X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosd-ldiskfs%2Fosd_io.c;h=7631d61248babf1ed5265ead6a67848694c42deb;hp=2164421eebebbc29364b7dbe06fe93acca6d58c5;hb=4aa17923f08bea425a20961cb6eaa72ad9af38c1;hpb=7b136af8f46024fd48a773e5a79b817bca05d9e8 diff --git a/lustre/osd-ldiskfs/osd_io.c b/lustre/osd-ldiskfs/osd_io.c index 2164421..7631d61 100644 --- a/lustre/osd-ldiskfs/osd_io.c +++ b/lustre/osd-ldiskfs/osd_io.c @@ -38,6 +38,8 @@ * */ +#define DEBUG_SUBSYSTEM S_OSD + /* prerequisite for linux/xattr.h */ #include /* prerequisite for linux/xattr.h */ @@ -100,12 +102,12 @@ static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf, i <<= 1; CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n", - (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages); + (unsigned int)(pages * sizeof(iobuf->dr_pages[0])), i, pages); pages = i; 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); + (unsigned int)(blocks * sizeof(iobuf->dr_blocks[0])), blocks); lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0])); iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf; @@ -141,24 +143,23 @@ static void osd_iobuf_add_page(struct osd_iobuf *iobuf, void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf) { - int rw = iobuf->dr_rw; - - if (iobuf->dr_elapsed_valid) { - iobuf->dr_elapsed_valid = 0; - LASSERT(iobuf->dr_dev == d); - LASSERT(iobuf->dr_frags > 0); - lprocfs_oh_tally(&d->od_brw_stats. - hist[BRW_R_DIO_FRAGS+rw], - iobuf->dr_frags); + int rw = iobuf->dr_rw; + + if (iobuf->dr_elapsed_valid) { + iobuf->dr_elapsed_valid = 0; + LASSERT(iobuf->dr_dev == d); + LASSERT(iobuf->dr_frags > 0); + 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], ktime_to_ms(iobuf->dr_elapsed)); - } + } } #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG static void dio_complete_routine(struct bio *bio) { - int error = bio->bi_status; + int error = blk_status_to_errno(bio->bi_status); #else static void dio_complete_routine(struct bio *bio, int error) { @@ -166,18 +167,12 @@ static void dio_complete_routine(struct bio *bio, int error) struct osd_iobuf *iobuf = bio->bi_private; struct bio_vec *bvl; - /* CAVEAT EMPTOR: possibly in IRQ context - * DO NOT record procfs stats here!!! */ + /* CAVEAT EMPTOR: possibly in IRQ context + * DO NOT record procfs stats here!!! + */ if (unlikely(iobuf == NULL)) { - 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 " - "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("***** 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 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, " __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, @@ -226,7 +221,8 @@ static void dio_complete_routine(struct bio *bio, int error) * filter_clear_dreq(). It was then possible to exhaust the biovec-256 * mempool when serious on-disk fragmentation was encountered, * deadlocking the OST. The bios are now released as soon as complete - * so the pool cannot be exhausted while IOs are competing. bug 10076 */ + * so the pool cannot be exhausted while IOs are competing. b=10076 + */ bio_put(bio); } @@ -306,9 +302,8 @@ static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_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", + 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; @@ -446,13 +441,13 @@ static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf, } static int osd_do_bio(struct osd_device *osd, struct inode *inode, - struct osd_iobuf *iobuf) + struct osd_iobuf *iobuf, sector_t start_blocks, + sector_t count) { 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; @@ -464,60 +459,69 @@ static int osd_do_bio(struct osd_device *osd, struct inode *inode, unsigned int page_offset; sector_t sector; int nblocks; - int block_idx; - int page_idx; + int block_idx, block_idx_end; + int page_idx, page_idx_start; int i; int rc = 0; bool fault_inject; bool integrity_enabled; struct blk_plug plug; + int blocks_left_page; + ENTRY; fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT); - LASSERT(iobuf->dr_npages == npages); + LASSERT(iobuf->dr_npages == npages); integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw); osd_brw_stats_update(osd, iobuf); iobuf->dr_start_time = ktime_get(); + if (!count) + count = npages * blocks_per_page; + block_idx_end = start_blocks + count; + blk_start_plug(&plug); - for (page_idx = 0, block_idx = 0; - page_idx < npages; - page_idx++, block_idx += blocks_per_page) { - - page = pages[page_idx]; - LASSERT(block_idx + blocks_per_page <= total_blocks); - - for (i = 0, page_offset = 0; - i < blocks_per_page; - i += nblocks, page_offset += blocksize * nblocks) { - - nblocks = 1; - - if (blocks[block_idx + i] == 0) { /* hole */ - LASSERTF(iobuf->dr_rw == 0, - "page_idx %u, block_idx %u, i %u\n", - page_idx, block_idx, i); - memset(kmap(page) + page_offset, 0, blocksize); - kunmap(page); - continue; - } - - sector = (sector_t)blocks[block_idx + i] << sector_bits; - - /* Additional contiguous file blocks? */ - while (i + nblocks < blocks_per_page && - (sector + (nblocks << sector_bits)) == - ((sector_t)blocks[block_idx + i + nblocks] << - sector_bits)) - nblocks++; - - if (bio != NULL && - can_be_merged(bio, sector) && - bio_add_page(bio, page, - blocksize * nblocks, page_offset) != 0) - continue; /* added this frag OK */ + + page_idx_start = start_blocks / blocks_per_page; + for (page_idx = page_idx_start, block_idx = start_blocks; + block_idx < block_idx_end; page_idx++, + block_idx += blocks_left_page) { + page = pages[page_idx]; + LASSERT(page_idx < iobuf->dr_npages); + + i = block_idx % blocks_per_page; + blocks_left_page = blocks_per_page - i; + for (page_offset = i * blocksize; i < blocks_left_page; + i += nblocks, page_offset += blocksize * nblocks) { + nblocks = 1; + + if (blocks[block_idx + i] == 0) { /* hole */ + LASSERTF(iobuf->dr_rw == 0, + "page_idx %u, block_idx %u, i %u," + "start_blocks: %llu, count: %llu, npages: %d\n", + page_idx, block_idx, i, + (unsigned long long)start_blocks, + (unsigned long long)count, npages); + memset(kmap(page) + page_offset, 0, blocksize); + kunmap(page); + continue; + } + + sector = (sector_t)blocks[block_idx + i] << sector_bits; + + /* Additional contiguous file blocks? */ + while (i + nblocks < blocks_left_page && + (sector + (nblocks << sector_bits)) == + ((sector_t)blocks[block_idx + i + nblocks] << + sector_bits)) + nblocks++; + + if (bio && can_be_merged(bio, sector) && + bio_add_page(bio, page, blocksize * nblocks, + page_offset) != 0) + continue; /* added this frag OK */ if (bio != NULL) { struct request_queue *q = bio_get_queue(bio); @@ -546,15 +550,15 @@ static int osd_do_bio(struct osd_device *osd, struct inode *inode, bio_start_page_idx = page_idx; /* allocate new bio */ bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES, - (npages - page_idx) * - blocks_per_page)); - if (bio == NULL) { - CERROR("Can't allocate bio %u*%u = %u pages\n", - (npages - page_idx), blocks_per_page, - (npages - page_idx) * blocks_per_page); - rc = -ENOMEM; - goto out; - } + (block_idx_end - block_idx + + blocks_left_page - 1))); + if (bio == NULL) { + CERROR("Can't allocate bio %u pages\n", + block_idx_end - block_idx + + blocks_left_page - 1); + rc = -ENOMEM; + goto out; + } bio_set_dev(bio, bdev); bio_set_sector(bio, sector); @@ -593,7 +597,8 @@ out: /* 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 */ + * see osd_trans_stop() for more details -bzzz + */ if (iobuf->dr_rw == 0 || fault_inject) { wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0); @@ -614,11 +619,11 @@ static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages, struct niobuf_local *lnb, int maxlnb) { int rc = 0; - ENTRY; + ENTRY; - *nrpages = 0; + *nrpages = 0; - while (len > 0) { + while (len > 0) { int poff = offset & (PAGE_SIZE - 1); int plen = PAGE_SIZE - poff; @@ -627,8 +632,8 @@ static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages, break; } - if (plen > len) - plen = len; + if (plen > len) + plen = len; lnb->lnb_file_offset = offset; lnb->lnb_page_offset = poff; lnb->lnb_len = plen; @@ -640,13 +645,13 @@ static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages, lnb->lnb_guard_disk = 0; lnb->lnb_locked = 0; - LASSERTF(plen <= len, "plen %u, len %lld\n", plen, - (long long) len); - offset += plen; - len -= plen; - lnb++; - (*nrpages)++; - } + LASSERTF(plen <= len, "plen %u, len %lld\n", plen, + (long long) len); + offset += plen; + len -= plen; + lnb++; + (*nrpages)++; + } RETURN(rc); } @@ -660,16 +665,18 @@ static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt, struct page *page; int cur; - LASSERT(inode); + LASSERT(inode); 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 + 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; } @@ -678,34 +685,29 @@ static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt, /* 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) + 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(!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); } - 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++; - - LASSERT(!PageLocked(page)); - lock_page(page); - - LASSERT(!page->mapping); - LASSERT(!PageWriteback(page)); ClearPageUptodate(page); - page->index = offset >> PAGE_SHIFT; + oti->oti_dio_pages_used++; return page; } @@ -757,12 +759,9 @@ static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt, continue; /* 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); + * to prevent reuse + */ + if (PagePrivate2(page)) { oti->oti_dio_pages_used--; } else { if (lnb[i].lnb_locked) @@ -770,7 +769,6 @@ static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt, if (pagevec_add(&pvec, page) == 0) pagevec_release(&pvec); } - dt_object_put(env, dt); lnb[i].lnb_page = NULL; } @@ -865,8 +863,8 @@ static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt, bypass_checks: if (!cache && unlikely(!oti->oti_dio_pages)) { - OBD_ALLOC(oti->oti_dio_pages, - sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES); + OBD_ALLOC_PTR_ARRAY_LARGE(oti->oti_dio_pages, + PTLRPC_MAX_BRW_PAGES); if (!oti->oti_dio_pages) return -ENOMEM; } @@ -881,10 +879,6 @@ bypass_checks: GOTO(cleanup, rc = -ENOMEM); lnb->lnb_locked = 1; - wait_on_page_writeback(lnb->lnb_page); - BUG_ON(PageWriteback(lnb->lnb_page)); - - lu_object_get(&dt->do_lu); } #if 0 @@ -905,17 +899,130 @@ cleanup: osd_bufs_put(env, dt, lnb - i, i); return rc; } +/* Borrow @ext4_chunk_trans_blocks */ +static int osd_chunk_trans_blocks(struct inode *inode, int nrblocks) +{ + ldiskfs_group_t groups; + int gdpblocks; + int idxblocks; + int depth; + int ret; + + depth = ext_depth(inode); + idxblocks = depth * 2; + + /* + * Now let's see how many group bitmaps and group descriptors need + * to account. + */ + groups = idxblocks + 1; + gdpblocks = groups; + if (groups > LDISKFS_SB(inode->i_sb)->s_groups_count) + groups = LDISKFS_SB(inode->i_sb)->s_groups_count; + if (gdpblocks > LDISKFS_SB(inode->i_sb)->s_gdb_count) + gdpblocks = LDISKFS_SB(inode->i_sb)->s_gdb_count; -static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page, - int pages, sector_t *blocks, - int create) + /* bitmaps and block group descriptor blocks */ + ret = idxblocks + groups + gdpblocks; + + /* Blocks for super block, inode, quota and xattr blocks */ + ret += LDISKFS_META_TRANS_BLOCKS(inode->i_sb); + + return ret; +} + +#ifdef HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS +static int osd_extend_trans(handle_t *handle, int needed, + struct inode *inode) +{ + return __ldiskfs_journal_ensure_credits(handle, needed, needed, + ldiskfs_trans_default_revoke_credits(inode->i_sb)); +} + +static int osd_extend_restart_trans(handle_t *handle, int needed, + struct inode *inode) +{ + int rc; + + rc = ldiskfs_journal_ensure_credits(handle, needed, + ldiskfs_trans_default_revoke_credits(inode->i_sb)); + /* this means journal has been restarted */ + if (rc > 0) + rc = 0; + + return rc; +} +#else +static int osd_extend_trans(handle_t *handle, int needed, + struct inode *inode) +{ + if (ldiskfs_handle_has_enough_credits(handle, needed)) + return 0; + + return ldiskfs_journal_extend(handle, + needed - handle->h_buffer_credits); +} + +static int osd_extend_restart_trans(handle_t *handle, int needed, + struct inode *inode) +{ + + int rc = osd_extend_trans(handle, needed, inode); + + if (rc <= 0) + return rc; + + return ldiskfs_journal_restart(handle, needed); +} +#endif /* HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS */ + +static int osd_ldiskfs_map_write(struct inode *inode, struct osd_iobuf *iobuf, + struct osd_device *osd, sector_t start_blocks, + sector_t count, loff_t *disk_size, + __u64 user_size) +{ + /* if file has grown, take user_size into account */ + if (user_size && *disk_size > user_size) + *disk_size = user_size; + + 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); + } + + /* + * We don't do stats here as in read path because + * write is async: we'll do this in osd_put_bufs() + */ + return osd_do_bio(osd, inode, iobuf, start_blocks, count); +} + + +static int osd_ldiskfs_map_inode_pages(struct inode *inode, + struct osd_iobuf *iobuf, + struct osd_device *osd, + int create, __u64 user_size, + int check_credits) { int blocks_per_page = PAGE_SIZE >> inode->i_blkbits; - int rc = 0, i = 0; + int rc = 0, i = 0, mapped_index = 0; struct page *fp = NULL; int clen = 0; pgoff_t max_page_index; handle_t *handle = NULL; + int credits; + sector_t start_blocks = 0, count = 0; + loff_t disk_size = 0; + struct page **page = iobuf->dr_pages; + int pages = iobuf->dr_npages; + sector_t *blocks = iobuf->dr_blocks; + struct niobuf_local *lnb1, *lnb2; + loff_t size1, size2; max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT; @@ -929,11 +1036,18 @@ static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page, rc = osd_attach_jinode(inode); if (rc) return rc; + 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; } /* pages are sorted already. so, we just have to find - * contig. space and process them properly */ + * contig. space and process them properly + */ while (i < pages) { - long blen, total = 0; + long blen, total = 0, previous_total = 0; struct ldiskfs_map_blocks map = { 0 }; if (fp == NULL) { /* start new extent */ @@ -954,31 +1068,89 @@ static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page, map.m_lblk = fp->index * blocks_per_page; map.m_len = blen = clen * blocks_per_page; cont_map: + /** + * We might restart transaction for block allocations, + * in order to make sure data ordered mode, issue IO, disk + * size update and block allocations need be within same + * transaction to make sure consistency. + */ + if (handle && check_credits) { + /* + * credits to insert 1 extent into extent tree. + */ + credits = osd_chunk_trans_blocks(inode, blen); + rc = osd_extend_trans(handle, credits, inode); + if (rc < 0) + GOTO(cleanup, rc); + /* + * only issue IO if restart transaction needed, + * as update disk size need hold inode lock, we + * want to avoid that as much as possible. + */ + if (rc > 0) { + WARN_ON_ONCE(start_blocks == 0); + rc = osd_ldiskfs_map_write(inode, + iobuf, osd, start_blocks, + count, &disk_size, user_size); + if (rc) + GOTO(cleanup, rc); +#ifdef HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS + rc = ldiskfs_journal_restart(handle, credits, + ldiskfs_trans_default_revoke_credits(inode->i_sb)); +#else + rc = ldiskfs_journal_restart(handle, credits); +#endif + if (rc) + GOTO(cleanup, rc); + start_blocks += count; + /* reset IO block count */ + count = 0; + } + } rc = ldiskfs_map_blocks(handle, inode, &map, create); if (rc >= 0) { int c = 0; + for (; total < blen && c < map.m_len; c++, total++) { if (rc == 0) { *(blocks + total) = 0; total++; break; - } else { - *(blocks + total) = map.m_pblk + c; - /* unmap any possible underlying - * metadata from the block device - * mapping. bug 6998. */ - if ((map.m_flags & LDISKFS_MAP_NEW) && - create) - clean_bdev_aliases( - inode->i_sb->s_bdev, - map.m_pblk + c, 1); } + *(blocks + total) = map.m_pblk + c; + /* unmap any possible underlying + * metadata from the block device + * mapping. b=6998. + */ + if ((map.m_flags & LDISKFS_MAP_NEW) && + create) + clean_bdev_aliases(inode->i_sb->s_bdev, + map.m_pblk + c, 1); } rc = 0; } + + if (rc == 0 && create) { + count += (total - previous_total); + mapped_index = (count + blocks_per_page - + 1) / blocks_per_page - 1; + lnb1 = iobuf->dr_lnbs[i - clen]; + lnb2 = iobuf->dr_lnbs[mapped_index]; + size1 = lnb1->lnb_file_offset - + (lnb1->lnb_file_offset % PAGE_SIZE) + + (total << inode->i_blkbits); + size2 = lnb2->lnb_file_offset + lnb2->lnb_len; + + if (size1 > size2) + size1 = size2; + if (size1 > disk_size) + disk_size = size1; + } + if (rc == 0 && total < blen) { map.m_lblk = fp->index * blocks_per_page + total; map.m_len = blen - total; + previous_total = total; goto cont_map; } if (rc != 0) @@ -989,23 +1161,29 @@ cont_map: blocks += blocks_per_page * clen; } cleanup: + if (rc == 0 && create && + start_blocks < pages * blocks_per_page) { + rc = osd_ldiskfs_map_write(inode, iobuf, osd, start_blocks, + count, &disk_size, user_size); + LASSERT(start_blocks + count == pages * blocks_per_page); + } return rc; } static int osd_write_prep(const struct lu_env *env, struct dt_object *dt, - struct niobuf_local *lnb, int npages) + struct niobuf_local *lnb, int npages) { - 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)); + 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)); ktime_t start, end; s64 timediff; ssize_t isize; __s64 maxidx; int i, rc = 0; - LASSERT(inode); + LASSERT(inode); rc = osd_init_iobuf(osd, iobuf, 0, npages); if (unlikely(rc != 0)) @@ -1047,17 +1225,16 @@ static int osd_write_prep(const struct lu_env *env, struct dt_object *dt, timediff = ktime_us_delta(end, start); lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff); - if (iobuf->dr_npages) { - rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages, - iobuf->dr_npages, - iobuf->dr_blocks, 0); - if (likely(rc == 0)) { - rc = osd_do_bio(osd, inode, iobuf); - /* do IO stats for preparation reads */ - osd_fini_iobuf(osd, iobuf); - } - } - RETURN(rc); + if (iobuf->dr_npages) { + rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0, + 0, 0); + if (likely(rc == 0)) { + rc = osd_do_bio(osd, inode, iobuf, 0, 0); + /* do IO stats for preparation reads */ + osd_fini_iobuf(osd, iobuf); + } + } + RETURN(rc); } struct osd_fextent { @@ -1114,42 +1291,31 @@ static int osd_is_mapped(struct dt_object *dt, __u64 offset, } static int osd_declare_write_commit(const struct lu_env *env, - struct dt_object *dt, - struct niobuf_local *lnb, int npages, - struct thandle *handle) + struct dt_object *dt, + 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 extents = 0; int depth; int i; - int newblocks; + int newblocks = 0; int rc = 0; - int flags = 0; int credits = 0; long long quota_space = 0; - struct osd_fextent extent = { 0 }; + struct osd_fextent mapped = { 0 }, 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); - - newblocks = npages; + LASSERT(handle != NULL); + oh = container_of(handle, struct osd_thandle, ot_super); + LASSERT(oh->ot_handle == NULL); - /* 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(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. * @@ -1158,55 +1324,85 @@ static int osd_declare_write_commit(const struct lu_env *env, * !(lnb[i].flags & OBD_BRW_SYNC) * * XXX we could handle this on per-lnb basis as done by - * grant. */ + * grant. + */ if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) || (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) == OBD_BRW_FROM_GRANT) declare_flags |= OSD_QID_FORCE; + + if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &mapped)) { + lnb[i].lnb_flags |= OBD_BRW_MAPPED; + continue; + } + + /* count only unmapped changes */ + newblocks++; + if (lnb[i].lnb_file_offset != extent.end || extent.end == 0) { + extents++; + extent.end = lnb[i].lnb_file_offset + lnb[i].lnb_len; + } else { + extent.end += lnb[i].lnb_len; + } + + quota_space += PAGE_SIZE; } - /* - * each extent can go into new leaf causing a split - * 5 is max tree depth: inode + 4 index blocks - * with blockmaps, depth is 3 at most - */ - if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) { - /* - * many concurrent threads may grow tree by the time - * our transaction starts. so, consider 2 is a min depth - */ - depth = ext_depth(inode); - depth = max(depth, 1) + 1; - newblocks += depth; - credits++; /* inode */ + credits++; /* inode */ + /* + * overwrite case, no need to modify tree and + * allocate blocks. + */ + if (!newblocks) + goto out_declare; + /* + * each extent can go into new leaf causing a split + * 5 is max tree depth: inode + 4 index blocks + * with blockmaps, depth is 3 at most + */ + if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) { + /* + * many concurrent threads may grow tree by the time + * our transaction starts. so, consider 2 is a min depth + */ + depth = ext_depth(inode); + depth = max(depth, 1) + 1; + newblocks += depth; credits += depth * 2 * extents; } else { depth = 3; newblocks += depth; - credits++; /* inode */ credits += depth * extents; } + /* + * try a bit more extents to avoid restart + * as much as possible in normal case. + */ + if (npages > 1 && extents) + extents <<= 1; + /* quota space for metadata blocks */ quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd)); /* quota space should be reported in 1K blocks */ quota_space = toqb(quota_space); - /* each new block can go in different group (bitmap + gd) */ + /* each new block can go in different group (bitmap + gd) */ - /* we can't dirty more bitmap blocks than exist */ - if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count) + /* we can't dirty more bitmap blocks than exist */ + if (extents > LDISKFS_SB(osd_sb(osd))->s_groups_count) credits += LDISKFS_SB(osd_sb(osd))->s_groups_count; - else - credits += newblocks; + else + credits += extents; /* we can't dirty more gd blocks than exist */ - if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count) + if (extents > LDISKFS_SB(osd_sb(osd))->s_gdb_count) credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count; else - credits += newblocks; + credits += extents; +out_declare: osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits); /* make sure the over quota flags were not set */ @@ -1214,16 +1410,17 @@ 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) + * revised. + */ + 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) @@ -1234,30 +1431,32 @@ static int osd_declare_write_commit(const struct lu_env *env, /* 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)); + int rc = 0, i, check_credits = 0; + struct osd_thandle *oh = container_of(thandle, + struct osd_thandle, ot_super); + unsigned int save_credits = oh->ot_credits; - LASSERT(inode); + LASSERT(inode); rc = osd_init_iobuf(osd, iobuf, 1, npages); if (unlikely(rc != 0)) RETURN(rc); - isize = i_size_read(inode); dquot_initialize(inode); - for (i = 0; i < npages; i++) { + for (i = 0; i < npages; i++) { if (lnb[i].lnb_rc == -ENOSPC && (lnb[i].lnb_flags & OBD_BRW_MAPPED)) { /* Allow the write to proceed if overwriting an - * existing block */ + * existing block + */ lnb[i].lnb_rc = 0; } @@ -1270,12 +1469,12 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, continue; } + if (!(lnb[i].lnb_flags & OBD_BRW_MAPPED)) + check_credits = 1; + 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; - /* * Since write and truncate are serialized by oo_sem, even * partial-page truncate should not leave dirty pages in the @@ -1286,39 +1485,35 @@ 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]); - } + } 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) { - 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; - } - - if (likely(rc == 0)) { - spin_lock(&inode->i_lock); - if (isize > i_size_read(inode)) { - i_size_write(inode, isize); - LDISKFS_I(inode)->i_disksize = isize; - spin_unlock(&inode->i_lock); - ll_dirty_inode(inode, I_DIRTY_DATASYNC); - } else { - spin_unlock(&inode->i_lock); - } + 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, osd, + 1, user_size, + check_credits); + /* + * Write might restart transaction, extend credits + * if needed for operations such as attribute set. + */ + if (rc == 0) { + handle_t *handle = ldiskfs_journal_current_handle(); - 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() */ + LASSERT(handle != NULL); + rc = osd_extend_restart_trans(handle, save_credits, + inode); + } } else { - osd_fini_iobuf(osd, iobuf); + /* no pages to write, no transno is needed */ + thandle->th_local = 1; } + if (rc != 0) + osd_fini_iobuf(osd, iobuf); + osd_trans_exec_check(env, thandle, OSD_OT_WRITE); if (unlikely(rc != 0)) { @@ -1326,9 +1521,11 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, 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); + } } } @@ -1336,18 +1533,18 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, } static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, - struct niobuf_local *lnb, int npages) + struct niobuf_local *lnb, int npages) { - 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)); + 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)); int rc = 0, i, cache_hits = 0, cache_misses = 0; ktime_t start, end; s64 timediff; loff_t isize; - LASSERT(inode); + LASSERT(inode); rc = osd_init_iobuf(osd, iobuf, 0, npages); if (unlikely(rc != 0)) @@ -1360,7 +1557,8 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, 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. */ + * lnb->lnb_rc == 0, so it's easy to detect later. + */ break; /* instead of looking if we go beyong isize, send complete @@ -1383,7 +1581,8 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, * 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 */ + * fine to have PG_locked and lnb_locked inconsistent here + */ lnb[i].lnb_locked = 0; } end = ktime_get(); @@ -1401,17 +1600,18 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, cache_hits + cache_misses); 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_ldiskfs_map_inode_pages(inode, iobuf, osd, 0, + 0, 0); + if (!rc) + rc = osd_do_bio(osd, inode, iobuf, 0, 0); /* IO stats will be done in osd_bufs_put() */ /* 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]); + if (!PagePrivate2(iobuf->dr_pages[i])) + unlock_page(iobuf->dr_pages[i]); } } @@ -1431,26 +1631,27 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, */ static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen) { - struct ldiskfs_inode_info *ei = LDISKFS_I(inode); + struct ldiskfs_inode_info *ei = LDISKFS_I(inode); - memcpy(buffer, (char *)ei->i_data, buflen); + memcpy(buffer, (char *)ei->i_data, buflen); - return buflen; + return buflen; } int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs) { - struct buffer_head *bh; - unsigned long block; - int osize; - int blocksize; - int csize; - int boffs; - - /* prevent reading after eof */ + struct buffer_head *bh; + unsigned long block; + int osize; + int blocksize; + int csize; + int boffs; + + /* prevent reading after eof */ spin_lock(&inode->i_lock); if (i_size_read(inode) < *offs + size) { loff_t diff = i_size_read(inode) - *offs; + spin_unlock(&inode->i_lock); if (diff < 0) { CDEBUG(D_OTHER, @@ -1466,17 +1667,16 @@ int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs) spin_unlock(&inode->i_lock); } - blocksize = 1 << inode->i_blkbits; - osize = size; - while (size > 0) { - block = *offs >> inode->i_blkbits; - boffs = *offs & (blocksize - 1); - csize = min(blocksize - boffs, size); + blocksize = 1 << inode->i_blkbits; + osize = size; + while (size > 0) { + block = *offs >> inode->i_blkbits; + boffs = *offs & (blocksize - 1); + 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", osd_ino2name(inode), - csize, *offs, inode->i_ino, + 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); } @@ -1488,29 +1688,37 @@ int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs) memset(buf, 0, csize); } - *offs += csize; - buf += csize; - size -= csize; - } - return osize; + *offs += csize; + buf += csize; + size -= csize; + } + return osize; } static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt, struct lu_buf *buf, loff_t *pos) { - struct inode *inode = osd_dt_obj(dt)->oo_inode; - int rc; - - /* 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); - - return rc; + struct inode *inode = osd_dt_obj(dt)->oo_inode; + int rc; + + /* 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); + + 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; } static inline int osd_extents_enabled(struct super_block *sb, @@ -1536,7 +1744,8 @@ int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode, /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself) * we do not expect blockmaps on the large files, - * so let's shrink it to 2 levels (4GB files) */ + * so let's shrink it to 2 levels (4GB files) + */ /* this is default reservation: 2 levels */ credits = (blocks + 2) * 3; @@ -1588,10 +1797,10 @@ 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; @@ -1599,7 +1808,8 @@ static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt, if (_pos == -1) { /* if this is an append, then we - * should expect cross-block record */ + * should expect cross-block record + */ pos = 0; } else { pos = _pos; @@ -1634,7 +1844,8 @@ static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt, depth = max(depth, 1) + 1; credits = depth; /* if not append, then split may need to modify - * existing blocks moving entries into the new ones */ + * existing blocks moving entries into the new ones + */ if (_pos != -1) credits += depth; /* blocks to store data: bitmap,gd,itself */ @@ -1643,7 +1854,8 @@ static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt, credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks); } /* if inode is created as part of the transaction, - * then it's counted already by the creation method */ + * then it's counted already by the creation method + */ if (inode != NULL) credits++; @@ -1653,7 +1865,8 @@ out: /* dt_declare_write() is usually called for system objects, such * as llog or last_rcvd files. We needn't enforce quota on those - * objects, so always set the lqi_space as 0. */ + * 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), @@ -1676,25 +1889,27 @@ static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen) LDISKFS_I(inode)->i_disksize = buflen; i_size_write(inode, buflen); spin_unlock(&inode->i_lock); - ll_dirty_inode(inode, I_DIRTY_DATASYNC); + osd_dirty_inode(inode, I_DIRTY_DATASYNC); 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 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; - int err = 0; - int size; - int boffs; - int dirty_inode = 0; + 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); - bool create, sparse; + int err = 0; + int size; + int boffs; + int dirty_inode = 0; + bool create, sparse, sync = false; if (write_NUL) { /* @@ -1706,13 +1921,15 @@ int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize, ++bufsize; } + 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; - bool sync; unsigned long last_block = (new_size == 0) ? 0 : (new_size - 1) >> inode->i_blkbits; @@ -1730,16 +1947,30 @@ int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize, 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, + CWARN( + "%s: adding bh without locking off %llu (block %lu, size %d, offs %llu)\n", + osd_ino2name(inode), offset, block, bufsize, *offs); if (IS_ERR_OR_NULL(bh)) { - bh = __ldiskfs_bread(handle, inode, block, 1); + 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) + if (sync) { up(&ei->i_append_sem); + sync = false; + } create = false; } if (IS_ERR_OR_NULL(bh)) { @@ -1750,40 +1981,45 @@ int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize, 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; - } - - err = ldiskfs_journal_get_write_access(handle, bh); - if (err) { - CERROR("journal_get_write_access() returned error %d\n", - err); - break; - } + CERROR( + "%s: error reading offset %llu (block %lu, size %d, offs %llu), credits %d/%d: rc = %d\n", + osd_ino2name(inode), offset, block, bufsize, + *offs, credits, handle->h_buffer_credits, err); + break; + } + + err = ldiskfs_journal_get_write_access(handle, bh); + if (err) { + CERROR("journal_get_write_access() returned error %d\n", + err); + break; + } LASSERTF(boffs + size <= bh->b_size, "boffs %d size %d bh->b_size %lu\n", boffs, size, (unsigned long)bh->b_size); if (create) { memset(bh->b_data, 0, bh->b_size); - if (sync) + 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; - - if (offset + size > new_size) - new_size = offset + size; - offset += size; - bufsize -= size; - buf += size; - } - if (bh) - brelse(bh); + if (err) + break; + + if (offset + size > new_size) + new_size = offset + size; + offset += size; + bufsize -= size; + buf += size; + } + if (sync) + up(&ei->i_append_sem); + + if (bh) + brelse(bh); if (write_NUL) --new_size; @@ -1792,18 +2028,18 @@ int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize, 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; - return err; + if (err == 0) + *offs = offset; + return err; } static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt, @@ -1815,17 +2051,17 @@ static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt, ssize_t result; int is_link; - LASSERT(dt_object_exists(dt)); + LASSERT(dt_object_exists(dt)); - LASSERT(handle != NULL); + LASSERT(handle != NULL); LASSERT(inode != NULL); dquot_initialize(inode); - /* XXX: don't check: one declared chunk can be used many times */ + /* 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); + 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 @@ -1837,9 +2073,8 @@ 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; @@ -1848,25 +2083,192 @@ static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt, return result; } +static int osd_declare_fallocate(const struct lu_env *env, + struct dt_object *dt, __u64 start, __u64 end, + int mode, struct thandle *th) +{ + struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super); + struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt)); + struct inode *inode = osd_dt_obj(dt)->oo_inode; + long long quota_space = 0; + /* 5 is max tree depth. (inode + 4 index blocks) */ + int depth = 5; + int rc; + + 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(th); + LASSERT(inode); + + /* quota space for metadata blocks + * approximate metadata estimate should be good enough. + */ + quota_space += PAGE_SIZE; + quota_space += depth * LDISKFS_BLOCK_SIZE(osd_sb(osd)); + + /* quota space should be reported in 1K blocks */ + quota_space = toqb(quota_space) + toqb(end - start) + + LDISKFS_META_TRANS_BLOCKS(inode->i_sb); + + /* We don't need to reserve credits for whole fallocate here. + * We reserve space only for metadata. Fallocate credits are + * extended as required + */ + 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), 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_thandle *oh = container_of(th, struct osd_thandle, ot_super); + handle_t *handle = ldiskfs_journal_current_handle(); + unsigned int save_credits = oh->ot_credits; + struct osd_object *obj = osd_dt_obj(dt); + struct inode *inode = obj->oo_inode; + struct ldiskfs_map_blocks map; + unsigned int credits; + ldiskfs_lblk_t blen; + ldiskfs_lblk_t boff; + loff_t new_size = 0; + int depth = 0; + int flags; + int rc = 0; + + ENTRY; + + LASSERT(dt_object_exists(dt)); + LASSERT(osd_invariant(obj)); + LASSERT(inode != NULL); + + CDEBUG(D_INODE, "fallocate: inode #%lu: start %llu end %llu mode %d\n", + inode->i_ino, start, end, mode); + + dquot_initialize(inode); + + LASSERT(th); + + boff = start >> inode->i_blkbits; + blen = (ALIGN(end, 1 << inode->i_blkbits) >> inode->i_blkbits) - boff; + + flags = LDISKFS_GET_BLOCKS_CREATE; + if (mode & FALLOC_FL_KEEP_SIZE) + flags |= LDISKFS_GET_BLOCKS_KEEP_SIZE; + + inode_lock(inode); + + /* + * We only support preallocation for extent-based file only. + */ + if (!(ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))) + GOTO(out, rc = -EOPNOTSUPP); + + if (!(mode & FALLOC_FL_KEEP_SIZE) && (end > i_size_read(inode) || + end > LDISKFS_I(inode)->i_disksize)) { + new_size = end; + rc = inode_newsize_ok(inode, new_size); + if (rc) + GOTO(out, rc); + } + + inode_dio_wait(inode); + + map.m_lblk = boff; + map.m_len = blen; + + /* Don't normalize the request if it can fit in one extent so + * that it doesn't get unnecessarily split into multiple extents. + */ + if (blen <= EXT_UNWRITTEN_MAX_LEN) + flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE; + + /* + * credits to insert 1 extent into extent tree. + */ + credits = osd_chunk_trans_blocks(inode, blen); + depth = ext_depth(inode); + + while (rc >= 0 && blen) { + loff_t epos; + + /* + * Recalculate credits when extent tree depth changes. + */ + if (depth != ext_depth(inode)) { + credits = osd_chunk_trans_blocks(inode, blen); + depth = ext_depth(inode); + } + + /* TODO: quota check */ + rc = osd_extend_restart_trans(handle, credits, inode); + if (rc) + break; + + rc = ldiskfs_map_blocks(handle, inode, &map, flags); + if (rc <= 0) { + CDEBUG(D_INODE, + "inode #%lu: block %u: len %u: ldiskfs_map_blocks returned %d\n", + inode->i_ino, map.m_lblk, map.m_len, rc); + ldiskfs_mark_inode_dirty(handle, inode); + break; + } + + map.m_lblk += rc; + map.m_len = blen = blen - rc; + epos = (loff_t)map.m_lblk << inode->i_blkbits; + inode->i_ctime = current_time(inode); + if (new_size) { + if (epos > end) + epos = end; + if (ldiskfs_update_inode_size(inode, epos) & 0x1) + inode->i_mtime = inode->i_ctime; + } else { + if (epos > inode->i_size) + ldiskfs_set_inode_flag(inode, + LDISKFS_INODE_EOFBLOCKS); + } + + ldiskfs_mark_inode_dirty(handle, inode); + } + +out: + /* extand credits if needed for operations such as attribute set */ + if (rc >= 0) + rc = osd_extend_restart_trans(handle, save_credits, inode); + + inode_unlock(inode); + + RETURN(rc); +} + static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt, - __u64 start, __u64 end, struct thandle *th) + __u64 start, __u64 end, struct thandle *th) { - struct osd_thandle *oh; + struct osd_thandle *oh; struct inode *inode; int rc; - ENTRY; - - LASSERT(th); - oh = container_of(th, struct osd_thandle, ot_super); - - /* - * we don't need to reserve credits for whole truncate - * it's not possible as truncate may need to free too many - * blocks and that won't fit a single transaction. instead - * we reserve credits to change i_size and put inode onto - * orphan list. if needed truncate will extend or restart - * transaction - */ + ENTRY; + + LASSERT(th); + oh = container_of(th, struct osd_thandle, ot_super); + + /* + * we don't need to reserve credits for whole truncate + * it's not possible as truncate may need to free too many + * blocks and that won't fit a single transaction. instead + * we reserve credits to change i_size and put inode onto + * orphan list. if needed truncate will extend or restart + * transaction + */ osd_trans_declare_op(env, oh, OSD_OT_PUNCH, osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3); @@ -1895,7 +2297,6 @@ static int osd_punch(const struct lu_env *env, struct dt_object *dt, bool grow = false; ENTRY; - LASSERT(end == OBD_OBJECT_EOF); LASSERT(dt_object_exists(dt)); LASSERT(osd_invariant(obj)); LASSERT(inode != NULL); @@ -1911,7 +2312,8 @@ static int osd_punch(const struct lu_env *env, struct dt_object *dt, * 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 */ + * to sent MDS_REINT (LU-11033) -bzzz + */ if (osd->od_is_ost && i_size_read(inode) == start) RETURN(0); @@ -1922,6 +2324,14 @@ static int osd_punch(const struct lu_env *env, struct dt_object *dt, grow = true; i_size_write(inode, start); spin_unlock(&inode->i_lock); + /* if object holds encrypted content, we need to make sure we truncate + * on an encryption unit boundary, or subsequent reads will get + * corrupted content + */ + if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL && + start & ~LUSTRE_ENCRYPTION_MASK) + start = (start & LUSTRE_ENCRYPTION_MASK) + + LUSTRE_ENCRYPTION_UNIT_SIZE; ll_truncate_pagecache(inode, start); /* optimize grow case */ @@ -1930,10 +2340,13 @@ static int osd_punch(const struct lu_env *env, struct dt_object *dt, 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 */ + * 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); @@ -2046,12 +2459,43 @@ static int osd_ladvise(const struct lu_env *env, struct dt_object *dt, RETURN(rc); } +static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt, + loff_t offset, int whence) +{ + struct osd_object *obj = osd_dt_obj(dt); + struct inode *inode = obj->oo_inode; + struct file *file; + loff_t result; + + ENTRY; + + LASSERT(dt_object_exists(dt)); + LASSERT(osd_invariant(obj)); + LASSERT(inode); + LASSERT(offset >= 0); + + file = osd_quasi_file(env, inode); + result = file->f_op->llseek(file, offset, whence); + + /* + * If 'offset' is beyond end of object file then treat it as not error + * but valid case for SEEK_HOLE and return 'offset' as result. + * LOV will decide if it is beyond real end of file or not. + */ + if (whence == SEEK_HOLE && result == -ENXIO) + result = offset; + + CDEBUG(D_INFO, "seek %s from %lld: %lld\n", whence == SEEK_HOLE ? + "hole" : "data", offset, result); + RETURN(result); +} + /* * in some cases we may need declare methods for objects being created * e.g., when we create symlink */ const struct dt_body_operations osd_body_ops_new = { - .dbo_declare_write = osd_declare_write, + .dbo_declare_write = osd_declare_write, }; const struct dt_body_operations osd_body_ops = { @@ -2068,6 +2512,9 @@ const struct dt_body_operations osd_body_ops = { .dbo_punch = osd_punch, .dbo_fiemap_get = osd_fiemap_get, .dbo_ladvise = osd_ladvise, + .dbo_declare_fallocate = osd_declare_fallocate, + .dbo_fallocate = osd_fallocate, + .dbo_lseek = osd_lseek, }; /** @@ -2112,20 +2559,23 @@ int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared) 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(struct list_head *list) +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); } @@ -2148,13 +2598,29 @@ void osd_execute_truncate(struct osd_object *obj) return; } + size = i_size_read(inode); + inode_lock(inode); + /* if object holds encrypted content, we need to make sure we truncate + * on an encryption unit boundary, or block content will get corrupted + */ + if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL && + size & ~LUSTRE_ENCRYPTION_MASK) + inode->i_size = (size & LUSTRE_ENCRYPTION_MASK) + + LUSTRE_ENCRYPTION_UNIT_SIZE; ldiskfs_truncate(inode); + inode_unlock(inode); + if (inode->i_size != size) { + spin_lock(&inode->i_lock); + i_size_write(inode, size); + LDISKFS_I(inode)->i_disksize = size; + spin_unlock(&inode->i_lock); + osd_dirty_inode(inode, I_DIRTY_DATASYNC); + } /* * 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)) {