X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosd-ldiskfs%2Fosd_io.c;h=961e10dc42f47b24ba9a49df4bae3c7406fa7006;hp=97655955e5af4a2a9b8524035f7dd52ae7064f57;hb=83d660436a164758fd4a29c1433d11c0f4591196;hpb=020941416419ab282f3d9b694014b2059d299d51 diff --git a/lustre/osd-ldiskfs/osd_io.c b/lustre/osd-ldiskfs/osd_io.c index 9765595..961e10d 100644 --- a/lustre/osd-ldiskfs/osd_io.c +++ b/lustre/osd-ldiskfs/osd_io.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2012, 2013, Intel Corporation. + * Copyright (c) 2012, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -42,12 +38,12 @@ * */ -/* LUSTRE_VERSION_CODE */ -#include /* prerequisite for linux/xattr.h */ #include /* prerequisite for linux/xattr.h */ #include +#include +#include /* * struct OBD_{ALLOC,FREE}*() @@ -60,6 +56,15 @@ /* ext_depth() */ #include +static inline bool osd_use_page_cache(struct osd_device *d) +{ + /* do not use pagecache if write and read caching are disabled */ + if (d->od_writethrough_cache + d->od_read_cache == 0) + return false; + /* use pagecache by default */ + return true; +} + static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf, int rw, int line, int pages) { @@ -77,12 +82,12 @@ static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf, iobuf->dr_error = 0; iobuf->dr_dev = d; iobuf->dr_frags = 0; - iobuf->dr_elapsed = 0; + iobuf->dr_elapsed = ktime_set(0, 0); /* must be counted before, so assert */ iobuf->dr_rw = rw; iobuf->dr_init_at = line; - blocks = pages * (PAGE_CACHE_SIZE >> osd_sb(d)->s_blocksize_bits); + blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits); if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) { LASSERT(iobuf->dr_pg_buf.lb_len >= pages * sizeof(iobuf->dr_pages[0])); @@ -97,7 +102,7 @@ static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf, CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n", (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages); pages = i; - blocks = pages * (PAGE_CACHE_SIZE >> osd_sb(d)->s_blocksize_bits); + blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits); iobuf->dr_max_pages = 0; CDEBUG(D_OTHER, "realloc %u for %u blocks\n", (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks); @@ -112,6 +117,12 @@ static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf, if (unlikely(iobuf->dr_pages == NULL)) return -ENOMEM; + lu_buf_realloc(&iobuf->dr_lnb_buf, + pages * sizeof(iobuf->dr_lnbs[0])); + iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf; + if (unlikely(iobuf->dr_lnbs == NULL)) + return -ENOMEM; + iobuf->dr_max_pages = pages; return 0; @@ -119,10 +130,13 @@ static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf, #define osd_init_iobuf(dev, iobuf, rw, pages) \ __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages) -static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page) +static void osd_iobuf_add_page(struct osd_iobuf *iobuf, + struct niobuf_local *lnb) { - LASSERT(iobuf->dr_npages < iobuf->dr_max_pages); - iobuf->dr_pages[iobuf->dr_npages++] = page; + LASSERT(iobuf->dr_npages < iobuf->dr_max_pages); + iobuf->dr_pages[iobuf->dr_npages] = lnb->lnb_page; + iobuf->dr_lnbs[iobuf->dr_npages] = lnb; + iobuf->dr_npages++; } void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf) @@ -136,25 +150,21 @@ void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf) lprocfs_oh_tally(&d->od_brw_stats. hist[BRW_R_DIO_FRAGS+rw], iobuf->dr_frags); - lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw], - iobuf->dr_elapsed); + lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw], + ktime_to_ms(iobuf->dr_elapsed)); } } -#ifndef REQ_WRITE /* pre-2.6.35 */ -#define __REQ_WRITE BIO_RW -#endif - +#ifdef HAVE_BIO_ENDIO_USES_ONE_ARG +static void dio_complete_routine(struct bio *bio) +{ + int error = bio->bi_status; +#else static void dio_complete_routine(struct bio *bio, int error) { +#endif struct osd_iobuf *iobuf = bio->bi_private; -#ifdef HAVE_BVEC_ITER - struct bvec_iter iter; - struct bio_vec bvl; -#else - int iter; struct bio_vec *bvl; -#endif /* CAVEAT EMPTOR: possibly in IRQ context * DO NOT record procfs stats here!!! */ @@ -163,26 +173,29 @@ static void dio_complete_routine(struct bio *bio, int error) 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 , along " + "report this to , 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"); - CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, " - "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, " - "bi_private: %p\n", bio->bi_next, bio->bi_flags, - bio->bi_rw, bio->bi_vcnt, bio_idx(bio), - bio_sectors(bio) << 9, bio->bi_end_io, - atomic_read(&bio->bi_cnt), bio->bi_private); + CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf) + ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n", + bio->bi_next, (unsigned long)bio->bi_flags, + (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio), + bio_sectors(bio) << 9, bio->bi_end_io, + atomic_read(&bio->__bi_cnt), + bio->bi_private); return; } /* the check is outside of the cycle for performance reason -bzzz */ - if (!test_bit(__REQ_WRITE, &bio->bi_rw)) { - bio_for_each_segment(bvl, bio, iter) { + if (!bio_data_dir(bio)) { + DECLARE_BVEC_ITER_ALL(iter_all); + + bio_for_each_segment_all(bvl, bio, iter_all) { if (likely(error == 0)) - SetPageUptodate(bvec_iter_page(&bvl, iter)); - LASSERT(PageLocked(bvec_iter_page(&bvl, iter))); + SetPageUptodate(bvl_to_page(bvl)); + LASSERT(PageLocked(bvl_to_page(bvl))); } atomic_dec(&iobuf->dr_dev->od_r_in_flight); } else { @@ -201,7 +214,9 @@ static void dio_complete_routine(struct bio *bio, int error) * call to OSD. */ if (atomic_read(&iobuf->dr_numreqs) == 1) { - iobuf->dr_elapsed = jiffies - iobuf->dr_start_time; + ktime_t now = ktime_get(); + + iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time); iobuf->dr_elapsed_valid = 1; } if (atomic_dec_and_test(&iobuf->dr_numreqs)) @@ -240,11 +255,13 @@ static void record_start_io(struct osd_iobuf *iobuf, int size) static void osd_submit_bio(int rw, struct bio *bio) { - LASSERTF(rw == 0 || rw == 1, "%x\n", rw); - if (rw == 0) - submit_bio(READ, bio); - else - submit_bio(WRITE, bio); + LASSERTF(rw == 0 || rw == 1, "%x\n", rw); +#ifdef HAVE_SUBMIT_BIO_2ARGS + submit_bio(rw ? WRITE : READ, bio); +#else + bio->bi_opf |= rw; + submit_bio(bio); +#endif } static int can_be_merged(struct bio *bio, sector_t sector) @@ -255,32 +272,216 @@ static int can_be_merged(struct bio *bio, sector_t sector) return bio_end_sector(bio) == sector ? 1 : 0; } +#if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) +/* + * This function will change the data written, thus it should only be + * used when checking data integrity feature + */ +static void bio_integrity_fault_inject(struct bio *bio) +{ + struct bio_vec *bvec; + DECLARE_BVEC_ITER_ALL(iter_all); + void *kaddr; + char *addr; + + bio_for_each_segment_all(bvec, bio, iter_all) { + struct page *page = bvec->bv_page; + + kaddr = kmap(page); + addr = kaddr; + *addr = ~(*addr); + kunmap(page); + break; + } +} + +static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf, + unsigned int sectors, int tuple_size) +{ + __u16 *expected_guard; + __u16 *bio_guard; + int i; + + expected_guard = expected_guard_buf; + for (i = 0; i < sectors; i++) { + bio_guard = (__u16 *)bio_prot_buf; + if (*bio_guard != *expected_guard) { + CERROR("unexpected guard tags on sector %d " + "expected guard %u, bio guard " + "%u, sectors %u, tuple size %d\n", + i, *expected_guard, *bio_guard, sectors, + tuple_size); + return -EIO; + } + expected_guard++; + bio_prot_buf += tuple_size; + } + return 0; +} + +static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev, + struct osd_iobuf *iobuf, int index) +{ + struct blk_integrity *bi = bdev_get_integrity(bdev); + struct bio_integrity_payload *bip = bio->bi_integrity; + struct niobuf_local *lnb; + unsigned short sector_size = blk_integrity_interval(bi); + void *bio_prot_buf = page_address(bip->bip_vec->bv_page) + + bip->bip_vec->bv_offset; + struct bio_vec *bv; + sector_t sector = bio_start_sector(bio); + unsigned int sectors, total; + DECLARE_BVEC_ITER_ALL(iter_all); + __u16 *expected_guard; + int rc; + + total = 0; + bio_for_each_segment_all(bv, bio, iter_all) { + lnb = iobuf->dr_lnbs[index]; + expected_guard = lnb->lnb_guards; + sectors = bv->bv_len / sector_size; + if (lnb->lnb_guard_rpc) { + rc = bio_dif_compare(expected_guard, bio_prot_buf, + sectors, bi->tuple_size); + if (rc) + return rc; + } + + sector += sectors; + bio_prot_buf += sectors * bi->tuple_size; + total += sectors * bi->tuple_size; + LASSERT(total <= bip_size(bio->bi_integrity)); + index++; + } + return 0; +} + +static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio, + struct osd_iobuf *iobuf, + int start_page_idx, bool fault_inject, + bool integrity_enabled) +{ + struct super_block *sb = osd_sb(osd); + integrity_gen_fn *generate_fn = NULL; + integrity_vrfy_fn *verify_fn = NULL; + int rc; + + ENTRY; + + if (!integrity_enabled) + RETURN(0); + + rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn); + if (rc) + RETURN(rc); + + rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn); + if (rc) + RETURN(rc); + + /* Verify and inject fault only when writing */ + if (iobuf->dr_rw == 1) { + if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) { + rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf, + start_page_idx); + if (rc) + RETURN(rc); + } + + if (unlikely(fault_inject)) + bio_integrity_fault_inject(bio); + } + + RETURN(0); +} + +#ifdef HAVE_BIO_INTEGRITY_PREP_FN +# ifdef HAVE_BIO_ENDIO_USES_ONE_ARG +static void dio_integrity_complete_routine(struct bio *bio) +# else +static void dio_integrity_complete_routine(struct bio *bio, int error) +# endif +{ + struct osd_bio_private *bio_private = bio->bi_private; + + bio->bi_private = bio_private->obp_iobuf; + osd_dio_complete_routine(bio, error); + + OBD_FREE_PTR(bio_private); +} +#endif /* HAVE_BIO_INTEGRITY_PREP_FN */ +#else /* !CONFIG_BLK_DEV_INTEGRITY */ +#define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \ + fault_inject, integrity_enabled) 0 +#endif /* CONFIG_BLK_DEV_INTEGRITY */ + +static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf, + bool integrity_enabled, int start_page_idx, + struct osd_bio_private **pprivate) +{ + ENTRY; + + *pprivate = NULL; + +#ifdef HAVE_BIO_INTEGRITY_PREP_FN + if (integrity_enabled) { + struct osd_bio_private *bio_private = NULL; + + OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO); + if (bio_private == NULL) + RETURN(-ENOMEM); + bio->bi_end_io = dio_integrity_complete_routine; + bio->bi_private = bio_private; + bio_private->obp_start_page_idx = start_page_idx; + bio_private->obp_iobuf = iobuf; + *pprivate = bio_private; + } else +#endif + { + bio->bi_end_io = dio_complete_routine; + bio->bi_private = iobuf; + } + + RETURN(0); +} + static int osd_do_bio(struct osd_device *osd, struct inode *inode, struct osd_iobuf *iobuf) { - int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits; - struct page **pages = iobuf->dr_pages; - int npages = iobuf->dr_npages; - unsigned long *blocks = iobuf->dr_blocks; - int total_blocks = npages * blocks_per_page; - int sector_bits = inode->i_sb->s_blocksize_bits - 9; - unsigned int blocksize = inode->i_sb->s_blocksize; - struct bio *bio = NULL; - struct page *page; - unsigned int page_offset; - sector_t sector; - int nblocks; - int block_idx; - int page_idx; - int i; - int rc = 0; - ENTRY; + int blocks_per_page = PAGE_SIZE >> inode->i_blkbits; + struct page **pages = iobuf->dr_pages; + int npages = iobuf->dr_npages; + sector_t *blocks = iobuf->dr_blocks; + int total_blocks = npages * blocks_per_page; + struct super_block *sb = inode->i_sb; + int sector_bits = sb->s_blocksize_bits - 9; + unsigned int blocksize = sb->s_blocksize; + struct block_device *bdev = sb->s_bdev; + struct osd_bio_private *bio_private = NULL; + struct bio *bio = NULL; + int bio_start_page_idx; + struct page *page; + unsigned int page_offset; + sector_t sector; + int nblocks; + int block_idx; + int page_idx; + int i; + int rc = 0; + bool fault_inject; + bool integrity_enabled; + struct blk_plug plug; + ENTRY; + fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT); LASSERT(iobuf->dr_npages == npages); - osd_brw_stats_update(osd, iobuf); - iobuf->dr_start_time = cfs_time_current(); + integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw); + osd_brw_stats_update(osd, iobuf); + iobuf->dr_start_time = ktime_get(); + + blk_start_plug(&plug); for (page_idx = 0, block_idx = 0; page_idx < npages; page_idx++, block_idx += blocks_per_page) { @@ -319,23 +520,30 @@ static int osd_do_bio(struct osd_device *osd, struct inode *inode, continue; /* added this frag OK */ if (bio != NULL) { - struct request_queue *q = - bdev_get_queue(bio->bi_bdev); + struct request_queue *q = bio_get_queue(bio); unsigned int bi_size = bio_sectors(bio) << 9; /* Dang! I have to fragment this I/O */ - CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) " - "sectors %d(%d) psg %d(%d) hsg %d(%d)\n", + CDEBUG(D_INODE, + "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n", bi_size, bio->bi_vcnt, bio->bi_max_vecs, bio_sectors(bio), queue_max_sectors(q), - bio_phys_segments(q, bio), - queue_max_phys_segments(q), - 0, queue_max_hw_segments(q)); + osd_bio_nr_segs(bio), + queue_max_segments(q)); + rc = osd_bio_integrity_handle(osd, bio, + iobuf, bio_start_page_idx, + fault_inject, integrity_enabled); + if (rc) { + bio_put(bio); + goto out; + } + record_start_io(iobuf, bi_size); osd_submit_bio(iobuf->dr_rw, bio); } + bio_start_page_idx = page_idx; /* allocate new bio */ bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES, (npages - page_idx) * @@ -348,11 +556,15 @@ static int osd_do_bio(struct osd_device *osd, struct inode *inode, goto out; } - bio->bi_bdev = inode->i_sb->s_bdev; + bio_set_dev(bio, bdev); bio_set_sector(bio, sector); - bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE; - bio->bi_end_io = dio_complete_routine; - bio->bi_private = iobuf; + bio->bi_opf = iobuf->dr_rw ? WRITE : READ; + rc = osd_bio_init(bio, iobuf, integrity_enabled, + bio_start_page_idx, &bio_private); + if (rc) { + bio_put(bio); + goto out; + } rc = bio_add_page(bio, page, blocksize * nblocks, page_offset); @@ -361,37 +573,59 @@ static int osd_do_bio(struct osd_device *osd, struct inode *inode, } if (bio != NULL) { + rc = osd_bio_integrity_handle(osd, bio, iobuf, + bio_start_page_idx, + fault_inject, + integrity_enabled); + if (rc) { + bio_put(bio); + goto out; + } + record_start_io(iobuf, bio_sectors(bio) << 9); osd_submit_bio(iobuf->dr_rw, bio); rc = 0; } out: + blk_finish_plug(&plug); + /* in order to achieve better IO throughput, we don't wait for writes * completion here. instead we proceed with transaction commit in * parallel and wait for IO completion once transaction is stopped * see osd_trans_stop() for more details -bzzz */ - if (iobuf->dr_rw == 0) { + if (iobuf->dr_rw == 0 || fault_inject) { wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0); osd_fini_iobuf(osd, iobuf); } - if (rc == 0) + if (rc == 0) { rc = iobuf->dr_error; + } else { + if (bio_private) + OBD_FREE_PTR(bio_private); + } + RETURN(rc); } static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages, - struct niobuf_local *lnb) + struct niobuf_local *lnb, int maxlnb) { + int rc = 0; ENTRY; *nrpages = 0; while (len > 0) { - int poff = offset & (PAGE_CACHE_SIZE - 1); - int plen = PAGE_CACHE_SIZE - poff; + int poff = offset & (PAGE_SIZE - 1); + int plen = PAGE_SIZE - poff; + + if (*nrpages >= maxlnb) { + rc = -EOVERFLOW; + break; + } if (plen > len) plen = len; @@ -402,6 +636,9 @@ static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages, lnb->lnb_flags = 0; lnb->lnb_page = NULL; lnb->lnb_rc = 0; + lnb->lnb_guard_rpc = 0; + lnb->lnb_guard_disk = 0; + lnb->lnb_locked = 0; LASSERTF(plen <= len, "plen %u, len %lld\n", plen, (long long) len); @@ -411,23 +648,63 @@ static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages, (*nrpages)++; } - RETURN(0); + RETURN(rc); } -struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw) +static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt, + loff_t offset, gfp_t gfp_mask, bool cache) { - struct inode *inode = osd_dt_obj(dt)->oo_inode; - struct osd_device *d = osd_obj2dev(osd_dt_obj(dt)); - struct page *page; + struct osd_thread_info *oti = osd_oti_get(env); + struct inode *inode = osd_dt_obj(dt)->oo_inode; + struct osd_device *d = osd_obj2dev(osd_dt_obj(dt)); + struct page *page; + int cur; LASSERT(inode); - page = find_or_create_page(inode->i_mapping, offset >> PAGE_CACHE_SHIFT, - GFP_NOFS | __GFP_HIGHMEM); - if (unlikely(page == NULL)) - lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1); + if (cache) { + page = find_or_create_page(inode->i_mapping, + offset >> PAGE_SHIFT, gfp_mask); + + if (likely(page)) { + LASSERT(!PagePrivate2(page)); + wait_on_page_writeback(page); + } else { + lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1); + } + + return page; + } - return page; + if (inode->i_mapping->nrpages) { + /* consult with pagecache, but do not create new pages */ + /* this is normally used once */ + page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT); + if (page) { + wait_on_page_writeback(page); + return page; + } + } + + LASSERT(oti->oti_dio_pages); + cur = oti->oti_dio_pages_used; + page = oti->oti_dio_pages[cur]; + + if (unlikely(!page)) { + LASSERT(cur < PTLRPC_MAX_BRW_PAGES); + page = alloc_page(gfp_mask); + if (!page) + return NULL; + oti->oti_dio_pages[cur] = page; + SetPagePrivate2(page); + lock_page(page); + } + + ClearPageUptodate(page); + page->index = offset >> PAGE_SHIFT; + oti->oti_dio_pages_used++; + + return page; } /* @@ -435,426 +712,215 @@ struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw) * journal_start * i_mutex * page lock - - * osd write path - * lock page(s) - * journal_start - * truncate_sem - + * + * osd write path: + * - lock page(s) + * - journal_start + * - truncate_sem + * * ext4 vmtruncate: - * lock pages, unlock - * journal_start - * lock partial page - * i_data_sem - -*/ -int osd_bufs_get(const struct lu_env *env, struct dt_object *d, loff_t pos, - ssize_t len, struct niobuf_local *lnb, int rw, - struct lustre_capa *capa) -{ - struct osd_object *obj = osd_dt_obj(d); - int npages, i, rc = 0; - - LASSERT(obj->oo_inode); - - osd_map_remote_to_local(pos, len, &npages, lnb); - - for (i = 0; i < npages; i++, lnb++) { - lnb->lnb_page = osd_get_page(d, lnb->lnb_file_offset, rw); - if (lnb->lnb_page == NULL) - GOTO(cleanup, rc = -ENOMEM); - - /* DLM locking protects us from write and truncate competing - * for same region, but truncate can leave dirty page in the - * cache. it's possible the writeout on a such a page is in - * progress when we access it. it's also possible that during - * this writeout we put new (partial) data, but then won't - * be able to proceed in filter_commitrw_write(). thus let's - * just wait for writeout completion, should be rare enough. - * -bzzz */ - wait_on_page_writeback(lnb->lnb_page); - BUG_ON(PageWriteback(lnb->lnb_page)); - - lu_object_get(&d->do_lu); - } - rc = i; - -cleanup: - RETURN(rc); -} + * - lock pages, unlock + * - journal_start + * - lock partial page + * - i_data_sem + * + */ +/** + * Unlock and release pages loaded by osd_bufs_get() + * + * Unlock \a npages pages from \a lnb and drop the refcount on them. + * + * \param env thread execution environment + * \param dt dt object undergoing IO (OSD object + methods) + * \param lnb array of pages undergoing IO + * \param npages number of pages in \a lnb + * + * \retval 0 always + */ static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt, struct niobuf_local *lnb, int npages) { + struct osd_thread_info *oti = osd_oti_get(env); + struct pagevec pvec; int i; - for (i = 0; i < npages; i++) { - if (lnb[i].lnb_page == NULL) - continue; - LASSERT(PageLocked(lnb[i].lnb_page)); - unlock_page(lnb[i].lnb_page); - page_cache_release(lnb[i].lnb_page); - lu_object_put(env, &dt->do_lu); - lnb[i].lnb_page = NULL; - } - - RETURN(0); -} + ll_pagevec_init(&pvec, 0); -#ifndef HAVE_LDISKFS_MAP_BLOCKS + for (i = 0; i < npages; i++) { + struct page *page = lnb[i].lnb_page; -#ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */ -#define ldiskfs_ext_pblock(ex) ext_pblock((ex)) -#endif + if (page == NULL) + continue; -struct bpointers { - unsigned long *blocks; - unsigned long start; - int num; - int init_num; - int create; -}; + /* if the page isn't cached, then reset uptodate + * to prevent reuse */ + if (PagePrivate2(page)) { + oti->oti_dio_pages_used--; + } else { + if (lnb[i].lnb_locked) + unlock_page(page); + if (pagevec_add(&pvec, page) == 0) + pagevec_release(&pvec); + } -static long ldiskfs_ext_find_goal(struct inode *inode, - struct ldiskfs_ext_path *path, - unsigned long block, int *aflags) -{ - struct ldiskfs_inode_info *ei = LDISKFS_I(inode); - unsigned long bg_start; - unsigned long colour; - int depth; - - if (path) { - struct ldiskfs_extent *ex; - depth = path->p_depth; - - /* try to predict block placement */ - if ((ex = path[depth].p_ext)) - return ldiskfs_ext_pblock(ex) + - (block - le32_to_cpu(ex->ee_block)); - - /* it looks index is empty - * try to find starting from index itself */ - if (path[depth].p_bh) - return path[depth].p_bh->b_blocknr; + lnb[i].lnb_page = NULL; } - /* OK. use inode's group */ - bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) + - le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block); - colour = (current->pid % 16) * - (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16); - return bg_start + colour + block; -} + LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used); -static unsigned long new_blocks(handle_t *handle, struct inode *inode, - struct ldiskfs_ext_path *path, - unsigned long block, unsigned long *count, - int *err) -{ - struct ldiskfs_allocation_request ar; - unsigned long pblock; - int aflags; - - /* find neighbour allocated blocks */ - ar.lleft = block; - *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft); - if (*err) - return 0; - ar.lright = block; - *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright); - if (*err) - return 0; + /* Release any partial pagevec */ + pagevec_release(&pvec); - /* allocate new block */ - ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags); - ar.inode = inode; - ar.logical = block; - ar.len = *count; - ar.flags = LDISKFS_MB_HINT_DATA; - pblock = ldiskfs_mb_new_blocks(handle, &ar, err); - *count = ar.len; - return pblock; + RETURN(0); } -static int ldiskfs_ext_new_extent_cb(struct inode *inode, - struct ldiskfs_ext_path *path, - struct ldiskfs_ext_cache *cex, -#ifdef HAVE_EXT_PREPARE_CB_EXTENT - struct ldiskfs_extent *ex, -#endif - void *cbdata) +/** + * Load and lock pages undergoing IO + * + * Pages as described in the \a lnb array are fetched (from disk or cache) + * and locked for IO by the caller. + * + * DLM locking protects us from write and truncate competing for same region, + * but partial-page truncate can leave dirty pages in the cache for ldiskfs. + * It's possible the writeout on a such a page is in progress when we access + * it. It's also possible that during this writeout we put new (partial) data + * into the page, but won't be able to proceed in filter_commitrw_write(). + * Therefore, just wait for writeout completion as it should be rare enough. + * + * \param env thread execution environment + * \param dt dt object undergoing IO (OSD object + methods) + * \param pos byte offset of IO start + * \param len number of bytes of IO + * \param lnb array of extents undergoing IO + * \param rw read or write operation, and other flags + * \param capa capabilities + * + * \retval pages (zero or more) loaded successfully + * \retval -ENOMEM on memory/page allocation error + */ +static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt, + loff_t pos, ssize_t len, struct niobuf_local *lnb, + int maxlnb, enum dt_bufs_type rw) { - struct bpointers *bp = cbdata; - struct ldiskfs_extent nex; - unsigned long pblock; - unsigned long tgen; - int err, i; - unsigned long count; - handle_t *handle; - -#ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */ - if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) { -#else - if ((cex->ec_len != 0) && (cex->ec_start != 0)) { -#endif - err = EXT_CONTINUE; - goto map; - } - - if (bp->create == 0) { - i = 0; - if (cex->ec_block < bp->start) - i = bp->start - cex->ec_block; - if (i >= cex->ec_len) - CERROR("nothing to do?! i = %d, e_num = %u\n", - i, cex->ec_len); - for (; i < cex->ec_len && bp->num; i++) { - *(bp->blocks) = 0; - bp->blocks++; - bp->num--; - bp->start++; - } - - return EXT_CONTINUE; - } - - tgen = LDISKFS_I(inode)->i_ext_generation; - count = ldiskfs_ext_calc_credits_for_insert(inode, path); - - handle = osd_journal_start(inode, LDISKFS_HT_MISC, - count + LDISKFS_ALLOC_NEEDED + 1); - if (IS_ERR(handle)) { - return PTR_ERR(handle); - } - - if (tgen != LDISKFS_I(inode)->i_ext_generation) { - /* the tree has changed. so path can be invalid at moment */ - ldiskfs_journal_stop(handle); - return EXT_REPEAT; - } + struct osd_thread_info *oti = osd_oti_get(env); + struct osd_object *obj = osd_dt_obj(dt); + struct osd_device *osd = osd_obj2dev(obj); + int npages, i, iosize, rc = 0; + bool cache, write; + loff_t fsize; + gfp_t gfp_mask; + + LASSERT(obj->oo_inode); + + rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb); + if (rc) + RETURN(rc); - /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not - * protected by i_data_sem as whole. so we patch it to store - * generation to path and now verify the tree hasn't changed */ - down_write((&LDISKFS_I(inode)->i_data_sem)); - - /* validate extent, make sure the extent tree does not changed */ - if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) { - /* cex is invalid, try again */ - up_write(&LDISKFS_I(inode)->i_data_sem); - ldiskfs_journal_stop(handle); - return EXT_REPEAT; - } + write = rw & DT_BUFS_TYPE_WRITE; - count = cex->ec_len; - pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err); - if (!pblock) - goto out; - BUG_ON(count > cex->ec_len); - - /* insert new extent */ - nex.ee_block = cpu_to_le32(cex->ec_block); - ldiskfs_ext_store_pblock(&nex, pblock); - nex.ee_len = cpu_to_le16(count); - err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0); - if (err) { - /* free data blocks we just allocated */ - /* not a good idea to call discard here directly, - * but otherwise we'd need to call it every free() */ - ldiskfs_discard_preallocations(inode); -#ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */ - ldiskfs_free_blocks(handle, inode, NULL, - ldiskfs_ext_pblock(&nex), - le16_to_cpu(nex.ee_len), 0); -#else - ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex), - le16_to_cpu(nex.ee_len), 0); -#endif - goto out; - } + fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len; + iosize = fsize - lnb[0].lnb_file_offset; + fsize = max(fsize, i_size_read(obj->oo_inode)); - /* - * Putting len of the actual extent we just inserted, - * we are asking ldiskfs_ext_walk_space() to continue - * scaning after that block - */ - cex->ec_len = le16_to_cpu(nex.ee_len); - cex->ec_start = ldiskfs_ext_pblock(&nex); - BUG_ON(le16_to_cpu(nex.ee_len) == 0); - BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block); + cache = rw & DT_BUFS_TYPE_READAHEAD; + if (cache) + goto bypass_checks; -out: - up_write((&LDISKFS_I(inode)->i_data_sem)); - ldiskfs_journal_stop(handle); -map: - if (err >= 0) { - /* map blocks */ - if (bp->num == 0) { - CERROR("hmm. why do we find this extent?\n"); - CERROR("initial space: %lu:%u\n", - bp->start, bp->init_num); -#ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */ - CERROR("current extent: %u/%u/%llu %d\n", - cex->ec_block, cex->ec_len, - (unsigned long long)cex->ec_start, - cex->ec_type); -#else - CERROR("current extent: %u/%u/%llu\n", - cex->ec_block, cex->ec_len, - (unsigned long long)cex->ec_start); -#endif - } - i = 0; - if (cex->ec_block < bp->start) - i = bp->start - cex->ec_block; - if (i >= cex->ec_len) - CERROR("nothing to do?! i = %d, e_num = %u\n", - i, cex->ec_len); - for (; i < cex->ec_len && bp->num; i++) { - *(bp->blocks) = cex->ec_start + i; -#ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */ - if (cex->ec_type != LDISKFS_EXT_CACHE_EXTENT) { -#else - if ((cex->ec_len == 0) || (cex->ec_start == 0)) { -#endif - /* unmap any possible underlying metadata from - * the block device mapping. bug 6998. */ - unmap_underlying_metadata(inode->i_sb->s_bdev, - *(bp->blocks)); + cache = osd_use_page_cache(osd); + while (cache) { + if (write) { + if (!osd->od_writethrough_cache) { + cache = false; + break; + } + if (iosize > osd->od_writethrough_max_iosize) { + cache = false; + break; + } + } else { + if (!osd->od_read_cache) { + cache = false; + break; + } + if (iosize > osd->od_readcache_max_iosize) { + cache = false; + break; } - bp->blocks++; - bp->num--; - bp->start++; } + /* don't use cache on large files */ + if (osd->od_readcache_max_filesize && + fsize > osd->od_readcache_max_filesize) + cache = false; + break; } - return err; -} - -int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long block, - unsigned long num, unsigned long *blocks, - int create) -{ - struct bpointers bp; - int err; - - CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n", - block, block + num - 1, (unsigned) inode->i_ino); - - bp.blocks = blocks; - bp.start = block; - bp.init_num = bp.num = num; - bp.create = create; - - err = ldiskfs_ext_walk_space(inode, block, num, - ldiskfs_ext_new_extent_cb, &bp); - ldiskfs_ext_invalidate_cache(inode); - return err; -} - -int osd_ldiskfs_map_bm_inode_pages(struct inode *inode, struct page **page, - int pages, unsigned long *blocks, - int create) -{ - int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits; - unsigned long *b; - int rc = 0, i; - - for (i = 0, b = blocks; i < pages; i++, page++) { - rc = ldiskfs_map_inode_page(inode, *page, b, create); - if (rc) { - CERROR("ino %lu, blk %lu create %d: rc %d\n", - inode->i_ino, *b, create, rc); - break; - } - b += blocks_per_page; +bypass_checks: + if (!cache && unlikely(!oti->oti_dio_pages)) { + OBD_ALLOC_PTR_ARRAY(oti->oti_dio_pages, PTLRPC_MAX_BRW_PAGES); + if (!oti->oti_dio_pages) + return -ENOMEM; } - return rc; -} - -int osd_ldiskfs_map_ext_inode_pages(struct inode *inode, struct page **page, - int pages, unsigned long *blocks, - int create) -{ - int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits; - int rc = 0, i = 0; - struct page *fp = NULL; - int clen = 0; - CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n", - inode->i_ino, pages, (*page)->index); - - /* pages are sorted already. so, we just have to find - * contig. space and process them properly */ - while (i < pages) { - if (fp == NULL) { - /* start new extent */ - fp = *page++; - clen = 1; - i++; - continue; - } else if (fp->index + clen == (*page)->index) { - /* continue the extent */ - page++; - clen++; - i++; - continue; - } + /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */ + gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) : + GFP_HIGHUSER; + for (i = 0; i < npages; i++, lnb++) { + lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset, + gfp_mask, cache); + if (lnb->lnb_page == NULL) + GOTO(cleanup, rc = -ENOMEM); - /* process found extent */ - rc = osd_ldiskfs_map_nblocks(inode, fp->index * blocks_per_page, - clen * blocks_per_page, blocks, - create); - if (rc) - GOTO(cleanup, rc); + lnb->lnb_locked = 1; + } - /* look for next extent */ - fp = NULL; - blocks += blocks_per_page * clen; +#if 0 + /* XXX: this version doesn't invalidate cached pages, but use them */ + if (!cache && write && obj->oo_inode->i_mapping->nrpages) { + /* do not allow data aliasing, invalidate pagecache */ + /* XXX: can be quite expensive in mixed case */ + invalidate_mapping_pages(obj->oo_inode->i_mapping, + lnb[0].lnb_file_offset >> PAGE_SHIFT, + lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT); } +#endif + + RETURN(i); - if (fp) - rc = osd_ldiskfs_map_nblocks(inode, fp->index * blocks_per_page, - clen * blocks_per_page, blocks, - create); cleanup: + if (i > 0) + osd_bufs_put(env, dt, lnb - i, i); return rc; } static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page, - int pages, unsigned long *blocks, - int create) -{ - int rc; - - if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) { - rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages, - blocks, create); - return rc; - } - rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create); - - return rc; -} -#else -static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page, - int pages, unsigned long *blocks, + int pages, sector_t *blocks, int create) { - int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits; + int blocks_per_page = PAGE_SIZE >> inode->i_blkbits; int rc = 0, i = 0; struct page *fp = NULL; int clen = 0; + pgoff_t max_page_index; + handle_t *handle = NULL; + + max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT; CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n", inode->i_ino, pages, (*page)->index); + if (create) { + create = LDISKFS_GET_BLOCKS_CREATE; + handle = ldiskfs_journal_current_handle(); + LASSERT(handle != NULL); + rc = osd_attach_jinode(inode); + if (rc) + return rc; + } /* pages are sorted already. so, we just have to find * contig. space and process them properly */ while (i < pages) { long blen, total = 0; - handle_t *handle = NULL; struct ldiskfs_map_blocks map = { 0 }; if (fp == NULL) { /* start new extent */ @@ -869,14 +935,11 @@ static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page, if (++i != pages) continue; } + if (fp->index + clen >= max_page_index) + GOTO(cleanup, rc = -EFBIG); /* process found extent */ map.m_lblk = fp->index * blocks_per_page; map.m_len = blen = clen * blocks_per_page; - if (create) { - create = LDISKFS_GET_BLOCKS_CREATE; - handle = ldiskfs_journal_current_handle(); - LASSERT(handle != NULL); - } cont_map: rc = ldiskfs_map_blocks(handle, inode, &map, create); if (rc >= 0) { @@ -893,9 +956,9 @@ cont_map: * mapping. bug 6998. */ if ((map.m_flags & LDISKFS_MAP_NEW) && create) - unmap_underlying_metadata( + clean_bdev_aliases( inode->i_sb->s_bdev, - map.m_pblk + c); + map.m_pblk + c, 1); } } rc = 0; @@ -915,7 +978,6 @@ cont_map: cleanup: return rc; } -#endif /* HAVE_LDISKFS_MAP_BLOCKS */ static int osd_write_prep(const struct lu_env *env, struct dt_object *dt, struct niobuf_local *lnb, int npages) @@ -924,14 +986,11 @@ static int osd_write_prep(const struct lu_env *env, struct dt_object *dt, struct osd_iobuf *iobuf = &oti->oti_iobuf; struct inode *inode = osd_dt_obj(dt)->oo_inode; struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt)); - struct timeval start; - struct timeval end; - unsigned long timediff; - ssize_t isize; - __s64 maxidx; - int rc = 0; - int i; - int cache = 0; + ktime_t start, end; + s64 timediff; + ssize_t isize; + __s64 maxidx; + int i, rc = 0; LASSERT(inode); @@ -940,20 +999,11 @@ static int osd_write_prep(const struct lu_env *env, struct dt_object *dt, RETURN(rc); isize = i_size_read(inode); - maxidx = ((isize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) - 1; - - if (osd->od_writethrough_cache) - cache = 1; - if (isize > osd->od_readcache_max_filesize) - cache = 0; + maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1; - do_gettimeofday(&start); + start = ktime_get(); for (i = 0; i < npages; i++) { - if (cache == 0) - generic_error_remove_page(inode->i_mapping, - lnb[i].lnb_page); - /* * till commit the content of the page is undefined * we'll set it uptodate once bulk is done. otherwise @@ -961,11 +1011,11 @@ static int osd_write_prep(const struct lu_env *env, struct dt_object *dt, */ ClearPageUptodate(lnb[i].lnb_page); - if (lnb[i].lnb_len == PAGE_CACHE_SIZE) + if (lnb[i].lnb_len == PAGE_SIZE) continue; if (maxidx >= lnb[i].lnb_page->index) { - osd_iobuf_add_page(iobuf, lnb[i].lnb_page); + osd_iobuf_add_page(iobuf, &lnb[i]); } else { long off; char *p = kmap(lnb[i].lnb_page); @@ -974,14 +1024,14 @@ static int osd_write_prep(const struct lu_env *env, struct dt_object *dt, if (off) memset(p, 0, off); off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) & - ~CFS_PAGE_MASK; + ~PAGE_MASK; if (off) - memset(p + off, 0, PAGE_CACHE_SIZE - off); + memset(p + off, 0, PAGE_SIZE - off); kunmap(lnb[i].lnb_page); } } - do_gettimeofday(&end); - timediff = cfs_timeval_sub(&end, &start, NULL); + end = ktime_get(); + timediff = ktime_us_delta(end, start); lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff); if (iobuf->dr_npages) { @@ -997,29 +1047,57 @@ static int osd_write_prep(const struct lu_env *env, struct dt_object *dt, RETURN(rc); } -/* Check if a block is allocated or not */ -static int osd_is_mapped(struct inode *inode, obd_size offset) -{ - sector_t (*fs_bmap)(struct address_space *, sector_t); +struct osd_fextent { + sector_t start; + sector_t end; + unsigned int mapped:1; +}; - fs_bmap = inode->i_mapping->a_ops->bmap; +static int osd_is_mapped(struct dt_object *dt, __u64 offset, + struct osd_fextent *cached_extent) +{ + struct inode *inode = osd_dt_obj(dt)->oo_inode; + sector_t block = offset >> inode->i_blkbits; + sector_t start; + struct fiemap_extent_info fei = { 0 }; + struct fiemap_extent fe = { 0 }; + mm_segment_t saved_fs; + int rc; - /* We can't know if we are overwriting or not */ - if (unlikely(fs_bmap == NULL)) - return 0; + if (block >= cached_extent->start && block < cached_extent->end) + return cached_extent->mapped; if (i_size_read(inode) == 0) return 0; /* Beyond EOF, must not be mapped */ - if (((i_size_read(inode) - 1) >> inode->i_blkbits) < - (offset >> inode->i_blkbits)) + if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block) return 0; - if (fs_bmap(inode->i_mapping, offset >> inode->i_blkbits) == 0) + fei.fi_extents_max = 1; + fei.fi_extents_start = &fe; + + saved_fs = get_fs(); + set_fs(KERNEL_DS); + rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset); + set_fs(saved_fs); + if (rc != 0) return 0; - return 1; + start = fe.fe_logical >> inode->i_blkbits; + + if (start > block) { + cached_extent->start = block; + cached_extent->end = start; + cached_extent->mapped = 0; + } else { + cached_extent->start = start; + cached_extent->end = (fe.fe_logical + fe.fe_length) >> + inode->i_blkbits; + cached_extent->mapped = 1; + } + + return cached_extent->mapped; } static int osd_declare_write_commit(const struct lu_env *env, @@ -1027,33 +1105,37 @@ static int osd_declare_write_commit(const struct lu_env *env, struct niobuf_local *lnb, int npages, struct thandle *handle) { - const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt)); - struct inode *inode = osd_dt_obj(dt)->oo_inode; - struct osd_thandle *oh; - int extents = 1; - int depth; - int i; - int newblocks; - int rc = 0; - int flags = 0; - bool ignore_quota = false; - long long quota_space = 0; + const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt)); + struct inode *inode = osd_dt_obj(dt)->oo_inode; + struct osd_thandle *oh; + int extents = 1; + int depth; + int i; + int newblocks; + int rc = 0; + int credits = 0; + long long quota_space = 0; + struct osd_fextent extent = { 0 }; + enum osd_quota_local_flags local_flags = 0; + enum osd_qid_declare_flags declare_flags = OSD_QID_BLK; ENTRY; - LASSERT(handle != NULL); - oh = container_of0(handle, struct osd_thandle, ot_super); - LASSERT(oh->ot_handle == NULL); + LASSERT(handle != NULL); + oh = container_of(handle, struct osd_thandle, ot_super); + LASSERT(oh->ot_handle == NULL); - newblocks = npages; + newblocks = npages; - /* calculate number of extents (probably better to pass nb) */ + /* calculate number of extents (probably better to pass nb) */ for (i = 0; i < npages; i++) { if (i && lnb[i].lnb_file_offset != lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len) extents++; - if (!osd_is_mapped(inode, lnb[i].lnb_file_offset)) - quota_space += PAGE_CACHE_SIZE; + if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent)) + lnb[i].lnb_flags |= OBD_BRW_MAPPED; + else + quota_space += PAGE_SIZE; /* ignore quota for the whole request if any page is from * client cache or written by root. @@ -1067,7 +1149,7 @@ static int osd_declare_write_commit(const struct lu_env *env, if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) || (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) == OBD_BRW_FROM_GRANT) - ignore_quota = true; + declare_flags |= OSD_QID_FORCE; } /* @@ -1083,14 +1165,14 @@ static int osd_declare_write_commit(const struct lu_env *env, depth = ext_depth(inode); depth = max(depth, 1) + 1; newblocks += depth; - oh->ot_credits++; /* inode */ - oh->ot_credits += depth * 2 * extents; - } else { - depth = 3; - newblocks += depth; - oh->ot_credits++; /* inode */ - oh->ot_credits += depth * extents; - } + credits++; /* inode */ + credits += depth * 2 * extents; + } else { + depth = 3; + newblocks += depth; + credits++; /* inode */ + credits += depth * extents; + } /* quota space for metadata blocks */ quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd)); @@ -1102,58 +1184,70 @@ static int osd_declare_write_commit(const struct lu_env *env, /* we can't dirty more bitmap blocks than exist */ if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count) - oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_groups_count; + credits += LDISKFS_SB(osd_sb(osd))->s_groups_count; else - oh->ot_credits += newblocks; + credits += newblocks; - /* we can't dirty more gd blocks than exist */ - if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count) - oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count; - else - oh->ot_credits += newblocks; + /* we can't dirty more gd blocks than exist */ + if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count) + credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count; + else + credits += newblocks; + + osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits); /* make sure the over quota flags were not set */ - lnb[0].lnb_flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA); + lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA; rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode), - quota_space, oh, osd_dt_obj(dt), true, - &flags, ignore_quota); + i_projid_read(inode), quota_space, oh, + osd_dt_obj(dt), &local_flags, declare_flags); /* we need only to store the overquota flags in the first lnb for * now, once we support multiple objects BRW, this code needs be * revised. */ - if (flags & QUOTA_FL_OVER_USRQUOTA) + if (local_flags & QUOTA_FL_OVER_USRQUOTA) lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA; - if (flags & QUOTA_FL_OVER_GRPQUOTA) + if (local_flags & QUOTA_FL_OVER_GRPQUOTA) lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA; + if (local_flags & QUOTA_FL_OVER_PRJQUOTA) + lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA; + + if (rc == 0) + rc = osd_trunc_lock(osd_dt_obj(dt), oh, true); RETURN(rc); } /* Check if a block is allocated or not */ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, - struct niobuf_local *lnb, int npages, - struct thandle *thandle) + struct niobuf_local *lnb, int npages, + struct thandle *thandle, __u64 user_size) { - struct osd_thread_info *oti = osd_oti_get(env); - struct osd_iobuf *iobuf = &oti->oti_iobuf; - struct inode *inode = osd_dt_obj(dt)->oo_inode; - struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt)); - loff_t isize; - int rc = 0, i; + struct osd_thread_info *oti = osd_oti_get(env); + struct osd_iobuf *iobuf = &oti->oti_iobuf; + struct inode *inode = osd_dt_obj(dt)->oo_inode; + struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt)); + loff_t disk_size; + int rc = 0, i; - LASSERT(inode); + LASSERT(inode); rc = osd_init_iobuf(osd, iobuf, 1, npages); if (unlikely(rc != 0)) RETURN(rc); - isize = i_size_read(inode); - ll_vfs_dq_init(inode); + disk_size = i_size_read(inode); + /* if disk_size is already bigger than specified user_size, + * ignore user_size + */ + if (disk_size > user_size) + user_size = 0; + dquot_initialize(inode); - for (i = 0; i < npages; i++) { + for (i = 0; i < npages; i++) { if (lnb[i].lnb_rc == -ENOSPC && - osd_is_mapped(inode, lnb[i].lnb_file_offset)) { + (lnb[i].lnb_flags & OBD_BRW_MAPPED)) { /* Allow the write to proceed if overwriting an * existing block */ lnb[i].lnb_rc = 0; @@ -1171,8 +1265,8 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, LASSERT(PageLocked(lnb[i].lnb_page)); LASSERT(!PageWriteback(lnb[i].lnb_page)); - if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize) - isize = lnb[i].lnb_file_offset + lnb[i].lnb_len; + if (lnb[i].lnb_file_offset + lnb[i].lnb_len > disk_size) + disk_size = lnb[i].lnb_file_offset + lnb[i].lnb_len; /* * Since write and truncate are serialized by oo_sem, even @@ -1183,42 +1277,55 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, SetPageUptodate(lnb[i].lnb_page); - osd_iobuf_add_page(iobuf, lnb[i].lnb_page); - } + osd_iobuf_add_page(iobuf, &lnb[i]); + } + /* if file has grown, take user_size into account */ + if (user_size && disk_size > user_size) + disk_size = user_size; + + osd_trans_exec_op(env, thandle, OSD_OT_WRITE); - if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) { - rc = -ENOSPC; - } else if (iobuf->dr_npages > 0) { + if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) { + rc = -ENOSPC; + } else if (iobuf->dr_npages > 0) { rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages, iobuf->dr_npages, iobuf->dr_blocks, 1); - } else { - /* no pages to write, no transno is needed */ - thandle->th_local = 1; - } + } else { + /* no pages to write, no transno is needed */ + thandle->th_local = 1; + } - if (likely(rc == 0)) { - if (isize > i_size_read(inode)) { - i_size_write(inode, isize); - LDISKFS_I(inode)->i_disksize = isize; - ll_dirty_inode(inode, I_DIRTY_DATASYNC); - } + if (likely(rc == 0)) { + spin_lock(&inode->i_lock); + if (disk_size > i_size_read(inode)) { + i_size_write(inode, disk_size); + LDISKFS_I(inode)->i_disksize = disk_size; + spin_unlock(&inode->i_lock); + osd_dirty_inode(inode, I_DIRTY_DATASYNC); + } else { + spin_unlock(&inode->i_lock); + } - rc = osd_do_bio(osd, inode, iobuf); - /* we don't do stats here as in read path because - * write is async: we'll do this in osd_put_bufs() */ + rc = osd_do_bio(osd, inode, iobuf); + /* we don't do stats here as in read path because + * write is async: we'll do this in osd_put_bufs() */ } else { osd_fini_iobuf(osd, iobuf); } - if (unlikely(rc != 0)) { - /* if write fails, we should drop pages from the cache */ - for (i = 0; i < npages; i++) { + osd_trans_exec_check(env, thandle, OSD_OT_WRITE); + + if (unlikely(rc != 0)) { + /* if write fails, we should drop pages from the cache */ + for (i = 0; i < npages; i++) { if (lnb[i].lnb_page == NULL) continue; - LASSERT(PageLocked(lnb[i].lnb_page)); - generic_error_remove_page(inode->i_mapping, - lnb[i].lnb_page); + if (!PagePrivate2(lnb[i].lnb_page)) { + LASSERT(PageLocked(lnb[i].lnb_page)); + generic_error_remove_page(inode->i_mapping, + lnb[i].lnb_page); + } } } @@ -1232,9 +1339,10 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, struct osd_iobuf *iobuf = &oti->oti_iobuf; struct inode *inode = osd_dt_obj(dt)->oo_inode; struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt)); - struct timeval start, end; - unsigned long timediff; - int rc = 0, i, m = 0, cache = 0, cache_hits = 0, cache_misses = 0; + int rc = 0, i, cache_hits = 0, cache_misses = 0; + ktime_t start, end; + s64 timediff; + loff_t isize; LASSERT(inode); @@ -1242,40 +1350,41 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, if (unlikely(rc != 0)) RETURN(rc); - if (osd->od_read_cache) - cache = 1; - if (i_size_read(inode) > osd->od_readcache_max_filesize) - cache = 0; + isize = i_size_read(inode); - do_gettimeofday(&start); + start = ktime_get(); for (i = 0; i < npages; i++) { - if (i_size_read(inode) <= lnb[i].lnb_file_offset) + if (isize <= lnb[i].lnb_file_offset) /* If there's no more data, abort early. * lnb->lnb_rc == 0, so it's easy to detect later. */ break; - if (i_size_read(inode) < - lnb[i].lnb_file_offset + lnb[i].lnb_len - 1) - lnb[i].lnb_rc = i_size_read(inode) - - lnb[i].lnb_file_offset; - else - lnb[i].lnb_rc = lnb[i].lnb_len; - m += lnb[i].lnb_len; + /* instead of looking if we go beyong isize, send complete + * pages all the time + */ + lnb[i].lnb_rc = lnb[i].lnb_len; + + /* Bypass disk read if fail_loc is set properly */ + if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW)) + SetPageUptodate(lnb[i].lnb_page); if (PageUptodate(lnb[i].lnb_page)) { cache_hits++; + unlock_page(lnb[i].lnb_page); } else { cache_misses++; - osd_iobuf_add_page(iobuf, lnb[i].lnb_page); + osd_iobuf_add_page(iobuf, &lnb[i]); } - - if (cache == 0) - generic_error_remove_page(inode->i_mapping, - lnb[i].lnb_page); + /* no need to unlock in osd_bufs_put(), the sooner page is + * unlocked, the earlier another client can access it. + * notice real unlock_page() can be called few lines + * below after osd_do_bio(). lnb is a per-thread, so it's + * fine to have PG_locked and lnb_locked inconsistent here */ + lnb[i].lnb_locked = 0; } - do_gettimeofday(&end); - timediff = cfs_timeval_sub(&end, &start, NULL); + end = ktime_get(); + timediff = ktime_us_delta(end, start); lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff); if (cache_hits != 0) @@ -1288,16 +1397,23 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS, cache_hits + cache_misses); - if (iobuf->dr_npages) { + if (iobuf->dr_npages) { rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages, iobuf->dr_npages, iobuf->dr_blocks, 0); - rc = osd_do_bio(osd, inode, iobuf); + rc = osd_do_bio(osd, inode, iobuf); - /* IO stats will be done in osd_bufs_put() */ - } + /* IO stats will be done in osd_bufs_put() */ - RETURN(rc); + /* early release to let others read data during the bulk */ + for (i = 0; i < iobuf->dr_npages; i++) { + LASSERT(PageLocked(iobuf->dr_pages[i])); + if (!PagePrivate2(iobuf->dr_pages[i])) + unlock_page(iobuf->dr_pages[i]); + } + } + + RETURN(rc); } /* @@ -1328,7 +1444,6 @@ int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs) int blocksize; int csize; int boffs; - int err; /* prevent reading after eof */ spin_lock(&inode->i_lock); @@ -1336,7 +1451,8 @@ int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs) loff_t diff = i_size_read(inode) - *offs; spin_unlock(&inode->i_lock); if (diff < 0) { - CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n", + CDEBUG(D_OTHER, + "size %llu is too short to read @%llu\n", i_size_read(inode), *offs); return -EBADR; } else if (diff == 0) { @@ -1354,16 +1470,21 @@ int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs) block = *offs >> inode->i_blkbits; boffs = *offs & (blocksize - 1); csize = min(blocksize - boffs, size); - bh = ldiskfs_bread(NULL, inode, block, 0, &err); - if (!bh) { - CERROR("%s: can't read %u@%llu on ino %lu: rc = %d\n", - LDISKFS_SB(inode->i_sb)->s_es->s_volume_name, - csize, *offs, inode->i_ino, err); - return err; - } + bh = __ldiskfs_bread(NULL, inode, block, 0); + if (IS_ERR(bh)) { + CERROR("%s: can't read %u@%llu on ino %lu: " + "rc = %ld\n", osd_ino2name(inode), + csize, *offs, inode->i_ino, + PTR_ERR(bh)); + return PTR_ERR(bh); + } - memcpy(buf, bh->b_data + boffs, csize); - brelse(bh); + if (bh != NULL) { + memcpy(buf, bh->b_data + boffs, csize); + brelse(bh); + } else { + memset(buf, 0, csize); + } *offs += csize; buf += csize; @@ -1373,25 +1494,29 @@ int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs) } static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt, - struct lu_buf *buf, loff_t *pos, - struct lustre_capa *capa) + struct lu_buf *buf, loff_t *pos) { - struct inode *inode = osd_dt_obj(dt)->oo_inode; - int rc; + struct inode *inode = osd_dt_obj(dt)->oo_inode; + int rc; - if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_READ)) - RETURN(-EACCES); + /* Read small symlink from inode body as we need to maintain correct + * on-disk symlinks for ldiskfs. + */ + if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) { + loff_t size = i_size_read(inode); - /* Read small symlink from inode body as we need to maintain correct - * on-disk symlinks for ldiskfs. - */ - if (S_ISLNK(dt->do_lu.lo_header->loh_attr) && - (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data))) - rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len); - else - rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos); + if (buf->lb_len < size) + return -EOVERFLOW; + + if (size < sizeof(LDISKFS_I(inode)->i_data)) + rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size); + else + rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos); + } else { + rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos); + } - return rc; + return rc; } static inline int osd_extents_enabled(struct super_block *sb, @@ -1400,17 +1525,15 @@ static inline int osd_extents_enabled(struct super_block *sb, if (inode != NULL) { if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) return 1; - } else if (test_opt(sb, EXTENTS)) { + } else if (ldiskfs_has_feature_extents(sb)) { return 1; } return 0; } -static inline int osd_calc_bkmap_credits(struct super_block *sb, - struct inode *inode, - const loff_t size, - const loff_t pos, - const int blocks) +int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode, + const loff_t size, const loff_t pos, + const int blocks) { int credits, bits, bs, i; @@ -1445,9 +1568,12 @@ static inline int osd_calc_bkmap_credits(struct super_block *sb, } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) { /* single indirect */ credits = blocks * 3; - /* probably indirect block has been allocated already */ - if (!inode || LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK]) + if (inode == NULL || + LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0) credits += 3; + else + /* The indirect block may be modified. */ + credits += 1; } return credits; @@ -1468,10 +1594,10 @@ static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt, ENTRY; LASSERT(buf != NULL); - LASSERT(handle != NULL); + LASSERT(handle != NULL); - oh = container_of0(handle, struct osd_thandle, ot_super); - LASSERT(oh->ot_handle == NULL); + oh = container_of(handle, struct osd_thandle, ot_super); + LASSERT(oh->ot_handle == NULL); size = buf->lb_len; bits = sb->s_blocksize_bits; @@ -1515,7 +1641,7 @@ static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt, credits = depth; /* if not append, then split may need to modify * existing blocks moving entries into the new ones */ - if (_pos == -1) + if (_pos != -1) credits += depth; /* blocks to store data: bitmap,gd,itself */ credits += blocks * 3; @@ -1536,8 +1662,13 @@ out: * objects, so always set the lqi_space as 0. */ if (inode != NULL) rc = osd_declare_inode_qid(env, i_uid_read(inode), - i_gid_read(inode), 0, oh, obj, true, - NULL, false); + i_gid_read(inode), + i_projid_read(inode), 0, + oh, obj, NULL, OSD_QID_BLK); + + if (rc == 0) + rc = osd_trunc_lock(obj, oh, true); + RETURN(rc); } @@ -1546,26 +1677,32 @@ static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen) /* LU-2634: clear the extent format for fast symlink */ ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS); - memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen); - LDISKFS_I(inode)->i_disksize = buflen; - i_size_write(inode, buflen); - ll_dirty_inode(inode, I_DIRTY_DATASYNC); + memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen); + spin_lock(&inode->i_lock); + LDISKFS_I(inode)->i_disksize = buflen; + i_size_write(inode, buflen); + spin_unlock(&inode->i_lock); + osd_dirty_inode(inode, I_DIRTY_DATASYNC); - return 0; + return 0; } -int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize, - int write_NUL, loff_t *offs, handle_t *handle) +static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf, + int bufsize, int write_NUL, loff_t *offs, + handle_t *handle) { + struct inode *inode = osd_dt_obj(dt)->oo_inode; struct buffer_head *bh = NULL; loff_t offset = *offs; loff_t new_size = i_size_read(inode); unsigned long block; int blocksize = 1 << inode->i_blkbits; + struct ldiskfs_inode_info *ei = LDISKFS_I(inode); int err = 0; int size; int boffs; int dirty_inode = 0; + bool create, sparse, sync = false; if (write_NUL) { /* @@ -1576,18 +1713,69 @@ int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize, ((char *)buf)[bufsize] = '\0'; ++bufsize; } - while (bufsize > 0) { - if (bh != NULL) - brelse(bh); - - block = offset >> inode->i_blkbits; - boffs = offset & (blocksize - 1); - size = min(blocksize - boffs, bufsize); - bh = ldiskfs_bread(handle, inode, block, 1, &err); - if (!bh) { - CERROR("%s: error reading offset %llu (block %lu): " - "rc = %d\n", - inode->i_sb->s_id, offset, block, err); + + dirty_inode = test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA, &ei->i_flags); + + /* sparse checking is racy, but sparse is very rare case, leave as is */ + sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) < + ((new_size - 1) >> inode->i_blkbits) + 1); + + while (bufsize > 0) { + int credits = handle->h_buffer_credits; + unsigned long last_block = (new_size == 0) ? 0 : + (new_size - 1) >> inode->i_blkbits; + + if (bh) + brelse(bh); + + block = offset >> inode->i_blkbits; + boffs = offset & (blocksize - 1); + size = min(blocksize - boffs, bufsize); + sync = (block > last_block || new_size == 0 || sparse); + + if (sync) + down(&ei->i_append_sem); + + bh = __ldiskfs_bread(handle, inode, block, 0); + + if (unlikely(IS_ERR_OR_NULL(bh) && !sync)) + CWARN("%s: adding bh without locking off %llu (block %lu, " + "size %d, offs %llu)\n", inode->i_sb->s_id, + offset, block, bufsize, *offs); + + if (IS_ERR_OR_NULL(bh)) { + struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt)); + int flags = LDISKFS_GET_BLOCKS_CREATE; + + /* while the file system is being mounted, avoid + * preallocation otherwise mount can take a long + * time as mballoc cache is cold. + * XXX: this is a workaround until we have a proper + * fix in mballoc + * XXX: works with extent-based files only */ + if (!osd->od_cl_seq) + flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE; + bh = __ldiskfs_bread(handle, inode, block, flags); + create = true; + } else { + if (sync) { + up(&ei->i_append_sem); + sync = false; + } + create = false; + } + if (IS_ERR_OR_NULL(bh)) { + if (bh == NULL) { + err = -EIO; + } else { + err = PTR_ERR(bh); + bh = NULL; + } + + CERROR("%s: error reading offset %llu (block %lu, " + "size %d, offs %llu), credits %d/%d: rc = %d\n", + inode->i_sb->s_id, offset, block, bufsize, *offs, + credits, handle->h_buffer_credits, err); break; } @@ -1600,8 +1788,15 @@ int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize, LASSERTF(boffs + size <= bh->b_size, "boffs %d size %d bh->b_size %lu\n", boffs, size, (unsigned long)bh->b_size); - memcpy(bh->b_data + boffs, buf, size); - err = ldiskfs_journal_dirty_metadata(handle, bh); + if (create) { + memset(bh->b_data, 0, bh->b_size); + if (sync) { + up(&ei->i_append_sem); + sync = false; + } + } + memcpy(bh->b_data + boffs, buf, size); + err = ldiskfs_handle_dirty_metadata(handle, NULL, bh); if (err) break; @@ -1611,24 +1806,27 @@ int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize, bufsize -= size; buf += size; } - if (bh) - brelse(bh); + if (sync) + up(&ei->i_append_sem); + + if (bh) + brelse(bh); if (write_NUL) --new_size; - /* correct in-core and on-disk sizes */ - if (new_size > i_size_read(inode)) { + /* correct in-core and on-disk sizes */ + if (new_size > i_size_read(inode)) { spin_lock(&inode->i_lock); if (new_size > i_size_read(inode)) i_size_write(inode, new_size); - if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) { - LDISKFS_I(inode)->i_disksize = i_size_read(inode); + if (i_size_read(inode) > ei->i_disksize) { + ei->i_disksize = i_size_read(inode); dirty_inode = 1; } spin_unlock(&inode->i_lock); - if (dirty_inode) - ll_dirty_inode(inode, I_DIRTY_DATASYNC); } + if (dirty_inode) + osd_dirty_inode(inode, I_DIRTY_DATASYNC); if (err == 0) *offs = offset; @@ -1636,9 +1834,8 @@ int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize, } static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt, - const struct lu_buf *buf, loff_t *pos, - struct thandle *handle, struct lustre_capa *capa, - int ignore_quota) + const struct lu_buf *buf, loff_t *pos, + struct thandle *handle) { struct inode *inode = osd_dt_obj(dt)->oo_inode; struct osd_thandle *oh; @@ -1647,18 +1844,17 @@ static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt, LASSERT(dt_object_exists(dt)); - if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_WRITE)) - return -EACCES; - LASSERT(handle != NULL); LASSERT(inode != NULL); - ll_vfs_dq_init(inode); + dquot_initialize(inode); /* XXX: don't check: one declared chunk can be used many times */ /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */ oh = container_of(handle, struct osd_thandle, ot_super); LASSERT(oh->ot_handle->h_transaction != NULL); + osd_trans_exec_op(env, handle, OSD_OT_WRITE); + /* Write small symlink to inode body as we need to maintain correct * on-disk symlinks for ldiskfs. * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len @@ -1668,12 +1864,77 @@ static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt, if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data))) result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len); else - result = osd_ldiskfs_write_record(inode, buf->lb_buf, - buf->lb_len, is_link, pos, - oh->ot_handle); - if (result == 0) - result = buf->lb_len; - return result; + result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len, + is_link, pos, oh->ot_handle); + if (result == 0) + result = buf->lb_len; + + osd_trans_exec_check(env, handle, OSD_OT_WRITE); + + return result; +} + +static int osd_declare_fallocate(const struct lu_env *env, + struct dt_object *dt, struct thandle *th) +{ + struct osd_thandle *oh; + struct inode *inode; + int rc; + ENTRY; + + LASSERT(th); + oh = container_of(th, struct osd_thandle, ot_super); + + osd_trans_declare_op(env, oh, OSD_OT_PREALLOC, + osd_dto_credits_noquota[DTO_WRITE_BLOCK]); + inode = osd_dt_obj(dt)->oo_inode; + LASSERT(inode); + + rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode), + i_projid_read(inode), 0, oh, osd_dt_obj(dt), + NULL, OSD_QID_BLK); + RETURN(rc); +} + +static int osd_fallocate(const struct lu_env *env, struct dt_object *dt, + __u64 start, __u64 end, int mode, struct thandle *th) +{ + struct osd_object *obj = osd_dt_obj(dt); + struct inode *inode = obj->oo_inode; + int rc = 0; + struct osd_thread_info *info = osd_oti_get(env); + struct dentry *dentry = &info->oti_obj_dentry; + struct file *file = &info->oti_file; + + ENTRY; + /* + * Only mode == 0 (which is standard prealloc) is supported now. + * Rest of mode options is not supported yet. + */ + if (mode & ~FALLOC_FL_KEEP_SIZE) + RETURN(-EOPNOTSUPP); + + LASSERT(dt_object_exists(dt)); + LASSERT(osd_invariant(obj)); + LASSERT(inode != NULL); + dquot_initialize(inode); + + LASSERT(th); + + osd_trans_exec_op(env, th, OSD_OT_PREALLOC); + + /* + * Because f_op->fallocate() does not have an inode arg + */ + dentry->d_inode = inode; + dentry->d_sb = inode->i_sb; + file->f_path.dentry = dentry; + file->f_mapping = inode->i_mapping; + file->f_op = inode->i_fop; + file->f_inode = inode; + rc = file->f_op->fallocate(file, mode, start, end - start); + + RETURN(rc); } static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt, @@ -1702,99 +1963,177 @@ static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt, LASSERT(inode); rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode), - 0, oh, osd_dt_obj(dt), true, NULL, false); + i_projid_read(inode), 0, oh, osd_dt_obj(dt), + NULL, OSD_QID_BLK); + + if (rc == 0) + rc = osd_trunc_lock(osd_dt_obj(dt), oh, false); + RETURN(rc); } static int osd_punch(const struct lu_env *env, struct dt_object *dt, - __u64 start, __u64 end, struct thandle *th, - struct lustre_capa *capa) + __u64 start, __u64 end, struct thandle *th) { + struct osd_object *obj = osd_dt_obj(dt); + struct osd_device *osd = osd_obj2dev(obj); + struct inode *inode = obj->oo_inode; + struct osd_access_lock *al; struct osd_thandle *oh; - struct osd_object *obj = osd_dt_obj(dt); - struct inode *inode = obj->oo_inode; - handle_t *h; - tid_t tid; - int rc = 0, rc2 = 0; + int rc = 0, found = 0; + bool grow = false; ENTRY; - LASSERT(end == OBD_OBJECT_EOF); LASSERT(dt_object_exists(dt)); LASSERT(osd_invariant(obj)); LASSERT(inode != NULL); - ll_vfs_dq_init(inode); + dquot_initialize(inode); LASSERT(th); oh = container_of(th, struct osd_thandle, ot_super); LASSERT(oh->ot_handle->h_transaction != NULL); - osd_trans_exec_op(env, th, OSD_OT_PUNCH); + /* we used to skip truncate to current size to + * optimize truncates on OST. with DoM we can + * get attr_set to set specific size (MDS_REINT) + * and then get truncate RPC which essentially + * would be skipped. this is bad.. so, disable + * this optimization on MDS till the client stop + * to sent MDS_REINT (LU-11033) -bzzz */ + if (osd->od_is_ost && i_size_read(inode) == start) + RETURN(0); - tid = oh->ot_handle->h_transaction->t_tid; + osd_trans_exec_op(env, th, OSD_OT_PUNCH); + spin_lock(&inode->i_lock); + if (i_size_read(inode) < start) + grow = true; i_size_write(inode, start); + spin_unlock(&inode->i_lock); ll_truncate_pagecache(inode, start); -#ifdef HAVE_INODEOPS_TRUNCATE - if (inode->i_op->truncate) { - inode->i_op->truncate(inode); - } else -#endif - ldiskfs_truncate(inode); + + /* optimize grow case */ + if (grow) { + osd_execute_truncate(obj); + GOTO(out, rc); + } + + inode_lock(inode); + /* add to orphan list to ensure truncate completion + * if this transaction succeed. ldiskfs_truncate() + * will take the inode out of the list */ + rc = ldiskfs_orphan_add(oh->ot_handle, inode); + inode_unlock(inode); + if (rc != 0) + GOTO(out, rc); + + list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) { + if (obj != al->tl_obj) + continue; + LASSERT(al->tl_shared == 0); + found = 1; + /* do actual truncate in osd_trans_stop() */ + al->tl_truncate = 1; + break; + } + LASSERT(found); + +out: + RETURN(rc); +} + +static int fiemap_check_ranges(struct inode *inode, + u64 start, u64 len, u64 *new_len) +{ + loff_t maxbytes; + + *new_len = len; + + if (len == 0) + return -EINVAL; + + if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS)) + maxbytes = inode->i_sb->s_maxbytes; + else + maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes; + + if (start > maxbytes) + return -EFBIG; /* - * For a partial-page truncate, flush the page to disk immediately to - * avoid data corruption during direct disk write. b=17397 + * Shrink request scope to what the fs can actually handle. */ - if ((start & ~CFS_PAGE_MASK) != 0) - rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1); + if (len > maxbytes || (maxbytes - len) < start) + *new_len = maxbytes - start; - h = journal_current_handle(); - LASSERT(h != NULL); - LASSERT(h == oh->ot_handle); + return 0; +} - if (tid != h->h_transaction->t_tid) { - int credits = oh->ot_credits; - /* - * transaction has changed during truncate - * we need to restart the handle with our credits - */ - if (h->h_buffer_credits < credits) { - if (ldiskfs_journal_extend(h, credits)) - rc2 = ldiskfs_journal_restart(h, credits); - } - } +/* So that the fiemap access checks can't overflow on 32 bit machines. */ +#define FIEMAP_MAX_EXTENTS (UINT_MAX / sizeof(struct fiemap_extent)) - RETURN(rc == 0 ? rc2 : rc); +static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt, + struct fiemap *fm) +{ + struct fiemap_extent_info fieinfo = {0, }; + struct inode *inode = osd_dt_obj(dt)->oo_inode; + u64 len; + int rc; + mm_segment_t cur_fs; + + LASSERT(inode); + if (inode->i_op->fiemap == NULL) + return -EOPNOTSUPP; + + if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS) + return -EINVAL; + + rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len); + if (rc) + return rc; + + fieinfo.fi_flags = fm->fm_flags; + fieinfo.fi_extents_max = fm->fm_extent_count; + fieinfo.fi_extents_start = fm->fm_extents; + + if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC) + filemap_write_and_wait(inode->i_mapping); + + /* Save previous value address limit */ + cur_fs = get_fs(); + /* Set the address limit of the kernel */ + set_fs(KERNEL_DS); + + rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len); + fm->fm_flags = fieinfo.fi_flags; + fm->fm_mapped_extents = fieinfo.fi_extents_mapped; + + /* Restore the previous address limt */ + set_fs(cur_fs); + + return rc; } -static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt, - struct ll_user_fiemap *fm) +static int osd_ladvise(const struct lu_env *env, struct dt_object *dt, + __u64 start, __u64 end, enum lu_ladvise_type advice) { - struct inode *inode = osd_dt_obj(dt)->oo_inode; - struct osd_thread_info *info = osd_oti_get(env); - struct dentry *dentry = &info->oti_obj_dentry; - struct file *file = &info->oti_file; - mm_segment_t saved_fs; - int rc; + struct osd_object *obj = osd_dt_obj(dt); + int rc = 0; + ENTRY; - LASSERT(inode); - dentry->d_inode = inode; - dentry->d_sb = inode->i_sb; - file->f_dentry = dentry; - file->f_mapping = inode->i_mapping; - file->f_op = inode->i_fop; - set_file_inode(file, inode); - - saved_fs = get_fs(); - set_fs(get_ds()); - /* ldiskfs_ioctl does not have a inode argument */ - if (inode->i_fop->unlocked_ioctl) - rc = inode->i_fop->unlocked_ioctl(file, FSFILT_IOC_FIEMAP, - (long)fm); - else - rc = -ENOTTY; - set_fs(saved_fs); - return rc; + switch (advice) { + case LU_LADVISE_DONTNEED: + if (end) + invalidate_mapping_pages(obj->oo_inode->i_mapping, + start >> PAGE_SHIFT, + (end - 1) >> PAGE_SHIFT); + break; + default: + rc = -ENOTSUPP; + break; + } + + RETURN(rc); } /* @@ -1806,17 +2145,135 @@ const struct dt_body_operations osd_body_ops_new = { }; const struct dt_body_operations osd_body_ops = { - .dbo_read = osd_read, - .dbo_declare_write = osd_declare_write, - .dbo_write = osd_write, - .dbo_bufs_get = osd_bufs_get, - .dbo_bufs_put = osd_bufs_put, - .dbo_write_prep = osd_write_prep, - .dbo_declare_write_commit = osd_declare_write_commit, - .dbo_write_commit = osd_write_commit, - .dbo_read_prep = osd_read_prep, - .dbo_declare_punch = osd_declare_punch, - .dbo_punch = osd_punch, - .dbo_fiemap_get = osd_fiemap_get, + .dbo_read = osd_read, + .dbo_declare_write = osd_declare_write, + .dbo_write = osd_write, + .dbo_bufs_get = osd_bufs_get, + .dbo_bufs_put = osd_bufs_put, + .dbo_write_prep = osd_write_prep, + .dbo_declare_write_commit = osd_declare_write_commit, + .dbo_write_commit = osd_write_commit, + .dbo_read_prep = osd_read_prep, + .dbo_declare_punch = osd_declare_punch, + .dbo_punch = osd_punch, + .dbo_fiemap_get = osd_fiemap_get, + .dbo_ladvise = osd_ladvise, + .dbo_declare_fallocate = osd_declare_fallocate, + .dbo_fallocate = osd_fallocate, }; +/** + * Get a truncate lock + * + * In order to take multi-transaction truncate out of main transaction we let + * the caller grab a lock on the object passed. the lock can be shared (for + * writes) and exclusive (for truncate). It's not allowed to mix truncate + * and write in the same transaction handle (do not confuse with big ldiskfs + * transaction containing lots of handles). + * The lock must be taken at declaration. + * + * \param obj object to lock + * \oh transaction + * \shared shared or exclusive + * + * \retval 0 lock is granted + * \retval -NOMEM no memory to allocate lock + */ +int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared) +{ + struct osd_access_lock *al, *tmp; + + LASSERT(obj); + LASSERT(oh); + + list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) { + if (tmp->tl_obj != obj) + continue; + LASSERT(tmp->tl_shared == shared); + /* found same lock */ + return 0; + } + + OBD_ALLOC_PTR(al); + if (unlikely(al == NULL)) + return -ENOMEM; + al->tl_obj = obj; + al->tl_truncate = false; + if (shared) + down_read(&obj->oo_ext_idx_sem); + else + down_write(&obj->oo_ext_idx_sem); + al->tl_shared = shared; + lu_object_get(&obj->oo_dt.do_lu); + + list_add(&al->tl_list, &oh->ot_trunc_locks); + + return 0; +} + +void osd_trunc_unlock_all(const struct lu_env *env, struct list_head *list) +{ + struct osd_access_lock *al, *tmp; + list_for_each_entry_safe(al, tmp, list, tl_list) { + if (al->tl_shared) + up_read(&al->tl_obj->oo_ext_idx_sem); + else + up_write(&al->tl_obj->oo_ext_idx_sem); + osd_object_put(env, al->tl_obj); + list_del(&al->tl_list); + OBD_FREE_PTR(al); + } +} + +void osd_execute_truncate(struct osd_object *obj) +{ + struct osd_device *d = osd_obj2dev(obj); + struct inode *inode = obj->oo_inode; + __u64 size; + + /* simulate crash before (in the middle) of delayed truncate */ + if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) { + struct ldiskfs_inode_info *ei = LDISKFS_I(inode); + struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb); + + mutex_lock(&sbi->s_orphan_lock); + list_del_init(&ei->i_orphan); + mutex_unlock(&sbi->s_orphan_lock); + return; + } + + inode_lock(inode); + ldiskfs_truncate(inode); + inode_unlock(inode); + + /* + * For a partial-page truncate, flush the page to disk immediately to + * avoid data corruption during direct disk write. b=17397 + */ + size = i_size_read(inode); + if ((size & ~PAGE_MASK) == 0) + return; + if (osd_use_page_cache(d)) { + filemap_fdatawrite_range(inode->i_mapping, size, size + 1); + } else { + /* Notice we use "wait" version to ensure I/O is complete */ + filemap_write_and_wait_range(inode->i_mapping, size, size + 1); + invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT, + size >> PAGE_SHIFT); + } +} + +void osd_process_truncates(struct list_head *list) +{ + struct osd_access_lock *al; + + LASSERT(journal_current_handle() == NULL); + + list_for_each_entry(al, list, tl_list) { + if (al->tl_shared) + continue; + if (!al->tl_truncate) + continue; + osd_execute_truncate(al->tl_obj); + } +}