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);
128 sub_obj = lovsub2cl(lov_r0(lov, index)->lo_sub[stripe]);
129 sub_io = &sub->sub_io;
131 sub_io->ci_obj = sub_obj;
132 sub_io->ci_result = 0;
134 sub_io->ci_parent = io;
135 sub_io->ci_lockreq = io->ci_lockreq;
136 sub_io->ci_type = io->ci_type;
137 sub_io->ci_no_srvlock = io->ci_no_srvlock;
138 sub_io->ci_noatime = io->ci_noatime;
139 sub_io->ci_lock_no_expand = io->ci_lock_no_expand;
140 sub_io->ci_ndelay = io->ci_ndelay;
141 sub_io->ci_layout_version = io->ci_layout_version;
143 result = cl_io_sub_init(sub->sub_env, sub_io, io->ci_type, sub_obj);
146 lov_io_sub_fini(env, lio, sub);
151 struct lov_io_sub *lov_sub_get(const struct lu_env *env,
152 struct lov_io *lio, int index)
154 struct lov_io_sub *sub;
159 list_for_each_entry(sub, &lio->lis_subios, sub_list) {
160 if (sub->sub_subio_index == index) {
167 sub = lov_sub_alloc(lio, index);
169 GOTO(out, rc = -ENOMEM);
171 rc = lov_io_sub_init(env, lio, sub);
173 lov_sub_free(lio, sub);
177 list_add_tail(&sub->sub_list, &lio->lis_subios);
178 lio->lis_nr_subios++;
186 /*****************************************************************************
192 int lov_page_index(const struct cl_page *page)
194 const struct cl_page_slice *slice;
197 slice = cl_page_at(page, &lov_device_type);
198 LASSERT(slice != NULL);
199 LASSERT(slice->cpl_obj != NULL);
201 RETURN(cl2lov_page(slice)->lps_index);
204 static int lov_io_subio_init(const struct lu_env *env, struct lov_io *lio,
209 LASSERT(lio->lis_object != NULL);
211 INIT_LIST_HEAD(&lio->lis_subios);
212 lio->lis_single_subio_index = -1;
213 lio->lis_nr_subios = 0;
219 * Decide if it will need write intent RPC
221 static int lov_io_mirror_write_intent(struct lov_io *lio,
222 struct lov_object *obj, struct cl_io *io)
224 struct lov_layout_composite *comp = &obj->u.composite;
225 struct lu_extent *ext = &io->ci_write_intent;
226 struct lov_mirror_entry *lre;
227 struct lov_mirror_entry *primary;
228 struct lov_layout_entry *lle;
232 *ext = (typeof(*ext)) { lio->lis_pos, lio->lis_endpos };
233 io->ci_need_write_intent = 0;
235 if (!(io->ci_type == CIT_WRITE || cl_io_is_trunc(io) ||
236 cl_io_is_mkwrite(io)))
240 * FLR: check if it needs to send a write intent RPC to server.
241 * Writing to sync_pending file needs write intent RPC to change
242 * the file state back to write_pending, so that the layout version
243 * can be increased when the state changes to sync_pending at a later
244 * time. Otherwise there exists a chance that an evicted client may
245 * dirty the file data while resync client is working on it.
246 * Designated I/O is allowed for resync workload.
248 if (lov_flr_state(obj) == LCM_FL_RDONLY ||
249 (lov_flr_state(obj) == LCM_FL_SYNC_PENDING &&
250 io->ci_designated_mirror == 0)) {
251 io->ci_need_write_intent = 1;
255 LASSERT((lov_flr_state(obj) == LCM_FL_WRITE_PENDING));
256 LASSERT(comp->lo_preferred_mirror >= 0);
259 * need to iterate all components to see if there are
260 * multiple components covering the writing component
262 primary = &comp->lo_mirrors[comp->lo_preferred_mirror];
263 LASSERT(!primary->lre_stale);
264 lov_foreach_mirror_layout_entry(obj, lle, primary) {
265 LASSERT(lle->lle_valid);
266 if (!lu_extent_is_overlapped(ext, lle->lle_extent))
269 ext->e_start = MIN(ext->e_start, lle->lle_extent->e_start);
270 ext->e_end = MAX(ext->e_end, lle->lle_extent->e_end);
274 CERROR(DFID ": cannot find any valid components covering "
275 "file extent "DEXT", mirror: %d\n",
276 PFID(lu_object_fid(lov2lu(obj))), PEXT(ext),
277 primary->lre_mirror_id);
282 lov_foreach_mirror_entry(obj, lre) {
286 lov_foreach_mirror_layout_entry(obj, lle, lre) {
290 if (lu_extent_is_overlapped(ext, lle->lle_extent)) {
297 CDEBUG(D_VFSTRACE, DFID "there are %zd components to be staled to "
298 "modify file extent "DEXT", iot: %d\n",
299 PFID(lu_object_fid(lov2lu(obj))), count, PEXT(ext), io->ci_type);
301 io->ci_need_write_intent = count > 0;
306 static int lov_io_mirror_init(struct lov_io *lio, struct lov_object *obj,
309 struct lov_layout_composite *comp = &obj->u.composite;
315 if (!lov_is_flr(obj)) {
316 LASSERT(comp->lo_preferred_mirror == 0);
317 lio->lis_mirror_index = comp->lo_preferred_mirror;
322 /* transfer the layout version for verification */
323 if (io->ci_layout_version == 0)
324 io->ci_layout_version = obj->lo_lsm->lsm_layout_gen;
326 /* find the corresponding mirror for designated mirror IO */
327 if (io->ci_designated_mirror > 0) {
328 struct lov_mirror_entry *entry;
330 LASSERT(!io->ci_ndelay);
332 CDEBUG(D_LAYOUT, "designated I/O mirror state: %d\n",
335 if ((cl_io_is_trunc(io) || io->ci_type == CIT_WRITE) &&
336 (io->ci_layout_version != obj->lo_lsm->lsm_layout_gen)) {
338 * For resync I/O, the ci_layout_version was the layout
339 * version when resync starts. If it doesn't match the
340 * current object layout version, it means the layout
346 io->ci_layout_version |= LU_LAYOUT_RESYNC;
349 lio->lis_mirror_index = -1;
350 lov_foreach_mirror_entry(obj, entry) {
351 if (entry->lre_mirror_id ==
352 io->ci_designated_mirror) {
353 lio->lis_mirror_index = index;
360 RETURN(lio->lis_mirror_index < 0 ? -EINVAL : 0);
363 result = lov_io_mirror_write_intent(lio, obj, io);
367 if (io->ci_need_write_intent) {
368 CDEBUG(D_VFSTRACE, DFID " need write intent for [%llu, %llu)\n",
369 PFID(lu_object_fid(lov2lu(obj))),
370 lio->lis_pos, lio->lis_endpos);
372 if (cl_io_is_trunc(io)) {
374 * for truncate, we uses [size, EOF) to judge whether
375 * a write intent needs to be send, but we need to
376 * restore the write extent to [0, size).
378 io->ci_write_intent.e_start = 0;
379 io->ci_write_intent.e_end =
380 io->u.ci_setattr.sa_attr.lvb_size;
382 /* stop cl_io_init() loop */
386 if (io->ci_ndelay_tried == 0 || /* first time to try */
387 /* reset the mirror index if layout has changed */
388 lio->lis_mirror_layout_gen != obj->lo_lsm->lsm_layout_gen) {
389 lio->lis_mirror_layout_gen = obj->lo_lsm->lsm_layout_gen;
390 index = lio->lis_mirror_index = comp->lo_preferred_mirror;
392 index = lio->lis_mirror_index;
395 /* move mirror index to the next one */
396 index = (index + 1) % comp->lo_mirror_count;
399 for (i = 0; i < comp->lo_mirror_count; i++) {
400 struct lu_extent ext = { .e_start = lio->lis_pos,
401 .e_end = lio->lis_pos + 1 };
402 struct lov_mirror_entry *lre;
403 struct lov_layout_entry *lle;
406 lre = &comp->lo_mirrors[(index + i) % comp->lo_mirror_count];
410 lov_foreach_mirror_layout_entry(obj, lle, lre) {
414 if (lu_extent_is_overlapped(&ext, lle->lle_extent)) {
421 index = (index + i) % comp->lo_mirror_count;
425 if (i == comp->lo_mirror_count) {
426 CERROR(DFID": failed to find a component covering "
427 "I/O region at %llu\n",
428 PFID(lu_object_fid(lov2lu(obj))), lio->lis_pos);
430 dump_lsm(D_ERROR, obj->lo_lsm);
435 CDEBUG(D_VFSTRACE, DFID ": flr state: %d, move mirror from %d to %d, "
436 "have retried: %d, mirror count: %d\n",
437 PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj),
438 lio->lis_mirror_index, index, io->ci_ndelay_tried,
439 comp->lo_mirror_count);
441 lio->lis_mirror_index = index;
444 * FLR: if all mirrors have been tried once, most likely the network
445 * of this client has been partitioned. We should relinquish CPU for
446 * a while before trying again.
448 ++io->ci_ndelay_tried;
449 if (io->ci_ndelay && io->ci_ndelay_tried >= comp->lo_mirror_count) {
450 set_current_state(TASK_INTERRUPTIBLE);
451 schedule_timeout(msecs_to_jiffies(MSEC_PER_SEC)); /* 10ms */
452 if (signal_pending(current))
455 /* reset retry counter */
456 io->ci_ndelay_tried = 1;
459 CDEBUG(D_VFSTRACE, "use %sdelayed RPC state for this IO\n",
460 io->ci_ndelay ? "non-" : "");
465 static int lov_io_slice_init(struct lov_io *lio,
466 struct lov_object *obj, struct cl_io *io)
473 lio->lis_object = obj;
475 LASSERT(obj->lo_lsm != NULL);
477 switch (io->ci_type) {
480 lio->lis_pos = io->u.ci_rw.crw_pos;
481 lio->lis_endpos = io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count;
482 lio->lis_io_endpos = lio->lis_endpos;
483 if (cl_io_is_append(io)) {
484 LASSERT(io->ci_type == CIT_WRITE);
487 * If there is LOV EA hole, then we may cannot locate
488 * the current file-tail exactly.
490 if (unlikely(obj->lo_lsm->lsm_entries[0]->lsme_pattern &
492 GOTO(out, result = -EIO);
495 lio->lis_endpos = OBD_OBJECT_EOF;
500 if (cl_io_is_trunc(io))
501 lio->lis_pos = io->u.ci_setattr.sa_attr.lvb_size;
504 lio->lis_endpos = OBD_OBJECT_EOF;
507 case CIT_DATA_VERSION:
509 lio->lis_endpos = OBD_OBJECT_EOF;
513 pgoff_t index = io->u.ci_fault.ft_index;
515 lio->lis_pos = cl_offset(io->ci_obj, index);
516 lio->lis_endpos = cl_offset(io->ci_obj, index + 1);
521 lio->lis_pos = io->u.ci_fsync.fi_start;
522 lio->lis_endpos = io->u.ci_fsync.fi_end;
527 lio->lis_pos = io->u.ci_ladvise.li_start;
528 lio->lis_endpos = io->u.ci_ladvise.li_end;
534 lio->lis_endpos = OBD_OBJECT_EOF;
536 if (lov_flr_state(obj) == LCM_FL_RDONLY &&
537 !OBD_FAIL_CHECK(OBD_FAIL_FLR_GLIMPSE_IMMUTABLE))
538 /* SoM is accurate, no need glimpse */
539 GOTO(out, result = 1);
544 lio->lis_endpos = OBD_OBJECT_EOF;
551 result = lov_io_mirror_init(lio, obj, io);
555 /* check if it needs to instantiate layout */
556 if (!(io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io) ||
557 (cl_io_is_trunc(io) && io->u.ci_setattr.sa_attr.lvb_size > 0)))
558 GOTO(out, result = 0);
561 * for truncate, it only needs to instantiate the components
562 * before the truncated size.
564 if (cl_io_is_trunc(io)) {
565 io->ci_write_intent.e_start = 0;
566 io->ci_write_intent.e_end = io->u.ci_setattr.sa_attr.lvb_size;
568 io->ci_write_intent.e_start = lio->lis_pos;
569 io->ci_write_intent.e_end = lio->lis_endpos;
573 lov_foreach_io_layout(index, lio, &io->ci_write_intent) {
574 if (!lsm_entry_inited(obj->lo_lsm, index)) {
575 io->ci_need_write_intent = 1;
580 if (io->ci_need_write_intent && io->ci_designated_mirror > 0) {
582 * REINT_SYNC RPC has already tried to instantiate all of the
583 * components involved, obviously it didn't succeed. Skip this
584 * mirror for now. The server won't be able to figure out
585 * which mirror it should instantiate components
587 CERROR(DFID": trying to instantiate components for designated "
588 "I/O, file state: %d\n",
589 PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj));
591 io->ci_need_write_intent = 0;
592 GOTO(out, result = -EIO);
595 if (io->ci_need_write_intent)
596 GOTO(out, result = 1);
604 static void lov_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
606 struct lov_io *lio = cl2lov_io(env, ios);
607 struct lov_object *lov = cl2lov(ios->cis_obj);
611 LASSERT(list_empty(&lio->lis_active));
613 while (!list_empty(&lio->lis_subios)) {
614 struct lov_io_sub *sub = list_entry(lio->lis_subios.next,
618 list_del_init(&sub->sub_list);
619 lio->lis_nr_subios--;
621 lov_io_sub_fini(env, lio, sub);
622 lov_sub_free(lio, sub);
624 LASSERT(lio->lis_nr_subios == 0);
626 LASSERT(atomic_read(&lov->lo_active_ios) > 0);
627 if (atomic_dec_and_test(&lov->lo_active_ios))
628 wake_up_all(&lov->lo_waitq);
632 static void lov_io_sub_inherit(struct lov_io_sub *sub, struct lov_io *lio,
633 loff_t start, loff_t end)
635 struct cl_io *io = &sub->sub_io;
636 struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
637 struct cl_io *parent = lio->lis_cl.cis_io;
638 int index = lov_comp_entry(sub->sub_subio_index);
639 int stripe = lov_comp_stripe(sub->sub_subio_index);
641 switch (io->ci_type) {
643 io->u.ci_setattr.sa_attr = parent->u.ci_setattr.sa_attr;
644 io->u.ci_setattr.sa_attr_flags =
645 parent->u.ci_setattr.sa_attr_flags;
646 io->u.ci_setattr.sa_avalid = parent->u.ci_setattr.sa_avalid;
647 io->u.ci_setattr.sa_xvalid = parent->u.ci_setattr.sa_xvalid;
648 io->u.ci_setattr.sa_stripe_index = stripe;
649 io->u.ci_setattr.sa_parent_fid =
650 parent->u.ci_setattr.sa_parent_fid;
651 if (cl_io_is_trunc(io)) {
652 loff_t new_size = parent->u.ci_setattr.sa_attr.lvb_size;
654 new_size = lov_size_to_stripe(lsm, index, new_size,
656 io->u.ci_setattr.sa_attr.lvb_size = new_size;
658 lov_lsm2layout(lsm, lsm->lsm_entries[index],
659 &io->u.ci_setattr.sa_layout);
662 case CIT_DATA_VERSION: {
663 io->u.ci_data_version.dv_data_version = 0;
664 io->u.ci_data_version.dv_flags =
665 parent->u.ci_data_version.dv_flags;
669 struct cl_object *obj = parent->ci_obj;
670 loff_t off = cl_offset(obj, parent->u.ci_fault.ft_index);
672 io->u.ci_fault = parent->u.ci_fault;
673 off = lov_size_to_stripe(lsm, index, off, stripe);
674 io->u.ci_fault.ft_index = cl_index(obj, off);
678 io->u.ci_fsync.fi_start = start;
679 io->u.ci_fsync.fi_end = end;
680 io->u.ci_fsync.fi_fid = parent->u.ci_fsync.fi_fid;
681 io->u.ci_fsync.fi_mode = parent->u.ci_fsync.fi_mode;
686 io->u.ci_wr.wr_sync = cl_io_is_sync_write(parent);
687 if (cl_io_is_append(parent)) {
688 io->u.ci_wr.wr_append = 1;
690 io->u.ci_rw.crw_pos = start;
691 io->u.ci_rw.crw_count = end - start;
696 io->u.ci_ladvise.li_start = start;
697 io->u.ci_ladvise.li_end = end;
698 io->u.ci_ladvise.li_fid = parent->u.ci_ladvise.li_fid;
699 io->u.ci_ladvise.li_advice = parent->u.ci_ladvise.li_advice;
700 io->u.ci_ladvise.li_flags = parent->u.ci_ladvise.li_flags;
710 static loff_t lov_offset_mod(loff_t val, int delta)
712 if (val != OBD_OBJECT_EOF)
717 static int lov_io_iter_init(const struct lu_env *env,
718 const struct cl_io_slice *ios)
720 struct lov_io *lio = cl2lov_io(env, ios);
721 struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
722 struct lov_io_sub *sub;
723 struct lu_extent ext;
729 ext.e_start = lio->lis_pos;
730 ext.e_end = lio->lis_endpos;
732 lov_foreach_io_layout(index, lio, &ext) {
733 struct lov_layout_entry *le = lov_entry(lio->lis_object, index);
734 struct lov_layout_raid0 *r0 = &le->lle_raid0;
739 CDEBUG(D_VFSTRACE, "component[%d] flags %#x\n",
740 index, lsm->lsm_entries[index]->lsme_flags);
741 if (!lsm_entry_inited(lsm, index)) {
743 * Read from uninitialized components should return
749 if (!le->lle_valid && !ios->cis_io->ci_designated_mirror) {
750 CERROR("I/O to invalid component: %d, mirror: %d\n",
751 index, lio->lis_mirror_index);
755 for (stripe = 0; stripe < r0->lo_nr; stripe++) {
756 if (!lov_stripe_intersects(lsm, index, stripe,
760 if (unlikely(!r0->lo_sub[stripe])) {
761 if (ios->cis_io->ci_type == CIT_READ ||
762 ios->cis_io->ci_type == CIT_WRITE ||
763 ios->cis_io->ci_type == CIT_FAULT)
769 end = lov_offset_mod(end, 1);
770 sub = lov_sub_get(env, lio,
771 lov_comp_index(index, stripe));
777 lov_io_sub_inherit(sub, lio, start, end);
778 rc = cl_io_iter_init(sub->sub_env, &sub->sub_io);
780 cl_io_iter_fini(sub->sub_env, &sub->sub_io);
784 CDEBUG(D_VFSTRACE, "shrink: %d [%llu, %llu)\n",
787 list_add_tail(&sub->sub_linkage, &lio->lis_active);
795 static int lov_io_rw_iter_init(const struct lu_env *env,
796 const struct cl_io_slice *ios)
798 struct lov_io *lio = cl2lov_io(env, ios);
799 struct cl_io *io = ios->cis_io;
800 struct lov_stripe_md_entry *lse;
801 loff_t start = io->u.ci_rw.crw_pos;
805 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
808 if (cl_io_is_append(io))
809 RETURN(lov_io_iter_init(env, ios));
811 index = lov_io_layout_at(lio, io->u.ci_rw.crw_pos);
812 if (index < 0) { /* non-existing layout component */
813 if (io->ci_type == CIT_READ) {
815 * TODO: it needs to detect the next component and
816 * then set the next pos
820 RETURN(lov_io_iter_init(env, ios));
826 if (!lov_entry(lio->lis_object, index)->lle_valid &&
827 !io->ci_designated_mirror)
828 RETURN(io->ci_type == CIT_READ ? -EAGAIN : -EIO);
830 lse = lov_lse(lio->lis_object, index);
832 next = MAX_LFS_FILESIZE;
833 if (lse->lsme_stripe_count > 1) {
834 unsigned long ssize = lse->lsme_stripe_size;
836 lov_do_div64(start, ssize);
837 next = (start + 1) * ssize;
838 if (next <= start * ssize)
839 next = MAX_LFS_FILESIZE;
842 LASSERTF(io->u.ci_rw.crw_pos >= lse->lsme_extent.e_start,
843 "pos %lld, [%lld, %lld)\n", io->u.ci_rw.crw_pos,
844 lse->lsme_extent.e_start, lse->lsme_extent.e_end);
845 next = min_t(__u64, next, lse->lsme_extent.e_end);
846 next = min_t(loff_t, next, lio->lis_io_endpos);
848 io->ci_continue = next < lio->lis_io_endpos;
849 io->u.ci_rw.crw_count = next - io->u.ci_rw.crw_pos;
850 lio->lis_pos = io->u.ci_rw.crw_pos;
851 lio->lis_endpos = io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count;
853 "stripe: %llu chunk: [%llu, %llu) %llu, %zd\n",
854 (__u64)start, lio->lis_pos, lio->lis_endpos,
855 (__u64)lio->lis_io_endpos, io->u.ci_rw.crw_count);
858 * XXX The following call should be optimized: we know, that
859 * [lio->lis_pos, lio->lis_endpos) intersects with exactly one stripe.
861 RETURN(lov_io_iter_init(env, ios));
864 static int lov_io_setattr_iter_init(const struct lu_env *env,
865 const struct cl_io_slice *ios)
867 struct lov_io *lio = cl2lov_io(env, ios);
868 struct cl_io *io = ios->cis_io;
872 if (cl_io_is_trunc(io) && lio->lis_pos > 0) {
873 index = lov_io_layout_at(lio, lio->lis_pos - 1);
874 /* no entry found for such offset */
876 RETURN(io->ci_result = -ENODATA);
879 RETURN(lov_io_iter_init(env, ios));
882 static int lov_io_call(const struct lu_env *env, struct lov_io *lio,
883 int (*iofunc)(const struct lu_env *, struct cl_io *))
885 struct cl_io *parent = lio->lis_cl.cis_io;
886 struct lov_io_sub *sub;
890 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
891 rc = iofunc(sub->sub_env, &sub->sub_io);
895 if (parent->ci_result == 0)
896 parent->ci_result = sub->sub_io.ci_result;
901 static int lov_io_lock(const struct lu_env *env, const struct cl_io_slice *ios)
904 RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_lock));
907 static int lov_io_start(const struct lu_env *env, const struct cl_io_slice *ios)
910 RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_start));
913 static int lov_io_end_wrapper(const struct lu_env *env, struct cl_io *io)
917 * It's possible that lov_io_start() wasn't called against this
918 * sub-io, either because previous sub-io failed, or upper layer
921 if (io->ci_state == CIS_IO_GOING)
924 io->ci_state = CIS_IO_FINISHED;
928 static int lov_io_iter_fini_wrapper(const struct lu_env *env, struct cl_io *io)
930 cl_io_iter_fini(env, io);
934 static int lov_io_unlock_wrapper(const struct lu_env *env, struct cl_io *io)
936 cl_io_unlock(env, io);
940 static void lov_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
944 rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_end_wrapper);
949 lov_io_data_version_end(const struct lu_env *env, const struct cl_io_slice *ios)
951 struct lov_io *lio = cl2lov_io(env, ios);
952 struct cl_io *parent = lio->lis_cl.cis_io;
953 struct cl_data_version_io *pdv = &parent->u.ci_data_version;
954 struct lov_io_sub *sub;
957 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
958 struct cl_data_version_io *sdv = &sub->sub_io.u.ci_data_version;
960 lov_io_end_wrapper(sub->sub_env, &sub->sub_io);
962 pdv->dv_data_version += sdv->dv_data_version;
963 if (pdv->dv_layout_version > sdv->dv_layout_version)
964 pdv->dv_layout_version = sdv->dv_layout_version;
966 if (parent->ci_result == 0)
967 parent->ci_result = sub->sub_io.ci_result;
973 static void lov_io_iter_fini(const struct lu_env *env,
974 const struct cl_io_slice *ios)
976 struct lov_io *lio = cl2lov_io(env, ios);
980 rc = lov_io_call(env, lio, lov_io_iter_fini_wrapper);
982 while (!list_empty(&lio->lis_active))
983 list_del_init(lio->lis_active.next);
987 static void lov_io_unlock(const struct lu_env *env,
988 const struct cl_io_slice *ios)
993 rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_unlock_wrapper);
998 static int lov_io_read_ahead(const struct lu_env *env,
999 const struct cl_io_slice *ios,
1000 pgoff_t start, struct cl_read_ahead *ra)
1002 struct lov_io *lio = cl2lov_io(env, ios);
1003 struct lov_object *loo = lio->lis_object;
1004 struct cl_object *obj = lov2cl(loo);
1005 struct lov_layout_raid0 *r0;
1006 struct lov_io_sub *sub;
1010 unsigned int pps; /* pages per stripe */
1016 offset = cl_offset(obj, start);
1017 index = lov_io_layout_at(lio, offset);
1018 if (index < 0 || !lsm_entry_inited(loo->lo_lsm, index))
1021 /* avoid readahead to expand to stale components */
1022 if (!lov_entry(loo, index)->lle_valid)
1025 stripe = lov_stripe_number(loo->lo_lsm, index, offset);
1027 r0 = lov_r0(loo, index);
1028 if (unlikely(!r0->lo_sub[stripe]))
1031 sub = lov_sub_get(env, lio, lov_comp_index(index, stripe));
1033 RETURN(PTR_ERR(sub));
1035 lov_stripe_offset(loo->lo_lsm, index, offset, stripe, &suboff);
1036 rc = cl_io_read_ahead(sub->sub_env, &sub->sub_io,
1037 cl_index(lovsub2cl(r0->lo_sub[stripe]), suboff),
1040 CDEBUG(D_READA, DFID " cra_end = %lu, stripes = %d, rc = %d\n",
1041 PFID(lu_object_fid(lov2lu(loo))), ra->cra_end, r0->lo_nr, rc);
1046 * Adjust the stripe index by layout of comp. ra->cra_end is the
1047 * maximum page index covered by an underlying DLM lock.
1048 * This function converts cra_end from stripe level to file level, and
1049 * make sure it's not beyond stripe and component boundary.
1052 /* cra_end is stripe level, convert it into file level */
1053 ra_end = ra->cra_end;
1054 if (ra_end != CL_PAGE_EOF)
1055 ra->cra_end = lov_stripe_pgoff(loo->lo_lsm, index,
1058 /* boundary of current component */
1059 ra_end = cl_index(obj, (loff_t)lov_io_extent(lio, index)->e_end);
1060 if (ra_end != CL_PAGE_EOF && ra->cra_end >= ra_end)
1061 ra->cra_end = ra_end - 1;
1063 if (r0->lo_nr == 1) /* single stripe file */
1066 pps = lov_lse(loo, index)->lsme_stripe_size >> PAGE_SHIFT;
1068 CDEBUG(D_READA, DFID " max_index = %lu, pps = %u, index = %u, "
1069 "stripe_size = %u, stripe no = %u, start index = %lu\n",
1070 PFID(lu_object_fid(lov2lu(loo))), ra->cra_end, pps, index,
1071 lov_lse(loo, index)->lsme_stripe_size, stripe, start);
1073 /* never exceed the end of the stripe */
1074 ra->cra_end = min_t(pgoff_t,
1075 ra->cra_end, start + pps - start % pps - 1);
1080 * lov implementation of cl_operations::cio_submit() method. It takes a list
1081 * of pages in \a queue, splits it into per-stripe sub-lists, invokes
1082 * cl_io_submit() on underlying devices to submit sub-lists, and then splices
1085 * Major complication of this function is a need to handle memory cleansing:
1086 * cl_io_submit() is called to write out pages as a part of VM memory
1087 * reclamation, and hence it may not fail due to memory shortages (system
1088 * dead-locks otherwise). To deal with this, some resources (sub-lists,
1089 * sub-environment, etc.) are allocated per-device on "startup" (i.e., in a
1090 * not-memory cleansing context), and in case of memory shortage, these
1091 * pre-allocated resources are used by lov_io_submit() under
1092 * lov_device::ld_mutex mutex.
1094 static int lov_io_submit(const struct lu_env *env,
1095 const struct cl_io_slice *ios,
1096 enum cl_req_type crt, struct cl_2queue *queue)
1098 struct cl_page_list *qin = &queue->c2_qin;
1099 struct lov_io *lio = cl2lov_io(env, ios);
1100 struct lov_io_sub *sub;
1101 struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
1102 struct cl_page *page;
1107 cl_page_list_init(plist);
1108 while (qin->pl_nr > 0) {
1109 struct cl_2queue *cl2q = &lov_env_info(env)->lti_cl2q;
1111 page = cl_page_list_first(qin);
1112 if (lov_page_is_empty(page)) {
1113 cl_page_list_move(&queue->c2_qout, qin, page);
1116 * it could only be mirror read to get here therefore
1117 * the pages will be transient. We don't care about
1118 * the return code of cl_page_prep() at all.
1120 (void) cl_page_prep(env, ios->cis_io, page, crt);
1121 cl_page_completion(env, page, crt, 0);
1125 cl_2queue_init(cl2q);
1126 cl_page_list_move(&cl2q->c2_qin, qin, page);
1128 index = lov_page_index(page);
1129 while (qin->pl_nr > 0) {
1130 page = cl_page_list_first(qin);
1131 if (index != lov_page_index(page))
1134 cl_page_list_move(&cl2q->c2_qin, qin, page);
1137 sub = lov_sub_get(env, lio, index);
1139 rc = cl_io_submit_rw(sub->sub_env, &sub->sub_io,
1145 cl_page_list_splice(&cl2q->c2_qin, plist);
1146 cl_page_list_splice(&cl2q->c2_qout, &queue->c2_qout);
1147 cl_2queue_fini(env, cl2q);
1153 cl_page_list_splice(plist, qin);
1154 cl_page_list_fini(env, plist);
1159 static int lov_io_commit_async(const struct lu_env *env,
1160 const struct cl_io_slice *ios,
1161 struct cl_page_list *queue, int from, int to,
1164 struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
1165 struct lov_io *lio = cl2lov_io(env, ios);
1166 struct lov_io_sub *sub;
1167 struct cl_page *page;
1171 if (lio->lis_nr_subios == 1) {
1172 int idx = lio->lis_single_subio_index;
1174 LASSERT(!lov_page_is_empty(cl_page_list_first(queue)));
1176 sub = lov_sub_get(env, lio, idx);
1177 LASSERT(!IS_ERR(sub));
1178 LASSERT(sub == &lio->lis_single_subio);
1179 rc = cl_io_commit_async(sub->sub_env, &sub->sub_io, queue,
1184 cl_page_list_init(plist);
1185 while (queue->pl_nr > 0) {
1189 LASSERT(plist->pl_nr == 0);
1190 page = cl_page_list_first(queue);
1191 LASSERT(!lov_page_is_empty(page));
1193 cl_page_list_move(plist, queue, page);
1195 index = lov_page_index(page);
1196 while (queue->pl_nr > 0) {
1197 page = cl_page_list_first(queue);
1198 if (index != lov_page_index(page))
1201 cl_page_list_move(plist, queue, page);
1204 if (queue->pl_nr > 0) /* still has more pages */
1205 stripe_to = PAGE_SIZE;
1207 sub = lov_sub_get(env, lio, index);
1209 rc = cl_io_commit_async(sub->sub_env, &sub->sub_io,
1210 plist, from, stripe_to, cb);
1216 if (plist->pl_nr > 0) /* short write */
1222 /* for error case, add the page back into the qin list */
1223 LASSERT(ergo(rc == 0, plist->pl_nr == 0));
1224 while (plist->pl_nr > 0) {
1225 /* error occurred, add the uncommitted pages back into queue */
1226 page = cl_page_list_last(plist);
1227 cl_page_list_move_head(queue, plist, page);
1233 static int lov_io_fault_start(const struct lu_env *env,
1234 const struct cl_io_slice *ios)
1236 struct cl_fault_io *fio;
1238 struct lov_io_sub *sub;
1242 fio = &ios->cis_io->u.ci_fault;
1243 lio = cl2lov_io(env, ios);
1244 sub = lov_sub_get(env, lio, lov_page_index(fio->ft_page));
1245 sub->sub_io.u.ci_fault.ft_nob = fio->ft_nob;
1247 RETURN(lov_io_start(env, ios));
1250 static void lov_io_fsync_end(const struct lu_env *env,
1251 const struct cl_io_slice *ios)
1253 struct lov_io *lio = cl2lov_io(env, ios);
1254 struct lov_io_sub *sub;
1255 unsigned int *written = &ios->cis_io->u.ci_fsync.fi_nr_written;
1259 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
1260 struct cl_io *subio = &sub->sub_io;
1262 lov_io_end_wrapper(sub->sub_env, subio);
1264 if (subio->ci_result == 0)
1265 *written += subio->u.ci_fsync.fi_nr_written;
1270 static const struct cl_io_operations lov_io_ops = {
1273 .cio_fini = lov_io_fini,
1274 .cio_iter_init = lov_io_rw_iter_init,
1275 .cio_iter_fini = lov_io_iter_fini,
1276 .cio_lock = lov_io_lock,
1277 .cio_unlock = lov_io_unlock,
1278 .cio_start = lov_io_start,
1279 .cio_end = lov_io_end
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_setattr_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
1299 [CIT_DATA_VERSION] = {
1300 .cio_fini = lov_io_fini,
1301 .cio_iter_init = lov_io_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_data_version_end,
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_fault_start,
1315 .cio_end = lov_io_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_start,
1324 .cio_end = lov_io_fsync_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_end
1336 .cio_fini = lov_io_fini,
1339 .cio_fini = lov_io_fini
1342 .cio_read_ahead = lov_io_read_ahead,
1343 .cio_submit = lov_io_submit,
1344 .cio_commit_async = lov_io_commit_async,
1347 /*****************************************************************************
1349 * Empty lov io operations.
1353 static void lov_empty_io_fini(const struct lu_env *env,
1354 const struct cl_io_slice *ios)
1356 struct lov_object *lov = cl2lov(ios->cis_obj);
1359 if (atomic_dec_and_test(&lov->lo_active_ios))
1360 wake_up_all(&lov->lo_waitq);
1364 static int lov_empty_io_submit(const struct lu_env *env,
1365 const struct cl_io_slice *ios,
1366 enum cl_req_type crt, struct cl_2queue *queue)
1371 static void lov_empty_impossible(const struct lu_env *env,
1372 struct cl_io_slice *ios)
1377 #define LOV_EMPTY_IMPOSSIBLE ((void *)lov_empty_impossible)
1380 * An io operation vector for files without stripes.
1382 static const struct cl_io_operations lov_empty_io_ops = {
1385 .cio_fini = lov_empty_io_fini,
1387 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1388 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1389 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1390 .cio_end = LOV_EMPTY_IMPOSSIBLE
1394 .cio_fini = lov_empty_io_fini,
1395 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1396 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1397 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1398 .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
1418 .cio_fini = lov_empty_io_fini
1421 .cio_fini = lov_empty_io_fini
1424 .cio_fini = lov_empty_io_fini
1427 .cio_submit = lov_empty_io_submit,
1428 .cio_commit_async = LOV_EMPTY_IMPOSSIBLE
1431 int lov_io_init_composite(const struct lu_env *env, struct cl_object *obj,
1434 struct lov_io *lio = lov_env_io(env);
1435 struct lov_object *lov = cl2lov(obj);
1440 INIT_LIST_HEAD(&lio->lis_active);
1441 result = lov_io_slice_init(lio, lov, io);
1445 result = lov_io_subio_init(env, lio, io);
1447 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_io_ops);
1448 atomic_inc(&lov->lo_active_ios);
1452 io->ci_result = result < 0 ? result : 0;
1456 int lov_io_init_empty(const struct lu_env *env, struct cl_object *obj,
1459 struct lov_object *lov = cl2lov(obj);
1460 struct lov_io *lio = lov_env_io(env);
1464 lio->lis_object = lov;
1465 switch (io->ci_type) {
1476 case CIT_DATA_VERSION:
1484 CERROR("Page fault on a file without stripes: "DFID"\n",
1485 PFID(lu_object_fid(&obj->co_lu)));
1489 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
1490 atomic_inc(&lov->lo_active_ios);
1493 io->ci_result = result < 0 ? result : 0;
1497 int lov_io_init_released(const struct lu_env *env, struct cl_object *obj,
1500 struct lov_object *lov = cl2lov(obj);
1501 struct lov_io *lio = lov_env_io(env);
1505 LASSERT(lov->lo_lsm != NULL);
1506 lio->lis_object = lov;
1508 switch (io->ci_type) {
1510 LASSERTF(0, "invalid type %d\n", io->ci_type);
1511 result = -EOPNOTSUPP;
1517 case CIT_DATA_VERSION:
1522 * the truncate to 0 is managed by MDT:
1523 * - in open, for open O_TRUNC
1524 * - in setattr, for truncate
1526 /* the truncate is for size > 0 so triggers a restore */
1527 if (cl_io_is_trunc(io)) {
1528 io->ci_restore_needed = 1;
1536 io->ci_restore_needed = 1;
1542 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
1543 atomic_inc(&lov->lo_active_ios);
1546 io->ci_result = result < 0 ? result : 0;
1551 * Return the index in composite:lo_entries by the file offset
1553 int lov_io_layout_at(struct lov_io *lio, __u64 offset)
1555 struct lov_object *lov = lio->lis_object;
1556 struct lov_layout_composite *comp = &lov->u.composite;
1557 int start_index = 0;
1558 int end_index = comp->lo_entry_count - 1;
1561 LASSERT(lov->lo_type == LLT_COMP);
1563 /* This is actual file offset so nothing can cover eof. */
1564 if (offset == LUSTRE_EOF)
1567 if (lov_is_flr(lov)) {
1568 struct lov_mirror_entry *lre;
1570 LASSERT(lio->lis_mirror_index >= 0);
1572 lre = &comp->lo_mirrors[lio->lis_mirror_index];
1573 start_index = lre->lre_start;
1574 end_index = lre->lre_end;
1577 for (i = start_index; i <= end_index; i++) {
1578 struct lov_layout_entry *lle = lov_entry(lov, i);
1580 if ((offset >= lle->lle_extent->e_start &&
1581 offset < lle->lle_extent->e_end) ||
1582 (offset == OBD_OBJECT_EOF &&
1583 lle->lle_extent->e_end == OBD_OBJECT_EOF))