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>
42 #include <linux/pagevec.h>
43 #include <linux/falloc.h>
45 #include "osc_internal.h"
51 /*****************************************************************************
57 static void osc_io_fini(const struct lu_env *env, const struct cl_io_slice *io)
61 void osc_read_ahead_release(const struct lu_env *env, struct cl_read_ahead *ra)
63 struct ldlm_lock *dlmlock = ra->cra_dlmlock;
64 struct osc_io *oio = ra->cra_oio;
65 struct lustre_handle lockh;
67 oio->oi_is_readahead = 0;
68 ldlm_lock2handle(dlmlock, &lockh);
69 ldlm_lock_decref(&lockh, LCK_PR);
70 LDLM_LOCK_PUT(dlmlock);
72 EXPORT_SYMBOL(osc_read_ahead_release);
74 static int osc_io_read_ahead(const struct lu_env *env,
75 const struct cl_io_slice *ios,
76 pgoff_t start, struct cl_read_ahead *ra)
78 struct osc_object *osc = cl2osc(ios->cis_obj);
79 struct osc_io *oio = cl2osc_io(env, ios);
80 struct ldlm_lock *dlmlock;
81 int result = -ENODATA;
85 oio->oi_is_readahead = true;
86 dlmlock = osc_dlmlock_at_pgoff(env, osc, start, 0);
87 if (dlmlock != NULL) {
88 LASSERT(dlmlock->l_ast_data == osc);
89 if (dlmlock->l_req_mode != LCK_PR) {
90 struct lustre_handle lockh;
91 ldlm_lock2handle(dlmlock, &lockh);
92 ldlm_lock_addref(&lockh, LCK_PR);
93 ldlm_lock_decref(&lockh, dlmlock->l_req_mode);
96 ra->cra_rpc_pages = osc_cli(osc)->cl_max_pages_per_rpc;
97 ra->cra_end_idx = cl_index(osc2cl(osc),
98 dlmlock->l_policy_data.l_extent.end);
99 ra->cra_release = osc_read_ahead_release;
100 ra->cra_dlmlock = dlmlock;
102 if (ra->cra_end_idx != CL_PAGE_EOF)
103 ra->cra_contention = true;
111 * An implementation of cl_io_operations::cio_io_submit() method for osc
112 * layer. Iterates over pages in the in-queue, prepares each for io by calling
113 * cl_page_prep() and then either submits them through osc_io_submit_page()
114 * or, if page is already submitted, changes osc flags through
115 * osc_set_async_flags().
117 int osc_io_submit(const struct lu_env *env, const struct cl_io_slice *ios,
118 enum cl_req_type crt, struct cl_2queue *queue)
120 struct cl_page *page;
122 struct client_obd *cli = NULL;
123 struct osc_object *osc = NULL; /* to keep gcc happy */
124 struct osc_page *opg;
128 struct cl_page_list *qin = &queue->c2_qin;
129 struct cl_page_list *qout = &queue->c2_qout;
130 unsigned int queued = 0;
133 unsigned int max_pages;
134 unsigned int ppc_bits; /* pages per chunk bits */
136 bool sync_queue = false;
138 LASSERT(qin->pl_nr > 0);
140 CDEBUG(D_CACHE|D_READA, "%d %d\n", qin->pl_nr, crt);
142 osc = cl2osc(ios->cis_obj);
144 max_pages = cli->cl_max_pages_per_rpc;
145 ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
148 brw_flags = osc_io_srvlock(cl2osc_io(env, ios)) ? OBD_BRW_SRVLOCK : 0;
149 brw_flags |= crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
150 if (crt == CRT_READ && ios->cis_io->ci_ndelay)
151 brw_flags |= OBD_BRW_NDELAY;
153 page = cl_page_list_first(qin);
154 if (page->cp_type == CPT_TRANSIENT)
155 brw_flags |= OBD_BRW_NOCACHE;
158 * NOTE: here @page is a top-level page. This is done to avoid
159 * creation of sub-page-list.
161 cl_page_list_for_each_safe(page, tmp, qin) {
162 struct osc_async_page *oap;
168 opg = osc_cl_page_osc(page, osc);
170 LASSERT(osc == oap->oap_obj);
172 if (!list_empty(&oap->oap_pending_item) ||
173 !list_empty(&oap->oap_rpc_item)) {
174 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
180 result = cl_page_prep(env, io, page, crt);
183 if (result != -EALREADY)
186 * Handle -EALREADY error: for read case, the page is
187 * already in UPTODATE state; for write, the page
194 spin_lock(&oap->oap_lock);
195 oap->oap_async_flags = ASYNC_URGENT|ASYNC_READY;
196 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
197 spin_unlock(&oap->oap_lock);
199 osc_page_submit(env, opg, crt, brw_flags);
200 list_add_tail(&oap->oap_pending_item, &list);
202 if (page->cp_sync_io != NULL)
203 cl_page_list_move(qout, qin, page);
205 cl_page_list_del(env, qin, page);
208 if (queued == max_pages) {
210 } else if (crt == CRT_WRITE) {
212 unsigned int next_chunks;
214 chunks = (queued + ppc - 1) >> ppc_bits;
215 /* chunk number if add another page */
216 next_chunks = (queued + ppc) >> ppc_bits;
218 /* next page will excceed write chunk limit */
219 if (chunks == osc_max_write_chunks(cli) &&
220 next_chunks > chunks)
225 result = osc_queue_sync_pages(env, io, osc, &list,
235 result = osc_queue_sync_pages(env, io, osc, &list, brw_flags);
237 /* Update c/mtime for sync write. LU-7310 */
238 if (crt == CRT_WRITE && qout->pl_nr > 0 && result == 0) {
239 struct cl_object *obj = ios->cis_obj;
240 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
242 cl_object_attr_lock(obj);
243 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
244 cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
245 cl_object_attr_unlock(obj);
248 CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
249 return qout->pl_nr > 0 ? 0 : result;
251 EXPORT_SYMBOL(osc_io_submit);
254 * This is called to update the attributes when modifying a specific page,
255 * both when making new pages and when doing updates to existing cached pages.
257 * Expand stripe KMS if necessary.
259 void osc_page_touch_at(const struct lu_env *env, struct cl_object *obj,
260 pgoff_t idx, size_t to)
262 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
263 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
269 /* offset within stripe */
270 kms = cl_offset(obj, idx) + to;
272 cl_object_attr_lock(obj);
273 CDEBUG(D_INODE, "stripe KMS %sincreasing %llu->%llu %llu\n",
274 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
275 loi->loi_lvb.lvb_size);
277 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
278 valid = CAT_MTIME | CAT_CTIME;
279 if (kms > loi->loi_kms) {
283 if (kms > loi->loi_lvb.lvb_size) {
284 attr->cat_size = kms;
287 cl_object_attr_update(env, obj, attr, valid);
288 cl_object_attr_unlock(obj);
293 int osc_io_commit_async(const struct lu_env *env,
294 const struct cl_io_slice *ios,
295 struct cl_page_list *qin, int from, int to,
298 struct cl_io *io = ios->cis_io;
299 struct osc_io *oio = cl2osc_io(env, ios);
300 struct osc_object *osc = cl2osc(ios->cis_obj);
301 struct cl_page *page;
302 struct cl_page *last_page;
303 struct osc_page *opg;
304 struct pagevec *pvec = &osc_env_info(env)->oti_pagevec;
308 LASSERT(qin->pl_nr > 0);
310 /* Handle partial page cases */
311 last_page = cl_page_list_last(qin);
312 if (oio->oi_lockless) {
313 page = cl_page_list_first(qin);
314 if (page == last_page) {
315 cl_page_clip(env, page, from, to);
318 cl_page_clip(env, page, from, PAGE_SIZE);
320 cl_page_clip(env, last_page, 0, to);
324 ll_pagevec_init(pvec, 0);
326 while (qin->pl_nr > 0) {
327 struct osc_async_page *oap;
329 page = cl_page_list_first(qin);
330 opg = osc_cl_page_osc(page, osc);
333 LASSERTF(osc == oap->oap_obj,
334 "obj mismatch: %p / %p\n", osc, oap->oap_obj);
336 if (!list_empty(&oap->oap_rpc_item)) {
337 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
343 /* The page may be already in dirty cache. */
344 if (list_empty(&oap->oap_pending_item)) {
345 result = osc_page_cache_add(env, opg, io, cb);
350 osc_page_touch_at(env, osc2cl(osc), osc_index(opg),
351 page == last_page ? to : PAGE_SIZE);
353 cl_page_list_del(env, qin, page);
355 /* if there are no more slots, do the callback & reinit */
356 if (pagevec_add(pvec, page->cp_vmpage) == 0) {
357 (*cb)(env, io, pvec);
358 pagevec_reinit(pvec);
362 /* Clean up any partially full pagevecs */
363 if (pagevec_count(pvec) != 0)
364 (*cb)(env, io, pvec);
366 /* Can't access these pages any more. Page can be in transfer and
367 * complete at any time. */
369 /* for sync write, kernel will wait for this page to be flushed before
370 * osc_io_end() is called, so release it earlier.
371 * for mkwrite(), it's known there is no further pages. */
372 if (cl_io_is_sync_write(io) && oio->oi_active != NULL) {
373 osc_extent_release(env, oio->oi_active);
374 oio->oi_active = NULL;
377 CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, result);
380 EXPORT_SYMBOL(osc_io_commit_async);
382 void osc_io_extent_release(const struct lu_env *env,
383 const struct cl_io_slice *ios)
385 struct osc_io *oio = cl2osc_io(env, ios);
387 if (oio->oi_active != NULL) {
388 osc_extent_release(env, oio->oi_active);
389 oio->oi_active = NULL;
392 EXPORT_SYMBOL(osc_io_extent_release);
394 static bool osc_import_not_healthy(struct obd_import *imp)
396 return imp->imp_invalid || imp->imp_deactive ||
397 !(imp->imp_state == LUSTRE_IMP_FULL ||
398 imp->imp_state == LUSTRE_IMP_IDLE);
401 int osc_io_iter_init(const struct lu_env *env, const struct cl_io_slice *ios)
403 struct osc_object *osc = cl2osc(ios->cis_obj);
404 struct obd_import *imp = osc_cli(osc)->cl_import;
405 struct osc_io *oio = osc_env_io(env);
410 spin_lock(&imp->imp_lock);
412 * check whether this OSC device is available for non-delay read,
413 * fast switching mirror if we haven't tried all mirrors.
415 if (ios->cis_io->ci_type == CIT_READ && ios->cis_io->ci_ndelay &&
416 !ios->cis_io->ci_tried_all_mirrors && osc_import_not_healthy(imp)) {
418 } else if (likely(!imp->imp_invalid)) {
419 atomic_inc(&osc->oo_nr_ios);
420 oio->oi_is_active = 1;
423 spin_unlock(&imp->imp_lock);
425 if (capable(CAP_SYS_RESOURCE))
426 oio->oi_cap_sys_resource = 1;
430 EXPORT_SYMBOL(osc_io_iter_init);
432 void osc_io_iter_fini(const struct lu_env *env,
433 const struct cl_io_slice *ios)
435 struct osc_io *oio = osc_env_io(env);
437 if (oio->oi_is_active) {
438 struct osc_object *osc = cl2osc(ios->cis_obj);
440 oio->oi_is_active = 0;
441 LASSERT(atomic_read(&osc->oo_nr_ios) > 0);
442 if (atomic_dec_and_test(&osc->oo_nr_ios))
443 wake_up(&osc->oo_io_waitq);
446 EXPORT_SYMBOL(osc_io_iter_fini);
448 void osc_io_rw_iter_fini(const struct lu_env *env,
449 const struct cl_io_slice *ios)
451 struct osc_io *oio = osc_env_io(env);
452 struct osc_object *osc = cl2osc(ios->cis_obj);
454 if (oio->oi_lru_reserved > 0) {
455 osc_lru_unreserve(osc_cli(osc), oio->oi_lru_reserved);
456 oio->oi_lru_reserved = 0;
458 oio->oi_write_osclock = NULL;
460 osc_io_iter_fini(env, ios);
462 EXPORT_SYMBOL(osc_io_rw_iter_fini);
464 int osc_io_fault_start(const struct lu_env *env, const struct cl_io_slice *ios)
467 struct cl_fault_io *fio;
471 fio = &io->u.ci_fault;
472 CDEBUG(D_INFO, "%lu %d %zu\n",
473 fio->ft_index, fio->ft_writable, fio->ft_nob);
475 * If mapping is writeable, adjust kms to cover this page,
476 * but do not extend kms beyond actual file size.
479 if (fio->ft_writable)
480 osc_page_touch_at(env, ios->cis_obj,
481 fio->ft_index, fio->ft_nob);
484 EXPORT_SYMBOL(osc_io_fault_start);
487 static int osc_async_upcall(void *a, int rc)
489 struct osc_async_cbargs *args = a;
492 complete(&args->opc_sync);
497 * Checks that there are no pages being written in the extent being truncated.
499 static bool trunc_check_cb(const struct lu_env *env, struct cl_io *io,
500 struct osc_page *ops , void *cbdata)
502 struct cl_page *page = ops->ops_cl.cpl_page;
503 struct osc_async_page *oap;
504 __u64 start = *(__u64 *)cbdata;
507 if (oap->oap_cmd & OBD_BRW_WRITE &&
508 !list_empty(&oap->oap_pending_item))
509 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %llu/%s.\n",
510 start, current->comm);
512 if (PageLocked(page->cp_vmpage))
513 CDEBUG(D_CACHE, "page %p index %lu locked for %d.\n",
514 ops, osc_index(ops), oap->oap_cmd & OBD_BRW_RWMASK);
519 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
520 struct osc_io *oio, __u64 size)
522 struct cl_object *clob;
526 clob = oio->oi_cl.cis_obj;
527 start = cl_index(clob, size);
528 partial = cl_offset(clob, start) < size;
531 * Complain if there are pages in the truncated region.
533 osc_page_gang_lookup(env, io, cl2osc(clob),
534 start + partial, CL_PAGE_EOF,
535 trunc_check_cb, (void *)&size);
539 * Flush affected pages prior punch.
540 * We shouldn't discard them locally first because that could be data loss
541 * if server doesn't support fallocate punch, we also need these data to be
542 * flushed first to prevent re-ordering with the punch
544 static int osc_punch_start(const struct lu_env *env, struct cl_io *io,
545 struct cl_object *obj)
547 struct osc_object *osc = cl2osc(obj);
548 pgoff_t pg_start = cl_index(obj, io->u.ci_setattr.sa_falloc_offset);
549 pgoff_t pg_end = cl_index(obj, io->u.ci_setattr.sa_falloc_end - 1);
553 rc = osc_cache_writeback_range(env, osc, pg_start, pg_end, 1, 0);
557 osc_page_gang_lookup(env, io, osc, pg_start, pg_end, osc_discard_cb,
562 static int osc_io_setattr_start(const struct lu_env *env,
563 const struct cl_io_slice *slice)
565 struct cl_io *io = slice->cis_io;
566 struct osc_io *oio = cl2osc_io(env, slice);
567 struct cl_object *obj = slice->cis_obj;
568 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
569 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
570 struct obdo *oa = &oio->oi_oa;
571 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
572 unsigned int ia_avalid = io->u.ci_setattr.sa_avalid;
573 enum op_xvalid ia_xvalid = io->u.ci_setattr.sa_xvalid;
575 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
576 bool io_is_falloc = cl_io_is_fallocate(io);
579 /* truncate cache dirty pages first */
580 if (cl_io_is_trunc(io))
581 result = osc_cache_truncate_start(env, cl2osc(obj), size,
583 /* flush local pages prior punching them on server */
585 io->u.ci_setattr.sa_falloc_mode & FALLOC_FL_PUNCH_HOLE)
586 result = osc_punch_start(env, io, obj);
588 if (result == 0 && oio->oi_lockless == 0) {
589 cl_object_attr_lock(obj);
590 result = cl_object_attr_get(env, obj, attr);
592 struct ost_lvb *lvb = &io->u.ci_setattr.sa_attr;
593 unsigned int cl_valid = 0;
595 if (ia_avalid & ATTR_SIZE) {
596 attr->cat_size = size;
597 attr->cat_kms = size;
598 cl_valid = (CAT_SIZE | CAT_KMS);
600 if (ia_avalid & ATTR_MTIME_SET) {
601 attr->cat_mtime = lvb->lvb_mtime;
602 cl_valid |= CAT_MTIME;
604 if (ia_avalid & ATTR_ATIME_SET) {
605 attr->cat_atime = lvb->lvb_atime;
606 cl_valid |= CAT_ATIME;
608 if (ia_xvalid & OP_XVALID_CTIME_SET) {
609 attr->cat_ctime = lvb->lvb_ctime;
610 cl_valid |= CAT_CTIME;
612 result = cl_object_attr_update(env, obj, attr,
615 cl_object_attr_unlock(obj);
617 memset(oa, 0, sizeof(*oa));
619 oa->o_oi = loi->loi_oi;
620 obdo_set_parent_fid(oa, io->u.ci_setattr.sa_parent_fid);
621 oa->o_stripe_idx = io->u.ci_setattr.sa_stripe_index;
622 oa->o_layout = io->u.ci_setattr.sa_layout;
623 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP |
625 if (ia_avalid & ATTR_CTIME) {
626 oa->o_valid |= OBD_MD_FLCTIME;
627 oa->o_ctime = attr->cat_ctime;
629 if (ia_avalid & ATTR_ATIME) {
630 oa->o_valid |= OBD_MD_FLATIME;
631 oa->o_atime = attr->cat_atime;
633 if (ia_avalid & ATTR_MTIME) {
634 oa->o_valid |= OBD_MD_FLMTIME;
635 oa->o_mtime = attr->cat_mtime;
638 if (ia_avalid & ATTR_SIZE || io_is_falloc) {
639 if (oio->oi_lockless) {
640 oa->o_flags = OBD_FL_SRVLOCK;
641 oa->o_valid |= OBD_MD_FLFLAGS;
644 if (io->ci_layout_version > 0) {
645 /* verify layout version */
646 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
647 oa->o_layout_version = io->ci_layout_version;
650 LASSERT(oio->oi_lockless == 0);
653 if (ia_xvalid & OP_XVALID_FLAGS) {
654 oa->o_flags = io->u.ci_setattr.sa_attr_flags;
655 oa->o_valid |= OBD_MD_FLFLAGS;
658 init_completion(&cbargs->opc_sync);
661 int falloc_mode = io->u.ci_setattr.sa_falloc_mode;
663 oa->o_size = io->u.ci_setattr.sa_falloc_offset;
664 oa->o_blocks = io->u.ci_setattr.sa_falloc_end;
665 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
666 result = osc_fallocate_base(osc_export(cl2osc(obj)),
667 oa, osc_async_upcall,
668 cbargs, falloc_mode);
669 } else if (ia_avalid & ATTR_SIZE) {
671 oa->o_blocks = OBD_OBJECT_EOF;
672 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
673 result = osc_punch_send(osc_export(cl2osc(obj)),
674 oa, osc_async_upcall, cbargs);
676 result = osc_setattr_async(osc_export(cl2osc(obj)),
677 oa, osc_async_upcall,
678 cbargs, PTLRPCD_SET);
680 cbargs->opc_rpc_sent = result == 0;
686 void osc_io_setattr_end(const struct lu_env *env,
687 const struct cl_io_slice *slice)
689 struct cl_io *io = slice->cis_io;
690 struct osc_io *oio = cl2osc_io(env, slice);
691 struct cl_object *obj = slice->cis_obj;
692 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
693 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
694 struct obdo *oa = &oio->oi_oa;
695 unsigned int cl_valid = 0;
698 if (cbargs->opc_rpc_sent) {
699 wait_for_completion(&cbargs->opc_sync);
700 result = io->ci_result = cbargs->opc_rc;
704 if (oio->oi_lockless) {
705 /* lockless truncate */
706 struct osc_device *osc = lu2osc_dev(obj->co_lu.lo_dev);
708 LASSERT(cl_io_is_trunc(io) || cl_io_is_fallocate(io));
709 /* XXX: Need a lock. */
710 osc->od_stats.os_lockless_truncates++;
714 if (cl_io_is_trunc(io)) {
715 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
718 cl_object_attr_lock(obj);
719 if (oa->o_valid & OBD_MD_FLBLOCKS) {
720 attr->cat_blocks = oa->o_blocks;
721 cl_valid |= CAT_BLOCKS;
724 cl_object_attr_update(env, obj, attr, cl_valid);
725 cl_object_attr_unlock(obj);
727 osc_trunc_check(env, io, oio, size);
728 osc_cache_truncate_end(env, oio->oi_trunc);
729 oio->oi_trunc = NULL;
732 if (cl_io_is_fallocate(io)) {
734 cl_object_attr_lock(obj);
736 if (oa->o_valid & OBD_MD_FLBLOCKS) {
737 attr->cat_blocks = oa->o_blocks;
738 cl_valid |= CAT_BLOCKS;
741 cl_object_attr_update(env, obj, attr, cl_valid);
742 cl_object_attr_unlock(obj);
746 EXPORT_SYMBOL(osc_io_setattr_end);
748 struct osc_data_version_args {
749 struct osc_io *dva_oio;
753 osc_data_version_interpret(const struct lu_env *env, struct ptlrpc_request *req,
756 struct osc_data_version_args *dva = args;
757 struct osc_io *oio = dva->dva_oio;
758 const struct ost_body *body;
764 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
766 GOTO(out, rc = -EPROTO);
768 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, &oio->oi_oa,
772 oio->oi_cbarg.opc_rc = rc;
773 complete(&oio->oi_cbarg.opc_sync);
778 static int osc_io_data_version_start(const struct lu_env *env,
779 const struct cl_io_slice *slice)
781 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
782 struct osc_io *oio = cl2osc_io(env, slice);
783 struct obdo *oa = &oio->oi_oa;
784 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
785 struct osc_object *obj = cl2osc(slice->cis_obj);
786 struct lov_oinfo *loi = obj->oo_oinfo;
787 struct obd_export *exp = osc_export(obj);
788 struct ptlrpc_request *req;
789 struct ost_body *body;
790 struct osc_data_version_args *dva;
794 memset(oa, 0, sizeof(*oa));
795 oa->o_oi = loi->loi_oi;
796 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
798 if (dv->dv_flags & (LL_DV_RD_FLUSH | LL_DV_WR_FLUSH)) {
799 oa->o_valid |= OBD_MD_FLFLAGS;
800 oa->o_flags |= OBD_FL_SRVLOCK;
801 if (dv->dv_flags & LL_DV_WR_FLUSH)
802 oa->o_flags |= OBD_FL_FLUSH;
805 init_completion(&cbargs->opc_sync);
807 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
811 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
813 ptlrpc_request_free(req);
817 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
818 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
820 ptlrpc_request_set_replen(req);
821 req->rq_interpret_reply = osc_data_version_interpret;
822 dva = ptlrpc_req_async_args(dva, req);
825 ptlrpcd_add_req(req);
830 static void osc_io_data_version_end(const struct lu_env *env,
831 const struct cl_io_slice *slice)
833 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
834 struct osc_io *oio = cl2osc_io(env, slice);
835 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
838 wait_for_completion(&cbargs->opc_sync);
840 if (cbargs->opc_rc != 0) {
841 slice->cis_io->ci_result = cbargs->opc_rc;
843 slice->cis_io->ci_result = 0;
844 if (!(oio->oi_oa.o_valid &
845 (OBD_MD_LAYOUT_VERSION | OBD_MD_FLDATAVERSION)))
846 slice->cis_io->ci_result = -ENOTSUPP;
848 if (oio->oi_oa.o_valid & OBD_MD_LAYOUT_VERSION)
849 dv->dv_layout_version = oio->oi_oa.o_layout_version;
850 if (oio->oi_oa.o_valid & OBD_MD_FLDATAVERSION)
851 dv->dv_data_version = oio->oi_oa.o_data_version;
857 int osc_io_read_start(const struct lu_env *env,
858 const struct cl_io_slice *slice)
860 struct cl_object *obj = slice->cis_obj;
861 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
865 if (!slice->cis_io->ci_noatime) {
866 cl_object_attr_lock(obj);
867 attr->cat_atime = ktime_get_real_seconds();
868 rc = cl_object_attr_update(env, obj, attr, CAT_ATIME);
869 cl_object_attr_unlock(obj);
874 EXPORT_SYMBOL(osc_io_read_start);
876 int osc_io_write_start(const struct lu_env *env,
877 const struct cl_io_slice *slice)
879 struct cl_object *obj = slice->cis_obj;
880 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
884 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_SETTIME, 1);
885 cl_object_attr_lock(obj);
886 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
887 rc = cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
888 cl_object_attr_unlock(obj);
892 EXPORT_SYMBOL(osc_io_write_start);
894 int osc_fsync_ost(const struct lu_env *env, struct osc_object *obj,
895 struct cl_fsync_io *fio)
897 struct osc_io *oio = osc_env_io(env);
898 struct obdo *oa = &oio->oi_oa;
899 struct lov_oinfo *loi = obj->oo_oinfo;
900 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
904 memset(oa, 0, sizeof(*oa));
905 oa->o_oi = loi->loi_oi;
906 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
908 /* reload size abd blocks for start and end of sync range */
909 oa->o_size = fio->fi_start;
910 oa->o_blocks = fio->fi_end;
911 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
913 obdo_set_parent_fid(oa, fio->fi_fid);
915 init_completion(&cbargs->opc_sync);
917 rc = osc_sync_base(obj, oa, osc_async_upcall, cbargs, PTLRPCD_SET);
920 EXPORT_SYMBOL(osc_fsync_ost);
922 int osc_io_fsync_start(const struct lu_env *env,
923 const struct cl_io_slice *slice)
925 struct cl_io *io = slice->cis_io;
926 struct cl_fsync_io *fio = &io->u.ci_fsync;
927 struct cl_object *obj = slice->cis_obj;
928 struct osc_object *osc = cl2osc(obj);
929 pgoff_t start = cl_index(obj, fio->fi_start);
930 pgoff_t end = cl_index(obj, fio->fi_end);
934 if (fio->fi_end == OBD_OBJECT_EOF)
937 result = osc_cache_writeback_range(env, osc, start, end, 0,
938 fio->fi_mode == CL_FSYNC_DISCARD);
940 fio->fi_nr_written += result;
943 if (fio->fi_mode == CL_FSYNC_ALL) {
946 /* we have to wait for writeback to finish before we can
947 * send OST_SYNC RPC. This is bad because it causes extents
948 * to be written osc by osc. However, we usually start
949 * writeback before CL_FSYNC_ALL so this won't have any real
951 rc = osc_cache_wait_range(env, osc, start, end);
954 rc = osc_fsync_ost(env, osc, fio);
962 void osc_io_fsync_end(const struct lu_env *env,
963 const struct cl_io_slice *slice)
965 struct cl_fsync_io *fio = &slice->cis_io->u.ci_fsync;
966 struct cl_object *obj = slice->cis_obj;
967 pgoff_t start = cl_index(obj, fio->fi_start);
968 pgoff_t end = cl_index(obj, fio->fi_end);
971 if (fio->fi_mode == CL_FSYNC_LOCAL) {
972 result = osc_cache_wait_range(env, cl2osc(obj), start, end);
973 } else if (fio->fi_mode == CL_FSYNC_ALL) {
974 struct osc_io *oio = cl2osc_io(env, slice);
975 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
977 wait_for_completion(&cbargs->opc_sync);
979 result = cbargs->opc_rc;
981 slice->cis_io->ci_result = result;
983 EXPORT_SYMBOL(osc_io_fsync_end);
985 static int osc_io_ladvise_start(const struct lu_env *env,
986 const struct cl_io_slice *slice)
989 struct cl_io *io = slice->cis_io;
990 struct osc_io *oio = cl2osc_io(env, slice);
991 struct cl_object *obj = slice->cis_obj;
992 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
993 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
994 struct obdo *oa = &oio->oi_oa;
995 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
996 struct lu_ladvise *ladvise;
997 struct ladvise_hdr *ladvise_hdr;
1002 /* TODO: add multiple ladvise support in CLIO */
1003 buf_size = offsetof(typeof(*ladvise_hdr), lah_advise[num_advise]);
1004 if (osc_env_info(env)->oti_ladvise_buf.lb_len < buf_size)
1005 lu_buf_realloc(&osc_env_info(env)->oti_ladvise_buf, buf_size);
1007 ladvise_hdr = osc_env_info(env)->oti_ladvise_buf.lb_buf;
1008 if (ladvise_hdr == NULL)
1011 memset(ladvise_hdr, 0, buf_size);
1012 ladvise_hdr->lah_magic = LADVISE_MAGIC;
1013 ladvise_hdr->lah_count = num_advise;
1014 ladvise_hdr->lah_flags = lio->li_flags;
1016 memset(oa, 0, sizeof(*oa));
1017 oa->o_oi = loi->loi_oi;
1018 oa->o_valid = OBD_MD_FLID;
1019 obdo_set_parent_fid(oa, lio->li_fid);
1021 ladvise = ladvise_hdr->lah_advise;
1022 ladvise->lla_start = lio->li_start;
1023 ladvise->lla_end = lio->li_end;
1024 ladvise->lla_advice = lio->li_advice;
1026 if (lio->li_flags & LF_ASYNC) {
1027 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1028 ladvise_hdr, NULL, NULL, NULL);
1030 init_completion(&cbargs->opc_sync);
1031 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1032 ladvise_hdr, osc_async_upcall,
1033 cbargs, PTLRPCD_SET);
1034 cbargs->opc_rpc_sent = result == 0;
1039 static void osc_io_ladvise_end(const struct lu_env *env,
1040 const struct cl_io_slice *slice)
1042 struct cl_io *io = slice->cis_io;
1043 struct osc_io *oio = cl2osc_io(env, slice);
1044 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1046 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
1048 if ((!(lio->li_flags & LF_ASYNC)) && cbargs->opc_rpc_sent) {
1049 wait_for_completion(&cbargs->opc_sync);
1050 result = cbargs->opc_rc;
1052 slice->cis_io->ci_result = result;
1055 void osc_io_end(const struct lu_env *env, const struct cl_io_slice *slice)
1057 struct osc_io *oio = cl2osc_io(env, slice);
1059 if (oio->oi_active) {
1060 osc_extent_release(env, oio->oi_active);
1061 oio->oi_active = NULL;
1064 EXPORT_SYMBOL(osc_io_end);
1066 struct osc_lseek_args {
1067 struct osc_io *lsa_oio;
1070 static int osc_lseek_interpret(const struct lu_env *env,
1071 struct ptlrpc_request *req,
1074 struct ost_body *reply;
1075 struct osc_lseek_args *lsa = arg;
1076 struct osc_io *oio = lsa->lsa_oio;
1077 struct cl_io *io = oio->oi_cl.cis_io;
1078 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1085 reply = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1087 GOTO(out, rc = -EPROTO);
1089 lsio->ls_result = reply->oa.o_size;
1091 osc_async_upcall(&oio->oi_cbarg, rc);
1095 int osc_io_lseek_start(const struct lu_env *env,
1096 const struct cl_io_slice *slice)
1098 struct cl_io *io = slice->cis_io;
1099 struct osc_io *oio = cl2osc_io(env, slice);
1100 struct cl_object *obj = slice->cis_obj;
1101 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1102 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1103 struct obdo *oa = &oio->oi_oa;
1104 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1105 struct obd_export *exp = osc_export(cl2osc(obj));
1106 struct ptlrpc_request *req;
1107 struct ost_body *body;
1108 struct osc_lseek_args *lsa;
1113 /* No negative values at this point */
1114 LASSERT(lsio->ls_start >= 0);
1115 LASSERT(lsio->ls_whence == SEEK_HOLE || lsio->ls_whence == SEEK_DATA);
1117 /* with IO lock taken we have object size in LVB and can check
1118 * boundaries prior sending LSEEK RPC
1120 if (lsio->ls_start >= loi->loi_lvb.lvb_size) {
1121 /* consider area beyond end of object as hole */
1122 if (lsio->ls_whence == SEEK_HOLE)
1123 lsio->ls_result = lsio->ls_start;
1125 lsio->ls_result = -ENXIO;
1129 /* if LSEEK RPC is not supported by server, consider whole stripe
1130 * object is data with hole after end of object
1132 if (!exp_connect_lseek(exp)) {
1133 if (lsio->ls_whence == SEEK_HOLE)
1134 lsio->ls_result = loi->loi_lvb.lvb_size;
1136 lsio->ls_result = lsio->ls_start;
1140 memset(oa, 0, sizeof(*oa));
1141 oa->o_oi = loi->loi_oi;
1142 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1143 oa->o_size = lsio->ls_start;
1144 oa->o_mode = lsio->ls_whence;
1145 if (oio->oi_lockless) {
1146 oa->o_flags = OBD_FL_SRVLOCK;
1147 oa->o_valid |= OBD_MD_FLFLAGS;
1150 init_completion(&cbargs->opc_sync);
1151 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SEEK);
1155 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SEEK);
1157 ptlrpc_request_free(req);
1161 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1162 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1163 ptlrpc_request_set_replen(req);
1164 req->rq_interpret_reply = osc_lseek_interpret;
1165 lsa = ptlrpc_req_async_args(lsa, req);
1168 ptlrpcd_add_req(req);
1169 cbargs->opc_rpc_sent = 1;
1173 EXPORT_SYMBOL(osc_io_lseek_start);
1175 void osc_io_lseek_end(const struct lu_env *env,
1176 const struct cl_io_slice *slice)
1178 struct osc_io *oio = cl2osc_io(env, slice);
1179 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1182 if (cbargs->opc_rpc_sent) {
1183 wait_for_completion(&cbargs->opc_sync);
1184 rc = cbargs->opc_rc;
1186 slice->cis_io->ci_result = rc;
1188 EXPORT_SYMBOL(osc_io_lseek_end);
1190 int osc_io_lru_reserve(const struct lu_env *env,
1191 const struct cl_io_slice *ios,
1192 loff_t pos, size_t bytes)
1194 struct osc_object *osc = cl2osc(ios->cis_obj);
1195 struct osc_io *oio = osc_env_io(env);
1196 unsigned long npages = 0;
1201 page_offset = pos & ~PAGE_MASK;
1204 if (bytes > PAGE_SIZE - page_offset)
1205 bytes -= (PAGE_SIZE - page_offset);
1209 npages += (bytes + PAGE_SIZE - 1) >> PAGE_SHIFT;
1210 oio->oi_lru_reserved = osc_lru_reserve(osc_cli(osc), npages);
1214 EXPORT_SYMBOL(osc_io_lru_reserve);
1216 static const struct cl_io_operations osc_io_ops = {
1219 .cio_iter_init = osc_io_iter_init,
1220 .cio_iter_fini = osc_io_rw_iter_fini,
1221 .cio_start = osc_io_read_start,
1222 .cio_fini = osc_io_fini
1225 .cio_iter_init = osc_io_iter_init,
1226 .cio_iter_fini = osc_io_rw_iter_fini,
1227 .cio_start = osc_io_write_start,
1228 .cio_end = osc_io_end,
1229 .cio_fini = osc_io_fini
1232 .cio_iter_init = osc_io_iter_init,
1233 .cio_iter_fini = osc_io_iter_fini,
1234 .cio_start = osc_io_setattr_start,
1235 .cio_end = osc_io_setattr_end
1237 [CIT_DATA_VERSION] = {
1238 .cio_start = osc_io_data_version_start,
1239 .cio_end = osc_io_data_version_end,
1242 .cio_iter_init = osc_io_iter_init,
1243 .cio_iter_fini = osc_io_iter_fini,
1244 .cio_start = osc_io_fault_start,
1245 .cio_end = osc_io_end,
1246 .cio_fini = osc_io_fini
1249 .cio_start = osc_io_fsync_start,
1250 .cio_end = osc_io_fsync_end,
1251 .cio_fini = osc_io_fini
1254 .cio_start = osc_io_ladvise_start,
1255 .cio_end = osc_io_ladvise_end,
1256 .cio_fini = osc_io_fini
1259 .cio_start = osc_io_lseek_start,
1260 .cio_end = osc_io_lseek_end,
1261 .cio_fini = osc_io_fini
1264 .cio_fini = osc_io_fini
1267 .cio_read_ahead = osc_io_read_ahead,
1268 .cio_lru_reserve = osc_io_lru_reserve,
1269 .cio_submit = osc_io_submit,
1270 .cio_commit_async = osc_io_commit_async,
1271 .cio_extent_release = osc_io_extent_release
1274 /*****************************************************************************
1276 * Transfer operations.
1280 int osc_io_init(const struct lu_env *env,
1281 struct cl_object *obj, struct cl_io *io)
1283 struct osc_io *oio = osc_env_io(env);
1285 CL_IO_SLICE_CLEAN(oio, oi_cl);
1286 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);