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++;
190 /*****************************************************************************
195 static int lov_io_subio_init(const struct lu_env *env, struct lov_io *lio,
200 LASSERT(lio->lis_object != NULL);
202 INIT_LIST_HEAD(&lio->lis_subios);
203 lio->lis_single_subio_index = -1;
204 lio->lis_nr_subios = 0;
210 * Decide if it will need write intent RPC
212 static int lov_io_mirror_write_intent(struct lov_io *lio,
213 struct lov_object *obj, struct cl_io *io)
215 struct lov_layout_composite *comp = &obj->u.composite;
216 struct lu_extent *ext = &io->ci_write_intent;
217 struct lov_mirror_entry *lre;
218 struct lov_mirror_entry *primary;
219 struct lov_layout_entry *lle;
223 *ext = (typeof(*ext)) { lio->lis_pos, lio->lis_endpos };
224 io->ci_need_write_intent = 0;
226 if (!(io->ci_type == CIT_WRITE || cl_io_is_trunc(io) ||
227 cl_io_is_mkwrite(io)))
231 * FLR: check if it needs to send a write intent RPC to server.
232 * Writing to sync_pending file needs write intent RPC to change
233 * the file state back to write_pending, so that the layout version
234 * can be increased when the state changes to sync_pending at a later
235 * time. Otherwise there exists a chance that an evicted client may
236 * dirty the file data while resync client is working on it.
237 * Designated I/O is allowed for resync workload.
239 if (lov_flr_state(obj) == LCM_FL_RDONLY ||
240 (lov_flr_state(obj) == LCM_FL_SYNC_PENDING &&
241 io->ci_designated_mirror == 0)) {
242 io->ci_need_write_intent = 1;
246 LASSERT((lov_flr_state(obj) == LCM_FL_WRITE_PENDING));
247 LASSERT(comp->lo_preferred_mirror >= 0);
250 * need to iterate all components to see if there are
251 * multiple components covering the writing component
253 primary = &comp->lo_mirrors[comp->lo_preferred_mirror];
254 LASSERT(!primary->lre_stale);
255 lov_foreach_mirror_layout_entry(obj, lle, primary) {
256 LASSERT(lle->lle_valid);
257 if (!lu_extent_is_overlapped(ext, lle->lle_extent))
260 ext->e_start = MIN(ext->e_start, lle->lle_extent->e_start);
261 ext->e_end = MAX(ext->e_end, lle->lle_extent->e_end);
265 CERROR(DFID ": cannot find any valid components covering "
266 "file extent "DEXT", mirror: %d\n",
267 PFID(lu_object_fid(lov2lu(obj))), PEXT(ext),
268 primary->lre_mirror_id);
273 lov_foreach_mirror_entry(obj, lre) {
277 lov_foreach_mirror_layout_entry(obj, lle, lre) {
281 if (lu_extent_is_overlapped(ext, lle->lle_extent)) {
288 CDEBUG(D_VFSTRACE, DFID "there are %zd components to be staled to "
289 "modify file extent "DEXT", iot: %d\n",
290 PFID(lu_object_fid(lov2lu(obj))), count, PEXT(ext), io->ci_type);
292 io->ci_need_write_intent = count > 0;
297 static int lov_io_mirror_init(struct lov_io *lio, struct lov_object *obj,
300 struct lov_layout_composite *comp = &obj->u.composite;
306 if (!lov_is_flr(obj)) {
307 LASSERT(comp->lo_preferred_mirror == 0);
308 lio->lis_mirror_index = comp->lo_preferred_mirror;
313 /* transfer the layout version for verification */
314 if (io->ci_layout_version == 0)
315 io->ci_layout_version = obj->lo_lsm->lsm_layout_gen;
317 /* find the corresponding mirror for designated mirror IO */
318 if (io->ci_designated_mirror > 0) {
319 struct lov_mirror_entry *entry;
321 LASSERT(!io->ci_ndelay);
323 CDEBUG(D_LAYOUT, "designated I/O mirror state: %d\n",
326 if ((cl_io_is_trunc(io) || io->ci_type == CIT_WRITE) &&
327 (io->ci_layout_version != obj->lo_lsm->lsm_layout_gen)) {
329 * For resync I/O, the ci_layout_version was the layout
330 * version when resync starts. If it doesn't match the
331 * current object layout version, it means the layout
337 io->ci_layout_version |= LU_LAYOUT_RESYNC;
340 lio->lis_mirror_index = -1;
341 lov_foreach_mirror_entry(obj, entry) {
342 if (entry->lre_mirror_id ==
343 io->ci_designated_mirror) {
344 lio->lis_mirror_index = index;
351 RETURN(lio->lis_mirror_index < 0 ? -EINVAL : 0);
354 result = lov_io_mirror_write_intent(lio, obj, io);
358 if (io->ci_need_write_intent) {
359 CDEBUG(D_VFSTRACE, DFID " need write intent for [%llu, %llu)\n",
360 PFID(lu_object_fid(lov2lu(obj))),
361 lio->lis_pos, lio->lis_endpos);
363 if (cl_io_is_trunc(io)) {
365 * for truncate, we uses [size, EOF) to judge whether
366 * a write intent needs to be send, but we need to
367 * restore the write extent to [0, size).
369 io->ci_write_intent.e_start = 0;
370 io->ci_write_intent.e_end =
371 io->u.ci_setattr.sa_attr.lvb_size;
373 /* stop cl_io_init() loop */
377 if (io->ci_ndelay_tried == 0 || /* first time to try */
378 /* reset the mirror index if layout has changed */
379 lio->lis_mirror_layout_gen != obj->lo_lsm->lsm_layout_gen) {
380 lio->lis_mirror_layout_gen = obj->lo_lsm->lsm_layout_gen;
381 index = lio->lis_mirror_index = comp->lo_preferred_mirror;
383 index = lio->lis_mirror_index;
386 /* move mirror index to the next one */
387 index = (index + 1) % comp->lo_mirror_count;
390 for (i = 0; i < comp->lo_mirror_count; i++) {
391 struct lu_extent ext = { .e_start = lio->lis_pos,
392 .e_end = lio->lis_pos + 1 };
393 struct lov_mirror_entry *lre;
394 struct lov_layout_entry *lle;
397 lre = &comp->lo_mirrors[(index + i) % comp->lo_mirror_count];
401 lov_foreach_mirror_layout_entry(obj, lle, lre) {
405 if (lu_extent_is_overlapped(&ext, lle->lle_extent)) {
409 } /* each component of the mirror */
411 index = (index + i) % comp->lo_mirror_count;
416 if (i == comp->lo_mirror_count) {
417 CERROR(DFID": failed to find a component covering "
418 "I/O region at %llu\n",
419 PFID(lu_object_fid(lov2lu(obj))), lio->lis_pos);
421 dump_lsm(D_ERROR, obj->lo_lsm);
426 CDEBUG(D_VFSTRACE, DFID ": flr state: %d, move mirror from %d to %d, "
427 "have retried: %d, mirror count: %d\n",
428 PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj),
429 lio->lis_mirror_index, index, io->ci_ndelay_tried,
430 comp->lo_mirror_count);
432 lio->lis_mirror_index = index;
435 * FLR: if all mirrors have been tried once, most likely the network
436 * of this client has been partitioned. We should relinquish CPU for
437 * a while before trying again.
439 if (io->ci_ndelay && io->ci_ndelay_tried > 0 &&
440 (io->ci_ndelay_tried % comp->lo_mirror_count == 0)) {
441 set_current_state(TASK_INTERRUPTIBLE);
442 schedule_timeout(cfs_time_seconds(1) / 100); /* 10ms */
443 if (signal_pending(current))
447 * we'd set ci_tried_all_mirrors to turn off fast mirror
448 * switching for read after we've tried all mirrors several
451 io->ci_tried_all_mirrors = io->ci_ndelay_tried %
452 (comp->lo_mirror_count * 4) == 0;
454 ++io->ci_ndelay_tried;
456 CDEBUG(D_VFSTRACE, "use %sdelayed RPC state for this IO\n",
457 io->ci_ndelay ? "non-" : "");
462 static int lov_io_slice_init(struct lov_io *lio,
463 struct lov_object *obj, struct cl_io *io)
470 lio->lis_object = obj;
472 LASSERT(obj->lo_lsm != NULL);
474 switch (io->ci_type) {
477 lio->lis_pos = io->u.ci_rw.crw_pos;
478 lio->lis_endpos = io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count;
479 lio->lis_io_endpos = lio->lis_endpos;
480 if (cl_io_is_append(io)) {
481 LASSERT(io->ci_type == CIT_WRITE);
484 * If there is LOV EA hole, then we may cannot locate
485 * the current file-tail exactly.
487 if (unlikely(obj->lo_lsm->lsm_entries[0]->lsme_pattern &
489 GOTO(out, result = -EIO);
492 lio->lis_endpos = OBD_OBJECT_EOF;
497 if (cl_io_is_trunc(io))
498 lio->lis_pos = io->u.ci_setattr.sa_attr.lvb_size;
501 lio->lis_endpos = OBD_OBJECT_EOF;
504 case CIT_DATA_VERSION:
506 lio->lis_endpos = OBD_OBJECT_EOF;
510 pgoff_t index = io->u.ci_fault.ft_index;
512 lio->lis_pos = cl_offset(io->ci_obj, index);
513 lio->lis_endpos = cl_offset(io->ci_obj, index + 1);
518 lio->lis_pos = io->u.ci_fsync.fi_start;
519 lio->lis_endpos = io->u.ci_fsync.fi_end;
524 lio->lis_pos = io->u.ci_ladvise.li_start;
525 lio->lis_endpos = io->u.ci_ladvise.li_end;
531 lio->lis_endpos = OBD_OBJECT_EOF;
533 if (lov_flr_state(obj) == LCM_FL_RDONLY &&
534 !OBD_FAIL_CHECK(OBD_FAIL_FLR_GLIMPSE_IMMUTABLE))
535 /* SoM is accurate, no need glimpse */
536 GOTO(out, result = 1);
541 lio->lis_endpos = OBD_OBJECT_EOF;
548 result = lov_io_mirror_init(lio, obj, io);
552 /* check if it needs to instantiate layout */
553 if (!(io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io) ||
554 (cl_io_is_trunc(io) && io->u.ci_setattr.sa_attr.lvb_size > 0)))
555 GOTO(out, result = 0);
558 * for truncate, it only needs to instantiate the components
559 * before the truncated size.
561 if (cl_io_is_trunc(io)) {
562 io->ci_write_intent.e_start = 0;
563 /* for writes, e_end is endpos, the location of the file
564 * pointer after the write is completed, so it is not accessed.
565 * For truncate, 'end' is the size, and *is* acccessed.
566 * In other words, writes are [start, end), but truncate is
567 * [start, size], where both are included. So add 1 to the
568 * size when creating the write intent to account for this.
570 io->ci_write_intent.e_end =
571 io->u.ci_setattr.sa_attr.lvb_size + 1;
573 io->ci_write_intent.e_start = lio->lis_pos;
574 io->ci_write_intent.e_end = lio->lis_endpos;
578 lov_foreach_io_layout(index, lio, &io->ci_write_intent) {
579 if (!lsm_entry_inited(obj->lo_lsm, index)) {
580 io->ci_need_write_intent = 1;
585 if (io->ci_need_write_intent && io->ci_designated_mirror > 0) {
587 * REINT_SYNC RPC has already tried to instantiate all of the
588 * components involved, obviously it didn't succeed. Skip this
589 * mirror for now. The server won't be able to figure out
590 * which mirror it should instantiate components
592 CERROR(DFID": trying to instantiate components for designated "
593 "I/O, file state: %d\n",
594 PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj));
596 io->ci_need_write_intent = 0;
597 GOTO(out, result = -EIO);
600 if (io->ci_need_write_intent)
601 GOTO(out, result = 1);
609 static void lov_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
611 struct lov_io *lio = cl2lov_io(env, ios);
612 struct lov_object *lov = cl2lov(ios->cis_obj);
616 LASSERT(list_empty(&lio->lis_active));
618 while (!list_empty(&lio->lis_subios)) {
619 struct lov_io_sub *sub = list_entry(lio->lis_subios.next,
623 list_del_init(&sub->sub_list);
624 lio->lis_nr_subios--;
626 lov_io_sub_fini(env, lio, sub);
627 lov_sub_free(lio, sub);
629 LASSERT(lio->lis_nr_subios == 0);
631 LASSERT(atomic_read(&lov->lo_active_ios) > 0);
632 if (atomic_dec_and_test(&lov->lo_active_ios))
633 wake_up_all(&lov->lo_waitq);
637 static void lov_io_sub_inherit(struct lov_io_sub *sub, struct lov_io *lio,
638 loff_t start, loff_t end)
640 struct cl_io *io = &sub->sub_io;
641 struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
642 struct cl_io *parent = lio->lis_cl.cis_io;
643 int index = lov_comp_entry(sub->sub_subio_index);
644 int stripe = lov_comp_stripe(sub->sub_subio_index);
646 switch (io->ci_type) {
648 io->u.ci_setattr.sa_attr = parent->u.ci_setattr.sa_attr;
649 io->u.ci_setattr.sa_attr_flags =
650 parent->u.ci_setattr.sa_attr_flags;
651 io->u.ci_setattr.sa_avalid = parent->u.ci_setattr.sa_avalid;
652 io->u.ci_setattr.sa_xvalid = parent->u.ci_setattr.sa_xvalid;
653 io->u.ci_setattr.sa_stripe_index = stripe;
654 io->u.ci_setattr.sa_parent_fid =
655 parent->u.ci_setattr.sa_parent_fid;
656 if (cl_io_is_trunc(io)) {
657 loff_t new_size = parent->u.ci_setattr.sa_attr.lvb_size;
659 new_size = lov_size_to_stripe(lsm, index, new_size,
661 io->u.ci_setattr.sa_attr.lvb_size = new_size;
663 lov_lsm2layout(lsm, lsm->lsm_entries[index],
664 &io->u.ci_setattr.sa_layout);
667 case CIT_DATA_VERSION: {
668 io->u.ci_data_version.dv_data_version = 0;
669 io->u.ci_data_version.dv_flags =
670 parent->u.ci_data_version.dv_flags;
674 struct cl_object *obj = parent->ci_obj;
675 loff_t off = cl_offset(obj, parent->u.ci_fault.ft_index);
677 io->u.ci_fault = parent->u.ci_fault;
678 off = lov_size_to_stripe(lsm, index, off, stripe);
679 io->u.ci_fault.ft_index = cl_index(obj, off);
683 io->u.ci_fsync.fi_start = start;
684 io->u.ci_fsync.fi_end = end;
685 io->u.ci_fsync.fi_fid = parent->u.ci_fsync.fi_fid;
686 io->u.ci_fsync.fi_mode = parent->u.ci_fsync.fi_mode;
691 io->u.ci_wr.wr_sync = cl_io_is_sync_write(parent);
692 io->ci_tried_all_mirrors = parent->ci_tried_all_mirrors;
693 if (cl_io_is_append(parent)) {
694 io->u.ci_wr.wr_append = 1;
696 io->u.ci_rw.crw_pos = start;
697 io->u.ci_rw.crw_count = end - start;
702 io->u.ci_ladvise.li_start = start;
703 io->u.ci_ladvise.li_end = end;
704 io->u.ci_ladvise.li_fid = parent->u.ci_ladvise.li_fid;
705 io->u.ci_ladvise.li_advice = parent->u.ci_ladvise.li_advice;
706 io->u.ci_ladvise.li_flags = parent->u.ci_ladvise.li_flags;
716 static loff_t lov_offset_mod(loff_t val, int delta)
718 if (val != OBD_OBJECT_EOF)
723 static int lov_io_iter_init(const struct lu_env *env,
724 const struct cl_io_slice *ios)
726 struct lov_io *lio = cl2lov_io(env, ios);
727 struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
728 struct lov_io_sub *sub;
729 struct lu_extent ext;
735 ext.e_start = lio->lis_pos;
736 ext.e_end = lio->lis_endpos;
738 lov_foreach_io_layout(index, lio, &ext) {
739 struct lov_layout_entry *le = lov_entry(lio->lis_object, index);
740 struct lov_layout_raid0 *r0 = &le->lle_raid0;
745 CDEBUG(D_VFSTRACE, "component[%d] flags %#x\n",
746 index, lsm->lsm_entries[index]->lsme_flags);
747 if (!lsm_entry_inited(lsm, index)) {
749 * Read from uninitialized components should return
755 if (!le->lle_valid && !ios->cis_io->ci_designated_mirror) {
756 CERROR("I/O to invalid component: %d, mirror: %d\n",
757 index, lio->lis_mirror_index);
761 for (stripe = 0; stripe < r0->lo_nr; stripe++) {
762 if (!lov_stripe_intersects(lsm, index, stripe,
766 if (unlikely(!r0->lo_sub[stripe])) {
767 if (ios->cis_io->ci_type == CIT_READ ||
768 ios->cis_io->ci_type == CIT_WRITE ||
769 ios->cis_io->ci_type == CIT_FAULT)
775 end = lov_offset_mod(end, 1);
776 sub = lov_sub_get(env, lio,
777 lov_comp_index(index, stripe));
783 lov_io_sub_inherit(sub, lio, start, end);
784 rc = cl_io_iter_init(sub->sub_env, &sub->sub_io);
786 cl_io_iter_fini(sub->sub_env, &sub->sub_io);
790 CDEBUG(D_VFSTRACE, "shrink: %d [%llu, %llu)\n",
793 list_add_tail(&sub->sub_linkage, &lio->lis_active);
801 static int lov_io_rw_iter_init(const struct lu_env *env,
802 const struct cl_io_slice *ios)
804 struct lov_io *lio = cl2lov_io(env, ios);
805 struct cl_io *io = ios->cis_io;
806 struct lov_stripe_md_entry *lse;
807 loff_t start = io->u.ci_rw.crw_pos;
811 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
814 if (cl_io_is_append(io))
815 RETURN(lov_io_iter_init(env, ios));
817 index = lov_io_layout_at(lio, io->u.ci_rw.crw_pos);
818 if (index < 0) { /* non-existing layout component */
819 if (io->ci_type == CIT_READ) {
821 * TODO: it needs to detect the next component and
822 * then set the next pos
826 RETURN(lov_io_iter_init(env, ios));
832 if (!lov_entry(lio->lis_object, index)->lle_valid &&
833 !io->ci_designated_mirror)
834 RETURN(io->ci_type == CIT_READ ? -EAGAIN : -EIO);
836 lse = lov_lse(lio->lis_object, index);
838 next = MAX_LFS_FILESIZE;
839 if (lse->lsme_stripe_count > 1) {
840 unsigned long ssize = lse->lsme_stripe_size;
842 lov_do_div64(start, ssize);
843 next = (start + 1) * ssize;
844 if (next <= start * ssize)
845 next = MAX_LFS_FILESIZE;
848 LASSERTF(io->u.ci_rw.crw_pos >= lse->lsme_extent.e_start,
849 "pos %lld, [%lld, %lld)\n", io->u.ci_rw.crw_pos,
850 lse->lsme_extent.e_start, lse->lsme_extent.e_end);
851 next = min_t(__u64, next, lse->lsme_extent.e_end);
852 next = min_t(loff_t, next, lio->lis_io_endpos);
854 io->ci_continue = next < lio->lis_io_endpos;
855 io->u.ci_rw.crw_count = next - io->u.ci_rw.crw_pos;
856 lio->lis_pos = io->u.ci_rw.crw_pos;
857 lio->lis_endpos = io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count;
859 "stripe: %llu chunk: [%llu, %llu) %llu, %zd\n",
860 (__u64)start, lio->lis_pos, lio->lis_endpos,
861 (__u64)lio->lis_io_endpos, io->u.ci_rw.crw_count);
864 * XXX The following call should be optimized: we know, that
865 * [lio->lis_pos, lio->lis_endpos) intersects with exactly one stripe.
867 RETURN(lov_io_iter_init(env, ios));
870 static int lov_io_setattr_iter_init(const struct lu_env *env,
871 const struct cl_io_slice *ios)
873 struct lov_io *lio = cl2lov_io(env, ios);
874 struct cl_io *io = ios->cis_io;
878 if (cl_io_is_trunc(io) && lio->lis_pos > 0) {
879 index = lov_io_layout_at(lio, lio->lis_pos - 1);
880 /* no entry found for such offset */
882 RETURN(io->ci_result = -ENODATA);
885 RETURN(lov_io_iter_init(env, ios));
888 static int lov_io_call(const struct lu_env *env, struct lov_io *lio,
889 int (*iofunc)(const struct lu_env *, struct cl_io *))
891 struct cl_io *parent = lio->lis_cl.cis_io;
892 struct lov_io_sub *sub;
896 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
897 rc = iofunc(sub->sub_env, &sub->sub_io);
901 if (parent->ci_result == 0)
902 parent->ci_result = sub->sub_io.ci_result;
907 static int lov_io_lock(const struct lu_env *env, const struct cl_io_slice *ios)
910 RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_lock));
913 static int lov_io_start(const struct lu_env *env, const struct cl_io_slice *ios)
916 RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_start));
919 static int lov_io_end_wrapper(const struct lu_env *env, struct cl_io *io)
923 * It's possible that lov_io_start() wasn't called against this
924 * sub-io, either because previous sub-io failed, or upper layer
927 if (io->ci_state == CIS_IO_GOING)
930 io->ci_state = CIS_IO_FINISHED;
934 static int lov_io_iter_fini_wrapper(const struct lu_env *env, struct cl_io *io)
936 cl_io_iter_fini(env, io);
940 static int lov_io_unlock_wrapper(const struct lu_env *env, struct cl_io *io)
942 cl_io_unlock(env, io);
946 static void lov_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
950 rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_end_wrapper);
955 lov_io_data_version_end(const struct lu_env *env, const struct cl_io_slice *ios)
957 struct lov_io *lio = cl2lov_io(env, ios);
958 struct cl_io *parent = lio->lis_cl.cis_io;
959 struct cl_data_version_io *pdv = &parent->u.ci_data_version;
960 struct lov_io_sub *sub;
963 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
964 struct cl_data_version_io *sdv = &sub->sub_io.u.ci_data_version;
966 lov_io_end_wrapper(sub->sub_env, &sub->sub_io);
968 pdv->dv_data_version += sdv->dv_data_version;
969 if (pdv->dv_layout_version > sdv->dv_layout_version)
970 pdv->dv_layout_version = sdv->dv_layout_version;
972 if (parent->ci_result == 0)
973 parent->ci_result = sub->sub_io.ci_result;
979 static void lov_io_iter_fini(const struct lu_env *env,
980 const struct cl_io_slice *ios)
982 struct lov_io *lio = cl2lov_io(env, ios);
986 rc = lov_io_call(env, lio, lov_io_iter_fini_wrapper);
988 while (!list_empty(&lio->lis_active))
989 list_del_init(lio->lis_active.next);
993 static void lov_io_unlock(const struct lu_env *env,
994 const struct cl_io_slice *ios)
999 rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_unlock_wrapper);
1004 static int lov_io_read_ahead(const struct lu_env *env,
1005 const struct cl_io_slice *ios,
1006 pgoff_t start, struct cl_read_ahead *ra)
1008 struct lov_io *lio = cl2lov_io(env, ios);
1009 struct lov_object *loo = lio->lis_object;
1010 struct cl_object *obj = lov2cl(loo);
1011 struct lov_layout_raid0 *r0;
1012 struct lov_io_sub *sub;
1016 unsigned int pps; /* pages per stripe */
1022 offset = cl_offset(obj, start);
1023 index = lov_io_layout_at(lio, offset);
1024 if (index < 0 || !lsm_entry_inited(loo->lo_lsm, index))
1027 /* avoid readahead to expand to stale components */
1028 if (!lov_entry(loo, index)->lle_valid)
1031 stripe = lov_stripe_number(loo->lo_lsm, index, offset);
1033 r0 = lov_r0(loo, index);
1034 if (unlikely(!r0->lo_sub[stripe]))
1037 sub = lov_sub_get(env, lio, lov_comp_index(index, stripe));
1039 RETURN(PTR_ERR(sub));
1041 lov_stripe_offset(loo->lo_lsm, index, offset, stripe, &suboff);
1042 rc = cl_io_read_ahead(sub->sub_env, &sub->sub_io,
1043 cl_index(lovsub2cl(r0->lo_sub[stripe]), suboff),
1046 CDEBUG(D_READA, DFID " cra_end = %lu, stripes = %d, rc = %d\n",
1047 PFID(lu_object_fid(lov2lu(loo))), ra->cra_end, r0->lo_nr, rc);
1052 * Adjust the stripe index by layout of comp. ra->cra_end is the
1053 * maximum page index covered by an underlying DLM lock.
1054 * This function converts cra_end from stripe level to file level, and
1055 * make sure it's not beyond stripe and component boundary.
1058 /* cra_end is stripe level, convert it into file level */
1059 ra_end = ra->cra_end;
1060 if (ra_end != CL_PAGE_EOF)
1061 ra->cra_end = lov_stripe_pgoff(loo->lo_lsm, index,
1064 /* boundary of current component */
1065 ra_end = cl_index(obj, (loff_t)lov_io_extent(lio, index)->e_end);
1066 if (ra_end != CL_PAGE_EOF && ra->cra_end >= ra_end)
1067 ra->cra_end = ra_end - 1;
1069 if (r0->lo_nr == 1) /* single stripe file */
1072 pps = lov_lse(loo, index)->lsme_stripe_size >> PAGE_SHIFT;
1074 CDEBUG(D_READA, DFID " max_index = %lu, pps = %u, index = %u, "
1075 "stripe_size = %u, stripe no = %u, start index = %lu\n",
1076 PFID(lu_object_fid(lov2lu(loo))), ra->cra_end, pps, index,
1077 lov_lse(loo, index)->lsme_stripe_size, stripe, start);
1079 /* never exceed the end of the stripe */
1080 ra->cra_end = min_t(pgoff_t,
1081 ra->cra_end, start + pps - start % pps - 1);
1086 * lov implementation of cl_operations::cio_submit() method. It takes a list
1087 * of pages in \a queue, splits it into per-stripe sub-lists, invokes
1088 * cl_io_submit() on underlying devices to submit sub-lists, and then splices
1091 * Major complication of this function is a need to handle memory cleansing:
1092 * cl_io_submit() is called to write out pages as a part of VM memory
1093 * reclamation, and hence it may not fail due to memory shortages (system
1094 * dead-locks otherwise). To deal with this, some resources (sub-lists,
1095 * sub-environment, etc.) are allocated per-device on "startup" (i.e., in a
1096 * not-memory cleansing context), and in case of memory shortage, these
1097 * pre-allocated resources are used by lov_io_submit() under
1098 * lov_device::ld_mutex mutex.
1100 static int lov_io_submit(const struct lu_env *env,
1101 const struct cl_io_slice *ios,
1102 enum cl_req_type crt, struct cl_2queue *queue)
1104 struct cl_page_list *qin = &queue->c2_qin;
1105 struct lov_io *lio = cl2lov_io(env, ios);
1106 struct lov_io_sub *sub;
1107 struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
1108 struct cl_page *page;
1109 struct cl_page *tmp;
1114 cl_page_list_init(plist);
1115 while (qin->pl_nr > 0) {
1116 struct cl_2queue *cl2q = &lov_env_info(env)->lti_cl2q;
1118 page = cl_page_list_first(qin);
1119 if (lov_page_is_empty(page)) {
1120 cl_page_list_move(&queue->c2_qout, qin, page);
1123 * it could only be mirror read to get here therefore
1124 * the pages will be transient. We don't care about
1125 * the return code of cl_page_prep() at all.
1127 (void) cl_page_prep(env, ios->cis_io, page, crt);
1128 cl_page_completion(env, page, crt, 0);
1132 cl_2queue_init(cl2q);
1133 cl_page_list_move(&cl2q->c2_qin, qin, page);
1135 index = page->cp_lov_index;
1136 cl_page_list_for_each_safe(page, tmp, qin) {
1137 /* this page is not on this stripe */
1138 if (index != page->cp_lov_index)
1141 cl_page_list_move(&cl2q->c2_qin, qin, page);
1144 sub = lov_sub_get(env, lio, index);
1146 rc = cl_io_submit_rw(sub->sub_env, &sub->sub_io,
1152 cl_page_list_splice(&cl2q->c2_qin, plist);
1153 cl_page_list_splice(&cl2q->c2_qout, &queue->c2_qout);
1154 cl_2queue_fini(env, cl2q);
1160 cl_page_list_splice(plist, qin);
1161 cl_page_list_fini(env, plist);
1166 static int lov_io_commit_async(const struct lu_env *env,
1167 const struct cl_io_slice *ios,
1168 struct cl_page_list *queue, int from, int to,
1171 struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
1172 struct lov_io *lio = cl2lov_io(env, ios);
1173 struct lov_io_sub *sub;
1174 struct cl_page *page;
1178 if (lio->lis_nr_subios == 1) {
1179 int idx = lio->lis_single_subio_index;
1181 LASSERT(!lov_page_is_empty(cl_page_list_first(queue)));
1183 sub = lov_sub_get(env, lio, idx);
1184 LASSERT(!IS_ERR(sub));
1185 LASSERT(sub == &lio->lis_single_subio);
1186 rc = cl_io_commit_async(sub->sub_env, &sub->sub_io, queue,
1191 cl_page_list_init(plist);
1192 while (queue->pl_nr > 0) {
1196 LASSERT(plist->pl_nr == 0);
1197 page = cl_page_list_first(queue);
1198 LASSERT(!lov_page_is_empty(page));
1200 cl_page_list_move(plist, queue, page);
1202 index = page->cp_lov_index;
1203 while (queue->pl_nr > 0) {
1204 page = cl_page_list_first(queue);
1205 if (index != page->cp_lov_index)
1208 cl_page_list_move(plist, queue, page);
1211 if (queue->pl_nr > 0) /* still has more pages */
1212 stripe_to = PAGE_SIZE;
1214 sub = lov_sub_get(env, lio, index);
1216 rc = cl_io_commit_async(sub->sub_env, &sub->sub_io,
1217 plist, from, stripe_to, cb);
1223 if (plist->pl_nr > 0) /* short write */
1229 /* for error case, add the page back into the qin list */
1230 LASSERT(ergo(rc == 0, plist->pl_nr == 0));
1231 while (plist->pl_nr > 0) {
1232 /* error occurred, add the uncommitted pages back into queue */
1233 page = cl_page_list_last(plist);
1234 cl_page_list_move_head(queue, plist, page);
1240 static int lov_io_fault_start(const struct lu_env *env,
1241 const struct cl_io_slice *ios)
1243 struct cl_fault_io *fio;
1245 struct lov_io_sub *sub;
1249 fio = &ios->cis_io->u.ci_fault;
1250 lio = cl2lov_io(env, ios);
1251 sub = lov_sub_get(env, lio, fio->ft_page->cp_lov_index);
1252 sub->sub_io.u.ci_fault.ft_nob = fio->ft_nob;
1254 RETURN(lov_io_start(env, ios));
1257 static void lov_io_fsync_end(const struct lu_env *env,
1258 const struct cl_io_slice *ios)
1260 struct lov_io *lio = cl2lov_io(env, ios);
1261 struct lov_io_sub *sub;
1262 unsigned int *written = &ios->cis_io->u.ci_fsync.fi_nr_written;
1266 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
1267 struct cl_io *subio = &sub->sub_io;
1269 lov_io_end_wrapper(sub->sub_env, subio);
1271 if (subio->ci_result == 0)
1272 *written += subio->u.ci_fsync.fi_nr_written;
1277 static const struct cl_io_operations lov_io_ops = {
1280 .cio_fini = lov_io_fini,
1281 .cio_iter_init = lov_io_rw_iter_init,
1282 .cio_iter_fini = lov_io_iter_fini,
1283 .cio_lock = lov_io_lock,
1284 .cio_unlock = lov_io_unlock,
1285 .cio_start = lov_io_start,
1286 .cio_end = lov_io_end
1289 .cio_fini = lov_io_fini,
1290 .cio_iter_init = lov_io_rw_iter_init,
1291 .cio_iter_fini = lov_io_iter_fini,
1292 .cio_lock = lov_io_lock,
1293 .cio_unlock = lov_io_unlock,
1294 .cio_start = lov_io_start,
1295 .cio_end = lov_io_end
1298 .cio_fini = lov_io_fini,
1299 .cio_iter_init = lov_io_setattr_iter_init,
1300 .cio_iter_fini = lov_io_iter_fini,
1301 .cio_lock = lov_io_lock,
1302 .cio_unlock = lov_io_unlock,
1303 .cio_start = lov_io_start,
1304 .cio_end = lov_io_end
1306 [CIT_DATA_VERSION] = {
1307 .cio_fini = lov_io_fini,
1308 .cio_iter_init = lov_io_iter_init,
1309 .cio_iter_fini = lov_io_iter_fini,
1310 .cio_lock = lov_io_lock,
1311 .cio_unlock = lov_io_unlock,
1312 .cio_start = lov_io_start,
1313 .cio_end = lov_io_data_version_end,
1316 .cio_fini = lov_io_fini,
1317 .cio_iter_init = lov_io_iter_init,
1318 .cio_iter_fini = lov_io_iter_fini,
1319 .cio_lock = lov_io_lock,
1320 .cio_unlock = lov_io_unlock,
1321 .cio_start = lov_io_fault_start,
1322 .cio_end = lov_io_end
1325 .cio_fini = lov_io_fini,
1326 .cio_iter_init = lov_io_iter_init,
1327 .cio_iter_fini = lov_io_iter_fini,
1328 .cio_lock = lov_io_lock,
1329 .cio_unlock = lov_io_unlock,
1330 .cio_start = lov_io_start,
1331 .cio_end = lov_io_fsync_end
1334 .cio_fini = lov_io_fini,
1335 .cio_iter_init = lov_io_iter_init,
1336 .cio_iter_fini = lov_io_iter_fini,
1337 .cio_lock = lov_io_lock,
1338 .cio_unlock = lov_io_unlock,
1339 .cio_start = lov_io_start,
1340 .cio_end = lov_io_end
1343 .cio_fini = lov_io_fini,
1346 .cio_fini = lov_io_fini
1349 .cio_read_ahead = lov_io_read_ahead,
1350 .cio_submit = lov_io_submit,
1351 .cio_commit_async = lov_io_commit_async,
1354 /*****************************************************************************
1356 * Empty lov io operations.
1360 static void lov_empty_io_fini(const struct lu_env *env,
1361 const struct cl_io_slice *ios)
1363 struct lov_object *lov = cl2lov(ios->cis_obj);
1366 if (atomic_dec_and_test(&lov->lo_active_ios))
1367 wake_up_all(&lov->lo_waitq);
1371 static int lov_empty_io_submit(const struct lu_env *env,
1372 const struct cl_io_slice *ios,
1373 enum cl_req_type crt, struct cl_2queue *queue)
1378 static void lov_empty_impossible(const struct lu_env *env,
1379 struct cl_io_slice *ios)
1384 #define LOV_EMPTY_IMPOSSIBLE ((void *)lov_empty_impossible)
1387 * An io operation vector for files without stripes.
1389 static const struct cl_io_operations lov_empty_io_ops = {
1392 .cio_fini = lov_empty_io_fini,
1394 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1395 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1396 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1397 .cio_end = LOV_EMPTY_IMPOSSIBLE
1401 .cio_fini = lov_empty_io_fini,
1402 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1403 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1404 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1405 .cio_end = LOV_EMPTY_IMPOSSIBLE
1408 .cio_fini = lov_empty_io_fini,
1409 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1410 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1411 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1412 .cio_end = LOV_EMPTY_IMPOSSIBLE
1415 .cio_fini = lov_empty_io_fini,
1416 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1417 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1418 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1419 .cio_end = LOV_EMPTY_IMPOSSIBLE
1422 .cio_fini = lov_empty_io_fini
1425 .cio_fini = lov_empty_io_fini
1428 .cio_fini = lov_empty_io_fini
1431 .cio_fini = lov_empty_io_fini
1434 .cio_submit = lov_empty_io_submit,
1435 .cio_commit_async = LOV_EMPTY_IMPOSSIBLE
1438 int lov_io_init_composite(const struct lu_env *env, struct cl_object *obj,
1441 struct lov_io *lio = lov_env_io(env);
1442 struct lov_object *lov = cl2lov(obj);
1447 INIT_LIST_HEAD(&lio->lis_active);
1448 result = lov_io_slice_init(lio, lov, io);
1452 result = lov_io_subio_init(env, lio, io);
1454 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_io_ops);
1455 atomic_inc(&lov->lo_active_ios);
1459 io->ci_result = result < 0 ? result : 0;
1463 int lov_io_init_empty(const struct lu_env *env, struct cl_object *obj,
1466 struct lov_object *lov = cl2lov(obj);
1467 struct lov_io *lio = lov_env_io(env);
1471 lio->lis_object = lov;
1472 switch (io->ci_type) {
1483 case CIT_DATA_VERSION:
1491 CERROR("Page fault on a file without stripes: "DFID"\n",
1492 PFID(lu_object_fid(&obj->co_lu)));
1496 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
1497 atomic_inc(&lov->lo_active_ios);
1500 io->ci_result = result < 0 ? result : 0;
1504 int lov_io_init_released(const struct lu_env *env, struct cl_object *obj,
1507 struct lov_object *lov = cl2lov(obj);
1508 struct lov_io *lio = lov_env_io(env);
1512 LASSERT(lov->lo_lsm != NULL);
1513 lio->lis_object = lov;
1515 switch (io->ci_type) {
1517 LASSERTF(0, "invalid type %d\n", io->ci_type);
1518 result = -EOPNOTSUPP;
1524 case CIT_DATA_VERSION:
1529 * the truncate to 0 is managed by MDT:
1530 * - in open, for open O_TRUNC
1531 * - in setattr, for truncate
1533 /* the truncate is for size > 0 so triggers a restore */
1534 if (cl_io_is_trunc(io)) {
1535 io->ci_restore_needed = 1;
1543 io->ci_restore_needed = 1;
1549 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
1550 atomic_inc(&lov->lo_active_ios);
1553 io->ci_result = result < 0 ? result : 0;
1558 * Return the index in composite:lo_entries by the file offset
1560 int lov_io_layout_at(struct lov_io *lio, __u64 offset)
1562 struct lov_object *lov = lio->lis_object;
1563 struct lov_layout_composite *comp = &lov->u.composite;
1564 int start_index = 0;
1565 int end_index = comp->lo_entry_count - 1;
1568 LASSERT(lov->lo_type == LLT_COMP);
1570 /* This is actual file offset so nothing can cover eof. */
1571 if (offset == LUSTRE_EOF)
1574 if (lov_is_flr(lov)) {
1575 struct lov_mirror_entry *lre;
1577 LASSERT(lio->lis_mirror_index >= 0);
1579 lre = &comp->lo_mirrors[lio->lis_mirror_index];
1580 start_index = lre->lre_start;
1581 end_index = lre->lre_end;
1584 for (i = start_index; i <= end_index; i++) {
1585 struct lov_layout_entry *lle = lov_entry(lov, i);
1587 if ((offset >= lle->lle_extent->e_start &&
1588 offset < lle->lle_extent->e_end) ||
1589 (offset == OBD_OBJECT_EOF &&
1590 lle->lle_extent->e_end == OBD_OBJECT_EOF))