv_flags;
675 struct cl_object *obj = parent->ci_obj;
676 loff_t off = cl_offset(obj, parent->u.ci_fault.ft_index);
678 io->u.ci_fault = parent->u.ci_fault;
679 off = lov_size_to_stripe(lsm, index, off, stripe);
680 io->u.ci_fault.ft_index = cl_index(obj, off);
684 io->u.ci_fsync.fi_start = start;
685 io->u.ci_fsync.fi_end = end;
686 io->u.ci_fsync.fi_fid = parent->u.ci_fsync.fi_fid;
687 io->u.ci_fsync.fi_mode = parent->u.ci_fsync.fi_mode;
692 io->u.ci_wr.wr_sync = cl_io_is_sync_write(parent);
693 io->ci_tried_all_mirrors = parent->ci_tried_all_mirrors;
694 if (cl_io_is_append(parent)) {
695 io->u.ci_wr.wr_append = 1;
697 io->u.ci_rw.crw_pos = start;
698 io->u.ci_rw.crw_count = end - start;
703 io->u.ci_ladvise.li_start = start;
704 io->u.ci_ladvise.li_end = end;
705 io->u.ci_ladvise.li_fid = parent->u.ci_ladvise.li_fid;
706 io->u.ci_ladvise.li_advice = parent->u.ci_ladvise.li_advice;
707 io->u.ci_ladvise.li_flags = parent->u.ci_ladvise.li_flags;
717 static loff_t lov_offset_mod(loff_t val, int delta)
719 if (val != OBD_OBJECT_EOF)
724 static int lov_io_iter_init(const struct lu_env *env,
725 const struct cl_io_slice *ios)
727 struct lov_io *lio = cl2lov_io(env, ios);
728 struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
729 struct lov_io_sub *sub;
730 struct lu_extent ext;
736 ext.e_start = lio->lis_pos;
737 ext.e_end = lio->lis_endpos;
739 lov_foreach_io_layout(index, lio, &ext) {
740 struct lov_layout_entry *le = lov_entry(lio->lis_object, index);
741 struct lov_layout_raid0 *r0 = &le->lle_raid0;
746 CDEBUG(D_VFSTRACE, "component[%d] flags %#x\n",
747 index, lsm->lsm_entries[index]->lsme_flags);
748 if (!lsm_entry_inited(lsm, index)) {
750 * Read from uninitialized components should return
756 if (!le->lle_valid && !ios->cis_io->ci_designated_mirror) {
757 CERROR("I/O to invalid component: %d, mirror: %d\n",
758 index, lio->lis_mirror_index);
762 for (stripe = 0; stripe < r0->lo_nr; stripe++) {
763 if (!lov_stripe_intersects(lsm, index, stripe,
767 if (unlikely(!r0->lo_sub[stripe])) {
768 if (ios->cis_io->ci_type == CIT_READ ||
769 ios->cis_io->ci_type == CIT_WRITE ||
770 ios->cis_io->ci_type == CIT_FAULT)
776 end = lov_offset_mod(end, 1);
777 sub = lov_sub_get(env, lio,
778 lov_comp_index(index, stripe));
784 lov_io_sub_inherit(sub, lio, start, end);
785 rc = cl_io_iter_init(sub->sub_env, &sub->sub_io);
787 cl_io_iter_fini(sub->sub_env, &sub->sub_io);
791 CDEBUG(D_VFSTRACE, "shrink: %d [%llu, %llu)\n",
794 list_add_tail(&sub->sub_linkage, &lio->lis_active);
802 static int lov_io_rw_iter_init(const struct lu_env *env,
803 const struct cl_io_slice *ios)
805 struct lov_io *lio = cl2lov_io(env, ios);
806 struct cl_io *io = ios->cis_io;
807 struct lov_stripe_md_entry *lse;
808 loff_t start = io->u.ci_rw.crw_pos;
812 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
815 if (cl_io_is_append(io))
816 RETURN(lov_io_iter_init(env, ios));
818 index = lov_io_layout_at(lio, io->u.ci_rw.crw_pos);
819 if (index < 0) { /* non-existing layout component */
820 if (io->ci_type == CIT_READ) {
822 * TODO: it needs to detect the next component and
823 * then set the next pos
827 RETURN(lov_io_iter_init(env, ios));
833 if (!lov_entry(lio->lis_object, index)->lle_valid &&
834 !io->ci_designated_mirror)
835 RETURN(io->ci_type == CIT_READ ? -EAGAIN : -EIO);
837 lse = lov_lse(lio->lis_object, index);
839 next = MAX_LFS_FILESIZE;
840 if (lse->lsme_stripe_count > 1) {
841 unsigned long ssize = lse->lsme_stripe_size;
843 lov_do_div64(start, ssize);
844 next = (start + 1) * ssize;
845 if (next <= start * ssize)
846 next = MAX_LFS_FILESIZE;
849 LASSERTF(io->u.ci_rw.crw_pos >= lse->lsme_extent.e_start,
850 "pos %lld, [%lld, %lld)\n", io->u.ci_rw.crw_pos,
851 lse->lsme_extent.e_start, lse->lsme_extent.e_end);
852 next = min_t(__u64, next, lse->lsme_extent.e_end);
853 next = min_t(loff_t, next, lio->lis_io_endpos);
855 io->ci_continue = next < lio->lis_io_endpos;
856 io->u.ci_rw.crw_count = next - io->u.ci_rw.crw_pos;
857 lio->lis_pos = io->u.ci_rw.crw_pos;
858 lio->lis_endpos = io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count;
860 "stripe: %llu chunk: [%llu, %llu) %llu, %zd\n",
861 (__u64)start, lio->lis_pos, lio->lis_endpos,
862 (__u64)lio->lis_io_endpos, io->u.ci_rw.crw_count);
865 * XXX The following call should be optimized: we know, that
866 * [lio->lis_pos, lio->lis_endpos) intersects with exactly one stripe.
868 RETURN(lov_io_iter_init(env, ios));
871 static int lov_io_setattr_iter_init(const struct lu_env *env,
872 const struct cl_io_slice *ios)
874 struct lov_io *lio = cl2lov_io(env, ios);
875 struct cl_io *io = ios->cis_io;
879 if (cl_io_is_trunc(io) && lio->lis_pos > 0) {
880 index = lov_io_layout_at(lio, lio->lis_pos - 1);
881 /* no entry found for such offset */
883 RETURN(io->ci_result = -ENODATA);
886 RETURN(lov_io_iter_init(env, ios));
889 static int lov_io_call(const struct lu_env *env, struct lov_io *lio,
890 int (*iofunc)(const struct lu_env *, struct cl_io *))
892 struct cl_io *parent = lio->lis_cl.cis_io;
893 struct lov_io_sub *sub;
897 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
898 rc = iofunc(sub->sub_env, &sub->sub_io);
902 if (parent->ci_result == 0)
903 parent->ci_result = sub->sub_io.ci_result;
908 static int lov_io_lock(const struct lu_env *env, const struct cl_io_slice *ios)
911 RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_lock));
914 static int lov_io_start(const struct lu_env *env, const struct cl_io_slice *ios)
917 RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_start));
920 static int lov_io_end_wrapper(const struct lu_env *env, struct cl_io *io)
924 * It's possible that lov_io_start() wasn't called against this
925 * sub-io, either because previous sub-io failed, or upper layer
928 if (io->ci_state == CIS_IO_GOING)
931 io->ci_state = CIS_IO_FINISHED;
935 static int lov_io_iter_fini_wrapper(const struct lu_env *env, struct cl_io *io)
937 cl_io_iter_fini(env, io);
941 static int lov_io_unlock_wrapper(const struct lu_env *env, struct cl_io *io)
943 cl_io_unlock(env, io);
947 static void lov_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
951 rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_end_wrapper);
956 lov_io_data_version_end(const struct lu_env *env, const struct cl_io_slice *ios)
958 struct lov_io *lio = cl2lov_io(env, ios);
959 struct cl_io *parent = lio->lis_cl.cis_io;
960 struct cl_data_version_io *pdv = &parent->u.ci_data_version;
961 struct lov_io_sub *sub;
964 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
965 struct cl_data_version_io *sdv = &sub->sub_io.u.ci_data_version;
967 lov_io_end_wrapper(sub->sub_env, &sub->sub_io);
969 pdv->dv_data_version += sdv->dv_data_version;
970 if (pdv->dv_layout_version > sdv->dv_layout_version)
971 pdv->dv_layout_version = sdv->dv_layout_version;
973 if (parent->ci_result == 0)
974 parent->ci_result = sub->sub_io.ci_result;
980 static void lov_io_iter_fini(const struct lu_env *env,
981 const struct cl_io_slice *ios)
983 struct lov_io *lio = cl2lov_io(env, ios);
987 rc = lov_io_call(env, lio, lov_io_iter_fini_wrapper);
989 while (!list_empty(&lio->lis_active))
990 list_del_init(lio->lis_active.next);
994 static void lov_io_unlock(const struct lu_env *env,
995 const struct cl_io_slice *ios)
1000 rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_unlock_wrapper);
1005 static int lov_io_read_ahead(const struct lu_env *env,
1006 const struct cl_io_slice *ios,
1007 pgoff_t start, struct cl_read_ahead *ra)
1009 struct lov_io *lio = cl2lov_io(env, ios);
1010 struct lov_object *loo = lio->lis_object;
1011 struct cl_object *obj = lov2cl(loo);
1012 struct lov_layout_raid0 *r0;
1013 struct lov_io_sub *sub;
1017 unsigned int pps; /* pages per stripe */
1023 offset = cl_offset(obj, start);
1024 index = lov_io_layout_at(lio, offset);
1025 if (index < 0 || !lsm_entry_inited(loo->lo_lsm, index))
1028 /* avoid readahead to expand to stale components */
1029 if (!lov_entry(loo, index)->lle_valid)
1032 stripe = lov_stripe_number(loo->lo_lsm, index, offset);
1034 r0 = lov_r0(loo, index);
1035 if (unlikely(!r0->lo_sub[stripe]))
1038 sub = lov_sub_get(env, lio, lov_comp_index(index, stripe));
1040 RETURN(PTR_ERR(sub));
1042 lov_stripe_offset(loo->lo_lsm, index, offset, stripe, &suboff);
1043 rc = cl_io_read_ahead(sub->sub_env, &sub->sub_io,
1044 cl_index(lovsub2cl(r0->lo_sub[stripe]), suboff),
1047 CDEBUG(D_READA, DFID " cra_end = %lu, stripes = %d, rc = %d\n",
1048 PFID(lu_object_fid(lov2lu(loo))), ra->cra_end_idx,
1054 * Adjust the stripe index by layout of comp. ra->cra_end is the
1055 * maximum page index covered by an underlying DLM lock.
1056 * This function converts cra_end from stripe level to file level, and
1057 * make sure it's not beyond stripe and component boundary.
1060 /* cra_end is stripe level, convert it into file level */
1061 ra_end = ra->cra_end_idx;
1062 if (ra_end != CL_PAGE_EOF)
1063 ra->cra_end_idx = lov_stripe_pgoff(loo->lo_lsm, index,
1066 /* boundary of current component */
1067 ra_end = cl_index(obj, (loff_t)lov_io_extent(lio, index)->e_end);
1068 if (ra_end != CL_PAGE_EOF && ra->cra_end_idx >= ra_end)
1069 ra->cra_end_idx = ra_end - 1;
1071 if (r0->lo_nr == 1) /* single stripe file */
1074 pps = lov_lse(loo, index)->lsme_stripe_size >> PAGE_SHIFT;
1076 CDEBUG(D_READA, DFID " max_index = %lu, pps = %u, index = %d, "
1077 "stripe_size = %u, stripe no = %u, start index = %lu\n",
1078 PFID(lu_object_fid(lov2lu(loo))), ra->cra_end_idx, pps, index,
1079 lov_lse(loo, index)->lsme_stripe_size, stripe, start);
1081 /* never exceed the end of the stripe */
1082 ra->cra_end_idx = min_t(pgoff_t, ra->cra_end_idx,
1083 start + pps - start % pps - 1);
1088 * lov implementation of cl_operations::cio_submit() method. It takes a list
1089 * of pages in \a queue, splits it into per-stripe sub-lists, invokes
1090 * cl_io_submit() on underlying devices to submit sub-lists, and then splices
1093 * Major complication of this function is a need to handle memory cleansing:
1094 * cl_io_submit() is called to write out pages as a part of VM memory
1095 * reclamation, and hence it may not fail due to memory shortages (system
1096 * dead-locks otherwise). To deal with this, some resources (sub-lists,
1097 * sub-environment, etc.) are allocated per-device on "startup" (i.e., in a
1098 * not-memory cleansing context), and in case of memory shortage, these
1099 * pre-allocated resources are used by lov_io_submit() under
1100 * lov_device::ld_mutex mutex.
1102 static int lov_io_submit(const struct lu_env *env,
1103 const struct cl_io_slice *ios,
1104 enum cl_req_type crt, struct cl_2queue *queue)
1106 struct cl_page_list *qin = &queue->c2_qin;
1107 struct lov_io *lio = cl2lov_io(env, ios);
1108 struct lov_io_sub *sub;
1109 struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
1110 struct cl_page *page;
1111 struct cl_page *tmp;
1116 cl_page_list_init(plist);
1117 while (qin->pl_nr > 0) {
1118 struct cl_2queue *cl2q = &lov_env_info(env)->lti_cl2q;
1120 page = cl_page_list_first(qin);
1121 if (lov_page_is_empty(page)) {
1122 cl_page_list_move(&queue->c2_qout, qin, page);
1125 * it could only be mirror read to get here therefore
1126 * the pages will be transient. We don't care about
1127 * the return code of cl_page_prep() at all.
1129 (void) cl_page_prep(env, ios->cis_io, page, crt);
1130 cl_page_completion(env, page, crt, 0);
1134 cl_2queue_init(cl2q);
1135 cl_page_list_move(&cl2q->c2_qin, qin, page);
1137 index = page->cp_lov_index;
1138 cl_page_list_for_each_safe(page, tmp, qin) {
1139 /* this page is not on this stripe */
1140 if (index != page->cp_lov_index)
1143 cl_page_list_move(&cl2q->c2_qin, qin, page);
1146 sub = lov_sub_get(env, lio, index);
1148 rc = cl_io_submit_rw(sub->sub_env, &sub->sub_io,
1154 cl_page_list_splice(&cl2q->c2_qin, plist);
1155 cl_page_list_splice(&cl2q->c2_qout, &queue->c2_qout);
1156 cl_2queue_fini(env, cl2q);
1162 cl_page_list_splice(plist, qin);
1163 cl_page_list_fini(env, plist);
1168 static int lov_io_commit_async(const struct lu_env *env,
1169 const struct cl_io_slice *ios,
1170 struct cl_page_list *queue, int from, int to,
1173 struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
1174 struct lov_io *lio = cl2lov_io(env, ios);
1175 struct lov_io_sub *sub;
1176 struct cl_page *page;
1180 if (lio->lis_nr_subios == 1) {
1181 int idx = lio->lis_single_subio_index;
1183 LASSERT(!lov_page_is_empty(cl_page_list_first(queue)));
1185 sub = lov_sub_get(env, lio, idx);
1186 LASSERT(!IS_ERR(sub));
1187 LASSERT(sub == &lio->lis_single_subio);
1188 rc = cl_io_commit_async(sub->sub_env, &sub->sub_io, queue,
1193 cl_page_list_init(plist);
1194 while (queue->pl_nr > 0) {
1198 LASSERT(plist->pl_nr == 0);
1199 page = cl_page_list_first(queue);
1200 LASSERT(!lov_page_is_empty(page));
1202 cl_page_list_move(plist, queue, page);
1204 index = page->cp_lov_index;
1205 while (queue->pl_nr > 0) {
1206 page = cl_page_list_first(queue);
1207 if (index != page->cp_lov_index)
1210 cl_page_list_move(plist, queue, page);
1213 if (queue->pl_nr > 0) /* still has more pages */
1214 stripe_to = PAGE_SIZE;
1216 sub = lov_sub_get(env, lio, index);
1218 rc = cl_io_commit_async(sub->sub_env, &sub->sub_io,
1219 plist, from, stripe_to, cb);
1225 if (plist->pl_nr > 0) /* short write */
1231 /* for error case, add the page back into the qin list */
1232 LASSERT(ergo(rc == 0, plist->pl_nr == 0));
1233 while (plist->pl_nr > 0) {
1234 /* error occurred, add the uncommitted pages back into queue */
1235 page = cl_page_list_last(plist);
1236 cl_page_list_move_head(queue, plist, page);
1242 static int lov_io_fault_start(const struct lu_env *env,
1243 const struct cl_io_slice *ios)
1245 struct cl_fault_io *fio;
1247 struct lov_io_sub *sub;
1251 fio = &ios->cis_io->u.ci_fault;
1252 lio = cl2lov_io(env, ios);
1253 sub = lov_sub_get(env, lio, fio->ft_page->cp_lov_index);
1254 sub->sub_io.u.ci_fault.ft_nob = fio->ft_nob;
1256 RETURN(lov_io_start(env, ios));
1259 static void lov_io_fsync_end(const struct lu_env *env,
1260 const struct cl_io_slice *ios)
1262 struct lov_io *lio = cl2lov_io(env, ios);
1263 struct lov_io_sub *sub;
1264 unsigned int *written = &ios->cis_io->u.ci_fsync.fi_nr_written;
1268 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
1269 struct cl_io *subio = &sub->sub_io;
1271 lov_io_end_wrapper(sub->sub_env, subio);
1273 if (subio->ci_result == 0)
1274 *written += subio->u.ci_fsync.fi_nr_written;
1279 static const struct cl_io_operations lov_io_ops = {
1282 .cio_fini = lov_io_fini,
1283 .cio_iter_init = lov_io_rw_iter_init,
1284 .cio_iter_fini = lov_io_iter_fini,
1285 .cio_lock = lov_io_lock,
1286 .cio_unlock = lov_io_unlock,
1287 .cio_start = lov_io_start,
1288 .cio_end = lov_io_end
1291 .cio_fini = lov_io_fini,
1292 .cio_iter_init = lov_io_rw_iter_init,
1293 .cio_iter_fini = lov_io_iter_fini,
1294 .cio_lock = lov_io_lock,
1295 .cio_unlock = lov_io_unlock,
1296 .cio_start = lov_io_start,
1297 .cio_end = lov_io_end
1300 .cio_fini = lov_io_fini,
1301 .cio_iter_init = lov_io_setattr_iter_init,
1302 .cio_iter_fini = lov_io_iter_fini,
1303 .cio_lock = lov_io_lock,
1304 .cio_unlock = lov_io_unlock,
1305 .cio_start = lov_io_start,
1306 .cio_end = lov_io_end
1308 [CIT_DATA_VERSION] = {
1309 .cio_fini = lov_io_fini,
1310 .cio_iter_init = lov_io_iter_init,
1311 .cio_iter_fini = lov_io_iter_fini,
1312 .cio_lock = lov_io_lock,
1313 .cio_unlock = lov_io_unlock,
1314 .cio_start = lov_io_start,
1315 .cio_end = lov_io_data_version_end,
1318 .cio_fini = lov_io_fini,
1319 .cio_iter_init = lov_io_iter_init,
1320 .cio_iter_fini = lov_io_iter_fini,
1321 .cio_lock = lov_io_lock,
1322 .cio_unlock = lov_io_unlock,
1323 .cio_start = lov_io_fault_start,
1324 .cio_end = lov_io_end
1327 .cio_fini = lov_io_fini,
1328 .cio_iter_init = lov_io_iter_init,
1329 .cio_iter_fini = lov_io_iter_fini,
1330 .cio_lock = lov_io_lock,
1331 .cio_unlock = lov_io_unlock,
1332 .cio_start = lov_io_start,
1333 .cio_end = lov_io_fsync_end
1336 .cio_fini = lov_io_fini,
1337 .cio_iter_init = lov_io_iter_init,
1338 .cio_iter_fini = lov_io_iter_fini,
1339 .cio_lock = lov_io_lock,
1340 .cio_unlock = lov_io_unlock,
1341 .cio_start = lov_io_start,
1342 .cio_end = lov_io_end
1345 .cio_fini = lov_io_fini,
1348 .cio_fini = lov_io_fini
1351 .cio_read_ahead = lov_io_read_ahead,
1352 .cio_submit = lov_io_submit,
1353 .cio_commit_async = lov_io_commit_async,
1356 /*****************************************************************************
1358 * Empty lov io operations.
1362 static void lov_empty_io_fini(const struct lu_env *env,
1363 const struct cl_io_slice *ios)
1365 struct lov_object *lov = cl2lov(ios->cis_obj);
1368 if (atomic_dec_and_test(&lov->lo_active_ios))
1369 wake_up_all(&lov->lo_waitq);
1373 static int lov_empty_io_submit(const struct lu_env *env,
1374 const struct cl_io_slice *ios,
1375 enum cl_req_type crt, struct cl_2queue *queue)
1380 static void lov_empty_impossible(const struct lu_env *env,
1381 struct cl_io_slice *ios)
1386 #define LOV_EMPTY_IMPOSSIBLE ((void *)lov_empty_impossible)
1389 * An io operation vector for files without stripes.
1391 static const struct cl_io_operations lov_empty_io_ops = {
1394 .cio_fini = lov_empty_io_fini,
1396 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1397 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1398 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1399 .cio_end = LOV_EMPTY_IMPOSSIBLE
1403 .cio_fini = lov_empty_io_fini,
1404 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1405 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1406 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1407 .cio_end = LOV_EMPTY_IMPOSSIBLE
1410 .cio_fini = lov_empty_io_fini,
1411 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1412 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1413 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1414 .cio_end = LOV_EMPTY_IMPOSSIBLE
1417 .cio_fini = lov_empty_io_fini,
1418 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1419 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1420 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1421 .cio_end = LOV_EMPTY_IMPOSSIBLE
1424 .cio_fini = lov_empty_io_fini
1427 .cio_fini = lov_empty_io_fini
1430 .cio_fini = lov_empty_io_fini
1433 .cio_fini = lov_empty_io_fini
1436 .cio_submit = lov_empty_io_submit,
1437 .cio_commit_async = LOV_EMPTY_IMPOSSIBLE
1440 int lov_io_init_composite(const struct lu_env *env, struct cl_object *obj,
1443 struct lov_io *lio = lov_env_io(env);
1444 struct lov_object *lov = cl2lov(obj);
1449 INIT_LIST_HEAD(&lio->lis_active);
1450 result = lov_io_slice_init(lio, lov, io);
1454 result = lov_io_subio_init(env, lio, io);
1456 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_io_ops);
1457 atomic_inc(&lov->lo_active_ios);
1461 io->ci_result = result < 0 ? result : 0;
1465 int lov_io_init_empty(const struct lu_env *env, struct cl_object *obj,
1468 struct lov_object *lov = cl2lov(obj);
1469 struct lov_io *lio = lov_env_io(env);
1473 lio->lis_object = lov;
1474 switch (io->ci_type) {
1485 case CIT_DATA_VERSION:
1493 CERROR("Page fault on a file without stripes: "DFID"\n",
1494 PFID(lu_object_fid(&obj->co_lu)));
1498 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
1499 atomic_inc(&lov->lo_active_ios);
1502 io->ci_result = result < 0 ? result : 0;
1506 int lov_io_init_released(const struct lu_env *env, struct cl_object *obj,
1509 struct lov_object *lov = cl2lov(obj);
1510 struct lov_io *lio = lov_env_io(env);
1514 LASSERT(lov->lo_lsm != NULL);
1515 lio->lis_object = lov;
1517 switch (io->ci_type) {
1519 LASSERTF(0, "invalid type %d\n", io->ci_type);
1520 result = -EOPNOTSUPP;
1526 case CIT_DATA_VERSION:
1531 * the truncate to 0 is managed by MDT:
1532 * - in open, for open O_TRUNC
1533 * - in setattr, for truncate
1535 /* the truncate is for size > 0 so triggers a restore */
1536 if (cl_io_is_trunc(io)) {
1537 io->ci_restore_needed = 1;
1545 io->ci_restore_needed = 1;
1551 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
1552 atomic_inc(&lov->lo_active_ios);
1555 io->ci_result = result < 0 ? result : 0;
1560 * Return the index in composite:lo_entries by the file offset
1562 int lov_io_layout_at(struct lov_io *lio, __u64 offset)
1564 struct lov_object *lov = lio->lis_object;
1565 struct lov_layout_composite *comp = &lov->u.composite;
1566 int start_index = 0;
1567 int end_index = comp->lo_entry_count - 1;
1570 LASSERT(lov->lo_type == LLT_COMP);
1572 /* This is actual file offset so nothing can cover eof. */
1573 if (offset == LUSTRE_EOF)
1576 if (lov_is_flr(lov)) {
1577 struct lov_mirror_entry *lre;
1579 LASSERT(lio->lis_mirror_index >= 0);
1581 lre = &comp->lo_mirrors[lio->lis_mirror_index];
1582 start_index = lre->lre_start;
1583 end_index = lre->lre_end;
1586 for (i = start_index; i <= end_index; i++) {
1587 struct lov_layout_entry *lle = lov_entry(lov, i);
1589 if ((offset >= lle->lle_extent->e_start &&
1590 offset < lle->lle_extent->e_end) ||
1591 (offset == OBD_OBJECT_EOF &&
1592 lle->lle_extent->e_end == OBD_OBJECT_EOF))