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"
50 /*****************************************************************************
56 static void osc_io_fini(const struct lu_env *env, const struct cl_io_slice *io)
60 void osc_read_ahead_release(const struct lu_env *env, struct cl_read_ahead *ra)
62 struct ldlm_lock *dlmlock = ra->cra_dlmlock;
63 struct osc_io *oio = ra->cra_oio;
64 struct lustre_handle lockh;
66 oio->oi_is_readahead = 0;
67 ldlm_lock2handle(dlmlock, &lockh);
68 ldlm_lock_decref(&lockh, LCK_PR);
69 LDLM_LOCK_PUT(dlmlock);
71 EXPORT_SYMBOL(osc_read_ahead_release);
73 static int osc_io_read_ahead(const struct lu_env *env,
74 const struct cl_io_slice *ios,
75 pgoff_t start, struct cl_read_ahead *ra)
77 struct osc_object *osc = cl2osc(ios->cis_obj);
78 struct osc_io *oio = cl2osc_io(env, ios);
79 struct ldlm_lock *dlmlock;
80 int result = -ENODATA;
84 oio->oi_is_readahead = true;
85 dlmlock = osc_dlmlock_at_pgoff(env, osc, start, 0);
86 if (dlmlock != NULL) {
87 LASSERT(dlmlock->l_ast_data == osc);
88 if (dlmlock->l_req_mode != LCK_PR) {
89 struct lustre_handle lockh;
90 ldlm_lock2handle(dlmlock, &lockh);
91 ldlm_lock_addref(&lockh, LCK_PR);
92 ldlm_lock_decref(&lockh, dlmlock->l_req_mode);
95 ra->cra_rpc_pages = osc_cli(osc)->cl_max_pages_per_rpc;
96 ra->cra_end_idx = cl_index(osc2cl(osc),
97 dlmlock->l_policy_data.l_extent.end);
98 ra->cra_release = osc_read_ahead_release;
99 ra->cra_dlmlock = dlmlock;
101 if (ra->cra_end_idx != CL_PAGE_EOF)
102 ra->cra_contention = true;
110 * An implementation of cl_io_operations::cio_io_submit() method for osc
111 * layer. Iterates over pages in the in-queue, prepares each for io by calling
112 * cl_page_prep() and then either submits them through osc_io_submit_page()
113 * or, if page is already submitted, changes osc flags through
114 * osc_set_async_flags().
116 int osc_io_submit(const struct lu_env *env, const struct cl_io_slice *ios,
117 enum cl_req_type crt, struct cl_2queue *queue)
119 struct cl_page *page;
121 struct client_obd *cli = NULL;
122 struct osc_object *osc = NULL; /* to keep gcc happy */
123 struct osc_page *opg;
127 struct cl_page_list *qin = &queue->c2_qin;
128 struct cl_page_list *qout = &queue->c2_qout;
129 unsigned int queued = 0;
132 unsigned int max_pages;
133 unsigned int ppc_bits; /* pages per chunk bits */
135 ktime_t submit_time = ktime_get();
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 if (page->cp_type != CPT_TRANSIENT) {
195 spin_lock(&oap->oap_lock);
196 oap->oap_async_flags = ASYNC_URGENT|ASYNC_READY;
197 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
198 spin_unlock(&oap->oap_lock);
201 osc_page_submit(env, opg, crt, brw_flags, submit_time);
202 list_add_tail(&oap->oap_pending_item, &list);
204 if (page->cp_sync_io != NULL)
205 cl_page_list_move(qout, qin, page);
207 cl_page_list_del(env, qin, page);
210 if (queued == max_pages) {
212 } else if (crt == CRT_WRITE) {
214 unsigned int next_chunks;
216 chunks = (queued + ppc - 1) >> ppc_bits;
217 /* chunk number if add another page */
218 next_chunks = (queued + ppc) >> ppc_bits;
220 /* next page will excceed write chunk limit */
221 if (chunks == osc_max_write_chunks(cli) &&
222 next_chunks > chunks)
227 result = osc_queue_sync_pages(env, io, osc, &list,
237 result = osc_queue_sync_pages(env, io, osc, &list, brw_flags);
239 /* Update c/mtime for sync write. LU-7310 */
240 if (crt == CRT_WRITE && qout->pl_nr > 0 && result == 0) {
241 struct cl_object *obj = ios->cis_obj;
242 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
244 cl_object_attr_lock(obj);
245 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
246 cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
247 cl_object_attr_unlock(obj);
250 CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
251 return qout->pl_nr > 0 ? 0 : result;
253 EXPORT_SYMBOL(osc_io_submit);
256 * This is called to update the attributes when modifying a specific page,
257 * both when making new pages and when doing updates to existing cached pages.
259 * Expand stripe KMS if necessary.
261 void osc_page_touch_at(const struct lu_env *env, struct cl_object *obj,
262 pgoff_t idx, size_t to)
264 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
265 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
271 /* offset within stripe */
272 kms = cl_offset(obj, idx) + to;
274 cl_object_attr_lock(obj);
275 CDEBUG(D_INODE, "stripe KMS %sincreasing %llu->%llu %llu\n",
276 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
277 loi->loi_lvb.lvb_size);
279 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
280 valid = CAT_MTIME | CAT_CTIME;
281 if (kms > loi->loi_kms) {
285 if (kms > loi->loi_lvb.lvb_size) {
286 attr->cat_size = kms;
289 cl_object_attr_update(env, obj, attr, valid);
290 cl_object_attr_unlock(obj);
295 int osc_io_commit_async(const struct lu_env *env,
296 const struct cl_io_slice *ios,
297 struct cl_page_list *qin, int from, int to,
300 struct cl_io *io = ios->cis_io;
301 struct osc_io *oio = cl2osc_io(env, ios);
302 struct osc_object *osc = cl2osc(ios->cis_obj);
303 struct cl_page *page;
304 struct cl_page *last_page;
305 struct osc_page *opg;
306 struct pagevec *pvec = &osc_env_info(env)->oti_pagevec;
310 LASSERT(qin->pl_nr > 0);
312 /* Handle partial page cases */
313 last_page = cl_page_list_last(qin);
314 if (oio->oi_lockless) {
315 page = cl_page_list_first(qin);
316 if (page == last_page) {
317 cl_page_clip(env, page, from, to);
320 cl_page_clip(env, page, from, PAGE_SIZE);
322 cl_page_clip(env, last_page, 0, to);
326 ll_pagevec_init(pvec, 0);
328 while (qin->pl_nr > 0) {
329 struct osc_async_page *oap;
331 page = cl_page_list_first(qin);
332 opg = osc_cl_page_osc(page, osc);
335 LASSERTF(osc == oap->oap_obj,
336 "obj mismatch: %p / %p\n", osc, oap->oap_obj);
338 if (!list_empty(&oap->oap_rpc_item)) {
339 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
345 /* The page may be already in dirty cache. */
346 if (list_empty(&oap->oap_pending_item)) {
347 result = osc_page_cache_add(env, opg, io, cb);
352 osc_page_touch_at(env, osc2cl(osc), osc_index(opg),
353 page == last_page ? to : PAGE_SIZE);
355 cl_page_list_del(env, qin, page);
357 /* if there are no more slots, do the callback & reinit */
358 if (pagevec_add(pvec, page->cp_vmpage) == 0) {
359 (*cb)(env, io, pvec);
360 pagevec_reinit(pvec);
363 /* The shrink interval is in seconds, so we can update it once per
364 * write, rather than once per page.
366 osc_update_next_shrink(osc_cli(osc));
369 /* Clean up any partially full pagevecs */
370 if (pagevec_count(pvec) != 0)
371 (*cb)(env, io, pvec);
373 /* Can't access these pages any more. Page can be in transfer and
374 * complete at any time. */
376 /* for sync write, kernel will wait for this page to be flushed before
377 * osc_io_end() is called, so release it earlier.
378 * for mkwrite(), it's known there is no further pages. */
379 if (cl_io_is_sync_write(io) && oio->oi_active != NULL) {
380 osc_extent_release(env, oio->oi_active);
381 oio->oi_active = NULL;
384 CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, result);
387 EXPORT_SYMBOL(osc_io_commit_async);
389 void osc_io_extent_release(const struct lu_env *env,
390 const struct cl_io_slice *ios)
392 struct osc_io *oio = cl2osc_io(env, ios);
394 if (oio->oi_active != NULL) {
395 osc_extent_release(env, oio->oi_active);
396 oio->oi_active = NULL;
399 EXPORT_SYMBOL(osc_io_extent_release);
401 static bool osc_import_not_healthy(struct obd_import *imp)
403 return imp->imp_invalid || imp->imp_deactive ||
404 !(imp->imp_state == LUSTRE_IMP_FULL ||
405 imp->imp_state == LUSTRE_IMP_IDLE);
408 int osc_io_iter_init(const struct lu_env *env, const struct cl_io_slice *ios)
410 struct osc_object *osc = cl2osc(ios->cis_obj);
411 struct obd_import *imp = osc_cli(osc)->cl_import;
412 struct osc_io *oio = osc_env_io(env);
417 spin_lock(&imp->imp_lock);
419 * check whether this OSC device is available for non-delay read,
420 * fast switching mirror if we haven't tried all mirrors.
422 if (ios->cis_io->ci_type == CIT_READ && ios->cis_io->ci_ndelay &&
423 !ios->cis_io->ci_tried_all_mirrors && osc_import_not_healthy(imp)) {
425 } else if (likely(!imp->imp_invalid)) {
426 atomic_inc(&osc->oo_nr_ios);
427 oio->oi_is_active = 1;
430 spin_unlock(&imp->imp_lock);
432 if (capable(CAP_SYS_RESOURCE))
433 oio->oi_cap_sys_resource = 1;
437 EXPORT_SYMBOL(osc_io_iter_init);
439 void osc_io_iter_fini(const struct lu_env *env,
440 const struct cl_io_slice *ios)
442 struct osc_io *oio = osc_env_io(env);
444 if (oio->oi_is_active) {
445 struct osc_object *osc = cl2osc(ios->cis_obj);
447 oio->oi_is_active = 0;
448 LASSERT(atomic_read(&osc->oo_nr_ios) > 0);
449 if (atomic_dec_and_test(&osc->oo_nr_ios))
450 wake_up(&osc->oo_io_waitq);
453 EXPORT_SYMBOL(osc_io_iter_fini);
455 void osc_io_rw_iter_fini(const struct lu_env *env,
456 const struct cl_io_slice *ios)
458 struct osc_io *oio = osc_env_io(env);
459 struct osc_object *osc = cl2osc(ios->cis_obj);
461 if (oio->oi_lru_reserved > 0) {
462 osc_lru_unreserve(osc_cli(osc), oio->oi_lru_reserved);
463 oio->oi_lru_reserved = 0;
465 oio->oi_write_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 %d.\n",
526 oap->oap_cmd & OBD_BRW_RWMASK);
531 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
532 struct osc_io *oio, __u64 size)
534 struct cl_object *clob;
538 clob = oio->oi_cl.cis_obj;
539 start = cl_index(clob, size);
540 partial = cl_offset(clob, start) < size;
543 * Complain if there are pages in the truncated region.
545 osc_page_gang_lookup(env, io, cl2osc(clob),
546 start + partial, CL_PAGE_EOF,
547 trunc_check_cb, (void *)&size);
551 * Flush affected pages prior punch.
552 * We shouldn't discard them locally first because that could be data loss
553 * if server doesn't support fallocate punch, we also need these data to be
554 * flushed first to prevent re-ordering with the punch
556 int osc_punch_start(const struct lu_env *env, struct cl_io *io,
557 struct cl_object *obj)
559 struct osc_object *osc = cl2osc(obj);
560 pgoff_t pg_start = cl_index(obj, io->u.ci_setattr.sa_falloc_offset);
561 pgoff_t pg_end = cl_index(obj, io->u.ci_setattr.sa_falloc_end - 1);
565 rc = osc_cache_writeback_range(env, osc, pg_start, pg_end, 1, 0);
569 osc_page_gang_lookup(env, io, osc, pg_start, pg_end, osc_discard_cb,
573 EXPORT_SYMBOL(osc_punch_start);
575 static int osc_io_setattr_start(const struct lu_env *env,
576 const struct cl_io_slice *slice)
578 struct cl_io *io = slice->cis_io;
579 struct osc_io *oio = cl2osc_io(env, slice);
580 struct cl_object *obj = slice->cis_obj;
581 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
582 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
583 struct obdo *oa = &oio->oi_oa;
584 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
585 unsigned int ia_avalid = io->u.ci_setattr.sa_avalid;
586 enum op_xvalid ia_xvalid = io->u.ci_setattr.sa_xvalid;
588 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
589 bool io_is_falloc = cl_io_is_fallocate(io);
592 /* truncate cache dirty pages first */
593 if (cl_io_is_trunc(io))
594 result = osc_cache_truncate_start(env, cl2osc(obj), size,
596 /* flush local pages prior punching them on server */
598 io->u.ci_setattr.sa_falloc_mode & FALLOC_FL_PUNCH_HOLE)
599 result = osc_punch_start(env, io, obj);
601 if (result == 0 && oio->oi_lockless == 0) {
602 cl_object_attr_lock(obj);
603 result = cl_object_attr_get(env, obj, attr);
605 struct ost_lvb *lvb = &io->u.ci_setattr.sa_attr;
606 unsigned int cl_valid = 0;
608 if (ia_avalid & ATTR_SIZE) {
609 attr->cat_size = size;
610 attr->cat_kms = size;
611 cl_valid = (CAT_SIZE | CAT_KMS);
613 if (ia_avalid & ATTR_MTIME_SET) {
614 attr->cat_mtime = lvb->lvb_mtime;
615 cl_valid |= CAT_MTIME;
617 if (ia_avalid & ATTR_ATIME_SET) {
618 attr->cat_atime = lvb->lvb_atime;
619 cl_valid |= CAT_ATIME;
621 if (ia_xvalid & OP_XVALID_CTIME_SET) {
622 attr->cat_ctime = lvb->lvb_ctime;
623 cl_valid |= CAT_CTIME;
625 result = cl_object_attr_update(env, obj, attr,
628 cl_object_attr_unlock(obj);
630 memset(oa, 0, sizeof(*oa));
632 oa->o_oi = loi->loi_oi;
633 obdo_set_parent_fid(oa, io->u.ci_setattr.sa_parent_fid);
634 oa->o_stripe_idx = io->u.ci_setattr.sa_stripe_index;
635 oa->o_layout = io->u.ci_setattr.sa_layout;
636 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP |
638 if (ia_avalid & ATTR_CTIME) {
639 oa->o_valid |= OBD_MD_FLCTIME;
640 oa->o_ctime = attr->cat_ctime;
642 if (ia_avalid & ATTR_ATIME) {
643 oa->o_valid |= OBD_MD_FLATIME;
644 oa->o_atime = attr->cat_atime;
646 if (ia_avalid & ATTR_MTIME) {
647 oa->o_valid |= OBD_MD_FLMTIME;
648 oa->o_mtime = attr->cat_mtime;
651 if (ia_avalid & ATTR_SIZE || io_is_falloc) {
652 if (oio->oi_lockless) {
653 oa->o_flags = OBD_FL_SRVLOCK;
654 oa->o_valid |= OBD_MD_FLFLAGS;
657 if (io->ci_layout_version > 0) {
658 /* verify layout version */
659 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
660 oa->o_layout_version = io->ci_layout_version;
663 LASSERT(oio->oi_lockless == 0);
666 if (ia_xvalid & OP_XVALID_FLAGS) {
667 oa->o_flags = io->u.ci_setattr.sa_attr_flags;
668 oa->o_valid |= OBD_MD_FLFLAGS;
671 init_completion(&cbargs->opc_sync);
674 int falloc_mode = io->u.ci_setattr.sa_falloc_mode;
676 oa->o_size = io->u.ci_setattr.sa_falloc_offset;
677 oa->o_blocks = io->u.ci_setattr.sa_falloc_end;
678 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
679 result = osc_fallocate_base(osc_export(cl2osc(obj)),
680 oa, osc_async_upcall,
681 cbargs, falloc_mode);
682 } else if (ia_avalid & ATTR_SIZE) {
684 oa->o_blocks = OBD_OBJECT_EOF;
685 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
686 result = osc_punch_send(osc_export(cl2osc(obj)),
687 oa, osc_async_upcall, cbargs);
689 result = osc_setattr_async(osc_export(cl2osc(obj)),
690 oa, osc_async_upcall,
691 cbargs, PTLRPCD_SET);
693 cbargs->opc_rpc_sent = result == 0;
699 void osc_io_setattr_end(const struct lu_env *env,
700 const struct cl_io_slice *slice)
702 struct cl_io *io = slice->cis_io;
703 struct osc_io *oio = cl2osc_io(env, slice);
704 struct cl_object *obj = slice->cis_obj;
705 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
706 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
707 struct obdo *oa = &oio->oi_oa;
708 unsigned int cl_valid = 0;
711 if (cbargs->opc_rpc_sent) {
712 wait_for_completion(&cbargs->opc_sync);
713 result = io->ci_result = cbargs->opc_rc;
716 if (cl_io_is_trunc(io)) {
717 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
720 cl_object_attr_lock(obj);
721 if (oa->o_valid & OBD_MD_FLBLOCKS) {
722 attr->cat_blocks = oa->o_blocks;
723 cl_valid |= CAT_BLOCKS;
726 cl_object_attr_update(env, obj, attr, cl_valid);
727 cl_object_attr_unlock(obj);
729 osc_trunc_check(env, io, oio, size);
730 osc_cache_truncate_end(env, oio->oi_trunc);
731 oio->oi_trunc = NULL;
734 if (cl_io_is_fallocate(io)) {
736 cl_object_attr_lock(obj);
738 if (oa->o_valid & OBD_MD_FLBLOCKS) {
739 attr->cat_blocks = oa->o_blocks;
740 cl_valid |= CAT_BLOCKS;
743 cl_object_attr_update(env, obj, attr, cl_valid);
744 cl_object_attr_unlock(obj);
748 EXPORT_SYMBOL(osc_io_setattr_end);
750 struct osc_data_version_args {
751 struct osc_io *dva_oio;
755 osc_data_version_interpret(const struct lu_env *env, struct ptlrpc_request *req,
758 struct osc_data_version_args *dva = args;
759 struct osc_io *oio = dva->dva_oio;
760 const struct ost_body *body;
766 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
768 GOTO(out, rc = -EPROTO);
770 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, &oio->oi_oa,
774 oio->oi_cbarg.opc_rc = rc;
775 complete(&oio->oi_cbarg.opc_sync);
780 static int osc_io_data_version_start(const struct lu_env *env,
781 const struct cl_io_slice *slice)
783 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
784 struct osc_io *oio = cl2osc_io(env, slice);
785 struct obdo *oa = &oio->oi_oa;
786 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
787 struct osc_object *obj = cl2osc(slice->cis_obj);
788 struct lov_oinfo *loi = obj->oo_oinfo;
789 struct obd_export *exp = osc_export(obj);
790 struct ptlrpc_request *req;
791 struct ost_body *body;
792 struct osc_data_version_args *dva;
796 memset(oa, 0, sizeof(*oa));
797 oa->o_oi = loi->loi_oi;
798 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
800 if (dv->dv_flags & (LL_DV_RD_FLUSH | LL_DV_WR_FLUSH)) {
801 oa->o_valid |= OBD_MD_FLFLAGS;
802 oa->o_flags |= OBD_FL_SRVLOCK;
803 if (dv->dv_flags & LL_DV_WR_FLUSH)
804 oa->o_flags |= OBD_FL_FLUSH;
807 init_completion(&cbargs->opc_sync);
809 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
813 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
815 ptlrpc_request_free(req);
819 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
820 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
822 ptlrpc_request_set_replen(req);
823 req->rq_interpret_reply = osc_data_version_interpret;
824 dva = ptlrpc_req_async_args(dva, req);
827 ptlrpcd_add_req(req);
832 static void osc_io_data_version_end(const struct lu_env *env,
833 const struct cl_io_slice *slice)
835 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
836 struct osc_io *oio = cl2osc_io(env, slice);
837 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
840 wait_for_completion(&cbargs->opc_sync);
842 if (cbargs->opc_rc != 0) {
843 slice->cis_io->ci_result = cbargs->opc_rc;
845 slice->cis_io->ci_result = 0;
846 if (!(oio->oi_oa.o_valid &
847 (OBD_MD_LAYOUT_VERSION | OBD_MD_FLDATAVERSION)))
848 slice->cis_io->ci_result = -ENOTSUPP;
850 if (oio->oi_oa.o_valid & OBD_MD_LAYOUT_VERSION)
851 dv->dv_layout_version = oio->oi_oa.o_layout_version;
852 if (oio->oi_oa.o_valid & OBD_MD_FLDATAVERSION)
853 dv->dv_data_version = oio->oi_oa.o_data_version;
859 int osc_io_read_start(const struct lu_env *env,
860 const struct cl_io_slice *slice)
862 struct cl_object *obj = slice->cis_obj;
863 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
867 if (!slice->cis_io->ci_noatime) {
868 cl_object_attr_lock(obj);
869 attr->cat_atime = ktime_get_real_seconds();
870 rc = cl_object_attr_update(env, obj, attr, CAT_ATIME);
871 cl_object_attr_unlock(obj);
876 EXPORT_SYMBOL(osc_io_read_start);
878 int osc_io_write_start(const struct lu_env *env,
879 const struct cl_io_slice *slice)
881 struct cl_object *obj = slice->cis_obj;
882 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
886 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_SETTIME, 1);
887 cl_object_attr_lock(obj);
888 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
889 rc = cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
890 cl_object_attr_unlock(obj);
894 EXPORT_SYMBOL(osc_io_write_start);
896 int osc_fsync_ost(const struct lu_env *env, struct osc_object *obj,
897 struct cl_fsync_io *fio)
899 struct osc_io *oio = osc_env_io(env);
900 struct obdo *oa = &oio->oi_oa;
901 struct lov_oinfo *loi = obj->oo_oinfo;
902 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
906 memset(oa, 0, sizeof(*oa));
907 oa->o_oi = loi->loi_oi;
908 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
910 /* reload size abd blocks for start and end of sync range */
911 oa->o_size = fio->fi_start;
912 oa->o_blocks = fio->fi_end;
913 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
915 obdo_set_parent_fid(oa, fio->fi_fid);
917 init_completion(&cbargs->opc_sync);
919 rc = osc_sync_base(obj, oa, osc_async_upcall, cbargs, PTLRPCD_SET);
922 EXPORT_SYMBOL(osc_fsync_ost);
924 int osc_io_fsync_start(const struct lu_env *env,
925 const struct cl_io_slice *slice)
927 struct cl_io *io = slice->cis_io;
928 struct cl_fsync_io *fio = &io->u.ci_fsync;
929 struct cl_object *obj = slice->cis_obj;
930 struct osc_object *osc = cl2osc(obj);
931 pgoff_t start = cl_index(obj, fio->fi_start);
932 pgoff_t end = cl_index(obj, fio->fi_end);
936 if (fio->fi_end == OBD_OBJECT_EOF)
939 result = osc_cache_writeback_range(env, osc, start, end, 0,
940 fio->fi_mode == CL_FSYNC_DISCARD);
942 fio->fi_nr_written += result;
945 if (fio->fi_mode == CL_FSYNC_ALL) {
948 /* we have to wait for writeback to finish before we can
949 * send OST_SYNC RPC. This is bad because it causes extents
950 * to be written osc by osc. However, we usually start
951 * writeback before CL_FSYNC_ALL so this won't have any real
953 rc = osc_cache_wait_range(env, osc, start, end);
956 rc = osc_fsync_ost(env, osc, fio);
964 void osc_io_fsync_end(const struct lu_env *env,
965 const struct cl_io_slice *slice)
967 struct cl_fsync_io *fio = &slice->cis_io->u.ci_fsync;
968 struct cl_object *obj = slice->cis_obj;
969 pgoff_t start = cl_index(obj, fio->fi_start);
970 pgoff_t end = cl_index(obj, fio->fi_end);
973 if (fio->fi_mode == CL_FSYNC_LOCAL) {
974 result = osc_cache_wait_range(env, cl2osc(obj), start, end);
975 } else if (fio->fi_mode == CL_FSYNC_ALL) {
976 struct osc_io *oio = cl2osc_io(env, slice);
977 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
979 wait_for_completion(&cbargs->opc_sync);
981 result = cbargs->opc_rc;
983 slice->cis_io->ci_result = result;
985 EXPORT_SYMBOL(osc_io_fsync_end);
987 static int osc_io_ladvise_start(const struct lu_env *env,
988 const struct cl_io_slice *slice)
991 struct cl_io *io = slice->cis_io;
992 struct osc_io *oio = cl2osc_io(env, slice);
993 struct cl_object *obj = slice->cis_obj;
994 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
995 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
996 struct obdo *oa = &oio->oi_oa;
997 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
998 struct lu_ladvise *ladvise;
999 struct ladvise_hdr *ladvise_hdr;
1004 /* TODO: add multiple ladvise support in CLIO */
1005 buf_size = offsetof(typeof(*ladvise_hdr), lah_advise[num_advise]);
1006 if (osc_env_info(env)->oti_ladvise_buf.lb_len < buf_size)
1007 lu_buf_realloc(&osc_env_info(env)->oti_ladvise_buf, buf_size);
1009 ladvise_hdr = osc_env_info(env)->oti_ladvise_buf.lb_buf;
1010 if (ladvise_hdr == NULL)
1013 memset(ladvise_hdr, 0, buf_size);
1014 ladvise_hdr->lah_magic = LADVISE_MAGIC;
1015 ladvise_hdr->lah_count = num_advise;
1016 ladvise_hdr->lah_flags = lio->li_flags;
1018 memset(oa, 0, sizeof(*oa));
1019 oa->o_oi = loi->loi_oi;
1020 oa->o_valid = OBD_MD_FLID;
1021 obdo_set_parent_fid(oa, lio->li_fid);
1023 ladvise = ladvise_hdr->lah_advise;
1024 ladvise->lla_start = lio->li_start;
1025 ladvise->lla_end = lio->li_end;
1026 ladvise->lla_advice = lio->li_advice;
1028 if (lio->li_flags & LF_ASYNC) {
1029 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1030 ladvise_hdr, NULL, NULL, NULL);
1032 init_completion(&cbargs->opc_sync);
1033 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1034 ladvise_hdr, osc_async_upcall,
1035 cbargs, PTLRPCD_SET);
1036 cbargs->opc_rpc_sent = result == 0;
1041 static void osc_io_ladvise_end(const struct lu_env *env,
1042 const struct cl_io_slice *slice)
1044 struct cl_io *io = slice->cis_io;
1045 struct osc_io *oio = cl2osc_io(env, slice);
1046 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1048 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
1050 if ((!(lio->li_flags & LF_ASYNC)) && cbargs->opc_rpc_sent) {
1051 wait_for_completion(&cbargs->opc_sync);
1052 result = cbargs->opc_rc;
1054 slice->cis_io->ci_result = result;
1057 void osc_io_end(const struct lu_env *env, const struct cl_io_slice *slice)
1059 struct osc_io *oio = cl2osc_io(env, slice);
1061 if (oio->oi_active) {
1062 osc_extent_release(env, oio->oi_active);
1063 oio->oi_active = NULL;
1066 EXPORT_SYMBOL(osc_io_end);
1068 struct osc_lseek_args {
1069 struct osc_io *lsa_oio;
1072 static int osc_lseek_interpret(const struct lu_env *env,
1073 struct ptlrpc_request *req,
1076 struct ost_body *reply;
1077 struct osc_lseek_args *lsa = arg;
1078 struct osc_io *oio = lsa->lsa_oio;
1079 struct cl_io *io = oio->oi_cl.cis_io;
1080 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1087 reply = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1089 GOTO(out, rc = -EPROTO);
1091 lsio->ls_result = reply->oa.o_size;
1093 osc_async_upcall(&oio->oi_cbarg, rc);
1097 int osc_io_lseek_start(const struct lu_env *env,
1098 const struct cl_io_slice *slice)
1100 struct cl_io *io = slice->cis_io;
1101 struct osc_io *oio = cl2osc_io(env, slice);
1102 struct cl_object *obj = slice->cis_obj;
1103 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1104 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1105 struct obdo *oa = &oio->oi_oa;
1106 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1107 struct obd_export *exp = osc_export(cl2osc(obj));
1108 struct ptlrpc_request *req;
1109 struct ost_body *body;
1110 struct osc_lseek_args *lsa;
1115 /* No negative values at this point */
1116 LASSERT(lsio->ls_start >= 0);
1117 LASSERT(lsio->ls_whence == SEEK_HOLE || lsio->ls_whence == SEEK_DATA);
1119 /* with IO lock taken we have object size in LVB and can check
1120 * boundaries prior sending LSEEK RPC
1122 if (lsio->ls_start >= loi->loi_lvb.lvb_size) {
1123 /* consider area beyond end of object as hole */
1124 if (lsio->ls_whence == SEEK_HOLE)
1125 lsio->ls_result = lsio->ls_start;
1127 lsio->ls_result = -ENXIO;
1131 /* if LSEEK RPC is not supported by server, consider whole stripe
1132 * object is data with hole after end of object
1134 if (!exp_connect_lseek(exp)) {
1135 if (lsio->ls_whence == SEEK_HOLE)
1136 lsio->ls_result = loi->loi_lvb.lvb_size;
1138 lsio->ls_result = lsio->ls_start;
1142 memset(oa, 0, sizeof(*oa));
1143 oa->o_oi = loi->loi_oi;
1144 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1145 oa->o_size = lsio->ls_start;
1146 oa->o_mode = lsio->ls_whence;
1147 if (oio->oi_lockless) {
1148 oa->o_flags = OBD_FL_SRVLOCK;
1149 oa->o_valid |= OBD_MD_FLFLAGS;
1152 init_completion(&cbargs->opc_sync);
1153 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SEEK);
1157 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SEEK);
1159 ptlrpc_request_free(req);
1163 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1164 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1165 ptlrpc_request_set_replen(req);
1166 req->rq_interpret_reply = osc_lseek_interpret;
1167 lsa = ptlrpc_req_async_args(lsa, req);
1170 ptlrpcd_add_req(req);
1171 cbargs->opc_rpc_sent = 1;
1175 EXPORT_SYMBOL(osc_io_lseek_start);
1177 void osc_io_lseek_end(const struct lu_env *env,
1178 const struct cl_io_slice *slice)
1180 struct osc_io *oio = cl2osc_io(env, slice);
1181 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1184 if (cbargs->opc_rpc_sent) {
1185 wait_for_completion(&cbargs->opc_sync);
1186 rc = cbargs->opc_rc;
1188 slice->cis_io->ci_result = rc;
1190 EXPORT_SYMBOL(osc_io_lseek_end);
1192 int osc_io_lru_reserve(const struct lu_env *env,
1193 const struct cl_io_slice *ios,
1194 loff_t pos, size_t bytes)
1196 struct osc_object *osc = cl2osc(ios->cis_obj);
1197 struct osc_io *oio = osc_env_io(env);
1198 unsigned long npages = 0;
1203 page_offset = pos & ~PAGE_MASK;
1206 if (bytes > PAGE_SIZE - page_offset)
1207 bytes -= (PAGE_SIZE - page_offset);
1211 npages += (bytes + PAGE_SIZE - 1) >> PAGE_SHIFT;
1212 oio->oi_lru_reserved = osc_lru_reserve(osc_cli(osc), npages);
1216 EXPORT_SYMBOL(osc_io_lru_reserve);
1218 static const struct cl_io_operations osc_io_ops = {
1221 .cio_iter_init = osc_io_iter_init,
1222 .cio_iter_fini = osc_io_rw_iter_fini,
1223 .cio_start = osc_io_read_start,
1224 .cio_fini = osc_io_fini
1227 .cio_iter_init = osc_io_iter_init,
1228 .cio_iter_fini = osc_io_rw_iter_fini,
1229 .cio_start = osc_io_write_start,
1230 .cio_end = osc_io_end,
1231 .cio_fini = osc_io_fini
1234 .cio_iter_init = osc_io_iter_init,
1235 .cio_iter_fini = osc_io_iter_fini,
1236 .cio_start = osc_io_setattr_start,
1237 .cio_end = osc_io_setattr_end
1239 [CIT_DATA_VERSION] = {
1240 .cio_start = osc_io_data_version_start,
1241 .cio_end = osc_io_data_version_end,
1244 .cio_iter_init = osc_io_iter_init,
1245 .cio_iter_fini = osc_io_iter_fini,
1246 .cio_start = osc_io_fault_start,
1247 .cio_end = osc_io_end,
1248 .cio_fini = osc_io_fini
1251 .cio_start = osc_io_fsync_start,
1252 .cio_end = osc_io_fsync_end,
1253 .cio_fini = osc_io_fini
1256 .cio_start = osc_io_ladvise_start,
1257 .cio_end = osc_io_ladvise_end,
1258 .cio_fini = osc_io_fini
1261 .cio_start = osc_io_lseek_start,
1262 .cio_end = osc_io_lseek_end,
1263 .cio_fini = osc_io_fini
1266 .cio_fini = osc_io_fini
1269 .cio_read_ahead = osc_io_read_ahead,
1270 .cio_lru_reserve = osc_io_lru_reserve,
1271 .cio_submit = osc_io_submit,
1272 .cio_commit_async = osc_io_commit_async,
1273 .cio_extent_release = osc_io_extent_release
1276 /*****************************************************************************
1278 * Transfer operations.
1282 int osc_io_init(const struct lu_env *env,
1283 struct cl_object *obj, struct cl_io *io)
1285 struct osc_io *oio = osc_env_io(env);
1287 CL_IO_SLICE_CLEAN(oio, oi_cl);
1288 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);