4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * Implementation of cl_io for OSC layer.
33 * Author: Nikita Danilov <nikita.danilov@sun.com>
34 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
37 #define DEBUG_SUBSYSTEM S_OSC
39 #include <lustre_obdo.h>
40 #include <lustre_osc.h>
41 #include <linux/pagevec.h>
42 #include <linux/falloc.h>
44 #include "osc_internal.h"
45 #include <lnet/lnet_rdma.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;
156 if (lnet_is_rdma_only_page(page->cp_vmpage))
157 brw_flags |= OBD_BRW_RDMA_ONLY;
160 * NOTE: here @page is a top-level page. This is done to avoid
161 * creation of sub-page-list.
163 cl_page_list_for_each_safe(page, tmp, qin) {
164 struct osc_async_page *oap;
170 opg = osc_cl_page_osc(page, osc);
172 LASSERT(osc == oap->oap_obj);
174 if (!list_empty(&oap->oap_pending_item) ||
175 !list_empty(&oap->oap_rpc_item)) {
176 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
182 result = cl_page_prep(env, io, page, crt);
185 if (result != -EALREADY)
188 * Handle -EALREADY error: for read case, the page is
189 * already in UPTODATE state; for write, the page
196 if (page->cp_type != CPT_TRANSIENT) {
197 oap->oap_async_flags = ASYNC_URGENT|ASYNC_READY|ASYNC_COUNT_STABLE;
200 osc_page_submit(env, opg, crt, brw_flags);
201 list_add_tail(&oap->oap_pending_item, &list);
203 if (page->cp_sync_io != NULL)
204 cl_page_list_move(qout, qin, page);
206 cl_page_list_del(env, qin, page);
209 if (queued == max_pages) {
211 } else if (crt == CRT_WRITE) {
213 unsigned int next_chunks;
215 chunks = (queued + ppc - 1) >> ppc_bits;
216 /* chunk number if add another page */
217 next_chunks = (queued + ppc) >> ppc_bits;
219 /* next page will excceed write chunk limit */
220 if (chunks == osc_max_write_chunks(cli) &&
221 next_chunks > chunks)
226 result = osc_queue_sync_pages(env, io, osc, &list,
236 result = osc_queue_sync_pages(env, io, osc, &list, brw_flags);
238 /* Update c/mtime for sync write. LU-7310 */
239 if (crt == CRT_WRITE && qout->pl_nr > 0 && result == 0) {
240 struct cl_object *obj = ios->cis_obj;
241 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
243 cl_object_attr_lock(obj);
244 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
245 cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
246 cl_object_attr_unlock(obj);
249 CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
250 return qout->pl_nr > 0 ? 0 : result;
252 EXPORT_SYMBOL(osc_io_submit);
255 * This is called to update the attributes when modifying a specific page,
256 * both when making new pages and when doing updates to existing cached pages.
258 * Expand stripe KMS if necessary.
260 void osc_page_touch_at(const struct lu_env *env, struct cl_object *obj,
261 pgoff_t idx, size_t to)
263 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
264 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
270 /* offset within stripe */
271 kms = cl_offset(obj, idx) + to;
273 cl_object_attr_lock(obj);
274 CDEBUG(D_INODE, "stripe KMS %sincreasing %llu->%llu %llu\n",
275 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
276 loi->loi_lvb.lvb_size);
278 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
279 valid = CAT_MTIME | CAT_CTIME;
280 if (kms > loi->loi_kms) {
284 if (kms > loi->loi_lvb.lvb_size) {
285 attr->cat_size = kms;
288 cl_object_attr_update(env, obj, attr, valid);
289 cl_object_attr_unlock(obj);
294 int osc_io_commit_async(const struct lu_env *env,
295 const struct cl_io_slice *ios,
296 struct cl_page_list *qin, int from, int to,
299 struct cl_io *io = ios->cis_io;
300 struct osc_io *oio = cl2osc_io(env, ios);
301 struct osc_object *osc = cl2osc(ios->cis_obj);
302 struct cl_page *page;
303 struct cl_page *last_page;
304 struct osc_page *opg;
305 struct pagevec *pvec = &osc_env_info(env)->oti_pagevec;
309 LASSERT(qin->pl_nr > 0);
311 /* Handle partial page cases */
312 last_page = cl_page_list_last(qin);
313 if (oio->oi_lockless) {
314 page = cl_page_list_first(qin);
315 if (page == last_page) {
316 cl_page_clip(env, page, from, to);
319 cl_page_clip(env, page, from, PAGE_SIZE);
321 cl_page_clip(env, last_page, 0, to);
325 ll_pagevec_init(pvec, 0);
327 while (qin->pl_nr > 0) {
328 struct osc_async_page *oap;
330 page = cl_page_list_first(qin);
331 opg = osc_cl_page_osc(page, osc);
334 LASSERTF(osc == oap->oap_obj,
335 "obj mismatch: %p / %p\n", osc, oap->oap_obj);
337 if (!list_empty(&oap->oap_rpc_item)) {
338 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
344 /* The page may be already in dirty cache. */
345 if (list_empty(&oap->oap_pending_item)) {
346 result = osc_page_cache_add(env, opg, io, cb);
351 osc_page_touch_at(env, osc2cl(osc), osc_index(opg),
352 page == last_page ? to : PAGE_SIZE);
354 cl_page_list_del(env, qin, page);
356 /* if there are no more slots, do the callback & reinit */
357 if (pagevec_add(pvec, page->cp_vmpage) == 0) {
358 (*cb)(env, io, pvec);
359 pagevec_reinit(pvec);
362 /* The shrink interval is in seconds, so we can update it once per
363 * write, rather than once per page.
365 osc_update_next_shrink(osc_cli(osc));
368 /* Clean up any partially full pagevecs */
369 if (pagevec_count(pvec) != 0)
370 (*cb)(env, io, pvec);
372 /* Can't access these pages any more. Page can be in transfer and
373 * complete at any time. */
375 /* for sync write, kernel will wait for this page to be flushed before
376 * osc_io_end() is called, so release it earlier.
377 * for mkwrite(), it's known there is no further pages. */
378 if (cl_io_is_sync_write(io) && oio->oi_active != NULL) {
379 osc_extent_release(env, oio->oi_active);
380 oio->oi_active = NULL;
383 CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, result);
386 EXPORT_SYMBOL(osc_io_commit_async);
388 void osc_io_extent_release(const struct lu_env *env,
389 const struct cl_io_slice *ios)
391 struct osc_io *oio = cl2osc_io(env, ios);
393 if (oio->oi_active != NULL) {
394 osc_extent_release(env, oio->oi_active);
395 oio->oi_active = NULL;
398 EXPORT_SYMBOL(osc_io_extent_release);
400 static bool osc_import_not_healthy(struct obd_import *imp)
402 return imp->imp_invalid || imp->imp_deactive ||
403 !(imp->imp_state == LUSTRE_IMP_FULL ||
404 imp->imp_state == LUSTRE_IMP_IDLE);
407 int osc_io_iter_init(const struct lu_env *env, const struct cl_io_slice *ios)
409 struct osc_object *osc = cl2osc(ios->cis_obj);
410 struct obd_import *imp = osc_cli(osc)->cl_import;
411 struct osc_io *oio = osc_env_io(env);
416 spin_lock(&imp->imp_lock);
418 * check whether this OSC device is available for non-delay read,
419 * fast switching mirror if we haven't tried all mirrors.
421 if (ios->cis_io->ci_type == CIT_READ && ios->cis_io->ci_ndelay &&
422 !ios->cis_io->ci_tried_all_mirrors && osc_import_not_healthy(imp)) {
424 } else if (likely(!imp->imp_invalid)) {
425 atomic_inc(&osc->oo_nr_ios);
426 oio->oi_is_active = 1;
429 spin_unlock(&imp->imp_lock);
431 if (capable(CAP_SYS_RESOURCE))
432 oio->oi_cap_sys_resource = 1;
436 EXPORT_SYMBOL(osc_io_iter_init);
438 void osc_io_iter_fini(const struct lu_env *env,
439 const struct cl_io_slice *ios)
441 struct osc_io *oio = osc_env_io(env);
443 if (oio->oi_is_active) {
444 struct osc_object *osc = cl2osc(ios->cis_obj);
446 oio->oi_is_active = 0;
447 LASSERT(atomic_read(&osc->oo_nr_ios) > 0);
448 if (atomic_dec_and_test(&osc->oo_nr_ios))
449 wake_up(&osc->oo_io_waitq);
452 EXPORT_SYMBOL(osc_io_iter_fini);
454 void osc_io_rw_iter_fini(const struct lu_env *env,
455 const struct cl_io_slice *ios)
457 struct osc_io *oio = osc_env_io(env);
458 struct osc_object *osc = cl2osc(ios->cis_obj);
460 if (oio->oi_lru_reserved > 0) {
461 osc_lru_unreserve(osc_cli(osc), oio->oi_lru_reserved);
462 oio->oi_lru_reserved = 0;
464 oio->oi_write_osclock = NULL;
465 oio->oi_read_osclock = NULL;
467 osc_io_iter_fini(env, ios);
469 EXPORT_SYMBOL(osc_io_rw_iter_fini);
471 int osc_io_fault_start(const struct lu_env *env, const struct cl_io_slice *ios)
474 struct cl_fault_io *fio;
478 fio = &io->u.ci_fault;
479 CDEBUG(D_INFO, "%lu %d %zu\n",
480 fio->ft_index, fio->ft_writable, fio->ft_nob);
482 * If mapping is writeable, adjust kms to cover this page,
483 * but do not extend kms beyond actual file size.
486 if (fio->ft_writable)
487 osc_page_touch_at(env, ios->cis_obj,
488 fio->ft_index, fio->ft_nob);
491 EXPORT_SYMBOL(osc_io_fault_start);
494 static int osc_async_upcall(void *a, int rc)
496 struct osc_async_cbargs *args = a;
499 complete(&args->opc_sync);
504 * Checks that there are no pages being written in the extent being truncated.
506 static bool trunc_check_cb(const struct lu_env *env, struct cl_io *io,
507 void **pvec, int count, void *cbdata)
511 for (i = 0; i < count; i++) {
512 struct osc_page *ops = pvec[i];
513 struct cl_page *page = ops->ops_cl.cpl_page;
514 struct osc_async_page *oap;
515 __u64 start = *(__u64 *)cbdata;
518 if (oap->oap_cmd & OBD_BRW_WRITE &&
519 !list_empty(&oap->oap_pending_item))
520 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %llu/%s.\n",
521 start, current->comm);
523 if (PageLocked(page->cp_vmpage))
524 CDEBUG(D_CACHE, "page %p index %lu locked for cmd=%d\n",
525 ops, osc_index(ops), oap->oap_cmd);
530 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
531 struct osc_io *oio, __u64 size)
533 struct cl_object *clob;
537 clob = oio->oi_cl.cis_obj;
538 start = cl_index(clob, size);
539 partial = cl_offset(clob, start) < size;
542 * Complain if there are pages in the truncated region.
544 osc_page_gang_lookup(env, io, cl2osc(clob),
545 start + partial, CL_PAGE_EOF,
546 trunc_check_cb, (void *)&size);
550 * Flush affected pages prior punch.
551 * We shouldn't discard them locally first because that could be data loss
552 * if server doesn't support fallocate punch, we also need these data to be
553 * flushed first to prevent re-ordering with the punch
555 int osc_punch_start(const struct lu_env *env, struct cl_io *io,
556 struct cl_object *obj)
558 struct osc_object *osc = cl2osc(obj);
559 pgoff_t pg_start = cl_index(obj, io->u.ci_setattr.sa_falloc_offset);
560 pgoff_t pg_end = cl_index(obj, io->u.ci_setattr.sa_falloc_end - 1);
564 rc = osc_cache_writeback_range(env, osc, pg_start, pg_end, 1, 0);
568 osc_page_gang_lookup(env, io, osc, pg_start, pg_end, osc_discard_cb,
572 EXPORT_SYMBOL(osc_punch_start);
574 static int osc_io_setattr_start(const struct lu_env *env,
575 const struct cl_io_slice *slice)
577 struct cl_io *io = slice->cis_io;
578 struct osc_io *oio = cl2osc_io(env, slice);
579 struct cl_object *obj = slice->cis_obj;
580 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
581 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
582 struct obdo *oa = &oio->oi_oa;
583 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
584 unsigned int ia_avalid = io->u.ci_setattr.sa_avalid;
585 enum op_xvalid ia_xvalid = io->u.ci_setattr.sa_xvalid;
587 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
588 bool io_is_falloc = cl_io_is_fallocate(io);
591 /* truncate cache dirty pages first */
592 if (cl_io_is_trunc(io))
593 result = osc_cache_truncate_start(env, cl2osc(obj), size,
595 /* flush local pages prior punching them on server */
597 io->u.ci_setattr.sa_falloc_mode & FALLOC_FL_PUNCH_HOLE)
598 result = osc_punch_start(env, io, obj);
600 if (result == 0 && oio->oi_lockless == 0) {
601 cl_object_attr_lock(obj);
602 result = cl_object_attr_get(env, obj, attr);
604 struct ost_lvb *lvb = &io->u.ci_setattr.sa_attr;
605 unsigned int cl_valid = 0;
607 if (ia_avalid & ATTR_SIZE) {
608 attr->cat_size = size;
609 attr->cat_kms = size;
610 cl_valid = (CAT_SIZE | CAT_KMS);
612 if (ia_avalid & ATTR_MTIME_SET) {
613 attr->cat_mtime = lvb->lvb_mtime;
614 cl_valid |= CAT_MTIME;
616 if (ia_avalid & ATTR_ATIME_SET) {
617 attr->cat_atime = lvb->lvb_atime;
618 cl_valid |= CAT_ATIME;
620 if (ia_xvalid & OP_XVALID_CTIME_SET) {
621 attr->cat_ctime = lvb->lvb_ctime;
622 cl_valid |= CAT_CTIME;
624 result = cl_object_attr_update(env, obj, attr,
627 cl_object_attr_unlock(obj);
629 memset(oa, 0, sizeof(*oa));
631 oa->o_oi = loi->loi_oi;
632 obdo_set_parent_fid(oa, io->u.ci_setattr.sa_parent_fid);
633 oa->o_stripe_idx = io->u.ci_setattr.sa_stripe_index;
634 oa->o_layout = io->u.ci_setattr.sa_layout;
635 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP |
637 if (ia_avalid & ATTR_CTIME) {
638 oa->o_valid |= OBD_MD_FLCTIME;
639 oa->o_ctime = attr->cat_ctime;
641 if (ia_avalid & ATTR_ATIME) {
642 oa->o_valid |= OBD_MD_FLATIME;
643 oa->o_atime = attr->cat_atime;
645 if (ia_avalid & ATTR_MTIME) {
646 oa->o_valid |= OBD_MD_FLMTIME;
647 oa->o_mtime = attr->cat_mtime;
650 if (ia_avalid & ATTR_SIZE || io_is_falloc) {
651 if (oio->oi_lockless) {
652 oa->o_flags = OBD_FL_SRVLOCK;
653 oa->o_valid |= OBD_MD_FLFLAGS;
656 if (io->ci_layout_version > 0) {
657 /* verify layout version */
658 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
659 oa->o_layout_version = io->ci_layout_version;
662 LASSERT(oio->oi_lockless == 0);
665 if (ia_xvalid & OP_XVALID_FLAGS) {
666 oa->o_flags = io->u.ci_setattr.sa_attr_flags;
667 oa->o_valid |= OBD_MD_FLFLAGS;
670 init_completion(&cbargs->opc_sync);
673 int falloc_mode = io->u.ci_setattr.sa_falloc_mode;
675 oa->o_size = io->u.ci_setattr.sa_falloc_offset;
676 oa->o_blocks = io->u.ci_setattr.sa_falloc_end;
677 oa->o_uid = io->u.ci_setattr.sa_falloc_uid;
678 oa->o_gid = io->u.ci_setattr.sa_falloc_gid;
679 oa->o_projid = io->u.ci_setattr.sa_falloc_projid;
680 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
681 OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID;
684 "size %llu blocks %llu uid %u gid %u prjid %u\n",
685 oa->o_size, oa->o_blocks, oa->o_uid, oa->o_gid,
687 result = osc_fallocate_base(osc_export(cl2osc(obj)),
688 oa, osc_async_upcall,
689 cbargs, falloc_mode);
690 } else if (ia_avalid & ATTR_SIZE) {
692 oa->o_blocks = OBD_OBJECT_EOF;
693 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
694 result = osc_punch_send(osc_export(cl2osc(obj)),
695 oa, osc_async_upcall, cbargs);
697 result = osc_setattr_async(osc_export(cl2osc(obj)),
698 oa, osc_async_upcall,
699 cbargs, PTLRPCD_SET);
701 cbargs->opc_rpc_sent = result == 0;
707 void osc_io_setattr_end(const struct lu_env *env,
708 const struct cl_io_slice *slice)
710 struct cl_io *io = slice->cis_io;
711 struct osc_io *oio = cl2osc_io(env, slice);
712 struct cl_object *obj = slice->cis_obj;
713 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
714 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
715 struct obdo *oa = &oio->oi_oa;
716 unsigned int cl_valid = 0;
719 if (cbargs->opc_rpc_sent) {
720 wait_for_completion(&cbargs->opc_sync);
721 result = io->ci_result = cbargs->opc_rc;
724 if (cl_io_is_trunc(io)) {
725 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
728 cl_object_attr_lock(obj);
729 if (oa->o_valid & OBD_MD_FLBLOCKS) {
730 attr->cat_blocks = oa->o_blocks;
731 cl_valid |= CAT_BLOCKS;
734 cl_object_attr_update(env, obj, attr, cl_valid);
735 cl_object_attr_unlock(obj);
737 osc_trunc_check(env, io, oio, size);
738 osc_cache_truncate_end(env, oio->oi_trunc);
739 oio->oi_trunc = NULL;
742 if (cl_io_is_fallocate(io)) {
744 cl_object_attr_lock(obj);
746 if (oa->o_valid & OBD_MD_FLBLOCKS) {
747 attr->cat_blocks = oa->o_blocks;
748 cl_valid |= CAT_BLOCKS;
751 cl_object_attr_update(env, obj, attr, cl_valid);
752 cl_object_attr_unlock(obj);
756 EXPORT_SYMBOL(osc_io_setattr_end);
758 struct osc_data_version_args {
759 struct osc_io *dva_oio;
763 osc_data_version_interpret(const struct lu_env *env, struct ptlrpc_request *req,
766 struct osc_data_version_args *dva = args;
767 struct osc_io *oio = dva->dva_oio;
768 const struct ost_body *body;
774 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
776 GOTO(out, rc = -EPROTO);
778 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, &oio->oi_oa,
782 oio->oi_cbarg.opc_rc = rc;
783 complete(&oio->oi_cbarg.opc_sync);
788 static int osc_io_data_version_start(const struct lu_env *env,
789 const struct cl_io_slice *slice)
791 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
792 struct osc_io *oio = cl2osc_io(env, slice);
793 struct obdo *oa = &oio->oi_oa;
794 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
795 struct osc_object *obj = cl2osc(slice->cis_obj);
796 struct lov_oinfo *loi = obj->oo_oinfo;
797 struct obd_export *exp = osc_export(obj);
798 struct ptlrpc_request *req;
799 struct ost_body *body;
800 struct osc_data_version_args *dva;
804 memset(oa, 0, sizeof(*oa));
805 oa->o_oi = loi->loi_oi;
806 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
808 if (dv->dv_flags & (LL_DV_RD_FLUSH | LL_DV_WR_FLUSH)) {
809 oa->o_valid |= OBD_MD_FLFLAGS;
810 oa->o_flags |= OBD_FL_SRVLOCK;
811 if (dv->dv_flags & LL_DV_WR_FLUSH)
812 oa->o_flags |= OBD_FL_FLUSH;
815 init_completion(&cbargs->opc_sync);
817 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
821 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
823 ptlrpc_request_free(req);
827 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
828 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
830 ptlrpc_request_set_replen(req);
831 req->rq_interpret_reply = osc_data_version_interpret;
832 dva = ptlrpc_req_async_args(dva, req);
835 ptlrpcd_add_req(req);
840 static void osc_io_data_version_end(const struct lu_env *env,
841 const struct cl_io_slice *slice)
843 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
844 struct osc_io *oio = cl2osc_io(env, slice);
845 struct cl_object *obj = slice->cis_obj;
846 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
847 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
848 struct obdo *oa = &oio->oi_oa;
849 unsigned int cl_valid = 0;
852 wait_for_completion(&cbargs->opc_sync);
854 if (cbargs->opc_rc != 0) {
855 slice->cis_io->ci_result = cbargs->opc_rc;
857 slice->cis_io->ci_result = 0;
859 (OBD_MD_LAYOUT_VERSION | OBD_MD_FLDATAVERSION)))
860 slice->cis_io->ci_result = -ENOTSUPP;
862 if (oa->o_valid & OBD_MD_LAYOUT_VERSION)
863 dv->dv_layout_version = oa->o_layout_version;
864 if (oa->o_valid & OBD_MD_FLDATAVERSION)
865 dv->dv_data_version = oa->o_data_version;
867 if (dv->dv_flags & LL_DV_SZ_UPDATE) {
868 if (oa->o_valid & OBD_MD_FLSIZE) {
869 attr->cat_size = oa->o_size;
870 cl_valid |= CAT_SIZE;
873 if (oa->o_valid & OBD_MD_FLBLOCKS) {
874 attr->cat_blocks = oa->o_blocks;
875 cl_valid |= CAT_BLOCKS;
878 cl_object_attr_lock(obj);
879 cl_object_attr_update(env, obj, attr, cl_valid);
880 cl_object_attr_unlock(obj);
887 int osc_io_read_start(const struct lu_env *env,
888 const struct cl_io_slice *slice)
890 struct cl_object *obj = slice->cis_obj;
891 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
895 if (!slice->cis_io->ci_noatime) {
896 cl_object_attr_lock(obj);
897 attr->cat_atime = ktime_get_real_seconds();
898 rc = cl_object_attr_update(env, obj, attr, CAT_ATIME);
899 cl_object_attr_unlock(obj);
904 EXPORT_SYMBOL(osc_io_read_start);
906 int osc_io_write_start(const struct lu_env *env,
907 const struct cl_io_slice *slice)
909 struct cl_object *obj = slice->cis_obj;
910 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
914 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_SETTIME, 1);
915 cl_object_attr_lock(obj);
916 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
917 rc = cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
918 cl_object_attr_unlock(obj);
922 EXPORT_SYMBOL(osc_io_write_start);
924 int osc_fsync_ost(const struct lu_env *env, struct osc_object *obj,
925 struct cl_fsync_io *fio)
927 struct osc_io *oio = osc_env_io(env);
928 struct obdo *oa = &oio->oi_oa;
929 struct lov_oinfo *loi = obj->oo_oinfo;
930 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
934 memset(oa, 0, sizeof(*oa));
935 oa->o_oi = loi->loi_oi;
936 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
938 /* reload size abd blocks for start and end of sync range */
939 oa->o_size = fio->fi_start;
940 oa->o_blocks = fio->fi_end;
941 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
943 obdo_set_parent_fid(oa, fio->fi_fid);
945 init_completion(&cbargs->opc_sync);
947 rc = osc_sync_base(obj, oa, osc_async_upcall, cbargs, PTLRPCD_SET);
950 EXPORT_SYMBOL(osc_fsync_ost);
952 int osc_io_fsync_start(const struct lu_env *env,
953 const struct cl_io_slice *slice)
955 struct cl_io *io = slice->cis_io;
956 struct cl_fsync_io *fio = &io->u.ci_fsync;
957 struct cl_object *obj = slice->cis_obj;
958 struct osc_object *osc = cl2osc(obj);
959 pgoff_t start = cl_index(obj, fio->fi_start);
960 pgoff_t end = cl_index(obj, fio->fi_end);
964 if (fio->fi_end == OBD_OBJECT_EOF)
967 result = osc_cache_writeback_range(env, osc, start, end, 0,
968 fio->fi_mode == CL_FSYNC_DISCARD);
970 fio->fi_nr_written += result;
973 if (fio->fi_mode == CL_FSYNC_ALL) {
976 /* we have to wait for writeback to finish before we can
977 * send OST_SYNC RPC. This is bad because it causes extents
978 * to be written osc by osc. However, we usually start
979 * writeback before CL_FSYNC_ALL so this won't have any real
981 rc = osc_cache_wait_range(env, osc, start, end);
984 rc = osc_fsync_ost(env, osc, fio);
992 void osc_io_fsync_end(const struct lu_env *env,
993 const struct cl_io_slice *slice)
995 struct cl_fsync_io *fio = &slice->cis_io->u.ci_fsync;
996 struct cl_object *obj = slice->cis_obj;
997 pgoff_t start = cl_index(obj, fio->fi_start);
998 pgoff_t end = cl_index(obj, fio->fi_end);
1001 if (fio->fi_mode == CL_FSYNC_LOCAL) {
1002 result = osc_cache_wait_range(env, cl2osc(obj), start, end);
1003 } else if (fio->fi_mode == CL_FSYNC_ALL) {
1004 struct osc_io *oio = cl2osc_io(env, slice);
1005 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1007 wait_for_completion(&cbargs->opc_sync);
1009 result = cbargs->opc_rc;
1011 slice->cis_io->ci_result = result;
1013 EXPORT_SYMBOL(osc_io_fsync_end);
1015 static int osc_io_ladvise_start(const struct lu_env *env,
1016 const struct cl_io_slice *slice)
1019 struct cl_io *io = slice->cis_io;
1020 struct osc_io *oio = cl2osc_io(env, slice);
1021 struct cl_object *obj = slice->cis_obj;
1022 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1023 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
1024 struct obdo *oa = &oio->oi_oa;
1025 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1026 struct lu_ladvise *ladvise;
1027 struct ladvise_hdr *ladvise_hdr;
1032 /* TODO: add multiple ladvise support in CLIO */
1033 buf_size = offsetof(typeof(*ladvise_hdr), lah_advise[num_advise]);
1034 if (osc_env_info(env)->oti_ladvise_buf.lb_len < buf_size)
1035 lu_buf_realloc(&osc_env_info(env)->oti_ladvise_buf, buf_size);
1037 ladvise_hdr = osc_env_info(env)->oti_ladvise_buf.lb_buf;
1038 if (ladvise_hdr == NULL)
1041 memset(ladvise_hdr, 0, buf_size);
1042 ladvise_hdr->lah_magic = LADVISE_MAGIC;
1043 ladvise_hdr->lah_count = num_advise;
1044 ladvise_hdr->lah_flags = lio->li_flags;
1046 memset(oa, 0, sizeof(*oa));
1047 oa->o_oi = loi->loi_oi;
1048 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1049 obdo_set_parent_fid(oa, lio->li_fid);
1051 ladvise = ladvise_hdr->lah_advise;
1052 ladvise->lla_start = lio->li_start;
1053 ladvise->lla_end = lio->li_end;
1054 ladvise->lla_advice = lio->li_advice;
1056 if (lio->li_flags & LF_ASYNC) {
1057 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1058 ladvise_hdr, NULL, NULL, NULL);
1060 init_completion(&cbargs->opc_sync);
1061 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1062 ladvise_hdr, osc_async_upcall,
1063 cbargs, PTLRPCD_SET);
1064 cbargs->opc_rpc_sent = result == 0;
1069 static void osc_io_ladvise_end(const struct lu_env *env,
1070 const struct cl_io_slice *slice)
1072 struct cl_io *io = slice->cis_io;
1073 struct osc_io *oio = cl2osc_io(env, slice);
1074 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1076 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
1078 if ((!(lio->li_flags & LF_ASYNC)) && cbargs->opc_rpc_sent) {
1079 wait_for_completion(&cbargs->opc_sync);
1080 result = cbargs->opc_rc;
1082 slice->cis_io->ci_result = result;
1085 void osc_io_end(const struct lu_env *env, const struct cl_io_slice *slice)
1087 struct osc_io *oio = cl2osc_io(env, slice);
1089 if (oio->oi_active) {
1090 osc_extent_release(env, oio->oi_active);
1091 oio->oi_active = NULL;
1094 EXPORT_SYMBOL(osc_io_end);
1096 struct osc_lseek_args {
1097 struct osc_io *lsa_oio;
1100 static int osc_lseek_interpret(const struct lu_env *env,
1101 struct ptlrpc_request *req,
1104 struct ost_body *reply;
1105 struct osc_lseek_args *lsa = arg;
1106 struct osc_io *oio = lsa->lsa_oio;
1107 struct cl_io *io = oio->oi_cl.cis_io;
1108 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1115 reply = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1117 GOTO(out, rc = -EPROTO);
1119 lsio->ls_result = reply->oa.o_size;
1121 osc_async_upcall(&oio->oi_cbarg, rc);
1125 int osc_io_lseek_start(const struct lu_env *env,
1126 const struct cl_io_slice *slice)
1128 struct cl_io *io = slice->cis_io;
1129 struct osc_io *oio = cl2osc_io(env, slice);
1130 struct cl_object *obj = slice->cis_obj;
1131 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1132 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1133 struct obdo *oa = &oio->oi_oa;
1134 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1135 struct obd_export *exp = osc_export(cl2osc(obj));
1136 struct ptlrpc_request *req;
1137 struct ost_body *body;
1138 struct osc_lseek_args *lsa;
1143 /* No negative values at this point */
1144 LASSERT(lsio->ls_start >= 0);
1145 LASSERT(lsio->ls_whence == SEEK_HOLE || lsio->ls_whence == SEEK_DATA);
1147 /* with IO lock taken we have object size in LVB and can check
1148 * boundaries prior sending LSEEK RPC
1150 if (lsio->ls_start >= loi->loi_lvb.lvb_size) {
1151 /* consider area beyond end of object as hole */
1152 if (lsio->ls_whence == SEEK_HOLE)
1153 lsio->ls_result = lsio->ls_start;
1155 lsio->ls_result = -ENXIO;
1159 /* if LSEEK RPC is not supported by server, consider whole stripe
1160 * object is data with hole after end of object
1162 if (!exp_connect_lseek(exp)) {
1163 if (lsio->ls_whence == SEEK_HOLE)
1164 lsio->ls_result = loi->loi_lvb.lvb_size;
1166 lsio->ls_result = lsio->ls_start;
1170 memset(oa, 0, sizeof(*oa));
1171 oa->o_oi = loi->loi_oi;
1172 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1173 oa->o_size = lsio->ls_start;
1174 oa->o_mode = lsio->ls_whence;
1175 if (oio->oi_lockless) {
1176 oa->o_flags = OBD_FL_SRVLOCK;
1177 oa->o_valid |= OBD_MD_FLFLAGS;
1180 init_completion(&cbargs->opc_sync);
1181 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SEEK);
1185 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SEEK);
1187 ptlrpc_request_free(req);
1191 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1192 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1193 ptlrpc_request_set_replen(req);
1194 req->rq_interpret_reply = osc_lseek_interpret;
1195 lsa = ptlrpc_req_async_args(lsa, req);
1198 ptlrpcd_add_req(req);
1199 cbargs->opc_rpc_sent = 1;
1203 EXPORT_SYMBOL(osc_io_lseek_start);
1205 void osc_io_lseek_end(const struct lu_env *env,
1206 const struct cl_io_slice *slice)
1208 struct osc_io *oio = cl2osc_io(env, slice);
1209 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1212 if (cbargs->opc_rpc_sent) {
1213 wait_for_completion(&cbargs->opc_sync);
1214 rc = cbargs->opc_rc;
1216 slice->cis_io->ci_result = rc;
1218 EXPORT_SYMBOL(osc_io_lseek_end);
1220 int osc_io_lru_reserve(const struct lu_env *env,
1221 const struct cl_io_slice *ios,
1222 loff_t pos, size_t bytes)
1224 struct osc_object *osc = cl2osc(ios->cis_obj);
1225 struct osc_io *oio = osc_env_io(env);
1226 unsigned long npages = 0;
1231 page_offset = pos & ~PAGE_MASK;
1234 if (bytes > PAGE_SIZE - page_offset)
1235 bytes -= (PAGE_SIZE - page_offset);
1239 npages += (bytes + PAGE_SIZE - 1) >> PAGE_SHIFT;
1240 oio->oi_lru_reserved = osc_lru_reserve(osc_cli(osc), npages);
1244 EXPORT_SYMBOL(osc_io_lru_reserve);
1246 static const struct cl_io_operations osc_io_ops = {
1249 .cio_iter_init = osc_io_iter_init,
1250 .cio_iter_fini = osc_io_rw_iter_fini,
1251 .cio_start = osc_io_read_start,
1252 .cio_fini = osc_io_fini
1255 .cio_iter_init = osc_io_iter_init,
1256 .cio_iter_fini = osc_io_rw_iter_fini,
1257 .cio_start = osc_io_write_start,
1258 .cio_end = osc_io_end,
1259 .cio_fini = osc_io_fini
1262 .cio_iter_init = osc_io_iter_init,
1263 .cio_iter_fini = osc_io_iter_fini,
1264 .cio_start = osc_io_setattr_start,
1265 .cio_end = osc_io_setattr_end
1267 [CIT_DATA_VERSION] = {
1268 .cio_start = osc_io_data_version_start,
1269 .cio_end = osc_io_data_version_end,
1272 .cio_iter_init = osc_io_iter_init,
1273 .cio_iter_fini = osc_io_iter_fini,
1274 .cio_start = osc_io_fault_start,
1275 .cio_end = osc_io_end,
1276 .cio_fini = osc_io_fini
1279 .cio_start = osc_io_fsync_start,
1280 .cio_end = osc_io_fsync_end,
1281 .cio_fini = osc_io_fini
1284 .cio_start = osc_io_ladvise_start,
1285 .cio_end = osc_io_ladvise_end,
1286 .cio_fini = osc_io_fini
1289 .cio_start = osc_io_lseek_start,
1290 .cio_end = osc_io_lseek_end,
1291 .cio_fini = osc_io_fini
1294 .cio_fini = osc_io_fini
1297 .cio_read_ahead = osc_io_read_ahead,
1298 .cio_lru_reserve = osc_io_lru_reserve,
1299 .cio_submit = osc_io_submit,
1300 .cio_commit_async = osc_io_commit_async,
1301 .cio_extent_release = osc_io_extent_release
1304 /*****************************************************************************
1306 * Transfer operations.
1310 int osc_io_init(const struct lu_env *env,
1311 struct cl_object *obj, struct cl_io *io)
1313 struct osc_io *oio = osc_env_io(env);
1315 CL_IO_SLICE_CLEAN(oio, oi_cl);
1316 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);