4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Implementation of cl_io for 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>
41 #include <lustre_osc.h>
43 #include "osc_internal.h"
49 /*****************************************************************************
55 static void osc_io_fini(const struct lu_env *env, const struct cl_io_slice *io)
59 void osc_read_ahead_release(const struct lu_env *env, void *cbdata)
61 struct ldlm_lock *dlmlock = cbdata;
62 struct lustre_handle lockh;
64 ldlm_lock2handle(dlmlock, &lockh);
65 ldlm_lock_decref(&lockh, LCK_PR);
66 LDLM_LOCK_PUT(dlmlock);
68 EXPORT_SYMBOL(osc_read_ahead_release);
70 static int osc_io_read_ahead(const struct lu_env *env,
71 const struct cl_io_slice *ios,
72 pgoff_t start, struct cl_read_ahead *ra)
74 struct osc_object *osc = cl2osc(ios->cis_obj);
75 struct ldlm_lock *dlmlock;
76 int result = -ENODATA;
79 dlmlock = osc_dlmlock_at_pgoff(env, osc, start, 0);
80 if (dlmlock != NULL) {
81 LASSERT(dlmlock->l_ast_data == osc);
82 if (dlmlock->l_req_mode != LCK_PR) {
83 struct lustre_handle lockh;
84 ldlm_lock2handle(dlmlock, &lockh);
85 ldlm_lock_addref(&lockh, LCK_PR);
86 ldlm_lock_decref(&lockh, dlmlock->l_req_mode);
89 ra->cra_rpc_size = osc_cli(osc)->cl_max_pages_per_rpc;
90 ra->cra_end = cl_index(osc2cl(osc),
91 dlmlock->l_policy_data.l_extent.end);
92 ra->cra_release = osc_read_ahead_release;
93 ra->cra_cbdata = dlmlock;
101 * An implementation of cl_io_operations::cio_io_submit() method for osc
102 * layer. Iterates over pages in the in-queue, prepares each for io by calling
103 * cl_page_prep() and then either submits them through osc_io_submit_page()
104 * or, if page is already submitted, changes osc flags through
105 * osc_set_async_flags().
107 int osc_io_submit(const struct lu_env *env, const struct cl_io_slice *ios,
108 enum cl_req_type crt, struct cl_2queue *queue)
110 struct cl_page *page;
112 struct client_obd *cli = NULL;
113 struct osc_object *osc = NULL; /* to keep gcc happy */
114 struct osc_page *opg;
116 struct list_head list = LIST_HEAD_INIT(list);
118 struct cl_page_list *qin = &queue->c2_qin;
119 struct cl_page_list *qout = &queue->c2_qout;
120 unsigned int queued = 0;
123 unsigned int max_pages;
124 unsigned int ppc_bits; /* pages per chunk bits */
126 bool sync_queue = false;
128 LASSERT(qin->pl_nr > 0);
130 CDEBUG(D_CACHE|D_READA, "%d %d\n", qin->pl_nr, crt);
132 osc = cl2osc(ios->cis_obj);
134 max_pages = cli->cl_max_pages_per_rpc;
135 ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
138 brw_flags = osc_io_srvlock(cl2osc_io(env, ios)) ? OBD_BRW_SRVLOCK : 0;
139 brw_flags |= crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
140 if (crt == CRT_READ && ios->cis_io->ci_ndelay)
141 brw_flags |= OBD_BRW_NDELAY;
144 * NOTE: here @page is a top-level page. This is done to avoid
145 * creation of sub-page-list.
147 cl_page_list_for_each_safe(page, tmp, qin) {
148 struct osc_async_page *oap;
154 opg = osc_cl_page_osc(page, osc);
156 LASSERT(osc == oap->oap_obj);
158 if (!list_empty(&oap->oap_pending_item) ||
159 !list_empty(&oap->oap_rpc_item)) {
160 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
166 result = cl_page_prep(env, io, page, crt);
169 if (result != -EALREADY)
172 * Handle -EALREADY error: for read case, the page is
173 * already in UPTODATE state; for write, the page
180 spin_lock(&oap->oap_lock);
181 oap->oap_async_flags = ASYNC_URGENT|ASYNC_READY;
182 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
183 spin_unlock(&oap->oap_lock);
185 osc_page_submit(env, opg, crt, brw_flags);
186 list_add_tail(&oap->oap_pending_item, &list);
188 if (page->cp_sync_io != NULL)
189 cl_page_list_move(qout, qin, page);
191 cl_page_list_del(env, qin, page);
194 if (queued == max_pages) {
196 } else if (crt == CRT_WRITE) {
198 unsigned int next_chunks;
200 chunks = (queued + ppc - 1) >> ppc_bits;
201 /* chunk number if add another page */
202 next_chunks = (queued + ppc) >> ppc_bits;
204 /* next page will excceed write chunk limit */
205 if (chunks == osc_max_write_chunks(cli) &&
206 next_chunks > chunks)
211 result = osc_queue_sync_pages(env, io, osc, &list,
221 result = osc_queue_sync_pages(env, io, osc, &list, brw_flags);
223 /* Update c/mtime for sync write. LU-7310 */
224 if (crt == CRT_WRITE && qout->pl_nr > 0 && result == 0) {
225 struct cl_object *obj = ios->cis_obj;
226 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
228 cl_object_attr_lock(obj);
229 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
230 cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
231 cl_object_attr_unlock(obj);
234 CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
235 return qout->pl_nr > 0 ? 0 : result;
237 EXPORT_SYMBOL(osc_io_submit);
240 * This is called to update the attributes when modifying a specific page,
241 * both when making new pages and when doing updates to existing cached pages.
243 * Expand stripe KMS if necessary.
245 void osc_page_touch_at(const struct lu_env *env, struct cl_object *obj,
246 pgoff_t idx, size_t to)
248 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
249 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
255 /* offset within stripe */
256 kms = cl_offset(obj, idx) + to;
258 cl_object_attr_lock(obj);
259 CDEBUG(D_INODE, "stripe KMS %sincreasing %llu->%llu %llu\n",
260 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
261 loi->loi_lvb.lvb_size);
263 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
264 valid = CAT_MTIME | CAT_CTIME;
265 if (kms > loi->loi_kms) {
269 if (kms > loi->loi_lvb.lvb_size) {
270 attr->cat_size = kms;
273 cl_object_attr_update(env, obj, attr, valid);
274 cl_object_attr_unlock(obj);
279 int osc_io_commit_async(const struct lu_env *env,
280 const struct cl_io_slice *ios,
281 struct cl_page_list *qin, int from, int to,
284 struct cl_io *io = ios->cis_io;
285 struct osc_io *oio = cl2osc_io(env, ios);
286 struct osc_object *osc = cl2osc(ios->cis_obj);
287 struct cl_page *page;
288 struct cl_page *last_page;
289 struct osc_page *opg;
293 LASSERT(qin->pl_nr > 0);
295 /* Handle partial page cases */
296 last_page = cl_page_list_last(qin);
297 if (oio->oi_lockless) {
298 page = cl_page_list_first(qin);
299 if (page == last_page) {
300 cl_page_clip(env, page, from, to);
303 cl_page_clip(env, page, from, PAGE_SIZE);
305 cl_page_clip(env, last_page, 0, to);
309 while (qin->pl_nr > 0) {
310 struct osc_async_page *oap;
312 page = cl_page_list_first(qin);
313 opg = osc_cl_page_osc(page, osc);
316 LASSERTF(osc == oap->oap_obj,
317 "obj mismatch: %p / %p\n", osc, oap->oap_obj);
319 if (!list_empty(&oap->oap_rpc_item)) {
320 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
326 /* The page may be already in dirty cache. */
327 if (list_empty(&oap->oap_pending_item)) {
328 result = osc_page_cache_add(env, &opg->ops_cl, io);
333 osc_page_touch_at(env, osc2cl(osc), osc_index(opg),
334 page == last_page ? to : PAGE_SIZE);
336 cl_page_list_del(env, qin, page);
338 (*cb)(env, io, page);
339 /* Can't access page any more. Page can be in transfer and
340 * complete at any time. */
343 /* for sync write, kernel will wait for this page to be flushed before
344 * osc_io_end() is called, so release it earlier.
345 * for mkwrite(), it's known there is no further pages. */
346 if (cl_io_is_sync_write(io) && oio->oi_active != NULL) {
347 osc_extent_release(env, oio->oi_active);
348 oio->oi_active = NULL;
351 CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, result);
354 EXPORT_SYMBOL(osc_io_commit_async);
356 int osc_io_iter_init(const struct lu_env *env, const struct cl_io_slice *ios)
358 struct osc_object *osc = cl2osc(ios->cis_obj);
359 struct obd_import *imp = osc_cli(osc)->cl_import;
362 spin_lock(&imp->imp_lock);
363 if (likely(!imp->imp_invalid)) {
364 struct osc_io *oio = osc_env_io(env);
366 atomic_inc(&osc->oo_nr_ios);
367 oio->oi_is_active = 1;
370 spin_unlock(&imp->imp_lock);
374 EXPORT_SYMBOL(osc_io_iter_init);
376 int osc_io_write_iter_init(const struct lu_env *env,
377 const struct cl_io_slice *ios)
379 struct cl_io *io = ios->cis_io;
380 struct osc_io *oio = osc_env_io(env);
381 struct osc_object *osc = cl2osc(ios->cis_obj);
382 unsigned long npages;
385 if (cl_io_is_append(io))
386 RETURN(osc_io_iter_init(env, ios));
388 npages = io->u.ci_rw.rw_range.cir_count >> PAGE_SHIFT;
389 if (io->u.ci_rw.rw_range.cir_pos & ~PAGE_MASK)
392 oio->oi_lru_reserved = osc_lru_reserve(osc_cli(osc), npages);
394 RETURN(osc_io_iter_init(env, ios));
396 EXPORT_SYMBOL(osc_io_write_iter_init);
398 void osc_io_iter_fini(const struct lu_env *env,
399 const struct cl_io_slice *ios)
401 struct osc_io *oio = osc_env_io(env);
403 if (oio->oi_is_active) {
404 struct osc_object *osc = cl2osc(ios->cis_obj);
406 oio->oi_is_active = 0;
407 LASSERT(atomic_read(&osc->oo_nr_ios) > 0);
408 if (atomic_dec_and_test(&osc->oo_nr_ios))
409 wake_up_all(&osc->oo_io_waitq);
412 EXPORT_SYMBOL(osc_io_iter_fini);
414 void osc_io_write_iter_fini(const struct lu_env *env,
415 const struct cl_io_slice *ios)
417 struct osc_io *oio = osc_env_io(env);
418 struct osc_object *osc = cl2osc(ios->cis_obj);
420 if (oio->oi_lru_reserved > 0) {
421 osc_lru_unreserve(osc_cli(osc), oio->oi_lru_reserved);
422 oio->oi_lru_reserved = 0;
424 oio->oi_write_osclock = NULL;
426 osc_io_iter_fini(env, ios);
428 EXPORT_SYMBOL(osc_io_write_iter_fini);
430 int osc_io_fault_start(const struct lu_env *env, const struct cl_io_slice *ios)
433 struct cl_fault_io *fio;
437 fio = &io->u.ci_fault;
438 CDEBUG(D_INFO, "%lu %d %zu\n",
439 fio->ft_index, fio->ft_writable, fio->ft_nob);
441 * If mapping is writeable, adjust kms to cover this page,
442 * but do not extend kms beyond actual file size.
445 if (fio->ft_writable)
446 osc_page_touch_at(env, ios->cis_obj,
447 fio->ft_index, fio->ft_nob);
450 EXPORT_SYMBOL(osc_io_fault_start);
453 static int osc_async_upcall(void *a, int rc)
455 struct osc_async_cbargs *args = a;
458 complete(&args->opc_sync);
463 * Checks that there are no pages being written in the extent being truncated.
465 static int trunc_check_cb(const struct lu_env *env, struct cl_io *io,
466 struct osc_page *ops , void *cbdata)
468 struct cl_page *page = ops->ops_cl.cpl_page;
469 struct osc_async_page *oap;
470 __u64 start = *(__u64 *)cbdata;
473 if (oap->oap_cmd & OBD_BRW_WRITE &&
474 !list_empty(&oap->oap_pending_item))
475 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %llu/%s.\n",
476 start, current->comm);
478 if (PageLocked(page->cp_vmpage))
479 CDEBUG(D_CACHE, "page %p index %lu locked for %d.\n",
480 ops, osc_index(ops), oap->oap_cmd & OBD_BRW_RWMASK);
482 return CLP_GANG_OKAY;
485 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
486 struct osc_io *oio, __u64 size)
488 struct cl_object *clob;
492 clob = oio->oi_cl.cis_obj;
493 start = cl_index(clob, size);
494 partial = cl_offset(clob, start) < size;
497 * Complain if there are pages in the truncated region.
499 osc_page_gang_lookup(env, io, cl2osc(clob),
500 start + partial, CL_PAGE_EOF,
501 trunc_check_cb, (void *)&size);
504 static int osc_io_setattr_start(const struct lu_env *env,
505 const struct cl_io_slice *slice)
507 struct cl_io *io = slice->cis_io;
508 struct osc_io *oio = cl2osc_io(env, slice);
509 struct cl_object *obj = slice->cis_obj;
510 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
511 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
512 struct obdo *oa = &oio->oi_oa;
513 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
514 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
515 unsigned int ia_avalid = io->u.ci_setattr.sa_avalid;
516 enum op_xvalid ia_xvalid = io->u.ci_setattr.sa_xvalid;
520 /* truncate cache dirty pages first */
521 if (cl_io_is_trunc(io))
522 result = osc_cache_truncate_start(env, cl2osc(obj), size,
525 if (result == 0 && oio->oi_lockless == 0) {
526 cl_object_attr_lock(obj);
527 result = cl_object_attr_get(env, obj, attr);
529 struct ost_lvb *lvb = &io->u.ci_setattr.sa_attr;
530 unsigned int cl_valid = 0;
532 if (ia_avalid & ATTR_SIZE) {
533 attr->cat_size = size;
534 attr->cat_kms = size;
535 cl_valid = (CAT_SIZE | CAT_KMS);
537 if (ia_avalid & ATTR_MTIME_SET) {
538 attr->cat_mtime = lvb->lvb_mtime;
539 cl_valid |= CAT_MTIME;
541 if (ia_avalid & ATTR_ATIME_SET) {
542 attr->cat_atime = lvb->lvb_atime;
543 cl_valid |= CAT_ATIME;
545 if (ia_xvalid & OP_XVALID_CTIME_SET) {
546 attr->cat_ctime = lvb->lvb_ctime;
547 cl_valid |= CAT_CTIME;
549 result = cl_object_attr_update(env, obj, attr,
552 cl_object_attr_unlock(obj);
554 memset(oa, 0, sizeof(*oa));
556 oa->o_oi = loi->loi_oi;
557 obdo_set_parent_fid(oa, io->u.ci_setattr.sa_parent_fid);
558 oa->o_stripe_idx = io->u.ci_setattr.sa_stripe_index;
559 oa->o_layout = io->u.ci_setattr.sa_layout;
560 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP |
562 if (ia_avalid & ATTR_CTIME) {
563 oa->o_valid |= OBD_MD_FLCTIME;
564 oa->o_ctime = attr->cat_ctime;
566 if (ia_avalid & ATTR_ATIME) {
567 oa->o_valid |= OBD_MD_FLATIME;
568 oa->o_atime = attr->cat_atime;
570 if (ia_avalid & ATTR_MTIME) {
571 oa->o_valid |= OBD_MD_FLMTIME;
572 oa->o_mtime = attr->cat_mtime;
574 if (ia_avalid & ATTR_SIZE) {
576 oa->o_blocks = OBD_OBJECT_EOF;
577 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
579 if (oio->oi_lockless) {
580 oa->o_flags = OBD_FL_SRVLOCK;
581 oa->o_valid |= OBD_MD_FLFLAGS;
584 if (io->ci_layout_version > 0) {
585 /* verify layout version */
586 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
587 oa->o_layout_version = io->ci_layout_version;
590 LASSERT(oio->oi_lockless == 0);
593 if (ia_xvalid & OP_XVALID_FLAGS) {
594 oa->o_flags = io->u.ci_setattr.sa_attr_flags;
595 oa->o_valid |= OBD_MD_FLFLAGS;
598 init_completion(&cbargs->opc_sync);
600 if (ia_avalid & ATTR_SIZE)
601 result = osc_punch_send(osc_export(cl2osc(obj)),
602 oa, osc_async_upcall, cbargs);
604 result = osc_setattr_async(osc_export(cl2osc(obj)),
605 oa, osc_async_upcall,
606 cbargs, PTLRPCD_SET);
608 cbargs->opc_rpc_sent = result == 0;
614 void osc_io_setattr_end(const struct lu_env *env,
615 const struct cl_io_slice *slice)
617 struct cl_io *io = slice->cis_io;
618 struct osc_io *oio = cl2osc_io(env, slice);
619 struct cl_object *obj = slice->cis_obj;
620 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
621 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
622 struct obdo *oa = &oio->oi_oa;
623 unsigned int cl_valid = 0;
626 if (cbargs->opc_rpc_sent) {
627 wait_for_completion(&cbargs->opc_sync);
628 result = io->ci_result = cbargs->opc_rc;
632 if (oio->oi_lockless) {
633 /* lockless truncate */
634 struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
636 LASSERT(cl_io_is_trunc(io));
637 /* XXX: Need a lock. */
638 osd->od_stats.os_lockless_truncates++;
642 if (cl_io_is_trunc(io)) {
643 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
644 cl_object_attr_lock(obj);
645 if (oa->o_valid & OBD_MD_FLBLOCKS) {
646 attr->cat_blocks = oa->o_blocks;
647 cl_valid |= CAT_BLOCKS;
650 cl_object_attr_update(env, obj, attr, cl_valid);
651 cl_object_attr_unlock(obj);
652 osc_trunc_check(env, io, oio, size);
653 osc_cache_truncate_end(env, oio->oi_trunc);
654 oio->oi_trunc = NULL;
657 EXPORT_SYMBOL(osc_io_setattr_end);
659 struct osc_data_version_args {
660 struct osc_io *dva_oio;
664 osc_data_version_interpret(const struct lu_env *env, struct ptlrpc_request *req,
667 struct osc_data_version_args *dva = args;
668 struct osc_io *oio = dva->dva_oio;
669 const struct ost_body *body;
675 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
677 GOTO(out, rc = -EPROTO);
679 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, &oio->oi_oa,
683 oio->oi_cbarg.opc_rc = rc;
684 complete(&oio->oi_cbarg.opc_sync);
689 static int osc_io_data_version_start(const struct lu_env *env,
690 const struct cl_io_slice *slice)
692 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
693 struct osc_io *oio = cl2osc_io(env, slice);
694 struct obdo *oa = &oio->oi_oa;
695 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
696 struct osc_object *obj = cl2osc(slice->cis_obj);
697 struct lov_oinfo *loi = obj->oo_oinfo;
698 struct obd_export *exp = osc_export(obj);
699 struct ptlrpc_request *req;
700 struct ost_body *body;
701 struct osc_data_version_args *dva;
705 memset(oa, 0, sizeof(*oa));
706 oa->o_oi = loi->loi_oi;
707 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
709 if (dv->dv_flags & (LL_DV_RD_FLUSH | LL_DV_WR_FLUSH)) {
710 oa->o_valid |= OBD_MD_FLFLAGS;
711 oa->o_flags |= OBD_FL_SRVLOCK;
712 if (dv->dv_flags & LL_DV_WR_FLUSH)
713 oa->o_flags |= OBD_FL_FLUSH;
716 init_completion(&cbargs->opc_sync);
718 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
722 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
724 ptlrpc_request_free(req);
728 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
729 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
731 ptlrpc_request_set_replen(req);
732 req->rq_interpret_reply = osc_data_version_interpret;
733 CLASSERT(sizeof(*dva) <= sizeof(req->rq_async_args));
734 dva = ptlrpc_req_async_args(req);
737 ptlrpcd_add_req(req);
742 static void osc_io_data_version_end(const struct lu_env *env,
743 const struct cl_io_slice *slice)
745 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
746 struct osc_io *oio = cl2osc_io(env, slice);
747 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
750 wait_for_completion(&cbargs->opc_sync);
752 if (cbargs->opc_rc != 0) {
753 slice->cis_io->ci_result = cbargs->opc_rc;
755 slice->cis_io->ci_result = 0;
756 if (!(oio->oi_oa.o_valid &
757 (OBD_MD_LAYOUT_VERSION | OBD_MD_FLDATAVERSION)))
758 slice->cis_io->ci_result = -ENOTSUPP;
760 if (oio->oi_oa.o_valid & OBD_MD_LAYOUT_VERSION)
761 dv->dv_layout_version = oio->oi_oa.o_layout_version;
762 if (oio->oi_oa.o_valid & OBD_MD_FLDATAVERSION)
763 dv->dv_data_version = oio->oi_oa.o_data_version;
769 int osc_io_read_start(const struct lu_env *env,
770 const struct cl_io_slice *slice)
772 struct cl_object *obj = slice->cis_obj;
773 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
777 if (!slice->cis_io->ci_noatime) {
778 cl_object_attr_lock(obj);
779 attr->cat_atime = ktime_get_real_seconds();
780 rc = cl_object_attr_update(env, obj, attr, CAT_ATIME);
781 cl_object_attr_unlock(obj);
786 EXPORT_SYMBOL(osc_io_read_start);
788 int osc_io_write_start(const struct lu_env *env,
789 const struct cl_io_slice *slice)
791 struct cl_object *obj = slice->cis_obj;
792 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
796 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_SETTIME, 1);
797 cl_object_attr_lock(obj);
798 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
799 rc = cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
800 cl_object_attr_unlock(obj);
804 EXPORT_SYMBOL(osc_io_write_start);
806 int osc_fsync_ost(const struct lu_env *env, struct osc_object *obj,
807 struct cl_fsync_io *fio)
809 struct osc_io *oio = osc_env_io(env);
810 struct obdo *oa = &oio->oi_oa;
811 struct lov_oinfo *loi = obj->oo_oinfo;
812 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
816 memset(oa, 0, sizeof(*oa));
817 oa->o_oi = loi->loi_oi;
818 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
820 /* reload size abd blocks for start and end of sync range */
821 oa->o_size = fio->fi_start;
822 oa->o_blocks = fio->fi_end;
823 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
825 obdo_set_parent_fid(oa, fio->fi_fid);
827 init_completion(&cbargs->opc_sync);
829 rc = osc_sync_base(obj, oa, osc_async_upcall, cbargs, PTLRPCD_SET);
832 EXPORT_SYMBOL(osc_fsync_ost);
834 int osc_io_fsync_start(const struct lu_env *env,
835 const struct cl_io_slice *slice)
837 struct cl_io *io = slice->cis_io;
838 struct cl_fsync_io *fio = &io->u.ci_fsync;
839 struct cl_object *obj = slice->cis_obj;
840 struct osc_object *osc = cl2osc(obj);
841 pgoff_t start = cl_index(obj, fio->fi_start);
842 pgoff_t end = cl_index(obj, fio->fi_end);
846 if (fio->fi_end == OBD_OBJECT_EOF)
849 result = osc_cache_writeback_range(env, osc, start, end, 0,
850 fio->fi_mode == CL_FSYNC_DISCARD);
852 fio->fi_nr_written += result;
855 if (fio->fi_mode == CL_FSYNC_ALL) {
858 /* we have to wait for writeback to finish before we can
859 * send OST_SYNC RPC. This is bad because it causes extents
860 * to be written osc by osc. However, we usually start
861 * writeback before CL_FSYNC_ALL so this won't have any real
863 rc = osc_cache_wait_range(env, osc, start, end);
866 rc = osc_fsync_ost(env, osc, fio);
874 void osc_io_fsync_end(const struct lu_env *env,
875 const struct cl_io_slice *slice)
877 struct cl_fsync_io *fio = &slice->cis_io->u.ci_fsync;
878 struct cl_object *obj = slice->cis_obj;
879 pgoff_t start = cl_index(obj, fio->fi_start);
880 pgoff_t end = cl_index(obj, fio->fi_end);
883 if (fio->fi_mode == CL_FSYNC_LOCAL) {
884 result = osc_cache_wait_range(env, cl2osc(obj), start, end);
885 } else if (fio->fi_mode == CL_FSYNC_ALL) {
886 struct osc_io *oio = cl2osc_io(env, slice);
887 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
889 wait_for_completion(&cbargs->opc_sync);
891 result = cbargs->opc_rc;
893 slice->cis_io->ci_result = result;
895 EXPORT_SYMBOL(osc_io_fsync_end);
897 static int osc_io_ladvise_start(const struct lu_env *env,
898 const struct cl_io_slice *slice)
901 struct cl_io *io = slice->cis_io;
902 struct osc_io *oio = cl2osc_io(env, slice);
903 struct cl_object *obj = slice->cis_obj;
904 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
905 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
906 struct obdo *oa = &oio->oi_oa;
907 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
908 struct lu_ladvise *ladvise;
909 struct ladvise_hdr *ladvise_hdr;
914 /* TODO: add multiple ladvise support in CLIO */
915 buf_size = offsetof(typeof(*ladvise_hdr), lah_advise[num_advise]);
916 if (osc_env_info(env)->oti_ladvise_buf.lb_len < buf_size)
917 lu_buf_realloc(&osc_env_info(env)->oti_ladvise_buf, buf_size);
919 ladvise_hdr = osc_env_info(env)->oti_ladvise_buf.lb_buf;
920 if (ladvise_hdr == NULL)
923 memset(ladvise_hdr, 0, buf_size);
924 ladvise_hdr->lah_magic = LADVISE_MAGIC;
925 ladvise_hdr->lah_count = num_advise;
926 ladvise_hdr->lah_flags = lio->li_flags;
928 memset(oa, 0, sizeof(*oa));
929 oa->o_oi = loi->loi_oi;
930 oa->o_valid = OBD_MD_FLID;
931 obdo_set_parent_fid(oa, lio->li_fid);
933 ladvise = ladvise_hdr->lah_advise;
934 ladvise->lla_start = lio->li_start;
935 ladvise->lla_end = lio->li_end;
936 ladvise->lla_advice = lio->li_advice;
938 if (lio->li_flags & LF_ASYNC) {
939 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
940 ladvise_hdr, NULL, NULL, NULL);
942 init_completion(&cbargs->opc_sync);
943 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
944 ladvise_hdr, osc_async_upcall,
945 cbargs, PTLRPCD_SET);
946 cbargs->opc_rpc_sent = result == 0;
951 static void osc_io_ladvise_end(const struct lu_env *env,
952 const struct cl_io_slice *slice)
954 struct cl_io *io = slice->cis_io;
955 struct osc_io *oio = cl2osc_io(env, slice);
956 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
958 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
960 if ((!(lio->li_flags & LF_ASYNC)) && cbargs->opc_rpc_sent) {
961 wait_for_completion(&cbargs->opc_sync);
962 result = cbargs->opc_rc;
964 slice->cis_io->ci_result = result;
967 void osc_io_end(const struct lu_env *env, const struct cl_io_slice *slice)
969 struct osc_io *oio = cl2osc_io(env, slice);
971 if (oio->oi_active) {
972 osc_extent_release(env, oio->oi_active);
973 oio->oi_active = NULL;
976 EXPORT_SYMBOL(osc_io_end);
978 static const struct cl_io_operations osc_io_ops = {
981 .cio_iter_init = osc_io_iter_init,
982 .cio_iter_fini = osc_io_iter_fini,
983 .cio_start = osc_io_read_start,
984 .cio_fini = osc_io_fini
987 .cio_iter_init = osc_io_write_iter_init,
988 .cio_iter_fini = osc_io_write_iter_fini,
989 .cio_start = osc_io_write_start,
990 .cio_end = osc_io_end,
991 .cio_fini = osc_io_fini
994 .cio_iter_init = osc_io_iter_init,
995 .cio_iter_fini = osc_io_iter_fini,
996 .cio_start = osc_io_setattr_start,
997 .cio_end = osc_io_setattr_end
999 [CIT_DATA_VERSION] = {
1000 .cio_start = osc_io_data_version_start,
1001 .cio_end = osc_io_data_version_end,
1004 .cio_iter_init = osc_io_iter_init,
1005 .cio_iter_fini = osc_io_iter_fini,
1006 .cio_start = osc_io_fault_start,
1007 .cio_end = osc_io_end,
1008 .cio_fini = osc_io_fini
1011 .cio_start = osc_io_fsync_start,
1012 .cio_end = osc_io_fsync_end,
1013 .cio_fini = osc_io_fini
1016 .cio_start = osc_io_ladvise_start,
1017 .cio_end = osc_io_ladvise_end,
1018 .cio_fini = osc_io_fini
1021 .cio_fini = osc_io_fini
1024 .cio_read_ahead = osc_io_read_ahead,
1025 .cio_submit = osc_io_submit,
1026 .cio_commit_async = osc_io_commit_async
1029 /*****************************************************************************
1031 * Transfer operations.
1035 int osc_io_init(const struct lu_env *env,
1036 struct cl_object *obj, struct cl_io *io)
1038 struct osc_io *oio = osc_env_io(env);
1040 CL_IO_SLICE_CLEAN(oio, oi_cl);
1041 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);