X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosd-zfs%2Fosd_handler.c;h=f4f3ea85a8ff7247cff819f37ae6ece8613c609e;hp=a847f40498c45f9e3c17fbe76813cd9ea5ff0f37;hb=d527e812461baf9db2f6ed960a3b6cc12d4ab37c;hpb=a6f415ad648f2da0708612beba51f34c2594f5ec diff --git a/lustre/osd-zfs/osd_handler.c b/lustre/osd-zfs/osd_handler.c index a847f40..f4f3ea8 100644 --- a/lustre/osd-zfs/osd_handler.c +++ b/lustre/osd-zfs/osd_handler.c @@ -15,26 +15,18 @@ * * 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 */ /* * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. - */ -/* - * Copyright (c) 2012, 2013, Intel Corporation. - * Use is subject to license terms. * + * Copyright (c) 2012, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ - * Lustre is a trademark of Sun Microsystems, Inc. * * lustre/osd-zfs/osd_handler.c * Top-level entry points into osd module @@ -46,7 +38,6 @@ #define DEBUG_SUBSYSTEM S_OSD -#include #include #include #include @@ -54,7 +45,7 @@ #include #include #include -#include +#include #include #include "osd_internal.h" @@ -74,9 +65,14 @@ struct lu_context_key osd_key; +static int osd_txg_sync_delay_us = -1; + /* Slab for OSD object allocation */ struct kmem_cache *osd_object_kmem; +/* Slab to allocate osd_zap_it */ +struct kmem_cache *osd_zapit_cachep; + static struct lu_kmem_descr osd_caches[] = { { .ckd_cache = &osd_object_kmem, @@ -84,6 +80,11 @@ static struct lu_kmem_descr osd_caches[] = { .ckd_size = sizeof(struct osd_object) }, { + .ckd_cache = &osd_zapit_cachep, + .ckd_name = "osd_zapit_cache", + .ckd_size = sizeof(struct osd_zap_it) + }, + { .ckd_cache = NULL } }; @@ -143,11 +144,14 @@ static void osd_trans_commit_cb(void *cb_data, int error) osd_dt_dev(th->th_dev)->od_svname, th, error); } - dt_txn_hook_commit(th); - /* call per-transaction callbacks if any */ - list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage) + list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage) { + LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC, + "commit callback entry: magic=%x name='%s'\n", + dcb->dcb_magic, dcb->dcb_name); + list_del_init(&dcb->dcb_linkage); dcb->dcb_func(NULL, th, dcb, error); + } /* Unlike ldiskfs, zfs updates space accounting at commit time. * As a consequence, op_end is called only now to inform the quota slave @@ -155,23 +159,28 @@ static void osd_trans_commit_cb(void *cb_data, int error) * should be released. Quota space won't be adjusted at this point since * we can't provide a suitable environment. It will be performed * asynchronously by a lquota thread. */ - qsd_op_end(NULL, osd->od_quota_slave, &oh->ot_quota_trans); + qsd_op_end(NULL, osd->od_quota_slave_dt, &oh->ot_quota_trans); + if (osd->od_quota_slave_md != NULL) + qsd_op_end(NULL, osd->od_quota_slave_md, &oh->ot_quota_trans); lu_device_put(lud); th->th_dev = NULL; - lu_context_exit(&th->th_ctx); - lu_context_fini(&th->th_ctx); - thandle_put(&oh->ot_super); + OBD_FREE_PTR(oh); EXIT; } static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb) { - struct osd_thandle *oh; + struct osd_thandle *oh = container_of(th, struct osd_thandle, + ot_super); - oh = container_of0(th, struct osd_thandle, ot_super); - list_add(&dcb->dcb_linkage, &oh->ot_dcb_list); + LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC); + LASSERT(&dcb->dcb_func != NULL); + if (dcb->dcb_flags & DCB_TRANS_STOP) + list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list); + else + list_add(&dcb->dcb_linkage, &oh->ot_dcb_list); return 0; } @@ -182,31 +191,34 @@ static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb) static int osd_trans_start(const struct lu_env *env, struct dt_device *d, struct thandle *th) { - struct osd_thandle *oh; - int rc; + struct osd_device *osd = osd_dt_dev(d); + struct osd_thandle *oh; + int rc; + ENTRY; - oh = container_of0(th, struct osd_thandle, ot_super); + oh = container_of(th, struct osd_thandle, ot_super); LASSERT(oh); LASSERT(oh->ot_tx); rc = dt_txn_hook_start(env, d, th); - if (rc != 0) + if (rc != 0) { + CERROR("%s: dt_txn_hook_start failed: rc = %d\n", + osd->od_svname, rc); RETURN(rc); + } - if (oh->ot_write_commit && OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) + if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START)) /* Unlike ldiskfs, ZFS checks for available space and returns * -ENOSPC when assigning txg */ - RETURN(-ENOSPC); + RETURN(-EIO); rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT); if (unlikely(rc != 0)) { - struct osd_device *osd = osd_dt_dev(d); /* dmu will call commit callback with error code during abort */ if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC) - CERROR("%s: failed to start transaction due to ENOSPC. " - "Metadata overhead is underestimated or " - "grant_ratio is too low.\n", osd->od_svname); + CERROR("%s: failed to start transaction due to ENOSPC" + "\n", osd->od_svname); else CERROR("%s: can't assign tx: rc = %d\n", osd->od_svname, rc); @@ -214,14 +226,48 @@ static int osd_trans_start(const struct lu_env *env, struct dt_device *d, /* add commit callback */ dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh); oh->ot_assigned = 1; - lu_context_init(&th->th_ctx, th->th_tags); - lu_context_enter(&th->th_ctx); + osd_oti_get(env)->oti_in_trans = 1; lu_device_get(&d->dd_lu_dev); } RETURN(rc); } +static void osd_unlinked_list_emptify(const struct lu_env *env, + struct osd_device *osd, + struct list_head *list, bool free) +{ + struct osd_object *obj; + uint64_t oid; + + while (!list_empty(list)) { + obj = list_entry(list->next, + struct osd_object, oo_unlinked_linkage); + LASSERT(obj->oo_dn != NULL); + oid = obj->oo_dn->dn_object; + + list_del_init(&obj->oo_unlinked_linkage); + if (free) + (void)osd_unlinked_object_free(env, osd, oid); + } +} + +static void osd_trans_stop_cb(struct osd_thandle *oth, int result) +{ + struct dt_txn_commit_cb *dcb; + struct dt_txn_commit_cb *tmp; + + /* call per-transaction stop callbacks if any */ + list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list, + dcb_linkage) { + LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC, + "commit callback entry: magic=%x name='%s'\n", + dcb->dcb_magic, dcb->dcb_name); + list_del_init(&dcb->dcb_linkage); + dcb->dcb_func(NULL, &oth->ot_super, dcb, result); + } +} + /* * Concurrency: shouldn't matter. */ @@ -229,47 +275,62 @@ static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt, struct thandle *th) { struct osd_device *osd = osd_dt_dev(th->th_dev); + bool sync = (th->th_sync != 0); struct osd_thandle *oh; + LIST_HEAD(unlinked); uint64_t txg; int rc; ENTRY; - oh = container_of0(th, struct osd_thandle, ot_super); + oh = container_of(th, struct osd_thandle, ot_super); + list_splice_init(&oh->ot_unlinked_list, &unlinked); + + osd_oti_get(env)->oti_ins_cache_depth--; + /* reset OI cache for safety */ + if (osd_oti_get(env)->oti_ins_cache_depth == 0) + osd_oti_get(env)->oti_ins_cache_used = 0; if (oh->ot_assigned == 0) { LASSERT(oh->ot_tx); + CDEBUG(D_OTHER, "%s: transaction is aborted\n", osd->od_svname); + osd_trans_stop_cb(oh, th->th_result); dmu_tx_abort(oh->ot_tx); - osd_object_sa_dirty_rele(oh); + osd_object_sa_dirty_rele(env, oh); + osd_unlinked_list_emptify(env, osd, &unlinked, false); /* there won't be any commit, release reserved quota space now, * if any */ - qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans); - thandle_put(&oh->ot_super); + qsd_op_end(env, osd->od_quota_slave_dt, &oh->ot_quota_trans); + if (osd->od_quota_slave_md != NULL) + qsd_op_end(env, osd->od_quota_slave_md, + &oh->ot_quota_trans); + OBD_FREE_PTR(oh); RETURN(0); } - /* When doing our own inode accounting, the ZAPs storing per-uid/gid - * usage are updated at operation execution time, so we should call - * qsd_op_end() straight away. Otherwise (for blk accounting maintained - * by ZFS and when #inode is estimated from #blks) accounting is updated - * at commit time and the call to qsd_op_end() must be delayed */ - if (oh->ot_quota_trans.lqt_id_cnt > 0 && - !oh->ot_quota_trans.lqt_ids[0].lqi_is_blk && - !osd->od_quota_iused_est) - qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans); - rc = dt_txn_hook_stop(env, th); if (rc != 0) CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n", osd->od_svname, rc); + osd_trans_stop_cb(oh, rc); + LASSERT(oh->ot_tx); txg = oh->ot_tx->tx_txg; - osd_object_sa_dirty_rele(oh); + osd_object_sa_dirty_rele(env, oh); + /* XXX: Once dmu_tx_commit() called, oh/th could have been freed + * by osd_trans_commit_cb already. */ dmu_tx_commit(oh->ot_tx); + osd_oti_get(env)->oti_in_trans = 0; - if (th->th_sync) - txg_wait_synced(dmu_objset_pool(osd->od_os), txg); + osd_unlinked_list_emptify(env, osd, &unlinked, true); + + if (sync) { + if (osd_txg_sync_delay_us < 0) + txg_wait_synced(dmu_objset_pool(osd->od_os), txg); + else + udelay(osd_txg_sync_delay_us); + } RETURN(rc); } @@ -283,6 +344,14 @@ static struct thandle *osd_trans_create(const struct lu_env *env, dmu_tx_t *tx; ENTRY; + if (dt->dd_rdonly) { + CERROR("%s: someone try to start transaction under " + "readonly mode, should be disabled.\n", + osd_name(osd_dt_dev(dt))); + dump_stack(); + RETURN(ERR_PTR(-EROFS)); + } + tx = dmu_tx_create(osd->od_os); if (tx == NULL) RETURN(ERR_PTR(-ENOMEM)); @@ -296,92 +365,121 @@ static struct thandle *osd_trans_create(const struct lu_env *env, oh->ot_tx = tx; INIT_LIST_HEAD(&oh->ot_dcb_list); + INIT_LIST_HEAD(&oh->ot_stop_dcb_list); + INIT_LIST_HEAD(&oh->ot_unlinked_list); INIT_LIST_HEAD(&oh->ot_sa_list); - sema_init(&oh->ot_sa_lock, 1); memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans)); th = &oh->ot_super; th->th_dev = dt; th->th_result = 0; - th->th_tags = LCT_TX_HANDLE; - atomic_set(&th->th_refc, 1); - th->th_alloc_size = sizeof(*oh); + + osd_oti_get(env)->oti_ins_cache_depth++; + RETURN(th); } -/* Estimate the number of objects from a number of blocks */ -uint64_t osd_objs_count_estimate(uint64_t refdbytes, uint64_t usedobjs, - uint64_t nrblocks) +/* Estimate the total number of objects from a number of blocks */ +uint64_t osd_objs_count_estimate(uint64_t usedbytes, uint64_t usedobjs, + uint64_t nrblocks, uint64_t est_maxblockshift) { - uint64_t est_objs, est_refdblocks, est_usedobjs; - - /* Compute an nrblocks estimate based on the actual number of - * dnodes that could fit in the space. Since we don't know the - * overhead associated with each dnode (xattrs, SAs, VDEV overhead, - * etc) just using DNODE_SHIFT isn't going to give a good estimate. - * Instead, compute an estimate based on the average space usage per - * dnode, with an upper and lower cap. + uint64_t est_totobjs, est_usedblocks, est_usedobjs; + + /* + * If blocksize is below 64KB (e.g. MDT with recordsize=4096) then + * bump the free dnode estimate to assume blocks at least 64KB in + * case of a directory-heavy MDT (at 32KB/directory). + */ + if (est_maxblockshift < 16) { + nrblocks >>= (16 - est_maxblockshift); + est_maxblockshift = 16; + } + + /* + * Estimate the total number of dnodes from the total blocks count + * and the space used per dnode. Since we don't know the overhead + * associated with each dnode (xattrs, SAs, VDEV overhead, etc.) + * just using DNODE_SHIFT isn't going to give a good estimate. + * Instead, compute the current average space usage per dnode, with + * an upper and lower cap to avoid unrealistic estimates.. * * In case there aren't many dnodes or blocks used yet, add a small - * correction factor using OSD_DNODE_EST_SHIFT. This correction - * factor gradually disappears as the number of real dnodes grows. - * This also avoids the need to check for divide-by-zero later. + * correction factor (OSD_DNODE_EST_{COUNT,BLKSHIFT}). This factor + * gradually disappears as the number of real dnodes grows. It also + * avoids the need to check for divide-by-zero computing dn_per_block. */ - CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0); - CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0); - - est_refdblocks = (refdbytes >> SPA_MAXBLOCKSHIFT) + - (OSD_DNODE_EST_COUNT >> OSD_DNODE_EST_BLKSHIFT); - est_usedobjs = usedobjs + OSD_DNODE_EST_COUNT; - - /* Average space/dnode more than maximum dnode size, use max dnode - * size to estimate free dnodes from adjusted free blocks count. - * OSTs typically use more than one block dnode so this case applies. */ - if (est_usedobjs <= est_refdblocks * 2) { - est_objs = nrblocks; - - /* Average space/dnode smaller than min dnode size (probably due to - * metadnode compression), use min dnode size to estimate the number of - * objects. - * An MDT typically uses below 512 bytes/dnode so this case applies. */ - } else if (est_usedobjs >= (est_refdblocks << OSD_DNODE_MIN_BLKSHIFT)) { - est_objs = nrblocks << OSD_DNODE_MIN_BLKSHIFT; - - /* Between the extremes, we try to use the average size of - * existing dnodes to compute the number of dnodes that fit - * into nrblocks: + BUILD_BUG_ON(OSD_DNODE_MIN_BLKSHIFT <= 0); + BUILD_BUG_ON(OSD_DNODE_EST_BLKSHIFT <= 0); + + est_usedblocks = ((OSD_DNODE_EST_COUNT << OSD_DNODE_EST_BLKSHIFT) + + usedbytes) >> est_maxblockshift; + est_usedobjs = OSD_DNODE_EST_COUNT + usedobjs; + + if (est_usedobjs <= est_usedblocks) { + /* + * Average space/dnode more than maximum block size, use max + * block size to estimate free dnodes from adjusted free blocks + * count. OSTs typically use multiple blocks per dnode so this + * case applies. + */ + est_totobjs = nrblocks; + + } else if (est_usedobjs >= (est_usedblocks << OSD_DNODE_MIN_BLKSHIFT)) { + /* + * Average space/dnode smaller than min dnode size (probably + * due to metadnode compression), use min dnode size to + * estimate object count. MDTs may use only one block per node + * so this case applies. + */ + est_totobjs = nrblocks << OSD_DNODE_MIN_BLKSHIFT; + + } else { + /* + * Between the extremes, use average space per existing dnode + * to compute the number of dnodes that will fit into nrblocks: * - * est_objs = nrblocks * (est_usedobjs / est_refblocks); + * est_totobjs = nrblocks * (est_usedobjs / est_usedblocks) * - * but this may overflow 64 bits or become 0 if not handled well + * this may overflow 64 bits or become 0 if not handled well. * - * We know nrblocks is below (64 - 17 = 47) bits from - * SPA_MAXBLKSHIFT, and est_usedobjs is under 48 bits due to - * DN_MAX_OBJECT_SHIFT, which means that multiplying them may - * get as large as 2 ^ 95. + * We know nrblocks is below 2^(64 - blkbits) bits, and + * est_usedobjs is under 48 bits due to DN_MAX_OBJECT_SHIFT, + * which means that multiplying them may get as large as + * 2 ^ 96 for the minimum blocksize of 64KB allowed above. * - * We also know (est_usedobjs / est_refdblocks) is between 2 and - * 256, due to above checks, we can safely compute this first. + * The ratio of dnodes per block (est_usedobjs / est_usedblocks) + * is under 2^(blkbits - DNODE_SHIFT) = blocksize / 512 due to + * the limit checks above, so we can safely compute this first. * We care more about accuracy on the MDT (many dnodes/block) * which is good because this is where truncation errors are - * smallest. This adds 8 bits to nrblocks so we can use 7 bits - * to compute a fixed-point fraction and nrblocks can still fit - * in 64 bits. */ - } else { - unsigned dnodes_per_block = (est_usedobjs << 7)/est_refdblocks; + * smallest. Since both nrblocks and dn_per_block are a + * function of blkbits, their product is at most: + * + * 2^(64 - blkbits) * 2^(blkbits - DNODE_SHIFT) = 2^(64 - 9) + * + * so we can safely use 7 bits to compute a fixed-point + * fraction and est_totobjs can still fit in 64 bits. + */ + unsigned dn_per_block = (est_usedobjs << 7) / est_usedblocks; - est_objs = (nrblocks * dnodes_per_block) >> 7; + est_totobjs = (nrblocks * dn_per_block) >> 7; } - return est_objs; + return est_totobjs; } -static int osd_objset_statfs(struct objset *os, struct obd_statfs *osfs) +static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs) { - uint64_t refdbytes, availbytes, usedobjs, availobjs; + struct objset *os = osd->od_os; + uint64_t usedbytes, availbytes, usedobjs, availobjs; uint64_t est_availobjs; uint64_t reserved; + uint64_t bshift; + + dmu_objset_space(os, &usedbytes, &availbytes, &usedobjs, &availobjs); - dmu_objset_space(os, &refdbytes, &availbytes, &usedobjs, - &availobjs); + memset(osfs, 0, sizeof(*osfs)); + + /* We're a zfs filesystem. */ + osfs->os_type = UBERBLOCK_MAGIC; /* * ZFS allows multiple block sizes. For statfs, Linux makes no @@ -391,14 +489,16 @@ static int osd_objset_statfs(struct objset *os, struct obd_statfs *osfs) * largest possible block size as IO size for the optimum performance * and scale the free and used blocks count appropriately. */ - osfs->os_bsize = 1ULL << SPA_MAXBLOCKSHIFT; + osfs->os_bsize = osd->od_max_blksz; + bshift = fls64(osfs->os_bsize) - 1; - osfs->os_blocks = (refdbytes + availbytes) >> SPA_MAXBLOCKSHIFT; - osfs->os_bfree = availbytes >> SPA_MAXBLOCKSHIFT; + osfs->os_blocks = (usedbytes + availbytes) >> bshift; + osfs->os_bfree = availbytes >> bshift; osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */ /* Take replication (i.e. number of copies) into account */ - osfs->os_bavail /= os->os_copies; + if (os->os_copies != 0) + osfs->os_bavail /= os->os_copies; /* * Reserve some space so we don't run into ENOSPC due to grants not @@ -406,19 +506,17 @@ static int osd_objset_statfs(struct objset *os, struct obd_statfs *osfs) * Rather than report this via os_bavail (which makes users unhappy if * they can't fill the filesystem 100%), reduce os_blocks as well. * - * Reserve 0.78% of total space, at least 4MB for small filesystems, + * Reserve 0.78% of total space, at least 16MB for small filesystems, * for internal files to be created/unlinked when space is tight. */ - CLASSERT(OSD_STATFS_RESERVED_BLKS > 0); - if (likely(osfs->os_blocks >= - OSD_STATFS_RESERVED_BLKS << OSD_STATFS_RESERVED_SHIFT)) + BUILD_BUG_ON(OSD_STATFS_RESERVED_SIZE <= 0); + reserved = OSD_STATFS_RESERVED_SIZE >> bshift; + if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT)) reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT; - else - reserved = OSD_STATFS_RESERVED_BLKS; osfs->os_blocks -= reserved; - osfs->os_bfree -= MIN(reserved, osfs->os_bfree); - osfs->os_bavail -= MIN(reserved, osfs->os_bavail); + osfs->os_bfree -= min(reserved, osfs->os_bfree); + osfs->os_bavail -= min(reserved, osfs->os_bavail); /* * The availobjs value returned from dmu_objset_space() is largely @@ -427,8 +525,8 @@ static int osd_objset_statfs(struct objset *os, struct obd_statfs *osfs) * issues like how much space is actually available in the pool. * Compute a better estimate in udmu_objs_count_estimate(). */ - est_availobjs = osd_objs_count_estimate(refdbytes, usedobjs, - osfs->os_bfree); + est_availobjs = osd_objs_count_estimate(usedbytes, usedobjs, + osfs->os_bfree, bshift); osfs->os_ffree = min(availobjs, est_availobjs); osfs->os_files = osfs->os_ffree + usedobjs; @@ -436,18 +534,13 @@ static int osd_objset_statfs(struct objset *os, struct obd_statfs *osfs) /* ZFS XXX: fill in backing dataset FSID/UUID memcpy(osfs->os_fsid, .... );*/ - /* We're a zfs filesystem. */ - osfs->os_type = UBERBLOCK_MAGIC; - - /* ZFS XXX: fill in appropriate OS_STATE_{DEGRADED,READONLY} flags - osfs->os_state = vf_to_stf(vfsp->vfs_flag); - if (sb->s_flags & MS_RDONLY) - osfs->os_state = OS_STATE_READONLY; - */ - osfs->os_namelen = MAXNAMELEN; osfs->os_maxbytes = OBD_OBJECT_EOF; + if (!spa_writeable(dmu_objset_spa(os)) || + osd->od_dev_set_rdonly || osd->od_prop_rdonly) + osfs->os_state |= OS_STATFS_READONLY; + return 0; } @@ -455,33 +548,40 @@ static int osd_objset_statfs(struct objset *os, struct obd_statfs *osfs) * Concurrency: shouldn't matter. */ int osd_statfs(const struct lu_env *env, struct dt_device *d, - struct obd_statfs *osfs) + struct obd_statfs *osfs, struct obd_statfs_info *info) { struct osd_device *osd = osd_dt_dev(d); - int rc; + int rc; ENTRY; - rc = osd_objset_statfs(osd->od_os, osfs); + rc = osd_objset_statfs(osd, osfs); if (unlikely(rc != 0)) RETURN(rc); osfs->os_bavail -= min_t(u64, OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize, osfs->os_bavail); + + /* ZFS does not support reporting nonrotional status yet, so return + * flag only if user has set nonrotational. + */ + osfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0; + RETURN(0); } -static int osd_blk_insert_cost(void) +static int osd_blk_insert_cost(struct osd_device *osd) { - int max_blockshift, nr_blkptrshift; + int max_blockshift, nr_blkptrshift, bshift; /* max_blockshift is the log2 of the number of blocks needed to reach * the maximum filesize (that's to say 2^64) */ - max_blockshift = DN_MAX_OFFSET_SHIFT - SPA_MAXBLOCKSHIFT; + bshift = osd_spa_maxblockshift(dmu_objset_spa(osd->od_os)); + max_blockshift = DN_MAX_OFFSET_SHIFT - bshift; /* nr_blkptrshift is the log2 of the number of block pointers that can * be stored in an indirect block */ - CLASSERT(DN_MAX_INDBLKSHIFT > SPA_BLKPTRSHIFT); + BUILD_BUG_ON(DN_MAX_INDBLKSHIFT <= SPA_BLKPTRSHIFT); nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT; /* max_blockshift / nr_blkptrshift is thus the maximum depth of the @@ -505,31 +605,48 @@ static void osd_conf_get(const struct lu_env *env, */ param->ddp_max_name_len = MAXNAMELEN; param->ddp_max_nlink = 1 << 31; /* it's 8byte on a disk */ - param->ddp_block_shift = 12; /* XXX */ + param->ddp_symlink_max = PATH_MAX; param->ddp_mount_type = LDD_MT_ZFS; param->ddp_mntopts = MNTOPT_USERXATTR; if (osd->od_posix_acl) param->ddp_mntopts |= MNTOPT_ACL; - param->ddp_max_ea_size = DXATTR_MAX_ENTRY_SIZE; + /* Previously DXATTR_MAX_ENTRY_SIZE */ + param->ddp_max_ea_size = OBD_MAX_EA_SIZE; /* for maxbytes, report same value as ZPL */ param->ddp_maxbytes = MAX_LFS_FILESIZE; - /* Default reserved fraction of the available space that should be kept - * for error margin. Unfortunately, there are many factors that can - * impact the overhead with zfs, so let's be very cautious for now and - * reserve 20% of the available space which is not given out as grant. - * This tunable can be changed on a live system via procfs if needed. */ - param->ddp_grant_reserved = 20; - /* inodes are dynamically allocated, so we report the per-inode space * consumption to upper layers. This static value is not really accurate * and we should use the same logic as in udmu_objset_statfs() to * estimate the real size consumed by an object */ param->ddp_inodespace = OSD_DNODE_EST_COUNT; - /* per-fragment overhead to be used by the client code */ - param->ddp_grant_frag = osd_blk_insert_cost(); + /* Although ZFS isn't an extent-based filesystem, the metadata overhead + * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should + * not be accounted for every single new block insertion. + * Instead, the maximum extent size is set to the number of blocks that + * can fit into a single contiguous indirect block. There would be some + * cases where this crosses indirect blocks, but it also won't have 7 + * new levels of indirect blocks in that case either, so it will still + * have enough reserved space for the extra indirect block */ + param->ddp_max_extent_blks = + (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT)); + param->ddp_extent_tax = osd_blk_insert_cost(osd); + + /* Preferred RPC size for efficient disk IO. 1MB shows good + * all-around performance for ZFS, but use blocksize (recordsize) + * by default if larger to avoid read-modify-write. */ + if (osd->od_max_blksz > ONE_MB_BRW_SIZE) + param->ddp_brw_size = osd->od_max_blksz; + else + param->ddp_brw_size = ONE_MB_BRW_SIZE; + +#ifdef HAVE_DMU_OFFSET_NEXT + param->ddp_has_lseek_data_hole = osd->od_sync_on_lseek; +#else + param->ddp_has_lseek_data_hole = false; +#endif } /* @@ -537,10 +654,14 @@ static void osd_conf_get(const struct lu_env *env, */ static int osd_sync(const struct lu_env *env, struct dt_device *d) { - struct osd_device *osd = osd_dt_dev(d); - CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME); - txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL); - CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME); + if (!d->dd_rdonly) { + struct osd_device *osd = osd_dt_dev(d); + + CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME); + txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL); + CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME); + } + return 0; } @@ -571,42 +692,45 @@ static int osd_ro(const struct lu_env *env, struct dt_device *d) CERROR("%s: *** setting device %s read-only ***\n", osd->od_svname, LUSTRE_OSD_ZFS_NAME); - osd->od_rdonly = 1; + osd->od_dev_set_rdonly = 1; spa_freeze(dmu_objset_spa(osd->od_os)); RETURN(0); } -/* - * Concurrency: serialization provided by callers. - */ -static int osd_init_capa_ctxt(const struct lu_env *env, struct dt_device *d, - int mode, unsigned long timeout, __u32 alg, - struct lustre_capa_key *keys) +/* reserve or free quota for some operation */ +static int osd_reserve_or_free_quota(const struct lu_env *env, + struct dt_device *dev, + enum quota_type type, __u64 uid, + __u64 gid, __s64 count, bool is_md) { - struct osd_device *dev = osd_dt_dev(d); - ENTRY; + int rc; + struct osd_device *osd = osd_dt_dev(dev); + struct osd_thread_info *info = osd_oti_get(env); + struct lquota_id_info *qi = &info->oti_qi; + struct qsd_instance *qsd = NULL; - dev->od_fl_capa = mode; - dev->od_capa_timeout = timeout; - dev->od_capa_alg = alg; - dev->od_capa_keys = keys; + if (is_md) + qsd = osd->od_quota_slave_md; + else + qsd = osd->od_quota_slave_dt; - RETURN(0); + rc = quota_reserve_or_free(env, qsd, qi, type, uid, gid, count, is_md); + RETURN(rc); } -static struct dt_device_operations osd_dt_ops = { - .dt_root_get = osd_root_get, - .dt_statfs = osd_statfs, - .dt_trans_create = osd_trans_create, - .dt_trans_start = osd_trans_start, - .dt_trans_stop = osd_trans_stop, - .dt_trans_cb_add = osd_trans_cb_add, - .dt_conf_get = osd_conf_get, - .dt_sync = osd_sync, - .dt_commit_async = osd_commit_async, - .dt_ro = osd_ro, - .dt_init_capa_ctxt = osd_init_capa_ctxt, +static const struct dt_device_operations osd_dt_ops = { + .dt_root_get = osd_root_get, + .dt_statfs = osd_statfs, + .dt_trans_create = osd_trans_create, + .dt_trans_start = osd_trans_start, + .dt_trans_stop = osd_trans_stop, + .dt_trans_cb_add = osd_trans_cb_add, + .dt_conf_get = osd_conf_get, + .dt_sync = osd_sync, + .dt_commit_async = osd_commit_async, + .dt_ro = osd_ro, + .dt_reserve_or_free_quota = osd_reserve_or_free_quota, }; /* @@ -640,16 +764,21 @@ static void osd_key_fini(const struct lu_context *ctx, struct lu_context_key *key, void *data) { struct osd_thread_info *info = data; + struct osd_idmap_cache *idc = info->oti_ins_cache; + if (idc != NULL) { + LASSERT(info->oti_ins_cache_size > 0); + OBD_FREE_PTR_ARRAY_LARGE(idc, info->oti_ins_cache_size); + info->oti_ins_cache = NULL; + info->oti_ins_cache_size = 0; + } + lu_buf_free(&info->oti_xattr_lbuf); OBD_FREE_PTR(info); } static void osd_key_exit(const struct lu_context *ctx, struct lu_context_key *key, void *data) { - struct osd_thread_info *info = data; - - memset(info, 0, sizeof(*info)); } struct lu_context_key osd_key = { @@ -659,16 +788,38 @@ struct lu_context_key osd_key = { .lct_exit = osd_key_exit }; +static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd) +{ + if (osd->od_cl_seq == NULL) + return; + + seq_client_fini(osd->od_cl_seq); + OBD_FREE_PTR(osd->od_cl_seq); + osd->od_cl_seq = NULL; +} + static int osd_shutdown(const struct lu_env *env, struct osd_device *o) { ENTRY; /* shutdown quota slave instance associated with the device */ - if (o->od_quota_slave != NULL) { - qsd_fini(env, o->od_quota_slave); - o->od_quota_slave = NULL; + if (o->od_quota_slave_md != NULL) { + /* complete all in-flight callbacks */ + osd_sync(env, &o->od_dt_dev); + txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os))); + qsd_fini(env, o->od_quota_slave_md); + o->od_quota_slave_md = NULL; } + if (o->od_quota_slave_dt != NULL) { + /* complete all in-flight callbacks */ + osd_sync(env, &o->od_dt_dev); + txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os))); + qsd_fini(env, o->od_quota_slave_dt); + o->od_quota_slave_dt = NULL; + } + osd_fid_fini(env, o); + RETURN(0); } @@ -679,17 +830,119 @@ static void osd_xattr_changed_cb(void *arg, uint64_t newval) osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA); } +static void osd_recordsize_changed_cb(void *arg, uint64_t newval) +{ + struct osd_device *osd = arg; + + LASSERT(newval <= osd_spa_maxblocksize(dmu_objset_spa(osd->od_os))); + LASSERT(newval >= SPA_MINBLOCKSIZE); + LASSERT(ISP2(newval)); + + osd->od_max_blksz = newval; +} + +static void osd_readonly_changed_cb(void *arg, uint64_t newval) +{ + struct osd_device *osd = arg; + + osd->od_prop_rdonly = !!newval; +} + +#ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE +static void osd_dnodesize_changed_cb(void *arg, uint64_t newval) +{ + struct osd_device *osd = arg; + + osd->od_dnsize = newval; +} +#endif +/* + * This function unregisters all registered callbacks. It's harmless to + * unregister callbacks that were never registered so it is used to safely + * unwind a partially completed call to osd_objset_register_callbacks(). + */ +static void osd_objset_unregister_callbacks(struct osd_device *o) +{ + struct dsl_dataset *ds = dmu_objset_ds(o->od_os); + + (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR), + osd_xattr_changed_cb, o); + (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE), + osd_recordsize_changed_cb, o); + (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY), + osd_readonly_changed_cb, o); +#ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE + (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE), + osd_dnodesize_changed_cb, o); +#endif + + if (o->arc_prune_cb != NULL) { + arc_remove_prune_callback(o->arc_prune_cb); + o->arc_prune_cb = NULL; + } +} + +/* + * Register the required callbacks to be notified when zfs properties + * are modified using the 'zfs(8)' command line utility. + */ +static int osd_objset_register_callbacks(struct osd_device *o) +{ + struct dsl_dataset *ds = dmu_objset_ds(o->od_os); + dsl_pool_t *dp = dmu_objset_pool(o->od_os); + int rc; + + LASSERT(ds); + LASSERT(dp); + + dsl_pool_config_enter(dp, FTAG); + rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR), + osd_xattr_changed_cb, o); + if (rc) + GOTO(err, rc); + + rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE), + osd_recordsize_changed_cb, o); + if (rc) + GOTO(err, rc); + + rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY), + osd_readonly_changed_cb, o); + if (rc) + GOTO(err, rc); + +#ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE + rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE), + osd_dnodesize_changed_cb, o); + if (rc) + GOTO(err, rc); +#endif + + o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o); +err: + dsl_pool_config_exit(dp, FTAG); + if (rc) + osd_objset_unregister_callbacks(o); + + RETURN(rc); +} + static int osd_objset_open(struct osd_device *o) { uint64_t version = ZPL_VERSION; - uint64_t sa_obj; + uint64_t sa_obj, unlink_obj; int rc; ENTRY; - rc = -dmu_objset_own(o->od_mntdev, DMU_OST_ZFS, B_FALSE, o, &o->od_os); + rc = -osd_dmu_objset_own(o->od_mntdev, DMU_OST_ZFS, + o->od_dt_dev.dd_rdonly ? B_TRUE : B_FALSE, + B_TRUE, o, &o->od_os); + if (rc) { + CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev); o->od_os = NULL; - goto out; + + GOTO(out, rc); } /* Check ZFS version */ @@ -722,32 +975,131 @@ static int osd_objset_open(struct osd_device *o) GOTO(out, rc); } + rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET, + 8, 1, &unlink_obj); + if (rc) { + CERROR("%s: lookup for %s failed: rc = %d\n", + o->od_svname, ZFS_UNLINKED_SET, rc); + GOTO(out, rc); + } + /* Check that user/group usage tracking is supported */ if (!dmu_objset_userused_enabled(o->od_os) || DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED || DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) { CERROR("%s: Space accounting not supported by this target, " "aborting\n", o->od_svname); - GOTO(out, -ENOTSUPP); + GOTO(out, rc = -ENOTSUPP); + } + + rc = __osd_obj2dnode(o->od_os, unlink_obj, &o->od_unlinked); + if (rc) { + CERROR("%s: can't get dnode for unlinked: rc = %d\n", + o->od_svname, rc); + GOTO(out, rc); } out: - if (rc != 0 && o->od_os != NULL) - dmu_objset_disown(o->od_os, o); + if (rc != 0 && o->od_os != NULL) { + osd_dmu_objset_disown(o->od_os, B_TRUE, o); + o->od_os = NULL; + } RETURN(rc); } +int osd_unlinked_object_free(const struct lu_env *env, struct osd_device *osd, + uint64_t oid) +{ + char *key = osd_oti_get(env)->oti_str; + int rc; + dmu_tx_t *tx; + + if (osd->od_dt_dev.dd_rdonly) { + CERROR("%s: someone try to free objects under " + "readonly mode, should be disabled.\n", osd_name(osd)); + dump_stack(); + + return -EROFS; + } + + rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END); + if (rc != 0) { + CWARN("%s: Cannot truncate %llu: rc = %d\n", + osd->od_svname, oid, rc); + return rc; + } + + tx = dmu_tx_create(osd->od_os); + dmu_tx_mark_netfree(tx); + dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END); + osd_tx_hold_zap(tx, osd->od_unlinked->dn_object, osd->od_unlinked, + FALSE, NULL); + rc = -dmu_tx_assign(tx, TXG_WAIT); + if (rc != 0) { + CWARN("%s: Cannot assign tx for %llu: rc = %d\n", + osd->od_svname, oid, rc); + goto failed; + } + + snprintf(key, sizeof(osd_oti_get(env)->oti_str), "%llx", oid); + rc = osd_zap_remove(osd, osd->od_unlinked->dn_object, + osd->od_unlinked, key, tx); + if (rc != 0) { + CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n", + osd->od_svname, oid, rc); + goto failed; + } + + rc = -dmu_object_free(osd->od_os, oid, tx); + if (rc != 0) { + CWARN("%s: Cannot free %llu: rc = %d\n", + osd->od_svname, oid, rc); + goto failed; + } + dmu_tx_commit(tx); + + return 0; + +failed: + LASSERT(rc != 0); + dmu_tx_abort(tx); + + return rc; +} + +static void +osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd) +{ + zap_cursor_t zc; + zap_attribute_t *za = &osd_oti_get(env)->oti_za; + + zap_cursor_init(&zc, osd->od_os, osd->od_unlinked->dn_object); + + while (zap_cursor_retrieve(&zc, za) == 0) { + /* If cannot free the object, leave it in the unlinked set, + * until the OSD is mounted again when obd_unlinked_drain() + * will be called. */ + if (osd_unlinked_object_free(env, osd, za->za_first_integer)) + break; + zap_cursor_advance(&zc); + } + + zap_cursor_fini(&zc); +} + static int osd_mount(const struct lu_env *env, struct osd_device *o, struct lustre_cfg *cfg) { - struct dsl_dataset *ds; - char *mntdev = lustre_cfg_string(cfg, 1); - char *svname = lustre_cfg_string(cfg, 4); - dmu_buf_t *rootdb; - dsl_pool_t *dp; - const char *opts; - int rc; + char *mntdev = lustre_cfg_string(cfg, 1); + char *str = lustre_cfg_string(cfg, 2); + char *svname = lustre_cfg_string(cfg, 4); + time64_t interval = AS_DEFAULT; + dnode_t *rootdn; + const char *opts; + bool resetoi = false; + int rc; + ENTRY; if (o->od_os != NULL) @@ -764,43 +1116,65 @@ static int osd_mount(const struct lu_env *env, if (rc >= sizeof(o->od_svname)) RETURN(-E2BIG); + opts = lustre_cfg_string(cfg, 3); + + o->od_index_backup_stop = 0; + o->od_index = -1; /* -1 means index is invalid */ + rc = server_name2index(o->od_svname, &o->od_index, NULL); + str = strstr(str, ":"); + if (str) { + unsigned long flags; + + rc = kstrtoul(str + 1, 10, &flags); + if (rc) + RETURN(-EINVAL); + + if (flags & LMD_FLG_DEV_RDONLY) { + o->od_dt_dev.dd_rdonly = 1; + LCONSOLE_WARN("%s: set dev_rdonly on this device\n", + svname); + } + + if (flags & LMD_FLG_NOSCRUB) + interval = AS_NEVER; + } + if (server_name_is_ost(o->od_svname)) o->od_is_ost = 1; rc = osd_objset_open(o); - if (rc) { - CERROR("%s: can't open objset %s: rc = %d\n", o->od_svname, - o->od_mntdev, rc); + if (rc) RETURN(rc); - } - ds = dmu_objset_ds(o->od_os); - dp = dmu_objset_pool(o->od_os); - LASSERT(ds); - LASSERT(dp); - dsl_pool_config_enter(dp, FTAG); - rc = dsl_prop_register(ds, "xattr", osd_xattr_changed_cb, o); - dsl_pool_config_exit(dp, FTAG); - if (rc) - CWARN("%s: can't register xattr callback, ignore: rc=%d\n", - o->od_svname, rc); + o->od_xattr_in_sa = B_TRUE; + o->od_max_blksz = osd_spa_maxblocksize(o->od_os->os_spa); + o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE; - rc = __osd_obj2dbuf(env, o->od_os, o->od_rootid, &rootdb); - if (rc) { - CERROR("%s: obj2dbuf() failed: rc = %d\n", o->od_svname, rc); - dmu_objset_disown(o->od_os, o); - o->od_os = NULL; - RETURN(rc); - } + rc = __osd_obj2dnode(o->od_os, o->od_rootid, &rootdn); + if (rc) + GOTO(err, rc); + o->od_root = rootdn->dn_object; + osd_dnode_rele(rootdn); - o->od_root = rootdb->db_object; - sa_buf_rele(rootdb, osd_obj_tag); + rc = __osd_obj2dnode(o->od_os, DMU_USERUSED_OBJECT, + &o->od_userused_dn); + if (rc) + GOTO(err, rc); - /* 1. initialize oi before any file create or file open */ - rc = osd_oi_init(env, o); + rc = __osd_obj2dnode(o->od_os, DMU_GROUPUSED_OBJECT, + &o->od_groupused_dn); if (rc) GOTO(err, rc); +#ifdef ZFS_PROJINHERIT + if (dmu_objset_projectquota_enabled(o->od_os)) { + rc = __osd_obj2dnode(o->od_os, DMU_PROJECTUSED_OBJECT, + &o->od_projectused_dn); + if (rc && rc != -ENOENT) + GOTO(err, rc); + } +#endif + rc = lu_site_init(&o->od_site, osd2lu_dev(o)); if (rc) GOTO(err, rc); @@ -810,35 +1184,71 @@ static int osd_mount(const struct lu_env *env, if (rc) GOTO(err, rc); - rc = osd_convert_root_to_new_seq(env, o); + rc = osd_objset_register_callbacks(o); if (rc) GOTO(err, rc); - /* Use our own ZAP for inode accounting by default, this can be changed - * via procfs to estimate the inode usage from the block usage */ - o->od_quota_iused_est = 0; + if (opts && strstr(opts, "resetoi")) + resetoi = true; + + rc = lprocfs_init_brw_stats(&o->od_brw_stats); + if (rc) + GOTO(err, rc); + + o->od_in_init = 1; + rc = osd_scrub_setup(env, o, interval, resetoi); + o->od_in_init = 0; + if (rc) + GOTO(err, rc); rc = osd_procfs_init(o, o->od_svname); if (rc) GOTO(err, rc); - o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o); + /* currently it's no need to prepare qsd_instance_md for OST */ + if (!o->od_is_ost) { + o->od_quota_slave_md = qsd_init(env, o->od_svname, + &o->od_dt_dev, o->od_proc_entry, + true, false); + if (IS_ERR(o->od_quota_slave_md)) { + rc = PTR_ERR(o->od_quota_slave_md); + o->od_quota_slave_md = NULL; + GOTO(err, rc); + } + } + + o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev, + o->od_proc_entry, false, false); - /* initialize quota slave instance */ - o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev, - o->od_proc_entry); - if (IS_ERR(o->od_quota_slave)) { - rc = PTR_ERR(o->od_quota_slave); - o->od_quota_slave = NULL; + if (IS_ERR(o->od_quota_slave_dt)) { + if (o->od_quota_slave_md != NULL) { + qsd_fini(env, o->od_quota_slave_md); + o->od_quota_slave_md = NULL; + } + + rc = PTR_ERR(o->od_quota_slave_dt); + o->od_quota_slave_dt = NULL; GOTO(err, rc); } +#ifdef HAVE_DMU_USEROBJ_ACCOUNTING + if (!osd_dmu_userobj_accounting_available(o)) + CWARN("%s: dnode accounting not enabled: " + "enable feature@userobj_accounting in pool\n", + o->od_mntdev); +#endif + /* parse mount option "noacl", and enable ACL by default */ - opts = lustre_cfg_string(cfg, 3); if (opts == NULL || strstr(opts, "noacl") == NULL) o->od_posix_acl = 1; + osd_unlinked_drain(env, o); err: + if (rc && o->od_os) { + osd_dmu_objset_disown(o->od_os, B_TRUE, o); + o->od_os = NULL; + } + RETURN(rc); } @@ -856,13 +1266,33 @@ static void osd_umount(const struct lu_env *env, struct osd_device *o) CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname, atomic_read(&o->od_zerocopy_pin)); + if (o->od_unlinked) { + osd_dnode_rele(o->od_unlinked); + o->od_unlinked = NULL; + } + if (o->od_userused_dn) { + osd_dnode_rele(o->od_userused_dn); + o->od_userused_dn = NULL; + } + if (o->od_groupused_dn) { + osd_dnode_rele(o->od_groupused_dn); + o->od_groupused_dn = NULL; + } + +#ifdef ZFS_PROJINHERIT + if (o->od_projectused_dn) { + osd_dnode_rele(o->od_projectused_dn); + o->od_projectused_dn = NULL; + } +#endif + if (o->od_os != NULL) { - /* force a txg sync to get all commit callbacks */ - txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL); + if (!o->od_dt_dev.dd_rdonly) + /* force a txg sync to get all commit callbacks */ + txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL); /* close the object set */ - dmu_objset_disown(o->od_os, o); - + osd_dmu_objset_disown(o->od_os, B_TRUE, o); o->od_os = NULL; } @@ -883,10 +1313,14 @@ static int osd_device_init0(const struct lu_env *env, l->ld_ops = &osd_lu_ops; o->od_dt_dev.dd_ops = &osd_dt_ops; + sema_init(&o->od_otable_sem, 1); + INIT_LIST_HEAD(&o->od_ios_list); + o->od_sync_on_lseek = B_TRUE; - o->od_capa_hash = init_capa_hash(); - if (o->od_capa_hash == NULL) - GOTO(out, rc = -ENOMEM); + /* ZFS does not support reporting nonrotional status yet, so this flag + * is only set if explicitly set by the user. + */ + o->od_nonrotational = 0; out: RETURN(rc); @@ -899,13 +1333,23 @@ static struct lu_device *osd_device_alloc(const struct lu_env *env, struct lu_device_type *type, struct lustre_cfg *cfg) { - struct osd_device *dev; - int rc; + struct osd_device *dev; + struct osd_seq_list *osl; + int rc; OBD_ALLOC_PTR(dev); if (dev == NULL) return ERR_PTR(-ENOMEM); + osl = &dev->od_seq_list; + INIT_LIST_HEAD(&osl->osl_seq_list); + rwlock_init(&osl->osl_seq_list_lock); + sema_init(&osl->osl_seq_init_sem, 1); + INIT_LIST_HEAD(&dev->od_index_backup_list); + INIT_LIST_HEAD(&dev->od_index_restore_list); + spin_lock_init(&dev->od_lock); + dev->od_index_backup_policy = LIBP_NONE; + rc = dt_device_init(&dev->od_dt_dev, type); if (rc == 0) { rc = osd_device_init0(env, dev, cfg); @@ -930,14 +1374,11 @@ static struct lu_device *osd_device_free(const struct lu_env *env, struct osd_device *o = osd_dev(d); ENTRY; - cleanup_capa_hash(o->od_capa_hash); /* XXX: make osd top device in order to release reference */ d->ld_site->ls_top_dev = d; lu_site_purge(env, d->ld_site, -1); - if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) { - LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL); - lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer); - } + lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems, + D_ERROR, lu_cdebug_printer); lu_site_fini(&o->od_site); dt_device_fini(&o->od_dt_dev); OBD_FREE_PTR(o); @@ -949,28 +1390,23 @@ static struct lu_device *osd_device_fini(const struct lu_env *env, struct lu_device *d) { struct osd_device *o = osd_dev(d); - struct dsl_dataset *ds; int rc; ENTRY; - - osd_shutdown(env, o); - osd_oi_fini(env, o); - + osd_index_backup(env, o, false); if (o->od_os) { - ds = dmu_objset_ds(o->od_os); - rc = dsl_prop_unregister(ds, "xattr", osd_xattr_changed_cb, o); - if (rc) - CERROR("%s: dsl_prop_unregister xattr error %d\n", - o->od_svname, rc); - if (o->arc_prune_cb != NULL) { - arc_remove_prune_callback(o->arc_prune_cb); - o->arc_prune_cb = NULL; + osd_objset_unregister_callbacks(o); + if (!o->od_dt_dev.dd_rdonly) { + osd_sync(env, lu2dt_dev(d)); + txg_wait_callbacks( + spa_get_dsl(dmu_objset_spa(o->od_os))); } - osd_sync(env, lu2dt_dev(d)); - txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os))); } + /* now with all the callbacks completed we can cleanup the remainings */ + osd_shutdown(env, o); + osd_scrub_cleanup(env, o); + rc = osd_procfs_fini(o); if (rc) { CERROR("proc fini error %d\n", rc); @@ -996,27 +1432,37 @@ static int osd_device_init(const struct lu_env *env, struct lu_device *d, static int osd_process_config(const struct lu_env *env, struct lu_device *d, struct lustre_cfg *cfg) { - struct osd_device *o = osd_dev(d); - int rc; - ENTRY; + struct osd_device *o = osd_dev(d); + ssize_t count; + int rc; + ENTRY; switch(cfg->lcfg_command) { case LCFG_SETUP: rc = osd_mount(env, o, cfg); break; case LCFG_CLEANUP: + /* For the case LCFG_PRE_CLEANUP is not called in advance, + * that may happend if hit failure during mount process. */ + osd_index_backup(env, o, false); rc = osd_shutdown(env, o); break; case LCFG_PARAM: { LASSERT(&o->od_dt_dev); - rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars, - cfg, &o->od_dt_dev); - if (rc > 0 || rc == -ENOSYS) - rc = class_process_proc_param(PARAM_OST, - lprocfs_osd_obd_vars, - cfg, &o->od_dt_dev); + count = class_modify_config(cfg, PARAM_OSD, + &o->od_dt_dev.dd_kobj); + if (count < 0) + count = class_modify_config(cfg, PARAM_OST, + &o->od_dt_dev.dd_kobj); + rc = count > 0 ? 0 : count; break; } + case LCFG_PRE_CLEANUP: + osd_scrub_stop(o); + osd_index_backup(env, o, + o->od_index_backup_policy != LIBP_NONE); + rc = 0; + break; default: rc = -ENOTTY; } @@ -1030,12 +1476,14 @@ static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d) int rc = 0; ENTRY; - if (osd->od_quota_slave == NULL) + if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL) RETURN(0); /* start qsd instance on recovery completion, this notifies the quota * slave code that we are about to process new requests now */ - rc = qsd_start(env, osd->od_quota_slave); + rc = qsd_start(env, osd->od_quota_slave_dt); + if (rc == 0 && osd->od_quota_slave_md != NULL) + rc = qsd_start(env, osd->od_quota_slave_md); RETURN(rc); } @@ -1091,6 +1539,38 @@ static int osd_obd_disconnect(struct obd_export *exp) RETURN(rc); } +static int osd_fid_init(const struct lu_env *env, struct osd_device *osd) +{ + struct seq_server_site *ss = osd_seq_site(osd); + int rc = 0; + ENTRY; + + if (osd->od_is_ost || osd->od_cl_seq != NULL) + RETURN(0); + + if (unlikely(ss == NULL)) + RETURN(-ENODEV); + + OBD_ALLOC_PTR(osd->od_cl_seq); + if (osd->od_cl_seq == NULL) + RETURN(-ENOMEM); + + seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA, + osd->od_svname, ss->ss_server_seq); + + if (ss->ss_node_id == 0) { + /* + * If the OSD on the sequence controller(MDT0), then allocate + * sequence here, otherwise allocate sequence after connected + * to MDT0 (see mdt_register_lwp_callback()). + */ + rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv, + &osd->od_cl_seq->lcs_space, env); + } + + RETURN(rc); +} + static int osd_prepare(const struct lu_env *env, struct lu_device *pdev, struct lu_device *dev) { @@ -1098,18 +1578,47 @@ static int osd_prepare(const struct lu_env *env, struct lu_device *pdev, int rc = 0; ENTRY; - if (osd->od_quota_slave != NULL) + if (osd->od_quota_slave_md != NULL) { + /* set up quota slave objects */ + rc = qsd_prepare(env, osd->od_quota_slave_md); + if (rc != 0) + RETURN(rc); + } + + if (osd->od_quota_slave_dt != NULL) { /* set up quota slave objects */ - rc = qsd_prepare(env, osd->od_quota_slave); + rc = qsd_prepare(env, osd->od_quota_slave_dt); + if (rc != 0) + RETURN(rc); + } + + rc = osd_fid_init(env, osd); RETURN(rc); } -struct lu_device_operations osd_lu_ops = { +/** + * Implementation of lu_device_operations::ldo_fid_alloc() for OSD + * + * Allocate FID. + * + * see include/lu_object.h for the details. + */ +static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d, + struct lu_fid *fid, struct lu_object *parent, + const struct lu_name *name) +{ + struct osd_device *osd = osd_dev(d); + + return seq_client_alloc_fid(env, osd->od_cl_seq, fid); +} + +const struct lu_device_operations osd_lu_ops = { .ldo_object_alloc = osd_object_alloc, .ldo_process_config = osd_process_config, .ldo_recovery_complete = osd_recovery_complete, .ldo_prepare = osd_prepare, + .ldo_fid_alloc = osd_fid_alloc, }; static void osd_type_start(struct lu_device_type *t) @@ -1120,15 +1629,7 @@ static void osd_type_stop(struct lu_device_type *t) { } -int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp, - struct lu_fid *fid, struct md_op_data *op_data) -{ - struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev); - - return seq_client_alloc_fid(env, osd->od_cl_seq, fid); -} - -static struct lu_device_type_operations osd_device_type_ops = { +static const struct lu_device_type_operations osd_device_type_ops = { .ldto_init = osd_type_init, .ldto_fini = osd_type_fini, @@ -1150,14 +1651,13 @@ static struct lu_device_type osd_device_type = { }; -static struct obd_ops osd_obd_device_ops = { +static const struct obd_ops osd_obd_device_ops = { .o_owner = THIS_MODULE, .o_connect = osd_obd_connect, .o_disconnect = osd_obd_disconnect, - .o_fid_alloc = osd_fid_alloc }; -int __init osd_init(void) +static int __init osd_init(void) { int rc; @@ -1169,26 +1669,30 @@ int __init osd_init(void) if (rc) return rc; - rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL, + rc = class_register_type(&osd_obd_device_ops, NULL, true, LUSTRE_OSD_ZFS_NAME, &osd_device_type); if (rc) lu_kmem_fini(osd_caches); return rc; } -void __exit osd_exit(void) +static void __exit osd_exit(void) { class_unregister_type(LUSTRE_OSD_ZFS_NAME); lu_kmem_fini(osd_caches); } -extern unsigned int osd_oi_count; -CFS_MODULE_PARM(osd_oi_count, "i", int, 0444, - "Number of Object Index containers to be created, " - "it's only valid for new filesystem."); +module_param(osd_oi_count, int, 0444); +MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem."); + +module_param(osd_txg_sync_delay_us, int, 0644); +MODULE_PARM_DESC(osd_txg_sync_delay_us, + "When zero or larger delay N usec instead of doing TXG sync"); -MODULE_AUTHOR("Sun Microsystems, Inc. "); +MODULE_AUTHOR("OpenSFS, Inc. "); MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")"); +MODULE_VERSION(LUSTRE_VERSION_STRING); MODULE_LICENSE("GPL"); -cfs_module(osd, LUSTRE_VERSION_STRING, osd_init, osd_exit); +module_init(osd_init); +module_exit(osd_exit);