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, 2016, 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 != NULL && !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_pio = io->ci_pio;
140 sub_io->ci_lock_no_expand = io->ci_lock_no_expand;
141 sub_io->ci_ndelay = io->ci_ndelay;
142 sub_io->ci_layout_version = io->ci_layout_version;
144 result = cl_io_sub_init(sub->sub_env, sub_io, io->ci_type, sub_obj);
147 lov_io_sub_fini(env, lio, sub);
152 struct lov_io_sub *lov_sub_get(const struct lu_env *env,
153 struct lov_io *lio, int index)
155 struct lov_io_sub *sub;
160 list_for_each_entry(sub, &lio->lis_subios, sub_list) {
161 if (sub->sub_subio_index == index) {
168 sub = lov_sub_alloc(lio, index);
170 GOTO(out, rc = -ENOMEM);
172 rc = lov_io_sub_init(env, lio, sub);
174 lov_sub_free(lio, sub);
178 list_add_tail(&sub->sub_list, &lio->lis_subios);
179 lio->lis_nr_subios++;
187 /*****************************************************************************
193 int lov_page_index(const struct cl_page *page)
195 const struct cl_page_slice *slice;
198 slice = cl_page_at(page, &lov_device_type);
199 LASSERT(slice != NULL);
200 LASSERT(slice->cpl_obj != NULL);
202 RETURN(cl2lov_page(slice)->lps_index);
205 static int lov_io_subio_init(const struct lu_env *env, struct lov_io *lio,
210 LASSERT(lio->lis_object != NULL);
212 INIT_LIST_HEAD(&lio->lis_subios);
213 lio->lis_single_subio_index = -1;
214 lio->lis_nr_subios = 0;
220 * Decide if it will need write intent RPC
222 static int lov_io_mirror_write_intent(struct lov_io *lio,
223 struct lov_object *obj, struct cl_io *io)
225 struct lov_layout_composite *comp = &obj->u.composite;
226 struct lu_extent *ext = &io->ci_write_intent;
227 struct lov_mirror_entry *lre;
228 struct lov_mirror_entry *primary;
229 struct lov_layout_entry *lle;
233 *ext = (typeof(*ext)) { lio->lis_pos, lio->lis_endpos };
234 io->ci_need_write_intent = 0;
236 if (!(io->ci_type == CIT_WRITE || cl_io_is_trunc(io) ||
237 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);
258 /* need to iterate all components to see if there are
259 * multiple components covering the writing component */
260 primary = &comp->lo_mirrors[comp->lo_preferred_mirror];
261 LASSERT(!primary->lre_stale);
262 lov_foreach_mirror_layout_entry(obj, lle, primary) {
263 LASSERT(lle->lle_valid);
264 if (!lu_extent_is_overlapped(ext, lle->lle_extent))
267 ext->e_start = MIN(ext->e_start, lle->lle_extent->e_start);
268 ext->e_end = MAX(ext->e_end, lle->lle_extent->e_end);
272 CERROR(DFID ": cannot find any valid components covering "
273 "file extent "DEXT", mirror: %d\n",
274 PFID(lu_object_fid(lov2lu(obj))), PEXT(ext),
275 primary->lre_mirror_id);
280 lov_foreach_mirror_entry(obj, lre) {
284 lov_foreach_mirror_layout_entry(obj, lle, lre) {
288 if (lu_extent_is_overlapped(ext, lle->lle_extent)) {
295 CDEBUG(D_VFSTRACE, DFID "there are %zd components to be staled to "
296 "modify file extent "DEXT", iot: %d\n",
297 PFID(lu_object_fid(lov2lu(obj))), count, PEXT(ext), io->ci_type);
299 io->ci_need_write_intent = count > 0;
304 static int lov_io_mirror_init(struct lov_io *lio, struct lov_object *obj,
307 struct lov_layout_composite *comp = &obj->u.composite;
313 if (!lov_is_flr(obj)) {
314 LASSERT(comp->lo_preferred_mirror == 0);
315 lio->lis_mirror_index = comp->lo_preferred_mirror;
320 /* transfer the layout version for verification */
321 if (io->ci_layout_version == 0)
322 io->ci_layout_version = obj->lo_lsm->lsm_layout_gen;
324 /* find the corresponding mirror for designated mirror IO */
325 if (io->ci_designated_mirror > 0) {
326 struct lov_mirror_entry *entry;
328 LASSERT(!io->ci_ndelay);
330 CDEBUG(D_LAYOUT, "designated I/O mirror state: %d\n",
333 if ((cl_io_is_trunc(io) || io->ci_type == CIT_WRITE) &&
334 (io->ci_layout_version != obj->lo_lsm->lsm_layout_gen)) {
335 /* For resync I/O, the ci_layout_version was the layout
336 * version when resync starts. If it doesn't match the
337 * current object layout version, it means the layout
338 * has been changed */
342 io->ci_layout_version |= LU_LAYOUT_RESYNC;
345 lio->lis_mirror_index = -1;
346 lov_foreach_mirror_entry(obj, entry) {
347 if (entry->lre_mirror_id ==
348 io->ci_designated_mirror) {
349 lio->lis_mirror_index = index;
356 RETURN(lio->lis_mirror_index < 0 ? -EINVAL : 0);
359 result = lov_io_mirror_write_intent(lio, obj, io);
363 if (io->ci_need_write_intent) {
364 CDEBUG(D_VFSTRACE, DFID " need write intent for [%llu, %llu)\n",
365 PFID(lu_object_fid(lov2lu(obj))),
366 lio->lis_pos, lio->lis_endpos);
368 /* stop cl_io_init() loop */
372 if (io->ci_ndelay_tried == 0 || /* first time to try */
373 /* reset the mirror index if layout has changed */
374 lio->lis_mirror_layout_gen != obj->lo_lsm->lsm_layout_gen) {
375 lio->lis_mirror_layout_gen = obj->lo_lsm->lsm_layout_gen;
376 index = lio->lis_mirror_index = comp->lo_preferred_mirror;
378 index = lio->lis_mirror_index;
381 /* move mirror index to the next one */
382 index = (index + 1) % comp->lo_mirror_count;
385 for (i = 0; i < comp->lo_mirror_count; i++) {
386 struct lu_extent ext = { .e_start = lio->lis_pos,
387 .e_end = lio->lis_pos + 1 };
388 struct lov_mirror_entry *lre;
389 struct lov_layout_entry *lle;
392 lre = &comp->lo_mirrors[(index + i) % comp->lo_mirror_count];
396 lov_foreach_mirror_layout_entry(obj, lle, lre) {
400 if (lu_extent_is_overlapped(&ext, lle->lle_extent)) {
407 index = (index + i) % comp->lo_mirror_count;
411 if (i == comp->lo_mirror_count) {
412 CERROR(DFID": failed to find a component covering "
413 "I/O region at %llu\n",
414 PFID(lu_object_fid(lov2lu(obj))), lio->lis_pos);
416 dump_lsm(D_ERROR, obj->lo_lsm);
421 CDEBUG(D_VFSTRACE, DFID ": flr state: %d, move mirror from %d to %d, "
422 "have retried: %d, mirror count: %d\n",
423 PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj),
424 lio->lis_mirror_index, index, io->ci_ndelay_tried,
425 comp->lo_mirror_count);
427 lio->lis_mirror_index = index;
429 /* FLR: if all mirrors have been tried once, most likely the network
430 * of this client has been partitioned. We should relinquish CPU for
431 * a while before trying again. */
432 ++io->ci_ndelay_tried;
433 if (io->ci_ndelay && io->ci_ndelay_tried >= comp->lo_mirror_count) {
434 set_current_state(TASK_INTERRUPTIBLE);
435 schedule_timeout(msecs_to_jiffies(MSEC_PER_SEC)); /* 10ms */
436 if (signal_pending(current))
439 /* reset retry counter */
440 io->ci_ndelay_tried = 1;
443 CDEBUG(D_VFSTRACE, "use %sdelayed RPC state for this IO\n",
444 io->ci_ndelay ? "non-" : "");
449 static int lov_io_slice_init(struct lov_io *lio,
450 struct lov_object *obj, struct cl_io *io)
452 struct lu_extent ext;
458 lio->lis_object = obj;
460 LASSERT(obj->lo_lsm != NULL);
462 switch (io->ci_type) {
465 lio->lis_pos = io->u.ci_rw.rw_range.cir_pos;
466 lio->lis_endpos = lio->lis_pos + io->u.ci_rw.rw_range.cir_count;
467 lio->lis_io_endpos = lio->lis_endpos;
468 if (cl_io_is_append(io)) {
469 LASSERT(io->ci_type == CIT_WRITE);
471 /* If there is LOV EA hole, then we may cannot locate
472 * the current file-tail exactly. */
473 if (unlikely(obj->lo_lsm->lsm_entries[0]->lsme_pattern &
475 GOTO(out, result = -EIO);
478 lio->lis_endpos = OBD_OBJECT_EOF;
483 if (cl_io_is_trunc(io))
484 lio->lis_pos = io->u.ci_setattr.sa_attr.lvb_size;
487 lio->lis_endpos = OBD_OBJECT_EOF;
490 case CIT_DATA_VERSION:
492 lio->lis_endpos = OBD_OBJECT_EOF;
496 pgoff_t index = io->u.ci_fault.ft_index;
497 lio->lis_pos = cl_offset(io->ci_obj, index);
498 lio->lis_endpos = cl_offset(io->ci_obj, index + 1);
503 lio->lis_pos = io->u.ci_fsync.fi_start;
504 lio->lis_endpos = io->u.ci_fsync.fi_end;
509 lio->lis_pos = io->u.ci_ladvise.li_start;
510 lio->lis_endpos = io->u.ci_ladvise.li_end;
516 lio->lis_endpos = OBD_OBJECT_EOF;
518 if (lov_flr_state(obj) == LCM_FL_RDONLY &&
519 !OBD_FAIL_CHECK(OBD_FAIL_FLR_GLIMPSE_IMMUTABLE))
520 /* SoM is accurate, no need glimpse */
521 GOTO(out, result = 1);
526 lio->lis_endpos = OBD_OBJECT_EOF;
533 result = lov_io_mirror_init(lio, obj, io);
537 /* check if it needs to instantiate layout */
538 if (!(io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io) ||
539 (cl_io_is_trunc(io) && io->u.ci_setattr.sa_attr.lvb_size > 0)))
540 GOTO(out, result = 0);
542 io->ci_write_intent.e_start = lio->lis_pos;
543 io->ci_write_intent.e_end = lio->lis_endpos;
545 ext = io->ci_write_intent;
546 /* for truncate, it only needs to instantiate the components
547 * before the truncated size. */
548 if (cl_io_is_trunc(io)) {
550 ext.e_end = io->u.ci_setattr.sa_attr.lvb_size;
554 lov_foreach_io_layout(index, lio, &ext) {
555 if (!lsm_entry_inited(obj->lo_lsm, index)) {
556 io->ci_need_write_intent = 1;
561 if (io->ci_need_write_intent && io->ci_designated_mirror > 0) {
562 /* REINT_SYNC RPC has already tried to instantiate all of the
563 * components involved, obviously it didn't succeed. Skip this
564 * mirror for now. The server won't be able to figure out
565 * which mirror it should instantiate components */
566 CERROR(DFID": trying to instantiate components for designated "
567 "I/O, file state: %d\n",
568 PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj));
570 io->ci_need_write_intent = 0;
571 GOTO(out, result = -EIO);
574 if (io->ci_need_write_intent)
575 GOTO(out, result = 1);
583 static void lov_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
585 struct lov_io *lio = cl2lov_io(env, ios);
586 struct lov_object *lov = cl2lov(ios->cis_obj);
590 LASSERT(list_empty(&lio->lis_active));
592 while (!list_empty(&lio->lis_subios)) {
593 struct lov_io_sub *sub = list_entry(lio->lis_subios.next,
597 list_del_init(&sub->sub_list);
598 lio->lis_nr_subios--;
600 lov_io_sub_fini(env, lio, sub);
601 lov_sub_free(lio, sub);
603 LASSERT(lio->lis_nr_subios == 0);
605 LASSERT(atomic_read(&lov->lo_active_ios) > 0);
606 if (atomic_dec_and_test(&lov->lo_active_ios))
607 wake_up_all(&lov->lo_waitq);
611 static void lov_io_sub_inherit(struct lov_io_sub *sub, struct lov_io *lio,
612 loff_t start, loff_t end)
614 struct cl_io *io = &sub->sub_io;
615 struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
616 struct cl_io *parent = lio->lis_cl.cis_io;
617 int index = lov_comp_entry(sub->sub_subio_index);
618 int stripe = lov_comp_stripe(sub->sub_subio_index);
620 io->ci_pio = parent->ci_pio;
621 switch (io->ci_type) {
623 io->u.ci_setattr.sa_attr = parent->u.ci_setattr.sa_attr;
624 io->u.ci_setattr.sa_attr_flags =
625 parent->u.ci_setattr.sa_attr_flags;
626 io->u.ci_setattr.sa_valid = parent->u.ci_setattr.sa_valid;
627 io->u.ci_setattr.sa_stripe_index = stripe;
628 io->u.ci_setattr.sa_parent_fid =
629 parent->u.ci_setattr.sa_parent_fid;
630 if (cl_io_is_trunc(io)) {
631 loff_t new_size = parent->u.ci_setattr.sa_attr.lvb_size;
633 new_size = lov_size_to_stripe(lsm, index, new_size,
635 io->u.ci_setattr.sa_attr.lvb_size = new_size;
637 lov_lsm2layout(lsm, lsm->lsm_entries[index],
638 &io->u.ci_setattr.sa_layout);
641 case CIT_DATA_VERSION: {
642 io->u.ci_data_version.dv_data_version = 0;
643 io->u.ci_data_version.dv_flags =
644 parent->u.ci_data_version.dv_flags;
648 struct cl_object *obj = parent->ci_obj;
649 loff_t off = cl_offset(obj, parent->u.ci_fault.ft_index);
651 io->u.ci_fault = parent->u.ci_fault;
652 off = lov_size_to_stripe(lsm, index, off, stripe);
653 io->u.ci_fault.ft_index = cl_index(obj, off);
657 io->u.ci_fsync.fi_start = start;
658 io->u.ci_fsync.fi_end = end;
659 io->u.ci_fsync.fi_fid = parent->u.ci_fsync.fi_fid;
660 io->u.ci_fsync.fi_mode = parent->u.ci_fsync.fi_mode;
665 io->u.ci_rw.rw_ptask = parent->u.ci_rw.rw_ptask;
666 io->u.ci_rw.rw_iter = parent->u.ci_rw.rw_iter;
667 io->u.ci_rw.rw_iocb = parent->u.ci_rw.rw_iocb;
668 io->u.ci_rw.rw_file = parent->u.ci_rw.rw_file;
669 io->u.ci_rw.rw_sync = parent->u.ci_rw.rw_sync;
670 if (cl_io_is_append(parent)) {
671 io->u.ci_rw.rw_append = 1;
673 io->u.ci_rw.rw_range.cir_pos = start;
674 io->u.ci_rw.rw_range.cir_count = end - start;
679 io->u.ci_ladvise.li_start = start;
680 io->u.ci_ladvise.li_end = end;
681 io->u.ci_ladvise.li_fid = parent->u.ci_ladvise.li_fid;
682 io->u.ci_ladvise.li_advice = parent->u.ci_ladvise.li_advice;
683 io->u.ci_ladvise.li_flags = parent->u.ci_ladvise.li_flags;
693 static loff_t lov_offset_mod(loff_t val, int delta)
695 if (val != OBD_OBJECT_EOF)
700 static int lov_io_iter_init(const struct lu_env *env,
701 const struct cl_io_slice *ios)
703 struct lov_io *lio = cl2lov_io(env, ios);
704 struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
705 struct lov_io_sub *sub;
706 struct lu_extent ext;
712 ext.e_start = lio->lis_pos;
713 ext.e_end = lio->lis_endpos;
715 lov_foreach_io_layout(index, lio, &ext) {
716 struct lov_layout_entry *le = lov_entry(lio->lis_object, index);
717 struct lov_layout_raid0 *r0 = &le->lle_raid0;
722 CDEBUG(D_VFSTRACE, "component[%d] flags %#x\n",
723 index, lsm->lsm_entries[index]->lsme_flags);
724 if (!lsm_entry_inited(lsm, index)) {
725 /* Read from uninitialized components should return
726 * zero filled pages. */
730 if (!le->lle_valid && !ios->cis_io->ci_designated_mirror) {
731 CERROR("I/O to invalid component: %d, mirror: %d\n",
732 index, lio->lis_mirror_index);
736 for (stripe = 0; stripe < r0->lo_nr; stripe++) {
737 if (!lov_stripe_intersects(lsm, index, stripe,
741 if (unlikely(r0->lo_sub[stripe] == NULL)) {
742 if (ios->cis_io->ci_type == CIT_READ ||
743 ios->cis_io->ci_type == CIT_WRITE ||
744 ios->cis_io->ci_type == CIT_FAULT)
750 end = lov_offset_mod(end, 1);
751 sub = lov_sub_get(env, lio,
752 lov_comp_index(index, stripe));
758 lov_io_sub_inherit(sub, lio, start, end);
759 rc = cl_io_iter_init(sub->sub_env, &sub->sub_io);
761 cl_io_iter_fini(sub->sub_env, &sub->sub_io);
766 "shrink stripe: {%d, %d} range: [%llu, %llu)\n",
767 index, stripe, start, end);
769 list_add_tail(&sub->sub_linkage, &lio->lis_active);
777 static int lov_io_rw_iter_init(const struct lu_env *env,
778 const struct cl_io_slice *ios)
780 struct cl_io *io = ios->cis_io;
781 struct lov_io *lio = cl2lov_io(env, ios);
782 struct lov_stripe_md_entry *lse;
783 struct cl_io_range *range = &io->u.ci_rw.rw_range;
784 loff_t start = range->cir_pos;
788 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
791 if (cl_io_is_append(io))
792 RETURN(lov_io_iter_init(env, ios));
794 index = lov_io_layout_at(lio, range->cir_pos);
795 if (index < 0) { /* non-existing layout component */
796 if (io->ci_type == CIT_READ) {
797 /* TODO: it needs to detect the next component and
798 * then set the next pos */
800 /* execute it in main thread */
803 RETURN(lov_io_iter_init(env, ios));
809 if (!lov_entry(lio->lis_object, index)->lle_valid &&
810 !io->ci_designated_mirror)
811 RETURN(io->ci_type == CIT_READ ? -EAGAIN : -EIO);
813 lse = lov_lse(lio->lis_object, index);
815 next = MAX_LFS_FILESIZE;
816 if (lse->lsme_stripe_count > 1) {
817 unsigned long ssize = lse->lsme_stripe_size;
819 lov_do_div64(start, ssize);
820 next = (start + 1) * ssize;
821 if (next <= start * ssize)
822 next = MAX_LFS_FILESIZE;
825 LASSERTF(range->cir_pos >= lse->lsme_extent.e_start,
826 "pos %lld, [%lld, %lld)\n", range->cir_pos,
827 lse->lsme_extent.e_start, lse->lsme_extent.e_end);
828 next = min_t(__u64, next, lse->lsme_extent.e_end);
829 next = min_t(loff_t, next, lio->lis_io_endpos);
831 io->ci_continue = next < lio->lis_io_endpos;
832 range->cir_count = next - range->cir_pos;
833 lio->lis_pos = range->cir_pos;
834 lio->lis_endpos = range->cir_pos + range->cir_count;
836 "stripe: {%d, %llu} range: [%llu, %llu) end: %llu, count: %zd\n",
837 index, start, lio->lis_pos, lio->lis_endpos,
838 lio->lis_io_endpos, range->cir_count);
840 if (!io->ci_continue) {
841 /* the last piece of IO, execute it in main thread */
849 * XXX The following call should be optimized: we know, that
850 * [lio->lis_pos, lio->lis_endpos) intersects with exactly one stripe.
852 RETURN(lov_io_iter_init(env, ios));
855 static int lov_io_setattr_iter_init(const struct lu_env *env,
856 const struct cl_io_slice *ios)
858 struct lov_io *lio = cl2lov_io(env, ios);
859 struct cl_io *io = ios->cis_io;
863 if (cl_io_is_trunc(io) && lio->lis_pos > 0) {
864 index = lov_io_layout_at(lio, lio->lis_pos - 1);
865 /* no entry found for such offset */
867 RETURN(io->ci_result = -ENODATA);
870 RETURN(lov_io_iter_init(env, ios));
873 static int lov_io_call(const struct lu_env *env, struct lov_io *lio,
874 int (*iofunc)(const struct lu_env *, struct cl_io *))
876 struct cl_io *parent = lio->lis_cl.cis_io;
877 struct lov_io_sub *sub;
881 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
882 rc = iofunc(sub->sub_env, &sub->sub_io);
886 if (parent->ci_result == 0)
887 parent->ci_result = sub->sub_io.ci_result;
892 static int lov_io_lock(const struct lu_env *env, const struct cl_io_slice *ios)
895 RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_lock));
898 static int lov_io_start(const struct lu_env *env, const struct cl_io_slice *ios)
901 RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_start));
904 static int lov_io_end_wrapper(const struct lu_env *env, struct cl_io *io)
908 * It's possible that lov_io_start() wasn't called against this
909 * sub-io, either because previous sub-io failed, or upper layer
912 if (io->ci_state == CIS_IO_GOING)
915 io->ci_state = CIS_IO_FINISHED;
919 static int lov_io_iter_fini_wrapper(const struct lu_env *env, struct cl_io *io)
921 cl_io_iter_fini(env, io);
925 static int lov_io_unlock_wrapper(const struct lu_env *env, struct cl_io *io)
927 cl_io_unlock(env, io);
931 static void lov_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
935 rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_end_wrapper);
940 lov_io_data_version_end(const struct lu_env *env, const struct cl_io_slice *ios)
942 struct lov_io *lio = cl2lov_io(env, ios);
943 struct cl_io *parent = lio->lis_cl.cis_io;
944 struct cl_data_version_io *pdv = &parent->u.ci_data_version;
945 struct lov_io_sub *sub;
948 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
949 struct cl_data_version_io *sdv = &sub->sub_io.u.ci_data_version;
951 lov_io_end_wrapper(env, &sub->sub_io);
953 pdv->dv_data_version += sdv->dv_data_version;
954 if (pdv->dv_layout_version > sdv->dv_layout_version)
955 pdv->dv_layout_version = sdv->dv_layout_version;
957 if (parent->ci_result == 0)
958 parent->ci_result = sub->sub_io.ci_result;
964 static void lov_io_iter_fini(const struct lu_env *env,
965 const struct cl_io_slice *ios)
967 struct lov_io *lio = cl2lov_io(env, ios);
971 rc = lov_io_call(env, lio, lov_io_iter_fini_wrapper);
973 while (!list_empty(&lio->lis_active))
974 list_del_init(lio->lis_active.next);
978 static void lov_io_unlock(const struct lu_env *env,
979 const struct cl_io_slice *ios)
984 rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_unlock_wrapper);
989 static int lov_io_read_ahead(const struct lu_env *env,
990 const struct cl_io_slice *ios,
991 pgoff_t start, struct cl_read_ahead *ra)
993 struct lov_io *lio = cl2lov_io(env, ios);
994 struct lov_object *loo = lio->lis_object;
995 struct cl_object *obj = lov2cl(loo);
996 struct lov_layout_raid0 *r0;
997 struct lov_io_sub *sub;
1001 unsigned int pps; /* pages per stripe */
1007 offset = cl_offset(obj, start);
1008 index = lov_io_layout_at(lio, offset);
1009 if (index < 0 || !lsm_entry_inited(loo->lo_lsm, index))
1012 /* avoid readahead to expand to stale components */
1013 if (!lov_entry(loo, index)->lle_valid)
1016 stripe = lov_stripe_number(loo->lo_lsm, index, offset);
1018 r0 = lov_r0(loo, index);
1019 if (unlikely(r0->lo_sub[stripe] == NULL))
1022 sub = lov_sub_get(env, lio, lov_comp_index(index, stripe));
1024 RETURN(PTR_ERR(sub));
1026 lov_stripe_offset(loo->lo_lsm, index, offset, stripe, &suboff);
1027 rc = cl_io_read_ahead(sub->sub_env, &sub->sub_io,
1028 cl_index(lovsub2cl(r0->lo_sub[stripe]), suboff),
1031 CDEBUG(D_READA, DFID " cra_end = %lu, stripes = %d, rc = %d\n",
1032 PFID(lu_object_fid(lov2lu(loo))), ra->cra_end, r0->lo_nr, rc);
1037 * Adjust the stripe index by layout of comp. ra->cra_end is the
1038 * maximum page index covered by an underlying DLM lock.
1039 * This function converts cra_end from stripe level to file level, and
1040 * make sure it's not beyond stripe and component boundary.
1043 /* cra_end is stripe level, convert it into file level */
1044 ra_end = ra->cra_end;
1045 if (ra_end != CL_PAGE_EOF)
1046 ra->cra_end = lov_stripe_pgoff(loo->lo_lsm, index,
1049 /* boundary of current component */
1050 ra_end = cl_index(obj, (loff_t)lov_io_extent(lio, index)->e_end);
1051 if (ra_end != CL_PAGE_EOF && ra->cra_end >= ra_end)
1052 ra->cra_end = ra_end - 1;
1054 if (r0->lo_nr == 1) /* single stripe file */
1057 pps = lov_lse(loo, index)->lsme_stripe_size >> PAGE_SHIFT;
1059 CDEBUG(D_READA, DFID " max_index = %lu, pps = %u, index = %u, "
1060 "stripe_size = %u, stripe no = %u, start index = %lu\n",
1061 PFID(lu_object_fid(lov2lu(loo))), ra->cra_end, pps, index,
1062 lov_lse(loo, index)->lsme_stripe_size, stripe, start);
1064 /* never exceed the end of the stripe */
1065 ra->cra_end = min_t(pgoff_t,
1066 ra->cra_end, start + pps - start % pps - 1);
1071 * lov implementation of cl_operations::cio_submit() method. It takes a list
1072 * of pages in \a queue, splits it into per-stripe sub-lists, invokes
1073 * cl_io_submit() on underlying devices to submit sub-lists, and then splices
1076 * Major complication of this function is a need to handle memory cleansing:
1077 * cl_io_submit() is called to write out pages as a part of VM memory
1078 * reclamation, and hence it may not fail due to memory shortages (system
1079 * dead-locks otherwise). To deal with this, some resources (sub-lists,
1080 * sub-environment, etc.) are allocated per-device on "startup" (i.e., in a
1081 * not-memory cleansing context), and in case of memory shortage, these
1082 * pre-allocated resources are used by lov_io_submit() under
1083 * lov_device::ld_mutex mutex.
1085 static int lov_io_submit(const struct lu_env *env,
1086 const struct cl_io_slice *ios,
1087 enum cl_req_type crt, struct cl_2queue *queue)
1089 struct cl_page_list *qin = &queue->c2_qin;
1090 struct lov_io *lio = cl2lov_io(env, ios);
1091 struct lov_io_sub *sub;
1092 struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
1093 struct cl_page *page;
1098 cl_page_list_init(plist);
1099 while (qin->pl_nr > 0) {
1100 struct cl_2queue *cl2q = &lov_env_info(env)->lti_cl2q;
1102 page = cl_page_list_first(qin);
1103 if (lov_page_is_empty(page)) {
1104 cl_page_list_move(&queue->c2_qout, qin, page);
1106 /* it could only be mirror read to get here therefore
1107 * the pages will be transient. We don't care about
1108 * the return code of cl_page_prep() at all. */
1109 (void) cl_page_prep(env, ios->cis_io, page, crt);
1110 cl_page_completion(env, page, crt, 0);
1114 cl_2queue_init(cl2q);
1115 cl_page_list_move(&cl2q->c2_qin, qin, page);
1117 index = lov_page_index(page);
1118 while (qin->pl_nr > 0) {
1119 page = cl_page_list_first(qin);
1120 if (index != lov_page_index(page))
1123 cl_page_list_move(&cl2q->c2_qin, qin, page);
1126 sub = lov_sub_get(env, lio, index);
1128 rc = cl_io_submit_rw(sub->sub_env, &sub->sub_io,
1134 cl_page_list_splice(&cl2q->c2_qin, plist);
1135 cl_page_list_splice(&cl2q->c2_qout, &queue->c2_qout);
1136 cl_2queue_fini(env, cl2q);
1142 cl_page_list_splice(plist, qin);
1143 cl_page_list_fini(env, plist);
1148 static int lov_io_commit_async(const struct lu_env *env,
1149 const struct cl_io_slice *ios,
1150 struct cl_page_list *queue, int from, int to,
1153 struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
1154 struct lov_io *lio = cl2lov_io(env, ios);
1155 struct lov_io_sub *sub;
1156 struct cl_page *page;
1160 if (lio->lis_nr_subios == 1) {
1161 int idx = lio->lis_single_subio_index;
1163 LASSERT(!lov_page_is_empty(cl_page_list_first(queue)));
1165 sub = lov_sub_get(env, lio, idx);
1166 LASSERT(!IS_ERR(sub));
1167 LASSERT(sub == &lio->lis_single_subio);
1168 rc = cl_io_commit_async(sub->sub_env, &sub->sub_io, queue,
1173 cl_page_list_init(plist);
1174 while (queue->pl_nr > 0) {
1178 LASSERT(plist->pl_nr == 0);
1179 page = cl_page_list_first(queue);
1180 LASSERT(!lov_page_is_empty(page));
1182 cl_page_list_move(plist, queue, page);
1184 index = lov_page_index(page);
1185 while (queue->pl_nr > 0) {
1186 page = cl_page_list_first(queue);
1187 if (index != lov_page_index(page))
1190 cl_page_list_move(plist, queue, page);
1193 if (queue->pl_nr > 0) /* still has more pages */
1194 stripe_to = PAGE_SIZE;
1196 sub = lov_sub_get(env, lio, index);
1198 rc = cl_io_commit_async(sub->sub_env, &sub->sub_io,
1199 plist, from, stripe_to, cb);
1205 if (plist->pl_nr > 0) /* short write */
1211 /* for error case, add the page back into the qin list */
1212 LASSERT(ergo(rc == 0, plist->pl_nr == 0));
1213 while (plist->pl_nr > 0) {
1214 /* error occurred, add the uncommitted pages back into queue */
1215 page = cl_page_list_last(plist);
1216 cl_page_list_move_head(queue, plist, page);
1222 static int lov_io_fault_start(const struct lu_env *env,
1223 const struct cl_io_slice *ios)
1225 struct cl_fault_io *fio;
1227 struct lov_io_sub *sub;
1231 fio = &ios->cis_io->u.ci_fault;
1232 lio = cl2lov_io(env, ios);
1233 sub = lov_sub_get(env, lio, lov_page_index(fio->ft_page));
1234 sub->sub_io.u.ci_fault.ft_nob = fio->ft_nob;
1236 RETURN(lov_io_start(env, ios));
1239 static void lov_io_fsync_end(const struct lu_env *env,
1240 const struct cl_io_slice *ios)
1242 struct lov_io *lio = cl2lov_io(env, ios);
1243 struct lov_io_sub *sub;
1244 unsigned int *written = &ios->cis_io->u.ci_fsync.fi_nr_written;
1248 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
1249 struct cl_io *subio = &sub->sub_io;
1251 lov_io_end_wrapper(sub->sub_env, subio);
1253 if (subio->ci_result == 0)
1254 *written += subio->u.ci_fsync.fi_nr_written;
1259 static const struct cl_io_operations lov_io_ops = {
1262 .cio_fini = lov_io_fini,
1263 .cio_iter_init = lov_io_rw_iter_init,
1264 .cio_iter_fini = lov_io_iter_fini,
1265 .cio_lock = lov_io_lock,
1266 .cio_unlock = lov_io_unlock,
1267 .cio_start = lov_io_start,
1268 .cio_end = lov_io_end
1271 .cio_fini = lov_io_fini,
1272 .cio_iter_init = lov_io_rw_iter_init,
1273 .cio_iter_fini = lov_io_iter_fini,
1274 .cio_lock = lov_io_lock,
1275 .cio_unlock = lov_io_unlock,
1276 .cio_start = lov_io_start,
1277 .cio_end = lov_io_end
1280 .cio_fini = lov_io_fini,
1281 .cio_iter_init = lov_io_setattr_iter_init,
1282 .cio_iter_fini = lov_io_iter_fini,
1283 .cio_lock = lov_io_lock,
1284 .cio_unlock = lov_io_unlock,
1285 .cio_start = lov_io_start,
1286 .cio_end = lov_io_end
1288 [CIT_DATA_VERSION] = {
1289 .cio_fini = lov_io_fini,
1290 .cio_iter_init = lov_io_iter_init,
1291 .cio_iter_fini = lov_io_iter_fini,
1292 .cio_lock = lov_io_lock,
1293 .cio_unlock = lov_io_unlock,
1294 .cio_start = lov_io_start,
1295 .cio_end = lov_io_data_version_end,
1298 .cio_fini = lov_io_fini,
1299 .cio_iter_init = lov_io_iter_init,
1300 .cio_iter_fini = lov_io_iter_fini,
1301 .cio_lock = lov_io_lock,
1302 .cio_unlock = lov_io_unlock,
1303 .cio_start = lov_io_fault_start,
1304 .cio_end = lov_io_end
1307 .cio_fini = lov_io_fini,
1308 .cio_iter_init = lov_io_iter_init,
1309 .cio_iter_fini = lov_io_iter_fini,
1310 .cio_lock = lov_io_lock,
1311 .cio_unlock = lov_io_unlock,
1312 .cio_start = lov_io_start,
1313 .cio_end = lov_io_fsync_end
1316 .cio_fini = lov_io_fini,
1317 .cio_iter_init = lov_io_iter_init,
1318 .cio_iter_fini = lov_io_iter_fini,
1319 .cio_lock = lov_io_lock,
1320 .cio_unlock = lov_io_unlock,
1321 .cio_start = lov_io_start,
1322 .cio_end = lov_io_end
1325 .cio_fini = lov_io_fini,
1328 .cio_fini = lov_io_fini
1331 .cio_read_ahead = lov_io_read_ahead,
1332 .cio_submit = lov_io_submit,
1333 .cio_commit_async = lov_io_commit_async,
1336 /*****************************************************************************
1338 * Empty lov io operations.
1342 static void lov_empty_io_fini(const struct lu_env *env,
1343 const struct cl_io_slice *ios)
1345 struct lov_object *lov = cl2lov(ios->cis_obj);
1348 if (atomic_dec_and_test(&lov->lo_active_ios))
1349 wake_up_all(&lov->lo_waitq);
1353 static int lov_empty_io_submit(const struct lu_env *env,
1354 const struct cl_io_slice *ios,
1355 enum cl_req_type crt, struct cl_2queue *queue)
1360 static void lov_empty_impossible(const struct lu_env *env,
1361 struct cl_io_slice *ios)
1366 #define LOV_EMPTY_IMPOSSIBLE ((void *)lov_empty_impossible)
1369 * An io operation vector for files without stripes.
1371 static const struct cl_io_operations lov_empty_io_ops = {
1374 .cio_fini = lov_empty_io_fini,
1376 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1377 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1378 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1379 .cio_end = LOV_EMPTY_IMPOSSIBLE
1383 .cio_fini = lov_empty_io_fini,
1384 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1385 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1386 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1387 .cio_end = LOV_EMPTY_IMPOSSIBLE
1390 .cio_fini = lov_empty_io_fini,
1391 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1392 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1393 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1394 .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
1407 .cio_fini = lov_empty_io_fini
1410 .cio_fini = lov_empty_io_fini
1413 .cio_fini = lov_empty_io_fini
1416 .cio_submit = lov_empty_io_submit,
1417 .cio_commit_async = LOV_EMPTY_IMPOSSIBLE
1420 int lov_io_init_composite(const struct lu_env *env, struct cl_object *obj,
1423 struct lov_io *lio = lov_env_io(env);
1424 struct lov_object *lov = cl2lov(obj);
1428 INIT_LIST_HEAD(&lio->lis_active);
1429 result = lov_io_slice_init(lio, lov, io);
1433 result = lov_io_subio_init(env, lio, io);
1435 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_io_ops);
1436 atomic_inc(&lov->lo_active_ios);
1440 io->ci_result = result < 0 ? result : 0;
1444 int lov_io_init_empty(const struct lu_env *env, struct cl_object *obj,
1447 struct lov_object *lov = cl2lov(obj);
1448 struct lov_io *lio = lov_env_io(env);
1452 lio->lis_object = lov;
1453 switch (io->ci_type) {
1464 case CIT_DATA_VERSION:
1472 CERROR("Page fault on a file without stripes: "DFID"\n",
1473 PFID(lu_object_fid(&obj->co_lu)));
1477 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
1478 atomic_inc(&lov->lo_active_ios);
1481 io->ci_result = result < 0 ? result : 0;
1485 int lov_io_init_released(const struct lu_env *env, struct cl_object *obj,
1488 struct lov_object *lov = cl2lov(obj);
1489 struct lov_io *lio = lov_env_io(env);
1493 LASSERT(lov->lo_lsm != NULL);
1494 lio->lis_object = lov;
1496 switch (io->ci_type) {
1498 LASSERTF(0, "invalid type %d\n", io->ci_type);
1499 result = -EOPNOTSUPP;
1505 case CIT_DATA_VERSION:
1509 /* the truncate to 0 is managed by MDT:
1510 * - in open, for open O_TRUNC
1511 * - in setattr, for truncate
1513 /* the truncate is for size > 0 so triggers a restore */
1514 if (cl_io_is_trunc(io)) {
1515 io->ci_restore_needed = 1;
1523 io->ci_restore_needed = 1;
1529 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
1530 atomic_inc(&lov->lo_active_ios);
1533 io->ci_result = result < 0 ? result : 0;
1538 * Return the index in composite:lo_entries by the file offset
1540 int lov_io_layout_at(struct lov_io *lio, __u64 offset)
1542 struct lov_object *lov = lio->lis_object;
1543 struct lov_layout_composite *comp = &lov->u.composite;
1544 int start_index = 0;
1545 int end_index = comp->lo_entry_count - 1;
1548 LASSERT(lov->lo_type == LLT_COMP);
1550 /* This is actual file offset so nothing can cover eof. */
1551 if (offset == LUSTRE_EOF)
1554 if (lov_is_flr(lov)) {
1555 struct lov_mirror_entry *lre;
1557 LASSERT(lio->lis_mirror_index >= 0);
1559 lre = &comp->lo_mirrors[lio->lis_mirror_index];
1560 start_index = lre->lre_start;
1561 end_index = lre->lre_end;
1564 for (i = start_index; i <= end_index; i++) {
1565 struct lov_layout_entry *lle = lov_entry(lov, i);
1567 if ((offset >= lle->lle_extent->e_start &&
1568 offset < lle->lle_extent->e_end) ||
1569 (offset == OBD_OBJECT_EOF &&
1570 lle->lle_extent->e_end == OBD_OBJECT_EOF))