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, 2015, 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 OSC layer.
34 * Author: Nikita Danilov <nikita.danilov@sun.com>
35 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
38 #define DEBUG_SUBSYSTEM S_OSC
40 #include <lustre_obdo.h>
42 #include "osc_cl_internal.h"
48 /*****************************************************************************
54 static struct osc_io *cl2osc_io(const struct lu_env *env,
55 const struct cl_io_slice *slice)
57 struct osc_io *oio = container_of0(slice, struct osc_io, oi_cl);
59 LINVRNT(oio == osc_env_io(env));
63 /*****************************************************************************
69 static void osc_io_fini(const struct lu_env *env, const struct cl_io_slice *io)
73 static void osc_read_ahead_release(const struct lu_env *env, void *cbdata)
75 struct ldlm_lock *dlmlock = cbdata;
76 struct lustre_handle lockh;
78 ldlm_lock2handle(dlmlock, &lockh);
79 ldlm_lock_decref(&lockh, LCK_PR);
80 LDLM_LOCK_PUT(dlmlock);
83 static int osc_io_read_ahead(const struct lu_env *env,
84 const struct cl_io_slice *ios,
85 pgoff_t start, struct cl_read_ahead *ra)
87 struct osc_object *osc = cl2osc(ios->cis_obj);
88 struct ldlm_lock *dlmlock;
89 int result = -ENODATA;
91 dlmlock = osc_dlmlock_at_pgoff(env, osc, start, 0);
93 LASSERT(dlmlock->l_ast_data == osc);
94 if (dlmlock->l_req_mode != LCK_PR) {
95 struct lustre_handle lockh;
97 ldlm_lock2handle(dlmlock, &lockh);
98 ldlm_lock_addref(&lockh, LCK_PR);
99 ldlm_lock_decref(&lockh, dlmlock->l_req_mode);
102 ra->cra_rpc_size = osc_cli(osc)->cl_max_pages_per_rpc;
103 ra->cra_end = cl_index(osc2cl(osc),
104 dlmlock->l_policy_data.l_extent.end);
105 ra->cra_release = osc_read_ahead_release;
106 ra->cra_cbdata = dlmlock;
114 * An implementation of cl_io_operations::cio_io_submit() method for osc
115 * layer. Iterates over pages in the in-queue, prepares each for io by calling
116 * cl_page_prep() and then either submits them through osc_io_submit_page()
117 * or, if page is already submitted, changes osc flags through
118 * osc_set_async_flags().
120 static int osc_io_submit(const struct lu_env *env,
121 const struct cl_io_slice *ios,
122 enum cl_req_type crt, struct cl_2queue *queue)
124 struct cl_page *page;
126 struct client_obd *cli = NULL;
127 struct osc_object *osc = NULL; /* to keep gcc happy */
128 struct osc_page *opg;
132 struct cl_page_list *qin = &queue->c2_qin;
133 struct cl_page_list *qout = &queue->c2_qout;
134 unsigned int queued = 0;
138 unsigned int max_pages;
140 LASSERT(qin->pl_nr > 0);
142 CDEBUG(D_CACHE | D_READA, "%d %d\n", qin->pl_nr, crt);
144 osc = cl2osc(ios->cis_obj);
146 max_pages = cli->cl_max_pages_per_rpc;
148 cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
149 brw_flags = osc_io_srvlock(cl2osc_io(env, ios)) ? OBD_BRW_SRVLOCK : 0;
152 * NOTE: here @page is a top-level page. This is done to avoid
153 * creation of sub-page-list.
155 cl_page_list_for_each_safe(page, tmp, qin) {
156 struct osc_async_page *oap;
162 opg = osc_cl_page_osc(page, osc);
164 LASSERT(osc == oap->oap_obj);
166 if (!list_empty(&oap->oap_pending_item) ||
167 !list_empty(&oap->oap_rpc_item)) {
168 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
174 result = cl_page_prep(env, io, page, crt);
177 if (result != -EALREADY)
180 * Handle -EALREADY error: for read case, the page is
181 * already in UPTODATE state; for write, the page
188 spin_lock(&oap->oap_lock);
189 oap->oap_async_flags = ASYNC_URGENT | ASYNC_READY;
190 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
191 spin_unlock(&oap->oap_lock);
193 osc_page_submit(env, opg, crt, brw_flags);
194 list_add_tail(&oap->oap_pending_item, &list);
196 if (page->cp_sync_io)
197 cl_page_list_move(qout, qin, page);
199 cl_page_list_del(env, qin, page);
201 if (++queued == max_pages) {
203 result = osc_queue_sync_pages(env, osc, &list, cmd,
211 result = osc_queue_sync_pages(env, osc, &list, cmd, brw_flags);
213 /* Update c/mtime for sync write. LU-7310 */
214 if (qout->pl_nr > 0 && !result) {
215 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
216 struct cl_object *obj = ios->cis_obj;
218 cl_object_attr_lock(obj);
219 attr->cat_mtime = ktime_get_real_seconds();
220 attr->cat_ctime = attr->cat_mtime;
221 cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
222 cl_object_attr_unlock(obj);
225 CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
226 return qout->pl_nr > 0 ? 0 : result;
230 * This is called when a page is accessed within file in a way that creates
231 * new page, if one were missing (i.e., if there were a hole at that place in
232 * the file, or accessed page is beyond the current file size).
234 * Expand stripe KMS if necessary.
236 static void osc_page_touch_at(const struct lu_env *env,
237 struct cl_object *obj, pgoff_t idx, size_t to)
239 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
240 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
244 /* offset within stripe */
245 kms = cl_offset(obj, idx) + to;
247 cl_object_attr_lock(obj);
251 * ll_inode_size_lock(inode, 0); lov_stripe_lock(lsm);
255 CDEBUG(D_INODE, "stripe KMS %sincreasing %llu->%llu %llu\n",
256 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
257 loi->loi_lvb.lvb_size);
259 attr->cat_ctime = ktime_get_real_seconds();
260 attr->cat_mtime = attr->cat_ctime;
261 valid = CAT_MTIME | CAT_CTIME;
262 if (kms > loi->loi_kms) {
266 if (kms > loi->loi_lvb.lvb_size) {
267 attr->cat_size = kms;
270 cl_object_attr_update(env, obj, attr, valid);
271 cl_object_attr_unlock(obj);
274 static int osc_io_commit_async(const struct lu_env *env,
275 const struct cl_io_slice *ios,
276 struct cl_page_list *qin, int from, int to,
279 struct cl_io *io = ios->cis_io;
280 struct osc_io *oio = cl2osc_io(env, ios);
281 struct osc_object *osc = cl2osc(ios->cis_obj);
282 struct cl_page *page;
283 struct cl_page *last_page;
284 struct osc_page *opg;
287 LASSERT(qin->pl_nr > 0);
289 /* Handle partial page cases */
290 last_page = cl_page_list_last(qin);
291 if (oio->oi_lockless) {
292 page = cl_page_list_first(qin);
293 if (page == last_page) {
294 cl_page_clip(env, page, from, to);
297 cl_page_clip(env, page, from, PAGE_SIZE);
299 cl_page_clip(env, last_page, 0, to);
303 while (qin->pl_nr > 0) {
304 struct osc_async_page *oap;
306 page = cl_page_list_first(qin);
307 opg = osc_cl_page_osc(page, osc);
310 if (!list_empty(&oap->oap_rpc_item)) {
311 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
317 /* The page may be already in dirty cache. */
318 if (list_empty(&oap->oap_pending_item)) {
319 result = osc_page_cache_add(env, &opg->ops_cl, io);
324 osc_page_touch_at(env, osc2cl(osc), osc_index(opg),
325 page == last_page ? to : PAGE_SIZE);
327 cl_page_list_del(env, qin, page);
329 (*cb)(env, io, page);
330 /* Can't access page any more. Page can be in transfer and
331 * complete at any time.
335 /* for sync write, kernel will wait for this page to be flushed before
336 * osc_io_end() is called, so release it earlier.
337 * for mkwrite(), it's known there is no further pages.
339 if (cl_io_is_sync_write(io) && oio->oi_active) {
340 osc_extent_release(env, oio->oi_active);
341 oio->oi_active = NULL;
344 CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, result);
348 static int osc_io_iter_init(const struct lu_env *env,
349 const struct cl_io_slice *ios)
351 struct osc_object *osc = cl2osc(ios->cis_obj);
352 struct obd_import *imp = osc_cli(osc)->cl_import;
355 spin_lock(&imp->imp_lock);
356 if (likely(!imp->imp_invalid)) {
357 struct osc_io *oio = osc_env_io(env);
359 atomic_inc(&osc->oo_nr_ios);
360 oio->oi_is_active = 1;
363 spin_unlock(&imp->imp_lock);
368 static int osc_io_write_iter_init(const struct lu_env *env,
369 const struct cl_io_slice *ios)
371 struct cl_io *io = ios->cis_io;
372 struct osc_io *oio = osc_env_io(env);
373 struct osc_object *osc = cl2osc(ios->cis_obj);
374 unsigned long npages;
376 if (cl_io_is_append(io))
377 return osc_io_iter_init(env, ios);
379 npages = io->u.ci_rw.crw_count >> PAGE_SHIFT;
380 if (io->u.ci_rw.crw_pos & ~PAGE_MASK)
383 oio->oi_lru_reserved = osc_lru_reserve(osc_cli(osc), npages);
385 return osc_io_iter_init(env, ios);
388 static void osc_io_iter_fini(const struct lu_env *env,
389 const struct cl_io_slice *ios)
391 struct osc_io *oio = osc_env_io(env);
393 if (oio->oi_is_active) {
394 struct osc_object *osc = cl2osc(ios->cis_obj);
396 oio->oi_is_active = 0;
397 LASSERT(atomic_read(&osc->oo_nr_ios) > 0);
398 if (atomic_dec_and_test(&osc->oo_nr_ios))
399 wake_up_all(&osc->oo_io_waitq);
403 static void osc_io_write_iter_fini(const struct lu_env *env,
404 const struct cl_io_slice *ios)
406 struct osc_io *oio = osc_env_io(env);
407 struct osc_object *osc = cl2osc(ios->cis_obj);
409 if (oio->oi_lru_reserved > 0) {
410 osc_lru_unreserve(osc_cli(osc), oio->oi_lru_reserved);
411 oio->oi_lru_reserved = 0;
413 oio->oi_write_osclock = NULL;
415 osc_io_iter_fini(env, ios);
418 static int osc_io_fault_start(const struct lu_env *env,
419 const struct cl_io_slice *ios)
422 struct cl_fault_io *fio;
425 fio = &io->u.ci_fault;
426 CDEBUG(D_INFO, "%lu %d %zu\n",
427 fio->ft_index, fio->ft_writable, fio->ft_nob);
429 * If mapping is writeable, adjust kms to cover this page,
430 * but do not extend kms beyond actual file size.
433 if (fio->ft_writable)
434 osc_page_touch_at(env, ios->cis_obj,
435 fio->ft_index, fio->ft_nob);
439 static int osc_async_upcall(void *a, int rc)
441 struct osc_async_cbargs *args = a;
444 complete(&args->opc_sync);
449 * Checks that there are no pages being written in the extent being truncated.
451 static int trunc_check_cb(const struct lu_env *env, struct cl_io *io,
452 struct osc_page *ops, void *cbdata)
454 struct cl_page *page = ops->ops_cl.cpl_page;
455 struct osc_async_page *oap;
456 __u64 start = *(__u64 *)cbdata;
459 if (oap->oap_cmd & OBD_BRW_WRITE &&
460 !list_empty(&oap->oap_pending_item))
461 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %llu/%s.\n",
462 start, current->comm);
464 if (PageLocked(page->cp_vmpage))
465 CDEBUG(D_CACHE, "page %p index %lu locked for %d.\n",
466 ops, osc_index(ops), oap->oap_cmd & OBD_BRW_RWMASK);
468 return CLP_GANG_OKAY;
471 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
472 struct osc_io *oio, __u64 size)
474 struct cl_object *clob;
478 clob = oio->oi_cl.cis_obj;
479 start = cl_index(clob, size);
480 partial = cl_offset(clob, start) < size;
483 * Complain if there are pages in the truncated region.
485 osc_page_gang_lookup(env, io, cl2osc(clob),
486 start + partial, CL_PAGE_EOF,
487 trunc_check_cb, (void *)&size);
490 static int osc_io_setattr_start(const struct lu_env *env,
491 const struct cl_io_slice *slice)
493 struct cl_io *io = slice->cis_io;
494 struct osc_io *oio = cl2osc_io(env, slice);
495 struct cl_object *obj = slice->cis_obj;
496 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
497 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
498 struct obdo *oa = &oio->oi_oa;
499 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
500 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
501 unsigned int ia_valid = io->u.ci_setattr.sa_valid;
504 /* truncate cache dirty pages first */
505 if (cl_io_is_trunc(io))
506 result = osc_cache_truncate_start(env, cl2osc(obj), size,
509 if (result == 0 && oio->oi_lockless == 0) {
510 cl_object_attr_lock(obj);
511 result = cl_object_attr_get(env, obj, attr);
513 struct ost_lvb *lvb = &io->u.ci_setattr.sa_attr;
514 unsigned int cl_valid = 0;
516 if (ia_valid & ATTR_SIZE) {
517 attr->cat_size = size;
518 attr->cat_kms = size;
519 cl_valid = CAT_SIZE | CAT_KMS;
521 if (ia_valid & ATTR_MTIME_SET) {
522 attr->cat_mtime = lvb->lvb_mtime;
523 cl_valid |= CAT_MTIME;
525 if (ia_valid & ATTR_ATIME_SET) {
526 attr->cat_atime = lvb->lvb_atime;
527 cl_valid |= CAT_ATIME;
529 if (ia_valid & ATTR_CTIME_SET) {
530 attr->cat_ctime = lvb->lvb_ctime;
531 cl_valid |= CAT_CTIME;
533 result = cl_object_attr_update(env, obj, attr,
536 cl_object_attr_unlock(obj);
538 memset(oa, 0, sizeof(*oa));
540 oa->o_oi = loi->loi_oi;
541 obdo_set_parent_fid(oa, io->u.ci_setattr.sa_parent_fid);
542 oa->o_stripe_idx = io->u.ci_setattr.sa_stripe_index;
543 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
544 if (ia_valid & ATTR_CTIME) {
545 oa->o_valid |= OBD_MD_FLCTIME;
546 oa->o_ctime = attr->cat_ctime;
548 if (ia_valid & ATTR_ATIME) {
549 oa->o_valid |= OBD_MD_FLATIME;
550 oa->o_atime = attr->cat_atime;
552 if (ia_valid & ATTR_MTIME) {
553 oa->o_valid |= OBD_MD_FLMTIME;
554 oa->o_mtime = attr->cat_mtime;
556 if (ia_valid & ATTR_SIZE) {
558 oa->o_blocks = OBD_OBJECT_EOF;
559 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
561 if (oio->oi_lockless) {
562 oa->o_flags = OBD_FL_SRVLOCK;
563 oa->o_valid |= OBD_MD_FLFLAGS;
566 LASSERT(oio->oi_lockless == 0);
568 if (ia_valid & ATTR_ATTR_FLAG) {
569 oa->o_flags = io->u.ci_setattr.sa_attr_flags;
570 oa->o_valid |= OBD_MD_FLFLAGS;
573 init_completion(&cbargs->opc_sync);
575 if (ia_valid & ATTR_SIZE)
576 result = osc_punch_base(osc_export(cl2osc(obj)),
577 oa, osc_async_upcall,
578 cbargs, PTLRPCD_SET);
580 result = osc_setattr_async(osc_export(cl2osc(obj)),
581 oa, osc_async_upcall,
582 cbargs, PTLRPCD_SET);
583 cbargs->opc_rpc_sent = result == 0;
588 static void osc_io_setattr_end(const struct lu_env *env,
589 const struct cl_io_slice *slice)
591 struct cl_io *io = slice->cis_io;
592 struct osc_io *oio = cl2osc_io(env, slice);
593 struct cl_object *obj = slice->cis_obj;
594 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
597 if (cbargs->opc_rpc_sent) {
598 wait_for_completion(&cbargs->opc_sync);
599 result = cbargs->opc_rc;
600 io->ci_result = cbargs->opc_rc;
603 if (oio->oi_lockless) {
604 /* lockless truncate */
605 struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
607 LASSERT(cl_io_is_trunc(io));
608 /* XXX: Need a lock. */
609 osd->od_stats.os_lockless_truncates++;
613 if (cl_io_is_trunc(io)) {
614 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
616 osc_trunc_check(env, io, oio, size);
617 osc_cache_truncate_end(env, oio->oi_trunc);
618 oio->oi_trunc = NULL;
622 struct osc_data_version_args {
623 struct osc_io *dva_oio;
627 osc_data_version_interpret(const struct lu_env *env, struct ptlrpc_request *req,
630 struct osc_data_version_args *dva = arg;
631 struct osc_io *oio = dva->dva_oio;
632 const struct ost_body *body;
637 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
643 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, &oio->oi_oa,
646 oio->oi_cbarg.opc_rc = rc;
647 complete(&oio->oi_cbarg.opc_sync);
652 static int osc_io_data_version_start(const struct lu_env *env,
653 const struct cl_io_slice *slice)
655 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
656 struct osc_io *oio = cl2osc_io(env, slice);
657 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
658 struct osc_object *obj = cl2osc(slice->cis_obj);
659 struct obd_export *exp = osc_export(obj);
660 struct lov_oinfo *loi = obj->oo_oinfo;
661 struct osc_data_version_args *dva;
662 struct obdo *oa = &oio->oi_oa;
663 struct ptlrpc_request *req;
664 struct ost_body *body;
667 memset(oa, 0, sizeof(*oa));
668 oa->o_oi = loi->loi_oi;
669 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
671 if (dv->dv_flags & (LL_DV_RD_FLUSH | LL_DV_WR_FLUSH)) {
672 oa->o_valid |= OBD_MD_FLFLAGS;
673 oa->o_flags |= OBD_FL_SRVLOCK;
674 if (dv->dv_flags & LL_DV_WR_FLUSH)
675 oa->o_flags |= OBD_FL_FLUSH;
678 init_completion(&cbargs->opc_sync);
680 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
684 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
686 ptlrpc_request_free(req);
690 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
691 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
693 ptlrpc_request_set_replen(req);
694 req->rq_interpret_reply = osc_data_version_interpret;
695 BUILD_BUG_ON(sizeof(*dva) > sizeof(req->rq_async_args));
696 dva = ptlrpc_req_async_args(req);
699 ptlrpcd_add_req(req);
704 static void osc_io_data_version_end(const struct lu_env *env,
705 const struct cl_io_slice *slice)
707 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
708 struct osc_io *oio = cl2osc_io(env, slice);
709 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
711 wait_for_completion(&cbargs->opc_sync);
713 if (cbargs->opc_rc) {
714 slice->cis_io->ci_result = cbargs->opc_rc;
715 } else if (!(oio->oi_oa.o_valid & OBD_MD_FLDATAVERSION)) {
716 slice->cis_io->ci_result = -EOPNOTSUPP;
718 dv->dv_data_version = oio->oi_oa.o_data_version;
719 slice->cis_io->ci_result = 0;
723 static int osc_io_read_start(const struct lu_env *env,
724 const struct cl_io_slice *slice)
726 struct cl_object *obj = slice->cis_obj;
727 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
730 if (!slice->cis_io->ci_noatime) {
731 cl_object_attr_lock(obj);
732 attr->cat_atime = ktime_get_real_seconds();
733 rc = cl_object_attr_update(env, obj, attr, CAT_ATIME);
734 cl_object_attr_unlock(obj);
739 static int osc_io_write_start(const struct lu_env *env,
740 const struct cl_io_slice *slice)
742 struct cl_object *obj = slice->cis_obj;
743 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
746 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_SETTIME, 1);
747 cl_object_attr_lock(obj);
748 attr->cat_ctime = ktime_get_real_seconds();
749 attr->cat_mtime = attr->cat_ctime;
750 rc = cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
751 cl_object_attr_unlock(obj);
756 static int osc_fsync_ost(const struct lu_env *env, struct osc_object *obj,
757 struct cl_fsync_io *fio)
759 struct osc_io *oio = osc_env_io(env);
760 struct obdo *oa = &oio->oi_oa;
761 struct lov_oinfo *loi = obj->oo_oinfo;
762 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
765 memset(oa, 0, sizeof(*oa));
766 oa->o_oi = loi->loi_oi;
767 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
769 /* reload size abd blocks for start and end of sync range */
770 oa->o_size = fio->fi_start;
771 oa->o_blocks = fio->fi_end;
772 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
774 obdo_set_parent_fid(oa, fio->fi_fid);
776 init_completion(&cbargs->opc_sync);
778 rc = osc_sync_base(obj, oa, osc_async_upcall, cbargs, PTLRPCD_SET);
782 static int osc_io_fsync_start(const struct lu_env *env,
783 const struct cl_io_slice *slice)
785 struct cl_io *io = slice->cis_io;
786 struct cl_fsync_io *fio = &io->u.ci_fsync;
787 struct cl_object *obj = slice->cis_obj;
788 struct osc_object *osc = cl2osc(obj);
789 pgoff_t start = cl_index(obj, fio->fi_start);
790 pgoff_t end = cl_index(obj, fio->fi_end);
793 if (fio->fi_end == OBD_OBJECT_EOF)
796 result = osc_cache_writeback_range(env, osc, start, end, 0,
797 fio->fi_mode == CL_FSYNC_DISCARD);
799 fio->fi_nr_written += result;
802 if (fio->fi_mode == CL_FSYNC_ALL) {
805 /* we have to wait for writeback to finish before we can
806 * send OST_SYNC RPC. This is bad because it causes extents
807 * to be written osc by osc. However, we usually start
808 * writeback before CL_FSYNC_ALL so this won't have any real
811 rc = osc_cache_wait_range(env, osc, start, end);
814 rc = osc_fsync_ost(env, osc, fio);
822 static void osc_io_fsync_end(const struct lu_env *env,
823 const struct cl_io_slice *slice)
825 struct cl_fsync_io *fio = &slice->cis_io->u.ci_fsync;
826 struct cl_object *obj = slice->cis_obj;
827 pgoff_t start = cl_index(obj, fio->fi_start);
828 pgoff_t end = cl_index(obj, fio->fi_end);
831 if (fio->fi_mode == CL_FSYNC_LOCAL) {
832 result = osc_cache_wait_range(env, cl2osc(obj), start, end);
833 } else if (fio->fi_mode == CL_FSYNC_ALL) {
834 struct osc_io *oio = cl2osc_io(env, slice);
835 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
837 wait_for_completion(&cbargs->opc_sync);
839 result = cbargs->opc_rc;
841 slice->cis_io->ci_result = result;
844 static void osc_io_end(const struct lu_env *env,
845 const struct cl_io_slice *slice)
847 struct osc_io *oio = cl2osc_io(env, slice);
849 if (oio->oi_active) {
850 osc_extent_release(env, oio->oi_active);
851 oio->oi_active = NULL;
855 static const struct cl_io_operations osc_io_ops = {
858 .cio_iter_init = osc_io_iter_init,
859 .cio_iter_fini = osc_io_iter_fini,
860 .cio_start = osc_io_read_start,
861 .cio_fini = osc_io_fini
864 .cio_iter_init = osc_io_write_iter_init,
865 .cio_iter_fini = osc_io_write_iter_fini,
866 .cio_start = osc_io_write_start,
867 .cio_end = osc_io_end,
868 .cio_fini = osc_io_fini
871 .cio_iter_init = osc_io_iter_init,
872 .cio_iter_fini = osc_io_iter_fini,
873 .cio_start = osc_io_setattr_start,
874 .cio_end = osc_io_setattr_end
876 [CIT_DATA_VERSION] = {
877 .cio_start = osc_io_data_version_start,
878 .cio_end = osc_io_data_version_end,
881 .cio_iter_init = osc_io_iter_init,
882 .cio_iter_fini = osc_io_iter_fini,
883 .cio_start = osc_io_fault_start,
884 .cio_end = osc_io_end,
885 .cio_fini = osc_io_fini
888 .cio_start = osc_io_fsync_start,
889 .cio_end = osc_io_fsync_end,
890 .cio_fini = osc_io_fini
893 .cio_fini = osc_io_fini
896 .cio_read_ahead = osc_io_read_ahead,
897 .cio_submit = osc_io_submit,
898 .cio_commit_async = osc_io_commit_async
901 /*****************************************************************************
903 * Transfer operations.
907 int osc_io_init(const struct lu_env *env,
908 struct cl_object *obj, struct cl_io *io)
910 struct osc_io *oio = osc_env_io(env);
912 CL_IO_SLICE_CLEAN(oio, oi_cl);
913 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);