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_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)))
241 * FLR: check if it needs to send a write intent RPC to server.
242 * Writing to sync_pending file needs write intent RPC to change
243 * the file state back to write_pending, so that the layout version
244 * can be increased when the state changes to sync_pending at a later
245 * time. Otherwise there exists a chance that an evicted client may
246 * dirty the file data while resync client is working on it.
247 * Designated I/O is allowed for resync workload.
249 if (lov_flr_state(obj) == LCM_FL_RDONLY ||
250 (lov_flr_state(obj) == LCM_FL_SYNC_PENDING &&
251 io->ci_designated_mirror == 0)) {
252 io->ci_need_write_intent = 1;
256 LASSERT((lov_flr_state(obj) == LCM_FL_WRITE_PENDING));
257 LASSERT(comp->lo_preferred_mirror >= 0);
260 * need to iterate all components to see if there are
261 * multiple components covering the writing component
263 primary = &comp->lo_mirrors[comp->lo_preferred_mirror];
264 LASSERT(!primary->lre_stale);
265 lov_foreach_mirror_layout_entry(obj, lle, primary) {
266 LASSERT(lle->lle_valid);
267 if (!lu_extent_is_overlapped(ext, lle->lle_extent))
270 ext->e_start = MIN(ext->e_start, lle->lle_extent->e_start);
271 ext->e_end = MAX(ext->e_end, lle->lle_extent->e_end);
275 CERROR(DFID ": cannot find any valid components covering "
276 "file extent "DEXT", mirror: %d\n",
277 PFID(lu_object_fid(lov2lu(obj))), PEXT(ext),
278 primary->lre_mirror_id);
283 lov_foreach_mirror_entry(obj, lre) {
287 lov_foreach_mirror_layout_entry(obj, lle, lre) {
291 if (lu_extent_is_overlapped(ext, lle->lle_extent)) {
298 CDEBUG(D_VFSTRACE, DFID "there are %zd components to be staled to "
299 "modify file extent "DEXT", iot: %d\n",
300 PFID(lu_object_fid(lov2lu(obj))), count, PEXT(ext), io->ci_type);
302 io->ci_need_write_intent = count > 0;
307 static int lov_io_mirror_init(struct lov_io *lio, struct lov_object *obj,
310 struct lov_layout_composite *comp = &obj->u.composite;
316 if (!lov_is_flr(obj)) {
317 LASSERT(comp->lo_preferred_mirror == 0);
318 lio->lis_mirror_index = comp->lo_preferred_mirror;
323 /* transfer the layout version for verification */
324 if (io->ci_layout_version == 0)
325 io->ci_layout_version = obj->lo_lsm->lsm_layout_gen;
327 /* find the corresponding mirror for designated mirror IO */
328 if (io->ci_designated_mirror > 0) {
329 struct lov_mirror_entry *entry;
331 LASSERT(!io->ci_ndelay);
333 CDEBUG(D_LAYOUT, "designated I/O mirror state: %d\n",
336 if ((cl_io_is_trunc(io) || io->ci_type == CIT_WRITE) &&
337 (io->ci_layout_version != obj->lo_lsm->lsm_layout_gen)) {
339 * For resync I/O, the ci_layout_version was the layout
340 * version when resync starts. If it doesn't match the
341 * current object layout version, it means the layout
347 io->ci_layout_version |= LU_LAYOUT_RESYNC;
350 lio->lis_mirror_index = -1;
351 lov_foreach_mirror_entry(obj, entry) {
352 if (entry->lre_mirror_id ==
353 io->ci_designated_mirror) {
354 lio->lis_mirror_index = index;
361 RETURN(lio->lis_mirror_index < 0 ? -EINVAL : 0);
364 result = lov_io_mirror_write_intent(lio, obj, io);
368 if (io->ci_need_write_intent) {
369 CDEBUG(D_VFSTRACE, DFID " need write intent for [%llu, %llu)\n",
370 PFID(lu_object_fid(lov2lu(obj))),
371 lio->lis_pos, lio->lis_endpos);
373 if (cl_io_is_trunc(io)) {
375 * for truncate, we uses [size, EOF) to judge whether
376 * a write intent needs to be send, but we need to
377 * restore the write extent to [0, size).
379 io->ci_write_intent.e_start = 0;
380 io->ci_write_intent.e_end =
381 io->u.ci_setattr.sa_attr.lvb_size;
383 /* stop cl_io_init() loop */
387 if (io->ci_ndelay_tried == 0 || /* first time to try */
388 /* reset the mirror index if layout has changed */
389 lio->lis_mirror_layout_gen != obj->lo_lsm->lsm_layout_gen) {
390 lio->lis_mirror_layout_gen = obj->lo_lsm->lsm_layout_gen;
391 index = lio->lis_mirror_index = comp->lo_preferred_mirror;
393 index = lio->lis_mirror_index;
396 /* move mirror index to the next one */
397 index = (index + 1) % comp->lo_mirror_count;
400 for (i = 0; i < comp->lo_mirror_count; i++) {
401 struct lu_extent ext = { .e_start = lio->lis_pos,
402 .e_end = lio->lis_pos + 1 };
403 struct lov_mirror_entry *lre;
404 struct lov_layout_entry *lle;
407 lre = &comp->lo_mirrors[(index + i) % comp->lo_mirror_count];
411 lov_foreach_mirror_layout_entry(obj, lle, lre) {
415 if (lu_extent_is_overlapped(&ext, lle->lle_extent)) {
422 index = (index + i) % comp->lo_mirror_count;
426 if (i == comp->lo_mirror_count) {
427 CERROR(DFID": failed to find a component covering "
428 "I/O region at %llu\n",
429 PFID(lu_object_fid(lov2lu(obj))), lio->lis_pos);
431 dump_lsm(D_ERROR, obj->lo_lsm);
436 CDEBUG(D_VFSTRACE, DFID ": flr state: %d, move mirror from %d to %d, "
437 "have retried: %d, mirror count: %d\n",
438 PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj),
439 lio->lis_mirror_index, index, io->ci_ndelay_tried,
440 comp->lo_mirror_count);
442 lio->lis_mirror_index = index;
445 * FLR: if all mirrors have been tried once, most likely the network
446 * of this client has been partitioned. We should relinquish CPU for
447 * a while before trying again.
449 ++io->ci_ndelay_tried;
450 if (io->ci_ndelay && io->ci_ndelay_tried >= comp->lo_mirror_count) {
451 set_current_state(TASK_INTERRUPTIBLE);
452 schedule_timeout(msecs_to_jiffies(MSEC_PER_SEC)); /* 10ms */
453 if (signal_pending(current))
456 /* reset retry counter */
457 io->ci_ndelay_tried = 1;
460 CDEBUG(D_VFSTRACE, "use %sdelayed RPC state for this IO\n",
461 io->ci_ndelay ? "non-" : "");
466 static int lov_io_slice_init(struct lov_io *lio,
467 struct lov_object *obj, struct cl_io *io)
474 lio->lis_object = obj;
476 LASSERT(obj->lo_lsm != NULL);
478 switch (io->ci_type) {
481 lio->lis_pos = io->u.ci_rw.rw_range.cir_pos;
482 lio->lis_endpos = lio->lis_pos + io->u.ci_rw.rw_range.cir_count;
483 lio->lis_io_endpos = lio->lis_endpos;
484 if (cl_io_is_append(io)) {
485 LASSERT(io->ci_type == CIT_WRITE);
488 * If there is LOV EA hole, then we may cannot locate
489 * the current file-tail exactly.
491 if (unlikely(obj->lo_lsm->lsm_entries[0]->lsme_pattern &
493 GOTO(out, result = -EIO);
496 lio->lis_endpos = OBD_OBJECT_EOF;
501 if (cl_io_is_trunc(io))
502 lio->lis_pos = io->u.ci_setattr.sa_attr.lvb_size;
505 lio->lis_endpos = OBD_OBJECT_EOF;
508 case CIT_DATA_VERSION:
510 lio->lis_endpos = OBD_OBJECT_EOF;
514 pgoff_t index = io->u.ci_fault.ft_index;
516 lio->lis_pos = cl_offset(io->ci_obj, index);
517 lio->lis_endpos = cl_offset(io->ci_obj, index + 1);
522 lio->lis_pos = io->u.ci_fsync.fi_start;
523 lio->lis_endpos = io->u.ci_fsync.fi_end;
528 lio->lis_pos = io->u.ci_ladvise.li_start;
529 lio->lis_endpos = io->u.ci_ladvise.li_end;
535 lio->lis_endpos = OBD_OBJECT_EOF;
537 if (lov_flr_state(obj) == LCM_FL_RDONLY &&
538 !OBD_FAIL_CHECK(OBD_FAIL_FLR_GLIMPSE_IMMUTABLE))
539 /* SoM is accurate, no need glimpse */
540 GOTO(out, result = 1);
545 lio->lis_endpos = OBD_OBJECT_EOF;
552 result = lov_io_mirror_init(lio, obj, io);
556 /* check if it needs to instantiate layout */
557 if (!(io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io) ||
558 (cl_io_is_trunc(io) && io->u.ci_setattr.sa_attr.lvb_size > 0)))
559 GOTO(out, result = 0);
562 * for truncate, it only needs to instantiate the components
563 * before the truncated size.
565 if (cl_io_is_trunc(io)) {
566 io->ci_write_intent.e_start = 0;
567 io->ci_write_intent.e_end = io->u.ci_setattr.sa_attr.lvb_size;
569 io->ci_write_intent.e_start = lio->lis_pos;
570 io->ci_write_intent.e_end = lio->lis_endpos;
574 lov_foreach_io_layout(index, lio, &io->ci_write_intent) {
575 if (!lsm_entry_inited(obj->lo_lsm, index)) {
576 io->ci_need_write_intent = 1;
581 if (io->ci_need_write_intent && io->ci_designated_mirror > 0) {
583 * REINT_SYNC RPC has already tried to instantiate all of the
584 * components involved, obviously it didn't succeed. Skip this
585 * mirror for now. The server won't be able to figure out
586 * which mirror it should instantiate components
588 CERROR(DFID": trying to instantiate components for designated "
589 "I/O, file state: %d\n",
590 PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj));
592 io->ci_need_write_intent = 0;
593 GOTO(out, result = -EIO);
596 if (io->ci_need_write_intent)
597 GOTO(out, result = 1);
605 static void lov_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
607 struct lov_io *lio = cl2lov_io(env, ios);
608 struct lov_object *lov = cl2lov(ios->cis_obj);
612 LASSERT(list_empty(&lio->lis_active));
614 while (!list_empty(&lio->lis_subios)) {
615 struct lov_io_sub *sub = list_entry(lio->lis_subios.next,
619 list_del_init(&sub->sub_list);
620 lio->lis_nr_subios--;
622 lov_io_sub_fini(env, lio, sub);
623 lov_sub_free(lio, sub);
625 LASSERT(lio->lis_nr_subios == 0);
627 LASSERT(atomic_read(&lov->lo_active_ios) > 0);
628 if (atomic_dec_and_test(&lov->lo_active_ios))
629 wake_up_all(&lov->lo_waitq);
633 static void lov_io_sub_inherit(struct lov_io_sub *sub, struct lov_io *lio,
634 loff_t start, loff_t end)
636 struct cl_io *io = &sub->sub_io;
637 struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
638 struct cl_io *parent = lio->lis_cl.cis_io;
639 int index = lov_comp_entry(sub->sub_subio_index);
640 int stripe = lov_comp_stripe(sub->sub_subio_index);
642 io->ci_pio = parent->ci_pio;
643 switch (io->ci_type) {
645 io->u.ci_setattr.sa_attr = parent->u.ci_setattr.sa_attr;
646 io->u.ci_setattr.sa_attr_flags =
647 parent->u.ci_setattr.sa_attr_flags;
648 io->u.ci_setattr.sa_avalid = parent->u.ci_setattr.sa_avalid;
649 io->u.ci_setattr.sa_xvalid = parent->u.ci_setattr.sa_xvalid;
650 io->u.ci_setattr.sa_stripe_index = stripe;
651 io->u.ci_setattr.sa_parent_fid =
652 parent->u.ci_setattr.sa_parent_fid;
653 if (cl_io_is_trunc(io)) {
654 loff_t new_size = parent->u.ci_setattr.sa_attr.lvb_size;
656 new_size = lov_size_to_stripe(lsm, index, new_size,
658 io->u.ci_setattr.sa_attr.lvb_size = new_size;
660 lov_lsm2layout(lsm, lsm->lsm_entries[index],
661 &io->u.ci_setattr.sa_layout);
664 case CIT_DATA_VERSION: {
665 io->u.ci_data_version.dv_data_version = 0;
666 io->u.ci_data_version.dv_flags =
667 parent->u.ci_data_version.dv_flags;
671 struct cl_object *obj = parent->ci_obj;
672 loff_t off = cl_offset(obj, parent->u.ci_fault.ft_index);
674 io->u.ci_fault = parent->u.ci_fault;
675 off = lov_size_to_stripe(lsm, index, off, stripe);
676 io->u.ci_fault.ft_index = cl_index(obj, off);
680 io->u.ci_fsync.fi_start = start;
681 io->u.ci_fsync.fi_end = end;
682 io->u.ci_fsync.fi_fid = parent->u.ci_fsync.fi_fid;
683 io->u.ci_fsync.fi_mode = parent->u.ci_fsync.fi_mode;
688 io->u.ci_rw.rw_ptask = parent->u.ci_rw.rw_ptask;
689 io->u.ci_rw.rw_iter = parent->u.ci_rw.rw_iter;
690 io->u.ci_rw.rw_iocb = parent->u.ci_rw.rw_iocb;
691 io->u.ci_rw.rw_file = parent->u.ci_rw.rw_file;
692 io->u.ci_rw.rw_sync = parent->u.ci_rw.rw_sync;
693 if (cl_io_is_append(parent)) {
694 io->u.ci_rw.rw_append = 1;
696 io->u.ci_rw.rw_range.cir_pos = start;
697 io->u.ci_rw.rw_range.cir_count = end - start;
702 io->u.ci_ladvise.li_start = start;
703 io->u.ci_ladvise.li_end = end;
704 io->u.ci_ladvise.li_fid = parent->u.ci_ladvise.li_fid;
705 io->u.ci_ladvise.li_advice = parent->u.ci_ladvise.li_advice;
706 io->u.ci_ladvise.li_flags = parent->u.ci_ladvise.li_flags;
716 static loff_t lov_offset_mod(loff_t val, int delta)
718 if (val != OBD_OBJECT_EOF)
723 static int lov_io_iter_init(const struct lu_env *env,
724 const struct cl_io_slice *ios)
726 struct lov_io *lio = cl2lov_io(env, ios);
727 struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
728 struct lov_io_sub *sub;
729 struct lu_extent ext;
735 ext.e_start = lio->lis_pos;
736 ext.e_end = lio->lis_endpos;
738 lov_foreach_io_layout(index, lio, &ext) {
739 struct lov_layout_entry *le = lov_entry(lio->lis_object, index);
740 struct lov_layout_raid0 *r0 = &le->lle_raid0;
745 CDEBUG(D_VFSTRACE, "component[%d] flags %#x\n",
746 index, lsm->lsm_entries[index]->lsme_flags);
747 if (!lsm_entry_inited(lsm, index)) {
749 * Read from uninitialized components should return
755 if (!le->lle_valid && !ios->cis_io->ci_designated_mirror) {
756 CERROR("I/O to invalid component: %d, mirror: %d\n",
757 index, lio->lis_mirror_index);
761 for (stripe = 0; stripe < r0->lo_nr; stripe++) {
762 if (!lov_stripe_intersects(lsm, index, stripe,
766 if (unlikely(!r0->lo_sub[stripe])) {
767 if (ios->cis_io->ci_type == CIT_READ ||
768 ios->cis_io->ci_type == CIT_WRITE ||
769 ios->cis_io->ci_type == CIT_FAULT)
775 end = lov_offset_mod(end, 1);
776 sub = lov_sub_get(env, lio,
777 lov_comp_index(index, stripe));
783 lov_io_sub_inherit(sub, lio, start, end);
784 rc = cl_io_iter_init(sub->sub_env, &sub->sub_io);
786 cl_io_iter_fini(sub->sub_env, &sub->sub_io);
791 "shrink stripe: {%d, %d} range: [%llu, %llu)\n",
792 index, stripe, start, end);
794 list_add_tail(&sub->sub_linkage, &lio->lis_active);
802 static int lov_io_rw_iter_init(const struct lu_env *env,
803 const struct cl_io_slice *ios)
805 struct cl_io *io = ios->cis_io;
806 struct lov_io *lio = cl2lov_io(env, ios);
807 struct lov_stripe_md_entry *lse;
808 struct cl_io_range *range = &io->u.ci_rw.rw_range;
809 loff_t start = range->cir_pos;
813 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
816 if (cl_io_is_append(io))
817 RETURN(lov_io_iter_init(env, ios));
819 index = lov_io_layout_at(lio, range->cir_pos);
820 if (index < 0) { /* non-existing layout component */
821 if (io->ci_type == CIT_READ) {
823 * TODO: it needs to detect the next component and
824 * then set the next pos
827 /* execute it in main thread */
830 RETURN(lov_io_iter_init(env, ios));
836 if (!lov_entry(lio->lis_object, index)->lle_valid &&
837 !io->ci_designated_mirror)
838 RETURN(io->ci_type == CIT_READ ? -EAGAIN : -EIO);
840 lse = lov_lse(lio->lis_object, index);
842 next = MAX_LFS_FILESIZE;
843 if (lse->lsme_stripe_count > 1) {
844 unsigned long ssize = lse->lsme_stripe_size;
846 lov_do_div64(start, ssize);
847 next = (start + 1) * ssize;
848 if (next <= start * ssize)
849 next = MAX_LFS_FILESIZE;
852 LASSERTF(range->cir_pos >= lse->lsme_extent.e_start,
853 "pos %lld, [%lld, %lld)\n", range->cir_pos,
854 lse->lsme_extent.e_start, lse->lsme_extent.e_end);
855 next = min_t(__u64, next, lse->lsme_extent.e_end);
856 next = min_t(loff_t, next, lio->lis_io_endpos);
858 io->ci_continue = next < lio->lis_io_endpos;
859 range->cir_count = next - range->cir_pos;
860 lio->lis_pos = range->cir_pos;
861 lio->lis_endpos = range->cir_pos + range->cir_count;
863 "stripe: {%d, %llu} range: [%llu, %llu) end: %llu, count: %zd\n",
864 index, start, lio->lis_pos, lio->lis_endpos,
865 lio->lis_io_endpos, range->cir_count);
867 if (!io->ci_continue) {
868 /* the last piece of IO, execute it in main thread */
876 * XXX The following call should be optimized: we know, that
877 * [lio->lis_pos, lio->lis_endpos) intersects with exactly one stripe.
879 RETURN(lov_io_iter_init(env, ios));
882 static int lov_io_setattr_iter_init(const struct lu_env *env,
883 const struct cl_io_slice *ios)
885 struct lov_io *lio = cl2lov_io(env, ios);
886 struct cl_io *io = ios->cis_io;
890 if (cl_io_is_trunc(io) && lio->lis_pos > 0) {
891 index = lov_io_layout_at(lio, lio->lis_pos - 1);
892 /* no entry found for such offset */
894 RETURN(io->ci_result = -ENODATA);
897 RETURN(lov_io_iter_init(env, ios));
900 static int lov_io_call(const struct lu_env *env, struct lov_io *lio,
901 int (*iofunc)(const struct lu_env *, struct cl_io *))
903 struct cl_io *parent = lio->lis_cl.cis_io;
904 struct lov_io_sub *sub;
908 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
909 rc = iofunc(sub->sub_env, &sub->sub_io);
913 if (parent->ci_result == 0)
914 parent->ci_result = sub->sub_io.ci_result;
919 static int lov_io_lock(const struct lu_env *env, const struct cl_io_slice *ios)
922 RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_lock));
925 static int lov_io_start(const struct lu_env *env, const struct cl_io_slice *ios)
928 RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_start));
931 static int lov_io_end_wrapper(const struct lu_env *env, struct cl_io *io)
935 * It's possible that lov_io_start() wasn't called against this
936 * sub-io, either because previous sub-io failed, or upper layer
939 if (io->ci_state == CIS_IO_GOING)
942 io->ci_state = CIS_IO_FINISHED;
946 static int lov_io_iter_fini_wrapper(const struct lu_env *env, struct cl_io *io)
948 cl_io_iter_fini(env, io);
952 static int lov_io_unlock_wrapper(const struct lu_env *env, struct cl_io *io)
954 cl_io_unlock(env, io);
958 static void lov_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
962 rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_end_wrapper);
967 lov_io_data_version_end(const struct lu_env *env, const struct cl_io_slice *ios)
969 struct lov_io *lio = cl2lov_io(env, ios);
970 struct cl_io *parent = lio->lis_cl.cis_io;
971 struct cl_data_version_io *pdv = &parent->u.ci_data_version;
972 struct lov_io_sub *sub;
975 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
976 struct cl_data_version_io *sdv = &sub->sub_io.u.ci_data_version;
978 lov_io_end_wrapper(sub->sub_env, &sub->sub_io);
980 pdv->dv_data_version += sdv->dv_data_version;
981 if (pdv->dv_layout_version > sdv->dv_layout_version)
982 pdv->dv_layout_version = sdv->dv_layout_version;
984 if (parent->ci_result == 0)
985 parent->ci_result = sub->sub_io.ci_result;
991 static void lov_io_iter_fini(const struct lu_env *env,
992 const struct cl_io_slice *ios)
994 struct lov_io *lio = cl2lov_io(env, ios);
998 rc = lov_io_call(env, lio, lov_io_iter_fini_wrapper);
1000 while (!list_empty(&lio->lis_active))
1001 list_del_init(lio->lis_active.next);
1005 static void lov_io_unlock(const struct lu_env *env,
1006 const struct cl_io_slice *ios)
1011 rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_unlock_wrapper);
1016 static int lov_io_read_ahead(const struct lu_env *env,
1017 const struct cl_io_slice *ios,
1018 pgoff_t start, struct cl_read_ahead *ra)
1020 struct lov_io *lio = cl2lov_io(env, ios);
1021 struct lov_object *loo = lio->lis_object;
1022 struct cl_object *obj = lov2cl(loo);
1023 struct lov_layout_raid0 *r0;
1024 struct lov_io_sub *sub;
1028 unsigned int pps; /* pages per stripe */
1034 offset = cl_offset(obj, start);
1035 index = lov_io_layout_at(lio, offset);
1036 if (index < 0 || !lsm_entry_inited(loo->lo_lsm, index))
1039 /* avoid readahead to expand to stale components */
1040 if (!lov_entry(loo, index)->lle_valid)
1043 stripe = lov_stripe_number(loo->lo_lsm, index, offset);
1045 r0 = lov_r0(loo, index);
1046 if (unlikely(!r0->lo_sub[stripe]))
1049 sub = lov_sub_get(env, lio, lov_comp_index(index, stripe));
1051 RETURN(PTR_ERR(sub));
1053 lov_stripe_offset(loo->lo_lsm, index, offset, stripe, &suboff);
1054 rc = cl_io_read_ahead(sub->sub_env, &sub->sub_io,
1055 cl_index(lovsub2cl(r0->lo_sub[stripe]), suboff),
1058 CDEBUG(D_READA, DFID " cra_end = %lu, stripes = %d, rc = %d\n",
1059 PFID(lu_object_fid(lov2lu(loo))), ra->cra_end, r0->lo_nr, rc);
1064 * Adjust the stripe index by layout of comp. ra->cra_end is the
1065 * maximum page index covered by an underlying DLM lock.
1066 * This function converts cra_end from stripe level to file level, and
1067 * make sure it's not beyond stripe and component boundary.
1070 /* cra_end is stripe level, convert it into file level */
1071 ra_end = ra->cra_end;
1072 if (ra_end != CL_PAGE_EOF)
1073 ra->cra_end = lov_stripe_pgoff(loo->lo_lsm, index,
1076 /* boundary of current component */
1077 ra_end = cl_index(obj, (loff_t)lov_io_extent(lio, index)->e_end);
1078 if (ra_end != CL_PAGE_EOF && ra->cra_end >= ra_end)
1079 ra->cra_end = ra_end - 1;
1081 if (r0->lo_nr == 1) /* single stripe file */
1084 pps = lov_lse(loo, index)->lsme_stripe_size >> PAGE_SHIFT;
1086 CDEBUG(D_READA, DFID " max_index = %lu, pps = %u, index = %u, "
1087 "stripe_size = %u, stripe no = %u, start index = %lu\n",
1088 PFID(lu_object_fid(lov2lu(loo))), ra->cra_end, pps, index,
1089 lov_lse(loo, index)->lsme_stripe_size, stripe, start);
1091 /* never exceed the end of the stripe */
1092 ra->cra_end = min_t(pgoff_t,
1093 ra->cra_end, start + pps - start % pps - 1);
1098 * lov implementation of cl_operations::cio_submit() method. It takes a list
1099 * of pages in \a queue, splits it into per-stripe sub-lists, invokes
1100 * cl_io_submit() on underlying devices to submit sub-lists, and then splices
1103 * Major complication of this function is a need to handle memory cleansing:
1104 * cl_io_submit() is called to write out pages as a part of VM memory
1105 * reclamation, and hence it may not fail due to memory shortages (system
1106 * dead-locks otherwise). To deal with this, some resources (sub-lists,
1107 * sub-environment, etc.) are allocated per-device on "startup" (i.e., in a
1108 * not-memory cleansing context), and in case of memory shortage, these
1109 * pre-allocated resources are used by lov_io_submit() under
1110 * lov_device::ld_mutex mutex.
1112 static int lov_io_submit(const struct lu_env *env,
1113 const struct cl_io_slice *ios,
1114 enum cl_req_type crt, struct cl_2queue *queue)
1116 struct cl_page_list *qin = &queue->c2_qin;
1117 struct lov_io *lio = cl2lov_io(env, ios);
1118 struct lov_io_sub *sub;
1119 struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
1120 struct cl_page *page;
1125 cl_page_list_init(plist);
1126 while (qin->pl_nr > 0) {
1127 struct cl_2queue *cl2q = &lov_env_info(env)->lti_cl2q;
1129 page = cl_page_list_first(qin);
1130 if (lov_page_is_empty(page)) {
1131 cl_page_list_move(&queue->c2_qout, qin, page);
1134 * it could only be mirror read to get here therefore
1135 * the pages will be transient. We don't care about
1136 * the return code of cl_page_prep() at all.
1138 (void) cl_page_prep(env, ios->cis_io, page, crt);
1139 cl_page_completion(env, page, crt, 0);
1143 cl_2queue_init(cl2q);
1144 cl_page_list_move(&cl2q->c2_qin, qin, page);
1146 index = lov_page_index(page);
1147 while (qin->pl_nr > 0) {
1148 page = cl_page_list_first(qin);
1149 if (index != lov_page_index(page))
1152 cl_page_list_move(&cl2q->c2_qin, qin, page);
1155 sub = lov_sub_get(env, lio, index);
1157 rc = cl_io_submit_rw(sub->sub_env, &sub->sub_io,
1163 cl_page_list_splice(&cl2q->c2_qin, plist);
1164 cl_page_list_splice(&cl2q->c2_qout, &queue->c2_qout);
1165 cl_2queue_fini(env, cl2q);
1171 cl_page_list_splice(plist, qin);
1172 cl_page_list_fini(env, plist);
1177 static int lov_io_commit_async(const struct lu_env *env,
1178 const struct cl_io_slice *ios,
1179 struct cl_page_list *queue, int from, int to,
1182 struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
1183 struct lov_io *lio = cl2lov_io(env, ios);
1184 struct lov_io_sub *sub;
1185 struct cl_page *page;
1189 if (lio->lis_nr_subios == 1) {
1190 int idx = lio->lis_single_subio_index;
1192 LASSERT(!lov_page_is_empty(cl_page_list_first(queue)));
1194 sub = lov_sub_get(env, lio, idx);
1195 LASSERT(!IS_ERR(sub));
1196 LASSERT(sub == &lio->lis_single_subio);
1197 rc = cl_io_commit_async(sub->sub_env, &sub->sub_io, queue,
1202 cl_page_list_init(plist);
1203 while (queue->pl_nr > 0) {
1207 LASSERT(plist->pl_nr == 0);
1208 page = cl_page_list_first(queue);
1209 LASSERT(!lov_page_is_empty(page));
1211 cl_page_list_move(plist, queue, page);
1213 index = lov_page_index(page);
1214 while (queue->pl_nr > 0) {
1215 page = cl_page_list_first(queue);
1216 if (index != lov_page_index(page))
1219 cl_page_list_move(plist, queue, page);
1222 if (queue->pl_nr > 0) /* still has more pages */
1223 stripe_to = PAGE_SIZE;
1225 sub = lov_sub_get(env, lio, index);
1227 rc = cl_io_commit_async(sub->sub_env, &sub->sub_io,
1228 plist, from, stripe_to, cb);
1234 if (plist->pl_nr > 0) /* short write */
1240 /* for error case, add the page back into the qin list */
1241 LASSERT(ergo(rc == 0, plist->pl_nr == 0));
1242 while (plist->pl_nr > 0) {
1243 /* error occurred, add the uncommitted pages back into queue */
1244 page = cl_page_list_last(plist);
1245 cl_page_list_move_head(queue, plist, page);
1251 static int lov_io_fault_start(const struct lu_env *env,
1252 const struct cl_io_slice *ios)
1254 struct cl_fault_io *fio;
1256 struct lov_io_sub *sub;
1260 fio = &ios->cis_io->u.ci_fault;
1261 lio = cl2lov_io(env, ios);
1262 sub = lov_sub_get(env, lio, lov_page_index(fio->ft_page));
1263 sub->sub_io.u.ci_fault.ft_nob = fio->ft_nob;
1265 RETURN(lov_io_start(env, ios));
1268 static void lov_io_fsync_end(const struct lu_env *env,
1269 const struct cl_io_slice *ios)
1271 struct lov_io *lio = cl2lov_io(env, ios);
1272 struct lov_io_sub *sub;
1273 unsigned int *written = &ios->cis_io->u.ci_fsync.fi_nr_written;
1277 list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
1278 struct cl_io *subio = &sub->sub_io;
1280 lov_io_end_wrapper(sub->sub_env, subio);
1282 if (subio->ci_result == 0)
1283 *written += subio->u.ci_fsync.fi_nr_written;
1288 static const struct cl_io_operations lov_io_ops = {
1291 .cio_fini = lov_io_fini,
1292 .cio_iter_init = lov_io_rw_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
1300 .cio_fini = lov_io_fini,
1301 .cio_iter_init = lov_io_rw_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_end
1309 .cio_fini = lov_io_fini,
1310 .cio_iter_init = lov_io_setattr_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_start,
1315 .cio_end = lov_io_end
1317 [CIT_DATA_VERSION] = {
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_data_version_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_fault_start,
1333 .cio_end = lov_io_end
1336 .cio_fini = lov_io_fini,
1337 .cio_iter_init = lov_io_iter_init,
1338 .cio_iter_fini = lov_io_iter_fini,
1339 .cio_lock = lov_io_lock,
1340 .cio_unlock = lov_io_unlock,
1341 .cio_start = lov_io_start,
1342 .cio_end = lov_io_fsync_end
1345 .cio_fini = lov_io_fini,
1346 .cio_iter_init = lov_io_iter_init,
1347 .cio_iter_fini = lov_io_iter_fini,
1348 .cio_lock = lov_io_lock,
1349 .cio_unlock = lov_io_unlock,
1350 .cio_start = lov_io_start,
1351 .cio_end = lov_io_end
1354 .cio_fini = lov_io_fini,
1357 .cio_fini = lov_io_fini
1360 .cio_read_ahead = lov_io_read_ahead,
1361 .cio_submit = lov_io_submit,
1362 .cio_commit_async = lov_io_commit_async,
1365 /*****************************************************************************
1367 * Empty lov io operations.
1371 static void lov_empty_io_fini(const struct lu_env *env,
1372 const struct cl_io_slice *ios)
1374 struct lov_object *lov = cl2lov(ios->cis_obj);
1377 if (atomic_dec_and_test(&lov->lo_active_ios))
1378 wake_up_all(&lov->lo_waitq);
1382 static int lov_empty_io_submit(const struct lu_env *env,
1383 const struct cl_io_slice *ios,
1384 enum cl_req_type crt, struct cl_2queue *queue)
1389 static void lov_empty_impossible(const struct lu_env *env,
1390 struct cl_io_slice *ios)
1395 #define LOV_EMPTY_IMPOSSIBLE ((void *)lov_empty_impossible)
1398 * An io operation vector for files without stripes.
1400 static const struct cl_io_operations lov_empty_io_ops = {
1403 .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
1412 .cio_fini = lov_empty_io_fini,
1413 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1414 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1415 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1416 .cio_end = LOV_EMPTY_IMPOSSIBLE
1419 .cio_fini = lov_empty_io_fini,
1420 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1421 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1422 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1423 .cio_end = LOV_EMPTY_IMPOSSIBLE
1426 .cio_fini = lov_empty_io_fini,
1427 .cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
1428 .cio_lock = LOV_EMPTY_IMPOSSIBLE,
1429 .cio_start = LOV_EMPTY_IMPOSSIBLE,
1430 .cio_end = LOV_EMPTY_IMPOSSIBLE
1433 .cio_fini = lov_empty_io_fini
1436 .cio_fini = lov_empty_io_fini
1439 .cio_fini = lov_empty_io_fini
1442 .cio_fini = lov_empty_io_fini
1445 .cio_submit = lov_empty_io_submit,
1446 .cio_commit_async = LOV_EMPTY_IMPOSSIBLE
1449 int lov_io_init_composite(const struct lu_env *env, struct cl_object *obj,
1452 struct lov_io *lio = lov_env_io(env);
1453 struct lov_object *lov = cl2lov(obj);
1458 INIT_LIST_HEAD(&lio->lis_active);
1459 result = lov_io_slice_init(lio, lov, io);
1463 result = lov_io_subio_init(env, lio, io);
1465 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_io_ops);
1466 atomic_inc(&lov->lo_active_ios);
1470 io->ci_result = result < 0 ? result : 0;
1474 int lov_io_init_empty(const struct lu_env *env, struct cl_object *obj,
1477 struct lov_object *lov = cl2lov(obj);
1478 struct lov_io *lio = lov_env_io(env);
1482 lio->lis_object = lov;
1483 switch (io->ci_type) {
1494 case CIT_DATA_VERSION:
1502 CERROR("Page fault on a file without stripes: "DFID"\n",
1503 PFID(lu_object_fid(&obj->co_lu)));
1507 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
1508 atomic_inc(&lov->lo_active_ios);
1511 io->ci_result = result < 0 ? result : 0;
1515 int lov_io_init_released(const struct lu_env *env, struct cl_object *obj,
1518 struct lov_object *lov = cl2lov(obj);
1519 struct lov_io *lio = lov_env_io(env);
1523 LASSERT(lov->lo_lsm != NULL);
1524 lio->lis_object = lov;
1526 switch (io->ci_type) {
1528 LASSERTF(0, "invalid type %d\n", io->ci_type);
1529 result = -EOPNOTSUPP;
1535 case CIT_DATA_VERSION:
1540 * the truncate to 0 is managed by MDT:
1541 * - in open, for open O_TRUNC
1542 * - in setattr, for truncate
1544 /* the truncate is for size > 0 so triggers a restore */
1545 if (cl_io_is_trunc(io)) {
1546 io->ci_restore_needed = 1;
1554 io->ci_restore_needed = 1;
1560 cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
1561 atomic_inc(&lov->lo_active_ios);
1564 io->ci_result = result < 0 ? result : 0;
1569 * Return the index in composite:lo_entries by the file offset
1571 int lov_io_layout_at(struct lov_io *lio, __u64 offset)
1573 struct lov_object *lov = lio->lis_object;
1574 struct lov_layout_composite *comp = &lov->u.composite;
1575 int start_index = 0;
1576 int end_index = comp->lo_entry_count - 1;
1579 LASSERT(lov->lo_type == LLT_COMP);
1581 /* This is actual file offset so nothing can cover eof. */
1582 if (offset == LUSTRE_EOF)
1585 if (lov_is_flr(lov)) {
1586 struct lov_mirror_entry *lre;
1588 LASSERT(lio->lis_mirror_index >= 0);
1590 lre = &comp->lo_mirrors[lio->lis_mirror_index];
1591 start_index = lre->lre_start;
1592 end_index = lre->lre_end;
1595 for (i = start_index; i <= end_index; i++) {
1596 struct lov_layout_entry *lle = lov_entry(lov, i);
1598 if ((offset >= lle->lle_extent->e_start &&
1599 offset < lle->lle_extent->e_end) ||
1600 (offset == OBD_OBJECT_EOF &&
1601 lle->lle_extent->e_end == OBD_OBJECT_EOF))