15cabbc6bSPrashanth Sreenivasa /* 25cabbc6bSPrashanth Sreenivasa * CDDL HEADER START 35cabbc6bSPrashanth Sreenivasa * 45cabbc6bSPrashanth Sreenivasa * The contents of this file are subject to the terms of the 55cabbc6bSPrashanth Sreenivasa * Common Development and Distribution License (the "License"). 65cabbc6bSPrashanth Sreenivasa * You may not use this file except in compliance with the License. 75cabbc6bSPrashanth Sreenivasa * 85cabbc6bSPrashanth Sreenivasa * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 95cabbc6bSPrashanth Sreenivasa * or http://www.opensolaris.org/os/licensing. 105cabbc6bSPrashanth Sreenivasa * See the License for the specific language governing permissions 115cabbc6bSPrashanth Sreenivasa * and limitations under the License. 125cabbc6bSPrashanth Sreenivasa * 135cabbc6bSPrashanth Sreenivasa * When distributing Covered Code, include this CDDL HEADER in each 145cabbc6bSPrashanth Sreenivasa * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 155cabbc6bSPrashanth Sreenivasa * If applicable, add the following below this CDDL HEADER, with the 165cabbc6bSPrashanth Sreenivasa * fields enclosed by brackets "[]" replaced with your own identifying 175cabbc6bSPrashanth Sreenivasa * information: Portions Copyright [yyyy] [name of copyright owner] 185cabbc6bSPrashanth Sreenivasa * 195cabbc6bSPrashanth Sreenivasa * CDDL HEADER END 205cabbc6bSPrashanth Sreenivasa */ 215cabbc6bSPrashanth Sreenivasa 225cabbc6bSPrashanth Sreenivasa /* 235cabbc6bSPrashanth Sreenivasa * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 245cabbc6bSPrashanth Sreenivasa * Copyright (c) 2011, 2018 by Delphix. All rights reserved. 255cabbc6bSPrashanth Sreenivasa */ 265cabbc6bSPrashanth Sreenivasa 275cabbc6bSPrashanth Sreenivasa #include <sys/zfs_context.h> 285cabbc6bSPrashanth Sreenivasa #include <sys/spa_impl.h> 295cabbc6bSPrashanth Sreenivasa #include <sys/dmu.h> 305cabbc6bSPrashanth Sreenivasa #include <sys/dmu_tx.h> 315cabbc6bSPrashanth Sreenivasa #include <sys/zap.h> 325cabbc6bSPrashanth Sreenivasa #include <sys/vdev_impl.h> 335cabbc6bSPrashanth Sreenivasa #include <sys/metaslab.h> 345cabbc6bSPrashanth Sreenivasa #include <sys/metaslab_impl.h> 355cabbc6bSPrashanth Sreenivasa #include <sys/uberblock_impl.h> 365cabbc6bSPrashanth Sreenivasa #include <sys/txg.h> 375cabbc6bSPrashanth Sreenivasa #include <sys/avl.h> 385cabbc6bSPrashanth Sreenivasa #include <sys/bpobj.h> 395cabbc6bSPrashanth Sreenivasa #include <sys/dsl_pool.h> 405cabbc6bSPrashanth Sreenivasa #include <sys/dsl_synctask.h> 415cabbc6bSPrashanth Sreenivasa #include <sys/dsl_dir.h> 425cabbc6bSPrashanth Sreenivasa #include <sys/arc.h> 435cabbc6bSPrashanth Sreenivasa #include <sys/zfeature.h> 445cabbc6bSPrashanth Sreenivasa #include <sys/vdev_indirect_births.h> 455cabbc6bSPrashanth Sreenivasa #include <sys/vdev_indirect_mapping.h> 465cabbc6bSPrashanth Sreenivasa #include <sys/abd.h> 47094e47e9SGeorge Wilson #include <sys/vdev_initialize.h> 485cabbc6bSPrashanth Sreenivasa 495cabbc6bSPrashanth Sreenivasa /* 505cabbc6bSPrashanth Sreenivasa * This file contains the necessary logic to remove vdevs from a 515cabbc6bSPrashanth Sreenivasa * storage pool. Currently, the only devices that can be removed 525cabbc6bSPrashanth Sreenivasa * are log, cache, and spare devices; and top level vdevs from a pool 535cabbc6bSPrashanth Sreenivasa * w/o raidz. (Note that members of a mirror can also be removed 545cabbc6bSPrashanth Sreenivasa * by the detach operation.) 555cabbc6bSPrashanth Sreenivasa * 565cabbc6bSPrashanth Sreenivasa * Log vdevs are removed by evacuating them and then turning the vdev 575cabbc6bSPrashanth Sreenivasa * into a hole vdev while holding spa config locks. 585cabbc6bSPrashanth Sreenivasa * 595cabbc6bSPrashanth Sreenivasa * Top level vdevs are removed and converted into an indirect vdev via 605cabbc6bSPrashanth Sreenivasa * a multi-step process: 615cabbc6bSPrashanth Sreenivasa * 625cabbc6bSPrashanth Sreenivasa * - Disable allocations from this device (spa_vdev_remove_top). 635cabbc6bSPrashanth Sreenivasa * 645cabbc6bSPrashanth Sreenivasa * - From a new thread (spa_vdev_remove_thread), copy data from 655cabbc6bSPrashanth Sreenivasa * the removing vdev to a different vdev. The copy happens in open 665cabbc6bSPrashanth Sreenivasa * context (spa_vdev_copy_impl) and issues a sync task 675cabbc6bSPrashanth Sreenivasa * (vdev_mapping_sync) so the sync thread can update the partial 685cabbc6bSPrashanth Sreenivasa * indirect mappings in core and on disk. 695cabbc6bSPrashanth Sreenivasa * 705cabbc6bSPrashanth Sreenivasa * - If a free happens during a removal, it is freed from the 715cabbc6bSPrashanth Sreenivasa * removing vdev, and if it has already been copied, from the new 725cabbc6bSPrashanth Sreenivasa * location as well (free_from_removing_vdev). 735cabbc6bSPrashanth Sreenivasa * 745cabbc6bSPrashanth Sreenivasa * - After the removal is completed, the copy thread converts the vdev 755cabbc6bSPrashanth Sreenivasa * into an indirect vdev (vdev_remove_complete) before instructing 765cabbc6bSPrashanth Sreenivasa * the sync thread to destroy the space maps and finish the removal 775cabbc6bSPrashanth Sreenivasa * (spa_finish_removal). 785cabbc6bSPrashanth Sreenivasa */ 795cabbc6bSPrashanth Sreenivasa 805cabbc6bSPrashanth Sreenivasa typedef struct vdev_copy_arg { 815cabbc6bSPrashanth Sreenivasa metaslab_t *vca_msp; 825cabbc6bSPrashanth Sreenivasa uint64_t vca_outstanding_bytes; 835cabbc6bSPrashanth Sreenivasa kcondvar_t vca_cv; 845cabbc6bSPrashanth Sreenivasa kmutex_t vca_lock; 855cabbc6bSPrashanth Sreenivasa } vdev_copy_arg_t; 865cabbc6bSPrashanth Sreenivasa 875cabbc6bSPrashanth Sreenivasa /* 883a4b1be9SMatthew Ahrens * The maximum amount of memory we can use for outstanding i/o while 893a4b1be9SMatthew Ahrens * doing a device removal. This determines how much i/o we can have 903a4b1be9SMatthew Ahrens * in flight concurrently. 915cabbc6bSPrashanth Sreenivasa */ 923a4b1be9SMatthew Ahrens int zfs_remove_max_copy_bytes = 64 * 1024 * 1024; 935cabbc6bSPrashanth Sreenivasa 945cabbc6bSPrashanth Sreenivasa /* 955cabbc6bSPrashanth Sreenivasa * The largest contiguous segment that we will attempt to allocate when 965cabbc6bSPrashanth Sreenivasa * removing a device. This can be no larger than SPA_MAXBLOCKSIZE. If 975cabbc6bSPrashanth Sreenivasa * there is a performance problem with attempting to allocate large blocks, 985cabbc6bSPrashanth Sreenivasa * consider decreasing this. 995cabbc6bSPrashanth Sreenivasa * 1005cabbc6bSPrashanth Sreenivasa * Note: we will issue I/Os of up to this size. The mpt driver does not 1015cabbc6bSPrashanth Sreenivasa * respond well to I/Os larger than 1MB, so we set this to 1MB. (When 1025cabbc6bSPrashanth Sreenivasa * mpt processes an I/O larger than 1MB, it needs to do an allocation of 1035cabbc6bSPrashanth Sreenivasa * 2 physically contiguous pages; if this allocation fails, mpt will drop 1045cabbc6bSPrashanth Sreenivasa * the I/O and hang the device.) 1055cabbc6bSPrashanth Sreenivasa */ 1065cabbc6bSPrashanth Sreenivasa int zfs_remove_max_segment = 1024 * 1024; 1075cabbc6bSPrashanth Sreenivasa 108*cfd63e1bSMatthew Ahrens /* 109*cfd63e1bSMatthew Ahrens * Allow a remap segment to span free chunks of at most this size. The main 110*cfd63e1bSMatthew Ahrens * impact of a larger span is that we will read and write larger, more 111*cfd63e1bSMatthew Ahrens * contiguous chunks, with more "unnecessary" data -- trading off bandwidth 112*cfd63e1bSMatthew Ahrens * for iops. The value here was chosen to align with 113*cfd63e1bSMatthew Ahrens * zfs_vdev_read_gap_limit, which is a similar concept when doing regular 114*cfd63e1bSMatthew Ahrens * reads (but there's no reason it has to be the same). 115*cfd63e1bSMatthew Ahrens * 116*cfd63e1bSMatthew Ahrens * Additionally, a higher span will have the following relatively minor 117*cfd63e1bSMatthew Ahrens * effects: 118*cfd63e1bSMatthew Ahrens * - the mapping will be smaller, since one entry can cover more allocated 119*cfd63e1bSMatthew Ahrens * segments 120*cfd63e1bSMatthew Ahrens * - more of the fragmentation in the removing device will be preserved 121*cfd63e1bSMatthew Ahrens * - we'll do larger allocations, which may fail and fall back on smaller 122*cfd63e1bSMatthew Ahrens * allocations 123*cfd63e1bSMatthew Ahrens */ 124*cfd63e1bSMatthew Ahrens int vdev_removal_max_span = 32 * 1024; 125*cfd63e1bSMatthew Ahrens 12686714001SSerapheim Dimitropoulos /* 12786714001SSerapheim Dimitropoulos * This is used by the test suite so that it can ensure that certain 12886714001SSerapheim Dimitropoulos * actions happen while in the middle of a removal. 12986714001SSerapheim Dimitropoulos */ 13086714001SSerapheim Dimitropoulos uint64_t zfs_remove_max_bytes_pause = UINT64_MAX; 13186714001SSerapheim Dimitropoulos 1325cabbc6bSPrashanth Sreenivasa #define VDEV_REMOVAL_ZAP_OBJS "lzap" 1335cabbc6bSPrashanth Sreenivasa 1345cabbc6bSPrashanth Sreenivasa static void spa_vdev_remove_thread(void *arg); 1355cabbc6bSPrashanth Sreenivasa 1365cabbc6bSPrashanth Sreenivasa static void 1375cabbc6bSPrashanth Sreenivasa spa_sync_removing_state(spa_t *spa, dmu_tx_t *tx) 1385cabbc6bSPrashanth Sreenivasa { 1395cabbc6bSPrashanth Sreenivasa VERIFY0(zap_update(spa->spa_dsl_pool->dp_meta_objset, 1405cabbc6bSPrashanth Sreenivasa DMU_POOL_DIRECTORY_OBJECT, 1415cabbc6bSPrashanth Sreenivasa DMU_POOL_REMOVING, sizeof (uint64_t), 1425cabbc6bSPrashanth Sreenivasa sizeof (spa->spa_removing_phys) / sizeof (uint64_t), 1435cabbc6bSPrashanth Sreenivasa &spa->spa_removing_phys, tx)); 1445cabbc6bSPrashanth Sreenivasa } 1455cabbc6bSPrashanth Sreenivasa 1465cabbc6bSPrashanth Sreenivasa static nvlist_t * 1475cabbc6bSPrashanth Sreenivasa spa_nvlist_lookup_by_guid(nvlist_t **nvpp, int count, uint64_t target_guid) 1485cabbc6bSPrashanth Sreenivasa { 1495cabbc6bSPrashanth Sreenivasa for (int i = 0; i < count; i++) { 1505cabbc6bSPrashanth Sreenivasa uint64_t guid = 1515cabbc6bSPrashanth Sreenivasa fnvlist_lookup_uint64(nvpp[i], ZPOOL_CONFIG_GUID); 1525cabbc6bSPrashanth Sreenivasa 1535cabbc6bSPrashanth Sreenivasa if (guid == target_guid) 1545cabbc6bSPrashanth Sreenivasa return (nvpp[i]); 1555cabbc6bSPrashanth Sreenivasa } 1565cabbc6bSPrashanth Sreenivasa 1575cabbc6bSPrashanth Sreenivasa return (NULL); 1585cabbc6bSPrashanth Sreenivasa } 1595cabbc6bSPrashanth Sreenivasa 1605cabbc6bSPrashanth Sreenivasa static void 1615cabbc6bSPrashanth Sreenivasa spa_vdev_remove_aux(nvlist_t *config, char *name, nvlist_t **dev, int count, 1625cabbc6bSPrashanth Sreenivasa nvlist_t *dev_to_remove) 1635cabbc6bSPrashanth Sreenivasa { 1645cabbc6bSPrashanth Sreenivasa nvlist_t **newdev = NULL; 1655cabbc6bSPrashanth Sreenivasa 1665cabbc6bSPrashanth Sreenivasa if (count > 1) 1675cabbc6bSPrashanth Sreenivasa newdev = kmem_alloc((count - 1) * sizeof (void *), KM_SLEEP); 1685cabbc6bSPrashanth Sreenivasa 1695cabbc6bSPrashanth Sreenivasa for (int i = 0, j = 0; i < count; i++) { 1705cabbc6bSPrashanth Sreenivasa if (dev[i] == dev_to_remove) 1715cabbc6bSPrashanth Sreenivasa continue; 1725cabbc6bSPrashanth Sreenivasa VERIFY(nvlist_dup(dev[i], &newdev[j++], KM_SLEEP) == 0); 1735cabbc6bSPrashanth Sreenivasa } 1745cabbc6bSPrashanth Sreenivasa 1755cabbc6bSPrashanth Sreenivasa VERIFY(nvlist_remove(config, name, DATA_TYPE_NVLIST_ARRAY) == 0); 1765cabbc6bSPrashanth Sreenivasa VERIFY(nvlist_add_nvlist_array(config, name, newdev, count - 1) == 0); 1775cabbc6bSPrashanth Sreenivasa 1785cabbc6bSPrashanth Sreenivasa for (int i = 0; i < count - 1; i++) 1795cabbc6bSPrashanth Sreenivasa nvlist_free(newdev[i]); 1805cabbc6bSPrashanth Sreenivasa 1815cabbc6bSPrashanth Sreenivasa if (count > 1) 1825cabbc6bSPrashanth Sreenivasa kmem_free(newdev, (count - 1) * sizeof (void *)); 1835cabbc6bSPrashanth Sreenivasa } 1845cabbc6bSPrashanth Sreenivasa 1855cabbc6bSPrashanth Sreenivasa static spa_vdev_removal_t * 1865cabbc6bSPrashanth Sreenivasa spa_vdev_removal_create(vdev_t *vd) 1875cabbc6bSPrashanth Sreenivasa { 1885cabbc6bSPrashanth Sreenivasa spa_vdev_removal_t *svr = kmem_zalloc(sizeof (*svr), KM_SLEEP); 1895cabbc6bSPrashanth Sreenivasa mutex_init(&svr->svr_lock, NULL, MUTEX_DEFAULT, NULL); 1905cabbc6bSPrashanth Sreenivasa cv_init(&svr->svr_cv, NULL, CV_DEFAULT, NULL); 1915cabbc6bSPrashanth Sreenivasa svr->svr_allocd_segs = range_tree_create(NULL, NULL); 1923a4b1be9SMatthew Ahrens svr->svr_vdev_id = vd->vdev_id; 1935cabbc6bSPrashanth Sreenivasa 1945cabbc6bSPrashanth Sreenivasa for (int i = 0; i < TXG_SIZE; i++) { 1955cabbc6bSPrashanth Sreenivasa svr->svr_frees[i] = range_tree_create(NULL, NULL); 1965cabbc6bSPrashanth Sreenivasa list_create(&svr->svr_new_segments[i], 1975cabbc6bSPrashanth Sreenivasa sizeof (vdev_indirect_mapping_entry_t), 1985cabbc6bSPrashanth Sreenivasa offsetof(vdev_indirect_mapping_entry_t, vime_node)); 1995cabbc6bSPrashanth Sreenivasa } 2005cabbc6bSPrashanth Sreenivasa 2015cabbc6bSPrashanth Sreenivasa return (svr); 2025cabbc6bSPrashanth Sreenivasa } 2035cabbc6bSPrashanth Sreenivasa 2045cabbc6bSPrashanth Sreenivasa void 2055cabbc6bSPrashanth Sreenivasa spa_vdev_removal_destroy(spa_vdev_removal_t *svr) 2065cabbc6bSPrashanth Sreenivasa { 2075cabbc6bSPrashanth Sreenivasa for (int i = 0; i < TXG_SIZE; i++) { 2085cabbc6bSPrashanth Sreenivasa ASSERT0(svr->svr_bytes_done[i]); 2095cabbc6bSPrashanth Sreenivasa ASSERT0(svr->svr_max_offset_to_sync[i]); 2105cabbc6bSPrashanth Sreenivasa range_tree_destroy(svr->svr_frees[i]); 2115cabbc6bSPrashanth Sreenivasa list_destroy(&svr->svr_new_segments[i]); 2125cabbc6bSPrashanth Sreenivasa } 2135cabbc6bSPrashanth Sreenivasa 2145cabbc6bSPrashanth Sreenivasa range_tree_destroy(svr->svr_allocd_segs); 2155cabbc6bSPrashanth Sreenivasa mutex_destroy(&svr->svr_lock); 2165cabbc6bSPrashanth Sreenivasa cv_destroy(&svr->svr_cv); 2175cabbc6bSPrashanth Sreenivasa kmem_free(svr, sizeof (*svr)); 2185cabbc6bSPrashanth Sreenivasa } 2195cabbc6bSPrashanth Sreenivasa 2205cabbc6bSPrashanth Sreenivasa /* 2215cabbc6bSPrashanth Sreenivasa * This is called as a synctask in the txg in which we will mark this vdev 2225cabbc6bSPrashanth Sreenivasa * as removing (in the config stored in the MOS). 2235cabbc6bSPrashanth Sreenivasa * 2245cabbc6bSPrashanth Sreenivasa * It begins the evacuation of a toplevel vdev by: 2255cabbc6bSPrashanth Sreenivasa * - initializing the spa_removing_phys which tracks this removal 2265cabbc6bSPrashanth Sreenivasa * - computing the amount of space to remove for accounting purposes 2275cabbc6bSPrashanth Sreenivasa * - dirtying all dbufs in the spa_config_object 2285cabbc6bSPrashanth Sreenivasa * - creating the spa_vdev_removal 2295cabbc6bSPrashanth Sreenivasa * - starting the spa_vdev_remove_thread 2305cabbc6bSPrashanth Sreenivasa */ 2315cabbc6bSPrashanth Sreenivasa static void 2325cabbc6bSPrashanth Sreenivasa vdev_remove_initiate_sync(void *arg, dmu_tx_t *tx) 2335cabbc6bSPrashanth Sreenivasa { 2343a4b1be9SMatthew Ahrens int vdev_id = (uintptr_t)arg; 2353a4b1be9SMatthew Ahrens spa_t *spa = dmu_tx_pool(tx)->dp_spa; 2363a4b1be9SMatthew Ahrens vdev_t *vd = vdev_lookup_top(spa, vdev_id); 2375cabbc6bSPrashanth Sreenivasa vdev_indirect_config_t *vic = &vd->vdev_indirect_config; 2385cabbc6bSPrashanth Sreenivasa objset_t *mos = spa->spa_dsl_pool->dp_meta_objset; 2395cabbc6bSPrashanth Sreenivasa spa_vdev_removal_t *svr = NULL; 2405cabbc6bSPrashanth Sreenivasa uint64_t txg = dmu_tx_get_txg(tx); 2415cabbc6bSPrashanth Sreenivasa 2425cabbc6bSPrashanth Sreenivasa ASSERT3P(vd->vdev_ops, !=, &vdev_raidz_ops); 2435cabbc6bSPrashanth Sreenivasa svr = spa_vdev_removal_create(vd); 2445cabbc6bSPrashanth Sreenivasa 2455cabbc6bSPrashanth Sreenivasa ASSERT(vd->vdev_removing); 2465cabbc6bSPrashanth Sreenivasa ASSERT3P(vd->vdev_indirect_mapping, ==, NULL); 2475cabbc6bSPrashanth Sreenivasa 2485cabbc6bSPrashanth Sreenivasa spa_feature_incr(spa, SPA_FEATURE_DEVICE_REMOVAL, tx); 2495cabbc6bSPrashanth Sreenivasa if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) { 2505cabbc6bSPrashanth Sreenivasa /* 2515cabbc6bSPrashanth Sreenivasa * By activating the OBSOLETE_COUNTS feature, we prevent 2525cabbc6bSPrashanth Sreenivasa * the pool from being downgraded and ensure that the 2535cabbc6bSPrashanth Sreenivasa * refcounts are precise. 2545cabbc6bSPrashanth Sreenivasa */ 2555cabbc6bSPrashanth Sreenivasa spa_feature_incr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx); 2565cabbc6bSPrashanth Sreenivasa uint64_t one = 1; 2575cabbc6bSPrashanth Sreenivasa VERIFY0(zap_add(spa->spa_meta_objset, vd->vdev_top_zap, 2585cabbc6bSPrashanth Sreenivasa VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE, sizeof (one), 1, 2595cabbc6bSPrashanth Sreenivasa &one, tx)); 2605cabbc6bSPrashanth Sreenivasa ASSERT3U(vdev_obsolete_counts_are_precise(vd), !=, 0); 2615cabbc6bSPrashanth Sreenivasa } 2625cabbc6bSPrashanth Sreenivasa 2635cabbc6bSPrashanth Sreenivasa vic->vic_mapping_object = vdev_indirect_mapping_alloc(mos, tx); 2645cabbc6bSPrashanth Sreenivasa vd->vdev_indirect_mapping = 2655cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_open(mos, vic->vic_mapping_object); 2665cabbc6bSPrashanth Sreenivasa vic->vic_births_object = vdev_indirect_births_alloc(mos, tx); 2675cabbc6bSPrashanth Sreenivasa vd->vdev_indirect_births = 2685cabbc6bSPrashanth Sreenivasa vdev_indirect_births_open(mos, vic->vic_births_object); 2695cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_removing_vdev = vd->vdev_id; 2705cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_start_time = gethrestime_sec(); 2715cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_end_time = 0; 2725cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_state = DSS_SCANNING; 2735cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_to_copy = 0; 2745cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_copied = 0; 2755cabbc6bSPrashanth Sreenivasa 2765cabbc6bSPrashanth Sreenivasa /* 2775cabbc6bSPrashanth Sreenivasa * Note: We can't use vdev_stat's vs_alloc for sr_to_copy, because 2785cabbc6bSPrashanth Sreenivasa * there may be space in the defer tree, which is free, but still 2795cabbc6bSPrashanth Sreenivasa * counted in vs_alloc. 2805cabbc6bSPrashanth Sreenivasa */ 2815cabbc6bSPrashanth Sreenivasa for (uint64_t i = 0; i < vd->vdev_ms_count; i++) { 2825cabbc6bSPrashanth Sreenivasa metaslab_t *ms = vd->vdev_ms[i]; 2835cabbc6bSPrashanth Sreenivasa if (ms->ms_sm == NULL) 2845cabbc6bSPrashanth Sreenivasa continue; 2855cabbc6bSPrashanth Sreenivasa 2865cabbc6bSPrashanth Sreenivasa /* 2875cabbc6bSPrashanth Sreenivasa * Sync tasks happen before metaslab_sync(), therefore 2885cabbc6bSPrashanth Sreenivasa * smp_alloc and sm_alloc must be the same. 2895cabbc6bSPrashanth Sreenivasa */ 2905cabbc6bSPrashanth Sreenivasa ASSERT3U(space_map_allocated(ms->ms_sm), ==, 2915cabbc6bSPrashanth Sreenivasa ms->ms_sm->sm_phys->smp_alloc); 2925cabbc6bSPrashanth Sreenivasa 2935cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_to_copy += 2945cabbc6bSPrashanth Sreenivasa space_map_allocated(ms->ms_sm); 2955cabbc6bSPrashanth Sreenivasa 2965cabbc6bSPrashanth Sreenivasa /* 2975cabbc6bSPrashanth Sreenivasa * Space which we are freeing this txg does not need to 2985cabbc6bSPrashanth Sreenivasa * be copied. 2995cabbc6bSPrashanth Sreenivasa */ 3005cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_to_copy -= 30186714001SSerapheim Dimitropoulos range_tree_space(ms->ms_freeing); 3025cabbc6bSPrashanth Sreenivasa 30386714001SSerapheim Dimitropoulos ASSERT0(range_tree_space(ms->ms_freed)); 3045cabbc6bSPrashanth Sreenivasa for (int t = 0; t < TXG_SIZE; t++) 30586714001SSerapheim Dimitropoulos ASSERT0(range_tree_space(ms->ms_allocating[t])); 3065cabbc6bSPrashanth Sreenivasa } 3075cabbc6bSPrashanth Sreenivasa 3085cabbc6bSPrashanth Sreenivasa /* 3095cabbc6bSPrashanth Sreenivasa * Sync tasks are called before metaslab_sync(), so there should 3105cabbc6bSPrashanth Sreenivasa * be no already-synced metaslabs in the TXG_CLEAN list. 3115cabbc6bSPrashanth Sreenivasa */ 3125cabbc6bSPrashanth Sreenivasa ASSERT3P(txg_list_head(&vd->vdev_ms_list, TXG_CLEAN(txg)), ==, NULL); 3135cabbc6bSPrashanth Sreenivasa 3145cabbc6bSPrashanth Sreenivasa spa_sync_removing_state(spa, tx); 3155cabbc6bSPrashanth Sreenivasa 3165cabbc6bSPrashanth Sreenivasa /* 3175cabbc6bSPrashanth Sreenivasa * All blocks that we need to read the most recent mapping must be 3185cabbc6bSPrashanth Sreenivasa * stored on concrete vdevs. Therefore, we must dirty anything that 3195cabbc6bSPrashanth Sreenivasa * is read before spa_remove_init(). Specifically, the 3205cabbc6bSPrashanth Sreenivasa * spa_config_object. (Note that although we already modified the 3215cabbc6bSPrashanth Sreenivasa * spa_config_object in spa_sync_removing_state, that may not have 3225cabbc6bSPrashanth Sreenivasa * modified all blocks of the object.) 3235cabbc6bSPrashanth Sreenivasa */ 3245cabbc6bSPrashanth Sreenivasa dmu_object_info_t doi; 3255cabbc6bSPrashanth Sreenivasa VERIFY0(dmu_object_info(mos, DMU_POOL_DIRECTORY_OBJECT, &doi)); 3265cabbc6bSPrashanth Sreenivasa for (uint64_t offset = 0; offset < doi.doi_max_offset; ) { 3275cabbc6bSPrashanth Sreenivasa dmu_buf_t *dbuf; 3285cabbc6bSPrashanth Sreenivasa VERIFY0(dmu_buf_hold(mos, DMU_POOL_DIRECTORY_OBJECT, 3295cabbc6bSPrashanth Sreenivasa offset, FTAG, &dbuf, 0)); 3305cabbc6bSPrashanth Sreenivasa dmu_buf_will_dirty(dbuf, tx); 3315cabbc6bSPrashanth Sreenivasa offset += dbuf->db_size; 3325cabbc6bSPrashanth Sreenivasa dmu_buf_rele(dbuf, FTAG); 3335cabbc6bSPrashanth Sreenivasa } 3345cabbc6bSPrashanth Sreenivasa 3355cabbc6bSPrashanth Sreenivasa /* 3365cabbc6bSPrashanth Sreenivasa * Now that we've allocated the im_object, dirty the vdev to ensure 3375cabbc6bSPrashanth Sreenivasa * that the object gets written to the config on disk. 3385cabbc6bSPrashanth Sreenivasa */ 3395cabbc6bSPrashanth Sreenivasa vdev_config_dirty(vd); 3405cabbc6bSPrashanth Sreenivasa 3415cabbc6bSPrashanth Sreenivasa zfs_dbgmsg("starting removal thread for vdev %llu (%p) in txg %llu " 3425cabbc6bSPrashanth Sreenivasa "im_obj=%llu", vd->vdev_id, vd, dmu_tx_get_txg(tx), 3435cabbc6bSPrashanth Sreenivasa vic->vic_mapping_object); 3445cabbc6bSPrashanth Sreenivasa 3455cabbc6bSPrashanth Sreenivasa spa_history_log_internal(spa, "vdev remove started", tx, 3465cabbc6bSPrashanth Sreenivasa "%s vdev %llu %s", spa_name(spa), vd->vdev_id, 3475cabbc6bSPrashanth Sreenivasa (vd->vdev_path != NULL) ? vd->vdev_path : "-"); 3485cabbc6bSPrashanth Sreenivasa /* 3495cabbc6bSPrashanth Sreenivasa * Setting spa_vdev_removal causes subsequent frees to call 3505cabbc6bSPrashanth Sreenivasa * free_from_removing_vdev(). Note that we don't need any locking 3515cabbc6bSPrashanth Sreenivasa * because we are the sync thread, and metaslab_free_impl() is only 3525cabbc6bSPrashanth Sreenivasa * called from syncing context (potentially from a zio taskq thread, 3535cabbc6bSPrashanth Sreenivasa * but in any case only when there are outstanding free i/os, which 3545cabbc6bSPrashanth Sreenivasa * there are not). 3555cabbc6bSPrashanth Sreenivasa */ 3565cabbc6bSPrashanth Sreenivasa ASSERT3P(spa->spa_vdev_removal, ==, NULL); 3575cabbc6bSPrashanth Sreenivasa spa->spa_vdev_removal = svr; 3585cabbc6bSPrashanth Sreenivasa svr->svr_thread = thread_create(NULL, 0, 3593a4b1be9SMatthew Ahrens spa_vdev_remove_thread, spa, 0, &p0, TS_RUN, minclsyspri); 3605cabbc6bSPrashanth Sreenivasa } 3615cabbc6bSPrashanth Sreenivasa 3625cabbc6bSPrashanth Sreenivasa /* 3635cabbc6bSPrashanth Sreenivasa * When we are opening a pool, we must read the mapping for each 3645cabbc6bSPrashanth Sreenivasa * indirect vdev in order from most recently removed to least 3655cabbc6bSPrashanth Sreenivasa * recently removed. We do this because the blocks for the mapping 3665cabbc6bSPrashanth Sreenivasa * of older indirect vdevs may be stored on more recently removed vdevs. 3675cabbc6bSPrashanth Sreenivasa * In order to read each indirect mapping object, we must have 3685cabbc6bSPrashanth Sreenivasa * initialized all more recently removed vdevs. 3695cabbc6bSPrashanth Sreenivasa */ 3705cabbc6bSPrashanth Sreenivasa int 3715cabbc6bSPrashanth Sreenivasa spa_remove_init(spa_t *spa) 3725cabbc6bSPrashanth Sreenivasa { 3735cabbc6bSPrashanth Sreenivasa int error; 3745cabbc6bSPrashanth Sreenivasa 3755cabbc6bSPrashanth Sreenivasa error = zap_lookup(spa->spa_dsl_pool->dp_meta_objset, 3765cabbc6bSPrashanth Sreenivasa DMU_POOL_DIRECTORY_OBJECT, 3775cabbc6bSPrashanth Sreenivasa DMU_POOL_REMOVING, sizeof (uint64_t), 3785cabbc6bSPrashanth Sreenivasa sizeof (spa->spa_removing_phys) / sizeof (uint64_t), 3795cabbc6bSPrashanth Sreenivasa &spa->spa_removing_phys); 3805cabbc6bSPrashanth Sreenivasa 3815cabbc6bSPrashanth Sreenivasa if (error == ENOENT) { 3825cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_state = DSS_NONE; 3835cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_removing_vdev = -1; 3845cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_prev_indirect_vdev = -1; 38547b8d4b8SAlexander Motin spa->spa_indirect_vdevs_loaded = B_TRUE; 3865cabbc6bSPrashanth Sreenivasa return (0); 3875cabbc6bSPrashanth Sreenivasa } else if (error != 0) { 3885cabbc6bSPrashanth Sreenivasa return (error); 3895cabbc6bSPrashanth Sreenivasa } 3905cabbc6bSPrashanth Sreenivasa 3915cabbc6bSPrashanth Sreenivasa if (spa->spa_removing_phys.sr_state == DSS_SCANNING) { 3925cabbc6bSPrashanth Sreenivasa /* 3935cabbc6bSPrashanth Sreenivasa * We are currently removing a vdev. Create and 3945cabbc6bSPrashanth Sreenivasa * initialize a spa_vdev_removal_t from the bonus 3955cabbc6bSPrashanth Sreenivasa * buffer of the removing vdevs vdev_im_object, and 3965cabbc6bSPrashanth Sreenivasa * initialize its partial mapping. 3975cabbc6bSPrashanth Sreenivasa */ 3985cabbc6bSPrashanth Sreenivasa spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 3995cabbc6bSPrashanth Sreenivasa vdev_t *vd = vdev_lookup_top(spa, 4005cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_removing_vdev); 4015cabbc6bSPrashanth Sreenivasa 4023a4b1be9SMatthew Ahrens if (vd == NULL) { 4033a4b1be9SMatthew Ahrens spa_config_exit(spa, SCL_STATE, FTAG); 4045cabbc6bSPrashanth Sreenivasa return (EINVAL); 4053a4b1be9SMatthew Ahrens } 4065cabbc6bSPrashanth Sreenivasa 4075cabbc6bSPrashanth Sreenivasa vdev_indirect_config_t *vic = &vd->vdev_indirect_config; 4085cabbc6bSPrashanth Sreenivasa 4095cabbc6bSPrashanth Sreenivasa ASSERT(vdev_is_concrete(vd)); 4105cabbc6bSPrashanth Sreenivasa spa_vdev_removal_t *svr = spa_vdev_removal_create(vd); 4113a4b1be9SMatthew Ahrens ASSERT3U(svr->svr_vdev_id, ==, vd->vdev_id); 4123a4b1be9SMatthew Ahrens ASSERT(vd->vdev_removing); 4135cabbc6bSPrashanth Sreenivasa 4145cabbc6bSPrashanth Sreenivasa vd->vdev_indirect_mapping = vdev_indirect_mapping_open( 4155cabbc6bSPrashanth Sreenivasa spa->spa_meta_objset, vic->vic_mapping_object); 4165cabbc6bSPrashanth Sreenivasa vd->vdev_indirect_births = vdev_indirect_births_open( 4175cabbc6bSPrashanth Sreenivasa spa->spa_meta_objset, vic->vic_births_object); 4183a4b1be9SMatthew Ahrens spa_config_exit(spa, SCL_STATE, FTAG); 4195cabbc6bSPrashanth Sreenivasa 4205cabbc6bSPrashanth Sreenivasa spa->spa_vdev_removal = svr; 4215cabbc6bSPrashanth Sreenivasa } 4225cabbc6bSPrashanth Sreenivasa 4235cabbc6bSPrashanth Sreenivasa spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 4245cabbc6bSPrashanth Sreenivasa uint64_t indirect_vdev_id = 4255cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_prev_indirect_vdev; 4265cabbc6bSPrashanth Sreenivasa while (indirect_vdev_id != UINT64_MAX) { 4275cabbc6bSPrashanth Sreenivasa vdev_t *vd = vdev_lookup_top(spa, indirect_vdev_id); 4285cabbc6bSPrashanth Sreenivasa vdev_indirect_config_t *vic = &vd->vdev_indirect_config; 4295cabbc6bSPrashanth Sreenivasa 4305cabbc6bSPrashanth Sreenivasa ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops); 4315cabbc6bSPrashanth Sreenivasa vd->vdev_indirect_mapping = vdev_indirect_mapping_open( 4325cabbc6bSPrashanth Sreenivasa spa->spa_meta_objset, vic->vic_mapping_object); 4335cabbc6bSPrashanth Sreenivasa vd->vdev_indirect_births = vdev_indirect_births_open( 4345cabbc6bSPrashanth Sreenivasa spa->spa_meta_objset, vic->vic_births_object); 4355cabbc6bSPrashanth Sreenivasa 4365cabbc6bSPrashanth Sreenivasa indirect_vdev_id = vic->vic_prev_indirect_vdev; 4375cabbc6bSPrashanth Sreenivasa } 4385cabbc6bSPrashanth Sreenivasa spa_config_exit(spa, SCL_STATE, FTAG); 4395cabbc6bSPrashanth Sreenivasa 4405cabbc6bSPrashanth Sreenivasa /* 4415cabbc6bSPrashanth Sreenivasa * Now that we've loaded all the indirect mappings, we can allow 4425cabbc6bSPrashanth Sreenivasa * reads from other blocks (e.g. via predictive prefetch). 4435cabbc6bSPrashanth Sreenivasa */ 4445cabbc6bSPrashanth Sreenivasa spa->spa_indirect_vdevs_loaded = B_TRUE; 4455cabbc6bSPrashanth Sreenivasa return (0); 4465cabbc6bSPrashanth Sreenivasa } 4475cabbc6bSPrashanth Sreenivasa 4485cabbc6bSPrashanth Sreenivasa void 4495cabbc6bSPrashanth Sreenivasa spa_restart_removal(spa_t *spa) 4505cabbc6bSPrashanth Sreenivasa { 4515cabbc6bSPrashanth Sreenivasa spa_vdev_removal_t *svr = spa->spa_vdev_removal; 4525cabbc6bSPrashanth Sreenivasa 4535cabbc6bSPrashanth Sreenivasa if (svr == NULL) 4545cabbc6bSPrashanth Sreenivasa return; 4555cabbc6bSPrashanth Sreenivasa 4565cabbc6bSPrashanth Sreenivasa /* 4575cabbc6bSPrashanth Sreenivasa * In general when this function is called there is no 4585cabbc6bSPrashanth Sreenivasa * removal thread running. The only scenario where this 4595cabbc6bSPrashanth Sreenivasa * is not true is during spa_import() where this function 4605cabbc6bSPrashanth Sreenivasa * is called twice [once from spa_import_impl() and 4615cabbc6bSPrashanth Sreenivasa * spa_async_resume()]. Thus, in the scenario where we 4625cabbc6bSPrashanth Sreenivasa * import a pool that has an ongoing removal we don't 4635cabbc6bSPrashanth Sreenivasa * want to spawn a second thread. 4645cabbc6bSPrashanth Sreenivasa */ 4655cabbc6bSPrashanth Sreenivasa if (svr->svr_thread != NULL) 4665cabbc6bSPrashanth Sreenivasa return; 4675cabbc6bSPrashanth Sreenivasa 4685cabbc6bSPrashanth Sreenivasa if (!spa_writeable(spa)) 4695cabbc6bSPrashanth Sreenivasa return; 4705cabbc6bSPrashanth Sreenivasa 4713a4b1be9SMatthew Ahrens zfs_dbgmsg("restarting removal of %llu", svr->svr_vdev_id); 4723a4b1be9SMatthew Ahrens svr->svr_thread = thread_create(NULL, 0, spa_vdev_remove_thread, spa, 4735cabbc6bSPrashanth Sreenivasa 0, &p0, TS_RUN, minclsyspri); 4745cabbc6bSPrashanth Sreenivasa } 4755cabbc6bSPrashanth Sreenivasa 4765cabbc6bSPrashanth Sreenivasa /* 4775cabbc6bSPrashanth Sreenivasa * Process freeing from a device which is in the middle of being removed. 4785cabbc6bSPrashanth Sreenivasa * We must handle this carefully so that we attempt to copy freed data, 4795cabbc6bSPrashanth Sreenivasa * and we correctly free already-copied data. 4805cabbc6bSPrashanth Sreenivasa */ 4815cabbc6bSPrashanth Sreenivasa void 48286714001SSerapheim Dimitropoulos free_from_removing_vdev(vdev_t *vd, uint64_t offset, uint64_t size) 4835cabbc6bSPrashanth Sreenivasa { 4845cabbc6bSPrashanth Sreenivasa spa_t *spa = vd->vdev_spa; 4855cabbc6bSPrashanth Sreenivasa spa_vdev_removal_t *svr = spa->spa_vdev_removal; 4865cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping; 48786714001SSerapheim Dimitropoulos uint64_t txg = spa_syncing_txg(spa); 4885cabbc6bSPrashanth Sreenivasa uint64_t max_offset_yet = 0; 4895cabbc6bSPrashanth Sreenivasa 4905cabbc6bSPrashanth Sreenivasa ASSERT(vd->vdev_indirect_config.vic_mapping_object != 0); 4915cabbc6bSPrashanth Sreenivasa ASSERT3U(vd->vdev_indirect_config.vic_mapping_object, ==, 4925cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_object(vim)); 4933a4b1be9SMatthew Ahrens ASSERT3U(vd->vdev_id, ==, svr->svr_vdev_id); 4945cabbc6bSPrashanth Sreenivasa 4955cabbc6bSPrashanth Sreenivasa mutex_enter(&svr->svr_lock); 4965cabbc6bSPrashanth Sreenivasa 4975cabbc6bSPrashanth Sreenivasa /* 4985cabbc6bSPrashanth Sreenivasa * Remove the segment from the removing vdev's spacemap. This 4995cabbc6bSPrashanth Sreenivasa * ensures that we will not attempt to copy this space (if the 5005cabbc6bSPrashanth Sreenivasa * removal thread has not yet visited it), and also ensures 5015cabbc6bSPrashanth Sreenivasa * that we know what is actually allocated on the new vdevs 5025cabbc6bSPrashanth Sreenivasa * (needed if we cancel the removal). 5035cabbc6bSPrashanth Sreenivasa * 5045cabbc6bSPrashanth Sreenivasa * Note: we must do the metaslab_free_concrete() with the svr_lock 5055cabbc6bSPrashanth Sreenivasa * held, so that the remove_thread can not load this metaslab and then 5065cabbc6bSPrashanth Sreenivasa * visit this offset between the time that we metaslab_free_concrete() 5075cabbc6bSPrashanth Sreenivasa * and when we check to see if it has been visited. 50886714001SSerapheim Dimitropoulos * 50986714001SSerapheim Dimitropoulos * Note: The checkpoint flag is set to false as having/taking 51086714001SSerapheim Dimitropoulos * a checkpoint and removing a device can't happen at the same 51186714001SSerapheim Dimitropoulos * time. 5125cabbc6bSPrashanth Sreenivasa */ 51386714001SSerapheim Dimitropoulos ASSERT(!spa_has_checkpoint(spa)); 51486714001SSerapheim Dimitropoulos metaslab_free_concrete(vd, offset, size, B_FALSE); 5155cabbc6bSPrashanth Sreenivasa 5165cabbc6bSPrashanth Sreenivasa uint64_t synced_size = 0; 5175cabbc6bSPrashanth Sreenivasa uint64_t synced_offset = 0; 5185cabbc6bSPrashanth Sreenivasa uint64_t max_offset_synced = vdev_indirect_mapping_max_offset(vim); 5195cabbc6bSPrashanth Sreenivasa if (offset < max_offset_synced) { 5205cabbc6bSPrashanth Sreenivasa /* 5215cabbc6bSPrashanth Sreenivasa * The mapping for this offset is already on disk. 5225cabbc6bSPrashanth Sreenivasa * Free from the new location. 5235cabbc6bSPrashanth Sreenivasa * 5245cabbc6bSPrashanth Sreenivasa * Note that we use svr_max_synced_offset because it is 5255cabbc6bSPrashanth Sreenivasa * updated atomically with respect to the in-core mapping. 5265cabbc6bSPrashanth Sreenivasa * By contrast, vim_max_offset is not. 5275cabbc6bSPrashanth Sreenivasa * 5285cabbc6bSPrashanth Sreenivasa * This block may be split between a synced entry and an 5295cabbc6bSPrashanth Sreenivasa * in-flight or unvisited entry. Only process the synced 5305cabbc6bSPrashanth Sreenivasa * portion of it here. 5315cabbc6bSPrashanth Sreenivasa */ 5325cabbc6bSPrashanth Sreenivasa synced_size = MIN(size, max_offset_synced - offset); 5335cabbc6bSPrashanth Sreenivasa synced_offset = offset; 5345cabbc6bSPrashanth Sreenivasa 5355cabbc6bSPrashanth Sreenivasa ASSERT3U(max_offset_yet, <=, max_offset_synced); 5365cabbc6bSPrashanth Sreenivasa max_offset_yet = max_offset_synced; 5375cabbc6bSPrashanth Sreenivasa 5385cabbc6bSPrashanth Sreenivasa DTRACE_PROBE3(remove__free__synced, 5395cabbc6bSPrashanth Sreenivasa spa_t *, spa, 5405cabbc6bSPrashanth Sreenivasa uint64_t, offset, 5415cabbc6bSPrashanth Sreenivasa uint64_t, synced_size); 5425cabbc6bSPrashanth Sreenivasa 5435cabbc6bSPrashanth Sreenivasa size -= synced_size; 5445cabbc6bSPrashanth Sreenivasa offset += synced_size; 5455cabbc6bSPrashanth Sreenivasa } 5465cabbc6bSPrashanth Sreenivasa 5475cabbc6bSPrashanth Sreenivasa /* 5485cabbc6bSPrashanth Sreenivasa * Look at all in-flight txgs starting from the currently syncing one 5495cabbc6bSPrashanth Sreenivasa * and see if a section of this free is being copied. By starting from 5505cabbc6bSPrashanth Sreenivasa * this txg and iterating forward, we might find that this region 5515cabbc6bSPrashanth Sreenivasa * was copied in two different txgs and handle it appropriately. 5525cabbc6bSPrashanth Sreenivasa */ 5535cabbc6bSPrashanth Sreenivasa for (int i = 0; i < TXG_CONCURRENT_STATES; i++) { 5545cabbc6bSPrashanth Sreenivasa int txgoff = (txg + i) & TXG_MASK; 5555cabbc6bSPrashanth Sreenivasa if (size > 0 && offset < svr->svr_max_offset_to_sync[txgoff]) { 5565cabbc6bSPrashanth Sreenivasa /* 5575cabbc6bSPrashanth Sreenivasa * The mapping for this offset is in flight, and 5585cabbc6bSPrashanth Sreenivasa * will be synced in txg+i. 5595cabbc6bSPrashanth Sreenivasa */ 5605cabbc6bSPrashanth Sreenivasa uint64_t inflight_size = MIN(size, 5615cabbc6bSPrashanth Sreenivasa svr->svr_max_offset_to_sync[txgoff] - offset); 5625cabbc6bSPrashanth Sreenivasa 5635cabbc6bSPrashanth Sreenivasa DTRACE_PROBE4(remove__free__inflight, 5645cabbc6bSPrashanth Sreenivasa spa_t *, spa, 5655cabbc6bSPrashanth Sreenivasa uint64_t, offset, 5665cabbc6bSPrashanth Sreenivasa uint64_t, inflight_size, 5675cabbc6bSPrashanth Sreenivasa uint64_t, txg + i); 5685cabbc6bSPrashanth Sreenivasa 5695cabbc6bSPrashanth Sreenivasa /* 5705cabbc6bSPrashanth Sreenivasa * We copy data in order of increasing offset. 5715cabbc6bSPrashanth Sreenivasa * Therefore the max_offset_to_sync[] must increase 5725cabbc6bSPrashanth Sreenivasa * (or be zero, indicating that nothing is being 5735cabbc6bSPrashanth Sreenivasa * copied in that txg). 5745cabbc6bSPrashanth Sreenivasa */ 5755cabbc6bSPrashanth Sreenivasa if (svr->svr_max_offset_to_sync[txgoff] != 0) { 5765cabbc6bSPrashanth Sreenivasa ASSERT3U(svr->svr_max_offset_to_sync[txgoff], 5775cabbc6bSPrashanth Sreenivasa >=, max_offset_yet); 5785cabbc6bSPrashanth Sreenivasa max_offset_yet = 5795cabbc6bSPrashanth Sreenivasa svr->svr_max_offset_to_sync[txgoff]; 5805cabbc6bSPrashanth Sreenivasa } 5815cabbc6bSPrashanth Sreenivasa 5825cabbc6bSPrashanth Sreenivasa /* 5835cabbc6bSPrashanth Sreenivasa * We've already committed to copying this segment: 5845cabbc6bSPrashanth Sreenivasa * we have allocated space elsewhere in the pool for 5855cabbc6bSPrashanth Sreenivasa * it and have an IO outstanding to copy the data. We 5865cabbc6bSPrashanth Sreenivasa * cannot free the space before the copy has 5875cabbc6bSPrashanth Sreenivasa * completed, or else the copy IO might overwrite any 5885cabbc6bSPrashanth Sreenivasa * new data. To free that space, we record the 5895cabbc6bSPrashanth Sreenivasa * segment in the appropriate svr_frees tree and free 5905cabbc6bSPrashanth Sreenivasa * the mapped space later, in the txg where we have 5915cabbc6bSPrashanth Sreenivasa * completed the copy and synced the mapping (see 5925cabbc6bSPrashanth Sreenivasa * vdev_mapping_sync). 5935cabbc6bSPrashanth Sreenivasa */ 5945cabbc6bSPrashanth Sreenivasa range_tree_add(svr->svr_frees[txgoff], 5955cabbc6bSPrashanth Sreenivasa offset, inflight_size); 5965cabbc6bSPrashanth Sreenivasa size -= inflight_size; 5975cabbc6bSPrashanth Sreenivasa offset += inflight_size; 5985cabbc6bSPrashanth Sreenivasa 5995cabbc6bSPrashanth Sreenivasa /* 6005cabbc6bSPrashanth Sreenivasa * This space is already accounted for as being 6015cabbc6bSPrashanth Sreenivasa * done, because it is being copied in txg+i. 6025cabbc6bSPrashanth Sreenivasa * However, if i!=0, then it is being copied in 6035cabbc6bSPrashanth Sreenivasa * a future txg. If we crash after this txg 6045cabbc6bSPrashanth Sreenivasa * syncs but before txg+i syncs, then the space 6055cabbc6bSPrashanth Sreenivasa * will be free. Therefore we must account 6065cabbc6bSPrashanth Sreenivasa * for the space being done in *this* txg 6075cabbc6bSPrashanth Sreenivasa * (when it is freed) rather than the future txg 6085cabbc6bSPrashanth Sreenivasa * (when it will be copied). 6095cabbc6bSPrashanth Sreenivasa */ 6105cabbc6bSPrashanth Sreenivasa ASSERT3U(svr->svr_bytes_done[txgoff], >=, 6115cabbc6bSPrashanth Sreenivasa inflight_size); 6125cabbc6bSPrashanth Sreenivasa svr->svr_bytes_done[txgoff] -= inflight_size; 6135cabbc6bSPrashanth Sreenivasa svr->svr_bytes_done[txg & TXG_MASK] += inflight_size; 6145cabbc6bSPrashanth Sreenivasa } 6155cabbc6bSPrashanth Sreenivasa } 6165cabbc6bSPrashanth Sreenivasa ASSERT0(svr->svr_max_offset_to_sync[TXG_CLEAN(txg) & TXG_MASK]); 6175cabbc6bSPrashanth Sreenivasa 6185cabbc6bSPrashanth Sreenivasa if (size > 0) { 6195cabbc6bSPrashanth Sreenivasa /* 6205cabbc6bSPrashanth Sreenivasa * The copy thread has not yet visited this offset. Ensure 6215cabbc6bSPrashanth Sreenivasa * that it doesn't. 6225cabbc6bSPrashanth Sreenivasa */ 6235cabbc6bSPrashanth Sreenivasa 6245cabbc6bSPrashanth Sreenivasa DTRACE_PROBE3(remove__free__unvisited, 6255cabbc6bSPrashanth Sreenivasa spa_t *, spa, 6265cabbc6bSPrashanth Sreenivasa uint64_t, offset, 6275cabbc6bSPrashanth Sreenivasa uint64_t, size); 6285cabbc6bSPrashanth Sreenivasa 6295cabbc6bSPrashanth Sreenivasa if (svr->svr_allocd_segs != NULL) 6305cabbc6bSPrashanth Sreenivasa range_tree_clear(svr->svr_allocd_segs, offset, size); 6315cabbc6bSPrashanth Sreenivasa 6325cabbc6bSPrashanth Sreenivasa /* 6335cabbc6bSPrashanth Sreenivasa * Since we now do not need to copy this data, for 6345cabbc6bSPrashanth Sreenivasa * accounting purposes we have done our job and can count 6355cabbc6bSPrashanth Sreenivasa * it as completed. 6365cabbc6bSPrashanth Sreenivasa */ 6375cabbc6bSPrashanth Sreenivasa svr->svr_bytes_done[txg & TXG_MASK] += size; 6385cabbc6bSPrashanth Sreenivasa } 6395cabbc6bSPrashanth Sreenivasa mutex_exit(&svr->svr_lock); 6405cabbc6bSPrashanth Sreenivasa 6415cabbc6bSPrashanth Sreenivasa /* 6425cabbc6bSPrashanth Sreenivasa * Now that we have dropped svr_lock, process the synced portion 6435cabbc6bSPrashanth Sreenivasa * of this free. 6445cabbc6bSPrashanth Sreenivasa */ 6455cabbc6bSPrashanth Sreenivasa if (synced_size > 0) { 64686714001SSerapheim Dimitropoulos vdev_indirect_mark_obsolete(vd, synced_offset, synced_size); 64786714001SSerapheim Dimitropoulos 6485cabbc6bSPrashanth Sreenivasa /* 6495cabbc6bSPrashanth Sreenivasa * Note: this can only be called from syncing context, 6505cabbc6bSPrashanth Sreenivasa * and the vdev_indirect_mapping is only changed from the 6515cabbc6bSPrashanth Sreenivasa * sync thread, so we don't need svr_lock while doing 6525cabbc6bSPrashanth Sreenivasa * metaslab_free_impl_cb. 6535cabbc6bSPrashanth Sreenivasa */ 65486714001SSerapheim Dimitropoulos boolean_t checkpoint = B_FALSE; 6555cabbc6bSPrashanth Sreenivasa vdev_indirect_ops.vdev_op_remap(vd, synced_offset, synced_size, 65686714001SSerapheim Dimitropoulos metaslab_free_impl_cb, &checkpoint); 6575cabbc6bSPrashanth Sreenivasa } 6585cabbc6bSPrashanth Sreenivasa } 6595cabbc6bSPrashanth Sreenivasa 6605cabbc6bSPrashanth Sreenivasa /* 6615cabbc6bSPrashanth Sreenivasa * Stop an active removal and update the spa_removing phys. 6625cabbc6bSPrashanth Sreenivasa */ 6635cabbc6bSPrashanth Sreenivasa static void 6645cabbc6bSPrashanth Sreenivasa spa_finish_removal(spa_t *spa, dsl_scan_state_t state, dmu_tx_t *tx) 6655cabbc6bSPrashanth Sreenivasa { 6665cabbc6bSPrashanth Sreenivasa spa_vdev_removal_t *svr = spa->spa_vdev_removal; 6675cabbc6bSPrashanth Sreenivasa ASSERT3U(dmu_tx_get_txg(tx), ==, spa_syncing_txg(spa)); 6685cabbc6bSPrashanth Sreenivasa 6695cabbc6bSPrashanth Sreenivasa /* Ensure the removal thread has completed before we free the svr. */ 6705cabbc6bSPrashanth Sreenivasa spa_vdev_remove_suspend(spa); 6715cabbc6bSPrashanth Sreenivasa 6725cabbc6bSPrashanth Sreenivasa ASSERT(state == DSS_FINISHED || state == DSS_CANCELED); 6735cabbc6bSPrashanth Sreenivasa 6745cabbc6bSPrashanth Sreenivasa if (state == DSS_FINISHED) { 6755cabbc6bSPrashanth Sreenivasa spa_removing_phys_t *srp = &spa->spa_removing_phys; 6763a4b1be9SMatthew Ahrens vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id); 6775cabbc6bSPrashanth Sreenivasa vdev_indirect_config_t *vic = &vd->vdev_indirect_config; 6785cabbc6bSPrashanth Sreenivasa 6795cabbc6bSPrashanth Sreenivasa if (srp->sr_prev_indirect_vdev != UINT64_MAX) { 6805cabbc6bSPrashanth Sreenivasa vdev_t *pvd = vdev_lookup_top(spa, 6815cabbc6bSPrashanth Sreenivasa srp->sr_prev_indirect_vdev); 6825cabbc6bSPrashanth Sreenivasa ASSERT3P(pvd->vdev_ops, ==, &vdev_indirect_ops); 6835cabbc6bSPrashanth Sreenivasa } 6845cabbc6bSPrashanth Sreenivasa 6855cabbc6bSPrashanth Sreenivasa vic->vic_prev_indirect_vdev = srp->sr_prev_indirect_vdev; 6865cabbc6bSPrashanth Sreenivasa srp->sr_prev_indirect_vdev = vd->vdev_id; 6875cabbc6bSPrashanth Sreenivasa } 6885cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_state = state; 6895cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_end_time = gethrestime_sec(); 6905cabbc6bSPrashanth Sreenivasa 6915cabbc6bSPrashanth Sreenivasa spa->spa_vdev_removal = NULL; 6925cabbc6bSPrashanth Sreenivasa spa_vdev_removal_destroy(svr); 6935cabbc6bSPrashanth Sreenivasa 6945cabbc6bSPrashanth Sreenivasa spa_sync_removing_state(spa, tx); 6955cabbc6bSPrashanth Sreenivasa 6965cabbc6bSPrashanth Sreenivasa vdev_config_dirty(spa->spa_root_vdev); 6975cabbc6bSPrashanth Sreenivasa } 6985cabbc6bSPrashanth Sreenivasa 6995cabbc6bSPrashanth Sreenivasa static void 7005cabbc6bSPrashanth Sreenivasa free_mapped_segment_cb(void *arg, uint64_t offset, uint64_t size) 7015cabbc6bSPrashanth Sreenivasa { 7025cabbc6bSPrashanth Sreenivasa vdev_t *vd = arg; 70386714001SSerapheim Dimitropoulos vdev_indirect_mark_obsolete(vd, offset, size); 70486714001SSerapheim Dimitropoulos boolean_t checkpoint = B_FALSE; 7055cabbc6bSPrashanth Sreenivasa vdev_indirect_ops.vdev_op_remap(vd, offset, size, 70686714001SSerapheim Dimitropoulos metaslab_free_impl_cb, &checkpoint); 7075cabbc6bSPrashanth Sreenivasa } 7085cabbc6bSPrashanth Sreenivasa 7095cabbc6bSPrashanth Sreenivasa /* 7105cabbc6bSPrashanth Sreenivasa * On behalf of the removal thread, syncs an incremental bit more of 7115cabbc6bSPrashanth Sreenivasa * the indirect mapping to disk and updates the in-memory mapping. 7125cabbc6bSPrashanth Sreenivasa * Called as a sync task in every txg that the removal thread makes progress. 7135cabbc6bSPrashanth Sreenivasa */ 7145cabbc6bSPrashanth Sreenivasa static void 7155cabbc6bSPrashanth Sreenivasa vdev_mapping_sync(void *arg, dmu_tx_t *tx) 7165cabbc6bSPrashanth Sreenivasa { 7175cabbc6bSPrashanth Sreenivasa spa_vdev_removal_t *svr = arg; 7185cabbc6bSPrashanth Sreenivasa spa_t *spa = dmu_tx_pool(tx)->dp_spa; 7193a4b1be9SMatthew Ahrens vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id); 7205cabbc6bSPrashanth Sreenivasa vdev_indirect_config_t *vic = &vd->vdev_indirect_config; 7215cabbc6bSPrashanth Sreenivasa uint64_t txg = dmu_tx_get_txg(tx); 7225cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping; 7235cabbc6bSPrashanth Sreenivasa 7245cabbc6bSPrashanth Sreenivasa ASSERT(vic->vic_mapping_object != 0); 7255cabbc6bSPrashanth Sreenivasa ASSERT3U(txg, ==, spa_syncing_txg(spa)); 7265cabbc6bSPrashanth Sreenivasa 7275cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_add_entries(vim, 7285cabbc6bSPrashanth Sreenivasa &svr->svr_new_segments[txg & TXG_MASK], tx); 7295cabbc6bSPrashanth Sreenivasa vdev_indirect_births_add_entry(vd->vdev_indirect_births, 7305cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_max_offset(vim), dmu_tx_get_txg(tx), tx); 7315cabbc6bSPrashanth Sreenivasa 7325cabbc6bSPrashanth Sreenivasa /* 7335cabbc6bSPrashanth Sreenivasa * Free the copied data for anything that was freed while the 7345cabbc6bSPrashanth Sreenivasa * mapping entries were in flight. 7355cabbc6bSPrashanth Sreenivasa */ 7365cabbc6bSPrashanth Sreenivasa mutex_enter(&svr->svr_lock); 7375cabbc6bSPrashanth Sreenivasa range_tree_vacate(svr->svr_frees[txg & TXG_MASK], 7385cabbc6bSPrashanth Sreenivasa free_mapped_segment_cb, vd); 7395cabbc6bSPrashanth Sreenivasa ASSERT3U(svr->svr_max_offset_to_sync[txg & TXG_MASK], >=, 7405cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_max_offset(vim)); 7415cabbc6bSPrashanth Sreenivasa svr->svr_max_offset_to_sync[txg & TXG_MASK] = 0; 7425cabbc6bSPrashanth Sreenivasa mutex_exit(&svr->svr_lock); 7435cabbc6bSPrashanth Sreenivasa 7445cabbc6bSPrashanth Sreenivasa spa_sync_removing_state(spa, tx); 7455cabbc6bSPrashanth Sreenivasa } 7465cabbc6bSPrashanth Sreenivasa 747*cfd63e1bSMatthew Ahrens typedef struct vdev_copy_segment_arg { 748*cfd63e1bSMatthew Ahrens spa_t *vcsa_spa; 749*cfd63e1bSMatthew Ahrens dva_t *vcsa_dest_dva; 750*cfd63e1bSMatthew Ahrens uint64_t vcsa_txg; 751*cfd63e1bSMatthew Ahrens range_tree_t *vcsa_obsolete_segs; 752*cfd63e1bSMatthew Ahrens } vdev_copy_segment_arg_t; 753*cfd63e1bSMatthew Ahrens 754*cfd63e1bSMatthew Ahrens static void 755*cfd63e1bSMatthew Ahrens unalloc_seg(void *arg, uint64_t start, uint64_t size) 756*cfd63e1bSMatthew Ahrens { 757*cfd63e1bSMatthew Ahrens vdev_copy_segment_arg_t *vcsa = arg; 758*cfd63e1bSMatthew Ahrens spa_t *spa = vcsa->vcsa_spa; 759*cfd63e1bSMatthew Ahrens blkptr_t bp = { 0 }; 760*cfd63e1bSMatthew Ahrens 761*cfd63e1bSMatthew Ahrens BP_SET_BIRTH(&bp, TXG_INITIAL, TXG_INITIAL); 762*cfd63e1bSMatthew Ahrens BP_SET_LSIZE(&bp, size); 763*cfd63e1bSMatthew Ahrens BP_SET_PSIZE(&bp, size); 764*cfd63e1bSMatthew Ahrens BP_SET_COMPRESS(&bp, ZIO_COMPRESS_OFF); 765*cfd63e1bSMatthew Ahrens BP_SET_CHECKSUM(&bp, ZIO_CHECKSUM_OFF); 766*cfd63e1bSMatthew Ahrens BP_SET_TYPE(&bp, DMU_OT_NONE); 767*cfd63e1bSMatthew Ahrens BP_SET_LEVEL(&bp, 0); 768*cfd63e1bSMatthew Ahrens BP_SET_DEDUP(&bp, 0); 769*cfd63e1bSMatthew Ahrens BP_SET_BYTEORDER(&bp, ZFS_HOST_BYTEORDER); 770*cfd63e1bSMatthew Ahrens 771*cfd63e1bSMatthew Ahrens DVA_SET_VDEV(&bp.blk_dva[0], DVA_GET_VDEV(vcsa->vcsa_dest_dva)); 772*cfd63e1bSMatthew Ahrens DVA_SET_OFFSET(&bp.blk_dva[0], 773*cfd63e1bSMatthew Ahrens DVA_GET_OFFSET(vcsa->vcsa_dest_dva) + start); 774*cfd63e1bSMatthew Ahrens DVA_SET_ASIZE(&bp.blk_dva[0], size); 775*cfd63e1bSMatthew Ahrens 776*cfd63e1bSMatthew Ahrens zio_free(spa, vcsa->vcsa_txg, &bp); 777*cfd63e1bSMatthew Ahrens } 778*cfd63e1bSMatthew Ahrens 7793a4b1be9SMatthew Ahrens /* 7803a4b1be9SMatthew Ahrens * All reads and writes associated with a call to spa_vdev_copy_segment() 7813a4b1be9SMatthew Ahrens * are done. 7823a4b1be9SMatthew Ahrens */ 7833a4b1be9SMatthew Ahrens static void 784*cfd63e1bSMatthew Ahrens spa_vdev_copy_segment_done(zio_t *zio) 7853a4b1be9SMatthew Ahrens { 786*cfd63e1bSMatthew Ahrens vdev_copy_segment_arg_t *vcsa = zio->io_private; 787*cfd63e1bSMatthew Ahrens 788*cfd63e1bSMatthew Ahrens range_tree_vacate(vcsa->vcsa_obsolete_segs, 789*cfd63e1bSMatthew Ahrens unalloc_seg, vcsa); 790*cfd63e1bSMatthew Ahrens range_tree_destroy(vcsa->vcsa_obsolete_segs); 791*cfd63e1bSMatthew Ahrens kmem_free(vcsa, sizeof (*vcsa)); 792*cfd63e1bSMatthew Ahrens 7933a4b1be9SMatthew Ahrens spa_config_exit(zio->io_spa, SCL_STATE, zio->io_spa); 7943a4b1be9SMatthew Ahrens } 7953a4b1be9SMatthew Ahrens 7963a4b1be9SMatthew Ahrens /* 7973a4b1be9SMatthew Ahrens * The write of the new location is done. 7983a4b1be9SMatthew Ahrens */ 7995cabbc6bSPrashanth Sreenivasa static void 8005cabbc6bSPrashanth Sreenivasa spa_vdev_copy_segment_write_done(zio_t *zio) 8015cabbc6bSPrashanth Sreenivasa { 8023a4b1be9SMatthew Ahrens vdev_copy_arg_t *vca = zio->io_private; 8033a4b1be9SMatthew Ahrens 8045cabbc6bSPrashanth Sreenivasa abd_free(zio->io_abd); 8055cabbc6bSPrashanth Sreenivasa 8065cabbc6bSPrashanth Sreenivasa mutex_enter(&vca->vca_lock); 8075cabbc6bSPrashanth Sreenivasa vca->vca_outstanding_bytes -= zio->io_size; 8085cabbc6bSPrashanth Sreenivasa cv_signal(&vca->vca_cv); 8095cabbc6bSPrashanth Sreenivasa mutex_exit(&vca->vca_lock); 8105cabbc6bSPrashanth Sreenivasa } 8115cabbc6bSPrashanth Sreenivasa 8123a4b1be9SMatthew Ahrens /* 8133a4b1be9SMatthew Ahrens * The read of the old location is done. The parent zio is the write to 8143a4b1be9SMatthew Ahrens * the new location. Allow it to start. 8153a4b1be9SMatthew Ahrens */ 8165cabbc6bSPrashanth Sreenivasa static void 8175cabbc6bSPrashanth Sreenivasa spa_vdev_copy_segment_read_done(zio_t *zio) 8185cabbc6bSPrashanth Sreenivasa { 8193a4b1be9SMatthew Ahrens zio_nowait(zio_unique_parent(zio)); 8205cabbc6bSPrashanth Sreenivasa } 8215cabbc6bSPrashanth Sreenivasa 8223a4b1be9SMatthew Ahrens /* 8233a4b1be9SMatthew Ahrens * If the old and new vdevs are mirrors, we will read both sides of the old 8243a4b1be9SMatthew Ahrens * mirror, and write each copy to the corresponding side of the new mirror. 8253a4b1be9SMatthew Ahrens * If the old and new vdevs have a different number of children, we will do 8263a4b1be9SMatthew Ahrens * this as best as possible. Since we aren't verifying checksums, this 8273a4b1be9SMatthew Ahrens * ensures that as long as there's a good copy of the data, we'll have a 8283a4b1be9SMatthew Ahrens * good copy after the removal, even if there's silent damage to one side 8293a4b1be9SMatthew Ahrens * of the mirror. If we're removing a mirror that has some silent damage, 8303a4b1be9SMatthew Ahrens * we'll have exactly the same damage in the new location (assuming that 8313a4b1be9SMatthew Ahrens * the new location is also a mirror). 8323a4b1be9SMatthew Ahrens * 8333a4b1be9SMatthew Ahrens * We accomplish this by creating a tree of zio_t's, with as many writes as 8343a4b1be9SMatthew Ahrens * there are "children" of the new vdev (a non-redundant vdev counts as one 8353a4b1be9SMatthew Ahrens * child, a 2-way mirror has 2 children, etc). Each write has an associated 8363a4b1be9SMatthew Ahrens * read from a child of the old vdev. Typically there will be the same 8373a4b1be9SMatthew Ahrens * number of children of the old and new vdevs. However, if there are more 8383a4b1be9SMatthew Ahrens * children of the new vdev, some child(ren) of the old vdev will be issued 8393a4b1be9SMatthew Ahrens * multiple reads. If there are more children of the old vdev, some copies 8403a4b1be9SMatthew Ahrens * will be dropped. 8413a4b1be9SMatthew Ahrens * 8423a4b1be9SMatthew Ahrens * For example, the tree of zio_t's for a 2-way mirror is: 8433a4b1be9SMatthew Ahrens * 8443a4b1be9SMatthew Ahrens * null 8453a4b1be9SMatthew Ahrens * / \ 8463a4b1be9SMatthew Ahrens * write(new vdev, child 0) write(new vdev, child 1) 8473a4b1be9SMatthew Ahrens * | | 8483a4b1be9SMatthew Ahrens * read(old vdev, child 0) read(old vdev, child 1) 8493a4b1be9SMatthew Ahrens * 8503a4b1be9SMatthew Ahrens * Child zio's complete before their parents complete. However, zio's 8513a4b1be9SMatthew Ahrens * created with zio_vdev_child_io() may be issued before their children 8523a4b1be9SMatthew Ahrens * complete. In this case we need to make sure that the children (reads) 8533a4b1be9SMatthew Ahrens * complete before the parents (writes) are *issued*. We do this by not 8543a4b1be9SMatthew Ahrens * calling zio_nowait() on each write until its corresponding read has 8553a4b1be9SMatthew Ahrens * completed. 8563a4b1be9SMatthew Ahrens * 8573a4b1be9SMatthew Ahrens * The spa_config_lock must be held while zio's created by 8583a4b1be9SMatthew Ahrens * zio_vdev_child_io() are in progress, to ensure that the vdev tree does 8593a4b1be9SMatthew Ahrens * not change (e.g. due to a concurrent "zpool attach/detach"). The "null" 8603a4b1be9SMatthew Ahrens * zio is needed to release the spa_config_lock after all the reads and 8613a4b1be9SMatthew Ahrens * writes complete. (Note that we can't grab the config lock for each read, 8623a4b1be9SMatthew Ahrens * because it is not reentrant - we could deadlock with a thread waiting 8633a4b1be9SMatthew Ahrens * for a write lock.) 8643a4b1be9SMatthew Ahrens */ 8653a4b1be9SMatthew Ahrens static void 8663a4b1be9SMatthew Ahrens spa_vdev_copy_one_child(vdev_copy_arg_t *vca, zio_t *nzio, 8673a4b1be9SMatthew Ahrens vdev_t *source_vd, uint64_t source_offset, 8683a4b1be9SMatthew Ahrens vdev_t *dest_child_vd, uint64_t dest_offset, int dest_id, uint64_t size) 8693a4b1be9SMatthew Ahrens { 8703a4b1be9SMatthew Ahrens ASSERT3U(spa_config_held(nzio->io_spa, SCL_ALL, RW_READER), !=, 0); 8713a4b1be9SMatthew Ahrens 8723a4b1be9SMatthew Ahrens mutex_enter(&vca->vca_lock); 8733a4b1be9SMatthew Ahrens vca->vca_outstanding_bytes += size; 8743a4b1be9SMatthew Ahrens mutex_exit(&vca->vca_lock); 8753a4b1be9SMatthew Ahrens 8763a4b1be9SMatthew Ahrens abd_t *abd = abd_alloc_for_io(size, B_FALSE); 8773a4b1be9SMatthew Ahrens 8783a4b1be9SMatthew Ahrens vdev_t *source_child_vd; 8793a4b1be9SMatthew Ahrens if (source_vd->vdev_ops == &vdev_mirror_ops && dest_id != -1) { 8803a4b1be9SMatthew Ahrens /* 8813a4b1be9SMatthew Ahrens * Source and dest are both mirrors. Copy from the same 8823a4b1be9SMatthew Ahrens * child id as we are copying to (wrapping around if there 8833a4b1be9SMatthew Ahrens * are more dest children than source children). 8843a4b1be9SMatthew Ahrens */ 8853a4b1be9SMatthew Ahrens source_child_vd = 8863a4b1be9SMatthew Ahrens source_vd->vdev_child[dest_id % source_vd->vdev_children]; 8873a4b1be9SMatthew Ahrens } else { 8883a4b1be9SMatthew Ahrens source_child_vd = source_vd; 8893a4b1be9SMatthew Ahrens } 8903a4b1be9SMatthew Ahrens 8913a4b1be9SMatthew Ahrens zio_t *write_zio = zio_vdev_child_io(nzio, NULL, 8923a4b1be9SMatthew Ahrens dest_child_vd, dest_offset, abd, size, 8933a4b1be9SMatthew Ahrens ZIO_TYPE_WRITE, ZIO_PRIORITY_REMOVAL, 8943a4b1be9SMatthew Ahrens ZIO_FLAG_CANFAIL, 8953a4b1be9SMatthew Ahrens spa_vdev_copy_segment_write_done, vca); 8963a4b1be9SMatthew Ahrens 8973a4b1be9SMatthew Ahrens zio_nowait(zio_vdev_child_io(write_zio, NULL, 8983a4b1be9SMatthew Ahrens source_child_vd, source_offset, abd, size, 8993a4b1be9SMatthew Ahrens ZIO_TYPE_READ, ZIO_PRIORITY_REMOVAL, 9003a4b1be9SMatthew Ahrens ZIO_FLAG_CANFAIL, 9013a4b1be9SMatthew Ahrens spa_vdev_copy_segment_read_done, vca)); 9023a4b1be9SMatthew Ahrens } 9033a4b1be9SMatthew Ahrens 9043a4b1be9SMatthew Ahrens /* 9053a4b1be9SMatthew Ahrens * Allocate a new location for this segment, and create the zio_t's to 9063a4b1be9SMatthew Ahrens * read from the old location and write to the new location. 9073a4b1be9SMatthew Ahrens */ 9085cabbc6bSPrashanth Sreenivasa static int 909*cfd63e1bSMatthew Ahrens spa_vdev_copy_segment(vdev_t *vd, range_tree_t *segs, 910*cfd63e1bSMatthew Ahrens uint64_t maxalloc, uint64_t txg, 9115cabbc6bSPrashanth Sreenivasa vdev_copy_arg_t *vca, zio_alloc_list_t *zal) 9125cabbc6bSPrashanth Sreenivasa { 9135cabbc6bSPrashanth Sreenivasa metaslab_group_t *mg = vd->vdev_mg; 9145cabbc6bSPrashanth Sreenivasa spa_t *spa = vd->vdev_spa; 9155cabbc6bSPrashanth Sreenivasa spa_vdev_removal_t *svr = spa->spa_vdev_removal; 9165cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_entry_t *entry; 9175cabbc6bSPrashanth Sreenivasa dva_t dst = { 0 }; 918*cfd63e1bSMatthew Ahrens uint64_t start = range_tree_min(segs); 9195cabbc6bSPrashanth Sreenivasa 920*cfd63e1bSMatthew Ahrens ASSERT3U(maxalloc, <=, SPA_MAXBLOCKSIZE); 921*cfd63e1bSMatthew Ahrens 922*cfd63e1bSMatthew Ahrens uint64_t size = range_tree_span(segs); 923*cfd63e1bSMatthew Ahrens if (range_tree_span(segs) > maxalloc) { 924*cfd63e1bSMatthew Ahrens /* 925*cfd63e1bSMatthew Ahrens * We can't allocate all the segments. Prefer to end 926*cfd63e1bSMatthew Ahrens * the allocation at the end of a segment, thus avoiding 927*cfd63e1bSMatthew Ahrens * additional split blocks. 928*cfd63e1bSMatthew Ahrens */ 929*cfd63e1bSMatthew Ahrens range_seg_t search; 930*cfd63e1bSMatthew Ahrens avl_index_t where; 931*cfd63e1bSMatthew Ahrens search.rs_start = start + maxalloc; 932*cfd63e1bSMatthew Ahrens search.rs_end = search.rs_start; 933*cfd63e1bSMatthew Ahrens range_seg_t *rs = avl_find(&segs->rt_root, &search, &where); 934*cfd63e1bSMatthew Ahrens if (rs == NULL) { 935*cfd63e1bSMatthew Ahrens rs = avl_nearest(&segs->rt_root, where, AVL_BEFORE); 936*cfd63e1bSMatthew Ahrens } else { 937*cfd63e1bSMatthew Ahrens rs = AVL_PREV(&segs->rt_root, rs); 938*cfd63e1bSMatthew Ahrens } 939*cfd63e1bSMatthew Ahrens if (rs != NULL) { 940*cfd63e1bSMatthew Ahrens size = rs->rs_end - start; 941*cfd63e1bSMatthew Ahrens } else { 942*cfd63e1bSMatthew Ahrens /* 943*cfd63e1bSMatthew Ahrens * There are no segments that end before maxalloc. 944*cfd63e1bSMatthew Ahrens * I.e. the first segment is larger than maxalloc, 945*cfd63e1bSMatthew Ahrens * so we must split it. 946*cfd63e1bSMatthew Ahrens */ 947*cfd63e1bSMatthew Ahrens size = maxalloc; 948*cfd63e1bSMatthew Ahrens } 949*cfd63e1bSMatthew Ahrens } 950*cfd63e1bSMatthew Ahrens ASSERT3U(size, <=, maxalloc); 9515cabbc6bSPrashanth Sreenivasa 952f78cdc34SPaul Dagnelie /* 953f78cdc34SPaul Dagnelie * We use allocator 0 for this I/O because we don't expect device remap 954f78cdc34SPaul Dagnelie * to be the steady state of the system, so parallelizing is not as 955f78cdc34SPaul Dagnelie * critical as it is for other allocation types. We also want to ensure 956f78cdc34SPaul Dagnelie * that the IOs are allocated together as much as possible, to reduce 957f78cdc34SPaul Dagnelie * mapping sizes. 958f78cdc34SPaul Dagnelie */ 9595cabbc6bSPrashanth Sreenivasa int error = metaslab_alloc_dva(spa, mg->mg_class, size, 960f78cdc34SPaul Dagnelie &dst, 0, NULL, txg, 0, zal, 0); 9615cabbc6bSPrashanth Sreenivasa if (error != 0) 9625cabbc6bSPrashanth Sreenivasa return (error); 9635cabbc6bSPrashanth Sreenivasa 964*cfd63e1bSMatthew Ahrens /* 965*cfd63e1bSMatthew Ahrens * Determine the ranges that are not actually needed. Offsets are 966*cfd63e1bSMatthew Ahrens * relative to the start of the range to be copied (i.e. relative to the 967*cfd63e1bSMatthew Ahrens * local variable "start"). 968*cfd63e1bSMatthew Ahrens */ 969*cfd63e1bSMatthew Ahrens range_tree_t *obsolete_segs = range_tree_create(NULL, NULL); 970*cfd63e1bSMatthew Ahrens 971*cfd63e1bSMatthew Ahrens range_seg_t *rs = avl_first(&segs->rt_root); 972*cfd63e1bSMatthew Ahrens ASSERT3U(rs->rs_start, ==, start); 973*cfd63e1bSMatthew Ahrens uint64_t prev_seg_end = rs->rs_end; 974*cfd63e1bSMatthew Ahrens while ((rs = AVL_NEXT(&segs->rt_root, rs)) != NULL) { 975*cfd63e1bSMatthew Ahrens if (rs->rs_start >= start + size) { 976*cfd63e1bSMatthew Ahrens break; 977*cfd63e1bSMatthew Ahrens } else { 978*cfd63e1bSMatthew Ahrens range_tree_add(obsolete_segs, 979*cfd63e1bSMatthew Ahrens prev_seg_end - start, 980*cfd63e1bSMatthew Ahrens rs->rs_start - prev_seg_end); 981*cfd63e1bSMatthew Ahrens } 982*cfd63e1bSMatthew Ahrens prev_seg_end = rs->rs_end; 983*cfd63e1bSMatthew Ahrens } 984*cfd63e1bSMatthew Ahrens /* We don't end in the middle of an obsolete range */ 985*cfd63e1bSMatthew Ahrens ASSERT3U(start + size, <=, prev_seg_end); 986*cfd63e1bSMatthew Ahrens 987*cfd63e1bSMatthew Ahrens range_tree_clear(segs, start, size); 988*cfd63e1bSMatthew Ahrens 9895cabbc6bSPrashanth Sreenivasa /* 9905cabbc6bSPrashanth Sreenivasa * We can't have any padding of the allocated size, otherwise we will 9915cabbc6bSPrashanth Sreenivasa * misunderstand what's allocated, and the size of the mapping. 9925cabbc6bSPrashanth Sreenivasa * The caller ensures this will be true by passing in a size that is 9935cabbc6bSPrashanth Sreenivasa * aligned to the worst (highest) ashift in the pool. 9945cabbc6bSPrashanth Sreenivasa */ 9955cabbc6bSPrashanth Sreenivasa ASSERT3U(DVA_GET_ASIZE(&dst), ==, size); 9965cabbc6bSPrashanth Sreenivasa 9975cabbc6bSPrashanth Sreenivasa entry = kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t), KM_SLEEP); 9985cabbc6bSPrashanth Sreenivasa DVA_MAPPING_SET_SRC_OFFSET(&entry->vime_mapping, start); 9995cabbc6bSPrashanth Sreenivasa entry->vime_mapping.vimep_dst = dst; 1000*cfd63e1bSMatthew Ahrens if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) { 1001*cfd63e1bSMatthew Ahrens entry->vime_obsolete_count = range_tree_space(obsolete_segs); 1002*cfd63e1bSMatthew Ahrens } 1003*cfd63e1bSMatthew Ahrens 1004*cfd63e1bSMatthew Ahrens vdev_copy_segment_arg_t *vcsa = kmem_zalloc(sizeof (*vcsa), KM_SLEEP); 1005*cfd63e1bSMatthew Ahrens vcsa->vcsa_dest_dva = &entry->vime_mapping.vimep_dst; 1006*cfd63e1bSMatthew Ahrens vcsa->vcsa_obsolete_segs = obsolete_segs; 1007*cfd63e1bSMatthew Ahrens vcsa->vcsa_spa = spa; 1008*cfd63e1bSMatthew Ahrens vcsa->vcsa_txg = txg; 10095cabbc6bSPrashanth Sreenivasa 10105cabbc6bSPrashanth Sreenivasa /* 10113a4b1be9SMatthew Ahrens * See comment before spa_vdev_copy_one_child(). 10125cabbc6bSPrashanth Sreenivasa */ 10133a4b1be9SMatthew Ahrens spa_config_enter(spa, SCL_STATE, spa, RW_READER); 10143a4b1be9SMatthew Ahrens zio_t *nzio = zio_null(spa->spa_txg_zio[txg & TXG_MASK], spa, NULL, 1015*cfd63e1bSMatthew Ahrens spa_vdev_copy_segment_done, vcsa, 0); 10163a4b1be9SMatthew Ahrens vdev_t *dest_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dst)); 10173a4b1be9SMatthew Ahrens if (dest_vd->vdev_ops == &vdev_mirror_ops) { 10183a4b1be9SMatthew Ahrens for (int i = 0; i < dest_vd->vdev_children; i++) { 10193a4b1be9SMatthew Ahrens vdev_t *child = dest_vd->vdev_child[i]; 10203a4b1be9SMatthew Ahrens spa_vdev_copy_one_child(vca, nzio, vd, start, 10213a4b1be9SMatthew Ahrens child, DVA_GET_OFFSET(&dst), i, size); 10223a4b1be9SMatthew Ahrens } 10233a4b1be9SMatthew Ahrens } else { 10243a4b1be9SMatthew Ahrens spa_vdev_copy_one_child(vca, nzio, vd, start, 10253a4b1be9SMatthew Ahrens dest_vd, DVA_GET_OFFSET(&dst), -1, size); 10263a4b1be9SMatthew Ahrens } 10273a4b1be9SMatthew Ahrens zio_nowait(nzio); 10285cabbc6bSPrashanth Sreenivasa 10295cabbc6bSPrashanth Sreenivasa list_insert_tail(&svr->svr_new_segments[txg & TXG_MASK], entry); 10305cabbc6bSPrashanth Sreenivasa ASSERT3U(start + size, <=, vd->vdev_ms_count << vd->vdev_ms_shift); 10315cabbc6bSPrashanth Sreenivasa vdev_dirty(vd, 0, NULL, txg); 10325cabbc6bSPrashanth Sreenivasa 10335cabbc6bSPrashanth Sreenivasa return (0); 10345cabbc6bSPrashanth Sreenivasa } 10355cabbc6bSPrashanth Sreenivasa 10365cabbc6bSPrashanth Sreenivasa /* 10375cabbc6bSPrashanth Sreenivasa * Complete the removal of a toplevel vdev. This is called as a 10385cabbc6bSPrashanth Sreenivasa * synctask in the same txg that we will sync out the new config (to the 10395cabbc6bSPrashanth Sreenivasa * MOS object) which indicates that this vdev is indirect. 10405cabbc6bSPrashanth Sreenivasa */ 10415cabbc6bSPrashanth Sreenivasa static void 10425cabbc6bSPrashanth Sreenivasa vdev_remove_complete_sync(void *arg, dmu_tx_t *tx) 10435cabbc6bSPrashanth Sreenivasa { 10445cabbc6bSPrashanth Sreenivasa spa_vdev_removal_t *svr = arg; 10453a4b1be9SMatthew Ahrens spa_t *spa = dmu_tx_pool(tx)->dp_spa; 10463a4b1be9SMatthew Ahrens vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id); 10475cabbc6bSPrashanth Sreenivasa 10485cabbc6bSPrashanth Sreenivasa ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops); 10495cabbc6bSPrashanth Sreenivasa 10505cabbc6bSPrashanth Sreenivasa for (int i = 0; i < TXG_SIZE; i++) { 10515cabbc6bSPrashanth Sreenivasa ASSERT0(svr->svr_bytes_done[i]); 10525cabbc6bSPrashanth Sreenivasa } 10535cabbc6bSPrashanth Sreenivasa 10545cabbc6bSPrashanth Sreenivasa ASSERT3U(spa->spa_removing_phys.sr_copied, ==, 10555cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_to_copy); 10565cabbc6bSPrashanth Sreenivasa 10575cabbc6bSPrashanth Sreenivasa vdev_destroy_spacemaps(vd, tx); 10585cabbc6bSPrashanth Sreenivasa 10595cabbc6bSPrashanth Sreenivasa /* destroy leaf zaps, if any */ 10605cabbc6bSPrashanth Sreenivasa ASSERT3P(svr->svr_zaplist, !=, NULL); 10615cabbc6bSPrashanth Sreenivasa for (nvpair_t *pair = nvlist_next_nvpair(svr->svr_zaplist, NULL); 10625cabbc6bSPrashanth Sreenivasa pair != NULL; 10635cabbc6bSPrashanth Sreenivasa pair = nvlist_next_nvpair(svr->svr_zaplist, pair)) { 10645cabbc6bSPrashanth Sreenivasa vdev_destroy_unlink_zap(vd, fnvpair_value_uint64(pair), tx); 10655cabbc6bSPrashanth Sreenivasa } 10665cabbc6bSPrashanth Sreenivasa fnvlist_free(svr->svr_zaplist); 10675cabbc6bSPrashanth Sreenivasa 10685cabbc6bSPrashanth Sreenivasa spa_finish_removal(dmu_tx_pool(tx)->dp_spa, DSS_FINISHED, tx); 10695cabbc6bSPrashanth Sreenivasa /* vd->vdev_path is not available here */ 10705cabbc6bSPrashanth Sreenivasa spa_history_log_internal(spa, "vdev remove completed", tx, 10715cabbc6bSPrashanth Sreenivasa "%s vdev %llu", spa_name(spa), vd->vdev_id); 10725cabbc6bSPrashanth Sreenivasa } 10735cabbc6bSPrashanth Sreenivasa 10745cabbc6bSPrashanth Sreenivasa static void 10755cabbc6bSPrashanth Sreenivasa vdev_remove_enlist_zaps(vdev_t *vd, nvlist_t *zlist) 10765cabbc6bSPrashanth Sreenivasa { 10775cabbc6bSPrashanth Sreenivasa ASSERT3P(zlist, !=, NULL); 10785cabbc6bSPrashanth Sreenivasa ASSERT3P(vd->vdev_ops, !=, &vdev_raidz_ops); 10795cabbc6bSPrashanth Sreenivasa 10805cabbc6bSPrashanth Sreenivasa if (vd->vdev_leaf_zap != 0) { 10815cabbc6bSPrashanth Sreenivasa char zkey[32]; 10825cabbc6bSPrashanth Sreenivasa (void) snprintf(zkey, sizeof (zkey), "%s-%"PRIu64, 10835cabbc6bSPrashanth Sreenivasa VDEV_REMOVAL_ZAP_OBJS, vd->vdev_leaf_zap); 10845cabbc6bSPrashanth Sreenivasa fnvlist_add_uint64(zlist, zkey, vd->vdev_leaf_zap); 10855cabbc6bSPrashanth Sreenivasa } 10865cabbc6bSPrashanth Sreenivasa 10875cabbc6bSPrashanth Sreenivasa for (uint64_t id = 0; id < vd->vdev_children; id++) { 10885cabbc6bSPrashanth Sreenivasa vdev_remove_enlist_zaps(vd->vdev_child[id], zlist); 10895cabbc6bSPrashanth Sreenivasa } 10905cabbc6bSPrashanth Sreenivasa } 10915cabbc6bSPrashanth Sreenivasa 10925cabbc6bSPrashanth Sreenivasa static void 10935cabbc6bSPrashanth Sreenivasa vdev_remove_replace_with_indirect(vdev_t *vd, uint64_t txg) 10945cabbc6bSPrashanth Sreenivasa { 10955cabbc6bSPrashanth Sreenivasa vdev_t *ivd; 10965cabbc6bSPrashanth Sreenivasa dmu_tx_t *tx; 10975cabbc6bSPrashanth Sreenivasa spa_t *spa = vd->vdev_spa; 10985cabbc6bSPrashanth Sreenivasa spa_vdev_removal_t *svr = spa->spa_vdev_removal; 10995cabbc6bSPrashanth Sreenivasa 11005cabbc6bSPrashanth Sreenivasa /* 11015cabbc6bSPrashanth Sreenivasa * First, build a list of leaf zaps to be destroyed. 11025cabbc6bSPrashanth Sreenivasa * This is passed to the sync context thread, 11035cabbc6bSPrashanth Sreenivasa * which does the actual unlinking. 11045cabbc6bSPrashanth Sreenivasa */ 11055cabbc6bSPrashanth Sreenivasa svr->svr_zaplist = fnvlist_alloc(); 11065cabbc6bSPrashanth Sreenivasa vdev_remove_enlist_zaps(vd, svr->svr_zaplist); 11075cabbc6bSPrashanth Sreenivasa 11085cabbc6bSPrashanth Sreenivasa ivd = vdev_add_parent(vd, &vdev_indirect_ops); 11093a4b1be9SMatthew Ahrens ivd->vdev_removing = 0; 11105cabbc6bSPrashanth Sreenivasa 11115cabbc6bSPrashanth Sreenivasa vd->vdev_leaf_zap = 0; 11125cabbc6bSPrashanth Sreenivasa 11135cabbc6bSPrashanth Sreenivasa vdev_remove_child(ivd, vd); 11145cabbc6bSPrashanth Sreenivasa vdev_compact_children(ivd); 11155cabbc6bSPrashanth Sreenivasa 11165cabbc6bSPrashanth Sreenivasa ASSERT(!list_link_active(&vd->vdev_state_dirty_node)); 11175cabbc6bSPrashanth Sreenivasa 11185cabbc6bSPrashanth Sreenivasa tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg); 11195cabbc6bSPrashanth Sreenivasa dsl_sync_task_nowait(spa->spa_dsl_pool, vdev_remove_complete_sync, svr, 11205cabbc6bSPrashanth Sreenivasa 0, ZFS_SPACE_CHECK_NONE, tx); 11215cabbc6bSPrashanth Sreenivasa dmu_tx_commit(tx); 11225cabbc6bSPrashanth Sreenivasa 11235cabbc6bSPrashanth Sreenivasa /* 11245cabbc6bSPrashanth Sreenivasa * Indicate that this thread has exited. 11255cabbc6bSPrashanth Sreenivasa * After this, we can not use svr. 11265cabbc6bSPrashanth Sreenivasa */ 11275cabbc6bSPrashanth Sreenivasa mutex_enter(&svr->svr_lock); 11285cabbc6bSPrashanth Sreenivasa svr->svr_thread = NULL; 11295cabbc6bSPrashanth Sreenivasa cv_broadcast(&svr->svr_cv); 11305cabbc6bSPrashanth Sreenivasa mutex_exit(&svr->svr_lock); 11315cabbc6bSPrashanth Sreenivasa } 11325cabbc6bSPrashanth Sreenivasa 11335cabbc6bSPrashanth Sreenivasa /* 11345cabbc6bSPrashanth Sreenivasa * Complete the removal of a toplevel vdev. This is called in open 11355cabbc6bSPrashanth Sreenivasa * context by the removal thread after we have copied all vdev's data. 11365cabbc6bSPrashanth Sreenivasa */ 11375cabbc6bSPrashanth Sreenivasa static void 11383a4b1be9SMatthew Ahrens vdev_remove_complete(spa_t *spa) 11395cabbc6bSPrashanth Sreenivasa { 11405cabbc6bSPrashanth Sreenivasa uint64_t txg; 11415cabbc6bSPrashanth Sreenivasa 11425cabbc6bSPrashanth Sreenivasa /* 11435cabbc6bSPrashanth Sreenivasa * Wait for any deferred frees to be synced before we call 11445cabbc6bSPrashanth Sreenivasa * vdev_metaslab_fini() 11455cabbc6bSPrashanth Sreenivasa */ 11465cabbc6bSPrashanth Sreenivasa txg_wait_synced(spa->spa_dsl_pool, 0); 11475cabbc6bSPrashanth Sreenivasa txg = spa_vdev_enter(spa); 11483a4b1be9SMatthew Ahrens vdev_t *vd = vdev_lookup_top(spa, spa->spa_vdev_removal->svr_vdev_id); 1149094e47e9SGeorge Wilson ASSERT3P(vd->vdev_initialize_thread, ==, NULL); 11503a4b1be9SMatthew Ahrens 11513a4b1be9SMatthew Ahrens sysevent_t *ev = spa_event_create(spa, vd, NULL, 11523a4b1be9SMatthew Ahrens ESC_ZFS_VDEV_REMOVE_DEV); 11533a4b1be9SMatthew Ahrens 11545cabbc6bSPrashanth Sreenivasa zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu", 11555cabbc6bSPrashanth Sreenivasa vd->vdev_id, txg); 11565cabbc6bSPrashanth Sreenivasa 11575cabbc6bSPrashanth Sreenivasa /* 11585cabbc6bSPrashanth Sreenivasa * Discard allocation state. 11595cabbc6bSPrashanth Sreenivasa */ 11605cabbc6bSPrashanth Sreenivasa if (vd->vdev_mg != NULL) { 11615cabbc6bSPrashanth Sreenivasa vdev_metaslab_fini(vd); 11625cabbc6bSPrashanth Sreenivasa metaslab_group_destroy(vd->vdev_mg); 11635cabbc6bSPrashanth Sreenivasa vd->vdev_mg = NULL; 11645cabbc6bSPrashanth Sreenivasa } 11655cabbc6bSPrashanth Sreenivasa ASSERT0(vd->vdev_stat.vs_space); 11665cabbc6bSPrashanth Sreenivasa ASSERT0(vd->vdev_stat.vs_dspace); 11675cabbc6bSPrashanth Sreenivasa 11685cabbc6bSPrashanth Sreenivasa vdev_remove_replace_with_indirect(vd, txg); 11695cabbc6bSPrashanth Sreenivasa 11705cabbc6bSPrashanth Sreenivasa /* 11715cabbc6bSPrashanth Sreenivasa * We now release the locks, allowing spa_sync to run and finish the 11725cabbc6bSPrashanth Sreenivasa * removal via vdev_remove_complete_sync in syncing context. 11733a4b1be9SMatthew Ahrens * 11743a4b1be9SMatthew Ahrens * Note that we hold on to the vdev_t that has been replaced. Since 11753a4b1be9SMatthew Ahrens * it isn't part of the vdev tree any longer, it can't be concurrently 11763a4b1be9SMatthew Ahrens * manipulated, even while we don't have the config lock. 11775cabbc6bSPrashanth Sreenivasa */ 11785cabbc6bSPrashanth Sreenivasa (void) spa_vdev_exit(spa, NULL, txg, 0); 11795cabbc6bSPrashanth Sreenivasa 11805cabbc6bSPrashanth Sreenivasa /* 11815cabbc6bSPrashanth Sreenivasa * Top ZAP should have been transferred to the indirect vdev in 11825cabbc6bSPrashanth Sreenivasa * vdev_remove_replace_with_indirect. 11835cabbc6bSPrashanth Sreenivasa */ 11845cabbc6bSPrashanth Sreenivasa ASSERT0(vd->vdev_top_zap); 11855cabbc6bSPrashanth Sreenivasa 11865cabbc6bSPrashanth Sreenivasa /* 11875cabbc6bSPrashanth Sreenivasa * Leaf ZAP should have been moved in vdev_remove_replace_with_indirect. 11885cabbc6bSPrashanth Sreenivasa */ 11895cabbc6bSPrashanth Sreenivasa ASSERT0(vd->vdev_leaf_zap); 11905cabbc6bSPrashanth Sreenivasa 11915cabbc6bSPrashanth Sreenivasa txg = spa_vdev_enter(spa); 11925cabbc6bSPrashanth Sreenivasa (void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE); 11935cabbc6bSPrashanth Sreenivasa /* 11945cabbc6bSPrashanth Sreenivasa * Request to update the config and the config cachefile. 11955cabbc6bSPrashanth Sreenivasa */ 11965cabbc6bSPrashanth Sreenivasa vdev_config_dirty(spa->spa_root_vdev); 11975cabbc6bSPrashanth Sreenivasa (void) spa_vdev_exit(spa, vd, txg, 0); 11983a4b1be9SMatthew Ahrens 11993a4b1be9SMatthew Ahrens spa_event_post(ev); 12005cabbc6bSPrashanth Sreenivasa } 12015cabbc6bSPrashanth Sreenivasa 12025cabbc6bSPrashanth Sreenivasa /* 12035cabbc6bSPrashanth Sreenivasa * Evacuates a segment of size at most max_alloc from the vdev 12045cabbc6bSPrashanth Sreenivasa * via repeated calls to spa_vdev_copy_segment. If an allocation 12055cabbc6bSPrashanth Sreenivasa * fails, the pool is probably too fragmented to handle such a 12065cabbc6bSPrashanth Sreenivasa * large size, so decrease max_alloc so that the caller will not try 12075cabbc6bSPrashanth Sreenivasa * this size again this txg. 12085cabbc6bSPrashanth Sreenivasa */ 12095cabbc6bSPrashanth Sreenivasa static void 12103a4b1be9SMatthew Ahrens spa_vdev_copy_impl(vdev_t *vd, spa_vdev_removal_t *svr, vdev_copy_arg_t *vca, 12115cabbc6bSPrashanth Sreenivasa uint64_t *max_alloc, dmu_tx_t *tx) 12125cabbc6bSPrashanth Sreenivasa { 12135cabbc6bSPrashanth Sreenivasa uint64_t txg = dmu_tx_get_txg(tx); 12145cabbc6bSPrashanth Sreenivasa spa_t *spa = dmu_tx_pool(tx)->dp_spa; 12155cabbc6bSPrashanth Sreenivasa 12165cabbc6bSPrashanth Sreenivasa mutex_enter(&svr->svr_lock); 12175cabbc6bSPrashanth Sreenivasa 1218*cfd63e1bSMatthew Ahrens /* 1219*cfd63e1bSMatthew Ahrens * Determine how big of a chunk to copy. We can allocate up 1220*cfd63e1bSMatthew Ahrens * to max_alloc bytes, and we can span up to vdev_removal_max_span 1221*cfd63e1bSMatthew Ahrens * bytes of unallocated space at a time. "segs" will track the 1222*cfd63e1bSMatthew Ahrens * allocated segments that we are copying. We may also be copying 1223*cfd63e1bSMatthew Ahrens * free segments (of up to vdev_removal_max_span bytes). 1224*cfd63e1bSMatthew Ahrens */ 1225*cfd63e1bSMatthew Ahrens range_tree_t *segs = range_tree_create(NULL, NULL); 1226*cfd63e1bSMatthew Ahrens for (;;) { 1227*cfd63e1bSMatthew Ahrens range_seg_t *rs = avl_first(&svr->svr_allocd_segs->rt_root); 1228*cfd63e1bSMatthew Ahrens if (rs == NULL) 1229*cfd63e1bSMatthew Ahrens break; 1230*cfd63e1bSMatthew Ahrens 1231*cfd63e1bSMatthew Ahrens uint64_t seg_length; 1232*cfd63e1bSMatthew Ahrens 1233*cfd63e1bSMatthew Ahrens if (range_tree_is_empty(segs)) { 1234*cfd63e1bSMatthew Ahrens /* need to truncate the first seg based on max_alloc */ 1235*cfd63e1bSMatthew Ahrens seg_length = 1236*cfd63e1bSMatthew Ahrens MIN(rs->rs_end - rs->rs_start, *max_alloc); 1237*cfd63e1bSMatthew Ahrens } else { 1238*cfd63e1bSMatthew Ahrens if (rs->rs_start - range_tree_max(segs) > 1239*cfd63e1bSMatthew Ahrens vdev_removal_max_span) { 1240*cfd63e1bSMatthew Ahrens /* 1241*cfd63e1bSMatthew Ahrens * Including this segment would cause us to 1242*cfd63e1bSMatthew Ahrens * copy a larger unneeded chunk than is allowed. 1243*cfd63e1bSMatthew Ahrens */ 1244*cfd63e1bSMatthew Ahrens break; 1245*cfd63e1bSMatthew Ahrens } else if (rs->rs_end - range_tree_min(segs) > 1246*cfd63e1bSMatthew Ahrens *max_alloc) { 1247*cfd63e1bSMatthew Ahrens /* 1248*cfd63e1bSMatthew Ahrens * This additional segment would extend past 1249*cfd63e1bSMatthew Ahrens * max_alloc. Rather than splitting this 1250*cfd63e1bSMatthew Ahrens * segment, leave it for the next mapping. 1251*cfd63e1bSMatthew Ahrens */ 1252*cfd63e1bSMatthew Ahrens break; 1253*cfd63e1bSMatthew Ahrens } else { 1254*cfd63e1bSMatthew Ahrens seg_length = rs->rs_end - rs->rs_start; 1255*cfd63e1bSMatthew Ahrens } 1256*cfd63e1bSMatthew Ahrens } 1257*cfd63e1bSMatthew Ahrens 1258*cfd63e1bSMatthew Ahrens range_tree_add(segs, rs->rs_start, seg_length); 1259*cfd63e1bSMatthew Ahrens range_tree_remove(svr->svr_allocd_segs, 1260*cfd63e1bSMatthew Ahrens rs->rs_start, seg_length); 1261*cfd63e1bSMatthew Ahrens } 1262*cfd63e1bSMatthew Ahrens 1263*cfd63e1bSMatthew Ahrens if (range_tree_is_empty(segs)) { 12645cabbc6bSPrashanth Sreenivasa mutex_exit(&svr->svr_lock); 1265*cfd63e1bSMatthew Ahrens range_tree_destroy(segs); 12665cabbc6bSPrashanth Sreenivasa return; 12675cabbc6bSPrashanth Sreenivasa } 12685cabbc6bSPrashanth Sreenivasa 12695cabbc6bSPrashanth Sreenivasa if (svr->svr_max_offset_to_sync[txg & TXG_MASK] == 0) { 12705cabbc6bSPrashanth Sreenivasa dsl_sync_task_nowait(dmu_tx_pool(tx), vdev_mapping_sync, 12715cabbc6bSPrashanth Sreenivasa svr, 0, ZFS_SPACE_CHECK_NONE, tx); 12725cabbc6bSPrashanth Sreenivasa } 12735cabbc6bSPrashanth Sreenivasa 1274*cfd63e1bSMatthew Ahrens svr->svr_max_offset_to_sync[txg & TXG_MASK] = range_tree_max(segs); 12755cabbc6bSPrashanth Sreenivasa 12765cabbc6bSPrashanth Sreenivasa /* 12775cabbc6bSPrashanth Sreenivasa * Note: this is the amount of *allocated* space 12785cabbc6bSPrashanth Sreenivasa * that we are taking care of each txg. 12795cabbc6bSPrashanth Sreenivasa */ 1280*cfd63e1bSMatthew Ahrens svr->svr_bytes_done[txg & TXG_MASK] += range_tree_space(segs); 12815cabbc6bSPrashanth Sreenivasa 12825cabbc6bSPrashanth Sreenivasa mutex_exit(&svr->svr_lock); 12835cabbc6bSPrashanth Sreenivasa 12845cabbc6bSPrashanth Sreenivasa zio_alloc_list_t zal; 12855cabbc6bSPrashanth Sreenivasa metaslab_trace_init(&zal); 1286*cfd63e1bSMatthew Ahrens uint64_t thismax = SPA_MAXBLOCKSIZE; 1287*cfd63e1bSMatthew Ahrens while (!range_tree_is_empty(segs)) { 12883a4b1be9SMatthew Ahrens int error = spa_vdev_copy_segment(vd, 1289*cfd63e1bSMatthew Ahrens segs, thismax, txg, vca, &zal); 12905cabbc6bSPrashanth Sreenivasa 12915cabbc6bSPrashanth Sreenivasa if (error == ENOSPC) { 12925cabbc6bSPrashanth Sreenivasa /* 12935cabbc6bSPrashanth Sreenivasa * Cut our segment in half, and don't try this 12945cabbc6bSPrashanth Sreenivasa * segment size again this txg. Note that the 12955cabbc6bSPrashanth Sreenivasa * allocation size must be aligned to the highest 12965cabbc6bSPrashanth Sreenivasa * ashift in the pool, so that the allocation will 12975cabbc6bSPrashanth Sreenivasa * not be padded out to a multiple of the ashift, 12985cabbc6bSPrashanth Sreenivasa * which could cause us to think that this mapping 12995cabbc6bSPrashanth Sreenivasa * is larger than we intended. 13005cabbc6bSPrashanth Sreenivasa */ 13015cabbc6bSPrashanth Sreenivasa ASSERT3U(spa->spa_max_ashift, >=, SPA_MINBLOCKSHIFT); 13025cabbc6bSPrashanth Sreenivasa ASSERT3U(spa->spa_max_ashift, ==, spa->spa_min_ashift); 1303*cfd63e1bSMatthew Ahrens uint64_t attempted = 1304*cfd63e1bSMatthew Ahrens MIN(range_tree_span(segs), thismax); 1305*cfd63e1bSMatthew Ahrens thismax = P2ROUNDUP(attempted / 2, 13065cabbc6bSPrashanth Sreenivasa 1 << spa->spa_max_ashift); 13075cabbc6bSPrashanth Sreenivasa /* 13085cabbc6bSPrashanth Sreenivasa * The minimum-size allocation can not fail. 13095cabbc6bSPrashanth Sreenivasa */ 1310*cfd63e1bSMatthew Ahrens ASSERT3U(attempted, >, 1 << spa->spa_max_ashift); 1311*cfd63e1bSMatthew Ahrens *max_alloc = attempted - (1 << spa->spa_max_ashift); 13125cabbc6bSPrashanth Sreenivasa } else { 13135cabbc6bSPrashanth Sreenivasa ASSERT0(error); 13145cabbc6bSPrashanth Sreenivasa 13155cabbc6bSPrashanth Sreenivasa /* 13165cabbc6bSPrashanth Sreenivasa * We've performed an allocation, so reset the 13175cabbc6bSPrashanth Sreenivasa * alloc trace list. 13185cabbc6bSPrashanth Sreenivasa */ 13195cabbc6bSPrashanth Sreenivasa metaslab_trace_fini(&zal); 13205cabbc6bSPrashanth Sreenivasa metaslab_trace_init(&zal); 13215cabbc6bSPrashanth Sreenivasa } 13225cabbc6bSPrashanth Sreenivasa } 13235cabbc6bSPrashanth Sreenivasa metaslab_trace_fini(&zal); 1324*cfd63e1bSMatthew Ahrens range_tree_destroy(segs); 13255cabbc6bSPrashanth Sreenivasa } 13265cabbc6bSPrashanth Sreenivasa 13275cabbc6bSPrashanth Sreenivasa /* 13285cabbc6bSPrashanth Sreenivasa * The removal thread operates in open context. It iterates over all 13295cabbc6bSPrashanth Sreenivasa * allocated space in the vdev, by loading each metaslab's spacemap. 13305cabbc6bSPrashanth Sreenivasa * For each contiguous segment of allocated space (capping the segment 13315cabbc6bSPrashanth Sreenivasa * size at SPA_MAXBLOCKSIZE), we: 13325cabbc6bSPrashanth Sreenivasa * - Allocate space for it on another vdev. 13335cabbc6bSPrashanth Sreenivasa * - Create a new mapping from the old location to the new location 13345cabbc6bSPrashanth Sreenivasa * (as a record in svr_new_segments). 13355cabbc6bSPrashanth Sreenivasa * - Initiate a logical read zio to get the data off the removing disk. 13365cabbc6bSPrashanth Sreenivasa * - In the read zio's done callback, initiate a logical write zio to 13375cabbc6bSPrashanth Sreenivasa * write it to the new vdev. 13385cabbc6bSPrashanth Sreenivasa * Note that all of this will take effect when a particular TXG syncs. 13395cabbc6bSPrashanth Sreenivasa * The sync thread ensures that all the phys reads and writes for the syncing 13405cabbc6bSPrashanth Sreenivasa * TXG have completed (see spa_txg_zio) and writes the new mappings to disk 13415cabbc6bSPrashanth Sreenivasa * (see vdev_mapping_sync()). 13425cabbc6bSPrashanth Sreenivasa */ 13435cabbc6bSPrashanth Sreenivasa static void 13445cabbc6bSPrashanth Sreenivasa spa_vdev_remove_thread(void *arg) 13455cabbc6bSPrashanth Sreenivasa { 13463a4b1be9SMatthew Ahrens spa_t *spa = arg; 13475cabbc6bSPrashanth Sreenivasa spa_vdev_removal_t *svr = spa->spa_vdev_removal; 13485cabbc6bSPrashanth Sreenivasa vdev_copy_arg_t vca; 13495cabbc6bSPrashanth Sreenivasa uint64_t max_alloc = zfs_remove_max_segment; 13505cabbc6bSPrashanth Sreenivasa uint64_t last_txg = 0; 13513a4b1be9SMatthew Ahrens 13523a4b1be9SMatthew Ahrens spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 13533a4b1be9SMatthew Ahrens vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id); 13545cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping; 13555cabbc6bSPrashanth Sreenivasa uint64_t start_offset = vdev_indirect_mapping_max_offset(vim); 13565cabbc6bSPrashanth Sreenivasa 13575cabbc6bSPrashanth Sreenivasa ASSERT3P(vd->vdev_ops, !=, &vdev_indirect_ops); 13585cabbc6bSPrashanth Sreenivasa ASSERT(vdev_is_concrete(vd)); 13595cabbc6bSPrashanth Sreenivasa ASSERT(vd->vdev_removing); 13605cabbc6bSPrashanth Sreenivasa ASSERT(vd->vdev_indirect_config.vic_mapping_object != 0); 13615cabbc6bSPrashanth Sreenivasa ASSERT(vim != NULL); 13625cabbc6bSPrashanth Sreenivasa 13635cabbc6bSPrashanth Sreenivasa mutex_init(&vca.vca_lock, NULL, MUTEX_DEFAULT, NULL); 13645cabbc6bSPrashanth Sreenivasa cv_init(&vca.vca_cv, NULL, CV_DEFAULT, NULL); 13655cabbc6bSPrashanth Sreenivasa vca.vca_outstanding_bytes = 0; 13665cabbc6bSPrashanth Sreenivasa 13675cabbc6bSPrashanth Sreenivasa mutex_enter(&svr->svr_lock); 13685cabbc6bSPrashanth Sreenivasa 13695cabbc6bSPrashanth Sreenivasa /* 13705cabbc6bSPrashanth Sreenivasa * Start from vim_max_offset so we pick up where we left off 13715cabbc6bSPrashanth Sreenivasa * if we are restarting the removal after opening the pool. 13725cabbc6bSPrashanth Sreenivasa */ 13735cabbc6bSPrashanth Sreenivasa uint64_t msi; 13745cabbc6bSPrashanth Sreenivasa for (msi = start_offset >> vd->vdev_ms_shift; 13755cabbc6bSPrashanth Sreenivasa msi < vd->vdev_ms_count && !svr->svr_thread_exit; msi++) { 13765cabbc6bSPrashanth Sreenivasa metaslab_t *msp = vd->vdev_ms[msi]; 13775cabbc6bSPrashanth Sreenivasa ASSERT3U(msi, <=, vd->vdev_ms_count); 13785cabbc6bSPrashanth Sreenivasa 13795cabbc6bSPrashanth Sreenivasa ASSERT0(range_tree_space(svr->svr_allocd_segs)); 13805cabbc6bSPrashanth Sreenivasa 13815cabbc6bSPrashanth Sreenivasa mutex_enter(&msp->ms_sync_lock); 13825cabbc6bSPrashanth Sreenivasa mutex_enter(&msp->ms_lock); 13835cabbc6bSPrashanth Sreenivasa 13845cabbc6bSPrashanth Sreenivasa /* 13855cabbc6bSPrashanth Sreenivasa * Assert nothing in flight -- ms_*tree is empty. 13865cabbc6bSPrashanth Sreenivasa */ 13875cabbc6bSPrashanth Sreenivasa for (int i = 0; i < TXG_SIZE; i++) { 138886714001SSerapheim Dimitropoulos ASSERT0(range_tree_space(msp->ms_allocating[i])); 13895cabbc6bSPrashanth Sreenivasa } 13905cabbc6bSPrashanth Sreenivasa 13915cabbc6bSPrashanth Sreenivasa /* 13925cabbc6bSPrashanth Sreenivasa * If the metaslab has ever been allocated from (ms_sm!=NULL), 13935cabbc6bSPrashanth Sreenivasa * read the allocated segments from the space map object 13945cabbc6bSPrashanth Sreenivasa * into svr_allocd_segs. Since we do this while holding 13955cabbc6bSPrashanth Sreenivasa * svr_lock and ms_sync_lock, concurrent frees (which 13965cabbc6bSPrashanth Sreenivasa * would have modified the space map) will wait for us 13975cabbc6bSPrashanth Sreenivasa * to finish loading the spacemap, and then take the 13985cabbc6bSPrashanth Sreenivasa * appropriate action (see free_from_removing_vdev()). 13995cabbc6bSPrashanth Sreenivasa */ 14005cabbc6bSPrashanth Sreenivasa if (msp->ms_sm != NULL) { 14015cabbc6bSPrashanth Sreenivasa space_map_t *sm = NULL; 14025cabbc6bSPrashanth Sreenivasa 14035cabbc6bSPrashanth Sreenivasa /* 14045cabbc6bSPrashanth Sreenivasa * We have to open a new space map here, because 14055cabbc6bSPrashanth Sreenivasa * ms_sm's sm_length and sm_alloc may not reflect 14065cabbc6bSPrashanth Sreenivasa * what's in the object contents, if we are in between 14075cabbc6bSPrashanth Sreenivasa * metaslab_sync() and metaslab_sync_done(). 14085cabbc6bSPrashanth Sreenivasa */ 14095cabbc6bSPrashanth Sreenivasa VERIFY0(space_map_open(&sm, 14105cabbc6bSPrashanth Sreenivasa spa->spa_dsl_pool->dp_meta_objset, 14115cabbc6bSPrashanth Sreenivasa msp->ms_sm->sm_object, msp->ms_sm->sm_start, 14125cabbc6bSPrashanth Sreenivasa msp->ms_sm->sm_size, msp->ms_sm->sm_shift)); 14135cabbc6bSPrashanth Sreenivasa space_map_update(sm); 14145cabbc6bSPrashanth Sreenivasa VERIFY0(space_map_load(sm, svr->svr_allocd_segs, 14155cabbc6bSPrashanth Sreenivasa SM_ALLOC)); 14165cabbc6bSPrashanth Sreenivasa space_map_close(sm); 14175cabbc6bSPrashanth Sreenivasa 141886714001SSerapheim Dimitropoulos range_tree_walk(msp->ms_freeing, 14195cabbc6bSPrashanth Sreenivasa range_tree_remove, svr->svr_allocd_segs); 14205cabbc6bSPrashanth Sreenivasa 14215cabbc6bSPrashanth Sreenivasa /* 14225cabbc6bSPrashanth Sreenivasa * When we are resuming from a paused removal (i.e. 14235cabbc6bSPrashanth Sreenivasa * when importing a pool with a removal in progress), 14245cabbc6bSPrashanth Sreenivasa * discard any state that we have already processed. 14255cabbc6bSPrashanth Sreenivasa */ 14265cabbc6bSPrashanth Sreenivasa range_tree_clear(svr->svr_allocd_segs, 0, start_offset); 14275cabbc6bSPrashanth Sreenivasa } 14285cabbc6bSPrashanth Sreenivasa mutex_exit(&msp->ms_lock); 14295cabbc6bSPrashanth Sreenivasa mutex_exit(&msp->ms_sync_lock); 14305cabbc6bSPrashanth Sreenivasa 14315cabbc6bSPrashanth Sreenivasa vca.vca_msp = msp; 14325cabbc6bSPrashanth Sreenivasa zfs_dbgmsg("copying %llu segments for metaslab %llu", 14335cabbc6bSPrashanth Sreenivasa avl_numnodes(&svr->svr_allocd_segs->rt_root), 14345cabbc6bSPrashanth Sreenivasa msp->ms_id); 14355cabbc6bSPrashanth Sreenivasa 14365cabbc6bSPrashanth Sreenivasa while (!svr->svr_thread_exit && 143786714001SSerapheim Dimitropoulos !range_tree_is_empty(svr->svr_allocd_segs)) { 14385cabbc6bSPrashanth Sreenivasa 14395cabbc6bSPrashanth Sreenivasa mutex_exit(&svr->svr_lock); 14405cabbc6bSPrashanth Sreenivasa 14413a4b1be9SMatthew Ahrens /* 14423a4b1be9SMatthew Ahrens * We need to periodically drop the config lock so that 14433a4b1be9SMatthew Ahrens * writers can get in. Additionally, we can't wait 14443a4b1be9SMatthew Ahrens * for a txg to sync while holding a config lock 14453a4b1be9SMatthew Ahrens * (since a waiting writer could cause a 3-way deadlock 14463a4b1be9SMatthew Ahrens * with the sync thread, which also gets a config 14473a4b1be9SMatthew Ahrens * lock for reader). So we can't hold the config lock 14483a4b1be9SMatthew Ahrens * while calling dmu_tx_assign(). 14493a4b1be9SMatthew Ahrens */ 14503a4b1be9SMatthew Ahrens spa_config_exit(spa, SCL_CONFIG, FTAG); 14513a4b1be9SMatthew Ahrens 145286714001SSerapheim Dimitropoulos /* 145386714001SSerapheim Dimitropoulos * This delay will pause the removal around the point 145486714001SSerapheim Dimitropoulos * specified by zfs_remove_max_bytes_pause. We do this 145586714001SSerapheim Dimitropoulos * solely from the test suite or during debugging. 145686714001SSerapheim Dimitropoulos */ 145786714001SSerapheim Dimitropoulos uint64_t bytes_copied = 145886714001SSerapheim Dimitropoulos spa->spa_removing_phys.sr_copied; 145986714001SSerapheim Dimitropoulos for (int i = 0; i < TXG_SIZE; i++) 146086714001SSerapheim Dimitropoulos bytes_copied += svr->svr_bytes_done[i]; 146186714001SSerapheim Dimitropoulos while (zfs_remove_max_bytes_pause <= bytes_copied && 146286714001SSerapheim Dimitropoulos !svr->svr_thread_exit) 146386714001SSerapheim Dimitropoulos delay(hz); 146486714001SSerapheim Dimitropoulos 14655cabbc6bSPrashanth Sreenivasa mutex_enter(&vca.vca_lock); 14665cabbc6bSPrashanth Sreenivasa while (vca.vca_outstanding_bytes > 14675cabbc6bSPrashanth Sreenivasa zfs_remove_max_copy_bytes) { 14685cabbc6bSPrashanth Sreenivasa cv_wait(&vca.vca_cv, &vca.vca_lock); 14695cabbc6bSPrashanth Sreenivasa } 14705cabbc6bSPrashanth Sreenivasa mutex_exit(&vca.vca_lock); 14715cabbc6bSPrashanth Sreenivasa 14725cabbc6bSPrashanth Sreenivasa dmu_tx_t *tx = 14735cabbc6bSPrashanth Sreenivasa dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir); 14745cabbc6bSPrashanth Sreenivasa 14755cabbc6bSPrashanth Sreenivasa VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); 14765cabbc6bSPrashanth Sreenivasa uint64_t txg = dmu_tx_get_txg(tx); 14775cabbc6bSPrashanth Sreenivasa 14783a4b1be9SMatthew Ahrens /* 14793a4b1be9SMatthew Ahrens * Reacquire the vdev_config lock. The vdev_t 14803a4b1be9SMatthew Ahrens * that we're removing may have changed, e.g. due 14813a4b1be9SMatthew Ahrens * to a vdev_attach or vdev_detach. 14823a4b1be9SMatthew Ahrens */ 14833a4b1be9SMatthew Ahrens spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 14843a4b1be9SMatthew Ahrens vd = vdev_lookup_top(spa, svr->svr_vdev_id); 14853a4b1be9SMatthew Ahrens 14865cabbc6bSPrashanth Sreenivasa if (txg != last_txg) 14875cabbc6bSPrashanth Sreenivasa max_alloc = zfs_remove_max_segment; 14885cabbc6bSPrashanth Sreenivasa last_txg = txg; 14895cabbc6bSPrashanth Sreenivasa 14903a4b1be9SMatthew Ahrens spa_vdev_copy_impl(vd, svr, &vca, &max_alloc, tx); 14915cabbc6bSPrashanth Sreenivasa 14925cabbc6bSPrashanth Sreenivasa dmu_tx_commit(tx); 14935cabbc6bSPrashanth Sreenivasa mutex_enter(&svr->svr_lock); 14945cabbc6bSPrashanth Sreenivasa } 14955cabbc6bSPrashanth Sreenivasa } 14965cabbc6bSPrashanth Sreenivasa 14975cabbc6bSPrashanth Sreenivasa mutex_exit(&svr->svr_lock); 14983a4b1be9SMatthew Ahrens 14993a4b1be9SMatthew Ahrens spa_config_exit(spa, SCL_CONFIG, FTAG); 15003a4b1be9SMatthew Ahrens 15015cabbc6bSPrashanth Sreenivasa /* 15025cabbc6bSPrashanth Sreenivasa * Wait for all copies to finish before cleaning up the vca. 15035cabbc6bSPrashanth Sreenivasa */ 15045cabbc6bSPrashanth Sreenivasa txg_wait_synced(spa->spa_dsl_pool, 0); 15055cabbc6bSPrashanth Sreenivasa ASSERT0(vca.vca_outstanding_bytes); 15065cabbc6bSPrashanth Sreenivasa 15075cabbc6bSPrashanth Sreenivasa mutex_destroy(&vca.vca_lock); 15085cabbc6bSPrashanth Sreenivasa cv_destroy(&vca.vca_cv); 15095cabbc6bSPrashanth Sreenivasa 15105cabbc6bSPrashanth Sreenivasa if (svr->svr_thread_exit) { 15115cabbc6bSPrashanth Sreenivasa mutex_enter(&svr->svr_lock); 15125cabbc6bSPrashanth Sreenivasa range_tree_vacate(svr->svr_allocd_segs, NULL, NULL); 15135cabbc6bSPrashanth Sreenivasa svr->svr_thread = NULL; 15145cabbc6bSPrashanth Sreenivasa cv_broadcast(&svr->svr_cv); 15155cabbc6bSPrashanth Sreenivasa mutex_exit(&svr->svr_lock); 15165cabbc6bSPrashanth Sreenivasa } else { 15175cabbc6bSPrashanth Sreenivasa ASSERT0(range_tree_space(svr->svr_allocd_segs)); 15183a4b1be9SMatthew Ahrens vdev_remove_complete(spa); 15195cabbc6bSPrashanth Sreenivasa } 15205cabbc6bSPrashanth Sreenivasa } 15215cabbc6bSPrashanth Sreenivasa 15225cabbc6bSPrashanth Sreenivasa void 15235cabbc6bSPrashanth Sreenivasa spa_vdev_remove_suspend(spa_t *spa) 15245cabbc6bSPrashanth Sreenivasa { 15255cabbc6bSPrashanth Sreenivasa spa_vdev_removal_t *svr = spa->spa_vdev_removal; 15265cabbc6bSPrashanth Sreenivasa 15275cabbc6bSPrashanth Sreenivasa if (svr == NULL) 15285cabbc6bSPrashanth Sreenivasa return; 15295cabbc6bSPrashanth Sreenivasa 15305cabbc6bSPrashanth Sreenivasa mutex_enter(&svr->svr_lock); 15315cabbc6bSPrashanth Sreenivasa svr->svr_thread_exit = B_TRUE; 15325cabbc6bSPrashanth Sreenivasa while (svr->svr_thread != NULL) 15335cabbc6bSPrashanth Sreenivasa cv_wait(&svr->svr_cv, &svr->svr_lock); 15345cabbc6bSPrashanth Sreenivasa svr->svr_thread_exit = B_FALSE; 15355cabbc6bSPrashanth Sreenivasa mutex_exit(&svr->svr_lock); 15365cabbc6bSPrashanth Sreenivasa } 15375cabbc6bSPrashanth Sreenivasa 15385cabbc6bSPrashanth Sreenivasa /* ARGSUSED */ 15395cabbc6bSPrashanth Sreenivasa static int 15405cabbc6bSPrashanth Sreenivasa spa_vdev_remove_cancel_check(void *arg, dmu_tx_t *tx) 15415cabbc6bSPrashanth Sreenivasa { 15425cabbc6bSPrashanth Sreenivasa spa_t *spa = dmu_tx_pool(tx)->dp_spa; 15435cabbc6bSPrashanth Sreenivasa 15445cabbc6bSPrashanth Sreenivasa if (spa->spa_vdev_removal == NULL) 15455cabbc6bSPrashanth Sreenivasa return (ENOTACTIVE); 15465cabbc6bSPrashanth Sreenivasa return (0); 15475cabbc6bSPrashanth Sreenivasa } 15485cabbc6bSPrashanth Sreenivasa 15495cabbc6bSPrashanth Sreenivasa /* 15505cabbc6bSPrashanth Sreenivasa * Cancel a removal by freeing all entries from the partial mapping 15515cabbc6bSPrashanth Sreenivasa * and marking the vdev as no longer being removing. 15525cabbc6bSPrashanth Sreenivasa */ 15535cabbc6bSPrashanth Sreenivasa /* ARGSUSED */ 15545cabbc6bSPrashanth Sreenivasa static void 15555cabbc6bSPrashanth Sreenivasa spa_vdev_remove_cancel_sync(void *arg, dmu_tx_t *tx) 15565cabbc6bSPrashanth Sreenivasa { 15575cabbc6bSPrashanth Sreenivasa spa_t *spa = dmu_tx_pool(tx)->dp_spa; 15585cabbc6bSPrashanth Sreenivasa spa_vdev_removal_t *svr = spa->spa_vdev_removal; 15593a4b1be9SMatthew Ahrens vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id); 15605cabbc6bSPrashanth Sreenivasa vdev_indirect_config_t *vic = &vd->vdev_indirect_config; 15615cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping; 15625cabbc6bSPrashanth Sreenivasa objset_t *mos = spa->spa_meta_objset; 15635cabbc6bSPrashanth Sreenivasa 15645cabbc6bSPrashanth Sreenivasa ASSERT3P(svr->svr_thread, ==, NULL); 15655cabbc6bSPrashanth Sreenivasa 15665cabbc6bSPrashanth Sreenivasa spa_feature_decr(spa, SPA_FEATURE_DEVICE_REMOVAL, tx); 15675cabbc6bSPrashanth Sreenivasa if (vdev_obsolete_counts_are_precise(vd)) { 15685cabbc6bSPrashanth Sreenivasa spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx); 15695cabbc6bSPrashanth Sreenivasa VERIFY0(zap_remove(spa->spa_meta_objset, vd->vdev_top_zap, 15705cabbc6bSPrashanth Sreenivasa VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE, tx)); 15715cabbc6bSPrashanth Sreenivasa } 15725cabbc6bSPrashanth Sreenivasa 15735cabbc6bSPrashanth Sreenivasa if (vdev_obsolete_sm_object(vd) != 0) { 15745cabbc6bSPrashanth Sreenivasa ASSERT(vd->vdev_obsolete_sm != NULL); 15755cabbc6bSPrashanth Sreenivasa ASSERT3U(vdev_obsolete_sm_object(vd), ==, 15765cabbc6bSPrashanth Sreenivasa space_map_object(vd->vdev_obsolete_sm)); 15775cabbc6bSPrashanth Sreenivasa 15785cabbc6bSPrashanth Sreenivasa space_map_free(vd->vdev_obsolete_sm, tx); 15795cabbc6bSPrashanth Sreenivasa VERIFY0(zap_remove(spa->spa_meta_objset, vd->vdev_top_zap, 15805cabbc6bSPrashanth Sreenivasa VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM, tx)); 15815cabbc6bSPrashanth Sreenivasa space_map_close(vd->vdev_obsolete_sm); 15825cabbc6bSPrashanth Sreenivasa vd->vdev_obsolete_sm = NULL; 15835cabbc6bSPrashanth Sreenivasa spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx); 15845cabbc6bSPrashanth Sreenivasa } 15855cabbc6bSPrashanth Sreenivasa for (int i = 0; i < TXG_SIZE; i++) { 15865cabbc6bSPrashanth Sreenivasa ASSERT(list_is_empty(&svr->svr_new_segments[i])); 15875cabbc6bSPrashanth Sreenivasa ASSERT3U(svr->svr_max_offset_to_sync[i], <=, 15885cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_max_offset(vim)); 15895cabbc6bSPrashanth Sreenivasa } 15905cabbc6bSPrashanth Sreenivasa 15915cabbc6bSPrashanth Sreenivasa for (uint64_t msi = 0; msi < vd->vdev_ms_count; msi++) { 15925cabbc6bSPrashanth Sreenivasa metaslab_t *msp = vd->vdev_ms[msi]; 15935cabbc6bSPrashanth Sreenivasa 15945cabbc6bSPrashanth Sreenivasa if (msp->ms_start >= vdev_indirect_mapping_max_offset(vim)) 15955cabbc6bSPrashanth Sreenivasa break; 15965cabbc6bSPrashanth Sreenivasa 15975cabbc6bSPrashanth Sreenivasa ASSERT0(range_tree_space(svr->svr_allocd_segs)); 15985cabbc6bSPrashanth Sreenivasa 15995cabbc6bSPrashanth Sreenivasa mutex_enter(&msp->ms_lock); 16005cabbc6bSPrashanth Sreenivasa 16015cabbc6bSPrashanth Sreenivasa /* 16025cabbc6bSPrashanth Sreenivasa * Assert nothing in flight -- ms_*tree is empty. 16035cabbc6bSPrashanth Sreenivasa */ 16045cabbc6bSPrashanth Sreenivasa for (int i = 0; i < TXG_SIZE; i++) 160586714001SSerapheim Dimitropoulos ASSERT0(range_tree_space(msp->ms_allocating[i])); 16065cabbc6bSPrashanth Sreenivasa for (int i = 0; i < TXG_DEFER_SIZE; i++) 160786714001SSerapheim Dimitropoulos ASSERT0(range_tree_space(msp->ms_defer[i])); 160886714001SSerapheim Dimitropoulos ASSERT0(range_tree_space(msp->ms_freed)); 16095cabbc6bSPrashanth Sreenivasa 16105cabbc6bSPrashanth Sreenivasa if (msp->ms_sm != NULL) { 16115cabbc6bSPrashanth Sreenivasa /* 16125cabbc6bSPrashanth Sreenivasa * Assert that the in-core spacemap has the same 16135cabbc6bSPrashanth Sreenivasa * length as the on-disk one, so we can use the 16145cabbc6bSPrashanth Sreenivasa * existing in-core spacemap to load it from disk. 16155cabbc6bSPrashanth Sreenivasa */ 16165cabbc6bSPrashanth Sreenivasa ASSERT3U(msp->ms_sm->sm_alloc, ==, 16175cabbc6bSPrashanth Sreenivasa msp->ms_sm->sm_phys->smp_alloc); 16185cabbc6bSPrashanth Sreenivasa ASSERT3U(msp->ms_sm->sm_length, ==, 16195cabbc6bSPrashanth Sreenivasa msp->ms_sm->sm_phys->smp_objsize); 16205cabbc6bSPrashanth Sreenivasa 16215cabbc6bSPrashanth Sreenivasa mutex_enter(&svr->svr_lock); 16225cabbc6bSPrashanth Sreenivasa VERIFY0(space_map_load(msp->ms_sm, 16235cabbc6bSPrashanth Sreenivasa svr->svr_allocd_segs, SM_ALLOC)); 162486714001SSerapheim Dimitropoulos range_tree_walk(msp->ms_freeing, 16255cabbc6bSPrashanth Sreenivasa range_tree_remove, svr->svr_allocd_segs); 16265cabbc6bSPrashanth Sreenivasa 16275cabbc6bSPrashanth Sreenivasa /* 16285cabbc6bSPrashanth Sreenivasa * Clear everything past what has been synced, 16295cabbc6bSPrashanth Sreenivasa * because we have not allocated mappings for it yet. 16305cabbc6bSPrashanth Sreenivasa */ 16315cabbc6bSPrashanth Sreenivasa uint64_t syncd = vdev_indirect_mapping_max_offset(vim); 16323a4b1be9SMatthew Ahrens uint64_t sm_end = msp->ms_sm->sm_start + 16333a4b1be9SMatthew Ahrens msp->ms_sm->sm_size; 16343a4b1be9SMatthew Ahrens if (sm_end > syncd) 16353a4b1be9SMatthew Ahrens range_tree_clear(svr->svr_allocd_segs, 16363a4b1be9SMatthew Ahrens syncd, sm_end - syncd); 16375cabbc6bSPrashanth Sreenivasa 16385cabbc6bSPrashanth Sreenivasa mutex_exit(&svr->svr_lock); 16395cabbc6bSPrashanth Sreenivasa } 16405cabbc6bSPrashanth Sreenivasa mutex_exit(&msp->ms_lock); 16415cabbc6bSPrashanth Sreenivasa 16425cabbc6bSPrashanth Sreenivasa mutex_enter(&svr->svr_lock); 16435cabbc6bSPrashanth Sreenivasa range_tree_vacate(svr->svr_allocd_segs, 16445cabbc6bSPrashanth Sreenivasa free_mapped_segment_cb, vd); 16455cabbc6bSPrashanth Sreenivasa mutex_exit(&svr->svr_lock); 16465cabbc6bSPrashanth Sreenivasa } 16475cabbc6bSPrashanth Sreenivasa 16485cabbc6bSPrashanth Sreenivasa /* 16495cabbc6bSPrashanth Sreenivasa * Note: this must happen after we invoke free_mapped_segment_cb, 16505cabbc6bSPrashanth Sreenivasa * because it adds to the obsolete_segments. 16515cabbc6bSPrashanth Sreenivasa */ 16525cabbc6bSPrashanth Sreenivasa range_tree_vacate(vd->vdev_obsolete_segments, NULL, NULL); 16535cabbc6bSPrashanth Sreenivasa 16545cabbc6bSPrashanth Sreenivasa ASSERT3U(vic->vic_mapping_object, ==, 16555cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_object(vd->vdev_indirect_mapping)); 16565cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_close(vd->vdev_indirect_mapping); 16575cabbc6bSPrashanth Sreenivasa vd->vdev_indirect_mapping = NULL; 16585cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_free(mos, vic->vic_mapping_object, tx); 16595cabbc6bSPrashanth Sreenivasa vic->vic_mapping_object = 0; 16605cabbc6bSPrashanth Sreenivasa 16615cabbc6bSPrashanth Sreenivasa ASSERT3U(vic->vic_births_object, ==, 16625cabbc6bSPrashanth Sreenivasa vdev_indirect_births_object(vd->vdev_indirect_births)); 16635cabbc6bSPrashanth Sreenivasa vdev_indirect_births_close(vd->vdev_indirect_births); 16645cabbc6bSPrashanth Sreenivasa vd->vdev_indirect_births = NULL; 16655cabbc6bSPrashanth Sreenivasa vdev_indirect_births_free(mos, vic->vic_births_object, tx); 16665cabbc6bSPrashanth Sreenivasa vic->vic_births_object = 0; 16675cabbc6bSPrashanth Sreenivasa 16685cabbc6bSPrashanth Sreenivasa /* 16695cabbc6bSPrashanth Sreenivasa * We may have processed some frees from the removing vdev in this 16705cabbc6bSPrashanth Sreenivasa * txg, thus increasing svr_bytes_done; discard that here to 16715cabbc6bSPrashanth Sreenivasa * satisfy the assertions in spa_vdev_removal_destroy(). 16725cabbc6bSPrashanth Sreenivasa * Note that future txg's can not have any bytes_done, because 16735cabbc6bSPrashanth Sreenivasa * future TXG's are only modified from open context, and we have 16745cabbc6bSPrashanth Sreenivasa * already shut down the copying thread. 16755cabbc6bSPrashanth Sreenivasa */ 16765cabbc6bSPrashanth Sreenivasa svr->svr_bytes_done[dmu_tx_get_txg(tx) & TXG_MASK] = 0; 16775cabbc6bSPrashanth Sreenivasa spa_finish_removal(spa, DSS_CANCELED, tx); 16785cabbc6bSPrashanth Sreenivasa 16795cabbc6bSPrashanth Sreenivasa vd->vdev_removing = B_FALSE; 16805cabbc6bSPrashanth Sreenivasa vdev_config_dirty(vd); 16815cabbc6bSPrashanth Sreenivasa 16825cabbc6bSPrashanth Sreenivasa zfs_dbgmsg("canceled device removal for vdev %llu in %llu", 16835cabbc6bSPrashanth Sreenivasa vd->vdev_id, dmu_tx_get_txg(tx)); 16845cabbc6bSPrashanth Sreenivasa spa_history_log_internal(spa, "vdev remove canceled", tx, 16855cabbc6bSPrashanth Sreenivasa "%s vdev %llu %s", spa_name(spa), 16865cabbc6bSPrashanth Sreenivasa vd->vdev_id, (vd->vdev_path != NULL) ? vd->vdev_path : "-"); 16875cabbc6bSPrashanth Sreenivasa } 16885cabbc6bSPrashanth Sreenivasa 16895cabbc6bSPrashanth Sreenivasa int 16905cabbc6bSPrashanth Sreenivasa spa_vdev_remove_cancel(spa_t *spa) 16915cabbc6bSPrashanth Sreenivasa { 16925cabbc6bSPrashanth Sreenivasa spa_vdev_remove_suspend(spa); 16935cabbc6bSPrashanth Sreenivasa 16945cabbc6bSPrashanth Sreenivasa if (spa->spa_vdev_removal == NULL) 16955cabbc6bSPrashanth Sreenivasa return (ENOTACTIVE); 16965cabbc6bSPrashanth Sreenivasa 16973a4b1be9SMatthew Ahrens uint64_t vdid = spa->spa_vdev_removal->svr_vdev_id; 16985cabbc6bSPrashanth Sreenivasa 16995cabbc6bSPrashanth Sreenivasa int error = dsl_sync_task(spa->spa_name, spa_vdev_remove_cancel_check, 170086714001SSerapheim Dimitropoulos spa_vdev_remove_cancel_sync, NULL, 0, 170186714001SSerapheim Dimitropoulos ZFS_SPACE_CHECK_EXTRA_RESERVED); 17025cabbc6bSPrashanth Sreenivasa 17035cabbc6bSPrashanth Sreenivasa if (error == 0) { 17045cabbc6bSPrashanth Sreenivasa spa_config_enter(spa, SCL_ALLOC | SCL_VDEV, FTAG, RW_WRITER); 17055cabbc6bSPrashanth Sreenivasa vdev_t *vd = vdev_lookup_top(spa, vdid); 17065cabbc6bSPrashanth Sreenivasa metaslab_group_activate(vd->vdev_mg); 17075cabbc6bSPrashanth Sreenivasa spa_config_exit(spa, SCL_ALLOC | SCL_VDEV, FTAG); 17085cabbc6bSPrashanth Sreenivasa } 17095cabbc6bSPrashanth Sreenivasa 17105cabbc6bSPrashanth Sreenivasa return (error); 17115cabbc6bSPrashanth Sreenivasa } 17125cabbc6bSPrashanth Sreenivasa 17135cabbc6bSPrashanth Sreenivasa /* 17145cabbc6bSPrashanth Sreenivasa * Called every sync pass of every txg if there's a svr. 17155cabbc6bSPrashanth Sreenivasa */ 17165cabbc6bSPrashanth Sreenivasa void 17175cabbc6bSPrashanth Sreenivasa svr_sync(spa_t *spa, dmu_tx_t *tx) 17185cabbc6bSPrashanth Sreenivasa { 17195cabbc6bSPrashanth Sreenivasa spa_vdev_removal_t *svr = spa->spa_vdev_removal; 17205cabbc6bSPrashanth Sreenivasa int txgoff = dmu_tx_get_txg(tx) & TXG_MASK; 17215cabbc6bSPrashanth Sreenivasa 17225cabbc6bSPrashanth Sreenivasa /* 17235cabbc6bSPrashanth Sreenivasa * This check is necessary so that we do not dirty the 17245cabbc6bSPrashanth Sreenivasa * DIRECTORY_OBJECT via spa_sync_removing_state() when there 17255cabbc6bSPrashanth Sreenivasa * is nothing to do. Dirtying it every time would prevent us 17265cabbc6bSPrashanth Sreenivasa * from syncing-to-convergence. 17275cabbc6bSPrashanth Sreenivasa */ 17285cabbc6bSPrashanth Sreenivasa if (svr->svr_bytes_done[txgoff] == 0) 17295cabbc6bSPrashanth Sreenivasa return; 17305cabbc6bSPrashanth Sreenivasa 17315cabbc6bSPrashanth Sreenivasa /* 17325cabbc6bSPrashanth Sreenivasa * Update progress accounting. 17335cabbc6bSPrashanth Sreenivasa */ 17345cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_copied += svr->svr_bytes_done[txgoff]; 17355cabbc6bSPrashanth Sreenivasa svr->svr_bytes_done[txgoff] = 0; 17365cabbc6bSPrashanth Sreenivasa 17375cabbc6bSPrashanth Sreenivasa spa_sync_removing_state(spa, tx); 17385cabbc6bSPrashanth Sreenivasa } 17395cabbc6bSPrashanth Sreenivasa 17405cabbc6bSPrashanth Sreenivasa static void 17415cabbc6bSPrashanth Sreenivasa vdev_remove_make_hole_and_free(vdev_t *vd) 17425cabbc6bSPrashanth Sreenivasa { 17435cabbc6bSPrashanth Sreenivasa uint64_t id = vd->vdev_id; 17445cabbc6bSPrashanth Sreenivasa spa_t *spa = vd->vdev_spa; 17455cabbc6bSPrashanth Sreenivasa vdev_t *rvd = spa->spa_root_vdev; 17465cabbc6bSPrashanth Sreenivasa boolean_t last_vdev = (id == (rvd->vdev_children - 1)); 17475cabbc6bSPrashanth Sreenivasa 17485cabbc6bSPrashanth Sreenivasa ASSERT(MUTEX_HELD(&spa_namespace_lock)); 17495cabbc6bSPrashanth Sreenivasa ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); 17505cabbc6bSPrashanth Sreenivasa 17515cabbc6bSPrashanth Sreenivasa vdev_free(vd); 17525cabbc6bSPrashanth Sreenivasa 17535cabbc6bSPrashanth Sreenivasa if (last_vdev) { 17545cabbc6bSPrashanth Sreenivasa vdev_compact_children(rvd); 17555cabbc6bSPrashanth Sreenivasa } else { 17565cabbc6bSPrashanth Sreenivasa vd = vdev_alloc_common(spa, id, 0, &vdev_hole_ops); 17575cabbc6bSPrashanth Sreenivasa vdev_add_child(rvd, vd); 17585cabbc6bSPrashanth Sreenivasa } 17595cabbc6bSPrashanth Sreenivasa vdev_config_dirty(rvd); 17605cabbc6bSPrashanth Sreenivasa 17615cabbc6bSPrashanth Sreenivasa /* 17625cabbc6bSPrashanth Sreenivasa * Reassess the health of our root vdev. 17635cabbc6bSPrashanth Sreenivasa */ 17645cabbc6bSPrashanth Sreenivasa vdev_reopen(rvd); 17655cabbc6bSPrashanth Sreenivasa } 17665cabbc6bSPrashanth Sreenivasa 17675cabbc6bSPrashanth Sreenivasa /* 17685cabbc6bSPrashanth Sreenivasa * Remove a log device. The config lock is held for the specified TXG. 17695cabbc6bSPrashanth Sreenivasa */ 17705cabbc6bSPrashanth Sreenivasa static int 17715cabbc6bSPrashanth Sreenivasa spa_vdev_remove_log(vdev_t *vd, uint64_t *txg) 17725cabbc6bSPrashanth Sreenivasa { 17735cabbc6bSPrashanth Sreenivasa metaslab_group_t *mg = vd->vdev_mg; 17745cabbc6bSPrashanth Sreenivasa spa_t *spa = vd->vdev_spa; 17755cabbc6bSPrashanth Sreenivasa int error = 0; 17765cabbc6bSPrashanth Sreenivasa 17775cabbc6bSPrashanth Sreenivasa ASSERT(vd->vdev_islog); 17785cabbc6bSPrashanth Sreenivasa ASSERT(vd == vd->vdev_top); 17795cabbc6bSPrashanth Sreenivasa 17805cabbc6bSPrashanth Sreenivasa /* 17815cabbc6bSPrashanth Sreenivasa * Stop allocating from this vdev. 17825cabbc6bSPrashanth Sreenivasa */ 17835cabbc6bSPrashanth Sreenivasa metaslab_group_passivate(mg); 17845cabbc6bSPrashanth Sreenivasa 17855cabbc6bSPrashanth Sreenivasa /* 17865cabbc6bSPrashanth Sreenivasa * Wait for the youngest allocations and frees to sync, 17875cabbc6bSPrashanth Sreenivasa * and then wait for the deferral of those frees to finish. 17885cabbc6bSPrashanth Sreenivasa */ 17895cabbc6bSPrashanth Sreenivasa spa_vdev_config_exit(spa, NULL, 17905cabbc6bSPrashanth Sreenivasa *txg + TXG_CONCURRENT_STATES + TXG_DEFER_SIZE, 0, FTAG); 17915cabbc6bSPrashanth Sreenivasa 17925cabbc6bSPrashanth Sreenivasa /* 17935cabbc6bSPrashanth Sreenivasa * Evacuate the device. We don't hold the config lock as writer 17945cabbc6bSPrashanth Sreenivasa * since we need to do I/O but we do keep the 17955cabbc6bSPrashanth Sreenivasa * spa_namespace_lock held. Once this completes the device 17965cabbc6bSPrashanth Sreenivasa * should no longer have any blocks allocated on it. 17975cabbc6bSPrashanth Sreenivasa */ 17985cabbc6bSPrashanth Sreenivasa if (vd->vdev_islog) { 17995cabbc6bSPrashanth Sreenivasa if (vd->vdev_stat.vs_alloc != 0) 18005cabbc6bSPrashanth Sreenivasa error = spa_reset_logs(spa); 18015cabbc6bSPrashanth Sreenivasa } 18025cabbc6bSPrashanth Sreenivasa 18035cabbc6bSPrashanth Sreenivasa *txg = spa_vdev_config_enter(spa); 18045cabbc6bSPrashanth Sreenivasa 18055cabbc6bSPrashanth Sreenivasa if (error != 0) { 18065cabbc6bSPrashanth Sreenivasa metaslab_group_activate(mg); 18075cabbc6bSPrashanth Sreenivasa return (error); 18085cabbc6bSPrashanth Sreenivasa } 18095cabbc6bSPrashanth Sreenivasa ASSERT0(vd->vdev_stat.vs_alloc); 18105cabbc6bSPrashanth Sreenivasa 18115cabbc6bSPrashanth Sreenivasa /* 18125cabbc6bSPrashanth Sreenivasa * The evacuation succeeded. Remove any remaining MOS metadata 18135cabbc6bSPrashanth Sreenivasa * associated with this vdev, and wait for these changes to sync. 18145cabbc6bSPrashanth Sreenivasa */ 18155cabbc6bSPrashanth Sreenivasa vd->vdev_removing = B_TRUE; 18165cabbc6bSPrashanth Sreenivasa 18175cabbc6bSPrashanth Sreenivasa vdev_dirty_leaves(vd, VDD_DTL, *txg); 18185cabbc6bSPrashanth Sreenivasa vdev_config_dirty(vd); 18195cabbc6bSPrashanth Sreenivasa 18205cabbc6bSPrashanth Sreenivasa spa_history_log_internal(spa, "vdev remove", NULL, 18215cabbc6bSPrashanth Sreenivasa "%s vdev %llu (log) %s", spa_name(spa), vd->vdev_id, 18225cabbc6bSPrashanth Sreenivasa (vd->vdev_path != NULL) ? vd->vdev_path : "-"); 18235cabbc6bSPrashanth Sreenivasa 18245cabbc6bSPrashanth Sreenivasa /* Make sure these changes are sync'ed */ 18255cabbc6bSPrashanth Sreenivasa spa_vdev_config_exit(spa, NULL, *txg, 0, FTAG); 18265cabbc6bSPrashanth Sreenivasa 1827094e47e9SGeorge Wilson /* Stop initializing */ 1828094e47e9SGeorge Wilson (void) vdev_initialize_stop_all(vd, VDEV_INITIALIZE_CANCELED); 1829094e47e9SGeorge Wilson 18305cabbc6bSPrashanth Sreenivasa *txg = spa_vdev_config_enter(spa); 18315cabbc6bSPrashanth Sreenivasa 18325cabbc6bSPrashanth Sreenivasa sysevent_t *ev = spa_event_create(spa, vd, NULL, 18335cabbc6bSPrashanth Sreenivasa ESC_ZFS_VDEV_REMOVE_DEV); 18345cabbc6bSPrashanth Sreenivasa ASSERT(MUTEX_HELD(&spa_namespace_lock)); 18355cabbc6bSPrashanth Sreenivasa ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); 18365cabbc6bSPrashanth Sreenivasa 18375cabbc6bSPrashanth Sreenivasa /* The top ZAP should have been destroyed by vdev_remove_empty. */ 18385cabbc6bSPrashanth Sreenivasa ASSERT0(vd->vdev_top_zap); 18395cabbc6bSPrashanth Sreenivasa /* The leaf ZAP should have been destroyed by vdev_dtl_sync. */ 18405cabbc6bSPrashanth Sreenivasa ASSERT0(vd->vdev_leaf_zap); 18415cabbc6bSPrashanth Sreenivasa 18425cabbc6bSPrashanth Sreenivasa (void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE); 18435cabbc6bSPrashanth Sreenivasa 18445cabbc6bSPrashanth Sreenivasa if (list_link_active(&vd->vdev_state_dirty_node)) 18455cabbc6bSPrashanth Sreenivasa vdev_state_clean(vd); 18465cabbc6bSPrashanth Sreenivasa if (list_link_active(&vd->vdev_config_dirty_node)) 18475cabbc6bSPrashanth Sreenivasa vdev_config_clean(vd); 18485cabbc6bSPrashanth Sreenivasa 18495cabbc6bSPrashanth Sreenivasa /* 18505cabbc6bSPrashanth Sreenivasa * Clean up the vdev namespace. 18515cabbc6bSPrashanth Sreenivasa */ 18525cabbc6bSPrashanth Sreenivasa vdev_remove_make_hole_and_free(vd); 18535cabbc6bSPrashanth Sreenivasa 18545cabbc6bSPrashanth Sreenivasa if (ev != NULL) 18555cabbc6bSPrashanth Sreenivasa spa_event_post(ev); 18565cabbc6bSPrashanth Sreenivasa 18575cabbc6bSPrashanth Sreenivasa return (0); 18585cabbc6bSPrashanth Sreenivasa } 18595cabbc6bSPrashanth Sreenivasa 18605cabbc6bSPrashanth Sreenivasa static int 18615cabbc6bSPrashanth Sreenivasa spa_vdev_remove_top_check(vdev_t *vd) 18625cabbc6bSPrashanth Sreenivasa { 18635cabbc6bSPrashanth Sreenivasa spa_t *spa = vd->vdev_spa; 18645cabbc6bSPrashanth Sreenivasa 18655cabbc6bSPrashanth Sreenivasa if (vd != vd->vdev_top) 18665cabbc6bSPrashanth Sreenivasa return (SET_ERROR(ENOTSUP)); 18675cabbc6bSPrashanth Sreenivasa 18685cabbc6bSPrashanth Sreenivasa if (!spa_feature_is_enabled(spa, SPA_FEATURE_DEVICE_REMOVAL)) 18695cabbc6bSPrashanth Sreenivasa return (SET_ERROR(ENOTSUP)); 18705cabbc6bSPrashanth Sreenivasa 18715cabbc6bSPrashanth Sreenivasa /* 18725cabbc6bSPrashanth Sreenivasa * There has to be enough free space to remove the 18735cabbc6bSPrashanth Sreenivasa * device and leave double the "slop" space (i.e. we 18745cabbc6bSPrashanth Sreenivasa * must leave at least 3% of the pool free, in addition to 18755cabbc6bSPrashanth Sreenivasa * the normal slop space). 18765cabbc6bSPrashanth Sreenivasa */ 18775cabbc6bSPrashanth Sreenivasa if (dsl_dir_space_available(spa->spa_dsl_pool->dp_root_dir, 18785cabbc6bSPrashanth Sreenivasa NULL, 0, B_TRUE) < 18795cabbc6bSPrashanth Sreenivasa vd->vdev_stat.vs_dspace + spa_get_slop_space(spa)) { 18805cabbc6bSPrashanth Sreenivasa return (SET_ERROR(ENOSPC)); 18815cabbc6bSPrashanth Sreenivasa } 18825cabbc6bSPrashanth Sreenivasa 18835cabbc6bSPrashanth Sreenivasa /* 18845cabbc6bSPrashanth Sreenivasa * There can not be a removal in progress. 18855cabbc6bSPrashanth Sreenivasa */ 18865cabbc6bSPrashanth Sreenivasa if (spa->spa_removing_phys.sr_state == DSS_SCANNING) 18875cabbc6bSPrashanth Sreenivasa return (SET_ERROR(EBUSY)); 18885cabbc6bSPrashanth Sreenivasa 18895cabbc6bSPrashanth Sreenivasa /* 18905cabbc6bSPrashanth Sreenivasa * The device must have all its data. 18915cabbc6bSPrashanth Sreenivasa */ 18925cabbc6bSPrashanth Sreenivasa if (!vdev_dtl_empty(vd, DTL_MISSING) || 18935cabbc6bSPrashanth Sreenivasa !vdev_dtl_empty(vd, DTL_OUTAGE)) 18945cabbc6bSPrashanth Sreenivasa return (SET_ERROR(EBUSY)); 18955cabbc6bSPrashanth Sreenivasa 18965cabbc6bSPrashanth Sreenivasa /* 18975cabbc6bSPrashanth Sreenivasa * The device must be healthy. 18985cabbc6bSPrashanth Sreenivasa */ 18995cabbc6bSPrashanth Sreenivasa if (!vdev_readable(vd)) 19005cabbc6bSPrashanth Sreenivasa return (SET_ERROR(EIO)); 19015cabbc6bSPrashanth Sreenivasa 19025cabbc6bSPrashanth Sreenivasa /* 19035cabbc6bSPrashanth Sreenivasa * All vdevs in normal class must have the same ashift. 19045cabbc6bSPrashanth Sreenivasa */ 19055cabbc6bSPrashanth Sreenivasa if (spa->spa_max_ashift != spa->spa_min_ashift) { 19065cabbc6bSPrashanth Sreenivasa return (SET_ERROR(EINVAL)); 19075cabbc6bSPrashanth Sreenivasa } 19085cabbc6bSPrashanth Sreenivasa 19095cabbc6bSPrashanth Sreenivasa /* 19105cabbc6bSPrashanth Sreenivasa * All vdevs in normal class must have the same ashift 19115cabbc6bSPrashanth Sreenivasa * and not be raidz. 19125cabbc6bSPrashanth Sreenivasa */ 19135cabbc6bSPrashanth Sreenivasa vdev_t *rvd = spa->spa_root_vdev; 19145cabbc6bSPrashanth Sreenivasa int num_indirect = 0; 19155cabbc6bSPrashanth Sreenivasa for (uint64_t id = 0; id < rvd->vdev_children; id++) { 19165cabbc6bSPrashanth Sreenivasa vdev_t *cvd = rvd->vdev_child[id]; 19175cabbc6bSPrashanth Sreenivasa if (cvd->vdev_ashift != 0 && !cvd->vdev_islog) 19185cabbc6bSPrashanth Sreenivasa ASSERT3U(cvd->vdev_ashift, ==, spa->spa_max_ashift); 19195cabbc6bSPrashanth Sreenivasa if (cvd->vdev_ops == &vdev_indirect_ops) 19205cabbc6bSPrashanth Sreenivasa num_indirect++; 19215cabbc6bSPrashanth Sreenivasa if (!vdev_is_concrete(cvd)) 19225cabbc6bSPrashanth Sreenivasa continue; 19235cabbc6bSPrashanth Sreenivasa if (cvd->vdev_ops == &vdev_raidz_ops) 19245cabbc6bSPrashanth Sreenivasa return (SET_ERROR(EINVAL)); 19255cabbc6bSPrashanth Sreenivasa /* 19265cabbc6bSPrashanth Sreenivasa * Need the mirror to be mirror of leaf vdevs only 19275cabbc6bSPrashanth Sreenivasa */ 19285cabbc6bSPrashanth Sreenivasa if (cvd->vdev_ops == &vdev_mirror_ops) { 19295cabbc6bSPrashanth Sreenivasa for (uint64_t cid = 0; 19305cabbc6bSPrashanth Sreenivasa cid < cvd->vdev_children; cid++) { 19315cabbc6bSPrashanth Sreenivasa vdev_t *tmp = cvd->vdev_child[cid]; 19325cabbc6bSPrashanth Sreenivasa if (!tmp->vdev_ops->vdev_op_leaf) 19335cabbc6bSPrashanth Sreenivasa return (SET_ERROR(EINVAL)); 19345cabbc6bSPrashanth Sreenivasa } 19355cabbc6bSPrashanth Sreenivasa } 19365cabbc6bSPrashanth Sreenivasa } 19375cabbc6bSPrashanth Sreenivasa 19385cabbc6bSPrashanth Sreenivasa return (0); 19395cabbc6bSPrashanth Sreenivasa } 19405cabbc6bSPrashanth Sreenivasa 19415cabbc6bSPrashanth Sreenivasa /* 19425cabbc6bSPrashanth Sreenivasa * Initiate removal of a top-level vdev, reducing the total space in the pool. 19435cabbc6bSPrashanth Sreenivasa * The config lock is held for the specified TXG. Once initiated, 19445cabbc6bSPrashanth Sreenivasa * evacuation of all allocated space (copying it to other vdevs) happens 19455cabbc6bSPrashanth Sreenivasa * in the background (see spa_vdev_remove_thread()), and can be canceled 19465cabbc6bSPrashanth Sreenivasa * (see spa_vdev_remove_cancel()). If successful, the vdev will 19475cabbc6bSPrashanth Sreenivasa * be transformed to an indirect vdev (see spa_vdev_remove_complete()). 19485cabbc6bSPrashanth Sreenivasa */ 19495cabbc6bSPrashanth Sreenivasa static int 19505cabbc6bSPrashanth Sreenivasa spa_vdev_remove_top(vdev_t *vd, uint64_t *txg) 19515cabbc6bSPrashanth Sreenivasa { 19525cabbc6bSPrashanth Sreenivasa spa_t *spa = vd->vdev_spa; 19535cabbc6bSPrashanth Sreenivasa int error; 19545cabbc6bSPrashanth Sreenivasa 19555cabbc6bSPrashanth Sreenivasa /* 19565cabbc6bSPrashanth Sreenivasa * Check for errors up-front, so that we don't waste time 19575cabbc6bSPrashanth Sreenivasa * passivating the metaslab group and clearing the ZIL if there 19585cabbc6bSPrashanth Sreenivasa * are errors. 19595cabbc6bSPrashanth Sreenivasa */ 19605cabbc6bSPrashanth Sreenivasa error = spa_vdev_remove_top_check(vd); 19615cabbc6bSPrashanth Sreenivasa if (error != 0) 19625cabbc6bSPrashanth Sreenivasa return (error); 19635cabbc6bSPrashanth Sreenivasa 19645cabbc6bSPrashanth Sreenivasa /* 19655cabbc6bSPrashanth Sreenivasa * Stop allocating from this vdev. Note that we must check 19665cabbc6bSPrashanth Sreenivasa * that this is not the only device in the pool before 19675cabbc6bSPrashanth Sreenivasa * passivating, otherwise we will not be able to make 19685cabbc6bSPrashanth Sreenivasa * progress because we can't allocate from any vdevs. 19695cabbc6bSPrashanth Sreenivasa * The above check for sufficient free space serves this 19705cabbc6bSPrashanth Sreenivasa * purpose. 19715cabbc6bSPrashanth Sreenivasa */ 19725cabbc6bSPrashanth Sreenivasa metaslab_group_t *mg = vd->vdev_mg; 19735cabbc6bSPrashanth Sreenivasa metaslab_group_passivate(mg); 19745cabbc6bSPrashanth Sreenivasa 19755cabbc6bSPrashanth Sreenivasa /* 19765cabbc6bSPrashanth Sreenivasa * Wait for the youngest allocations and frees to sync, 19775cabbc6bSPrashanth Sreenivasa * and then wait for the deferral of those frees to finish. 19785cabbc6bSPrashanth Sreenivasa */ 19795cabbc6bSPrashanth Sreenivasa spa_vdev_config_exit(spa, NULL, 19805cabbc6bSPrashanth Sreenivasa *txg + TXG_CONCURRENT_STATES + TXG_DEFER_SIZE, 0, FTAG); 19815cabbc6bSPrashanth Sreenivasa 19825cabbc6bSPrashanth Sreenivasa /* 19835cabbc6bSPrashanth Sreenivasa * We must ensure that no "stubby" log blocks are allocated 19845cabbc6bSPrashanth Sreenivasa * on the device to be removed. These blocks could be 19855cabbc6bSPrashanth Sreenivasa * written at any time, including while we are in the middle 19865cabbc6bSPrashanth Sreenivasa * of copying them. 19875cabbc6bSPrashanth Sreenivasa */ 19885cabbc6bSPrashanth Sreenivasa error = spa_reset_logs(spa); 19895cabbc6bSPrashanth Sreenivasa 1990094e47e9SGeorge Wilson /* 1991094e47e9SGeorge Wilson * We stop any initializing that is currently in progress but leave 1992094e47e9SGeorge Wilson * the state as "active". This will allow the initializing to resume 1993094e47e9SGeorge Wilson * if the removal is canceled sometime later. 1994094e47e9SGeorge Wilson */ 1995094e47e9SGeorge Wilson vdev_initialize_stop_all(vd, VDEV_INITIALIZE_ACTIVE); 1996094e47e9SGeorge Wilson 19975cabbc6bSPrashanth Sreenivasa *txg = spa_vdev_config_enter(spa); 19985cabbc6bSPrashanth Sreenivasa 19995cabbc6bSPrashanth Sreenivasa /* 20005cabbc6bSPrashanth Sreenivasa * Things might have changed while the config lock was dropped 20015cabbc6bSPrashanth Sreenivasa * (e.g. space usage). Check for errors again. 20025cabbc6bSPrashanth Sreenivasa */ 20035cabbc6bSPrashanth Sreenivasa if (error == 0) 20045cabbc6bSPrashanth Sreenivasa error = spa_vdev_remove_top_check(vd); 20055cabbc6bSPrashanth Sreenivasa 20065cabbc6bSPrashanth Sreenivasa if (error != 0) { 20075cabbc6bSPrashanth Sreenivasa metaslab_group_activate(mg); 2008094e47e9SGeorge Wilson spa_async_request(spa, SPA_ASYNC_INITIALIZE_RESTART); 20095cabbc6bSPrashanth Sreenivasa return (error); 20105cabbc6bSPrashanth Sreenivasa } 20115cabbc6bSPrashanth Sreenivasa 20125cabbc6bSPrashanth Sreenivasa vd->vdev_removing = B_TRUE; 20135cabbc6bSPrashanth Sreenivasa 20145cabbc6bSPrashanth Sreenivasa vdev_dirty_leaves(vd, VDD_DTL, *txg); 20155cabbc6bSPrashanth Sreenivasa vdev_config_dirty(vd); 20165cabbc6bSPrashanth Sreenivasa dmu_tx_t *tx = dmu_tx_create_assigned(spa->spa_dsl_pool, *txg); 20175cabbc6bSPrashanth Sreenivasa dsl_sync_task_nowait(spa->spa_dsl_pool, 20185cabbc6bSPrashanth Sreenivasa vdev_remove_initiate_sync, 20193a4b1be9SMatthew Ahrens (void *)(uintptr_t)vd->vdev_id, 0, ZFS_SPACE_CHECK_NONE, tx); 20205cabbc6bSPrashanth Sreenivasa dmu_tx_commit(tx); 20215cabbc6bSPrashanth Sreenivasa 20225cabbc6bSPrashanth Sreenivasa return (0); 20235cabbc6bSPrashanth Sreenivasa } 20245cabbc6bSPrashanth Sreenivasa 20255cabbc6bSPrashanth Sreenivasa /* 20265cabbc6bSPrashanth Sreenivasa * Remove a device from the pool. 20275cabbc6bSPrashanth Sreenivasa * 20285cabbc6bSPrashanth Sreenivasa * Removing a device from the vdev namespace requires several steps 20295cabbc6bSPrashanth Sreenivasa * and can take a significant amount of time. As a result we use 20305cabbc6bSPrashanth Sreenivasa * the spa_vdev_config_[enter/exit] functions which allow us to 20315cabbc6bSPrashanth Sreenivasa * grab and release the spa_config_lock while still holding the namespace 20325cabbc6bSPrashanth Sreenivasa * lock. During each step the configuration is synced out. 20335cabbc6bSPrashanth Sreenivasa */ 20345cabbc6bSPrashanth Sreenivasa int 20355cabbc6bSPrashanth Sreenivasa spa_vdev_remove(spa_t *spa, uint64_t guid, boolean_t unspare) 20365cabbc6bSPrashanth Sreenivasa { 20375cabbc6bSPrashanth Sreenivasa vdev_t *vd; 20385cabbc6bSPrashanth Sreenivasa nvlist_t **spares, **l2cache, *nv; 20395cabbc6bSPrashanth Sreenivasa uint64_t txg = 0; 20405cabbc6bSPrashanth Sreenivasa uint_t nspares, nl2cache; 20415cabbc6bSPrashanth Sreenivasa int error = 0; 20425cabbc6bSPrashanth Sreenivasa boolean_t locked = MUTEX_HELD(&spa_namespace_lock); 20435cabbc6bSPrashanth Sreenivasa sysevent_t *ev = NULL; 20445cabbc6bSPrashanth Sreenivasa 20455cabbc6bSPrashanth Sreenivasa ASSERT(spa_writeable(spa)); 20465cabbc6bSPrashanth Sreenivasa 20475cabbc6bSPrashanth Sreenivasa if (!locked) 20485cabbc6bSPrashanth Sreenivasa txg = spa_vdev_enter(spa); 20495cabbc6bSPrashanth Sreenivasa 205086714001SSerapheim Dimitropoulos ASSERT(MUTEX_HELD(&spa_namespace_lock)); 205186714001SSerapheim Dimitropoulos if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) { 205286714001SSerapheim Dimitropoulos error = (spa_has_checkpoint(spa)) ? 205386714001SSerapheim Dimitropoulos ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT; 205486714001SSerapheim Dimitropoulos 205586714001SSerapheim Dimitropoulos if (!locked) 205686714001SSerapheim Dimitropoulos return (spa_vdev_exit(spa, NULL, txg, error)); 205786714001SSerapheim Dimitropoulos 205886714001SSerapheim Dimitropoulos return (error); 205986714001SSerapheim Dimitropoulos } 206086714001SSerapheim Dimitropoulos 20615cabbc6bSPrashanth Sreenivasa vd = spa_lookup_by_guid(spa, guid, B_FALSE); 20625cabbc6bSPrashanth Sreenivasa 20635cabbc6bSPrashanth Sreenivasa if (spa->spa_spares.sav_vdevs != NULL && 20645cabbc6bSPrashanth Sreenivasa nvlist_lookup_nvlist_array(spa->spa_spares.sav_config, 20655cabbc6bSPrashanth Sreenivasa ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0 && 20665cabbc6bSPrashanth Sreenivasa (nv = spa_nvlist_lookup_by_guid(spares, nspares, guid)) != NULL) { 20675cabbc6bSPrashanth Sreenivasa /* 20685cabbc6bSPrashanth Sreenivasa * Only remove the hot spare if it's not currently in use 20695cabbc6bSPrashanth Sreenivasa * in this pool. 20705cabbc6bSPrashanth Sreenivasa */ 20715cabbc6bSPrashanth Sreenivasa if (vd == NULL || unspare) { 20725cabbc6bSPrashanth Sreenivasa char *nvstr = fnvlist_lookup_string(nv, 20735cabbc6bSPrashanth Sreenivasa ZPOOL_CONFIG_PATH); 20745cabbc6bSPrashanth Sreenivasa spa_history_log_internal(spa, "vdev remove", NULL, 20755cabbc6bSPrashanth Sreenivasa "%s vdev (%s) %s", spa_name(spa), 20765cabbc6bSPrashanth Sreenivasa VDEV_TYPE_SPARE, nvstr); 20775cabbc6bSPrashanth Sreenivasa if (vd == NULL) 20785cabbc6bSPrashanth Sreenivasa vd = spa_lookup_by_guid(spa, guid, B_TRUE); 20795cabbc6bSPrashanth Sreenivasa ev = spa_event_create(spa, vd, NULL, 20805cabbc6bSPrashanth Sreenivasa ESC_ZFS_VDEV_REMOVE_AUX); 20815cabbc6bSPrashanth Sreenivasa spa_vdev_remove_aux(spa->spa_spares.sav_config, 20825cabbc6bSPrashanth Sreenivasa ZPOOL_CONFIG_SPARES, spares, nspares, nv); 20835cabbc6bSPrashanth Sreenivasa spa_load_spares(spa); 20845cabbc6bSPrashanth Sreenivasa spa->spa_spares.sav_sync = B_TRUE; 20855cabbc6bSPrashanth Sreenivasa } else { 20865cabbc6bSPrashanth Sreenivasa error = SET_ERROR(EBUSY); 20875cabbc6bSPrashanth Sreenivasa } 20885cabbc6bSPrashanth Sreenivasa } else if (spa->spa_l2cache.sav_vdevs != NULL && 20895cabbc6bSPrashanth Sreenivasa nvlist_lookup_nvlist_array(spa->spa_l2cache.sav_config, 20905cabbc6bSPrashanth Sreenivasa ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0 && 20915cabbc6bSPrashanth Sreenivasa (nv = spa_nvlist_lookup_by_guid(l2cache, nl2cache, guid)) != NULL) { 20925cabbc6bSPrashanth Sreenivasa char *nvstr = fnvlist_lookup_string(nv, ZPOOL_CONFIG_PATH); 20935cabbc6bSPrashanth Sreenivasa spa_history_log_internal(spa, "vdev remove", NULL, 20945cabbc6bSPrashanth Sreenivasa "%s vdev (%s) %s", spa_name(spa), VDEV_TYPE_L2CACHE, nvstr); 20955cabbc6bSPrashanth Sreenivasa /* 20965cabbc6bSPrashanth Sreenivasa * Cache devices can always be removed. 20975cabbc6bSPrashanth Sreenivasa */ 20985cabbc6bSPrashanth Sreenivasa vd = spa_lookup_by_guid(spa, guid, B_TRUE); 20995cabbc6bSPrashanth Sreenivasa ev = spa_event_create(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE_AUX); 21005cabbc6bSPrashanth Sreenivasa spa_vdev_remove_aux(spa->spa_l2cache.sav_config, 21015cabbc6bSPrashanth Sreenivasa ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache, nv); 21025cabbc6bSPrashanth Sreenivasa spa_load_l2cache(spa); 21035cabbc6bSPrashanth Sreenivasa spa->spa_l2cache.sav_sync = B_TRUE; 21045cabbc6bSPrashanth Sreenivasa } else if (vd != NULL && vd->vdev_islog) { 21055cabbc6bSPrashanth Sreenivasa ASSERT(!locked); 21065cabbc6bSPrashanth Sreenivasa error = spa_vdev_remove_log(vd, &txg); 21075cabbc6bSPrashanth Sreenivasa } else if (vd != NULL) { 21085cabbc6bSPrashanth Sreenivasa ASSERT(!locked); 21095cabbc6bSPrashanth Sreenivasa error = spa_vdev_remove_top(vd, &txg); 21105cabbc6bSPrashanth Sreenivasa } else { 21115cabbc6bSPrashanth Sreenivasa /* 21125cabbc6bSPrashanth Sreenivasa * There is no vdev of any kind with the specified guid. 21135cabbc6bSPrashanth Sreenivasa */ 21145cabbc6bSPrashanth Sreenivasa error = SET_ERROR(ENOENT); 21155cabbc6bSPrashanth Sreenivasa } 21165cabbc6bSPrashanth Sreenivasa 21175cabbc6bSPrashanth Sreenivasa if (!locked) 21185cabbc6bSPrashanth Sreenivasa error = spa_vdev_exit(spa, NULL, txg, error); 21195cabbc6bSPrashanth Sreenivasa 21205cabbc6bSPrashanth Sreenivasa if (ev != NULL) { 21215cabbc6bSPrashanth Sreenivasa if (error != 0) { 21225cabbc6bSPrashanth Sreenivasa spa_event_discard(ev); 21235cabbc6bSPrashanth Sreenivasa } else { 21245cabbc6bSPrashanth Sreenivasa spa_event_post(ev); 21255cabbc6bSPrashanth Sreenivasa } 21265cabbc6bSPrashanth Sreenivasa } 21275cabbc6bSPrashanth Sreenivasa 21285cabbc6bSPrashanth Sreenivasa return (error); 21295cabbc6bSPrashanth Sreenivasa } 21305cabbc6bSPrashanth Sreenivasa 21315cabbc6bSPrashanth Sreenivasa int 21325cabbc6bSPrashanth Sreenivasa spa_removal_get_stats(spa_t *spa, pool_removal_stat_t *prs) 21335cabbc6bSPrashanth Sreenivasa { 21345cabbc6bSPrashanth Sreenivasa prs->prs_state = spa->spa_removing_phys.sr_state; 21355cabbc6bSPrashanth Sreenivasa 21365cabbc6bSPrashanth Sreenivasa if (prs->prs_state == DSS_NONE) 21375cabbc6bSPrashanth Sreenivasa return (SET_ERROR(ENOENT)); 21385cabbc6bSPrashanth Sreenivasa 21395cabbc6bSPrashanth Sreenivasa prs->prs_removing_vdev = spa->spa_removing_phys.sr_removing_vdev; 21405cabbc6bSPrashanth Sreenivasa prs->prs_start_time = spa->spa_removing_phys.sr_start_time; 21415cabbc6bSPrashanth Sreenivasa prs->prs_end_time = spa->spa_removing_phys.sr_end_time; 21425cabbc6bSPrashanth Sreenivasa prs->prs_to_copy = spa->spa_removing_phys.sr_to_copy; 21435cabbc6bSPrashanth Sreenivasa prs->prs_copied = spa->spa_removing_phys.sr_copied; 21445cabbc6bSPrashanth Sreenivasa 21455cabbc6bSPrashanth Sreenivasa if (spa->spa_vdev_removal != NULL) { 21465cabbc6bSPrashanth Sreenivasa for (int i = 0; i < TXG_SIZE; i++) { 21475cabbc6bSPrashanth Sreenivasa prs->prs_copied += 21485cabbc6bSPrashanth Sreenivasa spa->spa_vdev_removal->svr_bytes_done[i]; 21495cabbc6bSPrashanth Sreenivasa } 21505cabbc6bSPrashanth Sreenivasa } 21515cabbc6bSPrashanth Sreenivasa 21525cabbc6bSPrashanth Sreenivasa prs->prs_mapping_memory = 0; 21535cabbc6bSPrashanth Sreenivasa uint64_t indirect_vdev_id = 21545cabbc6bSPrashanth Sreenivasa spa->spa_removing_phys.sr_prev_indirect_vdev; 21555cabbc6bSPrashanth Sreenivasa while (indirect_vdev_id != -1) { 21565cabbc6bSPrashanth Sreenivasa vdev_t *vd = spa->spa_root_vdev->vdev_child[indirect_vdev_id]; 21575cabbc6bSPrashanth Sreenivasa vdev_indirect_config_t *vic = &vd->vdev_indirect_config; 21585cabbc6bSPrashanth Sreenivasa vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping; 21595cabbc6bSPrashanth Sreenivasa 21605cabbc6bSPrashanth Sreenivasa ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops); 21615cabbc6bSPrashanth Sreenivasa prs->prs_mapping_memory += vdev_indirect_mapping_size(vim); 21625cabbc6bSPrashanth Sreenivasa indirect_vdev_id = vic->vic_prev_indirect_vdev; 21635cabbc6bSPrashanth Sreenivasa } 21645cabbc6bSPrashanth Sreenivasa 21655cabbc6bSPrashanth Sreenivasa return (0); 21665cabbc6bSPrashanth Sreenivasa } 2167