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;
146 result = cl_io_sub_init(sub->sub_env, sub_io, io->ci_type, sub_obj);
149 lov_io_sub_fini(env, lio, sub);
154 struct lov_io_sub *lov_sub_get(const struct lu_env *env,
155 struct lov_io *lio, int index)
157 struct lov_io_sub *sub;
162 list_for_each_entry(sub, &lio->lis_subios, sub_list) {
163 if (sub->sub_subio_index == index) {
170 sub = lov_sub_alloc(lio, index);
172 GOTO(out, rc = -ENOMEM);
174 rc = lov_io_sub_init(env, lio, sub);
176 lov_sub_free(lio, sub);
180 list_add_tail(&sub->sub_list, &lio->lis_subios);
181 lio->lis_nr_subios++;
189 /*****************************************************************************
194 static int lov_io_subio_init(const struct lu_env *env, struct lov_io *lio,
199 LASSERT(lio->lis_object != NULL);
201 INIT_LIST_HEAD(&lio->lis_subios);
202 lio->lis_single_subio_index = -1;
203 lio->lis_nr_subios = 0;
209 * Decide if it will need write intent RPC
211 static int lov_io_mirror_write_intent(struct lov_io *lio,
212 struct lov_object *obj, struct cl_io *io)
214 struct lov_layout_composite *comp = &obj->u.composite;
215 struct lu_extent *ext = &io->ci_write_intent;
216 struct lov_mirror_entry *lre;
217 struct lov_mirror_entry *primary;
218 struct lov_layout_entry *lle;
222 *ext = (typeof(*ext)) { lio->lis_pos, lio->lis_endpos };
223 io->ci_need_write_intent = 0;
225 if (!(io->ci_type == CIT_WRITE || cl_io_is_trunc(io) ||
226 cl_io_is_mkwrite(io)))
230 * FLR: check if it needs to send a write intent RPC to server.
231 * Writing to sync_pending file needs write intent RPC to change
232 * the file state back to write_pending, so that the layout version
233 * can be increased when the state changes to sync_pending at a later
234 * time. Otherwise there exists a chance that an evicted client may
235 * dirty the file data while resync client is working on it.
236 * Designated I/O is allowed for resync workload.
238 if (lov_flr_state(obj) == LCM_FL_RDONLY ||
239 (lov_flr_state(obj) == LCM_FL_SYNC_PENDING &&
240 io->ci_designated_mirror == 0)) {
241 io->ci_need_write_intent = 1;
245 LASSERT((lov_flr_state(obj) == LCM_FL_WRITE_PENDING));
246 LASSERT(comp->lo_preferred_mirror >= 0);
249 * need to iterate all components to see if there are
250 * multiple components covering the writing component
252 primary = &comp->lo_mirrors[comp->lo_preferred_mirror];
253 LASSERT(!primary->lre_stale);
254 lov_foreach_mirror_layout_entry(obj, lle, primary) {
255 LASSERT(lle->lle_valid);
256 if (!lu_extent_is_overlapped(ext, lle->lle_extent))
259 ext->e_start = MIN(ext->e_start, lle->lle_extent->e_start);
260 ext->e_end = MAX(ext->e_end, lle->lle_extent->e_end);
264 CERROR(DFID ": cannot find any valid components covering "
265 "file extent "DEXT", mirror: %d\n",
266 PFID(lu_object_fid(lov2lu(obj))), PEXT(ext),
267 primary->lre_mirror_id);
272 lov_foreach_mirror_entry(obj, lre) {
276 lov_foreach_mirror_layout_entry(obj, lle, lre) {
280 if (lu_extent_is_overlapped(ext, lle->lle_extent)) {
287 CDEBUG(D_VFSTRACE, DFID "there are %zd components to be staled to "
288 "modify file extent "DEXT", iot: %d\n",
289 PFID(lu_object_fid(lov2lu(obj))), count, PEXT(ext), io->ci_type);
291 io->ci_need_write_intent = count > 0;
296 static int lov_io_mirror_init(struct lov_io *lio, struct lov_object *obj,
299 struct lov_layout_composite *comp = &obj->u.composite;
305 if (!lov_is_flr(obj)) {
306 LASSERT(comp->lo_preferred_mirror == 0);
307 lio->lis_mirror_index = comp->lo_preferred_mirror;
312 /* transfer the layout version for verification */
313 if (io->ci_layout_version == 0)
314 io->ci_layout_version = obj->lo_lsm->lsm_layout_gen;
316 /* find the corresponding mirror for designated mirror IO */
317 if (io->ci_designated_mirror > 0) {
318 struct lov_mirror_entry *entry;
320 LASSERT(!io->ci_ndelay);
322 CDEBUG(D_LAYOUT, "designated I/O mirror state: %d\n",
325 if ((cl_io_is_trunc(io) || io->ci_type == CIT_WRITE) &&
326 (io->ci_layout_version != obj->lo_lsm->lsm_layout_gen)) {
328 * For resync I/O, the ci_layout_version was the layout
329 * version when resync starts. If it doesn't match the
330 * current object layout version, it means the layout
336 io->ci_layout_version |= LU_LAYOUT_RESYNC;
339 lio->lis_mirror_index = -1;
340 lov_foreach_mirror_entry(obj, entry) {
341 if (entry->lre_mirror_id ==
342 io->ci_designated_mirror) {
343 lio->lis_mirror_index = index;
350 RETURN(lio->lis_mirror_index < 0 ? -EINVAL : 0);
353 result = lov_io_mirror_write_intent(lio, obj, io);
357 if (io->ci_need_write_intent) {
358 CDEBUG(D_VFSTRACE, DFID " need write intent for [%llu, %llu)\n",
359 PFID(lu_object_fid(lov2lu(obj))),
360 lio->lis_pos, lio->lis_endpos);
362 if (cl_io_is_trunc(io)) {
364 * for truncate, we uses [size, EOF) to judge whether
365 * a write intent needs to be send, but we need to
366 * restore the write extent to [0, size).
368 io->ci_write_intent.e_start = 0;
369 io->ci_write_intent.e_end =
370 io->u.ci_setattr.sa_attr.lvb_size;
372 /* stop cl_io_init() loop */
376 if (io->ci_ndelay_tried == 0 || /* first time to try */
377 /* reset the mirror index if layout has changed */
378 lio->lis_mirror_layout_gen != obj->lo_lsm->lsm_layout_gen) {
379 lio->lis_mirror_layout_gen = obj->lo_lsm->lsm_layout_gen;
380 index = lio->lis_mirror_index = comp->lo_preferred_mirror;
382 index = lio->lis_mirror_index;
385 /* move mirror index to the next one */
386 index = (index + 1) % comp->lo_mirror_count;
389 for (i = 0; i < comp->lo_mirror_count; i++) {
390 struct lu_extent ext = { .e_start = lio->lis_pos,
391 .e_end = lio->lis_pos + 1 };
392 struct lov_mirror_entry *lre;
393 struct lov_layout_entry *lle;
396 lre = &comp->lo_mirrors[(index + i) % comp->lo_mirror_count];
400 lov_foreach_mirror_layout_entry(obj, lle, lre) {
404 if (lu_extent_is_overlapped(&ext, lle->lle_extent)) {
411 index = (index + i) % comp->lo_mirror_count;
415 if (i == comp->lo_mirror_count) {
416 CERROR(DFID": failed to find a component covering "
417 "I/O region at %llu\n",
418 PFID(lu_object_fid(lov2lu(obj))), lio->lis_pos);
420 dump_lsm(D_ERROR, obj->lo_lsm);
425 CDEBUG(D_VFSTRACE, DFID ": flr state: %d, move mirror from %d to %d, "
426 "have retried: %d, mirror count: %d\n",
427 PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj),
428 lio->lis_mirror_index, index, io->ci_ndelay_tried,
429 comp->lo_mirror_count);
431 lio->lis_mirror_index = index;
434 * FLR: if all mirrors have been tried once, most likely the network
435 * of this client has been partitioned. We should relinquish CPU for
436 * a while before trying again.
438 ++io->ci_ndelay_tried;
439 if (io->ci_ndelay && io->ci_ndelay_tried >= comp->lo_mirror_count) {
440 set_current_state(TASK_INTERRUPTIBLE);
441 schedule_timeout(cfs_time_seconds(1)); /* 10ms */
442 if (signal_pending(current))
445 /* reset retry counter */
446 io->ci_ndelay_tried = 1;
449 CDEBUG(D_VFSTRACE, "use %sdelayed RPC state for this IO\n",
450 io->ci_ndelay ? "non-" : "");
455 static int lov_io_slice_init(struct lov_io *lio,
456 struct lov_object *obj, struct cl_io *io)
463 lio->lis_object = obj;
465 LASSERT(obj->lo_lsm != NULL);
467 switch (io->ci_type) {
470 lio->lis_pos = io->u.ci_rw.crw_pos;
471 lio->lis_endpos = io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count;
472 lio->lis_io_endpos = lio->lis_endpos;
473 if (cl_io_is_append(io)) {
474 LASSERT(io->ci_type == CIT_WRITE);
477 * If there is LOV EA hole, then we may cannot locate
478 * the current file-tail exactly.
480 if (unlikely(obj->lo_lsm->lsm_entries[0]->lsme_pattern &
482 GOTO(out, result = -EIO);
485 lio->lis_endpos = OBD_OBJECT_EOF;
490 if (cl_io_is_trunc(io))
491 lio->lis_pos = io->u.ci_setattr.sa_attr.lvb_size;
494 lio->lis_endpos = OBD_OBJECT_EOF;
497 case CIT_DATA_VERSION:
499 lio->lis_endpos = OBD_OBJECT_EOF;
503 pgoff_t index = io->u.ci_fault.ft_index;
505 lio->lis_pos = cl_offset(io->ci_obj, index);
506 lio->lis_endpos = cl_offset(io->ci_obj, index + 1);
511 lio->lis_pos = io->u.ci_fsync.fi_start;
512 lio->lis_endpos = io->u.ci_fsync.fi_end;
517 lio->lis_pos = io->u.ci_ladvise.li_start;
518 lio->lis_endpos = io->u.ci_ladvise.li_end;
524 lio->lis_endpos = OBD_OBJECT_EOF;
526 if (lov_flr_state(obj) == LCM_FL_RDONLY &&
527 !OBD_FAIL_CHECK(OBD_FAIL_FLR_GLIMPSE_IMMUTABLE))
528 /* SoM is accurate, no need glimpse */
529 GOTO(out, result = 1);
534 lio->lis_endpos = OBD_OBJECT_EOF;
541 result = lov_io_mirror_init(lio, obj, io);
545 /* check if it needs to instantiate layout */
546 if (!(io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io) ||
547 (cl_io_is_trunc(io) && io->u.ci_setattr.sa_attr.lvb_size > 0)))
548 GOTO(out, result = 0);
551 * for truncate, it only needs to instantiate the components
552 * before the truncated size.
554 if (cl_io_is_trunc(io)) {
555 io->ci_write_intent.e_start = 0;
556 /* for writes, e_end is endpos, the location of the file
557 * pointer after the write is completed, so it is not accessed.
558 * For truncate, 'end' is the size, and *is* acccessed.
559 * In other words, writes are [start, end), but truncate is
560 * [start, size], where both are included. So add 1 to the
561 * size when creating the write intent to account for this.
563 io->ci_write_intent.e_end =
564 io->u.ci_setattr.sa_attr.lvb_size + 1;
566 io->ci_write_intent.e_start = lio->lis_pos;
567 io->ci_write_intent.e_end = lio->lis_endpos;
571 lov_foreach_io_layout(index, lio, &io->ci_write_intent) {
572 if (!lsm_entry_inited(obj->lo_lsm, index)) {
573 io->ci_need_write_intent = 1;
578 if (io->ci_need_write_intent && io->ci_designated_mirror > 0) {
580 * REINT_SYNC RPC has already tried to instantiate all of the
581 * components involved, obviously it didn't succeed. Skip this
582 * mirror for now. The server won't be able to figure out
583 * which mirror it should instantiate components
585 CERROR(DFID": trying to instantiate components for designated "
586 "I/O, file state: %d\n",
587 PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj));
589 io->ci_need_write_intent = 0;
590 GOTO(out, result = -EIO);
593 if (io->ci_need_write_intent)
594 GOTO(out, result = 1);
602 static void lov_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
604 struct lov_io *lio = cl2lov_io(env, ios);
605 struct lov_object *lov = cl2lov(ios->cis_obj);
609 LASSERT(list_empty(&lio->lis_active));
611 while (!list_empty(&lio->lis_subios)) {
612 struct lov_io_sub *sub = list_entry(lio->lis_subios.next,
616 list_del_init(&sub->sub_list);
617 lio->lis_nr_subios--;
619 lov_io_sub_fini(env, lio, sub);
620 lov_sub_free(lio, sub);
622 LASSERT(lio->lis_nr_subios == 0);
624 LASSERT(atomic_read(&lov->lo_active_ios) > 0);
625 if (atomic_dec_and_test(&lov->lo_active_ios))
626 wake_up_all(&lov->lo_waitq);
630 static void lov_io_sub_inherit(struct lov_io_sub *sub, struct lov_io *lio,
631 loff_t start, loff_t end)
633 struct cl_io *io = &sub->sub_io;
634 struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
635 struct cl_io *parent = lio->lis_cl.cis_io;
636 int index = lov_comp_entry(sub->sub_subio_index);
637 int stripe = lov_comp_stripe(sub->sub_subio_index);
639 switch (io->ci_type) {
641 io->u.ci_setattr.sa_attr = parent->u.ci_setattr.sa_attr;
642 io->u.ci_setattr.sa_attr_flags =
643 parent->u.ci_setattr.sa_attr_flags;
644 io->u.ci_setattr.sa_avalid = parent->u.ci_setattr.sa_avalid;
645 io->u.ci_setattr.sa_xvalid = parent->u.ci_setattr.sa_xvalid;
646 io->u.ci_setattr.sa_stripe_index = stripe;
647 io->u.ci_setattr.sa_parent_fid =
648 parent->u.ci_setattr.sa_parent_fid;
649 if (cl_io_is_trunc(io)) {
650 loff_t new_size = parent->u.ci_setattr.sa_attr.lvb_size;
652 new_size = lov_size_to_stripe(lsm, index, new_size,
654 io->u.ci_setattr.sa_attr.lvb_size = new_size;
656 lov_lsm2layout(lsm, lsm->lsm_entries[index],
657 &io->u.ci_setattr.sa_layout);
660 case CIT_DATA_VERSION: {
661 io->u.ci_data_version.dv_data_version = 0;
662 io->u.ci_data_version.dv_flags =
663 parent->u.ci_data_version.dv_flags;
667 struct cl_object *obj = parent->ci_obj;
668 loff_t off = cl_offset(obj, parent->u.ci_fault.ft_index);
670 io->u.ci_fault = parent->u.ci_fault;
671 off = lov_size_to_stripe(lsm, index, off, stripe);
672 io->u.ci_fault.ft_index = cl_index(obj, off);
676 io->u.ci_fsync.fi_start = start;
677 io->u.ci_fsync.fi_end = end;
678 io->u.ci_fsync.fi_fid = parent->u.ci_fsync.fi_fid;
679 io->u.ci_fsync.fi_mode = parent->u.ci_fsync.fi_mode;
684 io->u.ci_wr.wr_sync = cl_io_is_sync_write(parent);
685 if (cl_io_is_append(parent)) {
686 io->u.ci_wr.wr_append = 1;
688 io->u.ci_rw.crw_pos = start;
689 io->u.ci_rw.crw_count = end - start;
694 io->u.ci_ladvise.li_start = start;
695 io->u.ci_ladvise.li_end = end;
696 io->u.ci_ladvise.li_fid = parent->u.ci_ladvise.li_fid;
697 io->u.ci_ladvise.li_advice = parent->u.ci_ladvise.li_advice;
698 io->u.ci_ladvise.li_flags = parent->u.ci_ladvise.li_flags;
708 static loff_t lov_offset_mod(loff_t val, int delta)
710 if (val != OBD_OBJECT_EOF)
715 static int lov_io_iter_init(const struct lu_env *env,
716 const struct cl_io_slice *ios)
718 struct lov_io *lio = cl2lov_io(env, ios);
719 struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
720 struct lov_io_sub *sub;
721 struct lu_extent ext;
727 ext.e_start = lio->lis_pos;
728 ext.e_end = lio->lis_endpos;
730 lov_foreach_io_layout(index, lio, &ext) {
731 struct lov_layout_entry *le = lov_entry(lio->lis_object, index);
732 struct lov_layout_raid0 *r0 = &le->lle_raid0;
737 CDEBUG(D_VFSTRACE, "component[%d] flags %#x\n",
738 index, lsm->lsm_entries[index]->lsme_flags);
739 if (!lsm_entry_inited(lsm, index)) {
741 * Read from uninitialized components should return
747 if (!le->lle_valid && !ios->cis_io->ci_designated_mirror) {
748 CERROR("I/O to invalid component: %d, mirror: %d\n",
749 index, lio->lis_mirror_index);
753 for (stripe = 0; stripe < r0->lo_nr; stripe++) {
754 if (!lov_stripe_intersects(lsm, index, stripe,
758 if (unlikely(!r0->lo_sub[stripe])) {
759 if (ios->cis_io->ci_type == CIT_READ ||
760 ios->cis_io->ci_type == CIT_WRITE ||
761 ios->cis_io->ci_type == CIT_FAULT)
767 end = lov_offset_mod(end, 1);
768 sub = lov_sub_get(env, lio,
769 lov_comp_index(index, stripe));
775 lov_io_sub_inherit(sub, lio, start, end);
776 rc = cl_io_iter_init(sub->sub_env, &sub->sub_io);
778 cl_io_iter_fini(sub->sub_env, &sub->sub_io);
782 CDEBUG(D_VFSTRACE, "shrink: %d [%llu, %llu)\n",
785 list_add_tail(&sub->sub_linkage, &lio->lis_active);
793 static int lov_io_rw_iter_init(const struct lu_env *env,
794 const struct cl_io_slice *ios)
796 struct lov_io *lio = cl2lov_io(env, ios);
797 struct cl_io *io = ios->cis_io;
798 struct lov_stripe_md_entry *lse;
799 loff_t start = io->u.ci_rw.crw_pos;
803 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
806 if (cl_io_is_append(io))
807 RETURN(lov_io_iter_init(env, ios));
809 index = lov_io_layout_at(lio, io->u.ci_rw.crw_pos);
810 if (index < 0) { /* non-existing layout component */
811 if (io->ci_type == CIT_READ) {
813 * TODO: it needs to detect the next component and
814 * then set the next pos
818 RETURN(lov_io_iter_init(env, ios));
824 if (!lov_entry(lio->lis_object, index)->lle_valid &&
825 !io->ci_designated_mirror)
826 RETURN(io->ci_type == CIT_READ ? -EAGAIN : -EIO);
828 lse = lov_lse(lio->lis_object, index);
830 next = MAX_LFS_FILESIZE;
831 if (lse->lsme_stripe_count > 1) {
832 unsigned long ssize = lse->lsme_stripe_size;
834 lov_do_div64(start, ssize);
835 next = (start + 1) * ssize;
836 if (next <= start * ssize)
837 next = MAX_LFS_FILESIZE;
840 LASSERTF(io->u.ci_rw.crw_pos >= lse->lsme_extent.e_start,
841 "pos %lld, [%lld, %lld)\n", io->u.ci_rw.crw_pos,
842 lse->lsme_extent.e_start, lse->lsme_extent.e_end);
843 next = min_t(__u64, next, lse->lsme_extent.e_end);
844 next = min_t(loff_t, next, lio->lis_io_endpos);
846 io->ci_continue = next < lio->lis_io_endpos;
847 io->u.ci_rw.crw_count = next - io->u.ci_rw.crw_pos;
848 lio->lis_pos = io->u.ci_rw.crw_pos;
849 lio->lis_endpos = io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count;
851 "stripe: %llu chunk: [%llu, %llu) %llu, %zd\n",
852 (__u64)start, lio->lis_pos, lio->lis_endpos,
853 (__u64)lio->lis_io_endpos, io->u.ci_rw.crw_count);
856 * XXX The following call should be optimized: we know, that
857 * [lio->lis_pos, lio->lis_endpos) intersects with exactly one stripe.
859 RETURN(lov_io_iter_init(env, ios));
862 static int lov_io_setattr_iter_init(const struct lu_env *env,
863 const struct cl_io_slice *ios)
865 struct lov_io *lio = cl2lov_io(env, ios);
866 struct cl_io *io = ios->cis_io;
870 if (cl_io_is_trunc(io) && lio->lis_pos > 0) {
871 index = lov_io_layout_at(lio, lio->lis_pos - 1);
872 /* no entry found for such offset */
874 RETURN(io->ci_result = -ENODATA);
877 RETURN(lov_io_iter_init(env, ios));
880 static int lov_io_call(const struct lu_env *env, struct lov_io *lio,
881 int (*iofunc)(const struct lu_env *, struct cl_io *))
883 struct cl_io *parent = lio->lis_cl.cis_io;
884 struct lov_io_sub *sub;
888 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
889 rc = iofunc(sub->sub_env, &sub->sub_io);
893 if (parent->ci_result == 0)
894 parent->ci_result = sub->sub_io.ci_result;
899 static int lov_io_lock(const struct lu_env *env, const struct cl_io_slice *ios)
902 RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_lock));
905 static int lov_io_start(const struct lu_env *env, const struct cl_io_slice *ios)
908 RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_start));
911 static int lov_io_end_wrapper(const struct lu_env *env, struct cl_io *io)
915 * It's possible that lov_io_start() wasn't called against this
916 * sub-io, either because previous sub-io failed, or upper layer
919 if (io->ci_state == CIS_IO_GOING)
922 io->ci_state = CIS_IO_FINISHED;
926 static int lov_io_iter_fini_wrapper(const struct lu_env *env, struct cl_io *io)
928 cl_io_iter_fini(env, io);
932 static int lov_io_unlock_wrapper(const struct lu_env *env, struct cl_io *io)
934 cl_io_unlock(env, io);
938 static void lov_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
942 rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_end_wrapper);
947 lov_io_data_version_end(const struct lu_env *env, const struct cl_io_slice *ios)
949 struct lov_io *lio = cl2lov_io(env, ios);
950 struct cl_io *parent = lio->lis_cl.cis_io;
951 struct cl_data_version_io *pdv = &parent->u.ci_data_version;
952 struct lov_io_sub *sub;
955 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
956 struct cl_data_version_io *sdv = &sub->sub_io.u.ci_data_version;
958 lov_io_end_wrapper(sub->sub_env, &sub->sub_io);
960 pdv->dv_data_version += sdv->dv_data_version;
961 if (pdv->dv_layout_version > sdv->dv_layout_version)
962 pdv->dv_layout_version = sdv->dv_layout_version;
964 if (parent->ci_result == 0)
965 parent->ci_result = sub->sub_io.ci_result;
971 static void lov_io_iter_fini(const struct lu_env *env,
972 const struct cl_io_slice *ios)
974 struct lov_io *lio = cl2lov_io(env, ios);
978 rc = lov_io_call(env, lio, lov_io_iter_fini_wrapper);
980 while (!list_empty(&lio->lis_active))
981 list_del_init(lio->lis_active.next);
985 static void lov_io_unlock(const struct lu_env *env,
986 const struct cl_io_slice *ios)
991 rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_unlock_wrapper);
996 static int lov_io_read_ahead(const struct lu_env *env,
997 const struct cl_io_slice *ios,
998 pgoff_t start, struct cl_read_ahead *ra)
1000 struct lov_io *lio = cl2lov_io(env, ios);
1001 struct lov_object *loo = lio->lis_object;
1002 struct cl_object *obj = lov2cl(loo);
1003 struct lov_layout_raid0 *r0;
1004 struct lov_io_sub *sub;
1008 unsigned int pps; /* pages per stripe */
1014 offset = cl_offset(obj, start);
1015 index = lov_io_layout_at(lio, offset);
1016 if (index < 0 || !lsm_entry_inited(loo->lo_lsm, index))
1019 /* avoid readahead to expand to stale components */
1020 if (!lov_entry(loo, index)->lle_valid)
1023 stripe = lov_stripe_number(loo->lo_lsm, index, offset);
1025 r0 = lov_r0(loo, index);
1026 if (unlikely(!r0->lo_sub[stripe]))
1029 sub = lov_sub_get(env, lio, lov_comp_index(index, stripe));
1031 RETURN(PTR_ERR(sub));
1033 lov_stripe_offset(loo->lo_lsm, index, offset, stripe, &suboff);
1034 rc = cl_io_read_ahead(sub->sub_env, &sub->sub_io,
1035 cl_index(lovsub2cl(r0->lo_sub[stripe]), suboff),
1038 CDEBUG(D_READA, DFID " cra_end = %lu, stripes = %d, rc = %d\n",
1039 PFID(lu_object_fid(lov2lu(loo))), ra->cra_end, r0->lo_nr, rc);
1044 * Adjust the stripe index by layout of comp. ra->cra_end is the
1045 * maximum page index covered by an underlying DLM lock.
1046 * This function converts cra_end from stripe level to file level, and
1047 * make sure it's not beyond stripe and component boundary.
1050 /* cra_end is stripe level, convert it into file level */
1051 ra_end = ra->cra_end;
1052 if (ra_end != CL_PAGE_EOF)
1053 ra->cra_end = lov_stripe_pgoff(loo->lo_lsm, index,
1056 /* boundary of current component */
1057 ra_end = cl_index(obj, (loff_t)lov_io_extent(lio, index)->e_end);
1058 if (ra_end != CL_PAGE_EOF && ra->cra_end >= ra_end)
1059 ra->cra_end = ra_end - 1;
1061 if (r0->lo_nr == 1) /* single stripe file */
1064 pps = lov_lse(loo, index)->lsme_stripe_size >> PAGE_SHIFT;
1066 CDEBUG(D_READA, DFID " max_index = %lu, pps = %u, index = %u, "
1067 "stripe_size = %u, stripe no = %u, start index = %lu\n",
1068 PFID(lu_object_fid(lov2lu(loo))), ra->cra_end, pps, index,
1069 lov_lse(loo, index)->lsme_stripe_size, stripe, start);
1071 /* never exceed the end of the stripe */
1072 ra->cra_end = min_t(pgoff_t,
1073 ra->cra_end, start + pps - start % pps - 1);
1078 * lov implementation of cl_operations::cio_submit() method. It takes a list
1079 * of pages in \a queue, splits it into per-stripe sub-lists, invokes
1080 * cl_io_submit() on underlying devices to submit sub-lists, and then splices
1083 * Major complication of this function is a need to handle memory cleansing:
1084 * cl_io_submit() is called to write out pages as a part of VM memory
1085 * reclamation, and hence it may not fail due to memory shortages (system
1086 * dead-locks otherwise). To deal with this, some resources (sub-lists,
1087 * sub-environment, etc.) are allocated per-device on "startup" (i.e., in a
1088 * not-memory cleansing context), and in case of memory shortage, these
1089 * pre-allocated resources are used by lov_io_submit() under
1090 * lov_device::ld_mutex mutex.
1092 static int lov_io_submit(const struct lu_env *env,
1093 const struct cl_io_slice *ios,
1094 enum cl_req_type crt, struct cl_2queue *queue)
1096 struct cl_page_list *qin = &queue->c2_qin;
1097 struct lov_io *lio = cl2lov_io(env, ios);
1098 struct lov_io_sub *sub;
1099 struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
1100 struct cl_page *page;
1101 struct cl_page *tmp;
1106 cl_page_list_init(plist);
1107 while (qin->pl_nr > 0) {
1108 struct cl_2queue *cl2q = &lov_env_info(env)->lti_cl2q;
1110 page = cl_page_list_first(qin);
1111 if (lov_page_is_empty(page)) {
1112 cl_page_list_move(&queue->c2_qout, qin, page);
1115 * it could only be mirror read to get here therefore
1116 * the pages will be transient. We don't care about
1117 * the return code of cl_page_prep() at all.
1119 (void) cl_page_prep(env, ios->cis_io, page, crt);
1120 cl_page_completion(env, page, crt, 0);
1124 cl_2queue_init(cl2q);
1125 cl_page_list_move(&cl2q->c2_qin, qin, page);
1127 index = page->cp_lov_index;
1128 cl_page_list_for_each_safe(page, tmp, qin) {
1129 /* this page is not on this stripe */
1130 if (index != page->cp_lov_index)
1133 cl_page_list_move(&cl2q->c2_qin, qin, page);
1136 sub = lov_sub_get(env, lio, index);
1138 rc = cl_io_submit_rw(sub->sub_env, &sub->sub_io,
1144 cl_page_list_splice(&cl2q->c2_qin, plist);
1145 cl_page_list_splice(&cl2q->c2_qout, &queue->c2_qout);
1146 cl_2queue_fini(env, cl2q);
1152 cl_page_list_splice(plist, qin);
1153 cl_page_list_fini(env, plist);
1158 static int lov_io_commit_async(const struct lu_env *env,
1159 const struct cl_io_slice *ios,
1160 struct cl_page_list *queue, int from, int to,
1163 struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
1164 struct lov_io *lio = cl2lov_io(env, ios);
1165 struct lov_io_sub *sub;
1166 struct cl_page *page;
1170 if (lio->lis_nr_subios == 1) {
1171 int idx = lio->lis_single_subio_index;
1173 LASSERT(!lov_page_is_empty(cl_page_list_first(queue)));
1175 sub = lov_sub_get(env, lio, idx);
1176 LASSERT(!IS_ERR(sub));
1177 LASSERT(sub == &lio->lis_single_subio);
1178 rc = cl_io_commit_async(sub->sub_env, &sub->sub_io, queue,
1183 cl_page_list_init(plist);
1184 while (queue->pl_nr > 0) {
1188 LASSERT(plist->pl_nr == 0);
1189 page = cl_page_list_first(queue);
1190 LASSERT(!lov_page_is_empty(page));
1192 cl_page_list_move(plist, queue, page);
1194 index = page->cp_lov_index;
1195 while (queue->pl_nr > 0) {
1196 page = cl_page_list_first(queue);
1197 if (index != page->cp_lov_index)
1200 cl_page_list_move(plist, queue, page);
1203 if (queue->pl_nr > 0) /* still has more pages */
1204 stripe_to = PAGE_SIZE;
1206 sub = lov_sub_get(env, lio, index);
1208 rc = cl_io_commit_async(sub->sub_env, &sub->sub_io,
1209 plist, from, stripe_to, cb);
1215 if (plist->pl_nr > 0) /* short write */
1221 /* for error case, add the page back into the qin list */
1222 LASSERT(ergo(rc == 0, plist->pl_nr == 0));
1223 while (plist->pl_nr > 0) {
1224 /* error occurred, add the uncommitted pages back into queue */
1225 page = cl_page_list_last(plist);
1226 cl_page_list_move_head(queue, plist, page);
1232 static int lov_io_fault_start(const struct lu_env *env,
1233 const struct cl_io_slice *ios)
1235 struct cl_fault_io *fio;
1237 struct lov_io_sub *sub;
1241 fio = &ios->cis_io->u.ci_fault;
1242 lio = cl2lov_io(env, ios);
1243 sub = lov_sub_get(env, lio, fio->ft_page->cp_lov_index);
1244 sub->sub_io.u.ci_fault.ft_nob = fio->ft_nob;
1246 RETURN(lov_io_start(env, ios));
1249 static void lov_io_fsync_end(const struct lu_env *env,
1250 const struct cl_io_slice *ios)
1252 struct lov_io *lio = cl2lov_io(env, ios);
1253 struct lov_io_sub *sub;
1254 unsigned int *written = &ios->cis_io->u.ci_fsync.fi_nr_written;
1258 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
1259 struct cl_io *subio = &sub->sub_io;
1261 lov_io_end_wrapper(sub->sub_env, subio);
1263 if (subio->ci_result == 0)
1264 *written += subio->u.ci_fsync.fi_nr_written;
1269 static const struct cl_io_operations lov_io_ops = {
1272 .cio_fini = lov_io_fini,
1273 .cio_iter_init = lov_io_rw_iter_init,
1274 .cio_iter_fini = lov_io_iter_fini,
1275 .cio_lock = lov_io_lock,
1276 .cio_unlock = lov_io_unlock,
1277 .cio_start = lov_io_start,
1278 .cio_end = lov_io_end
1281 .cio_fini = lov_io_fini,
1282 .cio_iter_init = lov_io_rw_iter_init,
1283 .cio_iter_fini = lov_io_iter_fini,
1284 .cio_lock = lov_io_lock,
1285 .cio_unlock = lov_io_unlock,
1286 .cio_start = lov_io_start,
1287 .cio_end = lov_io_end
1290 .cio_fini = lov_io_fini,
1291 .cio_iter_init = lov_io_setattr_iter_init,
1292 .cio_iter_fini = lov_io_iter_fini,
1293 .cio_lock = lov_io_lock,
1294 .cio_unlock = lov_io_unlock,
1295 .cio_start = lov_io_start,
1296 .cio_end = lov_io_end
1298 [CIT_DATA_VERSION] = {
1299 .cio_fini = lov_io_fini,
1300 .cio_iter_init = lov_io_iter_init,
1301 .cio_iter_fini = lov_io_iter_fini,
1302 .cio_lock = lov_io_lock,
1303 .cio_unlock = lov_io_unlock,
1304 .cio_start = lov_io_start,
1305 .cio_end = lov_io_data_version_end,
1308 .cio_fini = lov_io_fini,
1309 .cio_iter_init = lov_io_iter_init,
1310 .cio_iter_fini = lov_io_iter_fini,
1311 .cio_lock = lov_io_lock,
1312 .cio_unlock = lov_io_unlock,
1313 .cio_start = lov_io_fault_start,
1314 .cio_end = lov_io_end
1317 .cio_fini = lov_io_fini,
1318 .cio_iter_init = lov_io_iter_init,
1319 .cio_iter_fini = lov_io_iter_fini,
1320 .cio_lock = lov_io_lock,
1321 .cio_unlock = lov_io_unlock,
1322 .cio_start = lov_io_start,
1323 .cio_end = lov_io_fsync_end
1326 .cio_fini = lov_io_fini,
1327 .cio_iter_init = lov_io_iter_init,
1328 .cio_iter_fini = lov_io_iter_fini,
1329 .cio_lock = lov_io_lock,
1330 .cio_unlock = lov_io_unlock,
1331 .cio_start = lov_io_start,
1332 .cio_end = lov_io_end
1335 .cio_fini = lov_io_fini,
1338 .cio_fini = lov_io_fini
1341 .cio_read_ahead = lov_io_read_ahead,
1342 .cio_submit = lov_io_submit,
1343 .cio_commit_async = lov_io_commit_async,
1346 /*****************************************************************************
1348 * Empty lov io operations.
1352 static void lov_empty_io_fini(const struct lu_env *env,
1353 const struct cl_io_slice *ios)
1355 struct lov_object *lov = cl2lov(ios->cis_obj);
1358 if (atomic_dec_and_test(&lov->lo_active_ios))
1359 wake_up_all(&lov->lo_waitq);
1363 static int lov_empty_io_submit(const struct lu_env *env,
1364 const struct cl_io_slice *ios,
1365 enum cl_req_type crt, struct cl_2queue *queue)
1370 static void lov_empty_impossible(const struct lu_env *env,
1371 struct cl_io_slice *ios)
1376 #define LOV_EMPTY_IMPOSSIBLE ((void *)lov_empty_impossible)
1379 * An io operation vector for files without stripes.
1381 static const struct cl_io_operations lov_empty_io_ops = {
1384 .cio_fini = lov_empty_io_fini,
1386 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1387 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1388 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1389 .cio_end = LOV_EMPTY_IMPOSSIBLE
1393 .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
1400 .cio_fini = lov_empty_io_fini,
1401 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1402 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1403 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1404 .cio_end = LOV_EMPTY_IMPOSSIBLE
1407 .cio_fini = lov_empty_io_fini,
1408 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1409 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1410 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1411 .cio_end = LOV_EMPTY_IMPOSSIBLE
1414 .cio_fini = lov_empty_io_fini
1417 .cio_fini = lov_empty_io_fini
1420 .cio_fini = lov_empty_io_fini
1423 .cio_fini = lov_empty_io_fini
1426 .cio_submit = lov_empty_io_submit,
1427 .cio_commit_async = LOV_EMPTY_IMPOSSIBLE
1430 int lov_io_init_composite(const struct lu_env *env, struct cl_object *obj,
1433 struct lov_io *lio = lov_env_io(env);
1434 struct lov_object *lov = cl2lov(obj);
1439 INIT_LIST_HEAD(&lio->lis_active);
1440 result = lov_io_slice_init(lio, lov, io);
1444 result = lov_io_subio_init(env, lio, io);
1446 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_io_ops);
1447 atomic_inc(&lov->lo_active_ios);
1451 io->ci_result = result < 0 ? result : 0;
1455 int lov_io_init_empty(const struct lu_env *env, struct cl_object *obj,
1458 struct lov_object *lov = cl2lov(obj);
1459 struct lov_io *lio = lov_env_io(env);
1463 lio->lis_object = lov;
1464 switch (io->ci_type) {
1475 case CIT_DATA_VERSION:
1483 CERROR("Page fault on a file without stripes: "DFID"\n",
1484 PFID(lu_object_fid(&obj->co_lu)));
1488 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
1489 atomic_inc(&lov->lo_active_ios);
1492 io->ci_result = result < 0 ? result : 0;
1496 int lov_io_init_released(const struct lu_env *env, struct cl_object *obj,
1499 struct lov_object *lov = cl2lov(obj);
1500 struct lov_io *lio = lov_env_io(env);
1504 LASSERT(lov->lo_lsm != NULL);
1505 lio->lis_object = lov;
1507 switch (io->ci_type) {
1509 LASSERTF(0, "invalid type %d\n", io->ci_type);
1510 result = -EOPNOTSUPP;
1516 case CIT_DATA_VERSION:
1521 * the truncate to 0 is managed by MDT:
1522 * - in open, for open O_TRUNC
1523 * - in setattr, for truncate
1525 /* the truncate is for size > 0 so triggers a restore */
1526 if (cl_io_is_trunc(io)) {
1527 io->ci_restore_needed = 1;
1535 io->ci_restore_needed = 1;
1541 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
1542 atomic_inc(&lov->lo_active_ios);
1545 io->ci_result = result < 0 ? result : 0;
1550 * Return the index in composite:lo_entries by the file offset
1552 int lov_io_layout_at(struct lov_io *lio, __u64 offset)
1554 struct lov_object *lov = lio->lis_object;
1555 struct lov_layout_composite *comp = &lov->u.composite;
1556 int start_index = 0;
1557 int end_index = comp->lo_entry_count - 1;
1560 LASSERT(lov->lo_type == LLT_COMP);
1562 /* This is actual file offset so nothing can cover eof. */
1563 if (offset == LUSTRE_EOF)
1566 if (lov_is_flr(lov)) {
1567 struct lov_mirror_entry *lre;
1569 LASSERT(lio->lis_mirror_index >= 0);
1571 lre = &comp->lo_mirrors[lio->lis_mirror_index];
1572 start_index = lre->lre_start;
1573 end_index = lre->lre_end;
1576 for (i = start_index; i <= end_index; i++) {
1577 struct lov_layout_entry *lle = lov_entry(lov, i);
1579 if ((offset >= lle->lle_extent->e_start &&
1580 offset < lle->lle_extent->e_end) ||
1581 (offset == OBD_OBJECT_EOF &&
1582 lle->lle_extent->e_end == OBD_OBJECT_EOF))