X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fosd-zfs%2Fosd_io.c;h=a74e58c3222da6d749c9a45ab4a905585d7c7d85;hb=547e4021afb499f05cfbbf23e05b124587ca2ea8;hp=a253a9c16a1d454ef03b33a645411fcbfc0a58c4;hpb=72057a3af19ee02d9a686bd7e7d074917e381310;p=fs%2Flustre-release.git diff --git a/lustre/osd-zfs/osd_io.c b/lustre/osd-zfs/osd_io.c index a253a9c..a74e58c 100644 --- a/lustre/osd-zfs/osd_io.c +++ b/lustre/osd-zfs/osd_io.c @@ -23,7 +23,7 @@ * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2012, 2015, Intel Corporation. + * Copyright (c) 2012, 2016, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -62,7 +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) @@ -118,7 +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(); @@ -135,8 +135,8 @@ 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, - buf->lb_buf, DMU_READ_PREFETCH); + rc = osd_dmu_read(osd, obj->oo_dn, *pos, size, buf->lb_buf, + DMU_READ_PREFETCH); record_end_io(osd, READ, cfs_time_current() - start, size, size >> PAGE_SHIFT); @@ -166,7 +166,7 @@ static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt, /* size change (in dnode) will be declared by dmu_tx_hold_write() */ if (dt_object_exists(dt)) - oid = obj->oo_db->db_object; + oid = obj->oo_dn->dn_object; else oid = DMU_NEW_OBJECT; @@ -176,7 +176,7 @@ static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt, if (pos == -1) pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE, obj->oo_attr.la_size + (2 << 20)); - dmu_tx_hold_write(oh->ot_tx, oid, pos, buf->lb_len); + osd_tx_hold_write(oh->ot_tx, oid, obj->oo_dn, pos, buf->lb_len); /* dt_declare_write() is usually called for system objects, such * as llog or last_rcvd files. We needn't enforce quota on those @@ -199,15 +199,13 @@ 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, - (uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx); + osd_dmu_write(osd, obj->oo_dn, 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) { obj->oo_attr.la_size = offset + buf->lb_len; @@ -227,9 +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_SHIFT); - RETURN(rc); } @@ -252,7 +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) @@ -267,7 +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); @@ -283,6 +278,7 @@ static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt, static inline struct page *kmem_to_page(void *addr) { + LASSERT(!((unsigned long)addr & ~PAGE_MASK)); if (is_vmalloc_addr(addr)) return vmalloc_to_page(addr); else @@ -328,9 +324,9 @@ 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); @@ -379,7 +375,7 @@ 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, @@ -393,19 +389,40 @@ err: RETURN(rc); } +static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs) +{ + arc_buf_t *abuf; + + abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs); + if (unlikely(!abuf)) + return ERR_PTR(-ENOMEM); + +#if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0) + /** + * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio + * blocks smaller than (PAGE_SIZE << 2). This poses a problem of + * setting up page array for RDMA transfer. See LU-9305. + */ + if ((unsigned long)abuf->b_data & ~PAGE_MASK) { + dmu_return_arcbuf(abuf); + return NULL; + } +#endif + + return abuf; +} + static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj, loff_t off, ssize_t len, struct niobuf_local *lnb) { 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 @@ -416,13 +433,15 @@ static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj, off_in_block = off & (bs - 1); sz_in_block = min_t(int, bs - off_in_block, len); + abuf = NULL; if (sz_in_block == bs) { /* full block, try to use zerocopy */ + abuf = osd_request_arcbuf(dn, bs); + if (unlikely(IS_ERR(abuf))) + GOTO(out_err, rc = PTR_ERR(abuf)); + } - abuf = dmu_request_arcbuf(obj->oo_db, bs); - if (unlikely(abuf == NULL)) - GOTO(out_err, rc = -ENOMEM); - + if (abuf != NULL) { atomic_inc(&osd->od_zerocopy_loan); /* go over pages arcbuf contains, put them as @@ -505,7 +524,7 @@ 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); @@ -521,22 +540,11 @@ 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; } -static inline uint32_t osd_get_blocksz(struct osd_object *obj) -{ - uint32_t blksz; - u_longlong_t unused; - - LASSERT(obj->oo_db); - - dmu_object_size_from_db(obj->oo_db, &blksz, &unused); - return blksz; -} - static inline uint64_t osd_roundup2blocksz(uint64_t size, uint64_t offset, uint32_t blksz) @@ -545,7 +553,7 @@ static inline uint64_t osd_roundup2blocksz(uint64_t size, size += offset % blksz; - if (likely(IS_PO2(blksz))) + if (likely(is_power_of_2(blksz))) return PO2_ROUNDUP_TYPED(size, blksz, uint64_t); size += blksz - 1; @@ -563,7 +571,7 @@ static int osd_declare_write_commit(const struct lu_env *env, struct osd_thandle *oh; uint64_t offset = 0; uint32_t size = 0; - uint32_t blksz = osd_get_blocksz(obj); + uint32_t blksz = obj->oo_dn->dn_datablksz; int i, rc, flags = 0; bool ignore_quota = false, synced = false; long long space = 0; @@ -572,7 +580,7 @@ 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); @@ -614,8 +622,8 @@ static int osd_declare_write_commit(const struct lu_env *env, continue; } - dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, - offset, size); + osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, + obj->oo_dn, offset, size); /* Estimating space to be consumed by a write is rather * complicated with ZFS. As a consequence, we don't account for * indirect blocks and just use as a rough estimate the worse @@ -629,13 +637,11 @@ static int osd_declare_write_commit(const struct lu_env *env, } if (size) { - dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, + osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn, offset, size); space += osd_roundup2blocksz(size, offset, blksz); } - dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0); - oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */ /* backend zfs filesystem might be configured to store multiple data @@ -681,16 +687,12 @@ 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); - dmu_buf_impl_t *db = (dmu_buf_impl_t *)obj->oo_db; - dnode_t *dn; + dnode_t *dn = obj->oo_dn; uint32_t blksz; int rc = 0; ENTRY; - DB_DNODE_ENTER(db); - dn = DB_DNODE(db); - if (dn->dn_maxblkid > 0) /* can't change block size */ GOTO(out, rc); @@ -729,7 +731,6 @@ static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh, out_unlock: up_write(&obj->oo_guard); out: - DB_DNODE_EXIT(db); return rc; } @@ -746,7 +747,7 @@ 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); @@ -755,6 +756,32 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, (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, @@ -771,9 +798,9 @@ 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, - lnb[i].lnb_file_offset, lnb[i].lnb_len, - kmap(lnb[i].lnb_page), oh->ot_tx); + osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset, + lnb[i].lnb_len, kmap(lnb[i].lnb_page), + oh->ot_tx); kunmap(lnb[i].lnb_page); } else if (lnb[i].lnb_data) { LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0); @@ -781,7 +808,8 @@ 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! */ @@ -793,6 +821,7 @@ 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 */ @@ -829,7 +858,7 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt, 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; @@ -871,7 +900,7 @@ 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; @@ -885,8 +914,9 @@ 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; } @@ -915,7 +945,7 @@ 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) { @@ -948,14 +978,11 @@ 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); } - /* ... and we'll modify size attribute */ - dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0); - RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid, obj->oo_attr.la_gid, 0, oh, true, NULL, false));