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 spin_lock(&oap->oap_lock);
195 oap->oap_async_flags = ASYNC_URGENT|ASYNC_READY;
196 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
197 spin_unlock(&oap->oap_lock);
199 osc_page_submit(env, opg, crt, brw_flags, submit_time);
200 list_add_tail(&oap->oap_pending_item, &list);
202 if (page->cp_sync_io != NULL)
203 cl_page_list_move(qout, qin, page);
205 cl_page_list_del(env, qin, page);
208 if (queued == max_pages) {
210 } else if (crt == CRT_WRITE) {
212 unsigned int next_chunks;
214 chunks = (queued + ppc - 1) >> ppc_bits;
215 /* chunk number if add another page */
216 next_chunks = (queued + ppc) >> ppc_bits;
218 /* next page will excceed write chunk limit */
219 if (chunks == osc_max_write_chunks(cli) &&
220 next_chunks > chunks)
225 result = osc_queue_sync_pages(env, io, osc, &list,
235 result = osc_queue_sync_pages(env, io, osc, &list, brw_flags);
237 /* Update c/mtime for sync write. LU-7310 */
238 if (crt == CRT_WRITE && qout->pl_nr > 0 && result == 0) {
239 struct cl_object *obj = ios->cis_obj;
240 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
242 cl_object_attr_lock(obj);
243 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
244 cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
245 cl_object_attr_unlock(obj);
248 CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
249 return qout->pl_nr > 0 ? 0 : result;
251 EXPORT_SYMBOL(osc_io_submit);
254 * This is called to update the attributes when modifying a specific page,
255 * both when making new pages and when doing updates to existing cached pages.
257 * Expand stripe KMS if necessary.
259 void osc_page_touch_at(const struct lu_env *env, struct cl_object *obj,
260 pgoff_t idx, size_t to)
262 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
263 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
269 /* offset within stripe */
270 kms = cl_offset(obj, idx) + to;
272 cl_object_attr_lock(obj);
273 CDEBUG(D_INODE, "stripe KMS %sincreasing %llu->%llu %llu\n",
274 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
275 loi->loi_lvb.lvb_size);
277 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
278 valid = CAT_MTIME | CAT_CTIME;
279 if (kms > loi->loi_kms) {
283 if (kms > loi->loi_lvb.lvb_size) {
284 attr->cat_size = kms;
287 cl_object_attr_update(env, obj, attr, valid);
288 cl_object_attr_unlock(obj);
293 int osc_io_commit_async(const struct lu_env *env,
294 const struct cl_io_slice *ios,
295 struct cl_page_list *qin, int from, int to,
298 struct cl_io *io = ios->cis_io;
299 struct osc_io *oio = cl2osc_io(env, ios);
300 struct osc_object *osc = cl2osc(ios->cis_obj);
301 struct cl_page *page;
302 struct cl_page *last_page;
303 struct osc_page *opg;
304 struct pagevec *pvec = &osc_env_info(env)->oti_pagevec;
308 LASSERT(qin->pl_nr > 0);
310 /* Handle partial page cases */
311 last_page = cl_page_list_last(qin);
312 if (oio->oi_lockless) {
313 page = cl_page_list_first(qin);
314 if (page == last_page) {
315 cl_page_clip(env, page, from, to);
318 cl_page_clip(env, page, from, PAGE_SIZE);
320 cl_page_clip(env, last_page, 0, to);
324 ll_pagevec_init(pvec, 0);
326 while (qin->pl_nr > 0) {
327 struct osc_async_page *oap;
329 page = cl_page_list_first(qin);
330 opg = osc_cl_page_osc(page, osc);
333 LASSERTF(osc == oap->oap_obj,
334 "obj mismatch: %p / %p\n", osc, oap->oap_obj);
336 if (!list_empty(&oap->oap_rpc_item)) {
337 CDEBUG(D_CACHE, "Busy oap %p page %p for submit.\n",
343 /* The page may be already in dirty cache. */
344 if (list_empty(&oap->oap_pending_item)) {
345 result = osc_page_cache_add(env, opg, io, cb);
350 osc_page_touch_at(env, osc2cl(osc), osc_index(opg),
351 page == last_page ? to : PAGE_SIZE);
353 cl_page_list_del(env, qin, page);
355 /* if there are no more slots, do the callback & reinit */
356 if (pagevec_add(pvec, page->cp_vmpage) == 0) {
357 (*cb)(env, io, pvec);
358 pagevec_reinit(pvec);
361 /* The shrink interval is in seconds, so we can update it once per
362 * write, rather than once per page.
364 osc_update_next_shrink(osc_cli(osc));
367 /* Clean up any partially full pagevecs */
368 if (pagevec_count(pvec) != 0)
369 (*cb)(env, io, pvec);
371 /* Can't access these pages any more. Page can be in transfer and
372 * complete at any time. */
374 /* for sync write, kernel will wait for this page to be flushed before
375 * osc_io_end() is called, so release it earlier.
376 * for mkwrite(), it's known there is no further pages. */
377 if (cl_io_is_sync_write(io) && oio->oi_active != NULL) {
378 osc_extent_release(env, oio->oi_active);
379 oio->oi_active = NULL;
382 CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, result);
385 EXPORT_SYMBOL(osc_io_commit_async);
387 void osc_io_extent_release(const struct lu_env *env,
388 const struct cl_io_slice *ios)
390 struct osc_io *oio = cl2osc_io(env, ios);
392 if (oio->oi_active != NULL) {
393 osc_extent_release(env, oio->oi_active);
394 oio->oi_active = NULL;
397 EXPORT_SYMBOL(osc_io_extent_release);
399 static bool osc_import_not_healthy(struct obd_import *imp)
401 return imp->imp_invalid || imp->imp_deactive ||
402 !(imp->imp_state == LUSTRE_IMP_FULL ||
403 imp->imp_state == LUSTRE_IMP_IDLE);
406 int osc_io_iter_init(const struct lu_env *env, const struct cl_io_slice *ios)
408 struct osc_object *osc = cl2osc(ios->cis_obj);
409 struct obd_import *imp = osc_cli(osc)->cl_import;
410 struct osc_io *oio = osc_env_io(env);
415 spin_lock(&imp->imp_lock);
417 * check whether this OSC device is available for non-delay read,
418 * fast switching mirror if we haven't tried all mirrors.
420 if (ios->cis_io->ci_type == CIT_READ && ios->cis_io->ci_ndelay &&
421 !ios->cis_io->ci_tried_all_mirrors && osc_import_not_healthy(imp)) {
423 } else if (likely(!imp->imp_invalid)) {
424 atomic_inc(&osc->oo_nr_ios);
425 oio->oi_is_active = 1;
428 spin_unlock(&imp->imp_lock);
430 if (capable(CAP_SYS_RESOURCE))
431 oio->oi_cap_sys_resource = 1;
435 EXPORT_SYMBOL(osc_io_iter_init);
437 void osc_io_iter_fini(const struct lu_env *env,
438 const struct cl_io_slice *ios)
440 struct osc_io *oio = osc_env_io(env);
442 if (oio->oi_is_active) {
443 struct osc_object *osc = cl2osc(ios->cis_obj);
445 oio->oi_is_active = 0;
446 LASSERT(atomic_read(&osc->oo_nr_ios) > 0);
447 if (atomic_dec_and_test(&osc->oo_nr_ios))
448 wake_up(&osc->oo_io_waitq);
451 EXPORT_SYMBOL(osc_io_iter_fini);
453 void osc_io_rw_iter_fini(const struct lu_env *env,
454 const struct cl_io_slice *ios)
456 struct osc_io *oio = osc_env_io(env);
457 struct osc_object *osc = cl2osc(ios->cis_obj);
459 if (oio->oi_lru_reserved > 0) {
460 osc_lru_unreserve(osc_cli(osc), oio->oi_lru_reserved);
461 oio->oi_lru_reserved = 0;
463 oio->oi_write_osclock = NULL;
465 osc_io_iter_fini(env, ios);
467 EXPORT_SYMBOL(osc_io_rw_iter_fini);
469 int osc_io_fault_start(const struct lu_env *env, const struct cl_io_slice *ios)
472 struct cl_fault_io *fio;
476 fio = &io->u.ci_fault;
477 CDEBUG(D_INFO, "%lu %d %zu\n",
478 fio->ft_index, fio->ft_writable, fio->ft_nob);
480 * If mapping is writeable, adjust kms to cover this page,
481 * but do not extend kms beyond actual file size.
484 if (fio->ft_writable)
485 osc_page_touch_at(env, ios->cis_obj,
486 fio->ft_index, fio->ft_nob);
489 EXPORT_SYMBOL(osc_io_fault_start);
492 static int osc_async_upcall(void *a, int rc)
494 struct osc_async_cbargs *args = a;
497 complete(&args->opc_sync);
502 * Checks that there are no pages being written in the extent being truncated.
504 static bool trunc_check_cb(const struct lu_env *env, struct cl_io *io,
505 void **pvec, int count, void *cbdata)
509 for (i = 0; i < count; i++) {
510 struct osc_page *ops = pvec[i];
511 struct cl_page *page = ops->ops_cl.cpl_page;
512 struct osc_async_page *oap;
513 __u64 start = *(__u64 *)cbdata;
516 if (oap->oap_cmd & OBD_BRW_WRITE &&
517 !list_empty(&oap->oap_pending_item))
518 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %llu/%s.\n",
519 start, current->comm);
521 if (PageLocked(page->cp_vmpage))
522 CDEBUG(D_CACHE, "page %p index %lu locked for %d.\n",
524 oap->oap_cmd & OBD_BRW_RWMASK);
529 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
530 struct osc_io *oio, __u64 size)
532 struct cl_object *clob;
536 clob = oio->oi_cl.cis_obj;
537 start = cl_index(clob, size);
538 partial = cl_offset(clob, start) < size;
541 * Complain if there are pages in the truncated region.
543 osc_page_gang_lookup(env, io, cl2osc(clob),
544 start + partial, CL_PAGE_EOF,
545 trunc_check_cb, (void *)&size);
549 * Flush affected pages prior punch.
550 * We shouldn't discard them locally first because that could be data loss
551 * if server doesn't support fallocate punch, we also need these data to be
552 * flushed first to prevent re-ordering with the punch
554 static int osc_punch_start(const struct lu_env *env, struct cl_io *io,
555 struct cl_object *obj)
557 struct osc_object *osc = cl2osc(obj);
558 pgoff_t pg_start = cl_index(obj, io->u.ci_setattr.sa_falloc_offset);
559 pgoff_t pg_end = cl_index(obj, io->u.ci_setattr.sa_falloc_end - 1);
563 rc = osc_cache_writeback_range(env, osc, pg_start, pg_end, 1, 0);
567 osc_page_gang_lookup(env, io, osc, pg_start, pg_end, osc_discard_cb,
572 static int osc_io_setattr_start(const struct lu_env *env,
573 const struct cl_io_slice *slice)
575 struct cl_io *io = slice->cis_io;
576 struct osc_io *oio = cl2osc_io(env, slice);
577 struct cl_object *obj = slice->cis_obj;
578 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
579 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
580 struct obdo *oa = &oio->oi_oa;
581 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
582 unsigned int ia_avalid = io->u.ci_setattr.sa_avalid;
583 enum op_xvalid ia_xvalid = io->u.ci_setattr.sa_xvalid;
585 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
586 bool io_is_falloc = cl_io_is_fallocate(io);
589 /* truncate cache dirty pages first */
590 if (cl_io_is_trunc(io))
591 result = osc_cache_truncate_start(env, cl2osc(obj), size,
593 /* flush local pages prior punching them on server */
595 io->u.ci_setattr.sa_falloc_mode & FALLOC_FL_PUNCH_HOLE)
596 result = osc_punch_start(env, io, obj);
598 if (result == 0 && oio->oi_lockless == 0) {
599 cl_object_attr_lock(obj);
600 result = cl_object_attr_get(env, obj, attr);
602 struct ost_lvb *lvb = &io->u.ci_setattr.sa_attr;
603 unsigned int cl_valid = 0;
605 if (ia_avalid & ATTR_SIZE) {
606 attr->cat_size = size;
607 attr->cat_kms = size;
608 cl_valid = (CAT_SIZE | CAT_KMS);
610 if (ia_avalid & ATTR_MTIME_SET) {
611 attr->cat_mtime = lvb->lvb_mtime;
612 cl_valid |= CAT_MTIME;
614 if (ia_avalid & ATTR_ATIME_SET) {
615 attr->cat_atime = lvb->lvb_atime;
616 cl_valid |= CAT_ATIME;
618 if (ia_xvalid & OP_XVALID_CTIME_SET) {
619 attr->cat_ctime = lvb->lvb_ctime;
620 cl_valid |= CAT_CTIME;
622 result = cl_object_attr_update(env, obj, attr,
625 cl_object_attr_unlock(obj);
627 memset(oa, 0, sizeof(*oa));
629 oa->o_oi = loi->loi_oi;
630 obdo_set_parent_fid(oa, io->u.ci_setattr.sa_parent_fid);
631 oa->o_stripe_idx = io->u.ci_setattr.sa_stripe_index;
632 oa->o_layout = io->u.ci_setattr.sa_layout;
633 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP |
635 if (ia_avalid & ATTR_CTIME) {
636 oa->o_valid |= OBD_MD_FLCTIME;
637 oa->o_ctime = attr->cat_ctime;
639 if (ia_avalid & ATTR_ATIME) {
640 oa->o_valid |= OBD_MD_FLATIME;
641 oa->o_atime = attr->cat_atime;
643 if (ia_avalid & ATTR_MTIME) {
644 oa->o_valid |= OBD_MD_FLMTIME;
645 oa->o_mtime = attr->cat_mtime;
648 if (ia_avalid & ATTR_SIZE || io_is_falloc) {
649 if (oio->oi_lockless) {
650 oa->o_flags = OBD_FL_SRVLOCK;
651 oa->o_valid |= OBD_MD_FLFLAGS;
654 if (io->ci_layout_version > 0) {
655 /* verify layout version */
656 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
657 oa->o_layout_version = io->ci_layout_version;
660 LASSERT(oio->oi_lockless == 0);
663 if (ia_xvalid & OP_XVALID_FLAGS) {
664 oa->o_flags = io->u.ci_setattr.sa_attr_flags;
665 oa->o_valid |= OBD_MD_FLFLAGS;
668 init_completion(&cbargs->opc_sync);
671 int falloc_mode = io->u.ci_setattr.sa_falloc_mode;
673 oa->o_size = io->u.ci_setattr.sa_falloc_offset;
674 oa->o_blocks = io->u.ci_setattr.sa_falloc_end;
675 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
676 result = osc_fallocate_base(osc_export(cl2osc(obj)),
677 oa, osc_async_upcall,
678 cbargs, falloc_mode);
679 } else if (ia_avalid & ATTR_SIZE) {
681 oa->o_blocks = OBD_OBJECT_EOF;
682 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
683 result = osc_punch_send(osc_export(cl2osc(obj)),
684 oa, osc_async_upcall, cbargs);
686 result = osc_setattr_async(osc_export(cl2osc(obj)),
687 oa, osc_async_upcall,
688 cbargs, PTLRPCD_SET);
690 cbargs->opc_rpc_sent = result == 0;
696 void osc_io_setattr_end(const struct lu_env *env,
697 const struct cl_io_slice *slice)
699 struct cl_io *io = slice->cis_io;
700 struct osc_io *oio = cl2osc_io(env, slice);
701 struct cl_object *obj = slice->cis_obj;
702 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
703 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
704 struct obdo *oa = &oio->oi_oa;
705 unsigned int cl_valid = 0;
708 if (cbargs->opc_rpc_sent) {
709 wait_for_completion(&cbargs->opc_sync);
710 result = io->ci_result = cbargs->opc_rc;
714 if (oio->oi_lockless) {
715 /* lockless truncate */
716 struct osc_device *osc = lu2osc_dev(obj->co_lu.lo_dev);
718 LASSERT(cl_io_is_trunc(io) || cl_io_is_fallocate(io));
719 /* XXX: Need a lock. */
720 osc->od_stats.os_lockless_truncates++;
724 if (cl_io_is_trunc(io)) {
725 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
728 cl_object_attr_lock(obj);
729 if (oa->o_valid & OBD_MD_FLBLOCKS) {
730 attr->cat_blocks = oa->o_blocks;
731 cl_valid |= CAT_BLOCKS;
734 cl_object_attr_update(env, obj, attr, cl_valid);
735 cl_object_attr_unlock(obj);
737 osc_trunc_check(env, io, oio, size);
738 osc_cache_truncate_end(env, oio->oi_trunc);
739 oio->oi_trunc = NULL;
742 if (cl_io_is_fallocate(io)) {
744 cl_object_attr_lock(obj);
746 if (oa->o_valid & OBD_MD_FLBLOCKS) {
747 attr->cat_blocks = oa->o_blocks;
748 cl_valid |= CAT_BLOCKS;
751 cl_object_attr_update(env, obj, attr, cl_valid);
752 cl_object_attr_unlock(obj);
756 EXPORT_SYMBOL(osc_io_setattr_end);
758 struct osc_data_version_args {
759 struct osc_io *dva_oio;
763 osc_data_version_interpret(const struct lu_env *env, struct ptlrpc_request *req,
766 struct osc_data_version_args *dva = args;
767 struct osc_io *oio = dva->dva_oio;
768 const struct ost_body *body;
774 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
776 GOTO(out, rc = -EPROTO);
778 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, &oio->oi_oa,
782 oio->oi_cbarg.opc_rc = rc;
783 complete(&oio->oi_cbarg.opc_sync);
788 static int osc_io_data_version_start(const struct lu_env *env,
789 const struct cl_io_slice *slice)
791 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
792 struct osc_io *oio = cl2osc_io(env, slice);
793 struct obdo *oa = &oio->oi_oa;
794 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
795 struct osc_object *obj = cl2osc(slice->cis_obj);
796 struct lov_oinfo *loi = obj->oo_oinfo;
797 struct obd_export *exp = osc_export(obj);
798 struct ptlrpc_request *req;
799 struct ost_body *body;
800 struct osc_data_version_args *dva;
804 memset(oa, 0, sizeof(*oa));
805 oa->o_oi = loi->loi_oi;
806 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
808 if (dv->dv_flags & (LL_DV_RD_FLUSH | LL_DV_WR_FLUSH)) {
809 oa->o_valid |= OBD_MD_FLFLAGS;
810 oa->o_flags |= OBD_FL_SRVLOCK;
811 if (dv->dv_flags & LL_DV_WR_FLUSH)
812 oa->o_flags |= OBD_FL_FLUSH;
815 init_completion(&cbargs->opc_sync);
817 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
821 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
823 ptlrpc_request_free(req);
827 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
828 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
830 ptlrpc_request_set_replen(req);
831 req->rq_interpret_reply = osc_data_version_interpret;
832 dva = ptlrpc_req_async_args(dva, req);
835 ptlrpcd_add_req(req);
840 static void osc_io_data_version_end(const struct lu_env *env,
841 const struct cl_io_slice *slice)
843 struct cl_data_version_io *dv = &slice->cis_io->u.ci_data_version;
844 struct osc_io *oio = cl2osc_io(env, slice);
845 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
848 wait_for_completion(&cbargs->opc_sync);
850 if (cbargs->opc_rc != 0) {
851 slice->cis_io->ci_result = cbargs->opc_rc;
853 slice->cis_io->ci_result = 0;
854 if (!(oio->oi_oa.o_valid &
855 (OBD_MD_LAYOUT_VERSION | OBD_MD_FLDATAVERSION)))
856 slice->cis_io->ci_result = -ENOTSUPP;
858 if (oio->oi_oa.o_valid & OBD_MD_LAYOUT_VERSION)
859 dv->dv_layout_version = oio->oi_oa.o_layout_version;
860 if (oio->oi_oa.o_valid & OBD_MD_FLDATAVERSION)
861 dv->dv_data_version = oio->oi_oa.o_data_version;
867 int osc_io_read_start(const struct lu_env *env,
868 const struct cl_io_slice *slice)
870 struct cl_object *obj = slice->cis_obj;
871 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
875 if (!slice->cis_io->ci_noatime) {
876 cl_object_attr_lock(obj);
877 attr->cat_atime = ktime_get_real_seconds();
878 rc = cl_object_attr_update(env, obj, attr, CAT_ATIME);
879 cl_object_attr_unlock(obj);
884 EXPORT_SYMBOL(osc_io_read_start);
886 int osc_io_write_start(const struct lu_env *env,
887 const struct cl_io_slice *slice)
889 struct cl_object *obj = slice->cis_obj;
890 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
894 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_SETTIME, 1);
895 cl_object_attr_lock(obj);
896 attr->cat_mtime = attr->cat_ctime = ktime_get_real_seconds();
897 rc = cl_object_attr_update(env, obj, attr, CAT_MTIME | CAT_CTIME);
898 cl_object_attr_unlock(obj);
902 EXPORT_SYMBOL(osc_io_write_start);
904 int osc_fsync_ost(const struct lu_env *env, struct osc_object *obj,
905 struct cl_fsync_io *fio)
907 struct osc_io *oio = osc_env_io(env);
908 struct obdo *oa = &oio->oi_oa;
909 struct lov_oinfo *loi = obj->oo_oinfo;
910 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
914 memset(oa, 0, sizeof(*oa));
915 oa->o_oi = loi->loi_oi;
916 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
918 /* reload size abd blocks for start and end of sync range */
919 oa->o_size = fio->fi_start;
920 oa->o_blocks = fio->fi_end;
921 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
923 obdo_set_parent_fid(oa, fio->fi_fid);
925 init_completion(&cbargs->opc_sync);
927 rc = osc_sync_base(obj, oa, osc_async_upcall, cbargs, PTLRPCD_SET);
930 EXPORT_SYMBOL(osc_fsync_ost);
932 int osc_io_fsync_start(const struct lu_env *env,
933 const struct cl_io_slice *slice)
935 struct cl_io *io = slice->cis_io;
936 struct cl_fsync_io *fio = &io->u.ci_fsync;
937 struct cl_object *obj = slice->cis_obj;
938 struct osc_object *osc = cl2osc(obj);
939 pgoff_t start = cl_index(obj, fio->fi_start);
940 pgoff_t end = cl_index(obj, fio->fi_end);
944 if (fio->fi_end == OBD_OBJECT_EOF)
947 result = osc_cache_writeback_range(env, osc, start, end, 0,
948 fio->fi_mode == CL_FSYNC_DISCARD);
950 fio->fi_nr_written += result;
953 if (fio->fi_mode == CL_FSYNC_ALL) {
956 /* we have to wait for writeback to finish before we can
957 * send OST_SYNC RPC. This is bad because it causes extents
958 * to be written osc by osc. However, we usually start
959 * writeback before CL_FSYNC_ALL so this won't have any real
961 rc = osc_cache_wait_range(env, osc, start, end);
964 rc = osc_fsync_ost(env, osc, fio);
972 void osc_io_fsync_end(const struct lu_env *env,
973 const struct cl_io_slice *slice)
975 struct cl_fsync_io *fio = &slice->cis_io->u.ci_fsync;
976 struct cl_object *obj = slice->cis_obj;
977 pgoff_t start = cl_index(obj, fio->fi_start);
978 pgoff_t end = cl_index(obj, fio->fi_end);
981 if (fio->fi_mode == CL_FSYNC_LOCAL) {
982 result = osc_cache_wait_range(env, cl2osc(obj), start, end);
983 } else if (fio->fi_mode == CL_FSYNC_ALL) {
984 struct osc_io *oio = cl2osc_io(env, slice);
985 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
987 wait_for_completion(&cbargs->opc_sync);
989 result = cbargs->opc_rc;
991 slice->cis_io->ci_result = result;
993 EXPORT_SYMBOL(osc_io_fsync_end);
995 static int osc_io_ladvise_start(const struct lu_env *env,
996 const struct cl_io_slice *slice)
999 struct cl_io *io = slice->cis_io;
1000 struct osc_io *oio = cl2osc_io(env, slice);
1001 struct cl_object *obj = slice->cis_obj;
1002 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1003 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
1004 struct obdo *oa = &oio->oi_oa;
1005 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1006 struct lu_ladvise *ladvise;
1007 struct ladvise_hdr *ladvise_hdr;
1012 /* TODO: add multiple ladvise support in CLIO */
1013 buf_size = offsetof(typeof(*ladvise_hdr), lah_advise[num_advise]);
1014 if (osc_env_info(env)->oti_ladvise_buf.lb_len < buf_size)
1015 lu_buf_realloc(&osc_env_info(env)->oti_ladvise_buf, buf_size);
1017 ladvise_hdr = osc_env_info(env)->oti_ladvise_buf.lb_buf;
1018 if (ladvise_hdr == NULL)
1021 memset(ladvise_hdr, 0, buf_size);
1022 ladvise_hdr->lah_magic = LADVISE_MAGIC;
1023 ladvise_hdr->lah_count = num_advise;
1024 ladvise_hdr->lah_flags = lio->li_flags;
1026 memset(oa, 0, sizeof(*oa));
1027 oa->o_oi = loi->loi_oi;
1028 oa->o_valid = OBD_MD_FLID;
1029 obdo_set_parent_fid(oa, lio->li_fid);
1031 ladvise = ladvise_hdr->lah_advise;
1032 ladvise->lla_start = lio->li_start;
1033 ladvise->lla_end = lio->li_end;
1034 ladvise->lla_advice = lio->li_advice;
1036 if (lio->li_flags & LF_ASYNC) {
1037 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1038 ladvise_hdr, NULL, NULL, NULL);
1040 init_completion(&cbargs->opc_sync);
1041 result = osc_ladvise_base(osc_export(cl2osc(obj)), oa,
1042 ladvise_hdr, osc_async_upcall,
1043 cbargs, PTLRPCD_SET);
1044 cbargs->opc_rpc_sent = result == 0;
1049 static void osc_io_ladvise_end(const struct lu_env *env,
1050 const struct cl_io_slice *slice)
1052 struct cl_io *io = slice->cis_io;
1053 struct osc_io *oio = cl2osc_io(env, slice);
1054 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1056 struct cl_ladvise_io *lio = &io->u.ci_ladvise;
1058 if ((!(lio->li_flags & LF_ASYNC)) && cbargs->opc_rpc_sent) {
1059 wait_for_completion(&cbargs->opc_sync);
1060 result = cbargs->opc_rc;
1062 slice->cis_io->ci_result = result;
1065 void osc_io_end(const struct lu_env *env, const struct cl_io_slice *slice)
1067 struct osc_io *oio = cl2osc_io(env, slice);
1069 if (oio->oi_active) {
1070 osc_extent_release(env, oio->oi_active);
1071 oio->oi_active = NULL;
1074 EXPORT_SYMBOL(osc_io_end);
1076 struct osc_lseek_args {
1077 struct osc_io *lsa_oio;
1080 static int osc_lseek_interpret(const struct lu_env *env,
1081 struct ptlrpc_request *req,
1084 struct ost_body *reply;
1085 struct osc_lseek_args *lsa = arg;
1086 struct osc_io *oio = lsa->lsa_oio;
1087 struct cl_io *io = oio->oi_cl.cis_io;
1088 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1095 reply = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1097 GOTO(out, rc = -EPROTO);
1099 lsio->ls_result = reply->oa.o_size;
1101 osc_async_upcall(&oio->oi_cbarg, rc);
1105 int osc_io_lseek_start(const struct lu_env *env,
1106 const struct cl_io_slice *slice)
1108 struct cl_io *io = slice->cis_io;
1109 struct osc_io *oio = cl2osc_io(env, slice);
1110 struct cl_object *obj = slice->cis_obj;
1111 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1112 struct cl_lseek_io *lsio = &io->u.ci_lseek;
1113 struct obdo *oa = &oio->oi_oa;
1114 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1115 struct obd_export *exp = osc_export(cl2osc(obj));
1116 struct ptlrpc_request *req;
1117 struct ost_body *body;
1118 struct osc_lseek_args *lsa;
1123 /* No negative values at this point */
1124 LASSERT(lsio->ls_start >= 0);
1125 LASSERT(lsio->ls_whence == SEEK_HOLE || lsio->ls_whence == SEEK_DATA);
1127 /* with IO lock taken we have object size in LVB and can check
1128 * boundaries prior sending LSEEK RPC
1130 if (lsio->ls_start >= loi->loi_lvb.lvb_size) {
1131 /* consider area beyond end of object as hole */
1132 if (lsio->ls_whence == SEEK_HOLE)
1133 lsio->ls_result = lsio->ls_start;
1135 lsio->ls_result = -ENXIO;
1139 /* if LSEEK RPC is not supported by server, consider whole stripe
1140 * object is data with hole after end of object
1142 if (!exp_connect_lseek(exp)) {
1143 if (lsio->ls_whence == SEEK_HOLE)
1144 lsio->ls_result = loi->loi_lvb.lvb_size;
1146 lsio->ls_result = lsio->ls_start;
1150 memset(oa, 0, sizeof(*oa));
1151 oa->o_oi = loi->loi_oi;
1152 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1153 oa->o_size = lsio->ls_start;
1154 oa->o_mode = lsio->ls_whence;
1155 if (oio->oi_lockless) {
1156 oa->o_flags = OBD_FL_SRVLOCK;
1157 oa->o_valid |= OBD_MD_FLFLAGS;
1160 init_completion(&cbargs->opc_sync);
1161 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SEEK);
1165 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SEEK);
1167 ptlrpc_request_free(req);
1171 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1172 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1173 ptlrpc_request_set_replen(req);
1174 req->rq_interpret_reply = osc_lseek_interpret;
1175 lsa = ptlrpc_req_async_args(lsa, req);
1178 ptlrpcd_add_req(req);
1179 cbargs->opc_rpc_sent = 1;
1183 EXPORT_SYMBOL(osc_io_lseek_start);
1185 void osc_io_lseek_end(const struct lu_env *env,
1186 const struct cl_io_slice *slice)
1188 struct osc_io *oio = cl2osc_io(env, slice);
1189 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
1192 if (cbargs->opc_rpc_sent) {
1193 wait_for_completion(&cbargs->opc_sync);
1194 rc = cbargs->opc_rc;
1196 slice->cis_io->ci_result = rc;
1198 EXPORT_SYMBOL(osc_io_lseek_end);
1200 int osc_io_lru_reserve(const struct lu_env *env,
1201 const struct cl_io_slice *ios,
1202 loff_t pos, size_t bytes)
1204 struct osc_object *osc = cl2osc(ios->cis_obj);
1205 struct osc_io *oio = osc_env_io(env);
1206 unsigned long npages = 0;
1211 page_offset = pos & ~PAGE_MASK;
1214 if (bytes > PAGE_SIZE - page_offset)
1215 bytes -= (PAGE_SIZE - page_offset);
1219 npages += (bytes + PAGE_SIZE - 1) >> PAGE_SHIFT;
1220 oio->oi_lru_reserved = osc_lru_reserve(osc_cli(osc), npages);
1224 EXPORT_SYMBOL(osc_io_lru_reserve);
1226 static const struct cl_io_operations osc_io_ops = {
1229 .cio_iter_init = osc_io_iter_init,
1230 .cio_iter_fini = osc_io_rw_iter_fini,
1231 .cio_start = osc_io_read_start,
1232 .cio_fini = osc_io_fini
1235 .cio_iter_init = osc_io_iter_init,
1236 .cio_iter_fini = osc_io_rw_iter_fini,
1237 .cio_start = osc_io_write_start,
1238 .cio_end = osc_io_end,
1239 .cio_fini = osc_io_fini
1242 .cio_iter_init = osc_io_iter_init,
1243 .cio_iter_fini = osc_io_iter_fini,
1244 .cio_start = osc_io_setattr_start,
1245 .cio_end = osc_io_setattr_end
1247 [CIT_DATA_VERSION] = {
1248 .cio_start = osc_io_data_version_start,
1249 .cio_end = osc_io_data_version_end,
1252 .cio_iter_init = osc_io_iter_init,
1253 .cio_iter_fini = osc_io_iter_fini,
1254 .cio_start = osc_io_fault_start,
1255 .cio_end = osc_io_end,
1256 .cio_fini = osc_io_fini
1259 .cio_start = osc_io_fsync_start,
1260 .cio_end = osc_io_fsync_end,
1261 .cio_fini = osc_io_fini
1264 .cio_start = osc_io_ladvise_start,
1265 .cio_end = osc_io_ladvise_end,
1266 .cio_fini = osc_io_fini
1269 .cio_start = osc_io_lseek_start,
1270 .cio_end = osc_io_lseek_end,
1271 .cio_fini = osc_io_fini
1274 .cio_fini = osc_io_fini
1277 .cio_read_ahead = osc_io_read_ahead,
1278 .cio_lru_reserve = osc_io_lru_reserve,
1279 .cio_submit = osc_io_submit,
1280 .cio_commit_async = osc_io_commit_async,
1281 .cio_extent_release = osc_io_extent_release
1284 /*****************************************************************************
1286 * Transfer operations.
1290 int osc_io_init(const struct lu_env *env,
1291 struct cl_object *obj, struct cl_io *io)
1293 struct osc_io *oio = osc_env_io(env);
1295 CL_IO_SLICE_CLEAN(oio, oi_cl);
1296 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);