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) 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Author: Nikita Danilov <nikita@clusterfs.com>
37 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
41 /* prerequisite for linux/xattr.h */
42 #include <linux/types.h>
43 /* prerequisite for linux/xattr.h */
46 #include <linux/pagevec.h>
49 * struct OBD_{ALLOC,FREE}*()
52 #include <obd_support.h>
54 #include "osd_internal.h"
57 #include <ldiskfs/ldiskfs_extents.h>
59 static inline bool osd_use_page_cache(struct osd_device *d)
61 /* do not use pagecache if write and read caching are disabled */
62 if (d->od_writethrough_cache + d->od_read_cache == 0)
64 /* use pagecache by default */
68 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
69 int rw, int line, int pages)
73 LASSERTF(iobuf->dr_elapsed_valid == 0,
74 "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
75 atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
77 LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
79 init_waitqueue_head(&iobuf->dr_wait);
80 atomic_set(&iobuf->dr_numreqs, 0);
85 iobuf->dr_elapsed = ktime_set(0, 0);
86 /* must be counted before, so assert */
88 iobuf->dr_init_at = line;
90 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
91 if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
92 LASSERT(iobuf->dr_pg_buf.lb_len >=
93 pages * sizeof(iobuf->dr_pages[0]));
97 /* start with 1MB for 4K blocks */
99 while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
102 CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
103 (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
105 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
106 iobuf->dr_max_pages = 0;
107 CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
108 (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
110 lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
111 iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
112 if (unlikely(iobuf->dr_blocks == NULL))
115 lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
116 iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
117 if (unlikely(iobuf->dr_pages == NULL))
120 lu_buf_realloc(&iobuf->dr_lnb_buf,
121 pages * sizeof(iobuf->dr_lnbs[0]));
122 iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf;
123 if (unlikely(iobuf->dr_lnbs == NULL))
126 iobuf->dr_max_pages = pages;
130 #define osd_init_iobuf(dev, iobuf, rw, pages) \
131 __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
133 static void osd_iobuf_add_page(struct osd_iobuf *iobuf,
134 struct niobuf_local *lnb)
136 LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
137 iobuf->dr_pages[iobuf->dr_npages] = lnb->lnb_page;
138 iobuf->dr_lnbs[iobuf->dr_npages] = lnb;
142 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
144 int rw = iobuf->dr_rw;
146 if (iobuf->dr_elapsed_valid) {
147 iobuf->dr_elapsed_valid = 0;
148 LASSERT(iobuf->dr_dev == d);
149 LASSERT(iobuf->dr_frags > 0);
150 lprocfs_oh_tally(&d->od_brw_stats.
151 hist[BRW_R_DIO_FRAGS+rw],
153 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
154 ktime_to_ms(iobuf->dr_elapsed));
158 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
159 static void dio_complete_routine(struct bio *bio)
161 int error = bio->bi_status;
163 static void dio_complete_routine(struct bio *bio, int error)
166 struct osd_iobuf *iobuf = bio->bi_private;
169 /* CAVEAT EMPTOR: possibly in IRQ context
170 * DO NOT record procfs stats here!!! */
172 if (unlikely(iobuf == NULL)) {
173 CERROR("***** bio->bi_private is NULL! This should never "
174 "happen. Normally, I would crash here, but instead I "
175 "will dump the bio contents to the console. Please "
176 "report this to <https://jira.whamcloud.com/> , along "
177 "with any interesting messages leading up to this point "
178 "(like SCSI errors, perhaps). Because bi_private is "
179 "NULL, I can't wake up the thread that initiated this "
180 "IO - you will probably have to reboot this node.\n");
181 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
182 ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
183 bio->bi_next, (unsigned long)bio->bi_flags,
184 (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
185 bio_sectors(bio) << 9, bio->bi_end_io,
186 atomic_read(&bio->__bi_cnt),
191 /* the check is outside of the cycle for performance reason -bzzz */
192 if (!bio_data_dir(bio)) {
193 DECLARE_BVEC_ITER_ALL(iter_all);
195 bio_for_each_segment_all(bvl, bio, iter_all) {
196 if (likely(error == 0))
197 SetPageUptodate(bvl_to_page(bvl));
198 LASSERT(PageLocked(bvl_to_page(bvl)));
200 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
202 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
205 /* any real error is good enough -bzzz */
206 if (error != 0 && iobuf->dr_error == 0)
207 iobuf->dr_error = error;
210 * set dr_elapsed before dr_numreqs turns to 0, otherwise
211 * it's possible that service thread will see dr_numreqs
212 * is zero, but dr_elapsed is not set yet, leading to lost
213 * data in this processing and an assertion in a subsequent
216 if (atomic_read(&iobuf->dr_numreqs) == 1) {
217 ktime_t now = ktime_get();
219 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
220 iobuf->dr_elapsed_valid = 1;
222 if (atomic_dec_and_test(&iobuf->dr_numreqs))
223 wake_up(&iobuf->dr_wait);
225 /* Completed bios used to be chained off iobuf->dr_bios and freed in
226 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
227 * mempool when serious on-disk fragmentation was encountered,
228 * deadlocking the OST. The bios are now released as soon as complete
229 * so the pool cannot be exhausted while IOs are competing. bug 10076 */
233 static void record_start_io(struct osd_iobuf *iobuf, int size)
235 struct osd_device *osd = iobuf->dr_dev;
236 struct obd_histogram *h = osd->od_brw_stats.hist;
239 atomic_inc(&iobuf->dr_numreqs);
241 if (iobuf->dr_rw == 0) {
242 atomic_inc(&osd->od_r_in_flight);
243 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
244 atomic_read(&osd->od_r_in_flight));
245 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
246 } else if (iobuf->dr_rw == 1) {
247 atomic_inc(&osd->od_w_in_flight);
248 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
249 atomic_read(&osd->od_w_in_flight));
250 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
256 static void osd_submit_bio(int rw, struct bio *bio)
258 LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
259 #ifdef HAVE_SUBMIT_BIO_2ARGS
260 submit_bio(rw ? WRITE : READ, bio);
267 static int can_be_merged(struct bio *bio, sector_t sector)
272 return bio_end_sector(bio) == sector ? 1 : 0;
275 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
277 * This function will change the data written, thus it should only be
278 * used when checking data integrity feature
280 static void bio_integrity_fault_inject(struct bio *bio)
282 struct bio_vec *bvec;
283 DECLARE_BVEC_ITER_ALL(iter_all);
287 bio_for_each_segment_all(bvec, bio, iter_all) {
288 struct page *page = bvec->bv_page;
298 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
299 unsigned int sectors, int tuple_size)
301 __u16 *expected_guard;
305 expected_guard = expected_guard_buf;
306 for (i = 0; i < sectors; i++) {
307 bio_guard = (__u16 *)bio_prot_buf;
308 if (*bio_guard != *expected_guard) {
309 CERROR("unexpected guard tags on sector %d "
310 "expected guard %u, bio guard "
311 "%u, sectors %u, tuple size %d\n",
312 i, *expected_guard, *bio_guard, sectors,
317 bio_prot_buf += tuple_size;
322 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
323 struct osd_iobuf *iobuf, int index)
325 struct blk_integrity *bi = bdev_get_integrity(bdev);
326 struct bio_integrity_payload *bip = bio->bi_integrity;
327 struct niobuf_local *lnb;
328 unsigned short sector_size = blk_integrity_interval(bi);
329 void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
330 bip->bip_vec->bv_offset;
332 sector_t sector = bio_start_sector(bio);
333 unsigned int sectors, total;
334 DECLARE_BVEC_ITER_ALL(iter_all);
335 __u16 *expected_guard;
339 bio_for_each_segment_all(bv, bio, iter_all) {
340 lnb = iobuf->dr_lnbs[index];
341 expected_guard = lnb->lnb_guards;
342 sectors = bv->bv_len / sector_size;
343 if (lnb->lnb_guard_rpc) {
344 rc = bio_dif_compare(expected_guard, bio_prot_buf,
345 sectors, bi->tuple_size);
351 bio_prot_buf += sectors * bi->tuple_size;
352 total += sectors * bi->tuple_size;
353 LASSERT(total <= bip_size(bio->bi_integrity));
359 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
360 struct osd_iobuf *iobuf,
361 int start_page_idx, bool fault_inject,
362 bool integrity_enabled)
364 struct super_block *sb = osd_sb(osd);
365 integrity_gen_fn *generate_fn = NULL;
366 integrity_vrfy_fn *verify_fn = NULL;
371 if (!integrity_enabled)
374 rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
378 rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
382 /* Verify and inject fault only when writing */
383 if (iobuf->dr_rw == 1) {
384 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
385 rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
391 if (unlikely(fault_inject))
392 bio_integrity_fault_inject(bio);
398 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
399 # ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
400 static void dio_integrity_complete_routine(struct bio *bio)
402 static void dio_integrity_complete_routine(struct bio *bio, int error)
405 struct osd_bio_private *bio_private = bio->bi_private;
407 bio->bi_private = bio_private->obp_iobuf;
408 osd_dio_complete_routine(bio, error);
410 OBD_FREE_PTR(bio_private);
412 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
413 #else /* !CONFIG_BLK_DEV_INTEGRITY */
414 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
415 fault_inject, integrity_enabled) 0
416 #endif /* CONFIG_BLK_DEV_INTEGRITY */
418 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
419 bool integrity_enabled, int start_page_idx,
420 struct osd_bio_private **pprivate)
426 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
427 if (integrity_enabled) {
428 struct osd_bio_private *bio_private = NULL;
430 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
431 if (bio_private == NULL)
433 bio->bi_end_io = dio_integrity_complete_routine;
434 bio->bi_private = bio_private;
435 bio_private->obp_start_page_idx = start_page_idx;
436 bio_private->obp_iobuf = iobuf;
437 *pprivate = bio_private;
441 bio->bi_end_io = dio_complete_routine;
442 bio->bi_private = iobuf;
448 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
449 struct osd_iobuf *iobuf)
451 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
452 struct page **pages = iobuf->dr_pages;
453 int npages = iobuf->dr_npages;
454 sector_t *blocks = iobuf->dr_blocks;
455 int total_blocks = npages * blocks_per_page;
456 struct super_block *sb = inode->i_sb;
457 int sector_bits = sb->s_blocksize_bits - 9;
458 unsigned int blocksize = sb->s_blocksize;
459 struct block_device *bdev = sb->s_bdev;
460 struct osd_bio_private *bio_private = NULL;
461 struct bio *bio = NULL;
462 int bio_start_page_idx;
464 unsigned int page_offset;
472 bool integrity_enabled;
473 struct blk_plug plug;
476 fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
477 LASSERT(iobuf->dr_npages == npages);
479 integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
481 osd_brw_stats_update(osd, iobuf);
482 iobuf->dr_start_time = ktime_get();
484 blk_start_plug(&plug);
485 for (page_idx = 0, block_idx = 0;
487 page_idx++, block_idx += blocks_per_page) {
489 page = pages[page_idx];
490 LASSERT(block_idx + blocks_per_page <= total_blocks);
492 for (i = 0, page_offset = 0;
494 i += nblocks, page_offset += blocksize * nblocks) {
498 if (blocks[block_idx + i] == 0) { /* hole */
499 LASSERTF(iobuf->dr_rw == 0,
500 "page_idx %u, block_idx %u, i %u\n",
501 page_idx, block_idx, i);
502 memset(kmap(page) + page_offset, 0, blocksize);
507 sector = (sector_t)blocks[block_idx + i] << sector_bits;
509 /* Additional contiguous file blocks? */
510 while (i + nblocks < blocks_per_page &&
511 (sector + (nblocks << sector_bits)) ==
512 ((sector_t)blocks[block_idx + i + nblocks] <<
517 can_be_merged(bio, sector) &&
518 bio_add_page(bio, page,
519 blocksize * nblocks, page_offset) != 0)
520 continue; /* added this frag OK */
523 struct request_queue *q = bio_get_queue(bio);
524 unsigned int bi_size = bio_sectors(bio) << 9;
526 /* Dang! I have to fragment this I/O */
528 "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
529 bi_size, bio->bi_vcnt, bio->bi_max_vecs,
531 queue_max_sectors(q),
532 osd_bio_nr_segs(bio),
533 queue_max_segments(q));
534 rc = osd_bio_integrity_handle(osd, bio,
535 iobuf, bio_start_page_idx,
536 fault_inject, integrity_enabled);
542 record_start_io(iobuf, bi_size);
543 osd_submit_bio(iobuf->dr_rw, bio);
546 bio_start_page_idx = page_idx;
547 /* allocate new bio */
548 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
549 (npages - page_idx) *
552 CERROR("Can't allocate bio %u*%u = %u pages\n",
553 (npages - page_idx), blocks_per_page,
554 (npages - page_idx) * blocks_per_page);
559 bio_set_dev(bio, bdev);
560 bio_set_sector(bio, sector);
561 bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
562 rc = osd_bio_init(bio, iobuf, integrity_enabled,
563 bio_start_page_idx, &bio_private);
569 rc = bio_add_page(bio, page,
570 blocksize * nblocks, page_offset);
576 rc = osd_bio_integrity_handle(osd, bio, iobuf,
585 record_start_io(iobuf, bio_sectors(bio) << 9);
586 osd_submit_bio(iobuf->dr_rw, bio);
591 blk_finish_plug(&plug);
593 /* in order to achieve better IO throughput, we don't wait for writes
594 * completion here. instead we proceed with transaction commit in
595 * parallel and wait for IO completion once transaction is stopped
596 * see osd_trans_stop() for more details -bzzz */
597 if (iobuf->dr_rw == 0 || fault_inject) {
598 wait_event(iobuf->dr_wait,
599 atomic_read(&iobuf->dr_numreqs) == 0);
600 osd_fini_iobuf(osd, iobuf);
604 rc = iobuf->dr_error;
607 OBD_FREE_PTR(bio_private);
613 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
614 struct niobuf_local *lnb, int maxlnb)
622 int poff = offset & (PAGE_SIZE - 1);
623 int plen = PAGE_SIZE - poff;
625 if (*nrpages >= maxlnb) {
632 lnb->lnb_file_offset = offset;
633 lnb->lnb_page_offset = poff;
635 /* lnb->lnb_flags = rnb->rnb_flags; */
637 lnb->lnb_page = NULL;
639 lnb->lnb_guard_rpc = 0;
640 lnb->lnb_guard_disk = 0;
643 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
654 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
655 loff_t offset, gfp_t gfp_mask, bool cache)
657 struct osd_thread_info *oti = osd_oti_get(env);
658 struct inode *inode = osd_dt_obj(dt)->oo_inode;
659 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
666 page = find_or_create_page(inode->i_mapping,
667 offset >> PAGE_SHIFT, gfp_mask);
670 LASSERT(!test_bit(PG_private_2, &page->flags));
672 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
677 if (inode->i_mapping->nrpages) {
678 /* consult with pagecache, but do not create new pages */
679 /* this is normally used once */
680 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
685 LASSERT(oti->oti_dio_pages);
686 cur = oti->oti_dio_pages_used;
688 if (unlikely(!oti->oti_dio_pages[cur])) {
689 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
690 page = alloc_page(gfp_mask);
693 oti->oti_dio_pages[cur] = page;
696 page = oti->oti_dio_pages[cur];
697 LASSERT(!test_bit(PG_private_2, &page->flags));
698 set_bit(PG_private_2, &page->flags);
699 oti->oti_dio_pages_used++;
701 LASSERT(!PageLocked(page));
704 LASSERT(!page->mapping);
705 LASSERT(!PageWriteback(page));
706 ClearPageUptodate(page);
708 page->index = offset >> PAGE_SHIFT;
714 * there are following "locks":
725 * - lock pages, unlock
727 * - lock partial page
733 * Unlock and release pages loaded by osd_bufs_get()
735 * Unlock \a npages pages from \a lnb and drop the refcount on them.
737 * \param env thread execution environment
738 * \param dt dt object undergoing IO (OSD object + methods)
739 * \param lnb array of pages undergoing IO
740 * \param npages number of pages in \a lnb
744 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
745 struct niobuf_local *lnb, int npages)
747 struct osd_thread_info *oti = osd_oti_get(env);
751 ll_pagevec_init(&pvec, 0);
753 for (i = 0; i < npages; i++) {
754 struct page *page = lnb[i].lnb_page;
759 /* if the page isn't cached, then reset uptodate
760 * to prevent reuse */
761 if (test_bit(PG_private_2, &page->flags)) {
762 clear_bit(PG_private_2, &page->flags);
763 ClearPageUptodate(page);
764 if (lnb[i].lnb_locked)
766 oti->oti_dio_pages_used--;
768 if (lnb[i].lnb_locked)
770 if (pagevec_add(&pvec, page) == 0)
771 pagevec_release(&pvec);
774 lnb[i].lnb_page = NULL;
777 LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
779 /* Release any partial pagevec */
780 pagevec_release(&pvec);
786 * Load and lock pages undergoing IO
788 * Pages as described in the \a lnb array are fetched (from disk or cache)
789 * and locked for IO by the caller.
791 * DLM locking protects us from write and truncate competing for same region,
792 * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
793 * It's possible the writeout on a such a page is in progress when we access
794 * it. It's also possible that during this writeout we put new (partial) data
795 * into the page, but won't be able to proceed in filter_commitrw_write().
796 * Therefore, just wait for writeout completion as it should be rare enough.
798 * \param env thread execution environment
799 * \param dt dt object undergoing IO (OSD object + methods)
800 * \param pos byte offset of IO start
801 * \param len number of bytes of IO
802 * \param lnb array of extents undergoing IO
803 * \param rw read or write operation, and other flags
804 * \param capa capabilities
806 * \retval pages (zero or more) loaded successfully
807 * \retval -ENOMEM on memory/page allocation error
809 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
810 loff_t pos, ssize_t len, struct niobuf_local *lnb,
811 int maxlnb, enum dt_bufs_type rw)
813 struct osd_thread_info *oti = osd_oti_get(env);
814 struct osd_object *obj = osd_dt_obj(dt);
815 struct osd_device *osd = osd_obj2dev(obj);
816 int npages, i, iosize, rc = 0;
821 LASSERT(obj->oo_inode);
823 rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
827 write = rw & DT_BUFS_TYPE_WRITE;
829 fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
830 iosize = fsize - lnb[0].lnb_file_offset;
831 fsize = max(fsize, i_size_read(obj->oo_inode));
833 cache = rw & DT_BUFS_TYPE_READAHEAD;
837 cache = osd_use_page_cache(osd);
840 if (!osd->od_writethrough_cache) {
844 if (iosize > osd->od_writethrough_max_iosize) {
849 if (!osd->od_read_cache) {
853 if (iosize > osd->od_readcache_max_iosize) {
858 /* don't use cache on large files */
859 if (osd->od_readcache_max_filesize &&
860 fsize > osd->od_readcache_max_filesize)
866 if (!cache && unlikely(!oti->oti_dio_pages)) {
867 OBD_ALLOC(oti->oti_dio_pages,
868 sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
869 if (!oti->oti_dio_pages)
873 /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
874 gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
876 for (i = 0; i < npages; i++, lnb++) {
877 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
879 if (lnb->lnb_page == NULL)
880 GOTO(cleanup, rc = -ENOMEM);
883 wait_on_page_writeback(lnb->lnb_page);
884 BUG_ON(PageWriteback(lnb->lnb_page));
888 /* XXX: this version doesn't invalidate cached pages, but use them */
889 if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
890 /* do not allow data aliasing, invalidate pagecache */
891 /* XXX: can be quite expensive in mixed case */
892 invalidate_mapping_pages(obj->oo_inode->i_mapping,
893 lnb[0].lnb_file_offset >> PAGE_SHIFT,
894 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
902 osd_bufs_put(env, dt, lnb - i, i);
906 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
907 int pages, sector_t *blocks,
910 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
912 struct page *fp = NULL;
914 pgoff_t max_page_index;
915 handle_t *handle = NULL;
917 max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
919 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
920 inode->i_ino, pages, (*page)->index);
923 create = LDISKFS_GET_BLOCKS_CREATE;
924 handle = ldiskfs_journal_current_handle();
925 LASSERT(handle != NULL);
926 rc = osd_attach_jinode(inode);
930 /* pages are sorted already. so, we just have to find
931 * contig. space and process them properly */
933 long blen, total = 0;
934 struct ldiskfs_map_blocks map = { 0 };
936 if (fp == NULL) { /* start new extent */
941 } else if (fp->index + clen == (*page)->index) {
942 /* continue the extent */
948 if (fp->index + clen >= max_page_index)
949 GOTO(cleanup, rc = -EFBIG);
950 /* process found extent */
951 map.m_lblk = fp->index * blocks_per_page;
952 map.m_len = blen = clen * blocks_per_page;
954 rc = ldiskfs_map_blocks(handle, inode, &map, create);
957 for (; total < blen && c < map.m_len; c++, total++) {
959 *(blocks + total) = 0;
963 *(blocks + total) = map.m_pblk + c;
964 /* unmap any possible underlying
965 * metadata from the block device
966 * mapping. bug 6998. */
967 if ((map.m_flags & LDISKFS_MAP_NEW) &&
976 if (rc == 0 && total < blen) {
977 map.m_lblk = fp->index * blocks_per_page + total;
978 map.m_len = blen - total;
984 /* look for next extent */
986 blocks += blocks_per_page * clen;
992 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
993 struct niobuf_local *lnb, int npages)
995 struct osd_thread_info *oti = osd_oti_get(env);
996 struct osd_iobuf *iobuf = &oti->oti_iobuf;
997 struct inode *inode = osd_dt_obj(dt)->oo_inode;
998 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1007 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1008 if (unlikely(rc != 0))
1011 isize = i_size_read(inode);
1012 maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1014 start = ktime_get();
1015 for (i = 0; i < npages; i++) {
1018 * till commit the content of the page is undefined
1019 * we'll set it uptodate once bulk is done. otherwise
1020 * subsequent reads can access non-stable data
1022 ClearPageUptodate(lnb[i].lnb_page);
1024 if (lnb[i].lnb_len == PAGE_SIZE)
1027 if (maxidx >= lnb[i].lnb_page->index) {
1028 osd_iobuf_add_page(iobuf, &lnb[i]);
1031 char *p = kmap(lnb[i].lnb_page);
1033 off = lnb[i].lnb_page_offset;
1036 off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1039 memset(p + off, 0, PAGE_SIZE - off);
1040 kunmap(lnb[i].lnb_page);
1044 timediff = ktime_us_delta(end, start);
1045 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1047 if (iobuf->dr_npages) {
1048 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1050 iobuf->dr_blocks, 0);
1051 if (likely(rc == 0)) {
1052 rc = osd_do_bio(osd, inode, iobuf);
1053 /* do IO stats for preparation reads */
1054 osd_fini_iobuf(osd, iobuf);
1060 struct osd_fextent {
1063 unsigned int mapped:1;
1066 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1067 struct osd_fextent *cached_extent)
1069 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1070 sector_t block = offset >> inode->i_blkbits;
1072 struct fiemap_extent_info fei = { 0 };
1073 struct fiemap_extent fe = { 0 };
1074 mm_segment_t saved_fs;
1077 if (block >= cached_extent->start && block < cached_extent->end)
1078 return cached_extent->mapped;
1080 if (i_size_read(inode) == 0)
1083 /* Beyond EOF, must not be mapped */
1084 if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1087 fei.fi_extents_max = 1;
1088 fei.fi_extents_start = &fe;
1090 saved_fs = get_fs();
1092 rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1097 start = fe.fe_logical >> inode->i_blkbits;
1099 if (start > block) {
1100 cached_extent->start = block;
1101 cached_extent->end = start;
1102 cached_extent->mapped = 0;
1104 cached_extent->start = start;
1105 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1107 cached_extent->mapped = 1;
1110 return cached_extent->mapped;
1113 static int osd_declare_write_commit(const struct lu_env *env,
1114 struct dt_object *dt,
1115 struct niobuf_local *lnb, int npages,
1116 struct thandle *handle)
1118 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1119 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1120 struct osd_thandle *oh;
1127 long long quota_space = 0;
1128 struct osd_fextent extent = { 0 };
1129 enum osd_quota_local_flags local_flags = 0;
1130 enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1133 LASSERT(handle != NULL);
1134 oh = container_of0(handle, struct osd_thandle, ot_super);
1135 LASSERT(oh->ot_handle == NULL);
1139 /* calculate number of extents (probably better to pass nb) */
1140 for (i = 0; i < npages; i++) {
1141 if (i && lnb[i].lnb_file_offset !=
1142 lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1145 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1146 lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1148 quota_space += PAGE_SIZE;
1150 /* ignore quota for the whole request if any page is from
1151 * client cache or written by root.
1153 * XXX once we drop the 1.8 client support, the checking
1154 * for whether page is from cache can be simplified as:
1155 * !(lnb[i].flags & OBD_BRW_SYNC)
1157 * XXX we could handle this on per-lnb basis as done by
1159 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1160 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1162 declare_flags |= OSD_QID_FORCE;
1166 * each extent can go into new leaf causing a split
1167 * 5 is max tree depth: inode + 4 index blocks
1168 * with blockmaps, depth is 3 at most
1170 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1172 * many concurrent threads may grow tree by the time
1173 * our transaction starts. so, consider 2 is a min depth
1175 depth = ext_depth(inode);
1176 depth = max(depth, 1) + 1;
1178 credits++; /* inode */
1179 credits += depth * 2 * extents;
1183 credits++; /* inode */
1184 credits += depth * extents;
1187 /* quota space for metadata blocks */
1188 quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1190 /* quota space should be reported in 1K blocks */
1191 quota_space = toqb(quota_space);
1193 /* each new block can go in different group (bitmap + gd) */
1195 /* we can't dirty more bitmap blocks than exist */
1196 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1197 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1199 credits += newblocks;
1201 /* we can't dirty more gd blocks than exist */
1202 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1203 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1205 credits += newblocks;
1207 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1209 /* make sure the over quota flags were not set */
1210 lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1212 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1213 i_projid_read(inode), quota_space, oh,
1214 osd_dt_obj(dt), &local_flags, declare_flags);
1216 /* we need only to store the overquota flags in the first lnb for
1217 * now, once we support multiple objects BRW, this code needs be
1219 if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1220 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1221 if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1222 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1223 if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1224 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1227 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1232 /* Check if a block is allocated or not */
1233 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1234 struct niobuf_local *lnb, int npages,
1235 struct thandle *thandle)
1237 struct osd_thread_info *oti = osd_oti_get(env);
1238 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1239 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1240 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1246 rc = osd_init_iobuf(osd, iobuf, 1, npages);
1247 if (unlikely(rc != 0))
1250 isize = i_size_read(inode);
1251 dquot_initialize(inode);
1253 for (i = 0; i < npages; i++) {
1254 if (lnb[i].lnb_rc == -ENOSPC &&
1255 (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1256 /* Allow the write to proceed if overwriting an
1261 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1262 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1264 LASSERT(lnb[i].lnb_page);
1265 generic_error_remove_page(inode->i_mapping,
1270 LASSERT(PageLocked(lnb[i].lnb_page));
1271 LASSERT(!PageWriteback(lnb[i].lnb_page));
1273 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1274 isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1277 * Since write and truncate are serialized by oo_sem, even
1278 * partial-page truncate should not leave dirty pages in the
1281 LASSERT(!PageDirty(lnb[i].lnb_page));
1283 SetPageUptodate(lnb[i].lnb_page);
1285 osd_iobuf_add_page(iobuf, &lnb[i]);
1288 osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1290 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1292 } else if (iobuf->dr_npages > 0) {
1293 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1295 iobuf->dr_blocks, 1);
1297 /* no pages to write, no transno is needed */
1298 thandle->th_local = 1;
1301 if (likely(rc == 0)) {
1302 spin_lock(&inode->i_lock);
1303 if (isize > i_size_read(inode)) {
1304 i_size_write(inode, isize);
1305 LDISKFS_I(inode)->i_disksize = isize;
1306 spin_unlock(&inode->i_lock);
1307 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1309 spin_unlock(&inode->i_lock);
1312 rc = osd_do_bio(osd, inode, iobuf);
1313 /* we don't do stats here as in read path because
1314 * write is async: we'll do this in osd_put_bufs() */
1316 osd_fini_iobuf(osd, iobuf);
1319 osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1321 if (unlikely(rc != 0)) {
1322 /* if write fails, we should drop pages from the cache */
1323 for (i = 0; i < npages; i++) {
1324 if (lnb[i].lnb_page == NULL)
1326 LASSERT(PageLocked(lnb[i].lnb_page));
1327 generic_error_remove_page(inode->i_mapping,
1335 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1336 struct niobuf_local *lnb, int npages)
1338 struct osd_thread_info *oti = osd_oti_get(env);
1339 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1340 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1341 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1342 int rc = 0, i, cache_hits = 0, cache_misses = 0;
1349 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1350 if (unlikely(rc != 0))
1353 isize = i_size_read(inode);
1355 start = ktime_get();
1356 for (i = 0; i < npages; i++) {
1358 if (isize <= lnb[i].lnb_file_offset)
1359 /* If there's no more data, abort early.
1360 * lnb->lnb_rc == 0, so it's easy to detect later. */
1363 /* instead of looking if we go beyong isize, send complete
1364 * pages all the time
1366 lnb[i].lnb_rc = lnb[i].lnb_len;
1368 /* Bypass disk read if fail_loc is set properly */
1369 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1370 SetPageUptodate(lnb[i].lnb_page);
1372 if (PageUptodate(lnb[i].lnb_page)) {
1374 unlock_page(lnb[i].lnb_page);
1377 osd_iobuf_add_page(iobuf, &lnb[i]);
1379 /* no need to unlock in osd_bufs_put(), the sooner page is
1380 * unlocked, the earlier another client can access it.
1381 * notice real unlock_page() can be called few lines
1382 * below after osd_do_bio(). lnb is a per-thread, so it's
1383 * fine to have PG_locked and lnb_locked inconsistent here */
1384 lnb[i].lnb_locked = 0;
1387 timediff = ktime_us_delta(end, start);
1388 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1390 if (cache_hits != 0)
1391 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1393 if (cache_misses != 0)
1394 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1396 if (cache_hits + cache_misses != 0)
1397 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1398 cache_hits + cache_misses);
1400 if (iobuf->dr_npages) {
1401 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1403 iobuf->dr_blocks, 0);
1404 rc = osd_do_bio(osd, inode, iobuf);
1406 /* IO stats will be done in osd_bufs_put() */
1408 /* early release to let others read data during the bulk */
1409 for (i = 0; i < iobuf->dr_npages; i++) {
1410 LASSERT(PageLocked(iobuf->dr_pages[i]));
1411 unlock_page(iobuf->dr_pages[i]);
1419 * XXX: Another layering violation for now.
1421 * We don't want to use ->f_op->read methods, because generic file write
1423 * - serializes on ->i_sem, and
1425 * - does a lot of extra work like balance_dirty_pages(),
1427 * which doesn't work for globally shared files like /last_rcvd.
1429 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1431 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1433 memcpy(buffer, (char *)ei->i_data, buflen);
1438 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1440 struct buffer_head *bh;
1441 unsigned long block;
1447 /* prevent reading after eof */
1448 spin_lock(&inode->i_lock);
1449 if (i_size_read(inode) < *offs + size) {
1450 loff_t diff = i_size_read(inode) - *offs;
1451 spin_unlock(&inode->i_lock);
1454 "size %llu is too short to read @%llu\n",
1455 i_size_read(inode), *offs);
1457 } else if (diff == 0) {
1463 spin_unlock(&inode->i_lock);
1466 blocksize = 1 << inode->i_blkbits;
1469 block = *offs >> inode->i_blkbits;
1470 boffs = *offs & (blocksize - 1);
1471 csize = min(blocksize - boffs, size);
1472 bh = __ldiskfs_bread(NULL, inode, block, 0);
1474 CERROR("%s: can't read %u@%llu on ino %lu: "
1475 "rc = %ld\n", osd_ino2name(inode),
1476 csize, *offs, inode->i_ino,
1482 memcpy(buf, bh->b_data + boffs, csize);
1485 memset(buf, 0, csize);
1495 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1496 struct lu_buf *buf, loff_t *pos)
1498 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1501 /* Read small symlink from inode body as we need to maintain correct
1502 * on-disk symlinks for ldiskfs.
1504 if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1505 loff_t size = i_size_read(inode);
1507 if (buf->lb_len < size)
1510 if (size < sizeof(LDISKFS_I(inode)->i_data))
1511 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1513 rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1515 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1521 static inline int osd_extents_enabled(struct super_block *sb,
1522 struct inode *inode)
1524 if (inode != NULL) {
1525 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1527 } else if (ldiskfs_has_feature_extents(sb)) {
1533 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1534 const loff_t size, const loff_t pos,
1537 int credits, bits, bs, i;
1539 bits = sb->s_blocksize_bits;
1542 /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1543 * we do not expect blockmaps on the large files,
1544 * so let's shrink it to 2 levels (4GB files) */
1546 /* this is default reservation: 2 levels */
1547 credits = (blocks + 2) * 3;
1549 /* actual offset is unknown, hard to optimize */
1553 /* now check for few specific cases to optimize */
1554 if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1557 /* allocate if not allocated */
1558 if (inode == NULL) {
1559 credits += blocks * 2;
1562 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1563 LASSERT(i < LDISKFS_NDIR_BLOCKS);
1564 if (LDISKFS_I(inode)->i_data[i] == 0)
1567 } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1568 /* single indirect */
1569 credits = blocks * 3;
1570 if (inode == NULL ||
1571 LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1574 /* The indirect block may be modified. */
1581 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1582 const struct lu_buf *buf, loff_t _pos,
1583 struct thandle *handle)
1585 struct osd_object *obj = osd_dt_obj(dt);
1586 struct inode *inode = obj->oo_inode;
1587 struct super_block *sb = osd_sb(osd_obj2dev(obj));
1588 struct osd_thandle *oh;
1589 int rc = 0, est = 0, credits, blocks, allocated = 0;
1595 LASSERT(buf != NULL);
1596 LASSERT(handle != NULL);
1598 oh = container_of0(handle, struct osd_thandle, ot_super);
1599 LASSERT(oh->ot_handle == NULL);
1602 bits = sb->s_blocksize_bits;
1606 /* if this is an append, then we
1607 * should expect cross-block record */
1613 /* blocks to modify */
1614 blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1615 LASSERT(blocks > 0);
1617 if (inode != NULL && _pos != -1) {
1618 /* object size in blocks */
1619 est = (i_size_read(inode) + bs - 1) >> bits;
1620 allocated = inode->i_blocks >> (bits - 9);
1621 if (pos + size <= i_size_read(inode) && est <= allocated) {
1622 /* looks like an overwrite, no need to modify tree */
1624 /* no need to modify i_size */
1629 if (osd_extents_enabled(sb, inode)) {
1631 * many concurrent threads may grow tree by the time
1632 * our transaction starts. so, consider 2 is a min depth
1633 * for every level we may need to allocate a new block
1634 * and take some entries from the old one. so, 3 blocks
1635 * to allocate (bitmap, gd, itself) + old block - 4 per
1638 depth = inode != NULL ? ext_depth(inode) : 0;
1639 depth = max(depth, 1) + 1;
1641 /* if not append, then split may need to modify
1642 * existing blocks moving entries into the new ones */
1645 /* blocks to store data: bitmap,gd,itself */
1646 credits += blocks * 3;
1648 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1650 /* if inode is created as part of the transaction,
1651 * then it's counted already by the creation method */
1657 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1659 /* dt_declare_write() is usually called for system objects, such
1660 * as llog or last_rcvd files. We needn't enforce quota on those
1661 * objects, so always set the lqi_space as 0. */
1663 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1665 i_projid_read(inode), 0,
1666 oh, obj, NULL, OSD_QID_BLK);
1669 rc = osd_trunc_lock(obj, oh, true);
1674 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1676 /* LU-2634: clear the extent format for fast symlink */
1677 ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1679 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1680 spin_lock(&inode->i_lock);
1681 LDISKFS_I(inode)->i_disksize = buflen;
1682 i_size_write(inode, buflen);
1683 spin_unlock(&inode->i_lock);
1684 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1689 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
1690 int bufsize, int write_NUL, loff_t *offs,
1693 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1694 struct buffer_head *bh = NULL;
1695 loff_t offset = *offs;
1696 loff_t new_size = i_size_read(inode);
1697 unsigned long block;
1698 int blocksize = 1 << inode->i_blkbits;
1702 int dirty_inode = 0;
1703 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1704 bool create, sparse, sync = false;
1708 * long symlink write does not count the NUL terminator in
1709 * bufsize, we write it, and the inode's file size does not
1710 * count the NUL terminator as well.
1712 ((char *)buf)[bufsize] = '\0';
1716 /* sparse checking is racy, but sparse is very rare case, leave as is */
1717 sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
1718 ((new_size - 1) >> inode->i_blkbits) + 1);
1720 while (bufsize > 0) {
1721 int credits = handle->h_buffer_credits;
1722 unsigned long last_block = (new_size == 0) ? 0 :
1723 (new_size - 1) >> inode->i_blkbits;
1728 block = offset >> inode->i_blkbits;
1729 boffs = offset & (blocksize - 1);
1730 size = min(blocksize - boffs, bufsize);
1731 sync = (block > last_block || new_size == 0 || sparse);
1734 down(&ei->i_append_sem);
1736 bh = __ldiskfs_bread(handle, inode, block, 0);
1738 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
1739 CWARN("%s: adding bh without locking off %llu (block %lu, "
1740 "size %d, offs %llu)\n", inode->i_sb->s_id,
1741 offset, block, bufsize, *offs);
1743 if (IS_ERR_OR_NULL(bh)) {
1744 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1745 int flags = LDISKFS_GET_BLOCKS_CREATE;
1747 /* while the file system is being mounted, avoid
1748 * preallocation otherwise mount can take a long
1749 * time as mballoc cache is cold.
1750 * XXX: this is a workaround until we have a proper
1752 * XXX: works with extent-based files only */
1753 if (!osd->od_cl_seq)
1754 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
1755 bh = __ldiskfs_bread(handle, inode, block, flags);
1759 up(&ei->i_append_sem);
1764 if (IS_ERR_OR_NULL(bh)) {
1772 CERROR("%s: error reading offset %llu (block %lu, "
1773 "size %d, offs %llu), credits %d/%d: rc = %d\n",
1774 inode->i_sb->s_id, offset, block, bufsize, *offs,
1775 credits, handle->h_buffer_credits, err);
1779 err = ldiskfs_journal_get_write_access(handle, bh);
1781 CERROR("journal_get_write_access() returned error %d\n",
1785 LASSERTF(boffs + size <= bh->b_size,
1786 "boffs %d size %d bh->b_size %lu\n",
1787 boffs, size, (unsigned long)bh->b_size);
1789 memset(bh->b_data, 0, bh->b_size);
1791 up(&ei->i_append_sem);
1795 memcpy(bh->b_data + boffs, buf, size);
1796 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1800 if (offset + size > new_size)
1801 new_size = offset + size;
1807 up(&ei->i_append_sem);
1814 /* correct in-core and on-disk sizes */
1815 if (new_size > i_size_read(inode)) {
1816 spin_lock(&inode->i_lock);
1817 if (new_size > i_size_read(inode))
1818 i_size_write(inode, new_size);
1819 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1820 LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1823 spin_unlock(&inode->i_lock);
1825 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1833 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1834 const struct lu_buf *buf, loff_t *pos,
1835 struct thandle *handle)
1837 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1838 struct osd_thandle *oh;
1842 LASSERT(dt_object_exists(dt));
1844 LASSERT(handle != NULL);
1845 LASSERT(inode != NULL);
1846 dquot_initialize(inode);
1848 /* XXX: don't check: one declared chunk can be used many times */
1849 /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1851 oh = container_of(handle, struct osd_thandle, ot_super);
1852 LASSERT(oh->ot_handle->h_transaction != NULL);
1853 osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1855 /* Write small symlink to inode body as we need to maintain correct
1856 * on-disk symlinks for ldiskfs.
1857 * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1858 * does not count it in.
1860 is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1861 if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1862 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1864 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
1865 is_link, pos, oh->ot_handle);
1867 result = buf->lb_len;
1869 osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1874 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1875 __u64 start, __u64 end, struct thandle *th)
1877 struct osd_thandle *oh;
1878 struct inode *inode;
1883 oh = container_of(th, struct osd_thandle, ot_super);
1886 * we don't need to reserve credits for whole truncate
1887 * it's not possible as truncate may need to free too many
1888 * blocks and that won't fit a single transaction. instead
1889 * we reserve credits to change i_size and put inode onto
1890 * orphan list. if needed truncate will extend or restart
1893 osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1894 osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1896 inode = osd_dt_obj(dt)->oo_inode;
1899 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1900 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1904 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
1909 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1910 __u64 start, __u64 end, struct thandle *th)
1912 struct osd_object *obj = osd_dt_obj(dt);
1913 struct osd_device *osd = osd_obj2dev(obj);
1914 struct inode *inode = obj->oo_inode;
1915 struct osd_access_lock *al;
1916 struct osd_thandle *oh;
1917 int rc = 0, found = 0;
1921 LASSERT(end == OBD_OBJECT_EOF);
1922 LASSERT(dt_object_exists(dt));
1923 LASSERT(osd_invariant(obj));
1924 LASSERT(inode != NULL);
1925 dquot_initialize(inode);
1928 oh = container_of(th, struct osd_thandle, ot_super);
1929 LASSERT(oh->ot_handle->h_transaction != NULL);
1931 /* we used to skip truncate to current size to
1932 * optimize truncates on OST. with DoM we can
1933 * get attr_set to set specific size (MDS_REINT)
1934 * and then get truncate RPC which essentially
1935 * would be skipped. this is bad.. so, disable
1936 * this optimization on MDS till the client stop
1937 * to sent MDS_REINT (LU-11033) -bzzz */
1938 if (osd->od_is_ost && i_size_read(inode) == start)
1941 osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1943 spin_lock(&inode->i_lock);
1944 if (i_size_read(inode) < start)
1946 i_size_write(inode, start);
1947 spin_unlock(&inode->i_lock);
1948 ll_truncate_pagecache(inode, start);
1950 /* optimize grow case */
1952 osd_execute_truncate(obj);
1957 /* add to orphan list to ensure truncate completion
1958 * if this transaction succeed. ldiskfs_truncate()
1959 * will take the inode out of the list */
1960 rc = ldiskfs_orphan_add(oh->ot_handle, inode);
1961 inode_unlock(inode);
1965 list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
1966 if (obj != al->tl_obj)
1968 LASSERT(al->tl_shared == 0);
1970 /* do actual truncate in osd_trans_stop() */
1971 al->tl_truncate = 1;
1980 static int fiemap_check_ranges(struct inode *inode,
1981 u64 start, u64 len, u64 *new_len)
1990 if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1991 maxbytes = inode->i_sb->s_maxbytes;
1993 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1995 if (start > maxbytes)
1999 * Shrink request scope to what the fs can actually handle.
2001 if (len > maxbytes || (maxbytes - len) < start)
2002 *new_len = maxbytes - start;
2007 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2008 #define FIEMAP_MAX_EXTENTS (UINT_MAX / sizeof(struct fiemap_extent))
2010 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2013 struct fiemap_extent_info fieinfo = {0, };
2014 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2017 mm_segment_t cur_fs;
2020 if (inode->i_op->fiemap == NULL)
2023 if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2026 rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2030 fieinfo.fi_flags = fm->fm_flags;
2031 fieinfo.fi_extents_max = fm->fm_extent_count;
2032 fieinfo.fi_extents_start = fm->fm_extents;
2034 if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2035 filemap_write_and_wait(inode->i_mapping);
2037 /* Save previous value address limit */
2039 /* Set the address limit of the kernel */
2042 rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2043 fm->fm_flags = fieinfo.fi_flags;
2044 fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2046 /* Restore the previous address limt */
2052 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2053 __u64 start, __u64 end, enum lu_ladvise_type advice)
2055 struct osd_object *obj = osd_dt_obj(dt);
2060 case LU_LADVISE_DONTNEED:
2062 invalidate_mapping_pages(obj->oo_inode->i_mapping,
2063 start >> PAGE_SHIFT,
2064 (end - 1) >> PAGE_SHIFT);
2075 * in some cases we may need declare methods for objects being created
2076 * e.g., when we create symlink
2078 const struct dt_body_operations osd_body_ops_new = {
2079 .dbo_declare_write = osd_declare_write,
2082 const struct dt_body_operations osd_body_ops = {
2083 .dbo_read = osd_read,
2084 .dbo_declare_write = osd_declare_write,
2085 .dbo_write = osd_write,
2086 .dbo_bufs_get = osd_bufs_get,
2087 .dbo_bufs_put = osd_bufs_put,
2088 .dbo_write_prep = osd_write_prep,
2089 .dbo_declare_write_commit = osd_declare_write_commit,
2090 .dbo_write_commit = osd_write_commit,
2091 .dbo_read_prep = osd_read_prep,
2092 .dbo_declare_punch = osd_declare_punch,
2093 .dbo_punch = osd_punch,
2094 .dbo_fiemap_get = osd_fiemap_get,
2095 .dbo_ladvise = osd_ladvise,
2099 * Get a truncate lock
2101 * In order to take multi-transaction truncate out of main transaction we let
2102 * the caller grab a lock on the object passed. the lock can be shared (for
2103 * writes) and exclusive (for truncate). It's not allowed to mix truncate
2104 * and write in the same transaction handle (do not confuse with big ldiskfs
2105 * transaction containing lots of handles).
2106 * The lock must be taken at declaration.
2108 * \param obj object to lock
2110 * \shared shared or exclusive
2112 * \retval 0 lock is granted
2113 * \retval -NOMEM no memory to allocate lock
2115 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2117 struct osd_access_lock *al, *tmp;
2122 list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2123 if (tmp->tl_obj != obj)
2125 LASSERT(tmp->tl_shared == shared);
2126 /* found same lock */
2131 if (unlikely(al == NULL))
2134 al->tl_truncate = false;
2136 down_read(&obj->oo_ext_idx_sem);
2138 down_write(&obj->oo_ext_idx_sem);
2139 al->tl_shared = shared;
2141 list_add(&al->tl_list, &oh->ot_trunc_locks);
2146 void osd_trunc_unlock_all(struct list_head *list)
2148 struct osd_access_lock *al, *tmp;
2149 list_for_each_entry_safe(al, tmp, list, tl_list) {
2151 up_read(&al->tl_obj->oo_ext_idx_sem);
2153 up_write(&al->tl_obj->oo_ext_idx_sem);
2154 list_del(&al->tl_list);
2159 void osd_execute_truncate(struct osd_object *obj)
2161 struct osd_device *d = osd_obj2dev(obj);
2162 struct inode *inode = obj->oo_inode;
2165 /* simulate crash before (in the middle) of delayed truncate */
2166 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2167 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2168 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2170 mutex_lock(&sbi->s_orphan_lock);
2171 list_del_init(&ei->i_orphan);
2172 mutex_unlock(&sbi->s_orphan_lock);
2177 ldiskfs_truncate(inode);
2178 inode_unlock(inode);
2181 * For a partial-page truncate, flush the page to disk immediately to
2182 * avoid data corruption during direct disk write. b=17397
2184 size = i_size_read(inode);
2185 if ((size & ~PAGE_MASK) == 0)
2187 if (osd_use_page_cache(d)) {
2188 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2190 /* Notice we use "wait" version to ensure I/O is complete */
2191 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2192 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2193 size >> PAGE_SHIFT);
2197 void osd_process_truncates(struct list_head *list)
2199 struct osd_access_lock *al;
2201 LASSERT(journal_current_handle() == NULL);
2203 list_for_each_entry(al, list, tl_list) {
2206 if (!al->tl_truncate)
2208 osd_execute_truncate(al->tl_obj);