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 /*****************************************************************************
195 int lov_page_index(const struct cl_page *page)
197 const struct cl_page_slice *slice;
200 slice = cl_page_at(page, &lov_device_type);
201 LASSERT(slice != NULL);
202 LASSERT(slice->cpl_obj != NULL);
204 RETURN(cl2lov_page(slice)->lps_index);
207 static int lov_io_subio_init(const struct lu_env *env, struct lov_io *lio,
212 LASSERT(lio->lis_object != NULL);
214 INIT_LIST_HEAD(&lio->lis_subios);
215 lio->lis_single_subio_index = -1;
216 lio->lis_nr_subios = 0;
222 * Decide if it will need write intent RPC
224 static int lov_io_mirror_write_intent(struct lov_io *lio,
225 struct lov_object *obj, struct cl_io *io)
227 struct lov_layout_composite *comp = &obj->u.composite;
228 struct lu_extent *ext = &io->ci_write_intent;
229 struct lov_mirror_entry *lre;
230 struct lov_mirror_entry *primary;
231 struct lov_layout_entry *lle;
235 *ext = (typeof(*ext)) { lio->lis_pos, lio->lis_endpos };
236 io->ci_need_write_intent = 0;
238 if (!(io->ci_type == CIT_WRITE || cl_io_is_trunc(io) ||
239 cl_io_is_mkwrite(io)))
243 * FLR: check if it needs to send a write intent RPC to server.
244 * Writing to sync_pending file needs write intent RPC to change
245 * the file state back to write_pending, so that the layout version
246 * can be increased when the state changes to sync_pending at a later
247 * time. Otherwise there exists a chance that an evicted client may
248 * dirty the file data while resync client is working on it.
249 * Designated I/O is allowed for resync workload.
251 if (lov_flr_state(obj) == LCM_FL_RDONLY ||
252 (lov_flr_state(obj) == LCM_FL_SYNC_PENDING &&
253 io->ci_designated_mirror == 0)) {
254 io->ci_need_write_intent = 1;
258 LASSERT((lov_flr_state(obj) == LCM_FL_WRITE_PENDING));
259 LASSERT(comp->lo_preferred_mirror >= 0);
262 * need to iterate all components to see if there are
263 * multiple components covering the writing component
265 primary = &comp->lo_mirrors[comp->lo_preferred_mirror];
266 LASSERT(!primary->lre_stale);
267 lov_foreach_mirror_layout_entry(obj, lle, primary) {
268 LASSERT(lle->lle_valid);
269 if (!lu_extent_is_overlapped(ext, lle->lle_extent))
272 ext->e_start = MIN(ext->e_start, lle->lle_extent->e_start);
273 ext->e_end = MAX(ext->e_end, lle->lle_extent->e_end);
277 CERROR(DFID ": cannot find any valid components covering "
278 "file extent "DEXT", mirror: %d\n",
279 PFID(lu_object_fid(lov2lu(obj))), PEXT(ext),
280 primary->lre_mirror_id);
285 lov_foreach_mirror_entry(obj, lre) {
289 lov_foreach_mirror_layout_entry(obj, lle, lre) {
293 if (lu_extent_is_overlapped(ext, lle->lle_extent)) {
300 CDEBUG(D_VFSTRACE, DFID "there are %zd components to be staled to "
301 "modify file extent "DEXT", iot: %d\n",
302 PFID(lu_object_fid(lov2lu(obj))), count, PEXT(ext), io->ci_type);
304 io->ci_need_write_intent = count > 0;
309 static int lov_io_mirror_init(struct lov_io *lio, struct lov_object *obj,
312 struct lov_layout_composite *comp = &obj->u.composite;
318 if (!lov_is_flr(obj)) {
319 LASSERT(comp->lo_preferred_mirror == 0);
320 lio->lis_mirror_index = comp->lo_preferred_mirror;
325 /* transfer the layout version for verification */
326 if (io->ci_layout_version == 0)
327 io->ci_layout_version = obj->lo_lsm->lsm_layout_gen;
329 /* find the corresponding mirror for designated mirror IO */
330 if (io->ci_designated_mirror > 0) {
331 struct lov_mirror_entry *entry;
333 LASSERT(!io->ci_ndelay);
335 CDEBUG(D_LAYOUT, "designated I/O mirror state: %d\n",
338 if ((cl_io_is_trunc(io) || io->ci_type == CIT_WRITE) &&
339 (io->ci_layout_version != obj->lo_lsm->lsm_layout_gen)) {
341 * For resync I/O, the ci_layout_version was the layout
342 * version when resync starts. If it doesn't match the
343 * current object layout version, it means the layout
349 io->ci_layout_version |= LU_LAYOUT_RESYNC;
352 lio->lis_mirror_index = -1;
353 lov_foreach_mirror_entry(obj, entry) {
354 if (entry->lre_mirror_id ==
355 io->ci_designated_mirror) {
356 lio->lis_mirror_index = index;
363 RETURN(lio->lis_mirror_index < 0 ? -EINVAL : 0);
366 result = lov_io_mirror_write_intent(lio, obj, io);
370 if (io->ci_need_write_intent) {
371 CDEBUG(D_VFSTRACE, DFID " need write intent for [%llu, %llu)\n",
372 PFID(lu_object_fid(lov2lu(obj))),
373 lio->lis_pos, lio->lis_endpos);
375 if (cl_io_is_trunc(io)) {
377 * for truncate, we uses [size, EOF) to judge whether
378 * a write intent needs to be send, but we need to
379 * restore the write extent to [0, size).
381 io->ci_write_intent.e_start = 0;
382 io->ci_write_intent.e_end =
383 io->u.ci_setattr.sa_attr.lvb_size;
385 /* stop cl_io_init() loop */
389 if (io->ci_ndelay_tried == 0 || /* first time to try */
390 /* reset the mirror index if layout has changed */
391 lio->lis_mirror_layout_gen != obj->lo_lsm->lsm_layout_gen) {
392 lio->lis_mirror_layout_gen = obj->lo_lsm->lsm_layout_gen;
393 index = lio->lis_mirror_index = comp->lo_preferred_mirror;
395 index = lio->lis_mirror_index;
398 /* move mirror index to the next one */
399 index = (index + 1) % comp->lo_mirror_count;
402 for (i = 0; i < comp->lo_mirror_count; i++) {
403 struct lu_extent ext = { .e_start = lio->lis_pos,
404 .e_end = lio->lis_pos + 1 };
405 struct lov_mirror_entry *lre;
406 struct lov_layout_entry *lle;
409 lre = &comp->lo_mirrors[(index + i) % comp->lo_mirror_count];
413 lov_foreach_mirror_layout_entry(obj, lle, lre) {
417 if (lu_extent_is_overlapped(&ext, lle->lle_extent)) {
424 index = (index + i) % comp->lo_mirror_count;
428 if (i == comp->lo_mirror_count) {
429 CERROR(DFID": failed to find a component covering "
430 "I/O region at %llu\n",
431 PFID(lu_object_fid(lov2lu(obj))), lio->lis_pos);
433 dump_lsm(D_ERROR, obj->lo_lsm);
438 CDEBUG(D_VFSTRACE, DFID ": flr state: %d, move mirror from %d to %d, "
439 "have retried: %d, mirror count: %d\n",
440 PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj),
441 lio->lis_mirror_index, index, io->ci_ndelay_tried,
442 comp->lo_mirror_count);
444 lio->lis_mirror_index = index;
447 * FLR: if all mirrors have been tried once, most likely the network
448 * of this client has been partitioned. We should relinquish CPU for
449 * a while before trying again.
451 ++io->ci_ndelay_tried;
452 if (io->ci_ndelay && io->ci_ndelay_tried >= comp->lo_mirror_count) {
453 set_current_state(TASK_INTERRUPTIBLE);
454 schedule_timeout(msecs_to_jiffies(MSEC_PER_SEC)); /* 10ms */
455 if (signal_pending(current))
458 /* reset retry counter */
459 io->ci_ndelay_tried = 1;
462 CDEBUG(D_VFSTRACE, "use %sdelayed RPC state for this IO\n",
463 io->ci_ndelay ? "non-" : "");
468 static int lov_io_slice_init(struct lov_io *lio,
469 struct lov_object *obj, struct cl_io *io)
476 lio->lis_object = obj;
478 LASSERT(obj->lo_lsm != NULL);
480 switch (io->ci_type) {
483 lio->lis_pos = io->u.ci_rw.crw_pos;
484 lio->lis_endpos = io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count;
485 lio->lis_io_endpos = lio->lis_endpos;
486 if (cl_io_is_append(io)) {
487 LASSERT(io->ci_type == CIT_WRITE);
490 * If there is LOV EA hole, then we may cannot locate
491 * the current file-tail exactly.
493 if (unlikely(obj->lo_lsm->lsm_entries[0]->lsme_pattern &
495 GOTO(out, result = -EIO);
498 lio->lis_endpos = OBD_OBJECT_EOF;
503 if (cl_io_is_trunc(io))
504 lio->lis_pos = io->u.ci_setattr.sa_attr.lvb_size;
507 lio->lis_endpos = OBD_OBJECT_EOF;
510 case CIT_DATA_VERSION:
512 lio->lis_endpos = OBD_OBJECT_EOF;
516 pgoff_t index = io->u.ci_fault.ft_index;
518 lio->lis_pos = cl_offset(io->ci_obj, index);
519 lio->lis_endpos = cl_offset(io->ci_obj, index + 1);
524 lio->lis_pos = io->u.ci_fsync.fi_start;
525 lio->lis_endpos = io->u.ci_fsync.fi_end;
530 lio->lis_pos = io->u.ci_ladvise.li_start;
531 lio->lis_endpos = io->u.ci_ladvise.li_end;
537 lio->lis_endpos = OBD_OBJECT_EOF;
539 if (lov_flr_state(obj) == LCM_FL_RDONLY &&
540 !OBD_FAIL_CHECK(OBD_FAIL_FLR_GLIMPSE_IMMUTABLE))
541 /* SoM is accurate, no need glimpse */
542 GOTO(out, result = 1);
547 lio->lis_endpos = OBD_OBJECT_EOF;
554 result = lov_io_mirror_init(lio, obj, io);
558 /* check if it needs to instantiate layout */
559 if (!(io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io) ||
560 (cl_io_is_trunc(io) && io->u.ci_setattr.sa_attr.lvb_size > 0)))
561 GOTO(out, result = 0);
564 * for truncate, it only needs to instantiate the components
565 * before the truncated size.
567 if (cl_io_is_trunc(io)) {
568 io->ci_write_intent.e_start = 0;
569 io->ci_write_intent.e_end = io->u.ci_setattr.sa_attr.lvb_size;
571 io->ci_write_intent.e_start = lio->lis_pos;
572 io->ci_write_intent.e_end = lio->lis_endpos;
576 lov_foreach_io_layout(index, lio, &io->ci_write_intent) {
577 if (!lsm_entry_inited(obj->lo_lsm, index)) {
578 io->ci_need_write_intent = 1;
583 if (io->ci_need_write_intent && io->ci_designated_mirror > 0) {
585 * REINT_SYNC RPC has already tried to instantiate all of the
586 * components involved, obviously it didn't succeed. Skip this
587 * mirror for now. The server won't be able to figure out
588 * which mirror it should instantiate components
590 CERROR(DFID": trying to instantiate components for designated "
591 "I/O, file state: %d\n",
592 PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj));
594 io->ci_need_write_intent = 0;
595 GOTO(out, result = -EIO);
598 if (io->ci_need_write_intent)
599 GOTO(out, result = 1);
607 static void lov_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
609 struct lov_io *lio = cl2lov_io(env, ios);
610 struct lov_object *lov = cl2lov(ios->cis_obj);
614 LASSERT(list_empty(&lio->lis_active));
616 while (!list_empty(&lio->lis_subios)) {
617 struct lov_io_sub *sub = list_entry(lio->lis_subios.next,
621 list_del_init(&sub->sub_list);
622 lio->lis_nr_subios--;
624 lov_io_sub_fini(env, lio, sub);
625 lov_sub_free(lio, sub);
627 LASSERT(lio->lis_nr_subios == 0);
629 LASSERT(atomic_read(&lov->lo_active_ios) > 0);
630 if (atomic_dec_and_test(&lov->lo_active_ios))
631 wake_up_all(&lov->lo_waitq);
635 static void lov_io_sub_inherit(struct lov_io_sub *sub, struct lov_io *lio,
636 loff_t start, loff_t end)
638 struct cl_io *io = &sub->sub_io;
639 struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
640 struct cl_io *parent = lio->lis_cl.cis_io;
641 int index = lov_comp_entry(sub->sub_subio_index);
642 int stripe = lov_comp_stripe(sub->sub_subio_index);
644 switch (io->ci_type) {
646 io->u.ci_setattr.sa_attr = parent->u.ci_setattr.sa_attr;
647 io->u.ci_setattr.sa_attr_flags =
648 parent->u.ci_setattr.sa_attr_flags;
649 io->u.ci_setattr.sa_avalid = parent->u.ci_setattr.sa_avalid;
650 io->u.ci_setattr.sa_xvalid = parent->u.ci_setattr.sa_xvalid;
651 io->u.ci_setattr.sa_stripe_index = stripe;
652 io->u.ci_setattr.sa_parent_fid =
653 parent->u.ci_setattr.sa_parent_fid;
654 if (cl_io_is_trunc(io)) {
655 loff_t new_size = parent->u.ci_setattr.sa_attr.lvb_size;
657 new_size = lov_size_to_stripe(lsm, index, new_size,
659 io->u.ci_setattr.sa_attr.lvb_size = new_size;
661 lov_lsm2layout(lsm, lsm->lsm_entries[index],
662 &io->u.ci_setattr.sa_layout);
665 case CIT_DATA_VERSION: {
666 io->u.ci_data_version.dv_data_version = 0;
667 io->u.ci_data_version.dv_flags =
668 parent->u.ci_data_version.dv_flags;
672 struct cl_object *obj = parent->ci_obj;
673 loff_t off = cl_offset(obj, parent->u.ci_fault.ft_index);
675 io->u.ci_fault = parent->u.ci_fault;
676 off = lov_size_to_stripe(lsm, index, off, stripe);
677 io->u.ci_fault.ft_index = cl_index(obj, off);
681 io->u.ci_fsync.fi_start = start;
682 io->u.ci_fsync.fi_end = end;
683 io->u.ci_fsync.fi_fid = parent->u.ci_fsync.fi_fid;
684 io->u.ci_fsync.fi_mode = parent->u.ci_fsync.fi_mode;
689 io->u.ci_wr.wr_sync = cl_io_is_sync_write(parent);
690 if (cl_io_is_append(parent)) {
691 io->u.ci_wr.wr_append = 1;
693 io->u.ci_rw.crw_pos = start;
694 io->u.ci_rw.crw_count = end - start;
699 io->u.ci_ladvise.li_start = start;
700 io->u.ci_ladvise.li_end = end;
701 io->u.ci_ladvise.li_fid = parent->u.ci_ladvise.li_fid;
702 io->u.ci_ladvise.li_advice = parent->u.ci_ladvise.li_advice;
703 io->u.ci_ladvise.li_flags = parent->u.ci_ladvise.li_flags;
713 static loff_t lov_offset_mod(loff_t val, int delta)
715 if (val != OBD_OBJECT_EOF)
720 static int lov_io_iter_init(const struct lu_env *env,
721 const struct cl_io_slice *ios)
723 struct lov_io *lio = cl2lov_io(env, ios);
724 struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
725 struct lov_io_sub *sub;
726 struct lu_extent ext;
732 ext.e_start = lio->lis_pos;
733 ext.e_end = lio->lis_endpos;
735 lov_foreach_io_layout(index, lio, &ext) {
736 struct lov_layout_entry *le = lov_entry(lio->lis_object, index);
737 struct lov_layout_raid0 *r0 = &le->lle_raid0;
742 CDEBUG(D_VFSTRACE, "component[%d] flags %#x\n",
743 index, lsm->lsm_entries[index]->lsme_flags);
744 if (!lsm_entry_inited(lsm, index)) {
746 * Read from uninitialized components should return
752 if (!le->lle_valid && !ios->cis_io->ci_designated_mirror) {
753 CERROR("I/O to invalid component: %d, mirror: %d\n",
754 index, lio->lis_mirror_index);
758 for (stripe = 0; stripe < r0->lo_nr; stripe++) {
759 if (!lov_stripe_intersects(lsm, index, stripe,
763 if (unlikely(!r0->lo_sub[stripe])) {
764 if (ios->cis_io->ci_type == CIT_READ ||
765 ios->cis_io->ci_type == CIT_WRITE ||
766 ios->cis_io->ci_type == CIT_FAULT)
772 end = lov_offset_mod(end, 1);
773 sub = lov_sub_get(env, lio,
774 lov_comp_index(index, stripe));
780 lov_io_sub_inherit(sub, lio, start, end);
781 rc = cl_io_iter_init(sub->sub_env, &sub->sub_io);
783 cl_io_iter_fini(sub->sub_env, &sub->sub_io);
787 CDEBUG(D_VFSTRACE, "shrink: %d [%llu, %llu)\n",
790 list_add_tail(&sub->sub_linkage, &lio->lis_active);
798 static int lov_io_rw_iter_init(const struct lu_env *env,
799 const struct cl_io_slice *ios)
801 struct lov_io *lio = cl2lov_io(env, ios);
802 struct cl_io *io = ios->cis_io;
803 struct lov_stripe_md_entry *lse;
804 loff_t start = io->u.ci_rw.crw_pos;
808 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
811 if (cl_io_is_append(io))
812 RETURN(lov_io_iter_init(env, ios));
814 index = lov_io_layout_at(lio, io->u.ci_rw.crw_pos);
815 if (index < 0) { /* non-existing layout component */
816 if (io->ci_type == CIT_READ) {
818 * TODO: it needs to detect the next component and
819 * then set the next pos
823 RETURN(lov_io_iter_init(env, ios));
829 if (!lov_entry(lio->lis_object, index)->lle_valid &&
830 !io->ci_designated_mirror)
831 RETURN(io->ci_type == CIT_READ ? -EAGAIN : -EIO);
833 lse = lov_lse(lio->lis_object, index);
835 next = MAX_LFS_FILESIZE;
836 if (lse->lsme_stripe_count > 1) {
837 unsigned long ssize = lse->lsme_stripe_size;
839 lov_do_div64(start, ssize);
840 next = (start + 1) * ssize;
841 if (next <= start * ssize)
842 next = MAX_LFS_FILESIZE;
845 LASSERTF(io->u.ci_rw.crw_pos >= lse->lsme_extent.e_start,
846 "pos %lld, [%lld, %lld)\n", io->u.ci_rw.crw_pos,
847 lse->lsme_extent.e_start, lse->lsme_extent.e_end);
848 next = min_t(__u64, next, lse->lsme_extent.e_end);
849 next = min_t(loff_t, next, lio->lis_io_endpos);
851 io->ci_continue = next < lio->lis_io_endpos;
852 io->u.ci_rw.crw_count = next - io->u.ci_rw.crw_pos;
853 lio->lis_pos = io->u.ci_rw.crw_pos;
854 lio->lis_endpos = io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count;
856 "stripe: %llu chunk: [%llu, %llu) %llu, %zd\n",
857 (__u64)start, lio->lis_pos, lio->lis_endpos,
858 (__u64)lio->lis_io_endpos, io->u.ci_rw.crw_count);
861 * XXX The following call should be optimized: we know, that
862 * [lio->lis_pos, lio->lis_endpos) intersects with exactly one stripe.
864 RETURN(lov_io_iter_init(env, ios));
867 static int lov_io_setattr_iter_init(const struct lu_env *env,
868 const struct cl_io_slice *ios)
870 struct lov_io *lio = cl2lov_io(env, ios);
871 struct cl_io *io = ios->cis_io;
875 if (cl_io_is_trunc(io) && lio->lis_pos > 0) {
876 index = lov_io_layout_at(lio, lio->lis_pos - 1);
877 /* no entry found for such offset */
879 RETURN(io->ci_result = -ENODATA);
882 RETURN(lov_io_iter_init(env, ios));
885 static int lov_io_call(const struct lu_env *env, struct lov_io *lio,
886 int (*iofunc)(const struct lu_env *, struct cl_io *))
888 struct cl_io *parent = lio->lis_cl.cis_io;
889 struct lov_io_sub *sub;
893 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
894 rc = iofunc(sub->sub_env, &sub->sub_io);
898 if (parent->ci_result == 0)
899 parent->ci_result = sub->sub_io.ci_result;
904 static int lov_io_lock(const struct lu_env *env, const struct cl_io_slice *ios)
907 RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_lock));
910 static int lov_io_start(const struct lu_env *env, const struct cl_io_slice *ios)
913 RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_start));
916 static int lov_io_end_wrapper(const struct lu_env *env, struct cl_io *io)
920 * It's possible that lov_io_start() wasn't called against this
921 * sub-io, either because previous sub-io failed, or upper layer
924 if (io->ci_state == CIS_IO_GOING)
927 io->ci_state = CIS_IO_FINISHED;
931 static int lov_io_iter_fini_wrapper(const struct lu_env *env, struct cl_io *io)
933 cl_io_iter_fini(env, io);
937 static int lov_io_unlock_wrapper(const struct lu_env *env, struct cl_io *io)
939 cl_io_unlock(env, io);
943 static void lov_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
947 rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_end_wrapper);
952 lov_io_data_version_end(const struct lu_env *env, const struct cl_io_slice *ios)
954 struct lov_io *lio = cl2lov_io(env, ios);
955 struct cl_io *parent = lio->lis_cl.cis_io;
956 struct cl_data_version_io *pdv = &parent->u.ci_data_version;
957 struct lov_io_sub *sub;
960 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
961 struct cl_data_version_io *sdv = &sub->sub_io.u.ci_data_version;
963 lov_io_end_wrapper(sub->sub_env, &sub->sub_io);
965 pdv->dv_data_version += sdv->dv_data_version;
966 if (pdv->dv_layout_version > sdv->dv_layout_version)
967 pdv->dv_layout_version = sdv->dv_layout_version;
969 if (parent->ci_result == 0)
970 parent->ci_result = sub->sub_io.ci_result;
976 static void lov_io_iter_fini(const struct lu_env *env,
977 const struct cl_io_slice *ios)
979 struct lov_io *lio = cl2lov_io(env, ios);
983 rc = lov_io_call(env, lio, lov_io_iter_fini_wrapper);
985 while (!list_empty(&lio->lis_active))
986 list_del_init(lio->lis_active.next);
990 static void lov_io_unlock(const struct lu_env *env,
991 const struct cl_io_slice *ios)
996 rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_unlock_wrapper);
1001 static int lov_io_read_ahead(const struct lu_env *env,
1002 const struct cl_io_slice *ios,
1003 pgoff_t start, struct cl_read_ahead *ra)
1005 struct lov_io *lio = cl2lov_io(env, ios);
1006 struct lov_object *loo = lio->lis_object;
1007 struct cl_object *obj = lov2cl(loo);
1008 struct lov_layout_raid0 *r0;
1009 struct lov_io_sub *sub;
1013 unsigned int pps; /* pages per stripe */
1019 offset = cl_offset(obj, start);
1020 index = lov_io_layout_at(lio, offset);
1021 if (index < 0 || !lsm_entry_inited(loo->lo_lsm, index))
1024 /* avoid readahead to expand to stale components */
1025 if (!lov_entry(loo, index)->lle_valid)
1028 stripe = lov_stripe_number(loo->lo_lsm, index, offset);
1030 r0 = lov_r0(loo, index);
1031 if (unlikely(!r0->lo_sub[stripe]))
1034 sub = lov_sub_get(env, lio, lov_comp_index(index, stripe));
1036 RETURN(PTR_ERR(sub));
1038 lov_stripe_offset(loo->lo_lsm, index, offset, stripe, &suboff);
1039 rc = cl_io_read_ahead(sub->sub_env, &sub->sub_io,
1040 cl_index(lovsub2cl(r0->lo_sub[stripe]), suboff),
1043 CDEBUG(D_READA, DFID " cra_end = %lu, stripes = %d, rc = %d\n",
1044 PFID(lu_object_fid(lov2lu(loo))), ra->cra_end, r0->lo_nr, rc);
1049 * Adjust the stripe index by layout of comp. ra->cra_end is the
1050 * maximum page index covered by an underlying DLM lock.
1051 * This function converts cra_end from stripe level to file level, and
1052 * make sure it's not beyond stripe and component boundary.
1055 /* cra_end is stripe level, convert it into file level */
1056 ra_end = ra->cra_end;
1057 if (ra_end != CL_PAGE_EOF)
1058 ra->cra_end = lov_stripe_pgoff(loo->lo_lsm, index,
1061 /* boundary of current component */
1062 ra_end = cl_index(obj, (loff_t)lov_io_extent(lio, index)->e_end);
1063 if (ra_end != CL_PAGE_EOF && ra->cra_end >= ra_end)
1064 ra->cra_end = ra_end - 1;
1066 if (r0->lo_nr == 1) /* single stripe file */
1069 pps = lov_lse(loo, index)->lsme_stripe_size >> PAGE_SHIFT;
1071 CDEBUG(D_READA, DFID " max_index = %lu, pps = %u, index = %u, "
1072 "stripe_size = %u, stripe no = %u, start index = %lu\n",
1073 PFID(lu_object_fid(lov2lu(loo))), ra->cra_end, pps, index,
1074 lov_lse(loo, index)->lsme_stripe_size, stripe, start);
1076 /* never exceed the end of the stripe */
1077 ra->cra_end = min_t(pgoff_t,
1078 ra->cra_end, start + pps - start % pps - 1);
1083 * lov implementation of cl_operations::cio_submit() method. It takes a list
1084 * of pages in \a queue, splits it into per-stripe sub-lists, invokes
1085 * cl_io_submit() on underlying devices to submit sub-lists, and then splices
1088 * Major complication of this function is a need to handle memory cleansing:
1089 * cl_io_submit() is called to write out pages as a part of VM memory
1090 * reclamation, and hence it may not fail due to memory shortages (system
1091 * dead-locks otherwise). To deal with this, some resources (sub-lists,
1092 * sub-environment, etc.) are allocated per-device on "startup" (i.e., in a
1093 * not-memory cleansing context), and in case of memory shortage, these
1094 * pre-allocated resources are used by lov_io_submit() under
1095 * lov_device::ld_mutex mutex.
1097 static int lov_io_submit(const struct lu_env *env,
1098 const struct cl_io_slice *ios,
1099 enum cl_req_type crt, struct cl_2queue *queue)
1101 struct cl_page_list *qin = &queue->c2_qin;
1102 struct lov_io *lio = cl2lov_io(env, ios);
1103 struct lov_io_sub *sub;
1104 struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
1105 struct cl_page *page;
1110 cl_page_list_init(plist);
1111 while (qin->pl_nr > 0) {
1112 struct cl_2queue *cl2q = &lov_env_info(env)->lti_cl2q;
1114 page = cl_page_list_first(qin);
1115 if (lov_page_is_empty(page)) {
1116 cl_page_list_move(&queue->c2_qout, qin, page);
1119 * it could only be mirror read to get here therefore
1120 * the pages will be transient. We don't care about
1121 * the return code of cl_page_prep() at all.
1123 (void) cl_page_prep(env, ios->cis_io, page, crt);
1124 cl_page_completion(env, page, crt, 0);
1128 cl_2queue_init(cl2q);
1129 cl_page_list_move(&cl2q->c2_qin, qin, page);
1131 index = lov_page_index(page);
1132 while (qin->pl_nr > 0) {
1133 page = cl_page_list_first(qin);
1134 if (index != lov_page_index(page))
1137 cl_page_list_move(&cl2q->c2_qin, qin, page);
1140 sub = lov_sub_get(env, lio, index);
1142 rc = cl_io_submit_rw(sub->sub_env, &sub->sub_io,
1148 cl_page_list_splice(&cl2q->c2_qin, plist);
1149 cl_page_list_splice(&cl2q->c2_qout, &queue->c2_qout);
1150 cl_2queue_fini(env, cl2q);
1156 cl_page_list_splice(plist, qin);
1157 cl_page_list_fini(env, plist);
1162 static int lov_io_commit_async(const struct lu_env *env,
1163 const struct cl_io_slice *ios,
1164 struct cl_page_list *queue, int from, int to,
1167 struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
1168 struct lov_io *lio = cl2lov_io(env, ios);
1169 struct lov_io_sub *sub;
1170 struct cl_page *page;
1174 if (lio->lis_nr_subios == 1) {
1175 int idx = lio->lis_single_subio_index;
1177 LASSERT(!lov_page_is_empty(cl_page_list_first(queue)));
1179 sub = lov_sub_get(env, lio, idx);
1180 LASSERT(!IS_ERR(sub));
1181 LASSERT(sub == &lio->lis_single_subio);
1182 rc = cl_io_commit_async(sub->sub_env, &sub->sub_io, queue,
1187 cl_page_list_init(plist);
1188 while (queue->pl_nr > 0) {
1192 LASSERT(plist->pl_nr == 0);
1193 page = cl_page_list_first(queue);
1194 LASSERT(!lov_page_is_empty(page));
1196 cl_page_list_move(plist, queue, page);
1198 index = lov_page_index(page);
1199 while (queue->pl_nr > 0) {
1200 page = cl_page_list_first(queue);
1201 if (index != lov_page_index(page))
1204 cl_page_list_move(plist, queue, page);
1207 if (queue->pl_nr > 0) /* still has more pages */
1208 stripe_to = PAGE_SIZE;
1210 sub = lov_sub_get(env, lio, index);
1212 rc = cl_io_commit_async(sub->sub_env, &sub->sub_io,
1213 plist, from, stripe_to, cb);
1219 if (plist->pl_nr > 0) /* short write */
1225 /* for error case, add the page back into the qin list */
1226 LASSERT(ergo(rc == 0, plist->pl_nr == 0));
1227 while (plist->pl_nr > 0) {
1228 /* error occurred, add the uncommitted pages back into queue */
1229 page = cl_page_list_last(plist);
1230 cl_page_list_move_head(queue, plist, page);
1236 static int lov_io_fault_start(const struct lu_env *env,
1237 const struct cl_io_slice *ios)
1239 struct cl_fault_io *fio;
1241 struct lov_io_sub *sub;
1245 fio = &ios->cis_io->u.ci_fault;
1246 lio = cl2lov_io(env, ios);
1247 sub = lov_sub_get(env, lio, lov_page_index(fio->ft_page));
1248 sub->sub_io.u.ci_fault.ft_nob = fio->ft_nob;
1250 RETURN(lov_io_start(env, ios));
1253 static void lov_io_fsync_end(const struct lu_env *env,
1254 const struct cl_io_slice *ios)
1256 struct lov_io *lio = cl2lov_io(env, ios);
1257 struct lov_io_sub *sub;
1258 unsigned int *written = &ios->cis_io->u.ci_fsync.fi_nr_written;
1262 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
1263 struct cl_io *subio = &sub->sub_io;
1265 lov_io_end_wrapper(sub->sub_env, subio);
1267 if (subio->ci_result == 0)
1268 *written += subio->u.ci_fsync.fi_nr_written;
1273 static const struct cl_io_operations lov_io_ops = {
1276 .cio_fini = lov_io_fini,
1277 .cio_iter_init = lov_io_rw_iter_init,
1278 .cio_iter_fini = lov_io_iter_fini,
1279 .cio_lock = lov_io_lock,
1280 .cio_unlock = lov_io_unlock,
1281 .cio_start = lov_io_start,
1282 .cio_end = lov_io_end
1285 .cio_fini = lov_io_fini,
1286 .cio_iter_init = lov_io_rw_iter_init,
1287 .cio_iter_fini = lov_io_iter_fini,
1288 .cio_lock = lov_io_lock,
1289 .cio_unlock = lov_io_unlock,
1290 .cio_start = lov_io_start,
1291 .cio_end = lov_io_end
1294 .cio_fini = lov_io_fini,
1295 .cio_iter_init = lov_io_setattr_iter_init,
1296 .cio_iter_fini = lov_io_iter_fini,
1297 .cio_lock = lov_io_lock,
1298 .cio_unlock = lov_io_unlock,
1299 .cio_start = lov_io_start,
1300 .cio_end = lov_io_end
1302 [CIT_DATA_VERSION] = {
1303 .cio_fini = lov_io_fini,
1304 .cio_iter_init = lov_io_iter_init,
1305 .cio_iter_fini = lov_io_iter_fini,
1306 .cio_lock = lov_io_lock,
1307 .cio_unlock = lov_io_unlock,
1308 .cio_start = lov_io_start,
1309 .cio_end = lov_io_data_version_end,
1312 .cio_fini = lov_io_fini,
1313 .cio_iter_init = lov_io_iter_init,
1314 .cio_iter_fini = lov_io_iter_fini,
1315 .cio_lock = lov_io_lock,
1316 .cio_unlock = lov_io_unlock,
1317 .cio_start = lov_io_fault_start,
1318 .cio_end = lov_io_end
1321 .cio_fini = lov_io_fini,
1322 .cio_iter_init = lov_io_iter_init,
1323 .cio_iter_fini = lov_io_iter_fini,
1324 .cio_lock = lov_io_lock,
1325 .cio_unlock = lov_io_unlock,
1326 .cio_start = lov_io_start,
1327 .cio_end = lov_io_fsync_end
1330 .cio_fini = lov_io_fini,
1331 .cio_iter_init = lov_io_iter_init,
1332 .cio_iter_fini = lov_io_iter_fini,
1333 .cio_lock = lov_io_lock,
1334 .cio_unlock = lov_io_unlock,
1335 .cio_start = lov_io_start,
1336 .cio_end = lov_io_end
1339 .cio_fini = lov_io_fini,
1342 .cio_fini = lov_io_fini
1345 .cio_read_ahead = lov_io_read_ahead,
1346 .cio_submit = lov_io_submit,
1347 .cio_commit_async = lov_io_commit_async,
1350 /*****************************************************************************
1352 * Empty lov io operations.
1356 static void lov_empty_io_fini(const struct lu_env *env,
1357 const struct cl_io_slice *ios)
1359 struct lov_object *lov = cl2lov(ios->cis_obj);
1362 if (atomic_dec_and_test(&lov->lo_active_ios))
1363 wake_up_all(&lov->lo_waitq);
1367 static int lov_empty_io_submit(const struct lu_env *env,
1368 const struct cl_io_slice *ios,
1369 enum cl_req_type crt, struct cl_2queue *queue)
1374 static void lov_empty_impossible(const struct lu_env *env,
1375 struct cl_io_slice *ios)
1380 #define LOV_EMPTY_IMPOSSIBLE ((void *)lov_empty_impossible)
1383 * An io operation vector for files without stripes.
1385 static const struct cl_io_operations lov_empty_io_ops = {
1388 .cio_fini = lov_empty_io_fini,
1390 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1391 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1392 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1393 .cio_end = LOV_EMPTY_IMPOSSIBLE
1397 .cio_fini = lov_empty_io_fini,
1398 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1399 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1400 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1401 .cio_end = LOV_EMPTY_IMPOSSIBLE
1404 .cio_fini = lov_empty_io_fini,
1405 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1406 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1407 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1408 .cio_end = LOV_EMPTY_IMPOSSIBLE
1411 .cio_fini = lov_empty_io_fini,
1412 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1413 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1414 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1415 .cio_end = LOV_EMPTY_IMPOSSIBLE
1418 .cio_fini = lov_empty_io_fini
1421 .cio_fini = lov_empty_io_fini
1424 .cio_fini = lov_empty_io_fini
1427 .cio_fini = lov_empty_io_fini
1430 .cio_submit = lov_empty_io_submit,
1431 .cio_commit_async = LOV_EMPTY_IMPOSSIBLE
1434 int lov_io_init_composite(const struct lu_env *env, struct cl_object *obj,
1437 struct lov_io *lio = lov_env_io(env);
1438 struct lov_object *lov = cl2lov(obj);
1443 INIT_LIST_HEAD(&lio->lis_active);
1444 result = lov_io_slice_init(lio, lov, io);
1448 result = lov_io_subio_init(env, lio, io);
1450 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_io_ops);
1451 atomic_inc(&lov->lo_active_ios);
1455 io->ci_result = result < 0 ? result : 0;
1459 int lov_io_init_empty(const struct lu_env *env, struct cl_object *obj,
1462 struct lov_object *lov = cl2lov(obj);
1463 struct lov_io *lio = lov_env_io(env);
1467 lio->lis_object = lov;
1468 switch (io->ci_type) {
1479 case CIT_DATA_VERSION:
1487 CERROR("Page fault on a file without stripes: "DFID"\n",
1488 PFID(lu_object_fid(&obj->co_lu)));
1492 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
1493 atomic_inc(&lov->lo_active_ios);
1496 io->ci_result = result < 0 ? result : 0;
1500 int lov_io_init_released(const struct lu_env *env, struct cl_object *obj,
1503 struct lov_object *lov = cl2lov(obj);
1504 struct lov_io *lio = lov_env_io(env);
1508 LASSERT(lov->lo_lsm != NULL);
1509 lio->lis_object = lov;
1511 switch (io->ci_type) {
1513 LASSERTF(0, "invalid type %d\n", io->ci_type);
1514 result = -EOPNOTSUPP;
1520 case CIT_DATA_VERSION:
1525 * the truncate to 0 is managed by MDT:
1526 * - in open, for open O_TRUNC
1527 * - in setattr, for truncate
1529 /* the truncate is for size > 0 so triggers a restore */
1530 if (cl_io_is_trunc(io)) {
1531 io->ci_restore_needed = 1;
1539 io->ci_restore_needed = 1;
1545 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
1546 atomic_inc(&lov->lo_active_ios);
1549 io->ci_result = result < 0 ? result : 0;
1554 * Return the index in composite:lo_entries by the file offset
1556 int lov_io_layout_at(struct lov_io *lio, __u64 offset)
1558 struct lov_object *lov = lio->lis_object;
1559 struct lov_layout_composite *comp = &lov->u.composite;
1560 int start_index = 0;
1561 int end_index = comp->lo_entry_count - 1;
1564 LASSERT(lov->lo_type == LLT_COMP);
1566 /* This is actual file offset so nothing can cover eof. */
1567 if (offset == LUSTRE_EOF)
1570 if (lov_is_flr(lov)) {
1571 struct lov_mirror_entry *lre;
1573 LASSERT(lio->lis_mirror_index >= 0);
1575 lre = &comp->lo_mirrors[lio->lis_mirror_index];
1576 start_index = lre->lre_start;
1577 end_index = lre->lre_end;
1580 for (i = start_index; i <= end_index; i++) {
1581 struct lov_layout_entry *lle = lov_entry(lov, i);
1583 if ((offset >= lle->lle_extent->e_start &&
1584 offset < lle->lle_extent->e_end) ||
1585 (offset == OBD_OBJECT_EOF &&
1586 lle->lle_extent->e_end == OBD_OBJECT_EOF))