X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosd-zfs%2Fosd_io.c;h=1d3b23032f7bb5f0dc61a364cd23ca58367124a6;hp=9b650f5423988b8992803af28addfe8e9808fe03;hb=fc117cc92d224466c1ec45de422f9f6624bbb1da;hpb=dc689955366c895f9cdcc86d78f4221866fe0926 diff --git a/lustre/osd-zfs/osd_io.c b/lustre/osd-zfs/osd_io.c index 9b650f5..1d3b230 100644 --- a/lustre/osd-zfs/osd_io.c +++ b/lustre/osd-zfs/osd_io.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2012, 2014, Intel Corporation. + * Copyright (c) 2012, 2016, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -66,7 +62,7 @@ #include #include -static char *osd_zerocopy_tag = "zerocopy"; +static char *osd_0copy_tag = "zerocopy"; static void record_start_io(struct osd_device *osd, int rw, int discont_pages) @@ -122,7 +118,7 @@ static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt, unsigned long start; LASSERT(dt_object_exists(dt)); - LASSERT(obj->oo_db); + LASSERT(obj->oo_dn); start = cfs_time_current(); @@ -139,11 +135,11 @@ static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt, record_start_io(osd, READ, 0); - rc = -dmu_read(osd->od_os, obj->oo_db->db_object, *pos, size, + rc = -dmu_read(osd->od_os, obj->oo_dn->dn_object, *pos, size, buf->lb_buf, DMU_READ_PREFETCH); record_end_io(osd, READ, cfs_time_current() - start, size, - size >> PAGE_CACHE_SHIFT); + size >> PAGE_SHIFT); if (rc == 0) { rc = size; *pos += size; @@ -168,18 +164,11 @@ static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt, * LOHA_EXISTs is supposed to be the last step in the * initialization */ - /* declare possible size change. notice we can't check - * current size here as another thread can change it */ - - if (dt_object_exists(dt)) { - LASSERT(obj->oo_db); - oid = obj->oo_db->db_object; - - dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0); - } else { + /* size change (in dnode) will be declared by dmu_tx_hold_write() */ + if (dt_object_exists(dt)) + oid = obj->oo_dn->dn_object; + else oid = DMU_NEW_OBJECT; - dmu_tx_hold_sa_create(oh->ot_tx, ZFS_SA_BASE_ATTR_SIZE); - } /* XXX: we still miss for append declaration support in ZFS * -1 means append which is used by llog mostly, llog @@ -210,14 +199,12 @@ static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt, ENTRY; LASSERT(dt_object_exists(dt)); - LASSERT(obj->oo_db); + LASSERT(obj->oo_dn); LASSERT(th != NULL); oh = container_of0(th, struct osd_thandle, ot_super); - record_start_io(osd, WRITE, 0); - - dmu_write(osd->od_os, obj->oo_db->db_object, offset, + dmu_write(osd->od_os, obj->oo_dn->dn_object, offset, (uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx); write_lock(&obj->oo_attr_lock); if (obj->oo_attr.la_size < offset + buf->lb_len) { @@ -238,9 +225,6 @@ static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt, rc = buf->lb_len; out: - record_end_io(osd, WRITE, 0, buf->lb_len, - buf->lb_len >> PAGE_CACHE_SHIFT); - RETURN(rc); } @@ -263,7 +247,7 @@ static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt, int i; LASSERT(dt_object_exists(dt)); - LASSERT(obj->oo_db); + LASSERT(obj->oo_dn); for (i = 0; i < npages; i++) { if (lnb[i].lnb_page == NULL) @@ -278,7 +262,7 @@ static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt, ptr = (unsigned long)lnb[i].lnb_data; if (ptr & 1UL) { ptr &= ~1UL; - dmu_buf_rele((void *)ptr, osd_zerocopy_tag); + dmu_buf_rele((void *)ptr, osd_0copy_tag); atomic_dec(&osd->od_zerocopy_pin); } else if (lnb[i].lnb_data != NULL) { dmu_return_arcbuf(lnb[i].lnb_data); @@ -339,9 +323,9 @@ static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj, * can get own replacement for dmu_buf_hold_array_by_bonus(). */ while (len > 0) { - rc = -dmu_buf_hold_array_by_bonus(obj->oo_db, off, len, TRUE, - osd_zerocopy_tag, &numbufs, - &dbp); + rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db, + off, len, TRUE, osd_0copy_tag, + &numbufs, &dbp); if (unlikely(rc)) GOTO(err, rc); @@ -361,8 +345,8 @@ static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj, dbf = (void *) ((unsigned long)dbp[i] | 1); while (tocpy > 0) { - thispage = PAGE_CACHE_SIZE; - thispage -= bufoff & (PAGE_CACHE_SIZE - 1); + thispage = PAGE_SIZE; + thispage -= bufoff & (PAGE_SIZE - 1); thispage = min(tocpy, thispage); lnb->lnb_rc = 0; @@ -390,7 +374,7 @@ static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj, dbp[i] = NULL; } - dmu_buf_rele_array(dbp, numbufs, osd_zerocopy_tag); + dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag); } record_end_io(osd, READ, cfs_time_current() - start, @@ -410,13 +394,11 @@ static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj, struct osd_device *osd = osd_obj2dev(obj); int plen, off_in_block, sz_in_block; int rc, i = 0, npages = 0; - arc_buf_t *abuf; - uint32_t bs; - uint64_t dummy; + dnode_t *dn = obj->oo_dn; + arc_buf_t *abuf; + uint32_t bs = dn->dn_datablksz; ENTRY; - dmu_object_size_from_db(obj->oo_db, &bs, &dummy); - /* * currently only full blocks are subject to zerocopy approach: * so that we're sure nobody is trying to update the same block @@ -430,7 +412,7 @@ static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj, if (sz_in_block == bs) { /* full block, try to use zerocopy */ - abuf = dmu_request_arcbuf(obj->oo_db, bs); + abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs); if (unlikely(abuf == NULL)) GOTO(out_err, rc = -ENOMEM); @@ -439,7 +421,7 @@ static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj, /* go over pages arcbuf contains, put them as * local niobufs for ptlrpc's bulks */ while (sz_in_block > 0) { - plen = min_t(int, sz_in_block, PAGE_CACHE_SIZE); + plen = min_t(int, sz_in_block, PAGE_SIZE); lnb[i].lnb_file_offset = off; lnb[i].lnb_page_offset = 0; @@ -473,7 +455,7 @@ static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj, /* can't use zerocopy, allocate temp. buffers */ while (sz_in_block > 0) { - plen = min_t(int, sz_in_block, PAGE_CACHE_SIZE); + plen = min_t(int, sz_in_block, PAGE_SIZE); lnb[i].lnb_file_offset = off; lnb[i].lnb_page_offset = 0; @@ -516,7 +498,7 @@ static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt, int rc; LASSERT(dt_object_exists(dt)); - LASSERT(obj->oo_db); + LASSERT(obj->oo_dn); if (rw == 0) rc = osd_bufs_get_read(env, obj, offset, len, lnb); @@ -532,89 +514,38 @@ static int osd_write_prep(const struct lu_env *env, struct dt_object *dt, struct osd_object *obj = osd_dt_obj(dt); LASSERT(dt_object_exists(dt)); - LASSERT(obj->oo_db); + LASSERT(obj->oo_dn); return 0; } -/* Return number of blocks that aren't mapped in the [start, start + size] - * region */ -static int osd_count_not_mapped(struct osd_object *obj, uint64_t start, - uint32_t size) +static inline uint64_t osd_roundup2blocksz(uint64_t size, + uint64_t offset, + uint32_t blksz) { - dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)obj->oo_db; - dmu_buf_impl_t *db; - dnode_t *dn; - uint32_t blkshift; - uint64_t end, blkid; - int rc; - ENTRY; - - DB_DNODE_ENTER(dbi); - dn = DB_DNODE(dbi); - - if (dn->dn_maxblkid == 0) { - if (start + size <= dn->dn_datablksz) - GOTO(out, size = 0); - if (start < dn->dn_datablksz) - start = dn->dn_datablksz; - /* assume largest block size */ - blkshift = osd_spa_maxblockshift( - dmu_objset_spa(osd_obj2dev(obj)->od_os)); - } else { - /* blocksize can't change */ - blkshift = dn->dn_datablkshift; - } - - /* compute address of last block */ - end = (start + size - 1) >> blkshift; - /* align start on block boundaries */ - start >>= blkshift; - - /* size is null, can't be mapped */ - if (obj->oo_attr.la_size == 0 || dn->dn_maxblkid == 0) - GOTO(out, size = (end - start + 1) << blkshift); + LASSERT(blksz > 0); - /* beyond EOF, can't be mapped */ - if (start > dn->dn_maxblkid) - GOTO(out, size = (end - start + 1) << blkshift); - - size = 0; - for (blkid = start; blkid <= end; blkid++) { - if (blkid == dn->dn_maxblkid) - /* this one is mapped for sure */ - continue; - if (blkid > dn->dn_maxblkid) { - size += (end - blkid + 1) << blkshift; - GOTO(out, size); - } + size += offset % blksz; - rc = dbuf_hold_impl(dn, 0, blkid, TRUE, FTAG, &db); - if (rc) { - /* for ENOENT (block not mapped) and any other errors, - * assume the block isn't mapped */ - size += 1 << blkshift; - continue; - } - dbuf_rele(db, FTAG); - } + if (likely(is_power_of_2(blksz))) + return PO2_ROUNDUP_TYPED(size, blksz, uint64_t); - GOTO(out, size); -out: - DB_DNODE_EXIT(dbi); - return size; + size += blksz - 1; + do_div(size, blksz); + return size * blksz; } static int osd_declare_write_commit(const struct lu_env *env, - struct dt_object *dt, - struct niobuf_local *lnb, int npages, - struct thandle *th) + struct dt_object *dt, + struct niobuf_local *lnb, int npages, + struct thandle *th) { struct osd_object *obj = osd_dt_obj(dt); struct osd_device *osd = osd_obj2dev(obj); struct osd_thandle *oh; uint64_t offset = 0; uint32_t size = 0; + uint32_t blksz = obj->oo_dn->dn_datablksz; int i, rc, flags = 0; bool ignore_quota = false, synced = false; long long space = 0; @@ -623,7 +554,7 @@ static int osd_declare_write_commit(const struct lu_env *env, ENTRY; LASSERT(dt_object_exists(dt)); - LASSERT(obj->oo_db); + LASSERT(obj->oo_dn); LASSERT(lnb); LASSERT(npages > 0); @@ -665,22 +596,24 @@ static int osd_declare_write_commit(const struct lu_env *env, continue; } - dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, + dmu_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, offset, size); - /* estimating space that will be consumed by a write is rather + /* Estimating space to be consumed by a write is rather * complicated with ZFS. As a consequence, we don't account for - * indirect blocks and quota overrun will be adjusted once the - * operation is committed, if required. */ - space += osd_count_not_mapped(obj, offset, size); + * indirect blocks and just use as a rough estimate the worse + * case where the old space is being held by a snapshot. Quota + * overrun will be adjusted once the operation is committed, if + * required. */ + space += osd_roundup2blocksz(size, offset, blksz); offset = lnb[i].lnb_file_offset; size = lnb[i].lnb_len; } if (size) { - dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, + dmu_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, offset, size); - space += osd_count_not_mapped(obj, offset, size); + space += osd_roundup2blocksz(size, offset, blksz); } dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0); @@ -691,8 +624,8 @@ static int osd_declare_write_commit(const struct lu_env *env, * copies */ space *= osd->od_os->os_copies; space = toqb(space); - CDEBUG(D_QUOTA, "writting %d pages, reserving "LPD64"K of quota " - "space\n", npages, space); + CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n", + npages, space); record_start_io(osd, WRITE, discont_pages); retry: @@ -720,6 +653,63 @@ retry: RETURN(rc); } +/** + * Policy to grow ZFS block size by write pattern. + * For sequential write, it grows block size gradually until it reaches the + * maximum blocksize the dataset can support. Otherwise, it will pick a + * a block size by the writing region of this I/O. + */ +static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh, + uint64_t start, uint64_t end) +{ + struct osd_device *osd = osd_obj2dev(obj); + dnode_t *dn = obj->oo_dn; + uint32_t blksz; + int rc = 0; + + ENTRY; + + if (dn->dn_maxblkid > 0) /* can't change block size */ + GOTO(out, rc); + + if (dn->dn_datablksz >= osd->od_max_blksz) + GOTO(out, rc); + + down_write(&obj->oo_guard); + + blksz = dn->dn_datablksz; + if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */ + GOTO(out_unlock, rc); + + /* now ZFS can support up to 16MB block size, and if the write + * is sequential, it just increases the block size gradually */ + if (start <= blksz) { /* sequential */ + blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end); + } else { /* sparse, pick a block size by write region */ + blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, + end - start); + } + + if (!is_power_of_2(blksz)) + blksz = size_roundup_power2(blksz); + + if (blksz > dn->dn_datablksz) { + rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object, + blksz, 0, oh->ot_tx); + LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz)); + if (rc < 0) + CDEBUG(D_INODE, "object "DFID": change block size" + "%u -> %u error rc = %d\n", + PFID(lu_object_fid(&obj->oo_dt.do_lu)), + dn->dn_datablksz, blksz, rc); + } + EXIT; +out_unlock: + up_write(&obj->oo_guard); +out: + return rc; +} + static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, struct niobuf_local *lnb, int npages, struct thandle *th) @@ -733,11 +723,41 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, ENTRY; LASSERT(dt_object_exists(dt)); - LASSERT(obj->oo_db); + LASSERT(obj->oo_dn); LASSERT(th != NULL); oh = container_of0(th, struct osd_thandle, ot_super); + /* adjust block size. Assume the buffers are sorted. */ + (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset, + lnb[npages - 1].lnb_file_offset + + lnb[npages - 1].lnb_len); + + /* LU-8791: take oo_guard to avoid the deadlock that changing block + * size and assigning arcbuf take place at the same time. + * + * Thread 1: + * osd_write_commit() + * -> osd_grow_blocksize() with osd_object::oo_guard held + * -> dmu_object_set_blocksize() + * -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock + * write lock held + * -> dbuf_new_size() + * -> dmu_buf_will_dirty() + * -> dbuf_read() + * -> wait for the dbuf state to change + * Thread 2: + * osd_write_commit() + * -> dmu_assign_arcbuf() + * -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL + * -> dbuf_dirty() + * -> try to hold the read lock of dnode_t::dn_struct_rwlock + * + * By taking the read lock, it can avoid thread 2 to enter into the + * critical section of assigning the arcbuf, while thread 1 is + * changing the block size. + */ + down_read(&obj->oo_guard); for (i = 0; i < npages; i++) { CDEBUG(D_INODE, "write %u bytes at %u\n", (unsigned) lnb[i].lnb_len, @@ -754,7 +774,7 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, } if (lnb[i].lnb_page->mapping == (void *)obj) { - dmu_write(osd->od_os, obj->oo_db->db_object, + dmu_write(osd->od_os, obj->oo_dn->dn_object, lnb[i].lnb_file_offset, lnb[i].lnb_len, kmap(lnb[i].lnb_page), oh->ot_tx); kunmap(lnb[i].lnb_page); @@ -764,7 +784,8 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, * notice that dmu_assign_arcbuf() is smart * enough to recognize changed blocksize * in this case it fallbacks to dmu_write() */ - dmu_assign_arcbuf(obj->oo_db, lnb[i].lnb_file_offset, + dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db, + lnb[i].lnb_file_offset, lnb[i].lnb_data, oh->ot_tx); /* drop the reference, otherwise osd_put_bufs() * will be releasing it - bad! */ @@ -776,6 +797,7 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len; iosize += lnb[i].lnb_len; } + up_read(&obj->oo_guard); if (unlikely(new_size == 0)) { /* no pages to write, no transno is needed */ @@ -809,11 +831,10 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, { struct osd_object *obj = osd_dt_obj(dt); int i; - unsigned long size = 0; loff_t eof; LASSERT(dt_object_exists(dt)); - LASSERT(obj->oo_db); + LASSERT(obj->oo_dn); read_lock(&obj->oo_attr_lock); eof = obj->oo_attr.la_size; @@ -824,12 +845,12 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, continue; lnb[i].lnb_rc = lnb[i].lnb_len; - size += lnb[i].lnb_rc; - if (lnb[i].lnb_file_offset + lnb[i].lnb_len > eof) { - lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset; - if (lnb[i].lnb_rc < 0) + if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) { + if (eof <= lnb[i].lnb_file_offset) lnb[i].lnb_rc = 0; + else + lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset; /* all subsequent rc should be 0 */ while (++i < npages) @@ -855,7 +876,7 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free() * called and then assigned to a transaction group. */ -static int __osd_object_punch(objset_t *os, dmu_buf_t *db, dmu_tx_t *tx, +static int __osd_object_punch(objset_t *os, dnode_t *dn, dmu_tx_t *tx, uint64_t size, uint64_t off, uint64_t len) { int rc = 0; @@ -869,8 +890,9 @@ static int __osd_object_punch(objset_t *os, dmu_buf_t *db, dmu_tx_t *tx, if (len == DMU_OBJECT_END && size == off) return 0; + /* XXX: dnode_free_range() can be used to save on dnode lookup */ if (off < size) - rc = -dmu_free_range(os, db->db_object, off, len, tx); + dmu_free_range(os, dn->dn_object, off, len, tx); return rc; } @@ -899,7 +921,7 @@ static int osd_punch(const struct lu_env *env, struct dt_object *dt, len = end - start; write_unlock(&obj->oo_attr_lock); - rc = __osd_object_punch(osd->od_os, obj->oo_db, oh->ot_tx, + rc = __osd_object_punch(osd->od_os, obj->oo_dn, oh->ot_tx, obj->oo_attr.la_size, start, len); /* set new size */ if (len == DMU_OBJECT_END) { @@ -932,7 +954,7 @@ static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt, /* declare we'll free some blocks ... */ if (start < obj->oo_attr.la_size) { read_unlock(&obj->oo_attr_lock); - dmu_tx_hold_free(oh->ot_tx, obj->oo_db->db_object, start, len); + dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len); } else { read_unlock(&obj->oo_attr_lock); } @@ -945,6 +967,20 @@ static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt, false)); } +static int osd_ladvise(const struct lu_env *env, struct dt_object *dt, + __u64 start, __u64 end, enum lu_ladvise_type advice) +{ + int rc; + ENTRY; + + switch (advice) { + default: + rc = -ENOTSUPP; + break; + } + + RETURN(rc); +} struct dt_body_operations osd_body_ops = { .dbo_read = osd_read, @@ -958,4 +994,5 @@ struct dt_body_operations osd_body_ops = { .dbo_read_prep = osd_read_prep, .dbo_declare_punch = osd_declare_punch, .dbo_punch = osd_punch, + .dbo_ladvise = osd_ladvise, };