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 int)(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 int)(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.hist[BRW_R_DIO_FRAGS+rw],
152 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
153 ktime_to_ms(iobuf->dr_elapsed));
157 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
158 static void dio_complete_routine(struct bio *bio)
160 int error = bio->bi_status;
162 static void dio_complete_routine(struct bio *bio, int error)
165 struct osd_iobuf *iobuf = bio->bi_private;
168 /* CAVEAT EMPTOR: possibly in IRQ context
169 * DO NOT record procfs stats here!!!
172 if (unlikely(iobuf == NULL)) {
173 CERROR("***** bio->bi_private is NULL! This should never happen. Normally, I would crash here, but instead I will dump the bio contents to the console. Please report this to <https://jira.whamcloud.com/> , along with any interesting messages leading up to this point (like SCSI errors, perhaps). Because bi_private is NULL, I can't wake up the thread that initiated this IO - you will probably have to reboot this node.\n");
174 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
175 ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
176 bio->bi_next, (unsigned long)bio->bi_flags,
177 (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
178 bio_sectors(bio) << 9, bio->bi_end_io,
179 atomic_read(&bio->__bi_cnt),
184 /* the check is outside of the cycle for performance reason -bzzz */
185 if (!bio_data_dir(bio)) {
186 DECLARE_BVEC_ITER_ALL(iter_all);
188 bio_for_each_segment_all(bvl, bio, iter_all) {
189 if (likely(error == 0))
190 SetPageUptodate(bvl_to_page(bvl));
191 LASSERT(PageLocked(bvl_to_page(bvl)));
193 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
195 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
198 /* any real error is good enough -bzzz */
199 if (error != 0 && iobuf->dr_error == 0)
200 iobuf->dr_error = error;
203 * set dr_elapsed before dr_numreqs turns to 0, otherwise
204 * it's possible that service thread will see dr_numreqs
205 * is zero, but dr_elapsed is not set yet, leading to lost
206 * data in this processing and an assertion in a subsequent
209 if (atomic_read(&iobuf->dr_numreqs) == 1) {
210 ktime_t now = ktime_get();
212 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
213 iobuf->dr_elapsed_valid = 1;
215 if (atomic_dec_and_test(&iobuf->dr_numreqs))
216 wake_up(&iobuf->dr_wait);
218 /* Completed bios used to be chained off iobuf->dr_bios and freed in
219 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
220 * mempool when serious on-disk fragmentation was encountered,
221 * deadlocking the OST. The bios are now released as soon as complete
222 * so the pool cannot be exhausted while IOs are competing. b=10076
227 static void record_start_io(struct osd_iobuf *iobuf, int size)
229 struct osd_device *osd = iobuf->dr_dev;
230 struct obd_histogram *h = osd->od_brw_stats.hist;
233 atomic_inc(&iobuf->dr_numreqs);
235 if (iobuf->dr_rw == 0) {
236 atomic_inc(&osd->od_r_in_flight);
237 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
238 atomic_read(&osd->od_r_in_flight));
239 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
240 } else if (iobuf->dr_rw == 1) {
241 atomic_inc(&osd->od_w_in_flight);
242 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
243 atomic_read(&osd->od_w_in_flight));
244 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
250 static void osd_submit_bio(int rw, struct bio *bio)
252 LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
253 #ifdef HAVE_SUBMIT_BIO_2ARGS
254 submit_bio(rw ? WRITE : READ, bio);
261 static int can_be_merged(struct bio *bio, sector_t sector)
266 return bio_end_sector(bio) == sector ? 1 : 0;
269 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
271 * This function will change the data written, thus it should only be
272 * used when checking data integrity feature
274 static void bio_integrity_fault_inject(struct bio *bio)
276 struct bio_vec *bvec;
277 DECLARE_BVEC_ITER_ALL(iter_all);
281 bio_for_each_segment_all(bvec, bio, iter_all) {
282 struct page *page = bvec->bv_page;
292 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
293 unsigned int sectors, int tuple_size)
295 __u16 *expected_guard;
299 expected_guard = expected_guard_buf;
300 for (i = 0; i < sectors; i++) {
301 bio_guard = (__u16 *)bio_prot_buf;
302 if (*bio_guard != *expected_guard) {
304 "unexpected guard tags on sector %d expected guard %u, bio guard %u, sectors %u, tuple size %d\n",
305 i, *expected_guard, *bio_guard, sectors,
310 bio_prot_buf += tuple_size;
315 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
316 struct osd_iobuf *iobuf, int index)
318 struct blk_integrity *bi = bdev_get_integrity(bdev);
319 struct bio_integrity_payload *bip = bio->bi_integrity;
320 struct niobuf_local *lnb;
321 unsigned short sector_size = blk_integrity_interval(bi);
322 void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
323 bip->bip_vec->bv_offset;
325 sector_t sector = bio_start_sector(bio);
326 unsigned int sectors, total;
327 DECLARE_BVEC_ITER_ALL(iter_all);
328 __u16 *expected_guard;
332 bio_for_each_segment_all(bv, bio, iter_all) {
333 lnb = iobuf->dr_lnbs[index];
334 expected_guard = lnb->lnb_guards;
335 sectors = bv->bv_len / sector_size;
336 if (lnb->lnb_guard_rpc) {
337 rc = bio_dif_compare(expected_guard, bio_prot_buf,
338 sectors, bi->tuple_size);
344 bio_prot_buf += sectors * bi->tuple_size;
345 total += sectors * bi->tuple_size;
346 LASSERT(total <= bip_size(bio->bi_integrity));
352 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
353 struct osd_iobuf *iobuf,
354 int start_page_idx, bool fault_inject,
355 bool integrity_enabled)
357 struct super_block *sb = osd_sb(osd);
358 integrity_gen_fn *generate_fn = NULL;
359 integrity_vrfy_fn *verify_fn = NULL;
364 if (!integrity_enabled)
367 rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
371 rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
375 /* Verify and inject fault only when writing */
376 if (iobuf->dr_rw == 1) {
377 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
378 rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
384 if (unlikely(fault_inject))
385 bio_integrity_fault_inject(bio);
391 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
392 # ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
393 static void dio_integrity_complete_routine(struct bio *bio)
395 static void dio_integrity_complete_routine(struct bio *bio, int error)
398 struct osd_bio_private *bio_private = bio->bi_private;
400 bio->bi_private = bio_private->obp_iobuf;
401 osd_dio_complete_routine(bio, error);
403 OBD_FREE_PTR(bio_private);
405 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
406 #else /* !CONFIG_BLK_DEV_INTEGRITY */
407 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
408 fault_inject, integrity_enabled) 0
409 #endif /* CONFIG_BLK_DEV_INTEGRITY */
411 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
412 bool integrity_enabled, int start_page_idx,
413 struct osd_bio_private **pprivate)
419 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
420 if (integrity_enabled) {
421 struct osd_bio_private *bio_private = NULL;
423 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
424 if (bio_private == NULL)
426 bio->bi_end_io = dio_integrity_complete_routine;
427 bio->bi_private = bio_private;
428 bio_private->obp_start_page_idx = start_page_idx;
429 bio_private->obp_iobuf = iobuf;
430 *pprivate = bio_private;
434 bio->bi_end_io = dio_complete_routine;
435 bio->bi_private = iobuf;
441 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
442 struct osd_iobuf *iobuf)
444 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
445 struct page **pages = iobuf->dr_pages;
446 int npages = iobuf->dr_npages;
447 sector_t *blocks = iobuf->dr_blocks;
448 int total_blocks = npages * blocks_per_page;
449 struct super_block *sb = inode->i_sb;
450 int sector_bits = sb->s_blocksize_bits - 9;
451 unsigned int blocksize = sb->s_blocksize;
452 struct block_device *bdev = sb->s_bdev;
453 struct osd_bio_private *bio_private = NULL;
454 struct bio *bio = NULL;
455 int bio_start_page_idx;
457 unsigned int page_offset;
465 bool integrity_enabled;
466 struct blk_plug plug;
469 fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
470 LASSERT(iobuf->dr_npages == npages);
472 integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
474 osd_brw_stats_update(osd, iobuf);
475 iobuf->dr_start_time = ktime_get();
477 blk_start_plug(&plug);
478 for (page_idx = 0, block_idx = 0;
480 page_idx++, block_idx += blocks_per_page) {
481 page = pages[page_idx];
482 LASSERT(block_idx + blocks_per_page <= total_blocks);
484 for (i = 0, page_offset = 0;
486 i += nblocks, page_offset += blocksize * nblocks) {
489 if (blocks[block_idx + i] == 0) { /* hole */
490 LASSERTF(iobuf->dr_rw == 0,
491 "page_idx %u, block_idx %u, i %u\n",
492 page_idx, block_idx, i);
493 memset(kmap(page) + page_offset, 0, blocksize);
498 sector = (sector_t)blocks[block_idx + i] << sector_bits;
500 /* Additional contiguous file blocks? */
501 while (i + nblocks < blocks_per_page &&
502 (sector + (nblocks << sector_bits)) ==
503 ((sector_t)blocks[block_idx + i + nblocks] <<
507 if (bio && can_be_merged(bio, sector) &&
508 bio_add_page(bio, page, blocksize * nblocks,
510 continue; /* added this frag OK */
513 struct request_queue *q = bio_get_queue(bio);
514 unsigned int bi_size = bio_sectors(bio) << 9;
516 /* Dang! I have to fragment this I/O */
518 "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
519 bi_size, bio->bi_vcnt, bio->bi_max_vecs,
521 queue_max_sectors(q),
522 osd_bio_nr_segs(bio),
523 queue_max_segments(q));
524 rc = osd_bio_integrity_handle(osd, bio,
525 iobuf, bio_start_page_idx,
526 fault_inject, integrity_enabled);
532 record_start_io(iobuf, bi_size);
533 osd_submit_bio(iobuf->dr_rw, bio);
536 bio_start_page_idx = page_idx;
537 /* allocate new bio */
538 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
539 (npages - page_idx) *
542 CERROR("Can't allocate bio %u*%u = %u pages\n",
543 (npages - page_idx), blocks_per_page,
544 (npages - page_idx) * blocks_per_page);
549 bio_set_dev(bio, bdev);
550 bio_set_sector(bio, sector);
551 bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
552 rc = osd_bio_init(bio, iobuf, integrity_enabled,
553 bio_start_page_idx, &bio_private);
559 rc = bio_add_page(bio, page,
560 blocksize * nblocks, page_offset);
566 rc = osd_bio_integrity_handle(osd, bio, iobuf,
575 record_start_io(iobuf, bio_sectors(bio) << 9);
576 osd_submit_bio(iobuf->dr_rw, bio);
581 blk_finish_plug(&plug);
583 /* in order to achieve better IO throughput, we don't wait for writes
584 * completion here. instead we proceed with transaction commit in
585 * parallel and wait for IO completion once transaction is stopped
586 * see osd_trans_stop() for more details -bzzz
588 if (iobuf->dr_rw == 0 || fault_inject) {
589 wait_event(iobuf->dr_wait,
590 atomic_read(&iobuf->dr_numreqs) == 0);
591 osd_fini_iobuf(osd, iobuf);
595 rc = iobuf->dr_error;
598 OBD_FREE_PTR(bio_private);
604 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
605 struct niobuf_local *lnb, int maxlnb)
613 int poff = offset & (PAGE_SIZE - 1);
614 int plen = PAGE_SIZE - poff;
616 if (*nrpages >= maxlnb) {
623 lnb->lnb_file_offset = offset;
624 lnb->lnb_page_offset = poff;
626 /* lnb->lnb_flags = rnb->rnb_flags; */
628 lnb->lnb_page = NULL;
630 lnb->lnb_guard_rpc = 0;
631 lnb->lnb_guard_disk = 0;
634 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
645 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
646 loff_t offset, gfp_t gfp_mask, bool cache)
648 struct osd_thread_info *oti = osd_oti_get(env);
649 struct inode *inode = osd_dt_obj(dt)->oo_inode;
650 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
657 page = find_or_create_page(inode->i_mapping,
658 offset >> PAGE_SHIFT, gfp_mask);
661 LASSERT(!PagePrivate2(page));
662 wait_on_page_writeback(page);
664 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
670 if (inode->i_mapping->nrpages) {
671 /* consult with pagecache, but do not create new pages */
672 /* this is normally used once */
673 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
675 wait_on_page_writeback(page);
680 LASSERT(oti->oti_dio_pages);
681 cur = oti->oti_dio_pages_used;
682 page = oti->oti_dio_pages[cur];
684 if (unlikely(!page)) {
685 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
686 page = alloc_page(gfp_mask);
689 oti->oti_dio_pages[cur] = page;
690 SetPagePrivate2(page);
694 ClearPageUptodate(page);
695 page->index = offset >> PAGE_SHIFT;
696 oti->oti_dio_pages_used++;
702 * there are following "locks":
713 * - lock pages, unlock
715 * - lock partial page
721 * Unlock and release pages loaded by osd_bufs_get()
723 * Unlock \a npages pages from \a lnb and drop the refcount on them.
725 * \param env thread execution environment
726 * \param dt dt object undergoing IO (OSD object + methods)
727 * \param lnb array of pages undergoing IO
728 * \param npages number of pages in \a lnb
732 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
733 struct niobuf_local *lnb, int npages)
735 struct osd_thread_info *oti = osd_oti_get(env);
739 ll_pagevec_init(&pvec, 0);
741 for (i = 0; i < npages; i++) {
742 struct page *page = lnb[i].lnb_page;
747 /* if the page isn't cached, then reset uptodate
750 if (PagePrivate2(page)) {
751 oti->oti_dio_pages_used--;
753 if (lnb[i].lnb_locked)
755 if (pagevec_add(&pvec, page) == 0)
756 pagevec_release(&pvec);
759 lnb[i].lnb_page = NULL;
762 LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
764 /* Release any partial pagevec */
765 pagevec_release(&pvec);
771 * Load and lock pages undergoing IO
773 * Pages as described in the \a lnb array are fetched (from disk or cache)
774 * and locked for IO by the caller.
776 * DLM locking protects us from write and truncate competing for same region,
777 * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
778 * It's possible the writeout on a such a page is in progress when we access
779 * it. It's also possible that during this writeout we put new (partial) data
780 * into the page, but won't be able to proceed in filter_commitrw_write().
781 * Therefore, just wait for writeout completion as it should be rare enough.
783 * \param env thread execution environment
784 * \param dt dt object undergoing IO (OSD object + methods)
785 * \param pos byte offset of IO start
786 * \param len number of bytes of IO
787 * \param lnb array of extents undergoing IO
788 * \param rw read or write operation, and other flags
789 * \param capa capabilities
791 * \retval pages (zero or more) loaded successfully
792 * \retval -ENOMEM on memory/page allocation error
794 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
795 loff_t pos, ssize_t len, struct niobuf_local *lnb,
796 int maxlnb, enum dt_bufs_type rw)
798 struct osd_thread_info *oti = osd_oti_get(env);
799 struct osd_object *obj = osd_dt_obj(dt);
800 struct osd_device *osd = osd_obj2dev(obj);
801 int npages, i, iosize, rc = 0;
806 LASSERT(obj->oo_inode);
808 rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
812 write = rw & DT_BUFS_TYPE_WRITE;
814 fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
815 iosize = fsize - lnb[0].lnb_file_offset;
816 fsize = max(fsize, i_size_read(obj->oo_inode));
818 cache = rw & DT_BUFS_TYPE_READAHEAD;
822 cache = osd_use_page_cache(osd);
825 if (!osd->od_writethrough_cache) {
829 if (iosize > osd->od_writethrough_max_iosize) {
834 if (!osd->od_read_cache) {
838 if (iosize > osd->od_readcache_max_iosize) {
843 /* don't use cache on large files */
844 if (osd->od_readcache_max_filesize &&
845 fsize > osd->od_readcache_max_filesize)
851 if (!cache && unlikely(!oti->oti_dio_pages)) {
852 OBD_ALLOC_PTR_ARRAY_LARGE(oti->oti_dio_pages,
853 PTLRPC_MAX_BRW_PAGES);
854 if (!oti->oti_dio_pages)
858 /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
859 gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
861 for (i = 0; i < npages; i++, lnb++) {
862 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
864 if (lnb->lnb_page == NULL)
865 GOTO(cleanup, rc = -ENOMEM);
871 /* XXX: this version doesn't invalidate cached pages, but use them */
872 if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
873 /* do not allow data aliasing, invalidate pagecache */
874 /* XXX: can be quite expensive in mixed case */
875 invalidate_mapping_pages(obj->oo_inode->i_mapping,
876 lnb[0].lnb_file_offset >> PAGE_SHIFT,
877 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
885 osd_bufs_put(env, dt, lnb - i, i);
889 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
890 int pages, sector_t *blocks,
893 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
895 struct page *fp = NULL;
897 pgoff_t max_page_index;
898 handle_t *handle = NULL;
900 max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
902 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
903 inode->i_ino, pages, (*page)->index);
906 create = LDISKFS_GET_BLOCKS_CREATE;
907 handle = ldiskfs_journal_current_handle();
908 LASSERT(handle != NULL);
909 rc = osd_attach_jinode(inode);
913 /* pages are sorted already. so, we just have to find
914 * contig. space and process them properly
917 long blen, total = 0;
918 struct ldiskfs_map_blocks map = { 0 };
920 if (fp == NULL) { /* start new extent */
925 } else if (fp->index + clen == (*page)->index) {
926 /* continue the extent */
932 if (fp->index + clen >= max_page_index)
933 GOTO(cleanup, rc = -EFBIG);
934 /* process found extent */
935 map.m_lblk = fp->index * blocks_per_page;
936 map.m_len = blen = clen * blocks_per_page;
938 rc = ldiskfs_map_blocks(handle, inode, &map, create);
942 for (; total < blen && c < map.m_len; c++, total++) {
944 *(blocks + total) = 0;
948 *(blocks + total) = map.m_pblk + c;
949 /* unmap any possible underlying
950 * metadata from the block device
953 if ((map.m_flags & LDISKFS_MAP_NEW) &&
955 clean_bdev_aliases(inode->i_sb->s_bdev,
960 if (rc == 0 && total < blen) {
961 map.m_lblk = fp->index * blocks_per_page + total;
962 map.m_len = blen - total;
968 /* look for next extent */
970 blocks += blocks_per_page * clen;
976 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
977 struct niobuf_local *lnb, int npages)
979 struct osd_thread_info *oti = osd_oti_get(env);
980 struct osd_iobuf *iobuf = &oti->oti_iobuf;
981 struct inode *inode = osd_dt_obj(dt)->oo_inode;
982 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
991 rc = osd_init_iobuf(osd, iobuf, 0, npages);
992 if (unlikely(rc != 0))
995 isize = i_size_read(inode);
996 maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
999 for (i = 0; i < npages; i++) {
1002 * till commit the content of the page is undefined
1003 * we'll set it uptodate once bulk is done. otherwise
1004 * subsequent reads can access non-stable data
1006 ClearPageUptodate(lnb[i].lnb_page);
1008 if (lnb[i].lnb_len == PAGE_SIZE)
1011 if (maxidx >= lnb[i].lnb_page->index) {
1012 osd_iobuf_add_page(iobuf, &lnb[i]);
1015 char *p = kmap(lnb[i].lnb_page);
1017 off = lnb[i].lnb_page_offset;
1020 off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1023 memset(p + off, 0, PAGE_SIZE - off);
1024 kunmap(lnb[i].lnb_page);
1028 timediff = ktime_us_delta(end, start);
1029 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1031 if (iobuf->dr_npages) {
1032 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1034 iobuf->dr_blocks, 0);
1035 if (likely(rc == 0)) {
1036 rc = osd_do_bio(osd, inode, iobuf);
1037 /* do IO stats for preparation reads */
1038 osd_fini_iobuf(osd, iobuf);
1044 struct osd_fextent {
1047 unsigned int mapped:1;
1050 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1051 struct osd_fextent *cached_extent)
1053 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1054 sector_t block = offset >> inode->i_blkbits;
1056 struct fiemap_extent_info fei = { 0 };
1057 struct fiemap_extent fe = { 0 };
1058 mm_segment_t saved_fs;
1061 if (block >= cached_extent->start && block < cached_extent->end)
1062 return cached_extent->mapped;
1064 if (i_size_read(inode) == 0)
1067 /* Beyond EOF, must not be mapped */
1068 if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1071 fei.fi_extents_max = 1;
1072 fei.fi_extents_start = &fe;
1074 saved_fs = get_fs();
1076 rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1081 start = fe.fe_logical >> inode->i_blkbits;
1083 if (start > block) {
1084 cached_extent->start = block;
1085 cached_extent->end = start;
1086 cached_extent->mapped = 0;
1088 cached_extent->start = start;
1089 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1091 cached_extent->mapped = 1;
1094 return cached_extent->mapped;
1097 static int osd_declare_write_commit(const struct lu_env *env,
1098 struct dt_object *dt,
1099 struct niobuf_local *lnb, int npages,
1100 struct thandle *handle)
1102 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1103 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1104 struct osd_thandle *oh;
1111 long long quota_space = 0;
1112 struct osd_fextent extent = { 0 };
1113 enum osd_quota_local_flags local_flags = 0;
1114 enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1117 LASSERT(handle != NULL);
1118 oh = container_of(handle, struct osd_thandle, ot_super);
1119 LASSERT(oh->ot_handle == NULL);
1123 /* calculate number of extents (probably better to pass nb) */
1124 for (i = 0; i < npages; i++) {
1125 if (i && lnb[i].lnb_file_offset !=
1126 lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1129 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1130 lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1132 quota_space += PAGE_SIZE;
1134 /* ignore quota for the whole request if any page is from
1135 * client cache or written by root.
1137 * XXX once we drop the 1.8 client support, the checking
1138 * for whether page is from cache can be simplified as:
1139 * !(lnb[i].flags & OBD_BRW_SYNC)
1141 * XXX we could handle this on per-lnb basis as done by
1144 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1145 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1147 declare_flags |= OSD_QID_FORCE;
1151 * each extent can go into new leaf causing a split
1152 * 5 is max tree depth: inode + 4 index blocks
1153 * with blockmaps, depth is 3 at most
1155 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1157 * many concurrent threads may grow tree by the time
1158 * our transaction starts. so, consider 2 is a min depth
1160 depth = ext_depth(inode);
1161 depth = max(depth, 1) + 1;
1163 credits++; /* inode */
1164 credits += depth * 2 * extents;
1168 credits++; /* inode */
1169 credits += depth * extents;
1172 /* quota space for metadata blocks */
1173 quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1175 /* quota space should be reported in 1K blocks */
1176 quota_space = toqb(quota_space);
1178 /* each new block can go in different group (bitmap + gd) */
1180 /* we can't dirty more bitmap blocks than exist */
1181 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1182 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1184 credits += newblocks;
1186 /* we can't dirty more gd blocks than exist */
1187 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1188 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1190 credits += newblocks;
1192 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1194 /* make sure the over quota flags were not set */
1195 lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1197 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1198 i_projid_read(inode), quota_space, oh,
1199 osd_dt_obj(dt), &local_flags, declare_flags);
1201 /* we need only to store the overquota flags in the first lnb for
1202 * now, once we support multiple objects BRW, this code needs be
1205 if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1206 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1207 if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1208 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1209 if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1210 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1213 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1218 /* Check if a block is allocated or not */
1219 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1220 struct niobuf_local *lnb, int npages,
1221 struct thandle *thandle, __u64 user_size)
1223 struct osd_thread_info *oti = osd_oti_get(env);
1224 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1225 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1226 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1232 rc = osd_init_iobuf(osd, iobuf, 1, npages);
1233 if (unlikely(rc != 0))
1236 disk_size = i_size_read(inode);
1237 /* if disk_size is already bigger than specified user_size,
1240 if (disk_size > user_size)
1242 dquot_initialize(inode);
1244 for (i = 0; i < npages; i++) {
1245 if (lnb[i].lnb_rc == -ENOSPC &&
1246 (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1247 /* Allow the write to proceed if overwriting an
1253 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1254 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1256 LASSERT(lnb[i].lnb_page);
1257 generic_error_remove_page(inode->i_mapping,
1262 LASSERT(PageLocked(lnb[i].lnb_page));
1263 LASSERT(!PageWriteback(lnb[i].lnb_page));
1265 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > disk_size)
1266 disk_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1269 * Since write and truncate are serialized by oo_sem, even
1270 * partial-page truncate should not leave dirty pages in the
1273 LASSERT(!PageDirty(lnb[i].lnb_page));
1275 SetPageUptodate(lnb[i].lnb_page);
1277 osd_iobuf_add_page(iobuf, &lnb[i]);
1280 /* if file has grown, take user_size into account */
1281 if (user_size && disk_size > user_size)
1282 disk_size = user_size;
1284 osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1286 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1288 } else if (iobuf->dr_npages > 0) {
1289 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1291 iobuf->dr_blocks, 1);
1293 /* no pages to write, no transno is needed */
1294 thandle->th_local = 1;
1297 if (likely(rc == 0)) {
1298 spin_lock(&inode->i_lock);
1299 if (disk_size > i_size_read(inode)) {
1300 i_size_write(inode, disk_size);
1301 LDISKFS_I(inode)->i_disksize = disk_size;
1302 spin_unlock(&inode->i_lock);
1303 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1305 spin_unlock(&inode->i_lock);
1308 rc = osd_do_bio(osd, inode, iobuf);
1309 /* we don't do stats here as in read path because
1310 * write is async: we'll do this in osd_put_bufs()
1313 osd_fini_iobuf(osd, iobuf);
1316 osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1318 if (unlikely(rc != 0)) {
1319 /* if write fails, we should drop pages from the cache */
1320 for (i = 0; i < npages; i++) {
1321 if (lnb[i].lnb_page == NULL)
1323 if (!PagePrivate2(lnb[i].lnb_page)) {
1324 LASSERT(PageLocked(lnb[i].lnb_page));
1325 generic_error_remove_page(inode->i_mapping,
1334 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1335 struct niobuf_local *lnb, int npages)
1337 struct osd_thread_info *oti = osd_oti_get(env);
1338 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1339 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1340 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1341 int rc = 0, i, cache_hits = 0, cache_misses = 0;
1348 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1349 if (unlikely(rc != 0))
1352 isize = i_size_read(inode);
1354 start = ktime_get();
1355 for (i = 0; i < npages; i++) {
1357 if (isize <= lnb[i].lnb_file_offset)
1358 /* If there's no more data, abort early.
1359 * 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
1385 lnb[i].lnb_locked = 0;
1388 timediff = ktime_us_delta(end, start);
1389 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1391 if (cache_hits != 0)
1392 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1394 if (cache_misses != 0)
1395 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1397 if (cache_hits + cache_misses != 0)
1398 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1399 cache_hits + cache_misses);
1401 if (iobuf->dr_npages) {
1402 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1404 iobuf->dr_blocks, 0);
1405 rc = osd_do_bio(osd, inode, iobuf);
1407 /* IO stats will be done in osd_bufs_put() */
1409 /* early release to let others read data during the bulk */
1410 for (i = 0; i < iobuf->dr_npages; i++) {
1411 LASSERT(PageLocked(iobuf->dr_pages[i]));
1412 if (!PagePrivate2(iobuf->dr_pages[i]))
1413 unlock_page(iobuf->dr_pages[i]);
1421 * XXX: Another layering violation for now.
1423 * We don't want to use ->f_op->read methods, because generic file write
1425 * - serializes on ->i_sem, and
1427 * - does a lot of extra work like balance_dirty_pages(),
1429 * which doesn't work for globally shared files like /last_rcvd.
1431 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1433 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1435 memcpy(buffer, (char *)ei->i_data, buflen);
1440 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1442 struct buffer_head *bh;
1443 unsigned long block;
1449 /* prevent reading after eof */
1450 spin_lock(&inode->i_lock);
1451 if (i_size_read(inode) < *offs + size) {
1452 loff_t diff = i_size_read(inode) - *offs;
1454 spin_unlock(&inode->i_lock);
1457 "size %llu is too short to read @%llu\n",
1458 i_size_read(inode), *offs);
1460 } else if (diff == 0) {
1466 spin_unlock(&inode->i_lock);
1469 blocksize = 1 << inode->i_blkbits;
1472 block = *offs >> inode->i_blkbits;
1473 boffs = *offs & (blocksize - 1);
1474 csize = min(blocksize - boffs, size);
1475 bh = __ldiskfs_bread(NULL, inode, block, 0);
1477 CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1478 osd_ino2name(inode), csize, *offs, inode->i_ino,
1484 memcpy(buf, bh->b_data + boffs, csize);
1487 memset(buf, 0, csize);
1497 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1498 struct lu_buf *buf, loff_t *pos)
1500 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1503 /* Read small symlink from inode body as we need to maintain correct
1504 * on-disk symlinks for ldiskfs.
1506 if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1507 loff_t size = i_size_read(inode);
1509 if (buf->lb_len < size)
1512 if (size < sizeof(LDISKFS_I(inode)->i_data))
1513 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1515 rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1517 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1523 static inline int osd_extents_enabled(struct super_block *sb,
1524 struct inode *inode)
1526 if (inode != NULL) {
1527 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1529 } else if (ldiskfs_has_feature_extents(sb)) {
1535 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1536 const loff_t size, const loff_t pos,
1539 int credits, bits, bs, i;
1541 bits = sb->s_blocksize_bits;
1544 /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1545 * we do not expect blockmaps on the large files,
1546 * so let's shrink it to 2 levels (4GB files)
1549 /* this is default reservation: 2 levels */
1550 credits = (blocks + 2) * 3;
1552 /* actual offset is unknown, hard to optimize */
1556 /* now check for few specific cases to optimize */
1557 if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1560 /* allocate if not allocated */
1561 if (inode == NULL) {
1562 credits += blocks * 2;
1565 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1566 LASSERT(i < LDISKFS_NDIR_BLOCKS);
1567 if (LDISKFS_I(inode)->i_data[i] == 0)
1570 } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1571 /* single indirect */
1572 credits = blocks * 3;
1573 if (inode == NULL ||
1574 LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1577 /* The indirect block may be modified. */
1584 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1585 const struct lu_buf *buf, loff_t _pos,
1586 struct thandle *handle)
1588 struct osd_object *obj = osd_dt_obj(dt);
1589 struct inode *inode = obj->oo_inode;
1590 struct super_block *sb = osd_sb(osd_obj2dev(obj));
1591 struct osd_thandle *oh;
1592 int rc = 0, est = 0, credits, blocks, allocated = 0;
1598 LASSERT(buf != NULL);
1599 LASSERT(handle != NULL);
1601 oh = container_of(handle, struct osd_thandle, ot_super);
1602 LASSERT(oh->ot_handle == NULL);
1605 bits = sb->s_blocksize_bits;
1609 /* if this is an append, then we
1610 * should expect cross-block record
1617 /* blocks to modify */
1618 blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1619 LASSERT(blocks > 0);
1621 if (inode != NULL && _pos != -1) {
1622 /* object size in blocks */
1623 est = (i_size_read(inode) + bs - 1) >> bits;
1624 allocated = inode->i_blocks >> (bits - 9);
1625 if (pos + size <= i_size_read(inode) && est <= allocated) {
1626 /* looks like an overwrite, no need to modify tree */
1628 /* no need to modify i_size */
1633 if (osd_extents_enabled(sb, inode)) {
1635 * many concurrent threads may grow tree by the time
1636 * our transaction starts. so, consider 2 is a min depth
1637 * for every level we may need to allocate a new block
1638 * and take some entries from the old one. so, 3 blocks
1639 * to allocate (bitmap, gd, itself) + old block - 4 per
1642 depth = inode != NULL ? ext_depth(inode) : 0;
1643 depth = max(depth, 1) + 1;
1645 /* if not append, then split may need to modify
1646 * existing blocks moving entries into the new ones
1650 /* blocks to store data: bitmap,gd,itself */
1651 credits += blocks * 3;
1653 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1655 /* if inode is created as part of the transaction,
1656 * then it's counted already by the creation method
1663 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1665 /* dt_declare_write() is usually called for system objects, such
1666 * as llog or last_rcvd files. We needn't enforce quota on those
1667 * objects, so always set the lqi_space as 0.
1670 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1672 i_projid_read(inode), 0,
1673 oh, obj, NULL, OSD_QID_BLK);
1676 rc = osd_trunc_lock(obj, oh, true);
1681 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1683 /* LU-2634: clear the extent format for fast symlink */
1684 ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1686 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1687 spin_lock(&inode->i_lock);
1688 LDISKFS_I(inode)->i_disksize = buflen;
1689 i_size_write(inode, buflen);
1690 spin_unlock(&inode->i_lock);
1691 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1696 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
1697 int bufsize, int write_NUL, loff_t *offs,
1700 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1701 struct buffer_head *bh = NULL;
1702 loff_t offset = *offs;
1703 loff_t new_size = i_size_read(inode);
1704 unsigned long block;
1705 int blocksize = 1 << inode->i_blkbits;
1706 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1710 int dirty_inode = 0;
1711 bool create, sparse, sync = false;
1715 * long symlink write does not count the NUL terminator in
1716 * bufsize, we write it, and the inode's file size does not
1717 * count the NUL terminator as well.
1719 ((char *)buf)[bufsize] = '\0';
1723 dirty_inode = test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA,
1726 /* sparse checking is racy, but sparse is very rare case, leave as is */
1727 sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
1728 ((new_size - 1) >> inode->i_blkbits) + 1);
1730 while (bufsize > 0) {
1731 int credits = handle->h_buffer_credits;
1732 unsigned long last_block = (new_size == 0) ? 0 :
1733 (new_size - 1) >> inode->i_blkbits;
1738 block = offset >> inode->i_blkbits;
1739 boffs = offset & (blocksize - 1);
1740 size = min(blocksize - boffs, bufsize);
1741 sync = (block > last_block || new_size == 0 || sparse);
1744 down(&ei->i_append_sem);
1746 bh = __ldiskfs_bread(handle, inode, block, 0);
1748 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
1750 "%s: adding bh without locking off %llu (block %lu, size %d, offs %llu)\n",
1752 offset, block, bufsize, *offs);
1754 if (IS_ERR_OR_NULL(bh)) {
1755 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1756 int flags = LDISKFS_GET_BLOCKS_CREATE;
1758 /* while the file system is being mounted, avoid
1759 * preallocation otherwise mount can take a long
1760 * time as mballoc cache is cold.
1761 * XXX: this is a workaround until we have a proper
1763 * XXX: works with extent-based files only */
1764 if (!osd->od_cl_seq)
1765 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
1766 bh = __ldiskfs_bread(handle, inode, block, flags);
1770 up(&ei->i_append_sem);
1775 if (IS_ERR_OR_NULL(bh)) {
1784 "%s: error reading offset %llu (block %lu, size %d, offs %llu), credits %d/%d: rc = %d\n",
1785 inode->i_sb->s_id, offset, block, bufsize, *offs,
1786 credits, handle->h_buffer_credits, err);
1790 err = ldiskfs_journal_get_write_access(handle, bh);
1792 CERROR("journal_get_write_access() returned error %d\n",
1796 LASSERTF(boffs + size <= bh->b_size,
1797 "boffs %d size %d bh->b_size %lu\n",
1798 boffs, size, (unsigned long)bh->b_size);
1800 memset(bh->b_data, 0, bh->b_size);
1802 up(&ei->i_append_sem);
1806 memcpy(bh->b_data + boffs, buf, size);
1807 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1811 if (offset + size > new_size)
1812 new_size = offset + size;
1818 up(&ei->i_append_sem);
1825 /* correct in-core and on-disk sizes */
1826 if (new_size > i_size_read(inode)) {
1827 spin_lock(&inode->i_lock);
1828 if (new_size > i_size_read(inode))
1829 i_size_write(inode, new_size);
1830 if (i_size_read(inode) > ei->i_disksize) {
1831 ei->i_disksize = i_size_read(inode);
1834 spin_unlock(&inode->i_lock);
1837 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1844 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1845 const struct lu_buf *buf, loff_t *pos,
1846 struct thandle *handle)
1848 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1849 struct osd_thandle *oh;
1853 LASSERT(dt_object_exists(dt));
1855 LASSERT(handle != NULL);
1856 LASSERT(inode != NULL);
1857 dquot_initialize(inode);
1859 /* XXX: don't check: one declared chunk can be used many times */
1860 /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1862 oh = container_of(handle, struct osd_thandle, ot_super);
1863 LASSERT(oh->ot_handle->h_transaction != NULL);
1864 osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1866 /* Write small symlink to inode body as we need to maintain correct
1867 * on-disk symlinks for ldiskfs.
1868 * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1869 * does not count it in.
1871 is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1872 if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1873 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1875 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
1876 is_link, pos, oh->ot_handle);
1878 result = buf->lb_len;
1880 osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1885 static int osd_declare_fallocate(const struct lu_env *env,
1886 struct dt_object *dt, struct thandle *th)
1888 struct osd_thandle *oh;
1889 struct inode *inode;
1894 oh = container_of(th, struct osd_thandle, ot_super);
1896 osd_trans_declare_op(env, oh, OSD_OT_PREALLOC,
1897 osd_dto_credits_noquota[DTO_WRITE_BLOCK]);
1898 inode = osd_dt_obj(dt)->oo_inode;
1901 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1902 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1907 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
1908 __u64 start, __u64 end, int mode, struct thandle *th)
1910 struct osd_object *obj = osd_dt_obj(dt);
1911 struct inode *inode = obj->oo_inode;
1913 struct osd_thread_info *info = osd_oti_get(env);
1914 struct dentry *dentry = &info->oti_obj_dentry;
1915 struct file *file = &info->oti_file;
1919 * Only mode == 0 (which is standard prealloc) is supported now.
1920 * Rest of mode options is not supported yet.
1922 if (mode & ~FALLOC_FL_KEEP_SIZE)
1923 RETURN(-EOPNOTSUPP);
1925 LASSERT(dt_object_exists(dt));
1926 LASSERT(osd_invariant(obj));
1927 LASSERT(inode != NULL);
1928 dquot_initialize(inode);
1932 osd_trans_exec_op(env, th, OSD_OT_PREALLOC);
1935 * Because f_op->fallocate() does not have an inode arg
1937 dentry->d_inode = inode;
1938 dentry->d_sb = inode->i_sb;
1939 file->f_path.dentry = dentry;
1940 file->f_mapping = inode->i_mapping;
1941 file->f_op = inode->i_fop;
1942 file->f_inode = inode;
1943 rc = file->f_op->fallocate(file, mode, start, end - start);
1948 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1949 __u64 start, __u64 end, struct thandle *th)
1951 struct osd_thandle *oh;
1952 struct inode *inode;
1957 oh = container_of(th, struct osd_thandle, ot_super);
1960 * we don't need to reserve credits for whole truncate
1961 * it's not possible as truncate may need to free too many
1962 * blocks and that won't fit a single transaction. instead
1963 * we reserve credits to change i_size and put inode onto
1964 * orphan list. if needed truncate will extend or restart
1967 osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1968 osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1970 inode = osd_dt_obj(dt)->oo_inode;
1973 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1974 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1978 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
1983 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1984 __u64 start, __u64 end, struct thandle *th)
1986 struct osd_object *obj = osd_dt_obj(dt);
1987 struct osd_device *osd = osd_obj2dev(obj);
1988 struct inode *inode = obj->oo_inode;
1989 struct osd_access_lock *al;
1990 struct osd_thandle *oh;
1991 int rc = 0, found = 0;
1995 LASSERT(dt_object_exists(dt));
1996 LASSERT(osd_invariant(obj));
1997 LASSERT(inode != NULL);
1998 dquot_initialize(inode);
2001 oh = container_of(th, struct osd_thandle, ot_super);
2002 LASSERT(oh->ot_handle->h_transaction != NULL);
2004 /* we used to skip truncate to current size to
2005 * optimize truncates on OST. with DoM we can
2006 * get attr_set to set specific size (MDS_REINT)
2007 * and then get truncate RPC which essentially
2008 * would be skipped. this is bad.. so, disable
2009 * this optimization on MDS till the client stop
2010 * to sent MDS_REINT (LU-11033) -bzzz
2012 if (osd->od_is_ost && i_size_read(inode) == start)
2015 osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2017 spin_lock(&inode->i_lock);
2018 if (i_size_read(inode) < start)
2020 i_size_write(inode, start);
2021 spin_unlock(&inode->i_lock);
2022 /* if object holds encrypted content, we need to make sure we truncate
2023 * on an encryption unit boundary, or subsequent reads will get
2026 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2027 start & ~LUSTRE_ENCRYPTION_MASK)
2028 start = (start & LUSTRE_ENCRYPTION_MASK) +
2029 LUSTRE_ENCRYPTION_UNIT_SIZE;
2030 ll_truncate_pagecache(inode, start);
2032 /* optimize grow case */
2034 osd_execute_truncate(obj);
2039 /* add to orphan list to ensure truncate completion
2040 * if this transaction succeed. ldiskfs_truncate()
2041 * will take the inode out of the list
2043 rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2044 inode_unlock(inode);
2048 list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2049 if (obj != al->tl_obj)
2051 LASSERT(al->tl_shared == 0);
2053 /* do actual truncate in osd_trans_stop() */
2054 al->tl_truncate = 1;
2063 static int fiemap_check_ranges(struct inode *inode,
2064 u64 start, u64 len, u64 *new_len)
2073 if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2074 maxbytes = inode->i_sb->s_maxbytes;
2076 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2078 if (start > maxbytes)
2082 * Shrink request scope to what the fs can actually handle.
2084 if (len > maxbytes || (maxbytes - len) < start)
2085 *new_len = maxbytes - start;
2090 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2091 #define FIEMAP_MAX_EXTENTS (UINT_MAX / sizeof(struct fiemap_extent))
2093 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2096 struct fiemap_extent_info fieinfo = {0, };
2097 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2100 mm_segment_t cur_fs;
2103 if (inode->i_op->fiemap == NULL)
2106 if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2109 rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2113 fieinfo.fi_flags = fm->fm_flags;
2114 fieinfo.fi_extents_max = fm->fm_extent_count;
2115 fieinfo.fi_extents_start = fm->fm_extents;
2117 if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2118 filemap_write_and_wait(inode->i_mapping);
2120 /* Save previous value address limit */
2122 /* Set the address limit of the kernel */
2125 rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2126 fm->fm_flags = fieinfo.fi_flags;
2127 fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2129 /* Restore the previous address limt */
2135 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2136 __u64 start, __u64 end, enum lu_ladvise_type advice)
2138 struct osd_object *obj = osd_dt_obj(dt);
2143 case LU_LADVISE_DONTNEED:
2145 invalidate_mapping_pages(obj->oo_inode->i_mapping,
2146 start >> PAGE_SHIFT,
2147 (end - 1) >> PAGE_SHIFT);
2158 * in some cases we may need declare methods for objects being created
2159 * e.g., when we create symlink
2161 const struct dt_body_operations osd_body_ops_new = {
2162 .dbo_declare_write = osd_declare_write,
2165 const struct dt_body_operations osd_body_ops = {
2166 .dbo_read = osd_read,
2167 .dbo_declare_write = osd_declare_write,
2168 .dbo_write = osd_write,
2169 .dbo_bufs_get = osd_bufs_get,
2170 .dbo_bufs_put = osd_bufs_put,
2171 .dbo_write_prep = osd_write_prep,
2172 .dbo_declare_write_commit = osd_declare_write_commit,
2173 .dbo_write_commit = osd_write_commit,
2174 .dbo_read_prep = osd_read_prep,
2175 .dbo_declare_punch = osd_declare_punch,
2176 .dbo_punch = osd_punch,
2177 .dbo_fiemap_get = osd_fiemap_get,
2178 .dbo_ladvise = osd_ladvise,
2179 .dbo_declare_fallocate = osd_declare_fallocate,
2180 .dbo_fallocate = osd_fallocate,
2184 * Get a truncate lock
2186 * In order to take multi-transaction truncate out of main transaction we let
2187 * the caller grab a lock on the object passed. the lock can be shared (for
2188 * writes) and exclusive (for truncate). It's not allowed to mix truncate
2189 * and write in the same transaction handle (do not confuse with big ldiskfs
2190 * transaction containing lots of handles).
2191 * The lock must be taken at declaration.
2193 * \param obj object to lock
2195 * \shared shared or exclusive
2197 * \retval 0 lock is granted
2198 * \retval -NOMEM no memory to allocate lock
2200 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2202 struct osd_access_lock *al, *tmp;
2207 list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2208 if (tmp->tl_obj != obj)
2210 LASSERT(tmp->tl_shared == shared);
2211 /* found same lock */
2216 if (unlikely(al == NULL))
2219 al->tl_truncate = false;
2221 down_read(&obj->oo_ext_idx_sem);
2223 down_write(&obj->oo_ext_idx_sem);
2224 al->tl_shared = shared;
2225 lu_object_get(&obj->oo_dt.do_lu);
2227 list_add(&al->tl_list, &oh->ot_trunc_locks);
2232 void osd_trunc_unlock_all(const struct lu_env *env, struct list_head *list)
2234 struct osd_access_lock *al, *tmp;
2236 list_for_each_entry_safe(al, tmp, list, tl_list) {
2238 up_read(&al->tl_obj->oo_ext_idx_sem);
2240 up_write(&al->tl_obj->oo_ext_idx_sem);
2241 osd_object_put(env, al->tl_obj);
2242 list_del(&al->tl_list);
2247 void osd_execute_truncate(struct osd_object *obj)
2249 struct osd_device *d = osd_obj2dev(obj);
2250 struct inode *inode = obj->oo_inode;
2253 /* simulate crash before (in the middle) of delayed truncate */
2254 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2255 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2256 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2258 mutex_lock(&sbi->s_orphan_lock);
2259 list_del_init(&ei->i_orphan);
2260 mutex_unlock(&sbi->s_orphan_lock);
2264 size = i_size_read(inode);
2266 /* if object holds encrypted content, we need to make sure we truncate
2267 * on an encryption unit boundary, or block content will get corrupted
2269 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2270 size & ~LUSTRE_ENCRYPTION_MASK)
2271 inode->i_size = (size & LUSTRE_ENCRYPTION_MASK) +
2272 LUSTRE_ENCRYPTION_UNIT_SIZE;
2273 ldiskfs_truncate(inode);
2274 inode_unlock(inode);
2275 if (inode->i_size != size) {
2276 spin_lock(&inode->i_lock);
2277 i_size_write(inode, size);
2278 LDISKFS_I(inode)->i_disksize = size;
2279 spin_unlock(&inode->i_lock);
2280 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2284 * For a partial-page truncate, flush the page to disk immediately to
2285 * avoid data corruption during direct disk write. b=17397
2287 if ((size & ~PAGE_MASK) == 0)
2289 if (osd_use_page_cache(d)) {
2290 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2292 /* Notice we use "wait" version to ensure I/O is complete */
2293 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2294 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2295 size >> PAGE_SHIFT);
2299 void osd_process_truncates(struct list_head *list)
2301 struct osd_access_lock *al;
2303 LASSERT(journal_current_handle() == NULL);
2305 list_for_each_entry(al, list, tl_list) {
2308 if (!al->tl_truncate)
2310 osd_execute_truncate(al->tl_obj);