4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Implementation of cl_io for 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);
58 LINVRNT(oio == osc_env_io(env));
62 /*****************************************************************************
68 static void osc_io_fini(const struct lu_env *env, const struct cl_io_slice *io)
72 static void osc_read_ahead_release(const struct lu_env *env,
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;
92 dlmlock = osc_dlmlock_at_pgoff(env, osc, start, 0);
93 if (dlmlock != NULL) {
94 LASSERT(dlmlock->l_ast_data == osc);
95 if (dlmlock->l_req_mode != LCK_PR) {
96 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;
130 struct list_head list = LIST_HEAD_INIT(list);
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 != NULL)
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 == 0) {
215 struct cl_object *obj = ios->cis_obj;
216 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
218 cl_object_attr_lock(obj);
219 attr->cat_mtime = attr->cat_ctime = LTIME_S(CURRENT_TIME);
220 cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
221 cl_object_attr_unlock(obj);
224 CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
225 return qout->pl_nr > 0 ? 0 : result;
229 * This is called when a page is accessed within file in a way that creates
230 * new page, if one were missing (i.e., if there were a hole at that place in
231 * the file, or accessed page is beyond the current file size).
233 * Expand stripe KMS if necessary.
235 static void osc_page_touch_at(const struct lu_env *env,
236 struct cl_object *obj, pgoff_t idx, size_t to)
238 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
239 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
243 /* offset within stripe */
244 kms = cl_offset(obj, idx) + to;
246 cl_object_attr_lock(obj);
250 * ll_inode_size_lock(inode, 0); lov_stripe_lock(lsm);
254 CDEBUG(D_INODE, "stripe KMS %sincreasing %llu->%llu %llu\n",
255 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
256 loi->loi_lvb.lvb_size);
258 attr->cat_mtime = attr->cat_ctime = LTIME_S(CURRENT_TIME);
259 valid = CAT_MTIME | CAT_CTIME;
260 if (kms > loi->loi_kms) {
264 if (kms > loi->loi_lvb.lvb_size) {
265 attr->cat_size = kms;
268 cl_object_attr_update(env, obj, attr, valid);
269 cl_object_attr_unlock(obj);
272 static int osc_io_commit_async(const struct lu_env *env,
273 const struct cl_io_slice *ios,
274 struct cl_page_list *qin, int from, int to,
277 struct cl_io *io = ios->cis_io;
278 struct osc_io *oio = cl2osc_io(env, ios);
279 struct osc_object *osc = cl2osc(ios->cis_obj);
280 struct cl_page *page;
281 struct cl_page *last_page;
282 struct osc_page *opg;
286 LASSERT(qin->pl_nr > 0);
288 /* Handle partial page cases */
289 last_page = cl_page_list_last(qin);
290 if (oio->oi_lockless) {
291 page = cl_page_list_first(qin);
292 if (page == last_page) {
293 cl_page_clip(env, page, from, to);
296 cl_page_clip(env, page, from, PAGE_SIZE);
298 cl_page_clip(env, last_page, 0, to);
302 while (qin->pl_nr > 0) {
303 struct osc_async_page *oap;
305 page = cl_page_list_first(qin);
306 opg = osc_cl_page_osc(page, osc);
309 if (!list_empty(&oap->oap_rpc_item)) {
310 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
316 /* The page may be already in dirty cache. */
317 if (list_empty(&oap->oap_pending_item)) {
318 result = osc_page_cache_add(env, &opg->ops_cl, io);
323 osc_page_touch_at(env, osc2cl(osc), osc_index(opg),
324 page == last_page ? to : PAGE_SIZE);
326 cl_page_list_del(env, qin, page);
328 (*cb)(env, io, page);
329 /* Can't access page any more. Page can be in transfer and
330 * complete at any time. */
333 /* for sync write, kernel will wait for this page to be flushed before
334 * osc_io_end() is called, so release it earlier.
335 * for mkwrite(), it's known there is no further pages. */
336 if (cl_io_is_sync_write(io) && oio->oi_active != NULL) {
337 osc_extent_release(env, oio->oi_active);
338 oio->oi_active = NULL;
341 CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, result);
345 static int osc_io_iter_init(const struct lu_env *env,
346 const struct cl_io_slice *ios)
348 struct osc_object *osc = cl2osc(ios->cis_obj);
349 struct obd_import *imp = osc_cli(osc)->cl_import;
352 spin_lock(&imp->imp_lock);
353 if (likely(!imp->imp_invalid)) {
354 struct osc_io *oio = osc_env_io(env);
356 atomic_inc(&osc->oo_nr_ios);
357 oio->oi_is_active = 1;
360 spin_unlock(&imp->imp_lock);
365 static int osc_io_write_iter_init(const struct lu_env *env,
366 const struct cl_io_slice *ios)
368 struct cl_io *io = ios->cis_io;
369 struct osc_io *oio = osc_env_io(env);
370 struct osc_object *osc = cl2osc(ios->cis_obj);
371 unsigned long npages;
374 if (cl_io_is_append(io))
375 RETURN(osc_io_iter_init(env, ios));
377 npages = io->u.ci_rw.crw_count >> PAGE_SHIFT;
378 if (io->u.ci_rw.crw_pos & ~PAGE_MASK)
381 oio->oi_lru_reserved = osc_lru_reserve(osc_cli(osc), npages);
383 RETURN(osc_io_iter_init(env, ios));
386 static void osc_io_iter_fini(const struct lu_env *env,
387 const struct cl_io_slice *ios)
389 struct osc_io *oio = osc_env_io(env);
391 if (oio->oi_is_active) {
392 struct osc_object *osc = cl2osc(ios->cis_obj);
394 oio->oi_is_active = 0;
395 LASSERT(atomic_read(&osc->oo_nr_ios) > 0);
396 if (atomic_dec_and_test(&osc->oo_nr_ios))
397 wake_up_all(&osc->oo_io_waitq);
401 static void osc_io_write_iter_fini(const struct lu_env *env,
402 const struct cl_io_slice *ios)
404 struct osc_io *oio = osc_env_io(env);
405 struct osc_object *osc = cl2osc(ios->cis_obj);
407 if (oio->oi_lru_reserved > 0) {
408 osc_lru_unreserve(osc_cli(osc), oio->oi_lru_reserved);
409 oio->oi_lru_reserved = 0;
411 oio->oi_write_osclock = NULL;
413 osc_io_iter_fini(env, ios);
416 static int osc_io_fault_start(const struct lu_env *env,
417 const struct cl_io_slice *ios)
420 struct cl_fault_io *fio;
424 fio = &io->u.ci_fault;
425 CDEBUG(D_INFO, "%lu %d %zu\n",
426 fio->ft_index, fio->ft_writable, fio->ft_nob);
428 * If mapping is writeable, adjust kms to cover this page,
429 * but do not extend kms beyond actual file size.
432 if (fio->ft_writable)
433 osc_page_touch_at(env, ios->cis_obj,
434 fio->ft_index, fio->ft_nob);
438 static int osc_async_upcall(void *a, int rc)
440 struct osc_async_cbargs *args = a;
443 complete(&args->opc_sync);
448 * Checks that there are no pages being written in the extent being truncated.
450 static int trunc_check_cb(const struct lu_env *env, struct cl_io *io,
451 struct osc_page *ops , void *cbdata)
453 struct cl_page *page = ops->ops_cl.cpl_page;
454 struct osc_async_page *oap;
455 __u64 start = *(__u64 *)cbdata;
458 if (oap->oap_cmd & OBD_BRW_WRITE &&
459 !list_empty(&oap->oap_pending_item))
460 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %llu/%s.\n",
461 start, current->comm);
463 if (PageLocked(page->cp_vmpage))
464 CDEBUG(D_CACHE, "page %p index %lu locked for %d.\n",
465 ops, osc_index(ops), oap->oap_cmd & OBD_BRW_RWMASK);
467 return CLP_GANG_OKAY;
470 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
471 struct osc_io *oio, __u64 size)
473 struct cl_object *clob;
477 clob = oio->oi_cl.cis_obj;
478 start = cl_index(clob, size);
479 partial = cl_offset(clob, start) < size;
482 * Complain if there are pages in the truncated region.
484 osc_page_gang_lookup(env, io, cl2osc(clob),
485 start + partial, CL_PAGE_EOF,
486 trunc_check_cb, (void *)&size);
489 static int osc_io_setattr_start(const struct lu_env *env,
490 const struct cl_io_slice *slice)
492 struct cl_io *io = slice->cis_io;
493 struct osc_io *oio = cl2osc_io(env, slice);
494 struct cl_object *obj = slice->cis_obj;
495 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
496 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
497 struct obdo *oa = &oio->oi_oa;
498 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
499 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
500 unsigned int ia_valid = io->u.ci_setattr.sa_valid;
503 /* truncate cache dirty pages first */
504 if (cl_io_is_trunc(io))
505 result = osc_cache_truncate_start(env, cl2osc(obj), size,
508 if (result == 0 && oio->oi_lockless == 0) {
509 cl_object_attr_lock(obj);
510 result = cl_object_attr_get(env, obj, attr);
512 struct ost_lvb *lvb = &io->u.ci_setattr.sa_attr;
513 unsigned int cl_valid = 0;
515 if (ia_valid & ATTR_SIZE) {
516 attr->cat_size = attr->cat_kms = size;
517 cl_valid = (CAT_SIZE | CAT_KMS);
519 if (ia_valid & ATTR_MTIME_SET) {
520 attr->cat_mtime = lvb->lvb_mtime;
521 cl_valid |= CAT_MTIME;
523 if (ia_valid & ATTR_ATIME_SET) {
524 attr->cat_atime = lvb->lvb_atime;
525 cl_valid |= CAT_ATIME;
527 if (ia_valid & ATTR_CTIME_SET) {
528 attr->cat_ctime = lvb->lvb_ctime;
529 cl_valid |= CAT_CTIME;
531 result = cl_object_attr_update(env, obj, attr,
534 cl_object_attr_unlock(obj);
536 memset(oa, 0, sizeof(*oa));
538 oa->o_oi = loi->loi_oi;
539 obdo_set_parent_fid(oa, io->u.ci_setattr.sa_parent_fid);
540 oa->o_stripe_idx = io->u.ci_setattr.sa_stripe_index;
541 oa->o_layout = io->u.ci_setattr.sa_layout;
542 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);
569 if (ia_valid & ATTR_ATTR_FLAG) {
570 oa->o_flags = io->u.ci_setattr.sa_attr_flags;
571 oa->o_valid |= OBD_MD_FLFLAGS;
574 init_completion(&cbargs->opc_sync);
576 if (ia_valid & ATTR_SIZE)
577 result = osc_punch_base(osc_export(cl2osc(obj)),
578 oa, osc_async_upcall,
579 cbargs, PTLRPCD_SET);
581 result = osc_setattr_async(osc_export(cl2osc(obj)),
582 oa, osc_async_upcall,
583 cbargs, PTLRPCD_SET);
585 cbargs->opc_rpc_sent = result == 0;
591 static void osc_io_setattr_end(const struct lu_env *env,
592 const struct cl_io_slice *slice)
594 struct cl_io *io = slice->cis_io;
595 struct osc_io *oio = cl2osc_io(env, slice);
596 struct cl_object *obj = slice->cis_obj;
597 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
600 if (cbargs->opc_rpc_sent) {
601 wait_for_completion(&cbargs->opc_sync);
602 result = io->ci_result = cbargs->opc_rc;
605 if (oio->oi_lockless) {
606 /* lockless truncate */
607 struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
609 LASSERT(cl_io_is_trunc(io));
610 /* XXX: Need a lock. */
611 osd->od_stats.os_lockless_truncates++;
615 if (cl_io_is_trunc(io)) {
616 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
617 osc_trunc_check(env, io, oio, size);
618 osc_cache_truncate_end(env, oio->oi_trunc);
619 oio->oi_trunc = NULL;
623 struct osc_data_version_args {
624 struct osc_io *dva_oio;
628 osc_data_version_interpret(const struct lu_env *env, struct ptlrpc_request *req,
631 struct osc_data_version_args *dva = arg;
632 struct osc_io *oio = dva->dva_oio;
633 const struct ost_body *body;
639 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
641 GOTO(out, rc = -EPROTO);
643 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, &oio->oi_oa,
647 oio->oi_cbarg.opc_rc = rc;
648 complete(&oio->oi_cbarg.opc_sync);
653 static int osc_io_data_version_start(const struct lu_env *env,
654 const struct cl_io_slice *slice)
656 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
657 struct osc_io *oio = cl2osc_io(env, slice);
658 struct obdo *oa = &oio->oi_oa;
659 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
660 struct osc_object *obj = cl2osc(slice->cis_obj);
661 struct lov_oinfo *loi = obj->oo_oinfo;
662 struct obd_export *exp = osc_export(obj);
663 struct ptlrpc_request *req;
664 struct ost_body *body;
665 struct osc_data_version_args *dva;
669 memset(oa, 0, sizeof(*oa));
670 oa->o_oi = loi->loi_oi;
671 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
673 if (dv->dv_flags & (LL_DV_RD_FLUSH | LL_DV_WR_FLUSH)) {
674 oa->o_valid |= OBD_MD_FLFLAGS;
675 oa->o_flags |= OBD_FL_SRVLOCK;
676 if (dv->dv_flags & LL_DV_WR_FLUSH)
677 oa->o_flags |= OBD_FL_FLUSH;
680 init_completion(&cbargs->opc_sync);
682 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
686 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
688 ptlrpc_request_free(req);
692 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
693 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
695 ptlrpc_request_set_replen(req);
696 req->rq_interpret_reply = osc_data_version_interpret;
697 CLASSERT(sizeof(*dva) <= sizeof(req->rq_async_args));
698 dva = ptlrpc_req_async_args(req);
701 ptlrpcd_add_req(req);
706 static void osc_io_data_version_end(const struct lu_env *env,
707 const struct cl_io_slice *slice)
709 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
710 struct osc_io *oio = cl2osc_io(env, slice);
711 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
714 wait_for_completion(&cbargs->opc_sync);
716 if (cbargs->opc_rc != 0) {
717 slice->cis_io->ci_result = cbargs->opc_rc;
718 } else if (!(oio->oi_oa.o_valid & OBD_MD_FLDATAVERSION)) {
719 slice->cis_io->ci_result = -EOPNOTSUPP;
721 dv->dv_data_version = oio->oi_oa.o_data_version;
722 slice->cis_io->ci_result = 0;
728 static int osc_io_read_start(const struct lu_env *env,
729 const struct cl_io_slice *slice)
731 struct cl_object *obj = slice->cis_obj;
732 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
736 if (!slice->cis_io->ci_noatime) {
737 cl_object_attr_lock(obj);
738 attr->cat_atime = ktime_get_real_seconds();
739 rc = cl_object_attr_update(env, obj, attr, CAT_ATIME);
740 cl_object_attr_unlock(obj);
746 static int osc_io_write_start(const struct lu_env *env,
747 const struct cl_io_slice *slice)
749 struct cl_object *obj = slice->cis_obj;
750 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
754 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_SETTIME, 1);
755 cl_object_attr_lock(obj);
756 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
757 rc = cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
758 cl_object_attr_unlock(obj);
763 static int osc_fsync_ost(const struct lu_env *env, struct osc_object *obj,
764 struct cl_fsync_io *fio)
766 struct osc_io *oio = osc_env_io(env);
767 struct obdo *oa = &oio->oi_oa;
768 struct lov_oinfo *loi = obj->oo_oinfo;
769 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
773 memset(oa, 0, sizeof(*oa));
774 oa->o_oi = loi->loi_oi;
775 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
777 /* reload size abd blocks for start and end of sync range */
778 oa->o_size = fio->fi_start;
779 oa->o_blocks = fio->fi_end;
780 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
782 obdo_set_parent_fid(oa, fio->fi_fid);
784 init_completion(&cbargs->opc_sync);
786 rc = osc_sync_base(obj, oa, osc_async_upcall, cbargs, PTLRPCD_SET);
790 static int osc_io_fsync_start(const struct lu_env *env,
791 const struct cl_io_slice *slice)
793 struct cl_io *io = slice->cis_io;
794 struct cl_fsync_io *fio = &io->u.ci_fsync;
795 struct cl_object *obj = slice->cis_obj;
796 struct osc_object *osc = cl2osc(obj);
797 pgoff_t start = cl_index(obj, fio->fi_start);
798 pgoff_t end = cl_index(obj, fio->fi_end);
802 if (fio->fi_end == OBD_OBJECT_EOF)
805 result = osc_cache_writeback_range(env, osc, start, end, 0,
806 fio->fi_mode == CL_FSYNC_DISCARD);
808 fio->fi_nr_written += result;
811 if (fio->fi_mode == CL_FSYNC_ALL) {
814 /* we have to wait for writeback to finish before we can
815 * send OST_SYNC RPC. This is bad because it causes extents
816 * to be written osc by osc. However, we usually start
817 * writeback before CL_FSYNC_ALL so this won't have any real
819 rc = osc_cache_wait_range(env, osc, start, end);
822 rc = osc_fsync_ost(env, osc, fio);
830 static void osc_io_fsync_end(const struct lu_env *env,
831 const struct cl_io_slice *slice)
833 struct cl_fsync_io *fio = &slice->cis_io->u.ci_fsync;
834 struct cl_object *obj = slice->cis_obj;
835 pgoff_t start = cl_index(obj, fio->fi_start);
836 pgoff_t end = cl_index(obj, fio->fi_end);
839 if (fio->fi_mode == CL_FSYNC_LOCAL) {
840 result = osc_cache_wait_range(env, cl2osc(obj), start, end);
841 } else if (fio->fi_mode == CL_FSYNC_ALL) {
842 struct osc_io *oio = cl2osc_io(env, slice);
843 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
845 wait_for_completion(&cbargs->opc_sync);
847 result = cbargs->opc_rc;
849 slice->cis_io->ci_result = result;
852 static int osc_io_ladvise_start(const struct lu_env *env,
853 const struct cl_io_slice *slice)
856 struct cl_io *io = slice->cis_io;
857 struct osc_io *oio = cl2osc_io(env, slice);
858 struct cl_object *obj = slice->cis_obj;
859 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
860 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
861 struct obdo *oa = &oio->oi_oa;
862 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
863 struct lu_ladvise *ladvise;
864 struct ladvise_hdr *ladvise_hdr;
869 /* TODO: add multiple ladvise support in CLIO */
870 buf_size = offsetof(typeof(*ladvise_hdr), lah_advise[num_advise]);
871 if (osc_env_info(env)->oti_ladvise_buf.lb_len < buf_size)
872 lu_buf_realloc(&osc_env_info(env)->oti_ladvise_buf, buf_size);
874 ladvise_hdr = osc_env_info(env)->oti_ladvise_buf.lb_buf;
875 if (ladvise_hdr == NULL)
878 memset(ladvise_hdr, 0, buf_size);
879 ladvise_hdr->lah_magic = LADVISE_MAGIC;
880 ladvise_hdr->lah_count = num_advise;
881 ladvise_hdr->lah_flags = lio->li_flags;
883 memset(oa, 0, sizeof(*oa));
884 oa->o_oi = loi->loi_oi;
885 oa->o_valid = OBD_MD_FLID;
886 obdo_set_parent_fid(oa, lio->li_fid);
888 ladvise = ladvise_hdr->lah_advise;
889 ladvise->lla_start = lio->li_start;
890 ladvise->lla_end = lio->li_end;
891 ladvise->lla_advice = lio->li_advice;
893 if (lio->li_flags & LF_ASYNC) {
894 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
895 ladvise_hdr, NULL, NULL, NULL);
897 init_completion(&cbargs->opc_sync);
898 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
899 ladvise_hdr, osc_async_upcall,
900 cbargs, PTLRPCD_SET);
901 cbargs->opc_rpc_sent = result == 0;
906 static void osc_io_ladvise_end(const struct lu_env *env,
907 const struct cl_io_slice *slice)
909 struct cl_io *io = slice->cis_io;
910 struct osc_io *oio = cl2osc_io(env, slice);
911 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
913 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
915 if ((!(lio->li_flags & LF_ASYNC)) && cbargs->opc_rpc_sent) {
916 wait_for_completion(&cbargs->opc_sync);
917 result = cbargs->opc_rc;
919 slice->cis_io->ci_result = result;
922 static void osc_io_end(const struct lu_env *env,
923 const struct cl_io_slice *slice)
925 struct osc_io *oio = cl2osc_io(env, slice);
927 if (oio->oi_active) {
928 osc_extent_release(env, oio->oi_active);
929 oio->oi_active = NULL;
933 static const struct cl_io_operations osc_io_ops = {
936 .cio_iter_init = osc_io_iter_init,
937 .cio_iter_fini = osc_io_iter_fini,
938 .cio_start = osc_io_read_start,
939 .cio_fini = osc_io_fini
942 .cio_iter_init = osc_io_write_iter_init,
943 .cio_iter_fini = osc_io_write_iter_fini,
944 .cio_start = osc_io_write_start,
945 .cio_end = osc_io_end,
946 .cio_fini = osc_io_fini
949 .cio_iter_init = osc_io_iter_init,
950 .cio_iter_fini = osc_io_iter_fini,
951 .cio_start = osc_io_setattr_start,
952 .cio_end = osc_io_setattr_end
954 [CIT_DATA_VERSION] = {
955 .cio_start = osc_io_data_version_start,
956 .cio_end = osc_io_data_version_end,
959 .cio_iter_init = osc_io_iter_init,
960 .cio_iter_fini = osc_io_iter_fini,
961 .cio_start = osc_io_fault_start,
962 .cio_end = osc_io_end,
963 .cio_fini = osc_io_fini
966 .cio_start = osc_io_fsync_start,
967 .cio_end = osc_io_fsync_end,
968 .cio_fini = osc_io_fini
971 .cio_start = osc_io_ladvise_start,
972 .cio_end = osc_io_ladvise_end,
973 .cio_fini = osc_io_fini
976 .cio_fini = osc_io_fini
979 .cio_read_ahead = osc_io_read_ahead,
980 .cio_submit = osc_io_submit,
981 .cio_commit_async = osc_io_commit_async
984 /*****************************************************************************
986 * Transfer operations.
990 int osc_io_init(const struct lu_env *env,
991 struct cl_object *obj, struct cl_io *io)
993 struct osc_io *oio = osc_env_io(env);
995 CL_IO_SLICE_CLEAN(oio, oi_cl);
996 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);