metaslab.c revision 03f8c366886542ed249a15d755ae78ea4e775d9d
fa9e4066f08beec538e775443c5be79dd423fcabahrens * CDDL HEADER START
fa9e4066f08beec538e775443c5be79dd423fcabahrens * The contents of this file are subject to the terms of the
ea8dc4b6d2251b437950c0056bc626b311c73c27eschrock * Common Development and Distribution License (the "License").
ea8dc4b6d2251b437950c0056bc626b311c73c27eschrock * You may not use this file except in compliance with the License.
fa9e4066f08beec538e775443c5be79dd423fcabahrens * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
fa9e4066f08beec538e775443c5be79dd423fcabahrens * See the License for the specific language governing permissions
fa9e4066f08beec538e775443c5be79dd423fcabahrens * and limitations under the License.
fa9e4066f08beec538e775443c5be79dd423fcabahrens * When distributing Covered Code, include this CDDL HEADER in each
fa9e4066f08beec538e775443c5be79dd423fcabahrens * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
fa9e4066f08beec538e775443c5be79dd423fcabahrens * If applicable, add the following below this CDDL HEADER, with the
fa9e4066f08beec538e775443c5be79dd423fcabahrens * fields enclosed by brackets "[]" replaced with your own identifying
fa9e4066f08beec538e775443c5be79dd423fcabahrens * information: Portions Copyright [yyyy] [name of copyright owner]
fa9e4066f08beec538e775443c5be79dd423fcabahrens * CDDL HEADER END
8d18220deb04ec7b12410cd90deb4d45e66d49bfMark J Musante * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
840345f69c01af33ad169d6b4a6ae7ea78b4218aGeorge Wilson * Copyright (c) 2012 by Delphix. All rights reserved.
9dc3941c735ef88de46e850f745aa556d3a071a5Sašo Kiselkov * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * Allow allocations to switch to gang blocks quickly. We do this to
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * avoid having to load lots of space_maps in a given txg. There are,
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * however, some cases where we want to avoid "fast" ganging and instead
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * we want to do an exhaustive search of all metaslabs on this device.
840345f69c01af33ad169d6b4a6ae7ea78b4218aGeorge Wilson * Currently we don't allow any gang, zil, or dump device related allocations
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * to "fast" gang.
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
e05725b117836db173257fae43fb0746eb857fb5bonwickuint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * The in-core space map representation is more compact than its on-disk form.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * The zfs_condense_pct determines how much more compact the in-core
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * space_map representation must be before we compact it on-disk.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * Values should be greater than or equal to 100.
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * This value defines the number of allowed allocation failures per vdev.
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * If a device reaches this threshold in a given txg then we consider skipping
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * allocations on that device.
b24ab6762772a3f6a89393947930c7fa61306783Jeff Bonwick * Metaslab debugging: when set, keeps all space maps in core to verify frees.
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson * Minimum size which forces the dynamic allocator to change
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * it's allocation strategy. Once the space map cannot satisfy
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson * an allocation of this size then it switches to using more
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson * aggressive strategy (i.e search by size rather than offset).
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilsonuint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson * The minimum free space, in percent, which must be available
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson * in a space map to continue allocations in a first-fit fashion.
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson * Once the space_map's free space drops below this level we dynamically
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson * switch to using best-fit allocations.
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * A metaslab is considered "free" if it contains a contiguous
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * segment which is greater than metaslab_min_alloc_size.
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilsonuint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * Max number of space_maps to prefetch.
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * Percentage bonus multiplier for metaslabs that are in the bonus area.
9dc3941c735ef88de46e850f745aa556d3a071a5Sašo Kiselkov * Should we be willing to write data to degraded vdevs?
fa9e4066f08beec538e775443c5be79dd423fcabahrens * ==========================================================================
fa9e4066f08beec538e775443c5be79dd423fcabahrens * Metaslab classes
fa9e4066f08beec538e775443c5be79dd423fcabahrens * ==========================================================================
88ecc943b4eb72f7c4fbbd8435997b85ef171fc3George Wilsonmetaslab_class_create(spa_t *spa, space_map_ops_t *ops)
fa9e4066f08beec538e775443c5be79dd423fcabahrens mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
88ecc943b4eb72f7c4fbbd8435997b85ef171fc3George Wilson * Must hold one of the spa_config locks.
88ecc943b4eb72f7c4fbbd8435997b85ef171fc3George Wilson ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
88ecc943b4eb72f7c4fbbd8435997b85ef171fc3George Wilson spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
b24ab6762772a3f6a89393947930c7fa61306783Jeff Bonwickmetaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
b24ab6762772a3f6a89393947930c7fa61306783Jeff Bonwick int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
b24ab6762772a3f6a89393947930c7fa61306783Jeff Bonwickmetaslab_class_get_deferred(metaslab_class_t *mc)
b24ab6762772a3f6a89393947930c7fa61306783Jeff Bonwick return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
fa9e4066f08beec538e775443c5be79dd423fcabahrens * ==========================================================================
fa9e4066f08beec538e775443c5be79dd423fcabahrens * Metaslab groups
fa9e4066f08beec538e775443c5be79dd423fcabahrens * ==========================================================================
fa9e4066f08beec538e775443c5be79dd423fcabahrens return (1);
fa9e4066f08beec538e775443c5be79dd423fcabahrens return (-1);
fa9e4066f08beec538e775443c5be79dd423fcabahrens * If the weights are identical, use the offset to force uniqueness.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson if (m1->ms_map->sm_start < m2->ms_map->sm_start)
fa9e4066f08beec538e775443c5be79dd423fcabahrens return (-1);
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson if (m1->ms_map->sm_start > m2->ms_map->sm_start)
fa9e4066f08beec538e775443c5be79dd423fcabahrens return (1);
fa9e4066f08beec538e775443c5be79dd423fcabahrens return (0);
fa9e4066f08beec538e775443c5be79dd423fcabahrensmetaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
fa9e4066f08beec538e775443c5be79dd423fcabahrens mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
fa9e4066f08beec538e775443c5be79dd423fcabahrens sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
a33cae9802e94744efee12a7a77c89360645eae8Tim Haley * We may have gone below zero with the activation count
a33cae9802e94744efee12a7a77c89360645eae8Tim Haley * either because we never activated in the first place or
a33cae9802e94744efee12a7a77c89360645eae8Tim Haley * because we're done, and possibly removing the vdev.
a15215608b8bd90f714f6db21ee623b584607cb6Jeff Bonwick ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
a15215608b8bd90f714f6db21ee623b584607cb6Jeff Bonwick mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
a15215608b8bd90f714f6db21ee623b584607cb6Jeff Bonwick ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwickmetaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
fa9e4066f08beec538e775443c5be79dd423fcabahrensmetaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
fa9e4066f08beec538e775443c5be79dd423fcabahrensmetaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
5f5f7a6f9c8e9c1587a54e690556d756ec67558cahrens * Although in principle the weight can be any value, in
5f5f7a6f9c8e9c1587a54e690556d756ec67558cahrens * practice we do not use values in the range [1, 510].
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * ==========================================================================
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * Common allocator routines
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * ==========================================================================
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilsonmetaslab_segsize_compare(const void *x1, const void *x2)
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson uint64_t ss_size1 = s1->ss_end - s1->ss_start;
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson uint64_t ss_size2 = s2->ss_end - s2->ss_start;
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson * This is a helper function that can be used by the allocator to find
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson * a suitable block to allocate. This will search the specified AVL
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson * tree looking for a block that matches the specified criteria.
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilsonmetaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * If we know we've searched the whole map (*cursor == 0), give up.
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * Otherwise, reset the cursor to the beginning and try again.
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick return (-1ULL);
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson return (metaslab_block_picker(t, cursor, size, align));
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson avl_create(sm->sm_pp_root, metaslab_segsize_compare,
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node));
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) {
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson /* tear down the tree */
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson kmem_free(sm->sm_pp_root, sizeof (avl_tree_t));
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick/* ARGSUSED */
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilsonmetaslab_pp_claim(space_map_t *sm, uint64_t start, uint64_t size)
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick /* No need to update cursor */
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick/* ARGSUSED */
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilsonmetaslab_pp_free(space_map_t *sm, uint64_t start, uint64_t size)
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick /* No need to update cursor */
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * Return the maximum contiguous segment within the metaslab.
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * ==========================================================================
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * The first-fit block allocator
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * ==========================================================================
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilsonmetaslab_ff_alloc(space_map_t *sm, uint64_t size)
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson return (metaslab_block_picker(t, cursor, size, align));
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson/* ARGSUSED */
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * ==========================================================================
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * Dynamic block allocator -
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * Uses the first fit allocation scheme until space get low and then
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * and metaslab_df_free_pct to determine when to switch the allocation scheme.
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * ==========================================================================
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilsonmetaslab_df_alloc(space_map_t *sm, uint64_t size)
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson int free_pct = sm->sm_space * 100 / sm->sm_size;
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson return (-1ULL);
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson * If we're running low on space switch to using the size
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson * sorted AVL tree (best-fit).
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson return (metaslab_block_picker(t, cursor, size, 1ULL));
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson int free_pct = sm->sm_space * 100 / sm->sm_size;
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson if (max_size >= metaslab_df_alloc_threshold &&
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * ==========================================================================
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * Other experimental allocators
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * ==========================================================================
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilsonmetaslab_cdf_alloc(space_map_t *sm, uint64_t size)
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson uint64_t *extent_end = (uint64_t *)sm->sm_ppd + 1;
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson return (-1ULL);
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * If we're running low on space switch to using the size
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * sorted AVL tree (best-fit).
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson rsize = MIN(metaslab_min_alloc_size, max_size);
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson offset = metaslab_block_picker(t, extent_end, rsize, 1ULL);
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson offset = metaslab_block_picker(t, cursor, rsize, 1ULL);
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson if (max_size > (metaslab_min_alloc_size * 10))
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilsonmetaslab_ndf_alloc(space_map_t *sm, uint64_t size)
8d18220deb04ec7b12410cd90deb4d45e66d49bfMark J Musante uint64_t *cursor = (uint64_t *)sm->sm_ppd + hbit - 1;
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson return (-1ULL);
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson if (ss == NULL || (ss->ss_start + size > ss->ss_end)) {
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson return (-1ULL);
8d18220deb04ec7b12410cd90deb4d45e66d49bfMark J Musante if (max_size > (metaslab_min_alloc_size << metaslab_ndf_clump_shift))
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilsonspace_map_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
fa9e4066f08beec538e775443c5be79dd423fcabahrens * ==========================================================================
fa9e4066f08beec538e775443c5be79dd423fcabahrens * Metaslabs
fa9e4066f08beec538e775443c5be79dd423fcabahrens * ==========================================================================
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwickmetaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * We create the main space map here, but we don't create the
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * allocmaps and freemaps until metaslab_sync_done(). This serves
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * two purposes: it allows metaslab_sync_done() to detect the
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * addition of new space; and for debugging, it ensures that we'd
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * data fault on any attempt to use this metaslab before it's ready.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson msp->ms_map = kmem_zalloc(sizeof (space_map_t), KM_SLEEP);
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson VERIFY(space_map_load(msp->ms_map, mg->mg_class->mc_ops,
b24ab6762772a3f6a89393947930c7fa61306783Jeff Bonwick SM_FREE, smo, spa_meta_objset(vd->vdev_spa)) == 0);
fa9e4066f08beec538e775443c5be79dd423fcabahrens * If we're opening an existing pool (txg == 0) or creating
fa9e4066f08beec538e775443c5be79dd423fcabahrens * a new one (txg == TXG_INITIAL), all space is available now.
fa9e4066f08beec538e775443c5be79dd423fcabahrens * If we're adding space to an existing pool, the new space
fa9e4066f08beec538e775443c5be79dd423fcabahrens * does not become available until after this txg has synced.
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick if (txg != 0) {
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson -msp->ms_smo.smo_alloc, 0, -msp->ms_map->sm_size);
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson kmem_free(msp->ms_map, sizeof (*msp->ms_map));
468c413a79615e77179e8d98f22a7e513a8135bdTim Haley for (int t = 0; t < TXG_SIZE; t++) {
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson kmem_free(msp->ms_allocmap[t], sizeof (*msp->ms_allocmap[t]));
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson kmem_free(msp->ms_freemap[t], sizeof (*msp->ms_freemap[t]));
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson for (int t = 0; t < TXG_DEFER_SIZE; t++) {
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson kmem_free(msp->ms_defermap[t], sizeof (*msp->ms_defermap[t]));
9eb57f7f3fbb970d4b9b89dcd5ecf543fe2414d5George Wilson * This vdev is in the process of being removed so there is nothing
9eb57f7f3fbb970d4b9b89dcd5ecf543fe2414d5George Wilson * for us to do here.
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * The baseline weight is the metaslab's free space.
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * Modern disks have uniform bit density and constant angular velocity.
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * Therefore, the outer recording zones are faster (higher bandwidth)
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * than the inner zones by the ratio of outer to inner track diameter,
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * which is typically around 2:1. We account for this by assigning
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * In effect, this means that we'll select the metaslab with the most
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * free bandwidth rather than simply the one with the most free space.
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick ((sm->sm_start >> vd->vdev_ms_shift) * weight) / vd->vdev_ms_count;
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * For locality, assign higher weight to metaslabs which have
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * a lower offset than what we've already activated.
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson weight <= 2 * (metaslab_smo_bonus_pct / 100) * space);
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson if (sm->sm_loaded && !sm->sm_ops->smop_fragmented(sm)) {
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * If this metaslab is one we're actively using, adjust its
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * weight to make it preferable to any inactive metaslab so
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * we'll polish it off.
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * Prefetch the next potential metaslabs
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson for (msp = avl_first(t), m = 0; msp; msp = AVL_NEXT(t, msp), m++) {
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson /* If we have reached our prefetch limit then we're done */
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson dmu_prefetch(spa_meta_objset(spa), smo->smo_object,
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilsonmetaslab_activate(metaslab_t *msp, uint64_t activation_weight)
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson space_map_ops_t *sm_ops = msp->ms_group->mg_class->mc_ops;
01f55e48fb4d524eaf70687728aa51b7762e2e97George Wilson int error = space_map_load(sm, sm_ops, SM_FREE, smo,
b24ab6762772a3f6a89393947930c7fa61306783Jeff Bonwick spa_meta_objset(msp->ms_group->mg_vd->vdev_spa));
468c413a79615e77179e8d98f22a7e513a8135bdTim Haley for (int t = 0; t < TXG_DEFER_SIZE; t++)
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * Track the bonus area as we activate new metaslabs.
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick return (0);
5f5f7a6f9c8e9c1587a54e690556d756ec67558cahrens * If size < SPA_MINBLOCKSIZE, then we will not allocate from
5f5f7a6f9c8e9c1587a54e690556d756ec67558cahrens * this metaslab again. In that case, it had better be empty,
5f5f7a6f9c8e9c1587a54e690556d756ec67558cahrens * or we would be leaving space on the table.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map->sm_space == 0);
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * Determine if the in-core space map representation can be condensed on-disk.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * We would like to use the following criteria to make our decision:
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * 1. The size of the space map object should not dramatically increase as a
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * result of writing out our in-core free map.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * 2. The minimal on-disk space map representation is zfs_condense_pct/100
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * times the size than the in-core representation (i.e. zfs_condense_pct = 110
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * and in-core = 1MB, minimal = 1.1.MB).
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * Checking the first condition is tricky since we don't want to walk
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * the entire AVL tree calculating the estimated on-disk size. Instead we
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * use the size-ordered AVL tree in the space map and calculate the
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * size required for the largest segment in our in-core free map. If the
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * size required to represent that segment on disk is larger than the space
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * map object then we avoid condensing this map.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * To determine the second criterion we use a best-case estimate and assume
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * each segment can be represented on-disk as a single 64-bit entry. We refer
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * to this best-case estimate as the space map's minimal form.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * Use the sm_pp_root AVL tree, which is ordered by size, to obtain
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * the largest segment in the in-core free map. If the tree is
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * empty then we should condense the map.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * Calculate the number of 64-bit entries this segment would
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * require when written to disk. If this single segment would be
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * larger on-disk than the entire current on-disk structure, then
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * clearly condensing will increase the on-disk structure size.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson size = (ss->ss_end - ss->ss_start) >> sm->sm_shift;
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) / 100);
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * Condense the on-disk space map representation to its minimized form.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * The minimized form consists of a small number of allocations followed by
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * the in-core free map.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilsonmetaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_t *freemap = msp->ms_freemap[txg & TXG_MASK];
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, "
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson (msp->ms_map->sm_start / msp->ms_map->sm_size), msp,
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson smo->smo_objsize, avl_numnodes(&sm->sm_root));
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * Create an map that is a 100% allocated map. We remove segments
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * that have been freed in this txg, any deferred frees that exist,
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * and any allocation in the future. Removing segments should be
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * a relatively inexpensive operation since we expect these maps to
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * a small number of nodes.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_create(&condense_map, sm->sm_start, sm->sm_size,
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_add(&condense_map, condense_map.sm_start,
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * Remove what's been freed in this txg from the condense_map.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * Since we're in sync_pass 1, we know that all the frees from
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * this txg are in the freemap.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_walk(freemap, space_map_remove, &condense_map);
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson for (int t = 0; t < TXG_DEFER_SIZE; t++)
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson for (int t = 1; t < TXG_CONCURRENT_STATES; t++)
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_walk(msp->ms_allocmap[(txg + t) & TXG_MASK],
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * We're about to drop the metaslab's lock thus allowing
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * other consumers to change it's content. Set the
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * space_map's sm_condensing flag to ensure that
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * allocations on this metaslab do not occur while we're
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * in the middle of committing it to disk. This is only critical
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * for the ms_map as all other space_maps use per txg
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * views of their content.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * While we would ideally like to create a space_map representation
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * that consists only of allocation records, doing so can be
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * prohibitively expensive because the in-core free map can be
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * large, and therefore computationally expensive to subtract
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * from the condense_map. Instead we sync out two maps, a cheap
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * allocation only map followed by the in-core free map. While not
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * optimal, this is typically close to optimal, and much cheaper to
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_sync(&condense_map, SM_ALLOC, smo, mos, tx);
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson spa_dbgmsg(spa, "condensed: txg %llu, msp[%llu] %p, "
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson (msp->ms_map->sm_start / msp->ms_map->sm_size), msp,
fa9e4066f08beec538e775443c5be79dd423fcabahrens * Write a metaslab to disk in the context of the specified transaction group.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_t *allocmap = msp->ms_allocmap[txg & TXG_MASK];
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_t **freemap = &msp->ms_freemap[txg & TXG_MASK];
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_t **freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * This metaslab has just been added so there's no work to do now.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson if (allocmap->sm_space == 0 && (*freemap)->sm_space == 0)
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * The only state that can actually be changing concurrently with
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * metaslab_sync() is the metaslab's ms_map. No other thread can
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * be modifying this txg's allocmap, freemap, freed_map, or smo.
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * Therefore, we only hold ms_lock to satify space_map ASSERTs.
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * We drop it whenever we call into the DMU, because the DMU
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * can call down to us (e.g. via zio_free()) at any time.
468c413a79615e77179e8d98f22a7e513a8135bdTim Haley tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson if (sm->sm_loaded && spa_sync_pass(spa) == 1 &&
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_sync(*freemap, SM_FREE, smo, mos, tx);
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * For sync pass 1, we avoid walking the entire space map and
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * instead will just swap the pointers for freemap and
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * freed_map. We can safely do this since the freed_map is
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * guaranteed to be empty on the initial pass.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson ASSERT0(avl_numnodes(&(*freed_map)->sm_root));
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_vacate(*freemap, space_map_add, *freed_map);
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson ASSERT0(msp->ms_allocmap[txg & TXG_MASK]->sm_space);
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson ASSERT0(msp->ms_freemap[txg & TXG_MASK]->sm_space);
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson VERIFY0(dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
fa9e4066f08beec538e775443c5be79dd423fcabahrens * Called after a transaction group has completely synced to mark
fa9e4066f08beec538e775443c5be79dd423fcabahrens * all of the metaslab's free space as usable.
9eb57f7f3fbb970d4b9b89dcd5ecf543fe2414d5George Wilson space_map_t **freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
9eb57f7f3fbb970d4b9b89dcd5ecf543fe2414d5George Wilson space_map_t **defer_map = &msp->ms_defermap[txg % TXG_DEFER_SIZE];
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * If this metaslab is just becoming available, initialize its
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson * allocmaps, freemaps, and defermap and add its capacity to the vdev.
468c413a79615e77179e8d98f22a7e513a8135bdTim Haley for (int t = 0; t < TXG_SIZE; t++) {
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson msp->ms_allocmap[t] = kmem_zalloc(sizeof (space_map_t),
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_create(msp->ms_allocmap[t], sm->sm_start,
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson msp->ms_freemap[t] = kmem_zalloc(sizeof (space_map_t),
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_create(msp->ms_freemap[t], sm->sm_start,
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson for (int t = 0; t < TXG_DEFER_SIZE; t++) {
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson msp->ms_defermap[t] = kmem_zalloc(sizeof (space_map_t),
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_create(msp->ms_defermap[t], sm->sm_start,
9eb57f7f3fbb970d4b9b89dcd5ecf543fe2414d5George Wilson freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
9eb57f7f3fbb970d4b9b89dcd5ecf543fe2414d5George Wilson defer_map = &msp->ms_defermap[txg % TXG_DEFER_SIZE];
468c413a79615e77179e8d98f22a7e513a8135bdTim Haley alloc_delta = smosync->smo_alloc - smo->smo_alloc;
9eb57f7f3fbb970d4b9b89dcd5ecf543fe2414d5George Wilson defer_delta = (*freed_map)->sm_space - (*defer_map)->sm_space;
b24ab6762772a3f6a89393947930c7fa61306783Jeff Bonwick vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson ASSERT(msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0);
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson ASSERT(msp->ms_freemap[txg & TXG_MASK]->sm_space == 0);
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * If there's a space_map_load() in progress, wait for it to complete
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * so that we have a consistent view of the in-core space map.
9eb57f7f3fbb970d4b9b89dcd5ecf543fe2414d5George Wilson * Move the frees from the defer_map to this map (if it's loaded).
9eb57f7f3fbb970d4b9b89dcd5ecf543fe2414d5George Wilson * Swap the freed_map and the defer_map -- this is safe to do
9eb57f7f3fbb970d4b9b89dcd5ecf543fe2414d5George Wilson * because we've just emptied out the defer_map.
9eb57f7f3fbb970d4b9b89dcd5ecf543fe2414d5George Wilson space_map_vacate(*defer_map, sm->sm_loaded ? space_map_free : NULL, sm);
9eb57f7f3fbb970d4b9b89dcd5ecf543fe2414d5George Wilson ASSERT0(avl_numnodes(&(*defer_map)->sm_root));
468c413a79615e77179e8d98f22a7e513a8135bdTim Haley * Keep syncing this metaslab until all deferred frees
468c413a79615e77179e8d98f22a7e513a8135bdTim Haley * are back in circulation.
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * If the map is loaded but no longer active, evict it as soon as all
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * future allocations have synced. (If we unloaded it now and then
ecc2d604e885a75cc75e647b5641af99d5a6f4a6bonwick * loaded a moment later, the map wouldn't reflect those allocations.)
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm if (sm->sm_loaded && (msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson if (msp->ms_allocmap[(txg + t) & TXG_MASK]->sm_space)
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * Re-evaluate all metaslabs which have lower offsets than the
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * bonus area.
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson if (msp->ms_map->sm_start > mg->mg_bonus_area)
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson metaslab_group_sort(mg, msp, metaslab_weight(msp));
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson atomic_add_64(&mg->mg_alloc_failures, -failures);
80eb36f241abf8c076119fb4c49a55fd61ebc710George Wilson * Prefetch the next potential metaslabs
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson uint64_t start = msp->ms_map->sm_start >> ms_shift;
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm return (0);
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilsonmetaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson uint64_t txg, uint64_t min_distance, dva_t *dva, int d, int flags)
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson for (i = 0; i < d; i++) {
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson "requirement: vdev %llu, txg %llu, mg %p, "
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson "msp %p, psize %llu, asize %llu, "
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson "failures %llu, weight %llu",
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm return (-1ULL);
03f8c366886542ed249a15d755ae78ea4e775d9dGeorge Wilson * If the selected metaslab is condensing, skip it.
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm for (i = 0; i < d; i++)
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm if (i == d)
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm return (-1ULL);
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * If we've already reached the allowable number of failed
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * allocation attempts on this metaslab group then we
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * consider skipping it. We skip it only if we're allowed
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * to "fast" gang, the physical size is larger than
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * a gang block, and we're attempting to allocate from
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * the primary metaslab.
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson if (mg->mg_alloc_failures > zfs_mg_alloc_failures &&
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson CAN_FASTGANG(flags) && psize > SPA_GANGBLOCKSIZE &&
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson activation_weight == METASLAB_WEIGHT_PRIMARY) {
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson spa_dbgmsg(spa, "%s: skipping metaslab group: "
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson "vdev %llu, txg %llu, mg %p, psize %llu, "
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson return (-1ULL);
aeb1c1b609b02f03e8e7448beb88384ebc713525gw * Ensure that the metaslab we have selected is still
aeb1c1b609b02f03e8e7448beb88384ebc713525gw * capable of handling our request. It's possible that
aeb1c1b609b02f03e8e7448beb88384ebc713525gw * another thread may have changed the weight while we
aeb1c1b609b02f03e8e7448beb88384ebc713525gw * were blocked on the metaslab lock.
d6e555bdd793b8bc8fe57d5f12c3d69c813d0661George Wilson activation_weight == METASLAB_WEIGHT_PRIMARY)) {
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson if (metaslab_activate(msp, activation_weight) != 0) {
03f8c366886542ed249a15d755ae78ea4e775d9dGeorge Wilson * If this metaslab is currently condensing then pick again as
03f8c366886542ed249a15d755ae78ea4e775d9dGeorge Wilson * we can't manipulate this metaslab until it's committed
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson if ((offset = space_map_alloc(msp->ms_map, asize)) != -1ULL)
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson metaslab_passivate(msp, space_map_maxsize(msp->ms_map));
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson if (msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0)
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_add(msp->ms_allocmap[txg & TXG_MASK], offset, asize);
fa9e4066f08beec538e775443c5be79dd423fcabahrens * Allocate a block for the specified i/o.
8654d0253136055bd4cc2423d87378e8a37f2eb5perrinmetaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
e14bb3258d05c1b1077e2db7cf77088924e56919Jeff Bonwick dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
e05725b117836db173257fae43fb0746eb857fb5bonwick * For testing, make some blocks above a certain size be gang blocks.
d3d50737e566cade9a08d73d2af95105ac7cd960Rafael Vanoni if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
fa9e4066f08beec538e775443c5be79dd423fcabahrens * Start at the rotor and loop through all mgs until we find something.
b24ab6762772a3f6a89393947930c7fa61306783Jeff Bonwick * Note that there's no locking on mc_rotor or mc_aliquot because
fa9e4066f08beec538e775443c5be79dd423fcabahrens * nothing actually breaks if we miss a few updates -- we just won't
fa9e4066f08beec538e775443c5be79dd423fcabahrens * allocate quite as evenly. It all balances out over time.
67bd71c6cc629bab3aa0d595c624a667f1574254perrin * If we are doing ditto or log blocks, try to spread them across
67bd71c6cc629bab3aa0d595c624a667f1574254perrin * consecutive vdevs. If we're forced to reuse a vdev before we've
67bd71c6cc629bab3aa0d595c624a667f1574254perrin * allocated all of our ditto blocks, then try and spread them out on
67bd71c6cc629bab3aa0d595c624a667f1574254perrin * that vdev as much as possible. If it turns out to not be possible,
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm * gradually lower our standards until anything becomes acceptable.
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm * Also, allocating on consecutive vdevs (as opposed to random vdevs)
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm * gives us hope of containing our fault domains to something we're
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm * able to reason about. Otherwise, any two top-level vdev failures
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm * will guarantee the loss of data. With consecutive allocation,
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm * only two adjacent top-level vdev failures will result in data loss.
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm * If we are doing gang blocks (hintdva is non-NULL), try to keep
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm * ourselves on the same vdev as our gang block header. That
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm * way, we can hope for locality in vdev_cache, plus it makes our
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm * fault domains something tractable.
88ecc943b4eb72f7c4fbbd8435997b85ef171fc3George Wilson * It's possible the vdev we're using as the hint no
88ecc943b4eb72f7c4fbbd8435997b85ef171fc3George Wilson * longer exists (i.e. removed). Consult the rotor when
88ecc943b4eb72f7c4fbbd8435997b85ef171fc3George Wilson * all else fails.
44cd46cadd9aab751dae6a4023c1cb5bf316d274billm } else if (d != 0) {
a15215608b8bd90f714f6db21ee623b584607cb6Jeff Bonwick * If the hint put us into the wrong metaslab class, or into a
a15215608b8bd90f714f6db21ee623b584607cb6Jeff Bonwick * metaslab group that has been passivated, just follow the rotor.
a15215608b8bd90f714f6db21ee623b584607cb6Jeff Bonwick if (mg->mg_class != mc || mg->mg_activation_count <= 0)
e14bb3258d05c1b1077e2db7cf77088924e56919Jeff Bonwick * Don't allocate from faulted devices.
8ad4d6dd86f5bc65fb3afa566c8133f3bac21648Jeff Bonwick spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
0a4e9518a44f226be6d39383330b5b1792d2f184gw * Avoid writing single-copy data to a failing vdev
9dc3941c735ef88de46e850f745aa556d3a071a5Sašo Kiselkov * unless the user instructs us that it is okay.
fa9e4066f08beec538e775443c5be79dd423fcabahrens ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
fa9e4066f08beec538e775443c5be79dd423fcabahrens * If we've just selected this metaslab group,
fa9e4066f08beec538e775443c5be79dd423fcabahrens * figure out whether the corresponding vdev is
fa9e4066f08beec538e775443c5be79dd423fcabahrens * over- or under-used relative to the pool,
fa9e4066f08beec538e775443c5be79dd423fcabahrens * and set an allocation bias to even it out.
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * Calculate how much more or less we should
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * try to allocate from this device during
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * this iteration around the rotor.
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * For example, if a device is 80% full
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * and the pool is 20% full then we should
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * reduce allocations by 60% on this device.
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * mg_bias = (20 - 80) * 512K / 100 = -307K
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson * This reduces allocations by 307K for this
b24ab6762772a3f6a89393947930c7fa61306783Jeff Bonwick if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
e14bb3258d05c1b1077e2db7cf77088924e56919Jeff Bonwick DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
fa9e4066f08beec538e775443c5be79dd423fcabahrens return (0);
fa9e4066f08beec538e775443c5be79dd423fcabahrens * Free the block represented by DVA in the context of the specified
fa9e4066f08beec538e775443c5be79dd423fcabahrens * transaction group.
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwickmetaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwick (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwick cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_remove(msp->ms_allocmap[txg & TXG_MASK],
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson if (msp->ms_freemap[txg & TXG_MASK]->sm_space == 0)
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_add(msp->ms_freemap[txg & TXG_MASK], offset, size);
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwick * Intent log support: upon opening the pool after a crash, notify the SPA
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwick * of blocks that the intent log has allocated for immediate write, but
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwick * which are still considered free by the SPA because the last transaction
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwick * group didn't commit yet.
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwickmetaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson if ((txg != 0 && spa_writeable(spa)) || !msp->ms_map->sm_loaded)
09c9d376e8ccb8fbba74f33cc268964464092b62George Wilson error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson if (error == 0 && !space_map_contains(msp->ms_map, offset, size))
e14bb3258d05c1b1077e2db7cf77088924e56919Jeff Bonwick if (error || txg == 0) { /* txg == 0 indicates dry run */
8ad4d6dd86f5bc65fb3afa566c8133f3bac21648Jeff Bonwick if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson if (msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0)
16a4a8074274d2d7cc408589cf6359f4a378c861George Wilson space_map_add(msp->ms_allocmap[txg & TXG_MASK], offset, size);
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwick return (0);
8654d0253136055bd4cc2423d87378e8a37f2eb5perrinmetaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
e14bb3258d05c1b1077e2db7cf77088924e56919Jeff Bonwick int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
e14bb3258d05c1b1077e2db7cf77088924e56919Jeff Bonwick spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
e14bb3258d05c1b1077e2db7cf77088924e56919Jeff Bonwick if (mc->mc_rotor == NULL) { /* no vdevs in this class */
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwick ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwick ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
e14bb3258d05c1b1077e2db7cf77088924e56919Jeff Bonwick for (int d = 0; d < ndvas; d++) {
8654d0253136055bd4cc2423d87378e8a37f2eb5perrin error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwick for (d--; d >= 0; d--) {
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwick return (0);
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwickmetaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
b24ab6762772a3f6a89393947930c7fa61306783Jeff Bonwick ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
e14bb3258d05c1b1077e2db7cf77088924e56919Jeff Bonwick spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
e14bb3258d05c1b1077e2db7cf77088924e56919Jeff Bonwick for (int d = 0; d < ndvas; d++)
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwickmetaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
e14bb3258d05c1b1077e2db7cf77088924e56919Jeff Bonwick * First do a dry run to make sure all DVAs are claimable,
e14bb3258d05c1b1077e2db7cf77088924e56919Jeff Bonwick * so we don't have to unwind from partial failures below.
e14bb3258d05c1b1077e2db7cf77088924e56919Jeff Bonwick spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
e14bb3258d05c1b1077e2db7cf77088924e56919Jeff Bonwick for (int d = 0; d < ndvas; d++)
d80c45e0f58fa434ba37259ea2e2b12e0380c19abonwick if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)