X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fosd-ldiskfs%2Fosd_io.c;h=64b995dd6d246ca398cb4982f753750ed8be3924;hb=7599dd3d20d6bb4ee89634c5a76730481ca62470;hp=c5c70e6e1edd4cf693e7bb50b3b16cbd2b7632ea;hpb=4ce3219eb8e6a07c5c37e4b425b29195488005c3;p=fs%2Flustre-release.git diff --git a/lustre/osd-ldiskfs/osd_io.c b/lustre/osd-ldiskfs/osd_io.c index c5c70e6..64b995d 100644 --- a/lustre/osd-ldiskfs/osd_io.c +++ b/lustre/osd-ldiskfs/osd_io.c @@ -23,7 +23,7 @@ * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2012, 2016, Intel Corporation. + * Copyright (c) 2012, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -38,12 +38,12 @@ * */ -/* LUSTRE_VERSION_CODE */ -#include /* prerequisite for linux/xattr.h */ #include /* prerequisite for linux/xattr.h */ #include +#include +#include /* * struct OBD_{ALLOC,FREE}*() @@ -56,6 +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) { @@ -73,7 +82,7 @@ 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; @@ -108,6 +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; @@ -115,10 +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) @@ -132,25 +150,20 @@ 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_error; + int error = bio->bi_status; #else static void dio_complete_routine(struct bio *bio, int error) { #endif struct osd_iobuf *iobuf = bio->bi_private; - int iter; struct bio_vec *bvl; /* CAVEAT EMPTOR: possibly in IRQ context @@ -160,29 +173,26 @@ 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, - (unsigned long)bio->bi_flags, - bio->bi_rw, bio->bi_vcnt, bio_idx(bio), - bio_sectors(bio) << 9, bio->bi_end_io, -#ifdef HAVE_BI_CNT - atomic_read(&bio->bi_cnt), -#else - atomic_read(&bio->__bi_cnt), -#endif - 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_all(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(bvl_to_page(bvl)); LASSERT(PageLocked(bvl_to_page(bvl))); @@ -204,7 +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)) @@ -243,11 +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) @@ -258,32 +272,214 @@ 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_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; - 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; - DECLARE_PLUG(plug); + 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; @@ -324,23 +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) * @@ -353,11 +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); @@ -366,6 +573,15 @@ 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; @@ -378,20 +594,26 @@ out: * 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; @@ -400,6 +622,11 @@ static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages, int poff = offset & (PAGE_SIZE - 1); int plen = PAGE_SIZE - poff; + if (*nrpages >= maxlnb) { + rc = -EOVERFLOW; + break; + } + if (plen > len) plen = len; lnb->lnb_file_offset = offset; @@ -409,6 +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); @@ -418,23 +648,66 @@ static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages, (*nrpages)++; } - RETURN(0); + RETURN(rc); } -static 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_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(!test_bit(PG_private_2, &page->flags)); + else + lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1); + + 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) + return page; + } + + LASSERT(oti->oti_dio_pages); + cur = oti->oti_dio_pages_used; + + if (unlikely(!oti->oti_dio_pages[cur])) { + LASSERT(cur < PTLRPC_MAX_BRW_PAGES); + page = alloc_page(gfp_mask); + if (!page) + return NULL; + oti->oti_dio_pages[cur] = page; + } + + page = oti->oti_dio_pages[cur]; + LASSERT(!test_bit(PG_private_2, &page->flags)); + set_bit(PG_private_2, &page->flags); + oti->oti_dio_pages_used++; - return page; + LASSERT(!PageLocked(page)); + lock_page(page); + + LASSERT(!page->mapping); + LASSERT(!PageWriteback(page)); + ClearPageUptodate(page); + + page->index = offset >> PAGE_SHIFT; + + return page; } /* @@ -471,18 +744,42 @@ static struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw) 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; + ll_pagevec_init(&pvec, 0); + for (i = 0; i < npages; i++) { - if (lnb[i].lnb_page == NULL) + struct page *page = lnb[i].lnb_page; + + if (page == NULL) continue; - LASSERT(PageLocked(lnb[i].lnb_page)); - unlock_page(lnb[i].lnb_page); - put_page(lnb[i].lnb_page); + + /* if the page isn't cached, then reset uptodate + * to prevent reuse */ + if (test_bit(PG_private_2, &page->flags)) { + clear_bit(PG_private_2, &page->flags); + ClearPageUptodate(page); + if (lnb[i].lnb_locked) + unlock_page(page); + oti->oti_dio_pages_used--; + } else { + if (lnb[i].lnb_locked) + unlock_page(page); + if (pagevec_add(&pvec, page) == 0) + pagevec_release(&pvec); + } dt_object_put(env, dt); + lnb[i].lnb_page = NULL; } + LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used); + + /* Release any partial pagevec */ + pagevec_release(&pvec); + RETURN(0); } @@ -504,7 +801,7 @@ static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt, * \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? + * \param rw read or write operation, and other flags * \param capa capabilities * * \retval pages (zero or more) loaded successfully @@ -512,357 +809,100 @@ static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt, */ 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 rw) + int maxlnb, enum dt_bufs_type rw) { - struct osd_object *obj = osd_dt_obj(dt); - int npages, i, rc = 0; + 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); - osd_map_remote_to_local(pos, len, &npages, lnb); - - for (i = 0; i < npages; i++, lnb++) { - lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset, rw); - if (lnb->lnb_page == NULL) - GOTO(cleanup, rc = -ENOMEM); - - wait_on_page_writeback(lnb->lnb_page); - BUG_ON(PageWriteback(lnb->lnb_page)); - - lu_object_get(&dt->do_lu); - } - - RETURN(i); - -cleanup: - if (i > 0) - osd_bufs_put(env, dt, lnb - i, i); - return rc; -} - -#ifndef HAVE_LDISKFS_MAP_BLOCKS - -#ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */ -#define ldiskfs_ext_pblock(ex) ext_pblock((ex)) -#endif - -struct bpointers { - sector_t *blocks; - unsigned long start; - int num; - int init_num; - int create; -}; - -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; - } - - /* 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; -} - -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; - - /* 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; -} - -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) -{ - struct bpointers *bp = cbdata; - struct ldiskfs_extent nex; - unsigned long pblock = 0; - 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; - } + 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; - if (pblock != 0) { - /* 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; -} - -static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index, - int clen, sector_t *blocks, int create) -{ - int blocks_per_page = PAGE_SIZE >> inode->i_blkbits; - struct bpointers bp; - int err; - - if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT) - return -EFBIG; - bp.blocks = blocks; - bp.start = index * blocks_per_page; - bp.init_num = bp.num = clen * blocks_per_page; - bp.create = create; - - CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n", - bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino); - - err = ldiskfs_ext_walk_space(inode, bp.start, bp.num, - ldiskfs_ext_new_extent_cb, &bp); - ldiskfs_ext_invalidate_cache(inode); - - return err; -} - -static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode, - struct page **page, int pages, - sector_t *blocks, int create) -{ - int blocks_per_page = PAGE_SIZE >> inode->i_blkbits; - pgoff_t bitmap_max_page_index; - sector_t *b; - int rc = 0, i; - - bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >> - PAGE_SHIFT; - for (i = 0, b = blocks; i < pages; i++, page++) { - if ((*page)->index + 1 >= bitmap_max_page_index) { - rc = -EFBIG; - break; - } - rc = ldiskfs_map_inode_page(inode, *page, b, create); - if (rc) { - CERROR("ino %lu, blk %llu create %d: rc %d\n", - inode->i_ino, - (unsigned long long)*b, create, rc); - break; - } - b += blocks_per_page; +bypass_checks: + if (!cache && unlikely(!oti->oti_dio_pages)) { + OBD_ALLOC(oti->oti_dio_pages, + sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES); + if (!oti->oti_dio_pages) + return -ENOMEM; } - return rc; -} - -static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode, - struct page **page, - int pages, sector_t *blocks, - int create) -{ - int rc = 0, i = 0, clen = 0; - struct page *fp = NULL; - CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n", - inode->i_ino, pages, (*page)->index); + /* 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); - /* 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; - } + lnb->lnb_locked = 1; + wait_on_page_writeback(lnb->lnb_page); + BUG_ON(PageWriteback(lnb->lnb_page)); - /* process found extent */ - rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen, - blocks, create); - if (rc) - GOTO(cleanup, rc); + lu_object_get(&dt->do_lu); + } - /* look for next extent */ - fp = NULL; - blocks += clen * (PAGE_SIZE >> inode->i_blkbits); +#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 - if (fp) - rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen, - blocks, create); + RETURN(i); cleanup: + if (i > 0) + osd_bufs_put(env, dt, lnb - i, i); return rc; } @@ -870,22 +910,6 @@ static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page, int pages, sector_t *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, sector_t *blocks, - int create) -{ int blocks_per_page = PAGE_SIZE >> inode->i_blkbits; int rc = 0, i = 0; struct page *fp = NULL; @@ -945,9 +969,9 @@ 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; @@ -967,7 +991,6 @@ 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) @@ -976,14 +999,11 @@ 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)); - ktime_t start; - ktime_t end; + ktime_t start, end; s64 timediff; - ssize_t isize; - __s64 maxidx; - int rc = 0; - int i; - int cache = 0; + ssize_t isize; + __s64 maxidx; + int i, rc = 0; LASSERT(inode); @@ -994,18 +1014,9 @@ static int osd_write_prep(const struct lu_env *env, struct dt_object *dt, isize = i_size_read(inode); maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1; - if (osd->od_writethrough_cache) - cache = 1; - if (isize > osd->od_readcache_max_filesize) - cache = 0; - 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 @@ -1017,7 +1028,7 @@ static int osd_write_prep(const struct lu_env *env, struct dt_object *dt, 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); @@ -1080,7 +1091,7 @@ static int osd_is_mapped(struct dt_object *dt, __u64 offset, fei.fi_extents_start = &fe; saved_fs = get_fs(); - set_fs(get_ds()); + set_fs(KERNEL_DS); rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset); set_fs(saved_fs); if (rc != 0) @@ -1115,10 +1126,10 @@ static int osd_declare_write_commit(const struct lu_env *env, int i; int newblocks; int rc = 0; - int flags = 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; @@ -1134,7 +1145,9 @@ static int osd_declare_write_commit(const struct lu_env *env, lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len) extents++; - if (!osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent)) + 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 @@ -1201,18 +1214,21 @@ static int osd_declare_write_commit(const struct lu_env *env, rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode), i_projid_read(inode), quota_space, oh, - osd_dt_obj(dt), &flags, declare_flags); + 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 (flags & QUOTA_FL_OVER_PRJQUOTA) + 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); } @@ -1227,7 +1243,6 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt)); loff_t isize; int rc = 0, i; - struct osd_fextent extent = { 0 }; LASSERT(inode); @@ -1236,11 +1251,11 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, RETURN(rc); isize = i_size_read(inode); - ll_vfs_dq_init(inode); + dquot_initialize(inode); for (i = 0; i < npages; i++) { if (lnb[i].lnb_rc == -ENOSPC && - osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent)) { + (lnb[i].lnb_flags & OBD_BRW_MAPPED)) { /* Allow the write to proceed if overwriting an * existing block */ lnb[i].lnb_rc = 0; @@ -1270,7 +1285,7 @@ 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]); } osd_trans_exec_op(env, thandle, OSD_OT_WRITE); @@ -1327,7 +1342,7 @@ 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)); - int rc = 0, i, 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; @@ -1340,11 +1355,6 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, isize = i_size_read(inode); - if (osd->od_read_cache) - cache = 1; - if (isize > osd->od_readcache_max_filesize) - cache = 0; - start = ktime_get(); for (i = 0; i < npages; i++) { @@ -1353,10 +1363,10 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, * lnb->lnb_rc == 0, so it's easy to detect later. */ break; - if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len) - lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset; - else - lnb[i].lnb_rc = 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)) @@ -1364,14 +1374,17 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, 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; } end = ktime_get(); timediff = ktime_us_delta(end, start); @@ -1387,16 +1400,22 @@ 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])); + unlock_page(iobuf->dr_pages[i]); + } + } + + RETURN(rc); } /* @@ -1434,7 +1453,8 @@ 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) { @@ -1454,8 +1474,8 @@ int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs) csize = min(blocksize - boffs, size); bh = __ldiskfs_bread(NULL, inode, block, 0); if (IS_ERR(bh)) { - CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n", - LDISKFS_SB(inode->i_sb)->s_es->s_volume_name, + 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); @@ -1499,8 +1519,7 @@ 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 (LDISKFS_HAS_INCOMPAT_FEATURE(sb, - LDISKFS_FEATURE_INCOMPAT_EXTENTS)) { + } else if (ldiskfs_has_feature_extents(sb)) { return 1; } return 0; @@ -1616,7 +1635,7 @@ 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; @@ -1640,6 +1659,10 @@ out: 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); } @@ -1658,9 +1681,11 @@ static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen) 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); @@ -1670,6 +1695,8 @@ int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize, int size; int boffs; int dirty_inode = 0; + struct ldiskfs_inode_info *ei = LDISKFS_I(inode); + bool create, sparse, sync = false; if (write_NUL) { /* @@ -1680,14 +1707,55 @@ 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); + + /* 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); - bh = __ldiskfs_bread(handle, inode, block, 1); + 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; @@ -1695,9 +1763,11 @@ int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize, err = PTR_ERR(bh); bh = NULL; } - CERROR("%s: error reading offset %llu (block %lu): " - "rc = %d\n", - inode->i_sb->s_id, offset, block, err); + + 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; } @@ -1710,7 +1780,14 @@ 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); + 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; @@ -1721,8 +1798,11 @@ 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; @@ -1747,7 +1827,7 @@ 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, int ignore_quota) + struct thandle *handle) { struct inode *inode = osd_dt_obj(dt)->oo_inode; struct osd_thandle *oh; @@ -1758,7 +1838,7 @@ static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt, 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); */ @@ -1776,9 +1856,8 @@ 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); + 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; @@ -1815,73 +1894,80 @@ static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt, 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); + + 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 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); - - /* - * For a partial-page truncate, flush the page to disk immediately to - * avoid data corruption during direct disk write. b=17397 - */ - if ((start & ~PAGE_MASK) != 0) - rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1); - h = journal_current_handle(); - LASSERT(h != NULL); - LASSERT(h == oh->ot_handle); + /* optimize grow case */ + if (grow) { + osd_execute_truncate(obj); + GOTO(out, rc); + } - /* do not check credits with osd_trans_exec_check() as the truncate - * can restart the transaction internally and we restart the - * transaction in this case */ + /* 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); + if (rc != 0) + GOTO(out, rc); - 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); - } - } + 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); - RETURN(rc == 0 ? rc2 : rc); +out: + RETURN(rc); } static int fiemap_check_ranges(struct inode *inode, @@ -1921,7 +2007,7 @@ static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt, 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) @@ -1941,27 +2027,34 @@ static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt, 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_ladvise(const struct lu_env *env, struct dt_object *dt, __u64 start, __u64 end, enum lu_ladvise_type advice) { - int rc = 0; - struct inode *inode = osd_dt_obj(dt)->oo_inode; + struct osd_object *obj = osd_dt_obj(dt); + int rc = 0; ENTRY; switch (advice) { case LU_LADVISE_DONTNEED: - if (end == 0) - break; - invalidate_mapping_pages(inode->i_mapping, - start >> PAGE_CACHE_SHIFT, - (end - 1) >> PAGE_CACHE_SHIFT); + if (end) + invalidate_mapping_pages(obj->oo_inode->i_mapping, + start >> PAGE_SHIFT, + (end - 1) >> PAGE_SHIFT); break; default: rc = -ENOTSUPP; @@ -1994,3 +2087,117 @@ const struct dt_body_operations osd_body_ops = { .dbo_fiemap_get = osd_fiemap_get, .dbo_ladvise = osd_ladvise, }; + +/** + * 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; + + list_add(&al->tl_list, &oh->ot_trunc_locks); + + return 0; +} + +void osd_trunc_unlock_all(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); + 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); + } +}