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/
35 * Author: Nikita Danilov <nikita@clusterfs.com>
36 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
40 #define DEBUG_SUBSYSTEM S_OSD
42 /* prerequisite for linux/xattr.h */
43 #include <linux/types.h>
44 /* prerequisite for linux/xattr.h */
47 #include <linux/pagevec.h>
50 * struct OBD_{ALLOC,FREE}*()
53 #include <obd_support.h>
55 #include "osd_internal.h"
58 #include <ldiskfs/ldiskfs_extents.h>
60 static inline bool osd_use_page_cache(struct osd_device *d)
62 /* do not use pagecache if write and read caching are disabled */
63 if (d->od_writethrough_cache + d->od_read_cache == 0)
65 /* use pagecache by default */
69 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
70 int rw, int line, int pages)
74 LASSERTF(iobuf->dr_elapsed_valid == 0,
75 "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
76 atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
78 LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
80 init_waitqueue_head(&iobuf->dr_wait);
81 atomic_set(&iobuf->dr_numreqs, 0);
86 iobuf->dr_elapsed = ktime_set(0, 0);
87 /* must be counted before, so assert */
89 iobuf->dr_init_at = line;
91 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
92 if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
93 LASSERT(iobuf->dr_pg_buf.lb_len >=
94 pages * sizeof(iobuf->dr_pages[0]));
98 /* start with 1MB for 4K blocks */
100 while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
103 CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
104 (unsigned int)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
106 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
107 iobuf->dr_max_pages = 0;
108 CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
109 (unsigned int)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
111 lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
112 iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
113 if (unlikely(iobuf->dr_blocks == NULL))
116 lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
117 iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
118 if (unlikely(iobuf->dr_pages == NULL))
121 lu_buf_realloc(&iobuf->dr_lnb_buf,
122 pages * sizeof(iobuf->dr_lnbs[0]));
123 iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf;
124 if (unlikely(iobuf->dr_lnbs == NULL))
127 iobuf->dr_max_pages = pages;
131 #define osd_init_iobuf(dev, iobuf, rw, pages) \
132 __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
134 static void osd_iobuf_add_page(struct osd_iobuf *iobuf,
135 struct niobuf_local *lnb)
137 LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
138 iobuf->dr_pages[iobuf->dr_npages] = lnb->lnb_page;
139 iobuf->dr_lnbs[iobuf->dr_npages] = lnb;
143 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
145 int rw = iobuf->dr_rw;
147 if (iobuf->dr_elapsed_valid) {
148 iobuf->dr_elapsed_valid = 0;
149 LASSERT(iobuf->dr_dev == d);
150 LASSERT(iobuf->dr_frags > 0);
151 lprocfs_oh_tally(&d->od_brw_stats.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 = blk_status_to_errno(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!!!
173 if (unlikely(iobuf == NULL)) {
174 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");
175 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
176 ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
177 bio->bi_next, (unsigned long)bio->bi_flags,
178 (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
179 bio_sectors(bio) << 9, bio->bi_end_io,
180 atomic_read(&bio->__bi_cnt),
185 /* the check is outside of the cycle for performance reason -bzzz */
186 if (!bio_data_dir(bio)) {
187 DECLARE_BVEC_ITER_ALL(iter_all);
189 bio_for_each_segment_all(bvl, bio, iter_all) {
190 if (likely(error == 0))
191 SetPageUptodate(bvl_to_page(bvl));
192 LASSERT(PageLocked(bvl_to_page(bvl)));
194 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
196 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
199 /* any real error is good enough -bzzz */
200 if (error != 0 && iobuf->dr_error == 0)
201 iobuf->dr_error = error;
204 * set dr_elapsed before dr_numreqs turns to 0, otherwise
205 * it's possible that service thread will see dr_numreqs
206 * is zero, but dr_elapsed is not set yet, leading to lost
207 * data in this processing and an assertion in a subsequent
210 if (atomic_read(&iobuf->dr_numreqs) == 1) {
211 ktime_t now = ktime_get();
213 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
214 iobuf->dr_elapsed_valid = 1;
216 if (atomic_dec_and_test(&iobuf->dr_numreqs))
217 wake_up(&iobuf->dr_wait);
219 /* Completed bios used to be chained off iobuf->dr_bios and freed in
220 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
221 * mempool when serious on-disk fragmentation was encountered,
222 * deadlocking the OST. The bios are now released as soon as complete
223 * so the pool cannot be exhausted while IOs are competing. b=10076
228 static void record_start_io(struct osd_iobuf *iobuf, int size)
230 struct osd_device *osd = iobuf->dr_dev;
231 struct obd_histogram *h = osd->od_brw_stats.hist;
234 atomic_inc(&iobuf->dr_numreqs);
236 if (iobuf->dr_rw == 0) {
237 atomic_inc(&osd->od_r_in_flight);
238 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
239 atomic_read(&osd->od_r_in_flight));
240 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
241 } else if (iobuf->dr_rw == 1) {
242 atomic_inc(&osd->od_w_in_flight);
243 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
244 atomic_read(&osd->od_w_in_flight));
245 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
251 static void osd_submit_bio(int rw, struct bio *bio)
253 LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
254 #ifdef HAVE_SUBMIT_BIO_2ARGS
255 submit_bio(rw ? WRITE : READ, bio);
262 static int can_be_merged(struct bio *bio, sector_t sector)
267 return bio_end_sector(bio) == sector ? 1 : 0;
270 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
272 * This function will change the data written, thus it should only be
273 * used when checking data integrity feature
275 static void bio_integrity_fault_inject(struct bio *bio)
277 struct bio_vec *bvec;
278 DECLARE_BVEC_ITER_ALL(iter_all);
282 bio_for_each_segment_all(bvec, bio, iter_all) {
283 struct page *page = bvec->bv_page;
293 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
294 unsigned int sectors, int tuple_size)
296 __u16 *expected_guard;
300 expected_guard = expected_guard_buf;
301 for (i = 0; i < sectors; i++) {
302 bio_guard = (__u16 *)bio_prot_buf;
303 if (*bio_guard != *expected_guard) {
305 "unexpected guard tags on sector %d expected guard %u, bio guard %u, sectors %u, tuple size %d\n",
306 i, *expected_guard, *bio_guard, sectors,
311 bio_prot_buf += tuple_size;
316 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
317 struct osd_iobuf *iobuf, int index)
319 struct blk_integrity *bi = bdev_get_integrity(bdev);
320 struct bio_integrity_payload *bip = bio->bi_integrity;
321 struct niobuf_local *lnb;
322 unsigned short sector_size = blk_integrity_interval(bi);
323 void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
324 bip->bip_vec->bv_offset;
326 sector_t sector = bio_start_sector(bio);
327 unsigned int sectors, total;
328 DECLARE_BVEC_ITER_ALL(iter_all);
329 __u16 *expected_guard;
333 bio_for_each_segment_all(bv, bio, iter_all) {
334 lnb = iobuf->dr_lnbs[index];
335 expected_guard = lnb->lnb_guards;
336 sectors = bv->bv_len / sector_size;
337 if (lnb->lnb_guard_rpc) {
338 rc = bio_dif_compare(expected_guard, bio_prot_buf,
339 sectors, bi->tuple_size);
345 bio_prot_buf += sectors * bi->tuple_size;
346 total += sectors * bi->tuple_size;
347 LASSERT(total <= bip_size(bio->bi_integrity));
353 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
354 struct osd_iobuf *iobuf,
355 int start_page_idx, bool fault_inject,
356 bool integrity_enabled)
358 struct super_block *sb = osd_sb(osd);
359 integrity_gen_fn *generate_fn = NULL;
360 integrity_vrfy_fn *verify_fn = NULL;
365 if (!integrity_enabled)
368 rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
372 rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
376 /* Verify and inject fault only when writing */
377 if (iobuf->dr_rw == 1) {
378 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
379 rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
385 if (unlikely(fault_inject))
386 bio_integrity_fault_inject(bio);
392 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
393 # ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
394 static void dio_integrity_complete_routine(struct bio *bio)
396 static void dio_integrity_complete_routine(struct bio *bio, int error)
399 struct osd_bio_private *bio_private = bio->bi_private;
401 bio->bi_private = bio_private->obp_iobuf;
402 osd_dio_complete_routine(bio, error);
404 OBD_FREE_PTR(bio_private);
406 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
407 #else /* !CONFIG_BLK_DEV_INTEGRITY */
408 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
409 fault_inject, integrity_enabled) 0
410 #endif /* CONFIG_BLK_DEV_INTEGRITY */
412 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
413 bool integrity_enabled, int start_page_idx,
414 struct osd_bio_private **pprivate)
420 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
421 if (integrity_enabled) {
422 struct osd_bio_private *bio_private = NULL;
424 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
425 if (bio_private == NULL)
427 bio->bi_end_io = dio_integrity_complete_routine;
428 bio->bi_private = bio_private;
429 bio_private->obp_start_page_idx = start_page_idx;
430 bio_private->obp_iobuf = iobuf;
431 *pprivate = bio_private;
435 bio->bi_end_io = dio_complete_routine;
436 bio->bi_private = iobuf;
442 static void osd_mark_page_io_done(struct osd_iobuf *iobuf,
444 sector_t start_blocks,
447 struct niobuf_local *lnb;
448 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
449 pgoff_t pg_start, pg_end;
451 pg_start = start_blocks / blocks_per_page;
452 if (start_blocks % blocks_per_page)
454 if (count >= blocks_per_page)
455 pg_end = (start_blocks + count -
456 blocks_per_page) / blocks_per_page;
458 return; /* nothing to mark */
459 for ( ; pg_start <= pg_end; pg_start++) {
460 lnb = iobuf->dr_lnbs[pg_start];
461 lnb->lnb_flags |= OBD_BRW_DONE;
465 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
466 struct osd_iobuf *iobuf, sector_t start_blocks,
469 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
470 struct page **pages = iobuf->dr_pages;
471 int npages = iobuf->dr_npages;
472 sector_t *blocks = iobuf->dr_blocks;
473 struct super_block *sb = inode->i_sb;
474 int sector_bits = sb->s_blocksize_bits - 9;
475 unsigned int blocksize = sb->s_blocksize;
476 struct block_device *bdev = sb->s_bdev;
477 struct osd_bio_private *bio_private = NULL;
478 struct bio *bio = NULL;
479 int bio_start_page_idx;
481 unsigned int page_offset;
484 int block_idx, block_idx_end;
485 int page_idx, page_idx_start;
489 bool integrity_enabled;
490 struct blk_plug plug;
491 int blocks_left_page;
495 fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
496 LASSERT(iobuf->dr_npages == npages);
498 integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
500 osd_brw_stats_update(osd, iobuf);
501 iobuf->dr_start_time = ktime_get();
504 count = npages * blocks_per_page;
505 block_idx_end = start_blocks + count;
507 blk_start_plug(&plug);
509 page_idx_start = start_blocks / blocks_per_page;
510 for (page_idx = page_idx_start, block_idx = start_blocks;
511 block_idx < block_idx_end; page_idx++,
512 block_idx += blocks_left_page) {
513 page = pages[page_idx];
514 LASSERT(page_idx < iobuf->dr_npages);
516 i = block_idx % blocks_per_page;
517 blocks_left_page = blocks_per_page - i;
518 for (page_offset = i * blocksize; i < blocks_left_page;
519 i += nblocks, page_offset += blocksize * nblocks) {
522 if (blocks[block_idx + i] == 0) { /* hole */
523 LASSERTF(iobuf->dr_rw == 0,
524 "page_idx %u, block_idx %u, i %u,"
525 "start_blocks: %llu, count: %llu, npages: %d\n",
526 page_idx, block_idx, i,
527 (unsigned long long)start_blocks,
528 (unsigned long long)count, npages);
529 memset(kmap(page) + page_offset, 0, blocksize);
534 sector = (sector_t)blocks[block_idx + i] << sector_bits;
536 /* Additional contiguous file blocks? */
537 while (i + nblocks < blocks_left_page &&
538 (sector + (nblocks << sector_bits)) ==
539 ((sector_t)blocks[block_idx + i + nblocks] <<
543 if (bio && can_be_merged(bio, sector) &&
544 bio_add_page(bio, page, blocksize * nblocks,
546 continue; /* added this frag OK */
549 struct request_queue *q = bio_get_queue(bio);
550 unsigned int bi_size = bio_sectors(bio) << 9;
552 /* Dang! I have to fragment this I/O */
554 "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
555 bi_size, bio->bi_vcnt, bio->bi_max_vecs,
557 queue_max_sectors(q),
558 osd_bio_nr_segs(bio),
559 queue_max_segments(q));
560 rc = osd_bio_integrity_handle(osd, bio,
561 iobuf, bio_start_page_idx,
562 fault_inject, integrity_enabled);
568 record_start_io(iobuf, bi_size);
569 osd_submit_bio(iobuf->dr_rw, bio);
572 bio_start_page_idx = page_idx;
573 /* allocate new bio */
574 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
575 (block_idx_end - block_idx +
576 blocks_left_page - 1)));
578 CERROR("Can't allocate bio %u pages\n",
579 block_idx_end - block_idx +
580 blocks_left_page - 1);
585 bio_set_dev(bio, bdev);
586 bio_set_sector(bio, sector);
587 bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
588 rc = osd_bio_init(bio, iobuf, integrity_enabled,
589 bio_start_page_idx, &bio_private);
595 rc = bio_add_page(bio, page,
596 blocksize * nblocks, page_offset);
602 rc = osd_bio_integrity_handle(osd, bio, iobuf,
611 record_start_io(iobuf, bio_sectors(bio) << 9);
612 osd_submit_bio(iobuf->dr_rw, bio);
617 blk_finish_plug(&plug);
619 /* in order to achieve better IO throughput, we don't wait for writes
620 * completion here. instead we proceed with transaction commit in
621 * parallel and wait for IO completion once transaction is stopped
622 * see osd_trans_stop() for more details -bzzz
624 if (iobuf->dr_rw == 0 || fault_inject) {
625 wait_event(iobuf->dr_wait,
626 atomic_read(&iobuf->dr_numreqs) == 0);
627 osd_fini_iobuf(osd, iobuf);
631 rc = iobuf->dr_error;
634 OBD_FREE_PTR(bio_private);
638 if (rc == 0 && iobuf->dr_rw)
639 osd_mark_page_io_done(iobuf, inode,
640 start_blocks, count);
645 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
646 struct niobuf_local *lnb, int maxlnb)
654 int poff = offset & (PAGE_SIZE - 1);
655 int plen = PAGE_SIZE - poff;
657 if (*nrpages >= maxlnb) {
664 lnb->lnb_file_offset = offset;
665 lnb->lnb_page_offset = poff;
667 /* lnb->lnb_flags = rnb->rnb_flags; */
669 lnb->lnb_page = NULL;
671 lnb->lnb_guard_rpc = 0;
672 lnb->lnb_guard_disk = 0;
675 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
686 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
687 loff_t offset, gfp_t gfp_mask, bool cache)
689 struct osd_thread_info *oti = osd_oti_get(env);
690 struct inode *inode = osd_dt_obj(dt)->oo_inode;
691 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
698 page = find_or_create_page(inode->i_mapping,
699 offset >> PAGE_SHIFT, gfp_mask);
702 LASSERT(!PagePrivate2(page));
703 wait_on_page_writeback(page);
705 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
711 if (inode->i_mapping->nrpages) {
712 /* consult with pagecache, but do not create new pages */
713 /* this is normally used once */
714 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
716 wait_on_page_writeback(page);
721 LASSERT(oti->oti_dio_pages);
722 cur = oti->oti_dio_pages_used;
723 page = oti->oti_dio_pages[cur];
725 if (unlikely(!page)) {
726 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
727 page = alloc_page(gfp_mask);
730 oti->oti_dio_pages[cur] = page;
731 SetPagePrivate2(page);
735 ClearPageUptodate(page);
736 page->index = offset >> PAGE_SHIFT;
737 oti->oti_dio_pages_used++;
743 * there are following "locks":
754 * - lock pages, unlock
756 * - lock partial page
762 * Unlock and release pages loaded by osd_bufs_get()
764 * Unlock \a npages pages from \a lnb and drop the refcount on them.
766 * \param env thread execution environment
767 * \param dt dt object undergoing IO (OSD object + methods)
768 * \param lnb array of pages undergoing IO
769 * \param npages number of pages in \a lnb
773 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
774 struct niobuf_local *lnb, int npages)
776 struct osd_thread_info *oti = osd_oti_get(env);
780 ll_pagevec_init(&pvec, 0);
782 for (i = 0; i < npages; i++) {
783 struct page *page = lnb[i].lnb_page;
788 /* if the page isn't cached, then reset uptodate
791 if (PagePrivate2(page)) {
792 oti->oti_dio_pages_used--;
794 if (lnb[i].lnb_locked)
796 if (pagevec_add(&pvec, page) == 0)
797 pagevec_release(&pvec);
800 lnb[i].lnb_page = NULL;
803 LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
805 /* Release any partial pagevec */
806 pagevec_release(&pvec);
812 * Load and lock pages undergoing IO
814 * Pages as described in the \a lnb array are fetched (from disk or cache)
815 * and locked for IO by the caller.
817 * DLM locking protects us from write and truncate competing for same region,
818 * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
819 * It's possible the writeout on a such a page is in progress when we access
820 * it. It's also possible that during this writeout we put new (partial) data
821 * into the page, but won't be able to proceed in filter_commitrw_write().
822 * Therefore, just wait for writeout completion as it should be rare enough.
824 * \param env thread execution environment
825 * \param dt dt object undergoing IO (OSD object + methods)
826 * \param pos byte offset of IO start
827 * \param len number of bytes of IO
828 * \param lnb array of extents undergoing IO
829 * \param rw read or write operation, and other flags
830 * \param capa capabilities
832 * \retval pages (zero or more) loaded successfully
833 * \retval -ENOMEM on memory/page allocation error
835 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
836 loff_t pos, ssize_t len, struct niobuf_local *lnb,
837 int maxlnb, enum dt_bufs_type rw)
839 struct osd_thread_info *oti = osd_oti_get(env);
840 struct osd_object *obj = osd_dt_obj(dt);
841 struct osd_device *osd = osd_obj2dev(obj);
842 int npages, i, iosize, rc = 0;
847 LASSERT(obj->oo_inode);
849 rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
853 write = rw & DT_BUFS_TYPE_WRITE;
855 fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
856 iosize = fsize - lnb[0].lnb_file_offset;
857 fsize = max(fsize, i_size_read(obj->oo_inode));
859 cache = rw & DT_BUFS_TYPE_READAHEAD;
863 cache = osd_use_page_cache(osd);
866 if (!osd->od_writethrough_cache) {
870 if (iosize > osd->od_writethrough_max_iosize) {
875 if (!osd->od_read_cache) {
879 if (iosize > osd->od_readcache_max_iosize) {
884 /* don't use cache on large files */
885 if (osd->od_readcache_max_filesize &&
886 fsize > osd->od_readcache_max_filesize)
892 if (!cache && unlikely(!oti->oti_dio_pages)) {
893 OBD_ALLOC_PTR_ARRAY_LARGE(oti->oti_dio_pages,
894 PTLRPC_MAX_BRW_PAGES);
895 if (!oti->oti_dio_pages)
899 /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
900 gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
902 for (i = 0; i < npages; i++, lnb++) {
903 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
905 if (lnb->lnb_page == NULL)
906 GOTO(cleanup, rc = -ENOMEM);
912 /* XXX: this version doesn't invalidate cached pages, but use them */
913 if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
914 /* do not allow data aliasing, invalidate pagecache */
915 /* XXX: can be quite expensive in mixed case */
916 invalidate_mapping_pages(obj->oo_inode->i_mapping,
917 lnb[0].lnb_file_offset >> PAGE_SHIFT,
918 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
926 osd_bufs_put(env, dt, lnb - i, i);
929 /* Borrow @ext4_chunk_trans_blocks */
930 static int osd_chunk_trans_blocks(struct inode *inode, int nrblocks)
932 ldiskfs_group_t groups;
938 depth = ext_depth(inode);
939 idxblocks = depth * 2;
942 * Now let's see how many group bitmaps and group descriptors need
945 groups = idxblocks + 1;
947 if (groups > LDISKFS_SB(inode->i_sb)->s_groups_count)
948 groups = LDISKFS_SB(inode->i_sb)->s_groups_count;
949 if (gdpblocks > LDISKFS_SB(inode->i_sb)->s_gdb_count)
950 gdpblocks = LDISKFS_SB(inode->i_sb)->s_gdb_count;
952 /* bitmaps and block group descriptor blocks */
953 ret = idxblocks + groups + gdpblocks;
955 /* Blocks for super block, inode, quota and xattr blocks */
956 ret += LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
961 #ifdef HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS
962 static int osd_extend_restart_trans(handle_t *handle, int needed,
967 rc = ldiskfs_journal_ensure_credits(handle, needed,
968 ldiskfs_trans_default_revoke_credits(inode->i_sb));
969 /* this means journal has been restarted */
976 static int osd_extend_restart_trans(handle_t *handle, int needed,
981 if (ldiskfs_handle_has_enough_credits(handle, needed))
983 rc = ldiskfs_journal_extend(handle,
984 needed - handle->h_buffer_credits);
988 return ldiskfs_journal_restart(handle, needed);
990 #endif /* HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS */
992 static int osd_ldiskfs_map_write(struct inode *inode, struct osd_iobuf *iobuf,
993 struct osd_device *osd, sector_t start_blocks,
994 sector_t count, loff_t *disk_size,
997 /* if file has grown, take user_size into account */
998 if (user_size && *disk_size > user_size)
999 *disk_size = user_size;
1001 spin_lock(&inode->i_lock);
1002 if (*disk_size > i_size_read(inode)) {
1003 i_size_write(inode, *disk_size);
1004 LDISKFS_I(inode)->i_disksize = *disk_size;
1005 spin_unlock(&inode->i_lock);
1006 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1008 spin_unlock(&inode->i_lock);
1012 * We don't do stats here as in read path because
1013 * write is async: we'll do this in osd_put_bufs()
1015 return osd_do_bio(osd, inode, iobuf, start_blocks, count);
1018 static unsigned int osd_extent_bytes(const struct osd_device *o)
1020 unsigned int *extent_bytes_ptr =
1021 raw_cpu_ptr(o->od_extent_bytes_percpu);
1023 if (likely(*extent_bytes_ptr))
1024 return *extent_bytes_ptr;
1026 /* initialize on first access or CPU hotplug */
1027 if (!ldiskfs_has_feature_extents(osd_sb(o)))
1028 *extent_bytes_ptr = 1 << osd_sb(o)->s_blocksize_bits;
1030 *extent_bytes_ptr = OSD_DEFAULT_EXTENT_BYTES;
1032 return *extent_bytes_ptr;
1035 #define EXTENT_BYTES_DECAY 64
1036 static void osd_decay_extent_bytes(struct osd_device *osd,
1037 unsigned int new_bytes)
1039 unsigned int old_bytes;
1041 if (!ldiskfs_has_feature_extents(osd_sb(osd)))
1044 old_bytes = osd_extent_bytes(osd);
1045 *raw_cpu_ptr(osd->od_extent_bytes_percpu) =
1046 (old_bytes * (EXTENT_BYTES_DECAY - 1) +
1047 min(new_bytes, OSD_DEFAULT_EXTENT_BYTES) +
1048 EXTENT_BYTES_DECAY - 1) / EXTENT_BYTES_DECAY;
1051 static int osd_ldiskfs_map_inode_pages(struct inode *inode,
1052 struct osd_iobuf *iobuf,
1053 struct osd_device *osd,
1054 int create, __u64 user_size,
1056 struct thandle *thandle)
1058 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1059 int rc = 0, i = 0, mapped_index = 0;
1060 struct page *fp = NULL;
1062 pgoff_t max_page_index;
1063 handle_t *handle = NULL;
1064 sector_t start_blocks = 0, count = 0;
1065 loff_t disk_size = 0;
1066 struct page **page = iobuf->dr_pages;
1067 int pages = iobuf->dr_npages;
1068 sector_t *blocks = iobuf->dr_blocks;
1069 struct niobuf_local *lnb1, *lnb2;
1070 loff_t size1, size2;
1072 max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
1074 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1075 inode->i_ino, pages, (*page)->index);
1078 create = LDISKFS_GET_BLOCKS_CREATE;
1079 handle = ldiskfs_journal_current_handle();
1080 LASSERT(handle != NULL);
1081 rc = osd_attach_jinode(inode);
1084 disk_size = i_size_read(inode);
1085 /* if disk_size is already bigger than specified user_size,
1088 if (disk_size > user_size)
1091 /* pages are sorted already. so, we just have to find
1092 * contig. space and process them properly
1095 long blen, total = 0, previous_total = 0;
1096 struct ldiskfs_map_blocks map = { 0 };
1098 if (fp == NULL) { /* start new extent */
1103 } else if (fp->index + clen == (*page)->index) {
1104 /* continue the extent */
1110 if (fp->index + clen >= max_page_index)
1111 GOTO(cleanup, rc = -EFBIG);
1112 /* process found extent */
1113 map.m_lblk = fp->index * blocks_per_page;
1114 map.m_len = blen = clen * blocks_per_page;
1117 * We might restart transaction for block allocations,
1118 * in order to make sure data ordered mode, issue IO, disk
1119 * size update and block allocations need be within same
1120 * transaction to make sure consistency.
1122 if (handle && check_credits) {
1123 struct osd_thandle *oh;
1125 LASSERT(thandle != NULL);
1126 oh = container_of(thandle, struct osd_thandle,
1129 * only issue IO if restart transaction needed,
1130 * as update disk size need hold inode lock, we
1131 * want to avoid that as much as possible.
1133 if (oh->oh_declared_ext <= 0) {
1134 rc = osd_ldiskfs_map_write(inode,
1135 iobuf, osd, start_blocks,
1136 count, &disk_size, user_size);
1139 thandle->th_restart_tran = 1;
1140 GOTO(cleanup, rc = -EAGAIN);
1143 if (OBD_FAIL_CHECK(OBD_FAIL_OST_RESTART_IO))
1144 oh->oh_declared_ext = 0;
1146 oh->oh_declared_ext--;
1148 rc = ldiskfs_map_blocks(handle, inode, &map, create);
1152 for (; total < blen && c < map.m_len; c++, total++) {
1154 *(blocks + total) = 0;
1158 if ((map.m_flags & LDISKFS_MAP_UNWRITTEN) &&
1160 /* don't try to read allocated, but
1161 * unwritten blocks, instead fill the
1162 * patches with zeros in osd_do_bio() */
1163 *(blocks + total) = 0;
1166 *(blocks + total) = map.m_pblk + c;
1167 /* unmap any possible underlying
1168 * metadata from the block device
1171 if ((map.m_flags & LDISKFS_MAP_NEW) &&
1173 clean_bdev_aliases(inode->i_sb->s_bdev,
1179 if (rc == 0 && create) {
1180 count += (total - previous_total);
1181 mapped_index = (count + blocks_per_page -
1182 1) / blocks_per_page - 1;
1183 lnb1 = iobuf->dr_lnbs[i - clen];
1184 lnb2 = iobuf->dr_lnbs[mapped_index];
1185 size1 = lnb1->lnb_file_offset -
1186 (lnb1->lnb_file_offset % PAGE_SIZE) +
1187 (total << inode->i_blkbits);
1188 size2 = lnb2->lnb_file_offset + lnb2->lnb_len;
1192 if (size1 > disk_size)
1196 if (rc == 0 && total < blen) {
1198 * decay extent blocks if we could not
1199 * allocate extent once.
1201 osd_decay_extent_bytes(osd,
1202 (total - previous_total) << inode->i_blkbits);
1203 map.m_lblk = fp->index * blocks_per_page + total;
1204 map.m_len = blen - total;
1205 previous_total = total;
1211 * decay extent blocks if we could allocate
1212 * good large extent.
1214 if (total - previous_total >=
1215 osd_extent_bytes(osd) >> inode->i_blkbits)
1216 osd_decay_extent_bytes(osd,
1217 (total - previous_total) << inode->i_blkbits);
1218 /* look for next extent */
1220 blocks += blocks_per_page * clen;
1223 if (rc == 0 && create &&
1224 start_blocks < pages * blocks_per_page) {
1225 rc = osd_ldiskfs_map_write(inode, iobuf, osd, start_blocks,
1226 count, &disk_size, user_size);
1227 LASSERT(start_blocks + count == pages * blocks_per_page);
1232 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1233 struct niobuf_local *lnb, int npages)
1235 struct osd_thread_info *oti = osd_oti_get(env);
1236 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1237 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1238 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1247 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1248 if (unlikely(rc != 0))
1251 isize = i_size_read(inode);
1252 maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1254 start = ktime_get();
1255 for (i = 0; i < npages; i++) {
1258 * till commit the content of the page is undefined
1259 * we'll set it uptodate once bulk is done. otherwise
1260 * subsequent reads can access non-stable data
1262 ClearPageUptodate(lnb[i].lnb_page);
1264 if (lnb[i].lnb_len == PAGE_SIZE)
1267 if (maxidx >= lnb[i].lnb_page->index) {
1268 osd_iobuf_add_page(iobuf, &lnb[i]);
1271 char *p = kmap(lnb[i].lnb_page);
1273 off = lnb[i].lnb_page_offset;
1276 off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1279 memset(p + off, 0, PAGE_SIZE - off);
1280 kunmap(lnb[i].lnb_page);
1284 timediff = ktime_us_delta(end, start);
1285 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1287 if (iobuf->dr_npages) {
1288 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
1290 if (likely(rc == 0)) {
1291 rc = osd_do_bio(osd, inode, iobuf, 0, 0);
1292 /* do IO stats for preparation reads */
1293 osd_fini_iobuf(osd, iobuf);
1299 struct osd_fextent {
1302 unsigned int mapped:1;
1305 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1306 struct osd_fextent *cached_extent)
1308 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1309 sector_t block = offset >> inode->i_blkbits;
1311 struct fiemap_extent_info fei = { 0 };
1312 struct fiemap_extent fe = { 0 };
1315 if (block >= cached_extent->start && block < cached_extent->end)
1316 return cached_extent->mapped;
1318 if (i_size_read(inode) == 0)
1321 /* Beyond EOF, must not be mapped */
1322 if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1325 fei.fi_extents_max = 1;
1326 fei.fi_extents_start = &fe;
1328 rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1332 start = fe.fe_logical >> inode->i_blkbits;
1333 if (fei.fi_extents_mapped == 0) {
1334 /* a special case - no extent found at this offset and forward.
1335 * we can consider this as a hole to EOF. it's safe to cache
1336 * as other threads can not allocate/punch blocks this thread
1337 * is working on (LDLM). */
1338 cached_extent->start = block;
1339 cached_extent->end = i_size_read(inode) >> inode->i_blkbits;
1340 cached_extent->mapped = 0;
1344 if (start > block) {
1345 cached_extent->start = block;
1346 cached_extent->end = start;
1347 cached_extent->mapped = 0;
1349 cached_extent->start = start;
1350 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1352 cached_extent->mapped = 1;
1355 return cached_extent->mapped;
1358 #define MAX_EXTENTS_PER_WRITE 100
1359 static int osd_declare_write_commit(const struct lu_env *env,
1360 struct dt_object *dt,
1361 struct niobuf_local *lnb, int npages,
1362 struct thandle *handle)
1364 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1365 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1366 struct osd_thandle *oh;
1373 long long quota_space = 0;
1374 struct osd_fextent mapped = { 0 }, extent = { 0 };
1375 enum osd_quota_local_flags local_flags = 0;
1376 enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1377 unsigned int extent_bytes;
1380 LASSERT(handle != NULL);
1381 oh = container_of(handle, struct osd_thandle, ot_super);
1382 LASSERT(oh->ot_handle == NULL);
1385 * We track a decaying average extent blocks per filesystem,
1386 * for most of time, it will be 1M, with filesystem becoming
1387 * heavily-fragmented, it will be reduced to 4K at the worst.
1389 extent_bytes = osd_extent_bytes(osd);
1390 LASSERT(extent_bytes >= (1 << osd_sb(osd)->s_blocksize));
1392 /* calculate number of extents (probably better to pass nb) */
1393 for (i = 0; i < npages; i++) {
1394 /* ignore quota for the whole request if any page is from
1395 * client cache or written by root.
1397 * XXX once we drop the 1.8 client support, the checking
1398 * for whether page is from cache can be simplified as:
1399 * !(lnb[i].flags & OBD_BRW_SYNC)
1401 * XXX we could handle this on per-lnb basis as done by
1404 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1405 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1407 declare_flags |= OSD_QID_FORCE;
1409 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &mapped)) {
1410 lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1414 if (lnb[i].lnb_flags & OBD_BRW_DONE) {
1415 lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1419 /* count only unmapped changes */
1421 if (lnb[i].lnb_file_offset != extent.end || extent.end == 0) {
1422 if (extent.end != 0)
1423 extents += (extent.end - extent.start +
1424 extent_bytes - 1) / extent_bytes;
1425 extent.start = lnb[i].lnb_file_offset;
1426 extent.end = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1428 extent.end += lnb[i].lnb_len;
1431 quota_space += PAGE_SIZE;
1434 credits++; /* inode */
1436 * overwrite case, no need to modify tree and
1442 extents += (extent.end - extent.start +
1443 extent_bytes - 1) / extent_bytes;
1445 * with system space usage growing up, mballoc codes won't
1446 * try best to scan block group to align best free extent as
1447 * we can. So extent bytes per extent could be decayed to a
1448 * very small value, this could make us reserve too many credits.
1449 * We could be more optimistic in the credit reservations, even
1450 * in a case where the filesystem is nearly full, it is extremely
1451 * unlikely that the worst case would ever be hit.
1453 if (extents > MAX_EXTENTS_PER_WRITE)
1454 extents = MAX_EXTENTS_PER_WRITE;
1457 * each extent can go into new leaf causing a split
1458 * 5 is max tree depth: inode + 4 index blocks
1459 * with blockmaps, depth is 3 at most
1461 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1463 * many concurrent threads may grow tree by the time
1464 * our transaction starts. so, consider 2 is a min depth
1466 depth = ext_depth(inode);
1467 depth = max(depth, 1) + 1;
1469 credits += depth * 2 * extents;
1473 credits += depth * extents;
1476 oh->oh_declared_ext = extents;
1478 /* quota space for metadata blocks */
1479 quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1481 /* quota space should be reported in 1K blocks */
1482 quota_space = toqb(quota_space);
1484 /* each new block can go in different group (bitmap + gd) */
1486 /* we can't dirty more bitmap blocks than exist */
1487 if (extents > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1488 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1492 /* we can't dirty more gd blocks than exist */
1493 if (extents > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1494 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1499 "%s: inode #%lu extent_bytes %u extents %d credits %d\n",
1500 osd_ino2name(inode), inode->i_ino, extent_bytes, extents,
1504 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1506 /* make sure the over quota flags were not set */
1507 lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1509 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1510 i_projid_read(inode), quota_space, oh,
1511 osd_dt_obj(dt), &local_flags, declare_flags);
1513 /* we need only to store the overquota flags in the first lnb for
1514 * now, once we support multiple objects BRW, this code needs be
1517 if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1518 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1519 if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1520 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1521 if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1522 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1525 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1530 /* Check if a block is allocated or not */
1531 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1532 struct niobuf_local *lnb, int npages,
1533 struct thandle *thandle, __u64 user_size)
1535 struct osd_thread_info *oti = osd_oti_get(env);
1536 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1537 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1538 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1539 int rc = 0, i, check_credits = 0;
1543 rc = osd_init_iobuf(osd, iobuf, 1, npages);
1544 if (unlikely(rc != 0))
1547 dquot_initialize(inode);
1549 for (i = 0; i < npages; i++) {
1550 if (lnb[i].lnb_rc == -ENOSPC &&
1551 (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1552 /* Allow the write to proceed if overwriting an
1558 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1559 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1561 LASSERT(lnb[i].lnb_page);
1562 generic_error_remove_page(inode->i_mapping,
1567 if (lnb[i].lnb_flags & OBD_BRW_DONE)
1570 if (!(lnb[i].lnb_flags & OBD_BRW_MAPPED))
1573 LASSERT(PageLocked(lnb[i].lnb_page));
1574 LASSERT(!PageWriteback(lnb[i].lnb_page));
1577 * Since write and truncate are serialized by oo_sem, even
1578 * partial-page truncate should not leave dirty pages in the
1581 LASSERT(!PageDirty(lnb[i].lnb_page));
1583 SetPageUptodate(lnb[i].lnb_page);
1585 osd_iobuf_add_page(iobuf, &lnb[i]);
1588 osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1590 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1592 } else if (iobuf->dr_npages > 0) {
1593 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd,
1598 /* no pages to write, no transno is needed */
1599 thandle->th_local = 1;
1602 if (rc != 0 && !thandle->th_restart_tran)
1603 osd_fini_iobuf(osd, iobuf);
1605 osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1607 if (unlikely(rc != 0 && !thandle->th_restart_tran)) {
1608 /* if write fails, we should drop pages from the cache */
1609 for (i = 0; i < npages; i++) {
1610 if (lnb[i].lnb_page == NULL)
1612 if (!PagePrivate2(lnb[i].lnb_page)) {
1613 LASSERT(PageLocked(lnb[i].lnb_page));
1614 generic_error_remove_page(inode->i_mapping,
1623 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1624 struct niobuf_local *lnb, int npages)
1626 struct osd_thread_info *oti = osd_oti_get(env);
1627 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1628 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1629 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1630 int rc = 0, i, cache_hits = 0, cache_misses = 0;
1637 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1638 if (unlikely(rc != 0))
1641 isize = i_size_read(inode);
1643 start = ktime_get();
1644 for (i = 0; i < npages; i++) {
1646 if (isize <= lnb[i].lnb_file_offset)
1647 /* If there's no more data, abort early.
1648 * lnb->lnb_rc == 0, so it's easy to detect later.
1652 /* instead of looking if we go beyong isize, send complete
1653 * pages all the time
1655 lnb[i].lnb_rc = lnb[i].lnb_len;
1657 /* Bypass disk read if fail_loc is set properly */
1658 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1659 SetPageUptodate(lnb[i].lnb_page);
1661 if (PageUptodate(lnb[i].lnb_page)) {
1663 unlock_page(lnb[i].lnb_page);
1666 osd_iobuf_add_page(iobuf, &lnb[i]);
1668 /* no need to unlock in osd_bufs_put(), the sooner page is
1669 * unlocked, the earlier another client can access it.
1670 * notice real unlock_page() can be called few lines
1671 * below after osd_do_bio(). lnb is a per-thread, so it's
1672 * fine to have PG_locked and lnb_locked inconsistent here
1674 lnb[i].lnb_locked = 0;
1677 timediff = ktime_us_delta(end, start);
1678 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1680 if (cache_hits != 0)
1681 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1683 if (cache_misses != 0)
1684 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1686 if (cache_hits + cache_misses != 0)
1687 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1688 cache_hits + cache_misses);
1690 if (iobuf->dr_npages) {
1691 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
1694 rc = osd_do_bio(osd, inode, iobuf, 0, 0);
1696 /* IO stats will be done in osd_bufs_put() */
1698 /* early release to let others read data during the bulk */
1699 for (i = 0; i < iobuf->dr_npages; i++) {
1700 LASSERT(PageLocked(iobuf->dr_pages[i]));
1701 if (!PagePrivate2(iobuf->dr_pages[i]))
1702 unlock_page(iobuf->dr_pages[i]);
1710 * XXX: Another layering violation for now.
1712 * We don't want to use ->f_op->read methods, because generic file write
1714 * - serializes on ->i_sem, and
1716 * - does a lot of extra work like balance_dirty_pages(),
1718 * which doesn't work for globally shared files like /last_rcvd.
1720 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1722 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1724 memcpy(buffer, (char *)ei->i_data, buflen);
1729 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1731 struct buffer_head *bh;
1732 unsigned long block;
1738 /* prevent reading after eof */
1739 spin_lock(&inode->i_lock);
1740 if (i_size_read(inode) < *offs + size) {
1741 loff_t diff = i_size_read(inode) - *offs;
1743 spin_unlock(&inode->i_lock);
1746 "size %llu is too short to read @%llu\n",
1747 i_size_read(inode), *offs);
1749 } else if (diff == 0) {
1755 spin_unlock(&inode->i_lock);
1758 blocksize = 1 << inode->i_blkbits;
1761 block = *offs >> inode->i_blkbits;
1762 boffs = *offs & (blocksize - 1);
1763 csize = min(blocksize - boffs, size);
1764 bh = __ldiskfs_bread(NULL, inode, block, 0);
1766 CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1767 osd_ino2name(inode), csize, *offs, inode->i_ino,
1773 memcpy(buf, bh->b_data + boffs, csize);
1776 memset(buf, 0, csize);
1786 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1787 struct lu_buf *buf, loff_t *pos)
1789 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1792 /* Read small symlink from inode body as we need to maintain correct
1793 * on-disk symlinks for ldiskfs.
1795 if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1796 loff_t size = i_size_read(inode);
1798 if (buf->lb_len < size)
1801 if (size < sizeof(LDISKFS_I(inode)->i_data))
1802 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1804 rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1806 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1812 static inline int osd_extents_enabled(struct super_block *sb,
1813 struct inode *inode)
1815 if (inode != NULL) {
1816 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1818 } else if (ldiskfs_has_feature_extents(sb)) {
1824 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1825 const loff_t size, const loff_t pos,
1828 int credits, bits, bs, i;
1830 bits = sb->s_blocksize_bits;
1833 /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1834 * we do not expect blockmaps on the large files,
1835 * so let's shrink it to 2 levels (4GB files)
1838 /* this is default reservation: 2 levels */
1839 credits = (blocks + 2) * 3;
1841 /* actual offset is unknown, hard to optimize */
1845 /* now check for few specific cases to optimize */
1846 if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1849 /* allocate if not allocated */
1850 if (inode == NULL) {
1851 credits += blocks * 2;
1854 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1855 LASSERT(i < LDISKFS_NDIR_BLOCKS);
1856 if (LDISKFS_I(inode)->i_data[i] == 0)
1859 } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1860 /* single indirect */
1861 credits = blocks * 3;
1862 if (inode == NULL ||
1863 LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1866 /* The indirect block may be modified. */
1873 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1874 const struct lu_buf *buf, loff_t _pos,
1875 struct thandle *handle)
1877 struct osd_object *obj = osd_dt_obj(dt);
1878 struct inode *inode = obj->oo_inode;
1879 struct super_block *sb = osd_sb(osd_obj2dev(obj));
1880 struct osd_thandle *oh;
1881 int rc = 0, est = 0, credits, blocks, allocated = 0;
1887 LASSERT(buf != NULL);
1888 LASSERT(handle != NULL);
1890 oh = container_of(handle, struct osd_thandle, ot_super);
1891 LASSERT(oh->ot_handle == NULL);
1894 bits = sb->s_blocksize_bits;
1898 /* if this is an append, then we
1899 * should expect cross-block record
1906 /* blocks to modify */
1907 blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1908 LASSERT(blocks > 0);
1910 if (inode != NULL && _pos != -1) {
1911 /* object size in blocks */
1912 est = (i_size_read(inode) + bs - 1) >> bits;
1913 allocated = inode->i_blocks >> (bits - 9);
1914 if (pos + size <= i_size_read(inode) && est <= allocated) {
1915 /* looks like an overwrite, no need to modify tree */
1917 /* no need to modify i_size */
1922 if (osd_extents_enabled(sb, inode)) {
1924 * many concurrent threads may grow tree by the time
1925 * our transaction starts. so, consider 2 is a min depth
1926 * for every level we may need to allocate a new block
1927 * and take some entries from the old one. so, 3 blocks
1928 * to allocate (bitmap, gd, itself) + old block - 4 per
1931 depth = inode != NULL ? ext_depth(inode) : 0;
1932 depth = max(depth, 1) + 1;
1934 /* if not append, then split may need to modify
1935 * existing blocks moving entries into the new ones
1939 /* blocks to store data: bitmap,gd,itself */
1940 credits += blocks * 3;
1942 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1944 /* if inode is created as part of the transaction,
1945 * then it's counted already by the creation method
1952 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1954 /* dt_declare_write() is usually called for system objects, such
1955 * as llog or last_rcvd files. We needn't enforce quota on those
1956 * objects, so always set the lqi_space as 0.
1959 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1961 i_projid_read(inode), 0,
1962 oh, obj, NULL, OSD_QID_BLK);
1965 rc = osd_trunc_lock(obj, oh, true);
1970 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1972 /* LU-2634: clear the extent format for fast symlink */
1973 ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1975 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1976 spin_lock(&inode->i_lock);
1977 LDISKFS_I(inode)->i_disksize = buflen;
1978 i_size_write(inode, buflen);
1979 spin_unlock(&inode->i_lock);
1980 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1985 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
1986 int bufsize, int write_NUL, loff_t *offs,
1989 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1990 struct buffer_head *bh = NULL;
1991 loff_t offset = *offs;
1992 loff_t new_size = i_size_read(inode);
1993 unsigned long block;
1994 int blocksize = 1 << inode->i_blkbits;
1995 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1999 int dirty_inode = 0;
2000 bool create, sparse, sync = false;
2004 * long symlink write does not count the NUL terminator in
2005 * bufsize, we write it, and the inode's file size does not
2006 * count the NUL terminator as well.
2008 ((char *)buf)[bufsize] = '\0';
2012 /* only the first flag-set matters */
2013 dirty_inode = !test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA,
2016 /* sparse checking is racy, but sparse is very rare case, leave as is */
2017 sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
2018 ((new_size - 1) >> inode->i_blkbits) + 1);
2020 while (bufsize > 0) {
2021 int credits = handle->h_buffer_credits;
2022 unsigned long last_block = (new_size == 0) ? 0 :
2023 (new_size - 1) >> inode->i_blkbits;
2028 block = offset >> inode->i_blkbits;
2029 boffs = offset & (blocksize - 1);
2030 size = min(blocksize - boffs, bufsize);
2031 sync = (block > last_block || new_size == 0 || sparse);
2034 down(&ei->i_append_sem);
2036 bh = __ldiskfs_bread(handle, inode, block, 0);
2038 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
2040 "%s: adding bh without locking off %llu (block %lu, size %d, offs %llu)\n",
2041 osd_ino2name(inode),
2042 offset, block, bufsize, *offs);
2044 if (IS_ERR_OR_NULL(bh)) {
2045 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
2046 int flags = LDISKFS_GET_BLOCKS_CREATE;
2048 /* while the file system is being mounted, avoid
2049 * preallocation otherwise mount can take a long
2050 * time as mballoc cache is cold.
2051 * XXX: this is a workaround until we have a proper
2053 * XXX: works with extent-based files only */
2054 if (!osd->od_cl_seq)
2055 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
2056 bh = __ldiskfs_bread(handle, inode, block, flags);
2060 up(&ei->i_append_sem);
2065 if (IS_ERR_OR_NULL(bh)) {
2074 "%s: error reading offset %llu (block %lu, size %d, offs %llu), credits %d/%d: rc = %d\n",
2075 osd_ino2name(inode), offset, block, bufsize,
2076 *offs, credits, handle->h_buffer_credits, err);
2080 err = ldiskfs_journal_get_write_access(handle, bh);
2082 CERROR("journal_get_write_access() returned error %d\n",
2086 LASSERTF(boffs + size <= bh->b_size,
2087 "boffs %d size %d bh->b_size %lu\n",
2088 boffs, size, (unsigned long)bh->b_size);
2090 memset(bh->b_data, 0, bh->b_size);
2092 up(&ei->i_append_sem);
2096 memcpy(bh->b_data + boffs, buf, size);
2097 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2101 if (offset + size > new_size)
2102 new_size = offset + size;
2108 up(&ei->i_append_sem);
2115 /* correct in-core and on-disk sizes */
2116 if (new_size > i_size_read(inode)) {
2117 spin_lock(&inode->i_lock);
2118 if (new_size > i_size_read(inode))
2119 i_size_write(inode, new_size);
2120 if (i_size_read(inode) > ei->i_disksize) {
2121 ei->i_disksize = i_size_read(inode);
2124 spin_unlock(&inode->i_lock);
2127 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2134 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
2135 const struct lu_buf *buf, loff_t *pos,
2136 struct thandle *handle)
2138 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2139 struct osd_thandle *oh;
2143 LASSERT(dt_object_exists(dt));
2145 LASSERT(handle != NULL);
2146 LASSERT(inode != NULL);
2147 dquot_initialize(inode);
2149 /* XXX: don't check: one declared chunk can be used many times */
2150 /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
2152 oh = container_of(handle, struct osd_thandle, ot_super);
2153 LASSERT(oh->ot_handle->h_transaction != NULL);
2154 osd_trans_exec_op(env, handle, OSD_OT_WRITE);
2156 /* Write small symlink to inode body as we need to maintain correct
2157 * on-disk symlinks for ldiskfs.
2158 * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
2159 * does not count it in.
2161 is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
2162 if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
2163 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
2165 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
2166 is_link, pos, oh->ot_handle);
2168 result = buf->lb_len;
2170 osd_trans_exec_check(env, handle, OSD_OT_WRITE);
2175 static int osd_declare_fallocate(const struct lu_env *env,
2176 struct dt_object *dt, __u64 start, __u64 end,
2177 int mode, struct thandle *th)
2179 struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
2180 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
2181 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2182 long long quota_space = 0;
2183 /* 5 is max tree depth. (inode + 4 index blocks) */
2190 * mode == 0 (which is standard prealloc) and PUNCH is supported
2191 * Rest of mode options is not supported yet.
2193 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2194 RETURN(-EOPNOTSUPP);
2196 /* disable fallocate completely */
2197 if (osd_dev(dt->do_lu.lo_dev)->od_fallocate_zero_blocks < 0)
2198 RETURN(-EOPNOTSUPP);
2203 if (mode & FALLOC_FL_PUNCH_HOLE) {
2204 rc = osd_declare_inode_qid(env, i_uid_read(inode),
2206 i_projid_read(inode), 0, oh,
2207 osd_dt_obj(dt), NULL, OSD_QID_BLK);
2209 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2213 /* quota space for metadata blocks
2214 * approximate metadata estimate should be good enough.
2216 quota_space += PAGE_SIZE;
2217 quota_space += depth * LDISKFS_BLOCK_SIZE(osd_sb(osd));
2219 /* quota space should be reported in 1K blocks */
2220 quota_space = toqb(quota_space) + toqb(end - start) +
2221 LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
2223 /* We don't need to reserve credits for whole fallocate here.
2224 * We reserve space only for metadata. Fallocate credits are
2225 * extended as required
2227 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2228 i_projid_read(inode), quota_space, oh,
2229 osd_dt_obj(dt), NULL, OSD_QID_BLK);
2233 static int osd_fallocate_preallocate(const struct lu_env *env,
2234 struct dt_object *dt,
2235 __u64 start, __u64 end, int mode,
2238 struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
2239 handle_t *handle = ldiskfs_journal_current_handle();
2240 unsigned int save_credits = oh->ot_credits;
2241 struct osd_object *obj = osd_dt_obj(dt);
2242 struct inode *inode = obj->oo_inode;
2243 struct ldiskfs_map_blocks map;
2244 unsigned int credits;
2245 ldiskfs_lblk_t blen;
2246 ldiskfs_lblk_t boff;
2247 loff_t new_size = 0;
2254 LASSERT(dt_object_exists(dt));
2255 LASSERT(osd_invariant(obj));
2256 LASSERT(inode != NULL);
2258 CDEBUG(D_INODE, "fallocate: inode #%lu: start %llu end %llu mode %d\n",
2259 inode->i_ino, start, end, mode);
2261 dquot_initialize(inode);
2265 boff = start >> inode->i_blkbits;
2266 blen = (ALIGN(end, 1 << inode->i_blkbits) >> inode->i_blkbits) - boff;
2268 /* Create and mark new extents as either zero or unwritten */
2269 flags = osd_dev(dt->do_lu.lo_dev)->od_fallocate_zero_blocks ?
2270 LDISKFS_GET_BLOCKS_CREATE_ZERO :
2271 LDISKFS_GET_BLOCKS_CREATE_UNWRIT_EXT;
2272 if (mode & FALLOC_FL_KEEP_SIZE)
2273 flags |= LDISKFS_GET_BLOCKS_KEEP_SIZE;
2278 * We only support preallocation for extent-based file only.
2280 if (!(ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS)))
2281 GOTO(out, rc = -EOPNOTSUPP);
2283 if (!(mode & FALLOC_FL_KEEP_SIZE) && (end > i_size_read(inode) ||
2284 end > LDISKFS_I(inode)->i_disksize)) {
2286 rc = inode_newsize_ok(inode, new_size);
2291 inode_dio_wait(inode);
2296 /* Don't normalize the request if it can fit in one extent so
2297 * that it doesn't get unnecessarily split into multiple extents.
2299 if (blen <= EXT_UNWRITTEN_MAX_LEN)
2300 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
2303 * credits to insert 1 extent into extent tree.
2305 credits = osd_chunk_trans_blocks(inode, blen);
2306 depth = ext_depth(inode);
2308 while (rc >= 0 && blen) {
2312 * Recalculate credits when extent tree depth changes.
2314 if (depth != ext_depth(inode)) {
2315 credits = osd_chunk_trans_blocks(inode, blen);
2316 depth = ext_depth(inode);
2319 /* TODO: quota check */
2320 rc = osd_extend_restart_trans(handle, credits, inode);
2324 rc = ldiskfs_map_blocks(handle, inode, &map, flags);
2327 "inode #%lu: block %u: len %u: ldiskfs_map_blocks returned %d\n",
2328 inode->i_ino, map.m_lblk, map.m_len, rc);
2329 ldiskfs_mark_inode_dirty(handle, inode);
2334 map.m_len = blen = blen - rc;
2335 epos = (loff_t)map.m_lblk << inode->i_blkbits;
2336 inode->i_ctime = current_time(inode);
2340 if (ldiskfs_update_inode_size(inode, epos) & 0x1)
2341 inode->i_mtime = inode->i_ctime;
2343 if (epos > inode->i_size)
2344 ldiskfs_set_inode_flag(inode,
2345 LDISKFS_INODE_EOFBLOCKS);
2348 ldiskfs_mark_inode_dirty(handle, inode);
2352 /* extand credits if needed for operations such as attribute set */
2354 rc = osd_extend_restart_trans(handle, save_credits, inode);
2356 inode_unlock(inode);
2361 static int osd_fallocate_punch(const struct lu_env *env, struct dt_object *dt,
2362 __u64 start, __u64 end, int mode,
2365 struct osd_object *obj = osd_dt_obj(dt);
2366 struct inode *inode = obj->oo_inode;
2367 struct osd_access_lock *al;
2368 struct osd_thandle *oh;
2369 int rc = 0, found = 0;
2373 LASSERT(dt_object_exists(dt));
2374 LASSERT(osd_invariant(obj));
2375 LASSERT(inode != NULL);
2377 dquot_initialize(inode);
2380 oh = container_of(th, struct osd_thandle, ot_super);
2381 LASSERT(oh->ot_handle->h_transaction != NULL);
2383 list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2384 if (obj != al->tl_obj)
2386 LASSERT(al->tl_shared == 0);
2388 /* do actual punch in osd_trans_stop() */
2389 al->tl_start = start;
2392 al->tl_punch = true;
2399 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
2400 __u64 start, __u64 end, int mode, struct thandle *th)
2406 if (mode & FALLOC_FL_PUNCH_HOLE) {
2408 rc = osd_fallocate_punch(env, dt, start, end, mode, th);
2410 /* standard preallocate */
2411 rc = osd_fallocate_preallocate(env, dt, start, end, mode, th);
2416 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
2417 __u64 start, __u64 end, struct thandle *th)
2419 struct osd_thandle *oh;
2420 struct inode *inode;
2425 oh = container_of(th, struct osd_thandle, ot_super);
2428 * we don't need to reserve credits for whole truncate
2429 * it's not possible as truncate may need to free too many
2430 * blocks and that won't fit a single transaction. instead
2431 * we reserve credits to change i_size and put inode onto
2432 * orphan list. if needed truncate will extend or restart
2435 osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
2436 osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
2438 inode = osd_dt_obj(dt)->oo_inode;
2441 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2442 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
2446 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2451 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
2452 __u64 start, __u64 end, struct thandle *th)
2454 struct osd_object *obj = osd_dt_obj(dt);
2455 struct osd_device *osd = osd_obj2dev(obj);
2456 struct inode *inode = obj->oo_inode;
2457 struct osd_access_lock *al;
2458 struct osd_thandle *oh;
2459 int rc = 0, found = 0;
2463 LASSERT(dt_object_exists(dt));
2464 LASSERT(osd_invariant(obj));
2465 LASSERT(inode != NULL);
2466 dquot_initialize(inode);
2469 oh = container_of(th, struct osd_thandle, ot_super);
2470 LASSERT(oh->ot_handle->h_transaction != NULL);
2472 /* we used to skip truncate to current size to
2473 * optimize truncates on OST. with DoM we can
2474 * get attr_set to set specific size (MDS_REINT)
2475 * and then get truncate RPC which essentially
2476 * would be skipped. this is bad.. so, disable
2477 * this optimization on MDS till the client stop
2478 * to sent MDS_REINT (LU-11033) -bzzz
2480 if (osd->od_is_ost && i_size_read(inode) == start)
2483 osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2485 spin_lock(&inode->i_lock);
2486 if (i_size_read(inode) < start)
2488 i_size_write(inode, start);
2489 spin_unlock(&inode->i_lock);
2490 /* if object holds encrypted content, we need to make sure we truncate
2491 * on an encryption unit boundary, or subsequent reads will get
2494 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2495 start & ~LUSTRE_ENCRYPTION_MASK)
2496 start = (start & LUSTRE_ENCRYPTION_MASK) +
2497 LUSTRE_ENCRYPTION_UNIT_SIZE;
2498 ll_truncate_pagecache(inode, start);
2500 /* optimize grow case */
2502 osd_execute_truncate(obj);
2507 /* add to orphan list to ensure truncate completion
2508 * if this transaction succeed. ldiskfs_truncate()
2509 * will take the inode out of the list
2511 rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2512 inode_unlock(inode);
2516 list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2517 if (obj != al->tl_obj)
2519 LASSERT(al->tl_shared == 0);
2521 /* do actual truncate in osd_trans_stop() */
2522 al->tl_truncate = 1;
2531 static int fiemap_check_ranges(struct inode *inode,
2532 u64 start, u64 len, u64 *new_len)
2541 if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2542 maxbytes = inode->i_sb->s_maxbytes;
2544 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2546 if (start > maxbytes)
2550 * Shrink request scope to what the fs can actually handle.
2552 if (len > maxbytes || (maxbytes - len) < start)
2553 *new_len = maxbytes - start;
2558 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2559 #define FIEMAP_MAX_EXTENTS (UINT_MAX / sizeof(struct fiemap_extent))
2561 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2564 struct fiemap_extent_info fieinfo = {0, };
2565 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2570 if (inode->i_op->fiemap == NULL)
2573 if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2576 rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2580 fieinfo.fi_flags = fm->fm_flags;
2581 fieinfo.fi_extents_max = fm->fm_extent_count;
2582 fieinfo.fi_extents_start = fm->fm_extents;
2584 if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2585 filemap_write_and_wait(inode->i_mapping);
2587 rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2588 fm->fm_flags = fieinfo.fi_flags;
2589 fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2594 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2595 __u64 start, __u64 end, enum lu_ladvise_type advice)
2597 struct osd_object *obj = osd_dt_obj(dt);
2602 case LU_LADVISE_DONTNEED:
2604 invalidate_mapping_pages(obj->oo_inode->i_mapping,
2605 start >> PAGE_SHIFT,
2606 (end - 1) >> PAGE_SHIFT);
2616 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
2617 loff_t offset, int whence)
2619 struct osd_object *obj = osd_dt_obj(dt);
2620 struct inode *inode = obj->oo_inode;
2626 LASSERT(dt_object_exists(dt));
2627 LASSERT(osd_invariant(obj));
2629 LASSERT(offset >= 0);
2631 file = osd_quasi_file(env, inode);
2632 result = file->f_op->llseek(file, offset, whence);
2635 * If 'offset' is beyond end of object file then treat it as not error
2636 * but valid case for SEEK_HOLE and return 'offset' as result.
2637 * LOV will decide if it is beyond real end of file or not.
2639 if (whence == SEEK_HOLE && result == -ENXIO)
2642 CDEBUG(D_INFO, "seek %s from %lld: %lld\n", whence == SEEK_HOLE ?
2643 "hole" : "data", offset, result);
2648 * in some cases we may need declare methods for objects being created
2649 * e.g., when we create symlink
2651 const struct dt_body_operations osd_body_ops_new = {
2652 .dbo_declare_write = osd_declare_write,
2655 const struct dt_body_operations osd_body_ops = {
2656 .dbo_read = osd_read,
2657 .dbo_declare_write = osd_declare_write,
2658 .dbo_write = osd_write,
2659 .dbo_bufs_get = osd_bufs_get,
2660 .dbo_bufs_put = osd_bufs_put,
2661 .dbo_write_prep = osd_write_prep,
2662 .dbo_declare_write_commit = osd_declare_write_commit,
2663 .dbo_write_commit = osd_write_commit,
2664 .dbo_read_prep = osd_read_prep,
2665 .dbo_declare_punch = osd_declare_punch,
2666 .dbo_punch = osd_punch,
2667 .dbo_fiemap_get = osd_fiemap_get,
2668 .dbo_ladvise = osd_ladvise,
2669 .dbo_declare_fallocate = osd_declare_fallocate,
2670 .dbo_fallocate = osd_fallocate,
2671 .dbo_lseek = osd_lseek,
2675 * Get a truncate lock
2677 * In order to take multi-transaction truncate out of main transaction we let
2678 * the caller grab a lock on the object passed. the lock can be shared (for
2679 * writes) and exclusive (for truncate). It's not allowed to mix truncate
2680 * and write in the same transaction handle (do not confuse with big ldiskfs
2681 * transaction containing lots of handles).
2682 * The lock must be taken at declaration.
2684 * \param obj object to lock
2686 * \shared shared or exclusive
2688 * \retval 0 lock is granted
2689 * \retval -NOMEM no memory to allocate lock
2691 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2693 struct osd_access_lock *al, *tmp;
2698 list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2699 if (tmp->tl_obj != obj)
2701 LASSERT(tmp->tl_shared == shared);
2702 /* found same lock */
2707 if (unlikely(al == NULL))
2710 al->tl_truncate = false;
2712 down_read(&obj->oo_ext_idx_sem);
2714 down_write(&obj->oo_ext_idx_sem);
2715 al->tl_shared = shared;
2716 lu_object_get(&obj->oo_dt.do_lu);
2718 list_add(&al->tl_list, &oh->ot_trunc_locks);
2723 void osd_trunc_unlock_all(const struct lu_env *env, struct list_head *list)
2725 struct osd_access_lock *al, *tmp;
2727 list_for_each_entry_safe(al, tmp, list, tl_list) {
2729 up_read(&al->tl_obj->oo_ext_idx_sem);
2731 up_write(&al->tl_obj->oo_ext_idx_sem);
2732 osd_object_put(env, al->tl_obj);
2733 list_del(&al->tl_list);
2739 * For a partial-page truncate, flush the page to disk immediately to
2740 * avoid data corruption during direct disk write. b=17397
2742 static void osd_partial_page_flush(struct osd_device *d, struct inode *inode,
2745 if (!(offset & ~PAGE_MASK))
2748 if (osd_use_page_cache(d)) {
2749 filemap_fdatawrite_range(inode->i_mapping, offset, offset + 1);
2751 /* Notice we use "wait" version to ensure I/O is complete */
2752 filemap_write_and_wait_range(inode->i_mapping, offset,
2754 invalidate_mapping_pages(inode->i_mapping, offset >> PAGE_SHIFT,
2755 offset >> PAGE_SHIFT);
2759 void osd_execute_truncate(struct osd_object *obj)
2761 struct osd_device *d = osd_obj2dev(obj);
2762 struct inode *inode = obj->oo_inode;
2765 /* simulate crash before (in the middle) of delayed truncate */
2766 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2767 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2768 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2770 mutex_lock(&sbi->s_orphan_lock);
2771 list_del_init(&ei->i_orphan);
2772 mutex_unlock(&sbi->s_orphan_lock);
2776 size = i_size_read(inode);
2778 /* if object holds encrypted content, we need to make sure we truncate
2779 * on an encryption unit boundary, or block content will get corrupted
2781 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2782 size & ~LUSTRE_ENCRYPTION_MASK)
2783 inode->i_size = (size & LUSTRE_ENCRYPTION_MASK) +
2784 LUSTRE_ENCRYPTION_UNIT_SIZE;
2785 ldiskfs_truncate(inode);
2786 inode_unlock(inode);
2787 if (inode->i_size != size) {
2788 spin_lock(&inode->i_lock);
2789 i_size_write(inode, size);
2790 LDISKFS_I(inode)->i_disksize = size;
2791 spin_unlock(&inode->i_lock);
2792 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2794 osd_partial_page_flush(d, inode, size);
2797 void osd_execute_punch(const struct lu_env *env, struct osd_object *obj,
2798 loff_t start, loff_t end, int mode)
2800 struct osd_device *d = osd_obj2dev(obj);
2801 struct inode *inode = obj->oo_inode;
2802 struct file *file = osd_quasi_file(env, inode);
2804 file->f_op->fallocate(file, mode, start, end - start);
2805 osd_partial_page_flush(d, inode, start);
2806 osd_partial_page_flush(d, inode, end - 1);
2809 void osd_process_truncates(const struct lu_env *env, struct list_head *list)
2811 struct osd_access_lock *al;
2813 LASSERT(journal_current_handle() == NULL);
2815 list_for_each_entry(al, list, tl_list) {
2818 if (al->tl_truncate)
2819 osd_execute_truncate(al->tl_obj);
2820 else if (al->tl_punch)
2821 osd_execute_punch(env, al->tl_obj, al->tl_start,
2822 al->tl_end, al->tl_mode);