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"
50 /*****************************************************************************
56 static void osc_io_fini(const struct lu_env *env, const struct cl_io_slice *io)
60 void osc_read_ahead_release(const struct lu_env *env, struct cl_read_ahead *ra)
62 struct ldlm_lock *dlmlock = ra->cra_dlmlock;
63 struct osc_io *oio = ra->cra_oio;
64 struct lustre_handle lockh;
66 oio->oi_is_readahead = 0;
67 ldlm_lock2handle(dlmlock, &lockh);
68 ldlm_lock_decref(&lockh, LCK_PR);
69 LDLM_LOCK_PUT(dlmlock);
71 EXPORT_SYMBOL(osc_read_ahead_release);
73 static int osc_io_read_ahead(const struct lu_env *env,
74 const struct cl_io_slice *ios,
75 pgoff_t start, struct cl_read_ahead *ra)
77 struct osc_object *osc = cl2osc(ios->cis_obj);
78 struct osc_io *oio = cl2osc_io(env, ios);
79 struct ldlm_lock *dlmlock;
80 int result = -ENODATA;
84 oio->oi_is_readahead = true;
85 dlmlock = osc_dlmlock_at_pgoff(env, osc, start, 0);
86 if (dlmlock != NULL) {
87 LASSERT(dlmlock->l_ast_data == osc);
88 if (dlmlock->l_req_mode != LCK_PR) {
89 struct lustre_handle lockh;
90 ldlm_lock2handle(dlmlock, &lockh);
91 ldlm_lock_addref(&lockh, LCK_PR);
92 ldlm_lock_decref(&lockh, dlmlock->l_req_mode);
95 ra->cra_rpc_pages = osc_cli(osc)->cl_max_pages_per_rpc;
96 ra->cra_end_idx = cl_index(osc2cl(osc),
97 dlmlock->l_policy_data.l_extent.end);
98 ra->cra_release = osc_read_ahead_release;
99 ra->cra_dlmlock = dlmlock;
101 if (ra->cra_end_idx != CL_PAGE_EOF)
102 ra->cra_contention = true;
110 * An implementation of cl_io_operations::cio_io_submit() method for osc
111 * layer. Iterates over pages in the in-queue, prepares each for io by calling
112 * cl_page_prep() and then either submits them through osc_io_submit_page()
113 * or, if page is already submitted, changes osc flags through
114 * osc_set_async_flags().
116 int osc_io_submit(const struct lu_env *env, const struct cl_io_slice *ios,
117 enum cl_req_type crt, struct cl_2queue *queue)
119 struct cl_page *page;
121 struct client_obd *cli = NULL;
122 struct osc_object *osc = NULL; /* to keep gcc happy */
123 struct osc_page *opg;
127 struct cl_page_list *qin = &queue->c2_qin;
128 struct cl_page_list *qout = &queue->c2_qout;
129 unsigned int queued = 0;
132 unsigned int max_pages;
133 unsigned int ppc_bits; /* pages per chunk bits */
135 bool sync_queue = false;
137 LASSERT(qin->pl_nr > 0);
139 CDEBUG(D_CACHE|D_READA, "%d %d\n", qin->pl_nr, crt);
141 osc = cl2osc(ios->cis_obj);
143 max_pages = cli->cl_max_pages_per_rpc;
144 ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
147 brw_flags = osc_io_srvlock(cl2osc_io(env, ios)) ? OBD_BRW_SRVLOCK : 0;
148 brw_flags |= crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
149 if (crt == CRT_READ && ios->cis_io->ci_ndelay)
150 brw_flags |= OBD_BRW_NDELAY;
152 page = cl_page_list_first(qin);
153 if (page->cp_type == CPT_TRANSIENT)
154 brw_flags |= OBD_BRW_NOCACHE;
157 * NOTE: here @page is a top-level page. This is done to avoid
158 * creation of sub-page-list.
160 cl_page_list_for_each_safe(page, tmp, qin) {
161 struct osc_async_page *oap;
167 opg = osc_cl_page_osc(page, osc);
169 LASSERT(osc == oap->oap_obj);
171 if (!list_empty(&oap->oap_pending_item) ||
172 !list_empty(&oap->oap_rpc_item)) {
173 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
179 result = cl_page_prep(env, io, page, crt);
182 if (result != -EALREADY)
185 * Handle -EALREADY error: for read case, the page is
186 * already in UPTODATE state; for write, the page
193 spin_lock(&oap->oap_lock);
194 oap->oap_async_flags = ASYNC_URGENT|ASYNC_READY;
195 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
196 spin_unlock(&oap->oap_lock);
198 osc_page_submit(env, opg, crt, brw_flags);
199 list_add_tail(&oap->oap_pending_item, &list);
201 if (page->cp_sync_io != NULL)
202 cl_page_list_move(qout, qin, page);
204 cl_page_list_del(env, qin, page);
207 if (queued == max_pages) {
209 } else if (crt == CRT_WRITE) {
211 unsigned int next_chunks;
213 chunks = (queued + ppc - 1) >> ppc_bits;
214 /* chunk number if add another page */
215 next_chunks = (queued + ppc) >> ppc_bits;
217 /* next page will excceed write chunk limit */
218 if (chunks == osc_max_write_chunks(cli) &&
219 next_chunks > chunks)
224 result = osc_queue_sync_pages(env, io, osc, &list,
234 result = osc_queue_sync_pages(env, io, osc, &list, brw_flags);
236 /* Update c/mtime for sync write. LU-7310 */
237 if (crt == CRT_WRITE && qout->pl_nr > 0 && result == 0) {
238 struct cl_object *obj = ios->cis_obj;
239 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
241 cl_object_attr_lock(obj);
242 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
243 cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
244 cl_object_attr_unlock(obj);
247 CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
248 return qout->pl_nr > 0 ? 0 : result;
250 EXPORT_SYMBOL(osc_io_submit);
253 * This is called to update the attributes when modifying a specific page,
254 * both when making new pages and when doing updates to existing cached pages.
256 * Expand stripe KMS if necessary.
258 void osc_page_touch_at(const struct lu_env *env, struct cl_object *obj,
259 pgoff_t idx, size_t to)
261 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
262 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
268 /* offset within stripe */
269 kms = cl_offset(obj, idx) + to;
271 cl_object_attr_lock(obj);
272 CDEBUG(D_INODE, "stripe KMS %sincreasing %llu->%llu %llu\n",
273 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
274 loi->loi_lvb.lvb_size);
276 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
277 valid = CAT_MTIME | CAT_CTIME;
278 if (kms > loi->loi_kms) {
282 if (kms > loi->loi_lvb.lvb_size) {
283 attr->cat_size = kms;
286 cl_object_attr_update(env, obj, attr, valid);
287 cl_object_attr_unlock(obj);
292 int osc_io_commit_async(const struct lu_env *env,
293 const struct cl_io_slice *ios,
294 struct cl_page_list *qin, int from, int to,
297 struct cl_io *io = ios->cis_io;
298 struct osc_io *oio = cl2osc_io(env, ios);
299 struct osc_object *osc = cl2osc(ios->cis_obj);
300 struct cl_page *page;
301 struct cl_page *last_page;
302 struct osc_page *opg;
303 struct pagevec *pvec = &osc_env_info(env)->oti_pagevec;
307 LASSERT(qin->pl_nr > 0);
309 /* Handle partial page cases */
310 last_page = cl_page_list_last(qin);
311 if (oio->oi_lockless) {
312 page = cl_page_list_first(qin);
313 if (page == last_page) {
314 cl_page_clip(env, page, from, to);
317 cl_page_clip(env, page, from, PAGE_SIZE);
319 cl_page_clip(env, last_page, 0, to);
323 ll_pagevec_init(pvec, 0);
325 while (qin->pl_nr > 0) {
326 struct osc_async_page *oap;
328 page = cl_page_list_first(qin);
329 opg = osc_cl_page_osc(page, osc);
332 LASSERTF(osc == oap->oap_obj,
333 "obj mismatch: %p / %p\n", osc, oap->oap_obj);
335 if (!list_empty(&oap->oap_rpc_item)) {
336 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
342 /* The page may be already in dirty cache. */
343 if (list_empty(&oap->oap_pending_item)) {
344 result = osc_page_cache_add(env, opg, io, cb);
349 osc_page_touch_at(env, osc2cl(osc), osc_index(opg),
350 page == last_page ? to : PAGE_SIZE);
352 cl_page_list_del(env, qin, page);
354 /* if there are no more slots, do the callback & reinit */
355 if (pagevec_add(pvec, page->cp_vmpage) == 0) {
356 (*cb)(env, io, pvec);
357 pagevec_reinit(pvec);
361 /* Clean up any partially full pagevecs */
362 if (pagevec_count(pvec) != 0)
363 (*cb)(env, io, pvec);
365 /* Can't access these pages any more. Page can be in transfer and
366 * complete at any time. */
368 /* for sync write, kernel will wait for this page to be flushed before
369 * osc_io_end() is called, so release it earlier.
370 * for mkwrite(), it's known there is no further pages. */
371 if (cl_io_is_sync_write(io) && oio->oi_active != NULL) {
372 osc_extent_release(env, oio->oi_active);
373 oio->oi_active = NULL;
376 CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, result);
379 EXPORT_SYMBOL(osc_io_commit_async);
381 void osc_io_extent_release(const struct lu_env *env,
382 const struct cl_io_slice *ios)
384 struct osc_io *oio = cl2osc_io(env, ios);
386 if (oio->oi_active != NULL) {
387 osc_extent_release(env, oio->oi_active);
388 oio->oi_active = NULL;
391 EXPORT_SYMBOL(osc_io_extent_release);
393 static bool osc_import_not_healthy(struct obd_import *imp)
395 return imp->imp_invalid || imp->imp_deactive ||
396 !(imp->imp_state == LUSTRE_IMP_FULL ||
397 imp->imp_state == LUSTRE_IMP_IDLE);
400 int osc_io_iter_init(const struct lu_env *env, const struct cl_io_slice *ios)
402 struct osc_object *osc = cl2osc(ios->cis_obj);
403 struct obd_import *imp = osc_cli(osc)->cl_import;
404 struct osc_io *oio = osc_env_io(env);
409 spin_lock(&imp->imp_lock);
411 * check whether this OSC device is available for non-delay read,
412 * fast switching mirror if we haven't tried all mirrors.
414 if (ios->cis_io->ci_type == CIT_READ && ios->cis_io->ci_ndelay &&
415 !ios->cis_io->ci_tried_all_mirrors && osc_import_not_healthy(imp)) {
417 } else if (likely(!imp->imp_invalid)) {
418 atomic_inc(&osc->oo_nr_ios);
419 oio->oi_is_active = 1;
422 spin_unlock(&imp->imp_lock);
424 if (capable(CAP_SYS_RESOURCE))
425 oio->oi_cap_sys_resource = 1;
429 EXPORT_SYMBOL(osc_io_iter_init);
431 void osc_io_iter_fini(const struct lu_env *env,
432 const struct cl_io_slice *ios)
434 struct osc_io *oio = osc_env_io(env);
436 if (oio->oi_is_active) {
437 struct osc_object *osc = cl2osc(ios->cis_obj);
439 oio->oi_is_active = 0;
440 LASSERT(atomic_read(&osc->oo_nr_ios) > 0);
441 if (atomic_dec_and_test(&osc->oo_nr_ios))
442 wake_up(&osc->oo_io_waitq);
445 EXPORT_SYMBOL(osc_io_iter_fini);
447 void osc_io_rw_iter_fini(const struct lu_env *env,
448 const struct cl_io_slice *ios)
450 struct osc_io *oio = osc_env_io(env);
451 struct osc_object *osc = cl2osc(ios->cis_obj);
453 if (oio->oi_lru_reserved > 0) {
454 osc_lru_unreserve(osc_cli(osc), oio->oi_lru_reserved);
455 oio->oi_lru_reserved = 0;
457 oio->oi_write_osclock = NULL;
459 osc_io_iter_fini(env, ios);
461 EXPORT_SYMBOL(osc_io_rw_iter_fini);
463 int osc_io_fault_start(const struct lu_env *env, const struct cl_io_slice *ios)
466 struct cl_fault_io *fio;
470 fio = &io->u.ci_fault;
471 CDEBUG(D_INFO, "%lu %d %zu\n",
472 fio->ft_index, fio->ft_writable, fio->ft_nob);
474 * If mapping is writeable, adjust kms to cover this page,
475 * but do not extend kms beyond actual file size.
478 if (fio->ft_writable)
479 osc_page_touch_at(env, ios->cis_obj,
480 fio->ft_index, fio->ft_nob);
483 EXPORT_SYMBOL(osc_io_fault_start);
486 static int osc_async_upcall(void *a, int rc)
488 struct osc_async_cbargs *args = a;
491 complete(&args->opc_sync);
496 * Checks that there are no pages being written in the extent being truncated.
498 static bool trunc_check_cb(const struct lu_env *env, struct cl_io *io,
499 void **pvec, int count, void *cbdata)
503 for (i = 0; i < count; i++) {
504 struct osc_page *ops = pvec[i];
505 struct cl_page *page = ops->ops_cl.cpl_page;
506 struct osc_async_page *oap;
507 __u64 start = *(__u64 *)cbdata;
510 if (oap->oap_cmd & OBD_BRW_WRITE &&
511 !list_empty(&oap->oap_pending_item))
512 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %llu/%s.\n",
513 start, current->comm);
515 if (PageLocked(page->cp_vmpage))
516 CDEBUG(D_CACHE, "page %p index %lu locked for %d.\n",
518 oap->oap_cmd & OBD_BRW_RWMASK);
523 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
524 struct osc_io *oio, __u64 size)
526 struct cl_object *clob;
530 clob = oio->oi_cl.cis_obj;
531 start = cl_index(clob, size);
532 partial = cl_offset(clob, start) < size;
535 * Complain if there are pages in the truncated region.
537 osc_page_gang_lookup(env, io, cl2osc(clob),
538 start + partial, CL_PAGE_EOF,
539 trunc_check_cb, (void *)&size);
543 * Flush affected pages prior punch.
544 * We shouldn't discard them locally first because that could be data loss
545 * if server doesn't support fallocate punch, we also need these data to be
546 * flushed first to prevent re-ordering with the punch
548 static int osc_punch_start(const struct lu_env *env, struct cl_io *io,
549 struct cl_object *obj)
551 struct osc_object *osc = cl2osc(obj);
552 pgoff_t pg_start = cl_index(obj, io->u.ci_setattr.sa_falloc_offset);
553 pgoff_t pg_end = cl_index(obj, io->u.ci_setattr.sa_falloc_end - 1);
557 rc = osc_cache_writeback_range(env, osc, pg_start, pg_end, 1, 0);
561 osc_page_gang_lookup(env, io, osc, pg_start, pg_end, osc_discard_cb,
566 static int osc_io_setattr_start(const struct lu_env *env,
567 const struct cl_io_slice *slice)
569 struct cl_io *io = slice->cis_io;
570 struct osc_io *oio = cl2osc_io(env, slice);
571 struct cl_object *obj = slice->cis_obj;
572 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
573 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
574 struct obdo *oa = &oio->oi_oa;
575 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
576 unsigned int ia_avalid = io->u.ci_setattr.sa_avalid;
577 enum op_xvalid ia_xvalid = io->u.ci_setattr.sa_xvalid;
579 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
580 bool io_is_falloc = cl_io_is_fallocate(io);
583 /* truncate cache dirty pages first */
584 if (cl_io_is_trunc(io))
585 result = osc_cache_truncate_start(env, cl2osc(obj), size,
587 /* flush local pages prior punching them on server */
589 io->u.ci_setattr.sa_falloc_mode & FALLOC_FL_PUNCH_HOLE)
590 result = osc_punch_start(env, io, obj);
592 if (result == 0 && oio->oi_lockless == 0) {
593 cl_object_attr_lock(obj);
594 result = cl_object_attr_get(env, obj, attr);
596 struct ost_lvb *lvb = &io->u.ci_setattr.sa_attr;
597 unsigned int cl_valid = 0;
599 if (ia_avalid & ATTR_SIZE) {
600 attr->cat_size = size;
601 attr->cat_kms = size;
602 cl_valid = (CAT_SIZE | CAT_KMS);
604 if (ia_avalid & ATTR_MTIME_SET) {
605 attr->cat_mtime = lvb->lvb_mtime;
606 cl_valid |= CAT_MTIME;
608 if (ia_avalid & ATTR_ATIME_SET) {
609 attr->cat_atime = lvb->lvb_atime;
610 cl_valid |= CAT_ATIME;
612 if (ia_xvalid & OP_XVALID_CTIME_SET) {
613 attr->cat_ctime = lvb->lvb_ctime;
614 cl_valid |= CAT_CTIME;
616 result = cl_object_attr_update(env, obj, attr,
619 cl_object_attr_unlock(obj);
621 memset(oa, 0, sizeof(*oa));
623 oa->o_oi = loi->loi_oi;
624 obdo_set_parent_fid(oa, io->u.ci_setattr.sa_parent_fid);
625 oa->o_stripe_idx = io->u.ci_setattr.sa_stripe_index;
626 oa->o_layout = io->u.ci_setattr.sa_layout;
627 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP |
629 if (ia_avalid & ATTR_CTIME) {
630 oa->o_valid |= OBD_MD_FLCTIME;
631 oa->o_ctime = attr->cat_ctime;
633 if (ia_avalid & ATTR_ATIME) {
634 oa->o_valid |= OBD_MD_FLATIME;
635 oa->o_atime = attr->cat_atime;
637 if (ia_avalid & ATTR_MTIME) {
638 oa->o_valid |= OBD_MD_FLMTIME;
639 oa->o_mtime = attr->cat_mtime;
642 if (ia_avalid & ATTR_SIZE || io_is_falloc) {
643 if (oio->oi_lockless) {
644 oa->o_flags = OBD_FL_SRVLOCK;
645 oa->o_valid |= OBD_MD_FLFLAGS;
648 if (io->ci_layout_version > 0) {
649 /* verify layout version */
650 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
651 oa->o_layout_version = io->ci_layout_version;
654 LASSERT(oio->oi_lockless == 0);
657 if (ia_xvalid & OP_XVALID_FLAGS) {
658 oa->o_flags = io->u.ci_setattr.sa_attr_flags;
659 oa->o_valid |= OBD_MD_FLFLAGS;
662 init_completion(&cbargs->opc_sync);
665 int falloc_mode = io->u.ci_setattr.sa_falloc_mode;
667 oa->o_size = io->u.ci_setattr.sa_falloc_offset;
668 oa->o_blocks = io->u.ci_setattr.sa_falloc_end;
669 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
670 result = osc_fallocate_base(osc_export(cl2osc(obj)),
671 oa, osc_async_upcall,
672 cbargs, falloc_mode);
673 } else if (ia_avalid & ATTR_SIZE) {
675 oa->o_blocks = OBD_OBJECT_EOF;
676 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
677 result = osc_punch_send(osc_export(cl2osc(obj)),
678 oa, osc_async_upcall, cbargs);
680 result = osc_setattr_async(osc_export(cl2osc(obj)),
681 oa, osc_async_upcall,
682 cbargs, PTLRPCD_SET);
684 cbargs->opc_rpc_sent = result == 0;
690 void osc_io_setattr_end(const struct lu_env *env,
691 const struct cl_io_slice *slice)
693 struct cl_io *io = slice->cis_io;
694 struct osc_io *oio = cl2osc_io(env, slice);
695 struct cl_object *obj = slice->cis_obj;
696 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
697 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
698 struct obdo *oa = &oio->oi_oa;
699 unsigned int cl_valid = 0;
702 if (cbargs->opc_rpc_sent) {
703 wait_for_completion(&cbargs->opc_sync);
704 result = io->ci_result = cbargs->opc_rc;
708 if (oio->oi_lockless) {
709 /* lockless truncate */
710 struct osc_device *osc = lu2osc_dev(obj->co_lu.lo_dev);
712 LASSERT(cl_io_is_trunc(io) || cl_io_is_fallocate(io));
713 /* XXX: Need a lock. */
714 osc->od_stats.os_lockless_truncates++;
718 if (cl_io_is_trunc(io)) {
719 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
722 cl_object_attr_lock(obj);
723 if (oa->o_valid & OBD_MD_FLBLOCKS) {
724 attr->cat_blocks = oa->o_blocks;
725 cl_valid |= CAT_BLOCKS;
728 cl_object_attr_update(env, obj, attr, cl_valid);
729 cl_object_attr_unlock(obj);
731 osc_trunc_check(env, io, oio, size);
732 osc_cache_truncate_end(env, oio->oi_trunc);
733 oio->oi_trunc = NULL;
736 if (cl_io_is_fallocate(io)) {
738 cl_object_attr_lock(obj);
740 if (oa->o_valid & OBD_MD_FLBLOCKS) {
741 attr->cat_blocks = oa->o_blocks;
742 cl_valid |= CAT_BLOCKS;
745 cl_object_attr_update(env, obj, attr, cl_valid);
746 cl_object_attr_unlock(obj);
750 EXPORT_SYMBOL(osc_io_setattr_end);
752 struct osc_data_version_args {
753 struct osc_io *dva_oio;
757 osc_data_version_interpret(const struct lu_env *env, struct ptlrpc_request *req,
760 struct osc_data_version_args *dva = args;
761 struct osc_io *oio = dva->dva_oio;
762 const struct ost_body *body;
768 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
770 GOTO(out, rc = -EPROTO);
772 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, &oio->oi_oa,
776 oio->oi_cbarg.opc_rc = rc;
777 complete(&oio->oi_cbarg.opc_sync);
782 static int osc_io_data_version_start(const struct lu_env *env,
783 const struct cl_io_slice *slice)
785 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
786 struct osc_io *oio = cl2osc_io(env, slice);
787 struct obdo *oa = &oio->oi_oa;
788 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
789 struct osc_object *obj = cl2osc(slice->cis_obj);
790 struct lov_oinfo *loi = obj->oo_oinfo;
791 struct obd_export *exp = osc_export(obj);
792 struct ptlrpc_request *req;
793 struct ost_body *body;
794 struct osc_data_version_args *dva;
798 memset(oa, 0, sizeof(*oa));
799 oa->o_oi = loi->loi_oi;
800 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
802 if (dv->dv_flags & (LL_DV_RD_FLUSH | LL_DV_WR_FLUSH)) {
803 oa->o_valid |= OBD_MD_FLFLAGS;
804 oa->o_flags |= OBD_FL_SRVLOCK;
805 if (dv->dv_flags & LL_DV_WR_FLUSH)
806 oa->o_flags |= OBD_FL_FLUSH;
809 init_completion(&cbargs->opc_sync);
811 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
815 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
817 ptlrpc_request_free(req);
821 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
822 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
824 ptlrpc_request_set_replen(req);
825 req->rq_interpret_reply = osc_data_version_interpret;
826 dva = ptlrpc_req_async_args(dva, req);
829 ptlrpcd_add_req(req);
834 static void osc_io_data_version_end(const struct lu_env *env,
835 const struct cl_io_slice *slice)
837 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
838 struct osc_io *oio = cl2osc_io(env, slice);
839 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
842 wait_for_completion(&cbargs->opc_sync);
844 if (cbargs->opc_rc != 0) {
845 slice->cis_io->ci_result = cbargs->opc_rc;
847 slice->cis_io->ci_result = 0;
848 if (!(oio->oi_oa.o_valid &
849 (OBD_MD_LAYOUT_VERSION | OBD_MD_FLDATAVERSION)))
850 slice->cis_io->ci_result = -ENOTSUPP;
852 if (oio->oi_oa.o_valid & OBD_MD_LAYOUT_VERSION)
853 dv->dv_layout_version = oio->oi_oa.o_layout_version;
854 if (oio->oi_oa.o_valid & OBD_MD_FLDATAVERSION)
855 dv->dv_data_version = oio->oi_oa.o_data_version;
861 int osc_io_read_start(const struct lu_env *env,
862 const struct cl_io_slice *slice)
864 struct cl_object *obj = slice->cis_obj;
865 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
869 if (!slice->cis_io->ci_noatime) {
870 cl_object_attr_lock(obj);
871 attr->cat_atime = ktime_get_real_seconds();
872 rc = cl_object_attr_update(env, obj, attr, CAT_ATIME);
873 cl_object_attr_unlock(obj);
878 EXPORT_SYMBOL(osc_io_read_start);
880 int osc_io_write_start(const struct lu_env *env,
881 const struct cl_io_slice *slice)
883 struct cl_object *obj = slice->cis_obj;
884 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
888 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_SETTIME, 1);
889 cl_object_attr_lock(obj);
890 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
891 rc = cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
892 cl_object_attr_unlock(obj);
896 EXPORT_SYMBOL(osc_io_write_start);
898 int osc_fsync_ost(const struct lu_env *env, struct osc_object *obj,
899 struct cl_fsync_io *fio)
901 struct osc_io *oio = osc_env_io(env);
902 struct obdo *oa = &oio->oi_oa;
903 struct lov_oinfo *loi = obj->oo_oinfo;
904 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
908 memset(oa, 0, sizeof(*oa));
909 oa->o_oi = loi->loi_oi;
910 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
912 /* reload size abd blocks for start and end of sync range */
913 oa->o_size = fio->fi_start;
914 oa->o_blocks = fio->fi_end;
915 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
917 obdo_set_parent_fid(oa, fio->fi_fid);
919 init_completion(&cbargs->opc_sync);
921 rc = osc_sync_base(obj, oa, osc_async_upcall, cbargs, PTLRPCD_SET);
924 EXPORT_SYMBOL(osc_fsync_ost);
926 int osc_io_fsync_start(const struct lu_env *env,
927 const struct cl_io_slice *slice)
929 struct cl_io *io = slice->cis_io;
930 struct cl_fsync_io *fio = &io->u.ci_fsync;
931 struct cl_object *obj = slice->cis_obj;
932 struct osc_object *osc = cl2osc(obj);
933 pgoff_t start = cl_index(obj, fio->fi_start);
934 pgoff_t end = cl_index(obj, fio->fi_end);
938 if (fio->fi_end == OBD_OBJECT_EOF)
941 result = osc_cache_writeback_range(env, osc, start, end, 0,
942 fio->fi_mode == CL_FSYNC_DISCARD);
944 fio->fi_nr_written += result;
947 if (fio->fi_mode == CL_FSYNC_ALL) {
950 /* we have to wait for writeback to finish before we can
951 * send OST_SYNC RPC. This is bad because it causes extents
952 * to be written osc by osc. However, we usually start
953 * writeback before CL_FSYNC_ALL so this won't have any real
955 rc = osc_cache_wait_range(env, osc, start, end);
958 rc = osc_fsync_ost(env, osc, fio);
966 void osc_io_fsync_end(const struct lu_env *env,
967 const struct cl_io_slice *slice)
969 struct cl_fsync_io *fio = &slice->cis_io->u.ci_fsync;
970 struct cl_object *obj = slice->cis_obj;
971 pgoff_t start = cl_index(obj, fio->fi_start);
972 pgoff_t end = cl_index(obj, fio->fi_end);
975 if (fio->fi_mode == CL_FSYNC_LOCAL) {
976 result = osc_cache_wait_range(env, cl2osc(obj), start, end);
977 } else if (fio->fi_mode == CL_FSYNC_ALL) {
978 struct osc_io *oio = cl2osc_io(env, slice);
979 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
981 wait_for_completion(&cbargs->opc_sync);
983 result = cbargs->opc_rc;
985 slice->cis_io->ci_result = result;
987 EXPORT_SYMBOL(osc_io_fsync_end);
989 static int osc_io_ladvise_start(const struct lu_env *env,
990 const struct cl_io_slice *slice)
993 struct cl_io *io = slice->cis_io;
994 struct osc_io *oio = cl2osc_io(env, slice);
995 struct cl_object *obj = slice->cis_obj;
996 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
997 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
998 struct obdo *oa = &oio->oi_oa;
999 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1000 struct lu_ladvise *ladvise;
1001 struct ladvise_hdr *ladvise_hdr;
1006 /* TODO: add multiple ladvise support in CLIO */
1007 buf_size = offsetof(typeof(*ladvise_hdr), lah_advise[num_advise]);
1008 if (osc_env_info(env)->oti_ladvise_buf.lb_len < buf_size)
1009 lu_buf_realloc(&osc_env_info(env)->oti_ladvise_buf, buf_size);
1011 ladvise_hdr = osc_env_info(env)->oti_ladvise_buf.lb_buf;
1012 if (ladvise_hdr == NULL)
1015 memset(ladvise_hdr, 0, buf_size);
1016 ladvise_hdr->lah_magic = LADVISE_MAGIC;
1017 ladvise_hdr->lah_count = num_advise;
1018 ladvise_hdr->lah_flags = lio->li_flags;
1020 memset(oa, 0, sizeof(*oa));
1021 oa->o_oi = loi->loi_oi;
1022 oa->o_valid = OBD_MD_FLID;
1023 obdo_set_parent_fid(oa, lio->li_fid);
1025 ladvise = ladvise_hdr->lah_advise;
1026 ladvise->lla_start = lio->li_start;
1027 ladvise->lla_end = lio->li_end;
1028 ladvise->lla_advice = lio->li_advice;
1030 if (lio->li_flags & LF_ASYNC) {
1031 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1032 ladvise_hdr, NULL, NULL, NULL);
1034 init_completion(&cbargs->opc_sync);
1035 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1036 ladvise_hdr, osc_async_upcall,
1037 cbargs, PTLRPCD_SET);
1038 cbargs->opc_rpc_sent = result == 0;
1043 static void osc_io_ladvise_end(const struct lu_env *env,
1044 const struct cl_io_slice *slice)
1046 struct cl_io *io = slice->cis_io;
1047 struct osc_io *oio = cl2osc_io(env, slice);
1048 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1050 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
1052 if ((!(lio->li_flags & LF_ASYNC)) && cbargs->opc_rpc_sent) {
1053 wait_for_completion(&cbargs->opc_sync);
1054 result = cbargs->opc_rc;
1056 slice->cis_io->ci_result = result;
1059 void osc_io_end(const struct lu_env *env, const struct cl_io_slice *slice)
1061 struct osc_io *oio = cl2osc_io(env, slice);
1063 if (oio->oi_active) {
1064 osc_extent_release(env, oio->oi_active);
1065 oio->oi_active = NULL;
1068 EXPORT_SYMBOL(osc_io_end);
1070 struct osc_lseek_args {
1071 struct osc_io *lsa_oio;
1074 static int osc_lseek_interpret(const struct lu_env *env,
1075 struct ptlrpc_request *req,
1078 struct ost_body *reply;
1079 struct osc_lseek_args *lsa = arg;
1080 struct osc_io *oio = lsa->lsa_oio;
1081 struct cl_io *io = oio->oi_cl.cis_io;
1082 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1089 reply = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1091 GOTO(out, rc = -EPROTO);
1093 lsio->ls_result = reply->oa.o_size;
1095 osc_async_upcall(&oio->oi_cbarg, rc);
1099 int osc_io_lseek_start(const struct lu_env *env,
1100 const struct cl_io_slice *slice)
1102 struct cl_io *io = slice->cis_io;
1103 struct osc_io *oio = cl2osc_io(env, slice);
1104 struct cl_object *obj = slice->cis_obj;
1105 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1106 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1107 struct obdo *oa = &oio->oi_oa;
1108 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1109 struct obd_export *exp = osc_export(cl2osc(obj));
1110 struct ptlrpc_request *req;
1111 struct ost_body *body;
1112 struct osc_lseek_args *lsa;
1117 /* No negative values at this point */
1118 LASSERT(lsio->ls_start >= 0);
1119 LASSERT(lsio->ls_whence == SEEK_HOLE || lsio->ls_whence == SEEK_DATA);
1121 /* with IO lock taken we have object size in LVB and can check
1122 * boundaries prior sending LSEEK RPC
1124 if (lsio->ls_start >= loi->loi_lvb.lvb_size) {
1125 /* consider area beyond end of object as hole */
1126 if (lsio->ls_whence == SEEK_HOLE)
1127 lsio->ls_result = lsio->ls_start;
1129 lsio->ls_result = -ENXIO;
1133 /* if LSEEK RPC is not supported by server, consider whole stripe
1134 * object is data with hole after end of object
1136 if (!exp_connect_lseek(exp)) {
1137 if (lsio->ls_whence == SEEK_HOLE)
1138 lsio->ls_result = loi->loi_lvb.lvb_size;
1140 lsio->ls_result = lsio->ls_start;
1144 memset(oa, 0, sizeof(*oa));
1145 oa->o_oi = loi->loi_oi;
1146 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1147 oa->o_size = lsio->ls_start;
1148 oa->o_mode = lsio->ls_whence;
1149 if (oio->oi_lockless) {
1150 oa->o_flags = OBD_FL_SRVLOCK;
1151 oa->o_valid |= OBD_MD_FLFLAGS;
1154 init_completion(&cbargs->opc_sync);
1155 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SEEK);
1159 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SEEK);
1161 ptlrpc_request_free(req);
1165 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1166 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1167 ptlrpc_request_set_replen(req);
1168 req->rq_interpret_reply = osc_lseek_interpret;
1169 lsa = ptlrpc_req_async_args(lsa, req);
1172 ptlrpcd_add_req(req);
1173 cbargs->opc_rpc_sent = 1;
1177 EXPORT_SYMBOL(osc_io_lseek_start);
1179 void osc_io_lseek_end(const struct lu_env *env,
1180 const struct cl_io_slice *slice)
1182 struct osc_io *oio = cl2osc_io(env, slice);
1183 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1186 if (cbargs->opc_rpc_sent) {
1187 wait_for_completion(&cbargs->opc_sync);
1188 rc = cbargs->opc_rc;
1190 slice->cis_io->ci_result = rc;
1192 EXPORT_SYMBOL(osc_io_lseek_end);
1194 int osc_io_lru_reserve(const struct lu_env *env,
1195 const struct cl_io_slice *ios,
1196 loff_t pos, size_t bytes)
1198 struct osc_object *osc = cl2osc(ios->cis_obj);
1199 struct osc_io *oio = osc_env_io(env);
1200 unsigned long npages = 0;
1205 page_offset = pos & ~PAGE_MASK;
1208 if (bytes > PAGE_SIZE - page_offset)
1209 bytes -= (PAGE_SIZE - page_offset);
1213 npages += (bytes + PAGE_SIZE - 1) >> PAGE_SHIFT;
1214 oio->oi_lru_reserved = osc_lru_reserve(osc_cli(osc), npages);
1218 EXPORT_SYMBOL(osc_io_lru_reserve);
1220 static const struct cl_io_operations osc_io_ops = {
1223 .cio_iter_init = osc_io_iter_init,
1224 .cio_iter_fini = osc_io_rw_iter_fini,
1225 .cio_start = osc_io_read_start,
1226 .cio_fini = osc_io_fini
1229 .cio_iter_init = osc_io_iter_init,
1230 .cio_iter_fini = osc_io_rw_iter_fini,
1231 .cio_start = osc_io_write_start,
1232 .cio_end = osc_io_end,
1233 .cio_fini = osc_io_fini
1236 .cio_iter_init = osc_io_iter_init,
1237 .cio_iter_fini = osc_io_iter_fini,
1238 .cio_start = osc_io_setattr_start,
1239 .cio_end = osc_io_setattr_end
1241 [CIT_DATA_VERSION] = {
1242 .cio_start = osc_io_data_version_start,
1243 .cio_end = osc_io_data_version_end,
1246 .cio_iter_init = osc_io_iter_init,
1247 .cio_iter_fini = osc_io_iter_fini,
1248 .cio_start = osc_io_fault_start,
1249 .cio_end = osc_io_end,
1250 .cio_fini = osc_io_fini
1253 .cio_start = osc_io_fsync_start,
1254 .cio_end = osc_io_fsync_end,
1255 .cio_fini = osc_io_fini
1258 .cio_start = osc_io_ladvise_start,
1259 .cio_end = osc_io_ladvise_end,
1260 .cio_fini = osc_io_fini
1263 .cio_start = osc_io_lseek_start,
1264 .cio_end = osc_io_lseek_end,
1265 .cio_fini = osc_io_fini
1268 .cio_fini = osc_io_fini
1271 .cio_read_ahead = osc_io_read_ahead,
1272 .cio_lru_reserve = osc_io_lru_reserve,
1273 .cio_submit = osc_io_submit,
1274 .cio_commit_async = osc_io_commit_async,
1275 .cio_extent_release = osc_io_extent_release
1278 /*****************************************************************************
1280 * Transfer operations.
1284 int osc_io_init(const struct lu_env *env,
1285 struct cl_object *obj, struct cl_io *io)
1287 struct osc_io *oio = osc_env_io(env);
1289 CL_IO_SLICE_CLEAN(oio, oi_cl);
1290 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);