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 spin_lock(&oap->oap_lock);
198 oap->oap_async_flags = ASYNC_URGENT|ASYNC_READY;
199 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
200 spin_unlock(&oap->oap_lock);
203 osc_page_submit(env, opg, crt, brw_flags);
204 list_add_tail(&oap->oap_pending_item, &list);
206 if (page->cp_sync_io != NULL)
207 cl_page_list_move(qout, qin, page);
209 cl_page_list_del(env, qin, page);
212 if (queued == max_pages) {
214 } else if (crt == CRT_WRITE) {
216 unsigned int next_chunks;
218 chunks = (queued + ppc - 1) >> ppc_bits;
219 /* chunk number if add another page */
220 next_chunks = (queued + ppc) >> ppc_bits;
222 /* next page will excceed write chunk limit */
223 if (chunks == osc_max_write_chunks(cli) &&
224 next_chunks > chunks)
229 result = osc_queue_sync_pages(env, io, osc, &list,
239 result = osc_queue_sync_pages(env, io, osc, &list, brw_flags);
241 /* Update c/mtime for sync write. LU-7310 */
242 if (crt == CRT_WRITE && qout->pl_nr > 0 && result == 0) {
243 struct cl_object *obj = ios->cis_obj;
244 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
246 cl_object_attr_lock(obj);
247 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
248 cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
249 cl_object_attr_unlock(obj);
252 CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
253 return qout->pl_nr > 0 ? 0 : result;
255 EXPORT_SYMBOL(osc_io_submit);
258 * This is called to update the attributes when modifying a specific page,
259 * both when making new pages and when doing updates to existing cached pages.
261 * Expand stripe KMS if necessary.
263 void osc_page_touch_at(const struct lu_env *env, struct cl_object *obj,
264 pgoff_t idx, size_t to)
266 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
267 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
273 /* offset within stripe */
274 kms = cl_offset(obj, idx) + to;
276 cl_object_attr_lock(obj);
277 CDEBUG(D_INODE, "stripe KMS %sincreasing %llu->%llu %llu\n",
278 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
279 loi->loi_lvb.lvb_size);
281 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
282 valid = CAT_MTIME | CAT_CTIME;
283 if (kms > loi->loi_kms) {
287 if (kms > loi->loi_lvb.lvb_size) {
288 attr->cat_size = kms;
291 cl_object_attr_update(env, obj, attr, valid);
292 cl_object_attr_unlock(obj);
297 int osc_io_commit_async(const struct lu_env *env,
298 const struct cl_io_slice *ios,
299 struct cl_page_list *qin, int from, int to,
302 struct cl_io *io = ios->cis_io;
303 struct osc_io *oio = cl2osc_io(env, ios);
304 struct osc_object *osc = cl2osc(ios->cis_obj);
305 struct cl_page *page;
306 struct cl_page *last_page;
307 struct osc_page *opg;
308 struct pagevec *pvec = &osc_env_info(env)->oti_pagevec;
312 LASSERT(qin->pl_nr > 0);
314 /* Handle partial page cases */
315 last_page = cl_page_list_last(qin);
316 if (oio->oi_lockless) {
317 page = cl_page_list_first(qin);
318 if (page == last_page) {
319 cl_page_clip(env, page, from, to);
322 cl_page_clip(env, page, from, PAGE_SIZE);
324 cl_page_clip(env, last_page, 0, to);
328 ll_pagevec_init(pvec, 0);
330 while (qin->pl_nr > 0) {
331 struct osc_async_page *oap;
333 page = cl_page_list_first(qin);
334 opg = osc_cl_page_osc(page, osc);
337 LASSERTF(osc == oap->oap_obj,
338 "obj mismatch: %p / %p\n", osc, oap->oap_obj);
340 if (!list_empty(&oap->oap_rpc_item)) {
341 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
347 /* The page may be already in dirty cache. */
348 if (list_empty(&oap->oap_pending_item)) {
349 result = osc_page_cache_add(env, opg, io, cb);
354 osc_page_touch_at(env, osc2cl(osc), osc_index(opg),
355 page == last_page ? to : PAGE_SIZE);
357 cl_page_list_del(env, qin, page);
359 /* if there are no more slots, do the callback & reinit */
360 if (pagevec_add(pvec, page->cp_vmpage) == 0) {
361 (*cb)(env, io, pvec);
362 pagevec_reinit(pvec);
365 /* The shrink interval is in seconds, so we can update it once per
366 * write, rather than once per page.
368 osc_update_next_shrink(osc_cli(osc));
371 /* Clean up any partially full pagevecs */
372 if (pagevec_count(pvec) != 0)
373 (*cb)(env, io, pvec);
375 /* Can't access these pages any more. Page can be in transfer and
376 * complete at any time. */
378 /* for sync write, kernel will wait for this page to be flushed before
379 * osc_io_end() is called, so release it earlier.
380 * for mkwrite(), it's known there is no further pages. */
381 if (cl_io_is_sync_write(io) && oio->oi_active != NULL) {
382 osc_extent_release(env, oio->oi_active);
383 oio->oi_active = NULL;
386 CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, result);
389 EXPORT_SYMBOL(osc_io_commit_async);
391 void osc_io_extent_release(const struct lu_env *env,
392 const struct cl_io_slice *ios)
394 struct osc_io *oio = cl2osc_io(env, ios);
396 if (oio->oi_active != NULL) {
397 osc_extent_release(env, oio->oi_active);
398 oio->oi_active = NULL;
401 EXPORT_SYMBOL(osc_io_extent_release);
403 static bool osc_import_not_healthy(struct obd_import *imp)
405 return imp->imp_invalid || imp->imp_deactive ||
406 !(imp->imp_state == LUSTRE_IMP_FULL ||
407 imp->imp_state == LUSTRE_IMP_IDLE);
410 int osc_io_iter_init(const struct lu_env *env, const struct cl_io_slice *ios)
412 struct osc_object *osc = cl2osc(ios->cis_obj);
413 struct obd_import *imp = osc_cli(osc)->cl_import;
414 struct osc_io *oio = osc_env_io(env);
419 spin_lock(&imp->imp_lock);
421 * check whether this OSC device is available for non-delay read,
422 * fast switching mirror if we haven't tried all mirrors.
424 if (ios->cis_io->ci_type == CIT_READ && ios->cis_io->ci_ndelay &&
425 !ios->cis_io->ci_tried_all_mirrors && osc_import_not_healthy(imp)) {
427 } else if (likely(!imp->imp_invalid)) {
428 atomic_inc(&osc->oo_nr_ios);
429 oio->oi_is_active = 1;
432 spin_unlock(&imp->imp_lock);
434 if (capable(CAP_SYS_RESOURCE))
435 oio->oi_cap_sys_resource = 1;
439 EXPORT_SYMBOL(osc_io_iter_init);
441 void osc_io_iter_fini(const struct lu_env *env,
442 const struct cl_io_slice *ios)
444 struct osc_io *oio = osc_env_io(env);
446 if (oio->oi_is_active) {
447 struct osc_object *osc = cl2osc(ios->cis_obj);
449 oio->oi_is_active = 0;
450 LASSERT(atomic_read(&osc->oo_nr_ios) > 0);
451 if (atomic_dec_and_test(&osc->oo_nr_ios))
452 wake_up(&osc->oo_io_waitq);
455 EXPORT_SYMBOL(osc_io_iter_fini);
457 void osc_io_rw_iter_fini(const struct lu_env *env,
458 const struct cl_io_slice *ios)
460 struct osc_io *oio = osc_env_io(env);
461 struct osc_object *osc = cl2osc(ios->cis_obj);
463 if (oio->oi_lru_reserved > 0) {
464 osc_lru_unreserve(osc_cli(osc), oio->oi_lru_reserved);
465 oio->oi_lru_reserved = 0;
467 oio->oi_write_osclock = NULL;
469 osc_io_iter_fini(env, ios);
471 EXPORT_SYMBOL(osc_io_rw_iter_fini);
473 int osc_io_fault_start(const struct lu_env *env, const struct cl_io_slice *ios)
476 struct cl_fault_io *fio;
480 fio = &io->u.ci_fault;
481 CDEBUG(D_INFO, "%lu %d %zu\n",
482 fio->ft_index, fio->ft_writable, fio->ft_nob);
484 * If mapping is writeable, adjust kms to cover this page,
485 * but do not extend kms beyond actual file size.
488 if (fio->ft_writable)
489 osc_page_touch_at(env, ios->cis_obj,
490 fio->ft_index, fio->ft_nob);
493 EXPORT_SYMBOL(osc_io_fault_start);
496 static int osc_async_upcall(void *a, int rc)
498 struct osc_async_cbargs *args = a;
501 complete(&args->opc_sync);
506 * Checks that there are no pages being written in the extent being truncated.
508 static bool trunc_check_cb(const struct lu_env *env, struct cl_io *io,
509 void **pvec, int count, void *cbdata)
513 for (i = 0; i < count; i++) {
514 struct osc_page *ops = pvec[i];
515 struct cl_page *page = ops->ops_cl.cpl_page;
516 struct osc_async_page *oap;
517 __u64 start = *(__u64 *)cbdata;
520 if (oap->oap_cmd & OBD_BRW_WRITE &&
521 !list_empty(&oap->oap_pending_item))
522 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %llu/%s.\n",
523 start, current->comm);
525 if (PageLocked(page->cp_vmpage))
526 CDEBUG(D_CACHE, "page %p index %lu locked for %d.\n",
528 oap->oap_cmd & OBD_BRW_RWMASK);
533 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
534 struct osc_io *oio, __u64 size)
536 struct cl_object *clob;
540 clob = oio->oi_cl.cis_obj;
541 start = cl_index(clob, size);
542 partial = cl_offset(clob, start) < size;
545 * Complain if there are pages in the truncated region.
547 osc_page_gang_lookup(env, io, cl2osc(clob),
548 start + partial, CL_PAGE_EOF,
549 trunc_check_cb, (void *)&size);
553 * Flush affected pages prior punch.
554 * We shouldn't discard them locally first because that could be data loss
555 * if server doesn't support fallocate punch, we also need these data to be
556 * flushed first to prevent re-ordering with the punch
558 int osc_punch_start(const struct lu_env *env, struct cl_io *io,
559 struct cl_object *obj)
561 struct osc_object *osc = cl2osc(obj);
562 pgoff_t pg_start = cl_index(obj, io->u.ci_setattr.sa_falloc_offset);
563 pgoff_t pg_end = cl_index(obj, io->u.ci_setattr.sa_falloc_end - 1);
567 rc = osc_cache_writeback_range(env, osc, pg_start, pg_end, 1, 0);
571 osc_page_gang_lookup(env, io, osc, pg_start, pg_end, osc_discard_cb,
575 EXPORT_SYMBOL(osc_punch_start);
577 static int osc_io_setattr_start(const struct lu_env *env,
578 const struct cl_io_slice *slice)
580 struct cl_io *io = slice->cis_io;
581 struct osc_io *oio = cl2osc_io(env, slice);
582 struct cl_object *obj = slice->cis_obj;
583 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
584 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
585 struct obdo *oa = &oio->oi_oa;
586 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
587 unsigned int ia_avalid = io->u.ci_setattr.sa_avalid;
588 enum op_xvalid ia_xvalid = io->u.ci_setattr.sa_xvalid;
590 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
591 bool io_is_falloc = cl_io_is_fallocate(io);
594 /* truncate cache dirty pages first */
595 if (cl_io_is_trunc(io))
596 result = osc_cache_truncate_start(env, cl2osc(obj), size,
598 /* flush local pages prior punching them on server */
600 io->u.ci_setattr.sa_falloc_mode & FALLOC_FL_PUNCH_HOLE)
601 result = osc_punch_start(env, io, obj);
603 if (result == 0 && oio->oi_lockless == 0) {
604 cl_object_attr_lock(obj);
605 result = cl_object_attr_get(env, obj, attr);
607 struct ost_lvb *lvb = &io->u.ci_setattr.sa_attr;
608 unsigned int cl_valid = 0;
610 if (ia_avalid & ATTR_SIZE) {
611 attr->cat_size = size;
612 attr->cat_kms = size;
613 cl_valid = (CAT_SIZE | CAT_KMS);
615 if (ia_avalid & ATTR_MTIME_SET) {
616 attr->cat_mtime = lvb->lvb_mtime;
617 cl_valid |= CAT_MTIME;
619 if (ia_avalid & ATTR_ATIME_SET) {
620 attr->cat_atime = lvb->lvb_atime;
621 cl_valid |= CAT_ATIME;
623 if (ia_xvalid & OP_XVALID_CTIME_SET) {
624 attr->cat_ctime = lvb->lvb_ctime;
625 cl_valid |= CAT_CTIME;
627 result = cl_object_attr_update(env, obj, attr,
630 cl_object_attr_unlock(obj);
632 memset(oa, 0, sizeof(*oa));
634 oa->o_oi = loi->loi_oi;
635 obdo_set_parent_fid(oa, io->u.ci_setattr.sa_parent_fid);
636 oa->o_stripe_idx = io->u.ci_setattr.sa_stripe_index;
637 oa->o_layout = io->u.ci_setattr.sa_layout;
638 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP |
640 if (ia_avalid & ATTR_CTIME) {
641 oa->o_valid |= OBD_MD_FLCTIME;
642 oa->o_ctime = attr->cat_ctime;
644 if (ia_avalid & ATTR_ATIME) {
645 oa->o_valid |= OBD_MD_FLATIME;
646 oa->o_atime = attr->cat_atime;
648 if (ia_avalid & ATTR_MTIME) {
649 oa->o_valid |= OBD_MD_FLMTIME;
650 oa->o_mtime = attr->cat_mtime;
653 if (ia_avalid & ATTR_SIZE || io_is_falloc) {
654 if (oio->oi_lockless) {
655 oa->o_flags = OBD_FL_SRVLOCK;
656 oa->o_valid |= OBD_MD_FLFLAGS;
659 if (io->ci_layout_version > 0) {
660 /* verify layout version */
661 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
662 oa->o_layout_version = io->ci_layout_version;
665 LASSERT(oio->oi_lockless == 0);
668 if (ia_xvalid & OP_XVALID_FLAGS) {
669 oa->o_flags = io->u.ci_setattr.sa_attr_flags;
670 oa->o_valid |= OBD_MD_FLFLAGS;
673 init_completion(&cbargs->opc_sync);
676 int falloc_mode = io->u.ci_setattr.sa_falloc_mode;
678 oa->o_size = io->u.ci_setattr.sa_falloc_offset;
679 oa->o_blocks = io->u.ci_setattr.sa_falloc_end;
680 oa->o_uid = io->u.ci_setattr.sa_falloc_uid;
681 oa->o_gid = io->u.ci_setattr.sa_falloc_gid;
682 oa->o_projid = io->u.ci_setattr.sa_falloc_projid;
683 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
684 OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID;
687 "size %llu blocks %llu uid %u gid %u prjid %u\n",
688 oa->o_size, oa->o_blocks, oa->o_uid, oa->o_gid,
690 result = osc_fallocate_base(osc_export(cl2osc(obj)),
691 oa, osc_async_upcall,
692 cbargs, falloc_mode);
693 } else if (ia_avalid & ATTR_SIZE) {
695 oa->o_blocks = OBD_OBJECT_EOF;
696 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
697 result = osc_punch_send(osc_export(cl2osc(obj)),
698 oa, osc_async_upcall, cbargs);
700 result = osc_setattr_async(osc_export(cl2osc(obj)),
701 oa, osc_async_upcall,
702 cbargs, PTLRPCD_SET);
704 cbargs->opc_rpc_sent = result == 0;
710 void osc_io_setattr_end(const struct lu_env *env,
711 const struct cl_io_slice *slice)
713 struct cl_io *io = slice->cis_io;
714 struct osc_io *oio = cl2osc_io(env, slice);
715 struct cl_object *obj = slice->cis_obj;
716 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
717 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
718 struct obdo *oa = &oio->oi_oa;
719 unsigned int cl_valid = 0;
722 if (cbargs->opc_rpc_sent) {
723 wait_for_completion(&cbargs->opc_sync);
724 result = io->ci_result = cbargs->opc_rc;
727 if (cl_io_is_trunc(io)) {
728 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
731 cl_object_attr_lock(obj);
732 if (oa->o_valid & OBD_MD_FLBLOCKS) {
733 attr->cat_blocks = oa->o_blocks;
734 cl_valid |= CAT_BLOCKS;
737 cl_object_attr_update(env, obj, attr, cl_valid);
738 cl_object_attr_unlock(obj);
740 osc_trunc_check(env, io, oio, size);
741 osc_cache_truncate_end(env, oio->oi_trunc);
742 oio->oi_trunc = NULL;
745 if (cl_io_is_fallocate(io)) {
747 cl_object_attr_lock(obj);
749 if (oa->o_valid & OBD_MD_FLBLOCKS) {
750 attr->cat_blocks = oa->o_blocks;
751 cl_valid |= CAT_BLOCKS;
754 cl_object_attr_update(env, obj, attr, cl_valid);
755 cl_object_attr_unlock(obj);
759 EXPORT_SYMBOL(osc_io_setattr_end);
761 struct osc_data_version_args {
762 struct osc_io *dva_oio;
766 osc_data_version_interpret(const struct lu_env *env, struct ptlrpc_request *req,
769 struct osc_data_version_args *dva = args;
770 struct osc_io *oio = dva->dva_oio;
771 const struct ost_body *body;
777 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
779 GOTO(out, rc = -EPROTO);
781 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, &oio->oi_oa,
785 oio->oi_cbarg.opc_rc = rc;
786 complete(&oio->oi_cbarg.opc_sync);
791 static int osc_io_data_version_start(const struct lu_env *env,
792 const struct cl_io_slice *slice)
794 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
795 struct osc_io *oio = cl2osc_io(env, slice);
796 struct obdo *oa = &oio->oi_oa;
797 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
798 struct osc_object *obj = cl2osc(slice->cis_obj);
799 struct lov_oinfo *loi = obj->oo_oinfo;
800 struct obd_export *exp = osc_export(obj);
801 struct ptlrpc_request *req;
802 struct ost_body *body;
803 struct osc_data_version_args *dva;
807 memset(oa, 0, sizeof(*oa));
808 oa->o_oi = loi->loi_oi;
809 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
811 if (dv->dv_flags & (LL_DV_RD_FLUSH | LL_DV_WR_FLUSH)) {
812 oa->o_valid |= OBD_MD_FLFLAGS;
813 oa->o_flags |= OBD_FL_SRVLOCK;
814 if (dv->dv_flags & LL_DV_WR_FLUSH)
815 oa->o_flags |= OBD_FL_FLUSH;
818 init_completion(&cbargs->opc_sync);
820 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
824 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
826 ptlrpc_request_free(req);
830 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
831 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
833 ptlrpc_request_set_replen(req);
834 req->rq_interpret_reply = osc_data_version_interpret;
835 dva = ptlrpc_req_async_args(dva, req);
838 ptlrpcd_add_req(req);
843 static void osc_io_data_version_end(const struct lu_env *env,
844 const struct cl_io_slice *slice)
846 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
847 struct osc_io *oio = cl2osc_io(env, slice);
848 struct cl_object *obj = slice->cis_obj;
849 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
850 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
851 struct obdo *oa = &oio->oi_oa;
852 unsigned int cl_valid = 0;
855 wait_for_completion(&cbargs->opc_sync);
857 if (cbargs->opc_rc != 0) {
858 slice->cis_io->ci_result = cbargs->opc_rc;
860 slice->cis_io->ci_result = 0;
862 (OBD_MD_LAYOUT_VERSION | OBD_MD_FLDATAVERSION)))
863 slice->cis_io->ci_result = -ENOTSUPP;
865 if (oa->o_valid & OBD_MD_LAYOUT_VERSION)
866 dv->dv_layout_version = oa->o_layout_version;
867 if (oa->o_valid & OBD_MD_FLDATAVERSION)
868 dv->dv_data_version = oa->o_data_version;
870 if (dv->dv_flags & LL_DV_SZ_UPDATE) {
871 if (oa->o_valid & OBD_MD_FLSIZE) {
872 attr->cat_size = oa->o_size;
873 cl_valid |= CAT_SIZE;
876 if (oa->o_valid & OBD_MD_FLBLOCKS) {
877 attr->cat_blocks = oa->o_blocks;
878 cl_valid |= CAT_BLOCKS;
881 cl_object_attr_lock(obj);
882 cl_object_attr_update(env, obj, attr, cl_valid);
883 cl_object_attr_unlock(obj);
890 int osc_io_read_start(const struct lu_env *env,
891 const struct cl_io_slice *slice)
893 struct cl_object *obj = slice->cis_obj;
894 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
898 if (!slice->cis_io->ci_noatime) {
899 cl_object_attr_lock(obj);
900 attr->cat_atime = ktime_get_real_seconds();
901 rc = cl_object_attr_update(env, obj, attr, CAT_ATIME);
902 cl_object_attr_unlock(obj);
907 EXPORT_SYMBOL(osc_io_read_start);
909 int osc_io_write_start(const struct lu_env *env,
910 const struct cl_io_slice *slice)
912 struct cl_object *obj = slice->cis_obj;
913 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
917 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_SETTIME, 1);
918 cl_object_attr_lock(obj);
919 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
920 rc = cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
921 cl_object_attr_unlock(obj);
925 EXPORT_SYMBOL(osc_io_write_start);
927 int osc_fsync_ost(const struct lu_env *env, struct osc_object *obj,
928 struct cl_fsync_io *fio)
930 struct osc_io *oio = osc_env_io(env);
931 struct obdo *oa = &oio->oi_oa;
932 struct lov_oinfo *loi = obj->oo_oinfo;
933 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
937 memset(oa, 0, sizeof(*oa));
938 oa->o_oi = loi->loi_oi;
939 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
941 /* reload size abd blocks for start and end of sync range */
942 oa->o_size = fio->fi_start;
943 oa->o_blocks = fio->fi_end;
944 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
946 obdo_set_parent_fid(oa, fio->fi_fid);
948 init_completion(&cbargs->opc_sync);
950 rc = osc_sync_base(obj, oa, osc_async_upcall, cbargs, PTLRPCD_SET);
953 EXPORT_SYMBOL(osc_fsync_ost);
955 int osc_io_fsync_start(const struct lu_env *env,
956 const struct cl_io_slice *slice)
958 struct cl_io *io = slice->cis_io;
959 struct cl_fsync_io *fio = &io->u.ci_fsync;
960 struct cl_object *obj = slice->cis_obj;
961 struct osc_object *osc = cl2osc(obj);
962 pgoff_t start = cl_index(obj, fio->fi_start);
963 pgoff_t end = cl_index(obj, fio->fi_end);
967 if (fio->fi_end == OBD_OBJECT_EOF)
970 result = osc_cache_writeback_range(env, osc, start, end, 0,
971 fio->fi_mode == CL_FSYNC_DISCARD);
973 fio->fi_nr_written += result;
976 if (fio->fi_mode == CL_FSYNC_ALL) {
979 /* we have to wait for writeback to finish before we can
980 * send OST_SYNC RPC. This is bad because it causes extents
981 * to be written osc by osc. However, we usually start
982 * writeback before CL_FSYNC_ALL so this won't have any real
984 rc = osc_cache_wait_range(env, osc, start, end);
987 rc = osc_fsync_ost(env, osc, fio);
995 void osc_io_fsync_end(const struct lu_env *env,
996 const struct cl_io_slice *slice)
998 struct cl_fsync_io *fio = &slice->cis_io->u.ci_fsync;
999 struct cl_object *obj = slice->cis_obj;
1000 pgoff_t start = cl_index(obj, fio->fi_start);
1001 pgoff_t end = cl_index(obj, fio->fi_end);
1004 if (fio->fi_mode == CL_FSYNC_LOCAL) {
1005 result = osc_cache_wait_range(env, cl2osc(obj), start, end);
1006 } else if (fio->fi_mode == CL_FSYNC_ALL) {
1007 struct osc_io *oio = cl2osc_io(env, slice);
1008 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1010 wait_for_completion(&cbargs->opc_sync);
1012 result = cbargs->opc_rc;
1014 slice->cis_io->ci_result = result;
1016 EXPORT_SYMBOL(osc_io_fsync_end);
1018 static int osc_io_ladvise_start(const struct lu_env *env,
1019 const struct cl_io_slice *slice)
1022 struct cl_io *io = slice->cis_io;
1023 struct osc_io *oio = cl2osc_io(env, slice);
1024 struct cl_object *obj = slice->cis_obj;
1025 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1026 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
1027 struct obdo *oa = &oio->oi_oa;
1028 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1029 struct lu_ladvise *ladvise;
1030 struct ladvise_hdr *ladvise_hdr;
1035 /* TODO: add multiple ladvise support in CLIO */
1036 buf_size = offsetof(typeof(*ladvise_hdr), lah_advise[num_advise]);
1037 if (osc_env_info(env)->oti_ladvise_buf.lb_len < buf_size)
1038 lu_buf_realloc(&osc_env_info(env)->oti_ladvise_buf, buf_size);
1040 ladvise_hdr = osc_env_info(env)->oti_ladvise_buf.lb_buf;
1041 if (ladvise_hdr == NULL)
1044 memset(ladvise_hdr, 0, buf_size);
1045 ladvise_hdr->lah_magic = LADVISE_MAGIC;
1046 ladvise_hdr->lah_count = num_advise;
1047 ladvise_hdr->lah_flags = lio->li_flags;
1049 memset(oa, 0, sizeof(*oa));
1050 oa->o_oi = loi->loi_oi;
1051 oa->o_valid = OBD_MD_FLID;
1052 obdo_set_parent_fid(oa, lio->li_fid);
1054 ladvise = ladvise_hdr->lah_advise;
1055 ladvise->lla_start = lio->li_start;
1056 ladvise->lla_end = lio->li_end;
1057 ladvise->lla_advice = lio->li_advice;
1059 if (lio->li_flags & LF_ASYNC) {
1060 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1061 ladvise_hdr, NULL, NULL, NULL);
1063 init_completion(&cbargs->opc_sync);
1064 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1065 ladvise_hdr, osc_async_upcall,
1066 cbargs, PTLRPCD_SET);
1067 cbargs->opc_rpc_sent = result == 0;
1072 static void osc_io_ladvise_end(const struct lu_env *env,
1073 const struct cl_io_slice *slice)
1075 struct cl_io *io = slice->cis_io;
1076 struct osc_io *oio = cl2osc_io(env, slice);
1077 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1079 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
1081 if ((!(lio->li_flags & LF_ASYNC)) && cbargs->opc_rpc_sent) {
1082 wait_for_completion(&cbargs->opc_sync);
1083 result = cbargs->opc_rc;
1085 slice->cis_io->ci_result = result;
1088 void osc_io_end(const struct lu_env *env, const struct cl_io_slice *slice)
1090 struct osc_io *oio = cl2osc_io(env, slice);
1092 if (oio->oi_active) {
1093 osc_extent_release(env, oio->oi_active);
1094 oio->oi_active = NULL;
1097 EXPORT_SYMBOL(osc_io_end);
1099 struct osc_lseek_args {
1100 struct osc_io *lsa_oio;
1103 static int osc_lseek_interpret(const struct lu_env *env,
1104 struct ptlrpc_request *req,
1107 struct ost_body *reply;
1108 struct osc_lseek_args *lsa = arg;
1109 struct osc_io *oio = lsa->lsa_oio;
1110 struct cl_io *io = oio->oi_cl.cis_io;
1111 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1118 reply = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1120 GOTO(out, rc = -EPROTO);
1122 lsio->ls_result = reply->oa.o_size;
1124 osc_async_upcall(&oio->oi_cbarg, rc);
1128 int osc_io_lseek_start(const struct lu_env *env,
1129 const struct cl_io_slice *slice)
1131 struct cl_io *io = slice->cis_io;
1132 struct osc_io *oio = cl2osc_io(env, slice);
1133 struct cl_object *obj = slice->cis_obj;
1134 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1135 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1136 struct obdo *oa = &oio->oi_oa;
1137 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1138 struct obd_export *exp = osc_export(cl2osc(obj));
1139 struct ptlrpc_request *req;
1140 struct ost_body *body;
1141 struct osc_lseek_args *lsa;
1146 /* No negative values at this point */
1147 LASSERT(lsio->ls_start >= 0);
1148 LASSERT(lsio->ls_whence == SEEK_HOLE || lsio->ls_whence == SEEK_DATA);
1150 /* with IO lock taken we have object size in LVB and can check
1151 * boundaries prior sending LSEEK RPC
1153 if (lsio->ls_start >= loi->loi_lvb.lvb_size) {
1154 /* consider area beyond end of object as hole */
1155 if (lsio->ls_whence == SEEK_HOLE)
1156 lsio->ls_result = lsio->ls_start;
1158 lsio->ls_result = -ENXIO;
1162 /* if LSEEK RPC is not supported by server, consider whole stripe
1163 * object is data with hole after end of object
1165 if (!exp_connect_lseek(exp)) {
1166 if (lsio->ls_whence == SEEK_HOLE)
1167 lsio->ls_result = loi->loi_lvb.lvb_size;
1169 lsio->ls_result = lsio->ls_start;
1173 memset(oa, 0, sizeof(*oa));
1174 oa->o_oi = loi->loi_oi;
1175 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1176 oa->o_size = lsio->ls_start;
1177 oa->o_mode = lsio->ls_whence;
1178 if (oio->oi_lockless) {
1179 oa->o_flags = OBD_FL_SRVLOCK;
1180 oa->o_valid |= OBD_MD_FLFLAGS;
1183 init_completion(&cbargs->opc_sync);
1184 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SEEK);
1188 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SEEK);
1190 ptlrpc_request_free(req);
1194 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1195 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1196 ptlrpc_request_set_replen(req);
1197 req->rq_interpret_reply = osc_lseek_interpret;
1198 lsa = ptlrpc_req_async_args(lsa, req);
1201 ptlrpcd_add_req(req);
1202 cbargs->opc_rpc_sent = 1;
1206 EXPORT_SYMBOL(osc_io_lseek_start);
1208 void osc_io_lseek_end(const struct lu_env *env,
1209 const struct cl_io_slice *slice)
1211 struct osc_io *oio = cl2osc_io(env, slice);
1212 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1215 if (cbargs->opc_rpc_sent) {
1216 wait_for_completion(&cbargs->opc_sync);
1217 rc = cbargs->opc_rc;
1219 slice->cis_io->ci_result = rc;
1221 EXPORT_SYMBOL(osc_io_lseek_end);
1223 int osc_io_lru_reserve(const struct lu_env *env,
1224 const struct cl_io_slice *ios,
1225 loff_t pos, size_t bytes)
1227 struct osc_object *osc = cl2osc(ios->cis_obj);
1228 struct osc_io *oio = osc_env_io(env);
1229 unsigned long npages = 0;
1234 page_offset = pos & ~PAGE_MASK;
1237 if (bytes > PAGE_SIZE - page_offset)
1238 bytes -= (PAGE_SIZE - page_offset);
1242 npages += (bytes + PAGE_SIZE - 1) >> PAGE_SHIFT;
1243 oio->oi_lru_reserved = osc_lru_reserve(osc_cli(osc), npages);
1247 EXPORT_SYMBOL(osc_io_lru_reserve);
1249 static const struct cl_io_operations osc_io_ops = {
1252 .cio_iter_init = osc_io_iter_init,
1253 .cio_iter_fini = osc_io_rw_iter_fini,
1254 .cio_start = osc_io_read_start,
1255 .cio_fini = osc_io_fini
1258 .cio_iter_init = osc_io_iter_init,
1259 .cio_iter_fini = osc_io_rw_iter_fini,
1260 .cio_start = osc_io_write_start,
1261 .cio_end = osc_io_end,
1262 .cio_fini = osc_io_fini
1265 .cio_iter_init = osc_io_iter_init,
1266 .cio_iter_fini = osc_io_iter_fini,
1267 .cio_start = osc_io_setattr_start,
1268 .cio_end = osc_io_setattr_end
1270 [CIT_DATA_VERSION] = {
1271 .cio_start = osc_io_data_version_start,
1272 .cio_end = osc_io_data_version_end,
1275 .cio_iter_init = osc_io_iter_init,
1276 .cio_iter_fini = osc_io_iter_fini,
1277 .cio_start = osc_io_fault_start,
1278 .cio_end = osc_io_end,
1279 .cio_fini = osc_io_fini
1282 .cio_start = osc_io_fsync_start,
1283 .cio_end = osc_io_fsync_end,
1284 .cio_fini = osc_io_fini
1287 .cio_start = osc_io_ladvise_start,
1288 .cio_end = osc_io_ladvise_end,
1289 .cio_fini = osc_io_fini
1292 .cio_start = osc_io_lseek_start,
1293 .cio_end = osc_io_lseek_end,
1294 .cio_fini = osc_io_fini
1297 .cio_fini = osc_io_fini
1300 .cio_read_ahead = osc_io_read_ahead,
1301 .cio_lru_reserve = osc_io_lru_reserve,
1302 .cio_submit = osc_io_submit,
1303 .cio_commit_async = osc_io_commit_async,
1304 .cio_extent_release = osc_io_extent_release
1307 /*****************************************************************************
1309 * Transfer operations.
1313 int osc_io_init(const struct lu_env *env,
1314 struct cl_object *obj, struct cl_io *io)
1316 struct osc_io *oio = osc_env_io(env);
1318 CL_IO_SLICE_CLEAN(oio, oi_cl);
1319 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);