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/
31 * Implementation of cl_io for OSC layer.
33 * Author: Nikita Danilov <nikita.danilov@sun.com>
34 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
37 #define DEBUG_SUBSYSTEM S_OSC
39 #include <lustre_obdo.h>
40 #include <lustre_osc.h>
41 #include <linux/pagevec.h>
42 #include <linux/falloc.h>
44 #include "osc_internal.h"
45 #include <lnet/lnet_rdma.h>
51 /*****************************************************************************
57 static void osc_io_fini(const struct lu_env *env, const struct cl_io_slice *io)
61 void osc_read_ahead_release(const struct lu_env *env, struct cl_read_ahead *ra)
63 struct ldlm_lock *dlmlock = ra->cra_dlmlock;
64 struct osc_io *oio = ra->cra_oio;
65 struct lustre_handle lockh;
67 oio->oi_is_readahead = 0;
68 ldlm_lock2handle(dlmlock, &lockh);
69 ldlm_lock_decref(&lockh, LCK_PR);
70 LDLM_LOCK_PUT(dlmlock);
72 EXPORT_SYMBOL(osc_read_ahead_release);
74 static int osc_io_read_ahead(const struct lu_env *env,
75 const struct cl_io_slice *ios,
76 pgoff_t start, struct cl_read_ahead *ra)
78 struct osc_object *osc = cl2osc(ios->cis_obj);
79 struct osc_io *oio = cl2osc_io(env, ios);
80 struct ldlm_lock *dlmlock;
81 int result = -ENODATA;
85 oio->oi_is_readahead = true;
86 dlmlock = osc_dlmlock_at_pgoff(env, osc, start, 0);
87 if (dlmlock != NULL) {
88 struct lov_oinfo *oinfo = osc->oo_oinfo;
90 LASSERT(dlmlock->l_ast_data == osc);
91 if (dlmlock->l_req_mode != LCK_PR) {
92 struct lustre_handle lockh;
93 ldlm_lock2handle(dlmlock, &lockh);
94 ldlm_lock_addref(&lockh, LCK_PR);
95 ldlm_lock_decref(&lockh, dlmlock->l_req_mode);
98 ra->cra_rpc_pages = osc_cli(osc)->cl_max_pages_per_rpc;
100 dlmlock->l_policy_data.l_extent.end >> PAGE_SHIFT;
101 ra->cra_release = osc_read_ahead_release;
102 ra->cra_dlmlock = dlmlock;
104 if (ra->cra_end_idx != CL_PAGE_EOF)
105 ra->cra_contention = true;
106 ra->cra_end_idx = min_t(pgoff_t,
108 (oinfo->loi_kms - 1) >> PAGE_SHIFT);
116 * An implementation of cl_io_operations::cio_io_submit() method for osc
117 * layer. Iterates over pages in the in-queue, prepares each for io by calling
118 * cl_page_prep() and then either submits them through osc_io_submit_page()
119 * or, if page is already submitted, changes osc flags through
120 * osc_set_async_flags().
122 int osc_io_submit(const struct lu_env *env, const struct cl_io_slice *ios,
123 enum cl_req_type crt, struct cl_2queue *queue)
125 struct cl_page *page;
127 struct client_obd *cli = NULL;
128 struct osc_object *osc = NULL; /* to keep gcc happy */
129 struct osc_page *opg;
133 struct cl_page_list *qin = &queue->c2_qin;
134 struct cl_page_list *qout = &queue->c2_qout;
135 unsigned int queued = 0;
138 unsigned int max_pages;
139 unsigned int ppc_bits; /* pages per chunk bits */
141 bool sync_queue = false;
143 LASSERT(qin->pl_nr > 0);
145 CDEBUG(D_CACHE|D_READA, "%d %d\n", qin->pl_nr, crt);
147 osc = cl2osc(ios->cis_obj);
149 max_pages = cli->cl_max_pages_per_rpc;
150 ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
153 brw_flags = osc_io_srvlock(cl2osc_io(env, ios)) ? OBD_BRW_SRVLOCK : 0;
154 brw_flags |= crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
155 if (crt == CRT_READ && ios->cis_io->ci_ndelay)
156 brw_flags |= OBD_BRW_NDELAY;
158 page = cl_page_list_first(qin);
159 if (page->cp_type == CPT_TRANSIENT)
160 brw_flags |= OBD_BRW_NOCACHE;
161 if (lnet_is_rdma_only_page(page->cp_vmpage))
162 brw_flags |= OBD_BRW_RDMA_ONLY;
165 * NOTE: here @page is a top-level page. This is done to avoid
166 * creation of sub-page-list.
168 cl_page_list_for_each_safe(page, tmp, qin) {
169 struct osc_async_page *oap;
175 opg = osc_cl_page_osc(page, osc);
177 LASSERT(osc == oap->oap_obj);
179 if (!list_empty(&oap->oap_pending_item) ||
180 !list_empty(&oap->oap_rpc_item)) {
181 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
187 result = cl_page_prep(env, io, page, crt);
190 if (result != -EALREADY)
193 * Handle -EALREADY error: for read case, the page is
194 * already in UPTODATE state; for write, the page
201 if (page->cp_type != CPT_TRANSIENT) {
202 oap->oap_async_flags = ASYNC_URGENT|ASYNC_READY|ASYNC_COUNT_STABLE;
205 osc_page_submit(env, opg, crt, brw_flags);
206 list_add_tail(&oap->oap_pending_item, &list);
208 if (page->cp_sync_io != NULL)
209 cl_page_list_move(qout, qin, page);
211 cl_page_list_del(env, qin, page, true);
214 if (queued == max_pages) {
216 } else if (crt == CRT_WRITE) {
218 unsigned int next_chunks;
220 chunks = (queued + ppc - 1) >> ppc_bits;
221 /* chunk number if add another page */
222 next_chunks = (queued + ppc) >> ppc_bits;
224 /* next page will excceed write chunk limit */
225 if (chunks == osc_max_write_chunks(cli) &&
226 next_chunks > chunks)
231 result = osc_queue_sync_pages(env, io, osc, &list,
241 result = osc_queue_sync_pages(env, io, osc, &list, brw_flags);
243 /* Update c/mtime for sync write. LU-7310 */
244 if (crt == CRT_WRITE && qout->pl_nr > 0 && result == 0) {
245 struct cl_object *obj = ios->cis_obj;
246 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
248 cl_object_attr_lock(obj);
249 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
250 cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
251 cl_object_attr_unlock(obj);
254 CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
255 return qout->pl_nr > 0 ? 0 : result;
257 EXPORT_SYMBOL(osc_io_submit);
260 * This is called to update the attributes when modifying a specific page,
261 * both when making new pages and when doing updates to existing cached pages.
263 * Expand stripe KMS if necessary.
265 void osc_page_touch_at(const struct lu_env *env, struct cl_object *obj,
266 pgoff_t idx, size_t to)
268 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
269 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
275 /* offset within stripe */
276 kms = (idx << PAGE_SHIFT) + to;
278 cl_object_attr_lock(obj);
279 CDEBUG(D_INODE, "stripe KMS %sincreasing %llu->%llu %llu\n",
280 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
281 loi->loi_lvb.lvb_size);
283 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
284 valid = CAT_MTIME | CAT_CTIME;
285 if (kms > loi->loi_kms) {
289 if (kms > loi->loi_lvb.lvb_size) {
290 attr->cat_size = kms;
293 cl_object_attr_update(env, obj, attr, valid);
294 cl_object_attr_unlock(obj);
299 int osc_io_commit_async(const struct lu_env *env,
300 const struct cl_io_slice *ios,
301 struct cl_page_list *qin, int from, int to,
304 struct cl_io *io = ios->cis_io;
305 struct osc_io *oio = cl2osc_io(env, ios);
306 struct osc_object *osc = cl2osc(ios->cis_obj);
307 struct cl_page *page;
308 struct cl_page *last_page;
309 struct osc_page *opg;
310 struct pagevec *pvec = &osc_env_info(env)->oti_pagevec;
314 LASSERT(qin->pl_nr > 0);
316 /* Handle partial page cases */
317 last_page = cl_page_list_last(qin);
318 if (oio->oi_lockless) {
319 page = cl_page_list_first(qin);
320 if (page == last_page) {
321 cl_page_clip(env, page, from, to);
324 cl_page_clip(env, page, from, PAGE_SIZE);
326 cl_page_clip(env, last_page, 0, to);
330 ll_pagevec_init(pvec, 0);
332 while (qin->pl_nr > 0) {
333 struct osc_async_page *oap;
335 page = cl_page_list_first(qin);
336 opg = osc_cl_page_osc(page, osc);
339 LASSERTF(osc == oap->oap_obj,
340 "obj mismatch: %px / %px\n", osc, oap->oap_obj);
342 if (!list_empty(&oap->oap_rpc_item)) {
343 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
349 /* The page may be already in dirty cache. */
350 if (list_empty(&oap->oap_pending_item)) {
351 result = osc_page_cache_add(env, opg, io, cb);
356 osc_page_touch_at(env, osc2cl(osc), osc_index(opg),
357 page == last_page ? to : PAGE_SIZE);
359 cl_page_list_del(env, qin, page, true);
361 /* if there are no more slots, do the callback & reinit */
362 if (pagevec_add(pvec, page->cp_vmpage) == 0) {
363 (*cb)(env, io, pvec);
364 pagevec_reinit(pvec);
367 /* The shrink interval is in seconds, so we can update it once per
368 * write, rather than once per page.
370 osc_update_next_shrink(osc_cli(osc));
373 /* Clean up any partially full pagevecs */
374 if (pagevec_count(pvec) != 0)
375 (*cb)(env, io, pvec);
377 /* Can't access these pages any more. Page can be in transfer and
378 * complete at any time. */
380 /* for sync write, kernel will wait for this page to be flushed before
381 * osc_io_end() is called, so release it earlier.
382 * for mkwrite(), it's known there is no further pages. */
383 if (cl_io_is_sync_write(io) && oio->oi_active != NULL) {
384 osc_extent_release(env, oio->oi_active);
385 oio->oi_active = NULL;
388 CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, result);
391 EXPORT_SYMBOL(osc_io_commit_async);
393 void osc_io_extent_release(const struct lu_env *env,
394 const struct cl_io_slice *ios)
396 struct osc_io *oio = cl2osc_io(env, ios);
398 if (oio->oi_active != NULL) {
399 osc_extent_release(env, oio->oi_active);
400 oio->oi_active = NULL;
403 EXPORT_SYMBOL(osc_io_extent_release);
405 static bool osc_import_not_healthy(struct obd_import *imp)
407 return imp->imp_invalid || imp->imp_deactive ||
408 !(imp->imp_state == LUSTRE_IMP_FULL ||
409 imp->imp_state == LUSTRE_IMP_IDLE);
412 int osc_io_iter_init(const struct lu_env *env, const struct cl_io_slice *ios)
414 struct osc_object *osc = cl2osc(ios->cis_obj);
415 struct obd_import *imp = osc_cli(osc)->cl_import;
416 struct osc_io *oio = osc_env_io(env);
421 spin_lock(&imp->imp_lock);
423 * check whether this OSC device is available for non-delay read,
424 * fast switching mirror if we haven't tried all mirrors.
426 if (ios->cis_io->ci_type == CIT_READ && ios->cis_io->ci_ndelay &&
427 !ios->cis_io->ci_tried_all_mirrors && osc_import_not_healthy(imp)) {
429 } else if (likely(!imp->imp_invalid)) {
430 atomic_inc(&osc->oo_nr_ios);
431 oio->oi_is_active = 1;
434 spin_unlock(&imp->imp_lock);
436 if (capable(CAP_SYS_RESOURCE))
437 oio->oi_cap_sys_resource = 1;
441 EXPORT_SYMBOL(osc_io_iter_init);
443 void osc_io_iter_fini(const struct lu_env *env,
444 const struct cl_io_slice *ios)
446 struct osc_io *oio = osc_env_io(env);
448 if (oio->oi_is_active) {
449 struct osc_object *osc = cl2osc(ios->cis_obj);
451 oio->oi_is_active = 0;
452 LASSERT(atomic_read(&osc->oo_nr_ios) > 0);
453 if (atomic_dec_and_test(&osc->oo_nr_ios))
454 wake_up(&osc->oo_io_waitq);
457 EXPORT_SYMBOL(osc_io_iter_fini);
459 void osc_io_rw_iter_fini(const struct lu_env *env,
460 const struct cl_io_slice *ios)
462 struct osc_io *oio = osc_env_io(env);
463 struct osc_object *osc = cl2osc(ios->cis_obj);
465 if (oio->oi_lru_reserved > 0) {
466 osc_lru_unreserve(osc_cli(osc), oio->oi_lru_reserved);
467 oio->oi_lru_reserved = 0;
469 oio->oi_write_osclock = NULL;
470 oio->oi_read_osclock = NULL;
472 osc_io_iter_fini(env, ios);
474 EXPORT_SYMBOL(osc_io_rw_iter_fini);
476 int osc_io_fault_start(const struct lu_env *env, const struct cl_io_slice *ios)
479 struct cl_fault_io *fio;
483 fio = &io->u.ci_fault;
484 CDEBUG(D_INFO, "%lu %d %zu\n",
485 fio->ft_index, fio->ft_writable, fio->ft_bytes);
487 * If mapping is writeable, adjust kms to cover this page,
488 * but do not extend kms beyond actual file size.
491 if (fio->ft_writable)
492 osc_page_touch_at(env, ios->cis_obj,
493 fio->ft_index, fio->ft_bytes);
496 EXPORT_SYMBOL(osc_io_fault_start);
499 static int osc_async_upcall(void *a, int rc)
501 struct osc_async_cbargs *args = a;
504 complete(&args->opc_sync);
509 * Checks that there are no pages being written in the extent being truncated.
511 static bool trunc_check_cb(const struct lu_env *env, struct cl_io *io,
512 void **pvec, int count, void *cbdata)
516 for (i = 0; i < count; i++) {
517 struct osc_page *ops = pvec[i];
518 struct cl_page *page = ops->ops_cl.cpl_page;
519 struct osc_async_page *oap;
520 __u64 start = *(__u64 *)cbdata;
523 if (oap->oap_cmd & OBD_BRW_WRITE &&
524 !list_empty(&oap->oap_pending_item))
525 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %llu/%s.\n",
526 start, current->comm);
528 if (PageLocked(page->cp_vmpage))
529 CDEBUG(D_CACHE, "page %p index %lu locked for cmd=%d\n",
530 ops, osc_index(ops), oap->oap_cmd);
535 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
536 struct osc_io *oio, __u64 size)
538 struct cl_object *clob;
542 clob = oio->oi_cl.cis_obj;
543 start = size >> PAGE_SHIFT;
544 partial = (start << PAGE_SHIFT) < size;
547 * Complain if there are pages in the truncated region.
549 osc_page_gang_lookup(env, io, cl2osc(clob),
550 start + partial, CL_PAGE_EOF,
551 trunc_check_cb, (void *)&size);
555 * Flush affected pages prior punch.
556 * We shouldn't discard them locally first because that could be data loss
557 * if server doesn't support fallocate punch, we also need these data to be
558 * flushed first to prevent re-ordering with the punch
560 int osc_punch_start(const struct lu_env *env, struct cl_io *io,
561 struct cl_object *obj)
563 struct osc_object *osc = cl2osc(obj);
564 pgoff_t pg_start = io->u.ci_setattr.sa_falloc_offset >> PAGE_SHIFT;
565 pgoff_t pg_end = (io->u.ci_setattr.sa_falloc_end - 1) >> PAGE_SHIFT;
569 rc = osc_cache_writeback_range(env, osc, pg_start, pg_end, 1, 0);
573 osc_page_gang_lookup(env, io, osc, pg_start, pg_end, osc_discard_cb,
577 EXPORT_SYMBOL(osc_punch_start);
579 static int osc_io_setattr_start(const struct lu_env *env,
580 const struct cl_io_slice *slice)
582 struct cl_io *io = slice->cis_io;
583 struct osc_io *oio = cl2osc_io(env, slice);
584 struct cl_object *obj = slice->cis_obj;
585 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
586 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
587 struct obdo *oa = &oio->oi_oa;
588 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
589 unsigned int ia_avalid = io->u.ci_setattr.sa_avalid;
590 enum op_xvalid ia_xvalid = io->u.ci_setattr.sa_xvalid;
592 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
593 bool io_is_falloc = cl_io_is_fallocate(io);
596 /* truncate cache dirty pages first */
597 if (cl_io_is_trunc(io))
598 result = osc_cache_truncate_start(env, cl2osc(obj), size,
600 /* flush local pages prior punching them on server */
602 io->u.ci_setattr.sa_falloc_mode & FALLOC_FL_PUNCH_HOLE)
603 result = osc_punch_start(env, io, obj);
605 if (result == 0 && oio->oi_lockless == 0) {
606 cl_object_attr_lock(obj);
607 result = cl_object_attr_get(env, obj, attr);
609 struct ost_lvb *lvb = &io->u.ci_setattr.sa_attr;
610 unsigned int cl_valid = 0;
612 if (ia_avalid & ATTR_SIZE) {
613 attr->cat_size = size;
614 attr->cat_kms = size;
615 cl_valid = (CAT_SIZE | CAT_KMS);
617 if (ia_avalid & ATTR_MTIME_SET) {
618 attr->cat_mtime = lvb->lvb_mtime;
619 cl_valid |= CAT_MTIME;
621 if (ia_avalid & ATTR_ATIME_SET) {
622 attr->cat_atime = lvb->lvb_atime;
623 cl_valid |= CAT_ATIME;
625 if (ia_xvalid & OP_XVALID_CTIME_SET) {
626 attr->cat_ctime = lvb->lvb_ctime;
627 cl_valid |= CAT_CTIME;
629 result = cl_object_attr_update(env, obj, attr,
632 cl_object_attr_unlock(obj);
634 memset(oa, 0, sizeof(*oa));
636 oa->o_oi = loi->loi_oi;
637 obdo_set_parent_fid(oa, io->u.ci_setattr.sa_parent_fid);
638 oa->o_stripe_idx = io->u.ci_setattr.sa_stripe_index;
639 oa->o_layout = io->u.ci_setattr.sa_layout;
640 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP |
642 if (ia_avalid & ATTR_CTIME) {
643 oa->o_valid |= OBD_MD_FLCTIME;
644 oa->o_ctime = attr->cat_ctime;
646 if (ia_avalid & ATTR_ATIME) {
647 oa->o_valid |= OBD_MD_FLATIME;
648 oa->o_atime = attr->cat_atime;
650 if (ia_avalid & ATTR_MTIME) {
651 oa->o_valid |= OBD_MD_FLMTIME;
652 oa->o_mtime = attr->cat_mtime;
655 if (ia_avalid & ATTR_SIZE || io_is_falloc) {
656 if (oio->oi_lockless) {
657 oa->o_flags = OBD_FL_SRVLOCK;
658 oa->o_valid |= OBD_MD_FLFLAGS;
661 if (io->ci_layout_version > 0) {
662 /* verify layout version */
663 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
664 oa->o_layout_version = io->ci_layout_version;
667 LASSERT(oio->oi_lockless == 0);
670 if (ia_xvalid & OP_XVALID_FLAGS) {
671 oa->o_flags = io->u.ci_setattr.sa_attr_flags;
672 oa->o_valid |= OBD_MD_FLFLAGS;
675 init_completion(&cbargs->opc_sync);
678 int falloc_mode = io->u.ci_setattr.sa_falloc_mode;
680 oa->o_size = io->u.ci_setattr.sa_falloc_offset;
681 oa->o_blocks = io->u.ci_setattr.sa_falloc_end;
682 oa->o_uid = io->u.ci_setattr.sa_falloc_uid;
683 oa->o_gid = io->u.ci_setattr.sa_falloc_gid;
684 oa->o_projid = io->u.ci_setattr.sa_falloc_projid;
685 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
686 OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID;
689 "size %llu blocks %llu uid %u gid %u prjid %u\n",
690 oa->o_size, oa->o_blocks, oa->o_uid, oa->o_gid,
692 result = osc_fallocate_base(osc_export(cl2osc(obj)),
693 oa, osc_async_upcall,
694 cbargs, falloc_mode);
695 } else if (ia_avalid & ATTR_SIZE) {
697 oa->o_blocks = OBD_OBJECT_EOF;
698 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
699 result = osc_punch_send(osc_export(cl2osc(obj)),
700 oa, osc_async_upcall, cbargs);
702 result = osc_setattr_async(osc_export(cl2osc(obj)),
703 oa, osc_async_upcall,
704 cbargs, PTLRPCD_SET);
706 cbargs->opc_rpc_sent = result == 0;
712 void osc_io_setattr_end(const struct lu_env *env,
713 const struct cl_io_slice *slice)
715 struct cl_io *io = slice->cis_io;
716 struct osc_io *oio = cl2osc_io(env, slice);
717 struct cl_object *obj = slice->cis_obj;
718 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
719 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
720 struct obdo *oa = &oio->oi_oa;
721 unsigned int cl_valid = 0;
724 if (cbargs->opc_rpc_sent) {
725 wait_for_completion(&cbargs->opc_sync);
726 result = io->ci_result = cbargs->opc_rc;
729 if (cl_io_is_trunc(io)) {
730 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
733 cl_object_attr_lock(obj);
734 if (oa->o_valid & OBD_MD_FLBLOCKS) {
735 attr->cat_blocks = oa->o_blocks;
736 cl_valid |= CAT_BLOCKS;
739 cl_object_attr_update(env, obj, attr, cl_valid);
740 cl_object_attr_unlock(obj);
742 osc_trunc_check(env, io, oio, size);
743 osc_cache_truncate_end(env, oio->oi_trunc);
744 oio->oi_trunc = NULL;
747 if (cl_io_is_fallocate(io)) {
749 cl_object_attr_lock(obj);
751 if (oa->o_valid & OBD_MD_FLBLOCKS) {
752 attr->cat_blocks = oa->o_blocks;
753 cl_valid |= CAT_BLOCKS;
756 cl_object_attr_update(env, obj, attr, cl_valid);
757 cl_object_attr_unlock(obj);
761 EXPORT_SYMBOL(osc_io_setattr_end);
763 struct osc_data_version_args {
764 struct osc_io *dva_oio;
768 osc_data_version_interpret(const struct lu_env *env, struct ptlrpc_request *req,
771 struct osc_data_version_args *dva = args;
772 struct osc_io *oio = dva->dva_oio;
773 const struct ost_body *body;
779 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
781 GOTO(out, rc = -EPROTO);
783 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, &oio->oi_oa,
787 oio->oi_cbarg.opc_rc = rc;
788 complete(&oio->oi_cbarg.opc_sync);
793 static int osc_io_data_version_start(const struct lu_env *env,
794 const struct cl_io_slice *slice)
796 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
797 struct osc_io *oio = cl2osc_io(env, slice);
798 struct obdo *oa = &oio->oi_oa;
799 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
800 struct osc_object *obj = cl2osc(slice->cis_obj);
801 struct lov_oinfo *loi = obj->oo_oinfo;
802 struct obd_export *exp = osc_export(obj);
803 struct ptlrpc_request *req;
804 struct ost_body *body;
805 struct osc_data_version_args *dva;
809 memset(oa, 0, sizeof(*oa));
810 oa->o_oi = loi->loi_oi;
811 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
813 if (dv->dv_flags & (LL_DV_RD_FLUSH | LL_DV_WR_FLUSH)) {
814 oa->o_valid |= OBD_MD_FLFLAGS;
815 oa->o_flags |= OBD_FL_SRVLOCK;
816 if (dv->dv_flags & LL_DV_WR_FLUSH)
817 oa->o_flags |= OBD_FL_FLUSH;
820 init_completion(&cbargs->opc_sync);
822 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
826 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
828 ptlrpc_request_free(req);
832 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
833 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
835 ptlrpc_request_set_replen(req);
836 req->rq_interpret_reply = osc_data_version_interpret;
837 dva = ptlrpc_req_async_args(dva, req);
840 ptlrpcd_add_req(req);
845 static void osc_io_data_version_end(const struct lu_env *env,
846 const struct cl_io_slice *slice)
848 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
849 struct osc_io *oio = cl2osc_io(env, slice);
850 struct cl_object *obj = slice->cis_obj;
851 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
852 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
853 struct obdo *oa = &oio->oi_oa;
854 unsigned int cl_valid = 0;
857 wait_for_completion(&cbargs->opc_sync);
859 if (cbargs->opc_rc != 0) {
860 slice->cis_io->ci_result = cbargs->opc_rc;
862 slice->cis_io->ci_result = 0;
864 (OBD_MD_LAYOUT_VERSION | OBD_MD_FLDATAVERSION)))
865 slice->cis_io->ci_result = -EOPNOTSUPP;
867 if (oa->o_valid & OBD_MD_LAYOUT_VERSION)
868 dv->dv_layout_version = oa->o_layout_version;
869 if (oa->o_valid & OBD_MD_FLDATAVERSION)
870 dv->dv_data_version = oa->o_data_version;
872 if (dv->dv_flags & LL_DV_SZ_UPDATE) {
873 if (oa->o_valid & OBD_MD_FLSIZE) {
874 attr->cat_size = oa->o_size;
875 cl_valid |= CAT_SIZE;
878 if (oa->o_valid & OBD_MD_FLBLOCKS) {
879 attr->cat_blocks = oa->o_blocks;
880 cl_valid |= CAT_BLOCKS;
883 cl_object_attr_lock(obj);
884 cl_object_attr_update(env, obj, attr, cl_valid);
885 cl_object_attr_unlock(obj);
892 int osc_io_read_start(const struct lu_env *env,
893 const struct cl_io_slice *slice)
895 struct cl_object *obj = slice->cis_obj;
896 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
900 if (!slice->cis_io->ci_noatime) {
901 cl_object_attr_lock(obj);
902 attr->cat_atime = ktime_get_real_seconds();
903 rc = cl_object_attr_update(env, obj, attr, CAT_ATIME);
904 cl_object_attr_unlock(obj);
909 EXPORT_SYMBOL(osc_io_read_start);
911 int osc_io_write_start(const struct lu_env *env,
912 const struct cl_io_slice *slice)
914 struct cl_object *obj = slice->cis_obj;
915 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
919 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_SETTIME, 1);
920 cl_object_attr_lock(obj);
921 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
922 rc = cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
923 cl_object_attr_unlock(obj);
927 EXPORT_SYMBOL(osc_io_write_start);
929 int osc_fsync_ost(const struct lu_env *env, struct osc_object *obj,
930 struct cl_fsync_io *fio)
932 struct osc_io *oio = osc_env_io(env);
933 struct obdo *oa = &oio->oi_oa;
934 struct lov_oinfo *loi = obj->oo_oinfo;
935 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
939 memset(oa, 0, sizeof(*oa));
940 oa->o_oi = loi->loi_oi;
941 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
943 /* reload size abd blocks for start and end of sync range */
944 oa->o_size = fio->fi_start;
945 oa->o_blocks = fio->fi_end;
946 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
948 obdo_set_parent_fid(oa, fio->fi_fid);
950 init_completion(&cbargs->opc_sync);
952 rc = osc_sync_base(obj, oa, osc_async_upcall, cbargs, PTLRPCD_SET);
955 EXPORT_SYMBOL(osc_fsync_ost);
957 static int osc_io_fsync_start(const struct lu_env *env,
958 const struct cl_io_slice *slice)
960 struct cl_io *io = slice->cis_io;
961 struct cl_fsync_io *fio = &io->u.ci_fsync;
962 struct cl_object *obj = slice->cis_obj;
963 struct osc_object *osc = cl2osc(obj);
964 pgoff_t start = fio->fi_start >> PAGE_SHIFT;
965 pgoff_t end = fio->fi_end >> PAGE_SHIFT;
970 if (fio->fi_mode == CL_FSYNC_RECLAIM) {
971 struct client_obd *cli = osc_cli(osc);
973 if (!atomic_long_read(&cli->cl_unstable_count)) {
974 /* Stop flush when there are no unstable pages? */
975 CDEBUG(D_CACHE, "unstable count is zero\n");
980 if (fio->fi_end == OBD_OBJECT_EOF)
983 result = osc_cache_writeback_range(env, osc, start, end, 0,
984 fio->fi_mode == CL_FSYNC_DISCARD);
985 if (result < 0 && fio->fi_mode == CL_FSYNC_DISCARD) {
987 "%s: ignore error %d on discarding "DFID":[%lu-%lu]\n",
988 cli_name(osc_cli(osc)), result, PFID(fio->fi_fid),
993 fio->fi_nr_written += result;
996 if (fio->fi_mode == CL_FSYNC_ALL || fio->fi_mode == CL_FSYNC_RECLAIM) {
997 struct osc_io *oio = cl2osc_io(env, slice);
998 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1001 /* we have to wait for writeback to finish before we can
1002 * send OST_SYNC RPC. This is bad because it causes extents
1003 * to be written osc by osc. However, we usually start
1004 * writeback before CL_FSYNC_ALL so this won't have any real
1006 * We do not have to wait for waitback to finish in the memory
1007 * reclaim environment.
1009 if (fio->fi_mode == CL_FSYNC_ALL) {
1010 rc = osc_cache_wait_range(env, osc, start, end);
1015 rc = osc_fsync_ost(env, osc, fio);
1017 cbargs->opc_rpc_sent = 1;
1025 void osc_io_fsync_end(const struct lu_env *env,
1026 const struct cl_io_slice *slice)
1028 struct cl_fsync_io *fio = &slice->cis_io->u.ci_fsync;
1029 struct cl_object *obj = slice->cis_obj;
1030 struct osc_io *oio = cl2osc_io(env, slice);
1031 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1032 pgoff_t start = fio->fi_start >> PAGE_SHIFT;
1033 pgoff_t end = fio->fi_end >> PAGE_SHIFT;
1036 if (fio->fi_mode == CL_FSYNC_LOCAL) {
1037 result = osc_cache_wait_range(env, cl2osc(obj), start, end);
1038 } else if (cbargs->opc_rpc_sent && (fio->fi_mode == CL_FSYNC_ALL ||
1039 fio->fi_mode == CL_FSYNC_RECLAIM)) {
1041 wait_for_completion(&cbargs->opc_sync);
1043 result = cbargs->opc_rc;
1045 slice->cis_io->ci_result = result;
1047 EXPORT_SYMBOL(osc_io_fsync_end);
1049 static int osc_io_ladvise_start(const struct lu_env *env,
1050 const struct cl_io_slice *slice)
1053 struct cl_io *io = slice->cis_io;
1054 struct osc_io *oio = cl2osc_io(env, slice);
1055 struct cl_object *obj = slice->cis_obj;
1056 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1057 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
1058 struct obdo *oa = &oio->oi_oa;
1059 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1060 struct lu_ladvise *ladvise;
1061 struct ladvise_hdr *ladvise_hdr;
1066 /* TODO: add multiple ladvise support in CLIO */
1067 buf_size = offsetof(typeof(*ladvise_hdr), lah_advise[num_advise]);
1068 if (osc_env_info(env)->oti_ladvise_buf.lb_len < buf_size)
1069 lu_buf_realloc(&osc_env_info(env)->oti_ladvise_buf, buf_size);
1071 ladvise_hdr = osc_env_info(env)->oti_ladvise_buf.lb_buf;
1072 if (ladvise_hdr == NULL)
1075 memset(ladvise_hdr, 0, buf_size);
1076 ladvise_hdr->lah_magic = LADVISE_MAGIC;
1077 ladvise_hdr->lah_count = num_advise;
1078 ladvise_hdr->lah_flags = lio->lio_flags;
1080 memset(oa, 0, sizeof(*oa));
1081 oa->o_oi = loi->loi_oi;
1082 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1083 obdo_set_parent_fid(oa, lio->lio_fid);
1085 ladvise = ladvise_hdr->lah_advise;
1086 ladvise->lla_start = lio->lio_start;
1087 ladvise->lla_end = lio->lio_end;
1088 ladvise->lla_advice = lio->lio_advice;
1090 if (lio->lio_flags & LF_ASYNC) {
1091 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1092 ladvise_hdr, NULL, NULL, NULL);
1094 init_completion(&cbargs->opc_sync);
1095 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1096 ladvise_hdr, osc_async_upcall,
1097 cbargs, PTLRPCD_SET);
1098 cbargs->opc_rpc_sent = result == 0;
1103 static void osc_io_ladvise_end(const struct lu_env *env,
1104 const struct cl_io_slice *slice)
1106 struct cl_io *io = slice->cis_io;
1107 struct osc_io *oio = cl2osc_io(env, slice);
1108 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1110 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
1112 if ((!(lio->lio_flags & LF_ASYNC)) && cbargs->opc_rpc_sent) {
1113 wait_for_completion(&cbargs->opc_sync);
1114 result = cbargs->opc_rc;
1116 slice->cis_io->ci_result = result;
1119 void osc_io_end(const struct lu_env *env, const struct cl_io_slice *slice)
1121 struct osc_io *oio = cl2osc_io(env, slice);
1123 if (oio->oi_active) {
1124 osc_extent_release(env, oio->oi_active);
1125 oio->oi_active = NULL;
1128 EXPORT_SYMBOL(osc_io_end);
1130 struct osc_lseek_args {
1131 struct osc_io *lsa_oio;
1134 static int osc_lseek_interpret(const struct lu_env *env,
1135 struct ptlrpc_request *req,
1138 struct ost_body *reply;
1139 struct osc_lseek_args *lsa = arg;
1140 struct osc_io *oio = lsa->lsa_oio;
1141 struct cl_io *io = oio->oi_cl.cis_io;
1142 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1149 reply = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1151 GOTO(out, rc = -EPROTO);
1153 lsio->ls_result = reply->oa.o_size;
1155 osc_async_upcall(&oio->oi_cbarg, rc);
1159 int osc_io_lseek_start(const struct lu_env *env,
1160 const struct cl_io_slice *slice)
1162 struct cl_io *io = slice->cis_io;
1163 struct osc_io *oio = cl2osc_io(env, slice);
1164 struct cl_object *obj = slice->cis_obj;
1165 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1166 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1167 struct obdo *oa = &oio->oi_oa;
1168 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1169 struct obd_export *exp = osc_export(cl2osc(obj));
1170 struct ptlrpc_request *req;
1171 struct ost_body *body;
1172 struct osc_lseek_args *lsa;
1177 /* No negative values at this point */
1178 LASSERT(lsio->ls_start >= 0);
1179 LASSERT(lsio->ls_whence == SEEK_HOLE || lsio->ls_whence == SEEK_DATA);
1181 /* with IO lock taken we have object size in LVB and can check
1182 * boundaries prior sending LSEEK RPC
1184 if (lsio->ls_start >= loi->loi_lvb.lvb_size) {
1185 /* consider area beyond end of object as hole */
1186 if (lsio->ls_whence == SEEK_HOLE)
1187 lsio->ls_result = lsio->ls_start;
1189 lsio->ls_result = -ENXIO;
1193 /* if LSEEK RPC is not supported by server, consider whole stripe
1194 * object is data with hole after end of object
1196 if (!exp_connect_lseek(exp)) {
1197 if (lsio->ls_whence == SEEK_HOLE)
1198 lsio->ls_result = loi->loi_lvb.lvb_size;
1200 lsio->ls_result = lsio->ls_start;
1204 memset(oa, 0, sizeof(*oa));
1205 oa->o_oi = loi->loi_oi;
1206 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1207 oa->o_size = lsio->ls_start;
1208 oa->o_mode = lsio->ls_whence;
1209 if (oio->oi_lockless) {
1210 oa->o_flags = OBD_FL_SRVLOCK;
1211 oa->o_valid |= OBD_MD_FLFLAGS;
1214 init_completion(&cbargs->opc_sync);
1215 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SEEK);
1219 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SEEK);
1221 ptlrpc_request_free(req);
1225 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1226 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1227 ptlrpc_request_set_replen(req);
1228 req->rq_interpret_reply = osc_lseek_interpret;
1229 lsa = ptlrpc_req_async_args(lsa, req);
1232 ptlrpcd_add_req(req);
1233 cbargs->opc_rpc_sent = 1;
1237 EXPORT_SYMBOL(osc_io_lseek_start);
1239 void osc_io_lseek_end(const struct lu_env *env,
1240 const struct cl_io_slice *slice)
1242 struct osc_io *oio = cl2osc_io(env, slice);
1243 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1246 if (cbargs->opc_rpc_sent) {
1247 wait_for_completion(&cbargs->opc_sync);
1248 rc = cbargs->opc_rc;
1250 slice->cis_io->ci_result = rc;
1252 EXPORT_SYMBOL(osc_io_lseek_end);
1254 int osc_io_lru_reserve(const struct lu_env *env,
1255 const struct cl_io_slice *ios,
1256 loff_t pos, size_t bytes)
1258 struct osc_object *osc = cl2osc(ios->cis_obj);
1259 struct osc_io *oio = osc_env_io(env);
1260 unsigned long npages = 0;
1265 page_offset = pos & ~PAGE_MASK;
1268 if (bytes > PAGE_SIZE - page_offset)
1269 bytes -= (PAGE_SIZE - page_offset);
1273 npages += (bytes + PAGE_SIZE - 1) >> PAGE_SHIFT;
1274 oio->oi_lru_reserved = osc_lru_reserve(osc_cli(osc), npages);
1278 EXPORT_SYMBOL(osc_io_lru_reserve);
1280 static const struct cl_io_operations osc_io_ops = {
1283 .cio_iter_init = osc_io_iter_init,
1284 .cio_iter_fini = osc_io_rw_iter_fini,
1285 .cio_start = osc_io_read_start,
1286 .cio_fini = osc_io_fini
1289 .cio_iter_init = osc_io_iter_init,
1290 .cio_iter_fini = osc_io_rw_iter_fini,
1291 .cio_start = osc_io_write_start,
1292 .cio_end = osc_io_end,
1293 .cio_fini = osc_io_fini
1296 .cio_iter_init = osc_io_iter_init,
1297 .cio_iter_fini = osc_io_iter_fini,
1298 .cio_start = osc_io_setattr_start,
1299 .cio_end = osc_io_setattr_end
1301 [CIT_DATA_VERSION] = {
1302 .cio_start = osc_io_data_version_start,
1303 .cio_end = osc_io_data_version_end,
1306 .cio_iter_init = osc_io_iter_init,
1307 .cio_iter_fini = osc_io_iter_fini,
1308 .cio_start = osc_io_fault_start,
1309 .cio_end = osc_io_end,
1310 .cio_fini = osc_io_fini
1313 .cio_start = osc_io_fsync_start,
1314 .cio_end = osc_io_fsync_end,
1315 .cio_fini = osc_io_fini
1318 .cio_start = osc_io_ladvise_start,
1319 .cio_end = osc_io_ladvise_end,
1320 .cio_fini = osc_io_fini
1323 .cio_start = osc_io_lseek_start,
1324 .cio_end = osc_io_lseek_end,
1325 .cio_fini = osc_io_fini
1328 .cio_fini = osc_io_fini
1331 .cio_read_ahead = osc_io_read_ahead,
1332 .cio_lru_reserve = osc_io_lru_reserve,
1333 .cio_submit = osc_io_submit,
1334 .cio_commit_async = osc_io_commit_async,
1335 .cio_extent_release = osc_io_extent_release
1338 /*****************************************************************************
1340 * Transfer operations.
1344 int osc_io_init(const struct lu_env *env,
1345 struct cl_object *obj, struct cl_io *io)
1347 struct obd_export *exp = osc_export(cl2osc(obj));
1348 struct osc_io *oio = osc_env_io(env);
1350 CL_IO_SLICE_CLEAN(oio, oi_cl);
1351 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);
1353 if (!exp_connect_unaligned_dio(exp))
1354 cl_io_top(io)->ci_allow_unaligned_dio = false;