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 ktime_t submit_time = ktime_get();
137 bool sync_queue = false;
139 LASSERT(qin->pl_nr > 0);
141 CDEBUG(D_CACHE|D_READA, "%d %d\n", qin->pl_nr, crt);
143 osc = cl2osc(ios->cis_obj);
145 max_pages = cli->cl_max_pages_per_rpc;
146 ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
149 brw_flags = osc_io_srvlock(cl2osc_io(env, ios)) ? OBD_BRW_SRVLOCK : 0;
150 brw_flags |= crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
151 if (crt == CRT_READ && ios->cis_io->ci_ndelay)
152 brw_flags |= OBD_BRW_NDELAY;
154 page = cl_page_list_first(qin);
155 if (page->cp_type == CPT_TRANSIENT)
156 brw_flags |= OBD_BRW_NOCACHE;
157 if (lnet_is_rdma_only_page(page->cp_vmpage))
158 brw_flags |= OBD_BRW_RDMA_ONLY;
161 * NOTE: here @page is a top-level page. This is done to avoid
162 * creation of sub-page-list.
164 cl_page_list_for_each_safe(page, tmp, qin) {
165 struct osc_async_page *oap;
171 opg = osc_cl_page_osc(page, osc);
173 LASSERT(osc == oap->oap_obj);
175 if (!list_empty(&oap->oap_pending_item) ||
176 !list_empty(&oap->oap_rpc_item)) {
177 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
183 result = cl_page_prep(env, io, page, crt);
186 if (result != -EALREADY)
189 * Handle -EALREADY error: for read case, the page is
190 * already in UPTODATE state; for write, the page
197 if (page->cp_type != CPT_TRANSIENT) {
198 spin_lock(&oap->oap_lock);
199 oap->oap_async_flags = ASYNC_URGENT|ASYNC_READY;
200 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
201 spin_unlock(&oap->oap_lock);
204 osc_page_submit(env, opg, crt, brw_flags, submit_time);
205 list_add_tail(&oap->oap_pending_item, &list);
207 if (page->cp_sync_io != NULL)
208 cl_page_list_move(qout, qin, page);
210 cl_page_list_del(env, qin, page);
213 if (queued == max_pages) {
215 } else if (crt == CRT_WRITE) {
217 unsigned int next_chunks;
219 chunks = (queued + ppc - 1) >> ppc_bits;
220 /* chunk number if add another page */
221 next_chunks = (queued + ppc) >> ppc_bits;
223 /* next page will excceed write chunk limit */
224 if (chunks == osc_max_write_chunks(cli) &&
225 next_chunks > chunks)
230 result = osc_queue_sync_pages(env, io, osc, &list,
240 result = osc_queue_sync_pages(env, io, osc, &list, brw_flags);
242 /* Update c/mtime for sync write. LU-7310 */
243 if (crt == CRT_WRITE && qout->pl_nr > 0 && result == 0) {
244 struct cl_object *obj = ios->cis_obj;
245 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
247 cl_object_attr_lock(obj);
248 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
249 cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
250 cl_object_attr_unlock(obj);
253 CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
254 return qout->pl_nr > 0 ? 0 : result;
256 EXPORT_SYMBOL(osc_io_submit);
259 * This is called to update the attributes when modifying a specific page,
260 * both when making new pages and when doing updates to existing cached pages.
262 * Expand stripe KMS if necessary.
264 void osc_page_touch_at(const struct lu_env *env, struct cl_object *obj,
265 pgoff_t idx, size_t to)
267 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
268 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
274 /* offset within stripe */
275 kms = cl_offset(obj, idx) + to;
277 cl_object_attr_lock(obj);
278 CDEBUG(D_INODE, "stripe KMS %sincreasing %llu->%llu %llu\n",
279 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
280 loi->loi_lvb.lvb_size);
282 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
283 valid = CAT_MTIME | CAT_CTIME;
284 if (kms > loi->loi_kms) {
288 if (kms > loi->loi_lvb.lvb_size) {
289 attr->cat_size = kms;
292 cl_object_attr_update(env, obj, attr, valid);
293 cl_object_attr_unlock(obj);
298 int osc_io_commit_async(const struct lu_env *env,
299 const struct cl_io_slice *ios,
300 struct cl_page_list *qin, int from, int to,
303 struct cl_io *io = ios->cis_io;
304 struct osc_io *oio = cl2osc_io(env, ios);
305 struct osc_object *osc = cl2osc(ios->cis_obj);
306 struct cl_page *page;
307 struct cl_page *last_page;
308 struct osc_page *opg;
309 struct pagevec *pvec = &osc_env_info(env)->oti_pagevec;
313 LASSERT(qin->pl_nr > 0);
315 /* Handle partial page cases */
316 last_page = cl_page_list_last(qin);
317 if (oio->oi_lockless) {
318 page = cl_page_list_first(qin);
319 if (page == last_page) {
320 cl_page_clip(env, page, from, to);
323 cl_page_clip(env, page, from, PAGE_SIZE);
325 cl_page_clip(env, last_page, 0, to);
329 ll_pagevec_init(pvec, 0);
331 while (qin->pl_nr > 0) {
332 struct osc_async_page *oap;
334 page = cl_page_list_first(qin);
335 opg = osc_cl_page_osc(page, osc);
338 LASSERTF(osc == oap->oap_obj,
339 "obj mismatch: %p / %p\n", osc, oap->oap_obj);
341 if (!list_empty(&oap->oap_rpc_item)) {
342 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
348 /* The page may be already in dirty cache. */
349 if (list_empty(&oap->oap_pending_item)) {
350 result = osc_page_cache_add(env, opg, io, cb);
355 osc_page_touch_at(env, osc2cl(osc), osc_index(opg),
356 page == last_page ? to : PAGE_SIZE);
358 cl_page_list_del(env, qin, page);
360 /* if there are no more slots, do the callback & reinit */
361 if (pagevec_add(pvec, page->cp_vmpage) == 0) {
362 (*cb)(env, io, pvec);
363 pagevec_reinit(pvec);
366 /* The shrink interval is in seconds, so we can update it once per
367 * write, rather than once per page.
369 osc_update_next_shrink(osc_cli(osc));
372 /* Clean up any partially full pagevecs */
373 if (pagevec_count(pvec) != 0)
374 (*cb)(env, io, pvec);
376 /* Can't access these pages any more. Page can be in transfer and
377 * complete at any time. */
379 /* for sync write, kernel will wait for this page to be flushed before
380 * osc_io_end() is called, so release it earlier.
381 * for mkwrite(), it's known there is no further pages. */
382 if (cl_io_is_sync_write(io) && oio->oi_active != NULL) {
383 osc_extent_release(env, oio->oi_active);
384 oio->oi_active = NULL;
387 CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, result);
390 EXPORT_SYMBOL(osc_io_commit_async);
392 void osc_io_extent_release(const struct lu_env *env,
393 const struct cl_io_slice *ios)
395 struct osc_io *oio = cl2osc_io(env, ios);
397 if (oio->oi_active != NULL) {
398 osc_extent_release(env, oio->oi_active);
399 oio->oi_active = NULL;
402 EXPORT_SYMBOL(osc_io_extent_release);
404 static bool osc_import_not_healthy(struct obd_import *imp)
406 return imp->imp_invalid || imp->imp_deactive ||
407 !(imp->imp_state == LUSTRE_IMP_FULL ||
408 imp->imp_state == LUSTRE_IMP_IDLE);
411 int osc_io_iter_init(const struct lu_env *env, const struct cl_io_slice *ios)
413 struct osc_object *osc = cl2osc(ios->cis_obj);
414 struct obd_import *imp = osc_cli(osc)->cl_import;
415 struct osc_io *oio = osc_env_io(env);
420 spin_lock(&imp->imp_lock);
422 * check whether this OSC device is available for non-delay read,
423 * fast switching mirror if we haven't tried all mirrors.
425 if (ios->cis_io->ci_type == CIT_READ && ios->cis_io->ci_ndelay &&
426 !ios->cis_io->ci_tried_all_mirrors && osc_import_not_healthy(imp)) {
428 } else if (likely(!imp->imp_invalid)) {
429 atomic_inc(&osc->oo_nr_ios);
430 oio->oi_is_active = 1;
433 spin_unlock(&imp->imp_lock);
435 if (capable(CAP_SYS_RESOURCE))
436 oio->oi_cap_sys_resource = 1;
440 EXPORT_SYMBOL(osc_io_iter_init);
442 void osc_io_iter_fini(const struct lu_env *env,
443 const struct cl_io_slice *ios)
445 struct osc_io *oio = osc_env_io(env);
447 if (oio->oi_is_active) {
448 struct osc_object *osc = cl2osc(ios->cis_obj);
450 oio->oi_is_active = 0;
451 LASSERT(atomic_read(&osc->oo_nr_ios) > 0);
452 if (atomic_dec_and_test(&osc->oo_nr_ios))
453 wake_up(&osc->oo_io_waitq);
456 EXPORT_SYMBOL(osc_io_iter_fini);
458 void osc_io_rw_iter_fini(const struct lu_env *env,
459 const struct cl_io_slice *ios)
461 struct osc_io *oio = osc_env_io(env);
462 struct osc_object *osc = cl2osc(ios->cis_obj);
464 if (oio->oi_lru_reserved > 0) {
465 osc_lru_unreserve(osc_cli(osc), oio->oi_lru_reserved);
466 oio->oi_lru_reserved = 0;
468 oio->oi_write_osclock = NULL;
470 osc_io_iter_fini(env, ios);
472 EXPORT_SYMBOL(osc_io_rw_iter_fini);
474 int osc_io_fault_start(const struct lu_env *env, const struct cl_io_slice *ios)
477 struct cl_fault_io *fio;
481 fio = &io->u.ci_fault;
482 CDEBUG(D_INFO, "%lu %d %zu\n",
483 fio->ft_index, fio->ft_writable, fio->ft_nob);
485 * If mapping is writeable, adjust kms to cover this page,
486 * but do not extend kms beyond actual file size.
489 if (fio->ft_writable)
490 osc_page_touch_at(env, ios->cis_obj,
491 fio->ft_index, fio->ft_nob);
494 EXPORT_SYMBOL(osc_io_fault_start);
497 static int osc_async_upcall(void *a, int rc)
499 struct osc_async_cbargs *args = a;
502 complete(&args->opc_sync);
507 * Checks that there are no pages being written in the extent being truncated.
509 static bool trunc_check_cb(const struct lu_env *env, struct cl_io *io,
510 void **pvec, int count, void *cbdata)
514 for (i = 0; i < count; i++) {
515 struct osc_page *ops = pvec[i];
516 struct cl_page *page = ops->ops_cl.cpl_page;
517 struct osc_async_page *oap;
518 __u64 start = *(__u64 *)cbdata;
521 if (oap->oap_cmd & OBD_BRW_WRITE &&
522 !list_empty(&oap->oap_pending_item))
523 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %llu/%s.\n",
524 start, current->comm);
526 if (PageLocked(page->cp_vmpage))
527 CDEBUG(D_CACHE, "page %p index %lu locked for %d.\n",
529 oap->oap_cmd & OBD_BRW_RWMASK);
534 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
535 struct osc_io *oio, __u64 size)
537 struct cl_object *clob;
541 clob = oio->oi_cl.cis_obj;
542 start = cl_index(clob, size);
543 partial = cl_offset(clob, start) < size;
546 * Complain if there are pages in the truncated region.
548 osc_page_gang_lookup(env, io, cl2osc(clob),
549 start + partial, CL_PAGE_EOF,
550 trunc_check_cb, (void *)&size);
554 * Flush affected pages prior punch.
555 * We shouldn't discard them locally first because that could be data loss
556 * if server doesn't support fallocate punch, we also need these data to be
557 * flushed first to prevent re-ordering with the punch
559 int osc_punch_start(const struct lu_env *env, struct cl_io *io,
560 struct cl_object *obj)
562 struct osc_object *osc = cl2osc(obj);
563 pgoff_t pg_start = cl_index(obj, io->u.ci_setattr.sa_falloc_offset);
564 pgoff_t pg_end = cl_index(obj, io->u.ci_setattr.sa_falloc_end - 1);
568 rc = osc_cache_writeback_range(env, osc, pg_start, pg_end, 1, 0);
572 osc_page_gang_lookup(env, io, osc, pg_start, pg_end, osc_discard_cb,
576 EXPORT_SYMBOL(osc_punch_start);
578 static int osc_io_setattr_start(const struct lu_env *env,
579 const struct cl_io_slice *slice)
581 struct cl_io *io = slice->cis_io;
582 struct osc_io *oio = cl2osc_io(env, slice);
583 struct cl_object *obj = slice->cis_obj;
584 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
585 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
586 struct obdo *oa = &oio->oi_oa;
587 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
588 unsigned int ia_avalid = io->u.ci_setattr.sa_avalid;
589 enum op_xvalid ia_xvalid = io->u.ci_setattr.sa_xvalid;
591 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
592 bool io_is_falloc = cl_io_is_fallocate(io);
595 /* truncate cache dirty pages first */
596 if (cl_io_is_trunc(io))
597 result = osc_cache_truncate_start(env, cl2osc(obj), size,
599 /* flush local pages prior punching them on server */
601 io->u.ci_setattr.sa_falloc_mode & FALLOC_FL_PUNCH_HOLE)
602 result = osc_punch_start(env, io, obj);
604 if (result == 0 && oio->oi_lockless == 0) {
605 cl_object_attr_lock(obj);
606 result = cl_object_attr_get(env, obj, attr);
608 struct ost_lvb *lvb = &io->u.ci_setattr.sa_attr;
609 unsigned int cl_valid = 0;
611 if (ia_avalid & ATTR_SIZE) {
612 attr->cat_size = size;
613 attr->cat_kms = size;
614 cl_valid = (CAT_SIZE | CAT_KMS);
616 if (ia_avalid & ATTR_MTIME_SET) {
617 attr->cat_mtime = lvb->lvb_mtime;
618 cl_valid |= CAT_MTIME;
620 if (ia_avalid & ATTR_ATIME_SET) {
621 attr->cat_atime = lvb->lvb_atime;
622 cl_valid |= CAT_ATIME;
624 if (ia_xvalid & OP_XVALID_CTIME_SET) {
625 attr->cat_ctime = lvb->lvb_ctime;
626 cl_valid |= CAT_CTIME;
628 result = cl_object_attr_update(env, obj, attr,
631 cl_object_attr_unlock(obj);
633 memset(oa, 0, sizeof(*oa));
635 oa->o_oi = loi->loi_oi;
636 obdo_set_parent_fid(oa, io->u.ci_setattr.sa_parent_fid);
637 oa->o_stripe_idx = io->u.ci_setattr.sa_stripe_index;
638 oa->o_layout = io->u.ci_setattr.sa_layout;
639 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP |
641 if (ia_avalid & ATTR_CTIME) {
642 oa->o_valid |= OBD_MD_FLCTIME;
643 oa->o_ctime = attr->cat_ctime;
645 if (ia_avalid & ATTR_ATIME) {
646 oa->o_valid |= OBD_MD_FLATIME;
647 oa->o_atime = attr->cat_atime;
649 if (ia_avalid & ATTR_MTIME) {
650 oa->o_valid |= OBD_MD_FLMTIME;
651 oa->o_mtime = attr->cat_mtime;
654 if (ia_avalid & ATTR_SIZE || io_is_falloc) {
655 if (oio->oi_lockless) {
656 oa->o_flags = OBD_FL_SRVLOCK;
657 oa->o_valid |= OBD_MD_FLFLAGS;
660 if (io->ci_layout_version > 0) {
661 /* verify layout version */
662 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
663 oa->o_layout_version = io->ci_layout_version;
666 LASSERT(oio->oi_lockless == 0);
669 if (ia_xvalid & OP_XVALID_FLAGS) {
670 oa->o_flags = io->u.ci_setattr.sa_attr_flags;
671 oa->o_valid |= OBD_MD_FLFLAGS;
674 init_completion(&cbargs->opc_sync);
677 int falloc_mode = io->u.ci_setattr.sa_falloc_mode;
679 oa->o_size = io->u.ci_setattr.sa_falloc_offset;
680 oa->o_blocks = io->u.ci_setattr.sa_falloc_end;
681 oa->o_uid = io->u.ci_setattr.sa_falloc_uid;
682 oa->o_gid = io->u.ci_setattr.sa_falloc_gid;
683 oa->o_projid = io->u.ci_setattr.sa_falloc_projid;
684 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
685 OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID;
688 "size %llu blocks %llu uid %u gid %u prjid %u\n",
689 oa->o_size, oa->o_blocks, oa->o_uid, oa->o_gid,
691 result = osc_fallocate_base(osc_export(cl2osc(obj)),
692 oa, osc_async_upcall,
693 cbargs, falloc_mode);
694 } else if (ia_avalid & ATTR_SIZE) {
696 oa->o_blocks = OBD_OBJECT_EOF;
697 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
698 result = osc_punch_send(osc_export(cl2osc(obj)),
699 oa, osc_async_upcall, cbargs);
701 result = osc_setattr_async(osc_export(cl2osc(obj)),
702 oa, osc_async_upcall,
703 cbargs, PTLRPCD_SET);
705 cbargs->opc_rpc_sent = result == 0;
711 void osc_io_setattr_end(const struct lu_env *env,
712 const struct cl_io_slice *slice)
714 struct cl_io *io = slice->cis_io;
715 struct osc_io *oio = cl2osc_io(env, slice);
716 struct cl_object *obj = slice->cis_obj;
717 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
718 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
719 struct obdo *oa = &oio->oi_oa;
720 unsigned int cl_valid = 0;
723 if (cbargs->opc_rpc_sent) {
724 wait_for_completion(&cbargs->opc_sync);
725 result = io->ci_result = cbargs->opc_rc;
728 if (cl_io_is_trunc(io)) {
729 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
732 cl_object_attr_lock(obj);
733 if (oa->o_valid & OBD_MD_FLBLOCKS) {
734 attr->cat_blocks = oa->o_blocks;
735 cl_valid |= CAT_BLOCKS;
738 cl_object_attr_update(env, obj, attr, cl_valid);
739 cl_object_attr_unlock(obj);
741 osc_trunc_check(env, io, oio, size);
742 osc_cache_truncate_end(env, oio->oi_trunc);
743 oio->oi_trunc = NULL;
746 if (cl_io_is_fallocate(io)) {
748 cl_object_attr_lock(obj);
750 if (oa->o_valid & OBD_MD_FLBLOCKS) {
751 attr->cat_blocks = oa->o_blocks;
752 cl_valid |= CAT_BLOCKS;
755 cl_object_attr_update(env, obj, attr, cl_valid);
756 cl_object_attr_unlock(obj);
760 EXPORT_SYMBOL(osc_io_setattr_end);
762 struct osc_data_version_args {
763 struct osc_io *dva_oio;
767 osc_data_version_interpret(const struct lu_env *env, struct ptlrpc_request *req,
770 struct osc_data_version_args *dva = args;
771 struct osc_io *oio = dva->dva_oio;
772 const struct ost_body *body;
778 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
780 GOTO(out, rc = -EPROTO);
782 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, &oio->oi_oa,
786 oio->oi_cbarg.opc_rc = rc;
787 complete(&oio->oi_cbarg.opc_sync);
792 static int osc_io_data_version_start(const struct lu_env *env,
793 const struct cl_io_slice *slice)
795 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
796 struct osc_io *oio = cl2osc_io(env, slice);
797 struct obdo *oa = &oio->oi_oa;
798 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
799 struct osc_object *obj = cl2osc(slice->cis_obj);
800 struct lov_oinfo *loi = obj->oo_oinfo;
801 struct obd_export *exp = osc_export(obj);
802 struct ptlrpc_request *req;
803 struct ost_body *body;
804 struct osc_data_version_args *dva;
808 memset(oa, 0, sizeof(*oa));
809 oa->o_oi = loi->loi_oi;
810 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
812 if (dv->dv_flags & (LL_DV_RD_FLUSH | LL_DV_WR_FLUSH)) {
813 oa->o_valid |= OBD_MD_FLFLAGS;
814 oa->o_flags |= OBD_FL_SRVLOCK;
815 if (dv->dv_flags & LL_DV_WR_FLUSH)
816 oa->o_flags |= OBD_FL_FLUSH;
819 init_completion(&cbargs->opc_sync);
821 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
825 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
827 ptlrpc_request_free(req);
831 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
832 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
834 ptlrpc_request_set_replen(req);
835 req->rq_interpret_reply = osc_data_version_interpret;
836 dva = ptlrpc_req_async_args(dva, req);
839 ptlrpcd_add_req(req);
844 static void osc_io_data_version_end(const struct lu_env *env,
845 const struct cl_io_slice *slice)
847 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
848 struct osc_io *oio = cl2osc_io(env, slice);
849 struct cl_object *obj = slice->cis_obj;
850 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
851 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
852 struct obdo *oa = &oio->oi_oa;
853 unsigned int cl_valid = 0;
856 wait_for_completion(&cbargs->opc_sync);
858 if (cbargs->opc_rc != 0) {
859 slice->cis_io->ci_result = cbargs->opc_rc;
861 slice->cis_io->ci_result = 0;
863 (OBD_MD_LAYOUT_VERSION | OBD_MD_FLDATAVERSION)))
864 slice->cis_io->ci_result = -ENOTSUPP;
866 if (oa->o_valid & OBD_MD_LAYOUT_VERSION)
867 dv->dv_layout_version = oa->o_layout_version;
868 if (oa->o_valid & OBD_MD_FLDATAVERSION)
869 dv->dv_data_version = oa->o_data_version;
871 if (dv->dv_flags & LL_DV_SZ_UPDATE) {
872 if (oa->o_valid & OBD_MD_FLSIZE) {
873 attr->cat_size = oa->o_size;
874 cl_valid |= CAT_SIZE;
877 if (oa->o_valid & OBD_MD_FLBLOCKS) {
878 attr->cat_blocks = oa->o_blocks;
879 cl_valid |= CAT_BLOCKS;
882 cl_object_attr_lock(obj);
883 cl_object_attr_update(env, obj, attr, cl_valid);
884 cl_object_attr_unlock(obj);
891 int osc_io_read_start(const struct lu_env *env,
892 const struct cl_io_slice *slice)
894 struct cl_object *obj = slice->cis_obj;
895 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
899 if (!slice->cis_io->ci_noatime) {
900 cl_object_attr_lock(obj);
901 attr->cat_atime = ktime_get_real_seconds();
902 rc = cl_object_attr_update(env, obj, attr, CAT_ATIME);
903 cl_object_attr_unlock(obj);
908 EXPORT_SYMBOL(osc_io_read_start);
910 int osc_io_write_start(const struct lu_env *env,
911 const struct cl_io_slice *slice)
913 struct cl_object *obj = slice->cis_obj;
914 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
918 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_SETTIME, 1);
919 cl_object_attr_lock(obj);
920 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
921 rc = cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
922 cl_object_attr_unlock(obj);
926 EXPORT_SYMBOL(osc_io_write_start);
928 int osc_fsync_ost(const struct lu_env *env, struct osc_object *obj,
929 struct cl_fsync_io *fio)
931 struct osc_io *oio = osc_env_io(env);
932 struct obdo *oa = &oio->oi_oa;
933 struct lov_oinfo *loi = obj->oo_oinfo;
934 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
938 memset(oa, 0, sizeof(*oa));
939 oa->o_oi = loi->loi_oi;
940 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
942 /* reload size abd blocks for start and end of sync range */
943 oa->o_size = fio->fi_start;
944 oa->o_blocks = fio->fi_end;
945 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
947 obdo_set_parent_fid(oa, fio->fi_fid);
949 init_completion(&cbargs->opc_sync);
951 rc = osc_sync_base(obj, oa, osc_async_upcall, cbargs, PTLRPCD_SET);
954 EXPORT_SYMBOL(osc_fsync_ost);
956 int osc_io_fsync_start(const struct lu_env *env,
957 const struct cl_io_slice *slice)
959 struct cl_io *io = slice->cis_io;
960 struct cl_fsync_io *fio = &io->u.ci_fsync;
961 struct cl_object *obj = slice->cis_obj;
962 struct osc_object *osc = cl2osc(obj);
963 pgoff_t start = cl_index(obj, fio->fi_start);
964 pgoff_t end = cl_index(obj, fio->fi_end);
968 if (fio->fi_end == OBD_OBJECT_EOF)
971 result = osc_cache_writeback_range(env, osc, start, end, 0,
972 fio->fi_mode == CL_FSYNC_DISCARD);
974 fio->fi_nr_written += result;
977 if (fio->fi_mode == CL_FSYNC_ALL) {
980 /* we have to wait for writeback to finish before we can
981 * send OST_SYNC RPC. This is bad because it causes extents
982 * to be written osc by osc. However, we usually start
983 * writeback before CL_FSYNC_ALL so this won't have any real
985 rc = osc_cache_wait_range(env, osc, start, end);
988 rc = osc_fsync_ost(env, osc, fio);
996 void osc_io_fsync_end(const struct lu_env *env,
997 const struct cl_io_slice *slice)
999 struct cl_fsync_io *fio = &slice->cis_io->u.ci_fsync;
1000 struct cl_object *obj = slice->cis_obj;
1001 pgoff_t start = cl_index(obj, fio->fi_start);
1002 pgoff_t end = cl_index(obj, fio->fi_end);
1005 if (fio->fi_mode == CL_FSYNC_LOCAL) {
1006 result = osc_cache_wait_range(env, cl2osc(obj), start, end);
1007 } else if (fio->fi_mode == CL_FSYNC_ALL) {
1008 struct osc_io *oio = cl2osc_io(env, slice);
1009 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1011 wait_for_completion(&cbargs->opc_sync);
1013 result = cbargs->opc_rc;
1015 slice->cis_io->ci_result = result;
1017 EXPORT_SYMBOL(osc_io_fsync_end);
1019 static int osc_io_ladvise_start(const struct lu_env *env,
1020 const struct cl_io_slice *slice)
1023 struct cl_io *io = slice->cis_io;
1024 struct osc_io *oio = cl2osc_io(env, slice);
1025 struct cl_object *obj = slice->cis_obj;
1026 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1027 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
1028 struct obdo *oa = &oio->oi_oa;
1029 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1030 struct lu_ladvise *ladvise;
1031 struct ladvise_hdr *ladvise_hdr;
1036 /* TODO: add multiple ladvise support in CLIO */
1037 buf_size = offsetof(typeof(*ladvise_hdr), lah_advise[num_advise]);
1038 if (osc_env_info(env)->oti_ladvise_buf.lb_len < buf_size)
1039 lu_buf_realloc(&osc_env_info(env)->oti_ladvise_buf, buf_size);
1041 ladvise_hdr = osc_env_info(env)->oti_ladvise_buf.lb_buf;
1042 if (ladvise_hdr == NULL)
1045 memset(ladvise_hdr, 0, buf_size);
1046 ladvise_hdr->lah_magic = LADVISE_MAGIC;
1047 ladvise_hdr->lah_count = num_advise;
1048 ladvise_hdr->lah_flags = lio->li_flags;
1050 memset(oa, 0, sizeof(*oa));
1051 oa->o_oi = loi->loi_oi;
1052 oa->o_valid = OBD_MD_FLID;
1053 obdo_set_parent_fid(oa, lio->li_fid);
1055 ladvise = ladvise_hdr->lah_advise;
1056 ladvise->lla_start = lio->li_start;
1057 ladvise->lla_end = lio->li_end;
1058 ladvise->lla_advice = lio->li_advice;
1060 if (lio->li_flags & LF_ASYNC) {
1061 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1062 ladvise_hdr, NULL, NULL, NULL);
1064 init_completion(&cbargs->opc_sync);
1065 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1066 ladvise_hdr, osc_async_upcall,
1067 cbargs, PTLRPCD_SET);
1068 cbargs->opc_rpc_sent = result == 0;
1073 static void osc_io_ladvise_end(const struct lu_env *env,
1074 const struct cl_io_slice *slice)
1076 struct cl_io *io = slice->cis_io;
1077 struct osc_io *oio = cl2osc_io(env, slice);
1078 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1080 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
1082 if ((!(lio->li_flags & LF_ASYNC)) && cbargs->opc_rpc_sent) {
1083 wait_for_completion(&cbargs->opc_sync);
1084 result = cbargs->opc_rc;
1086 slice->cis_io->ci_result = result;
1089 void osc_io_end(const struct lu_env *env, const struct cl_io_slice *slice)
1091 struct osc_io *oio = cl2osc_io(env, slice);
1093 if (oio->oi_active) {
1094 osc_extent_release(env, oio->oi_active);
1095 oio->oi_active = NULL;
1098 EXPORT_SYMBOL(osc_io_end);
1100 struct osc_lseek_args {
1101 struct osc_io *lsa_oio;
1104 static int osc_lseek_interpret(const struct lu_env *env,
1105 struct ptlrpc_request *req,
1108 struct ost_body *reply;
1109 struct osc_lseek_args *lsa = arg;
1110 struct osc_io *oio = lsa->lsa_oio;
1111 struct cl_io *io = oio->oi_cl.cis_io;
1112 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1119 reply = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1121 GOTO(out, rc = -EPROTO);
1123 lsio->ls_result = reply->oa.o_size;
1125 osc_async_upcall(&oio->oi_cbarg, rc);
1129 int osc_io_lseek_start(const struct lu_env *env,
1130 const struct cl_io_slice *slice)
1132 struct cl_io *io = slice->cis_io;
1133 struct osc_io *oio = cl2osc_io(env, slice);
1134 struct cl_object *obj = slice->cis_obj;
1135 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1136 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1137 struct obdo *oa = &oio->oi_oa;
1138 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1139 struct obd_export *exp = osc_export(cl2osc(obj));
1140 struct ptlrpc_request *req;
1141 struct ost_body *body;
1142 struct osc_lseek_args *lsa;
1147 /* No negative values at this point */
1148 LASSERT(lsio->ls_start >= 0);
1149 LASSERT(lsio->ls_whence == SEEK_HOLE || lsio->ls_whence == SEEK_DATA);
1151 /* with IO lock taken we have object size in LVB and can check
1152 * boundaries prior sending LSEEK RPC
1154 if (lsio->ls_start >= loi->loi_lvb.lvb_size) {
1155 /* consider area beyond end of object as hole */
1156 if (lsio->ls_whence == SEEK_HOLE)
1157 lsio->ls_result = lsio->ls_start;
1159 lsio->ls_result = -ENXIO;
1163 /* if LSEEK RPC is not supported by server, consider whole stripe
1164 * object is data with hole after end of object
1166 if (!exp_connect_lseek(exp)) {
1167 if (lsio->ls_whence == SEEK_HOLE)
1168 lsio->ls_result = loi->loi_lvb.lvb_size;
1170 lsio->ls_result = lsio->ls_start;
1174 memset(oa, 0, sizeof(*oa));
1175 oa->o_oi = loi->loi_oi;
1176 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1177 oa->o_size = lsio->ls_start;
1178 oa->o_mode = lsio->ls_whence;
1179 if (oio->oi_lockless) {
1180 oa->o_flags = OBD_FL_SRVLOCK;
1181 oa->o_valid |= OBD_MD_FLFLAGS;
1184 init_completion(&cbargs->opc_sync);
1185 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SEEK);
1189 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SEEK);
1191 ptlrpc_request_free(req);
1195 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1196 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1197 ptlrpc_request_set_replen(req);
1198 req->rq_interpret_reply = osc_lseek_interpret;
1199 lsa = ptlrpc_req_async_args(lsa, req);
1202 ptlrpcd_add_req(req);
1203 cbargs->opc_rpc_sent = 1;
1207 EXPORT_SYMBOL(osc_io_lseek_start);
1209 void osc_io_lseek_end(const struct lu_env *env,
1210 const struct cl_io_slice *slice)
1212 struct osc_io *oio = cl2osc_io(env, slice);
1213 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1216 if (cbargs->opc_rpc_sent) {
1217 wait_for_completion(&cbargs->opc_sync);
1218 rc = cbargs->opc_rc;
1220 slice->cis_io->ci_result = rc;
1222 EXPORT_SYMBOL(osc_io_lseek_end);
1224 int osc_io_lru_reserve(const struct lu_env *env,
1225 const struct cl_io_slice *ios,
1226 loff_t pos, size_t bytes)
1228 struct osc_object *osc = cl2osc(ios->cis_obj);
1229 struct osc_io *oio = osc_env_io(env);
1230 unsigned long npages = 0;
1235 page_offset = pos & ~PAGE_MASK;
1238 if (bytes > PAGE_SIZE - page_offset)
1239 bytes -= (PAGE_SIZE - page_offset);
1243 npages += (bytes + PAGE_SIZE - 1) >> PAGE_SHIFT;
1244 oio->oi_lru_reserved = osc_lru_reserve(osc_cli(osc), npages);
1248 EXPORT_SYMBOL(osc_io_lru_reserve);
1250 static const struct cl_io_operations osc_io_ops = {
1253 .cio_iter_init = osc_io_iter_init,
1254 .cio_iter_fini = osc_io_rw_iter_fini,
1255 .cio_start = osc_io_read_start,
1256 .cio_fini = osc_io_fini
1259 .cio_iter_init = osc_io_iter_init,
1260 .cio_iter_fini = osc_io_rw_iter_fini,
1261 .cio_start = osc_io_write_start,
1262 .cio_end = osc_io_end,
1263 .cio_fini = osc_io_fini
1266 .cio_iter_init = osc_io_iter_init,
1267 .cio_iter_fini = osc_io_iter_fini,
1268 .cio_start = osc_io_setattr_start,
1269 .cio_end = osc_io_setattr_end
1271 [CIT_DATA_VERSION] = {
1272 .cio_start = osc_io_data_version_start,
1273 .cio_end = osc_io_data_version_end,
1276 .cio_iter_init = osc_io_iter_init,
1277 .cio_iter_fini = osc_io_iter_fini,
1278 .cio_start = osc_io_fault_start,
1279 .cio_end = osc_io_end,
1280 .cio_fini = osc_io_fini
1283 .cio_start = osc_io_fsync_start,
1284 .cio_end = osc_io_fsync_end,
1285 .cio_fini = osc_io_fini
1288 .cio_start = osc_io_ladvise_start,
1289 .cio_end = osc_io_ladvise_end,
1290 .cio_fini = osc_io_fini
1293 .cio_start = osc_io_lseek_start,
1294 .cio_end = osc_io_lseek_end,
1295 .cio_fini = osc_io_fini
1298 .cio_fini = osc_io_fini
1301 .cio_read_ahead = osc_io_read_ahead,
1302 .cio_lru_reserve = osc_io_lru_reserve,
1303 .cio_submit = osc_io_submit,
1304 .cio_commit_async = osc_io_commit_async,
1305 .cio_extent_release = osc_io_extent_release
1308 /*****************************************************************************
1310 * Transfer operations.
1314 int osc_io_init(const struct lu_env *env,
1315 struct cl_object *obj, struct cl_io *io)
1317 struct osc_io *oio = osc_env_io(env);
1319 CL_IO_SLICE_CLEAN(oio, oi_cl);
1320 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);