From 67a1b0379159c46bcd60a462a2790248046c8804 Mon Sep 17 00:00:00 2001 From: Pawel Jakub Dawidek Date: Fri, 10 Mar 2023 20:59:53 +0100 Subject: [PATCH] Implementation of block cloning for ZFS Block Cloning allows to manually clone a file (or a subset of its blocks) into another (or the same) file by just creating additional references to the data blocks without copying the data itself. Those references are kept in the Block Reference Tables (BRTs). The whole design of block cloning is documented in module/zfs/brt.c. Reviewed-by: Alexander Motin Reviewed-by: Christian Schwarz Reviewed-by: Brian Behlendorf Reviewed-by: Rich Ercolani Signed-off-by: Pawel Jakub Dawidek Closes #13392 --- cmd/zdb/zdb_il.c | 19 + cmd/ztest.c | 2 +- include/Makefile.am | 2 + include/os/freebsd/zfs/sys/zfs_znode_impl.h | 5 +- include/os/linux/kernel/linux/mod_compat.h | 1 + include/os/linux/zfs/sys/zfs_znode_impl.h | 3 +- include/sys/bitmap.h | 93 + include/sys/brt.h | 62 + include/sys/dbuf.h | 1 + include/sys/ddt.h | 2 + include/sys/dmu.h | 8 + include/sys/dmu_tx.h | 1 + include/sys/fs/zfs.h | 3 + include/sys/spa_impl.h | 1 + include/sys/zfs_debug.h | 1 + include/sys/zfs_vnops.h | 4 + include/sys/zfs_znode.h | 3 + include/sys/zil.h | 25 +- include/sys/zio.h | 3 +- include/sys/zio_impl.h | 41 +- include/zfeature_common.h | 1 + lib/libzfs/libzfs.abi | 14 +- lib/libzfs/libzfs_pool.c | 3 + lib/libzpool/Makefile.am | 1 + man/man7/zpool-features.7 | 14 + man/man7/zpoolprops.7 | 27 +- module/Kbuild.in | 1 + module/Makefile.bsd | 3 +- module/os/freebsd/zfs/sysctl_os.c | 2 + module/os/freebsd/zfs/zfs_vfsops.c | 7 +- module/os/freebsd/zfs/zfs_vnops_os.c | 95 +- module/os/freebsd/zfs/zfs_znode.c | 26 + module/zcommon/zfeature_common.c | 6 + module/zcommon/zpool_prop.c | 9 + module/zfs/brt.c | 1884 +++++++++++++++++ module/zfs/dbuf.c | 126 +- module/zfs/ddt.c | 55 + module/zfs/dmu.c | 153 +- module/zfs/dmu_tx.c | 47 +- module/zfs/dsl_scan.c | 8 +- module/zfs/spa.c | 39 + module/zfs/spa_misc.c | 5 +- module/zfs/zfs_ioctl.c | 2 +- module/zfs/zfs_log.c | 52 + module/zfs/zfs_quota.c | 3 +- module/zfs/zfs_replay.c | 30 + module/zfs/zfs_vnops.c | 467 +++- module/zfs/zil.c | 126 +- module/zfs/zio.c | 55 +- module/zfs/zvol.c | 55 + .../cli_root/zpool_get/zpool_get.cfg | 4 + 51 files changed, 3480 insertions(+), 120 deletions(-) create mode 100644 include/sys/bitmap.h create mode 100644 include/sys/brt.h create mode 100644 module/zfs/brt.c diff --git a/cmd/zdb/zdb_il.c b/cmd/zdb/zdb_il.c index 55df1f559f6e..970c45c9b3bb 100644 --- a/cmd/zdb/zdb_il.c +++ b/cmd/zdb/zdb_il.c @@ -307,6 +307,23 @@ zil_prt_rec_acl(zilog_t *zilog, int txtype, const void *arg) (u_longlong_t)lr->lr_foid, (u_longlong_t)lr->lr_aclcnt); } +static void +zil_prt_rec_clone_range(zilog_t *zilog, int txtype, const void *arg) +{ + (void) zilog, (void) txtype; + const lr_clone_range_t *lr = arg; + + (void) printf("%sfoid %llu, offset %llx, length %llx, blksize %llx\n", + tab_prefix, (u_longlong_t)lr->lr_foid, (u_longlong_t)lr->lr_offset, + (u_longlong_t)lr->lr_length, (u_longlong_t)lr->lr_blksz); + + for (unsigned int i = 0; i < lr->lr_nbps; i++) { + (void) printf("%s[%u/%llu] ", tab_prefix, i + 1, + (u_longlong_t)lr->lr_nbps); + print_log_bp(&lr->lr_bps[i], ""); + } +} + typedef void (*zil_prt_rec_func_t)(zilog_t *, int, const void *); typedef struct zil_rec_info { zil_prt_rec_func_t zri_print; @@ -340,6 +357,8 @@ static zil_rec_info_t zil_rec_info[TX_MAX_TYPE] = { .zri_name = "TX_SETSAXATTR "}, {.zri_print = zil_prt_rec_rename, .zri_name = "TX_RENAME_EXCHANGE "}, {.zri_print = zil_prt_rec_rename, .zri_name = "TX_RENAME_WHITEOUT "}, + {.zri_print = zil_prt_rec_clone_range, + .zri_name = "TX_CLONE_RANGE "}, }; static int diff --git a/cmd/ztest.c b/cmd/ztest.c index fb9f83032e8f..b6b99bfff6db 100644 --- a/cmd/ztest.c +++ b/cmd/ztest.c @@ -1902,7 +1902,7 @@ ztest_log_write(ztest_ds_t *zd, dmu_tx_t *tx, lr_write_t *lr) if (zil_replaying(zd->zd_zilog, tx)) return; - if (lr->lr_length > zil_max_log_data(zd->zd_zilog)) + if (lr->lr_length > zil_max_log_data(zd->zd_zilog, sizeof (lr_write_t))) write_state = WR_INDIRECT; itx = zil_itx_create(TX_WRITE, diff --git a/include/Makefile.am b/include/Makefile.am index 6897e3c5e337..569de6dfa781 100644 --- a/include/Makefile.am +++ b/include/Makefile.am @@ -23,6 +23,7 @@ COMMON_H = \ sys/asm_linkage.h \ sys/avl.h \ sys/avl_impl.h \ + sys/bitmap.h \ sys/bitops.h \ sys/blake3.h \ sys/blkptr.h \ @@ -31,6 +32,7 @@ COMMON_H = \ sys/bptree.h \ sys/bqueue.h \ sys/btree.h \ + sys/brt.h \ sys/dataset_kstats.h \ sys/dbuf.h \ sys/ddt.h \ diff --git a/include/os/freebsd/zfs/sys/zfs_znode_impl.h b/include/os/freebsd/zfs/sys/zfs_znode_impl.h index 8cde33dbcbbb..050fc3036f87 100644 --- a/include/os/freebsd/zfs/sys/zfs_znode_impl.h +++ b/include/os/freebsd/zfs/sys/zfs_znode_impl.h @@ -119,7 +119,8 @@ typedef struct zfs_soft_state { #define zn_has_cached_data(zp, start, end) \ vn_has_cached_data(ZTOV(zp)) #define zn_flush_cached_data(zp, sync) vn_flush_cached_data(ZTOV(zp), sync) -#define zn_rlimit_fsize(zp, uio) \ +#define zn_rlimit_fsize(size) zfs_rlimit_fsize(size) +#define zn_rlimit_fsize_uio(zp, uio) \ vn_rlimit_fsize(ZTOV(zp), GET_UIO_STRUCT(uio), zfs_uio_td(uio)) /* Called on entry to each ZFS vnode and vfs operation */ @@ -179,6 +180,8 @@ extern zil_replay_func_t *const zfs_replay_vector[TX_MAX_TYPE]; extern int zfs_znode_parent_and_name(struct znode *zp, struct znode **dzpp, char *buf); + +extern int zfs_rlimit_fsize(off_t fsize); #ifdef __cplusplus } #endif diff --git a/include/os/linux/kernel/linux/mod_compat.h b/include/os/linux/kernel/linux/mod_compat.h index 67b4fc90653c..09d109d191bf 100644 --- a/include/os/linux/kernel/linux/mod_compat.h +++ b/include/os/linux/kernel/linux/mod_compat.h @@ -47,6 +47,7 @@ typedef const struct kernel_param zfs_kernel_param_t; enum scope_prefix_types { zfs, zfs_arc, + zfs_brt, zfs_condense, zfs_dbuf, zfs_dbuf_cache, diff --git a/include/os/linux/zfs/sys/zfs_znode_impl.h b/include/os/linux/zfs/sys/zfs_znode_impl.h index 81607ef2a25e..0be2c445ab76 100644 --- a/include/os/linux/zfs/sys/zfs_znode_impl.h +++ b/include/os/linux/zfs/sys/zfs_znode_impl.h @@ -86,7 +86,8 @@ extern "C" { #endif #define zn_flush_cached_data(zp, sync) write_inode_now(ZTOI(zp), sync) -#define zn_rlimit_fsize(zp, uio) (0) +#define zn_rlimit_fsize(size) (0) +#define zn_rlimit_fsize_uio(zp, uio) (0) /* * zhold() wraps igrab() on Linux, and igrab() may fail when the diff --git a/include/sys/bitmap.h b/include/sys/bitmap.h new file mode 100644 index 000000000000..7b92507a7930 --- /dev/null +++ b/include/sys/bitmap.h @@ -0,0 +1,93 @@ +/* + * CDDL HEADER START + * + * The contents of this file are subject to the terms of the + * Common Development and Distribution License (the "License"). + * You may not use this file except in compliance with the License. + * + * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE + * or http://www.opensolaris.org/os/licensing. + * See the License for the specific language governing permissions + * and limitations under the License. + * + * When distributing Covered Code, include this CDDL HEADER in each + * file and include the License file at usr/src/OPENSOLARIS.LICENSE. + * If applicable, add the following below this CDDL HEADER, with the + * fields enclosed by brackets "[]" replaced with your own identifying + * information: Portions Copyright [yyyy] [name of copyright owner] + * + * CDDL HEADER END + */ + +/* + * Copyright 2006 Sun Microsystems, Inc. All rights reserved. + * Use is subject to license terms. + */ + +/* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */ +/* All Rights Reserved */ + + +#ifndef _SYS_BITMAP_H +#define _SYS_BITMAP_H + +#ifdef __cplusplus +extern "C" { +#endif + +/* + * Operations on bitmaps of arbitrary size + * A bitmap is a vector of 1 or more ulong_t's. + * The user of the package is responsible for range checks and keeping + * track of sizes. + */ + +#ifdef _LP64 +#define BT_ULSHIFT 6 /* log base 2 of BT_NBIPUL, to extract word index */ +#define BT_ULSHIFT32 5 /* log base 2 of BT_NBIPUL, to extract word index */ +#else +#define BT_ULSHIFT 5 /* log base 2 of BT_NBIPUL, to extract word index */ +#endif + +#define BT_NBIPUL (1 << BT_ULSHIFT) /* n bits per ulong_t */ +#define BT_ULMASK (BT_NBIPUL - 1) /* to extract bit index */ + +/* + * bitmap is a ulong_t *, bitindex an index_t + * + * The macros BT_WIM and BT_BIW internal; there is no need + * for users of this package to use them. + */ + +/* + * word in map + */ +#define BT_WIM(bitmap, bitindex) \ + ((bitmap)[(bitindex) >> BT_ULSHIFT]) +/* + * bit in word + */ +#define BT_BIW(bitindex) \ + (1UL << ((bitindex) & BT_ULMASK)) + +/* + * These are public macros + * + * BT_BITOUL == n bits to n ulong_t's + */ +#define BT_BITOUL(nbits) \ + (((nbits) + BT_NBIPUL - 1l) / BT_NBIPUL) +#define BT_SIZEOFMAP(nbits) \ + (BT_BITOUL(nbits) * sizeof (ulong_t)) +#define BT_TEST(bitmap, bitindex) \ + ((BT_WIM((bitmap), (bitindex)) & BT_BIW(bitindex)) ? 1 : 0) +#define BT_SET(bitmap, bitindex) \ + { BT_WIM((bitmap), (bitindex)) |= BT_BIW(bitindex); } +#define BT_CLEAR(bitmap, bitindex) \ + { BT_WIM((bitmap), (bitindex)) &= ~BT_BIW(bitindex); } + +#ifdef __cplusplus +} +#endif + +#endif /* _SYS_BITMAP_H */ diff --git a/include/sys/brt.h b/include/sys/brt.h new file mode 100644 index 000000000000..b1f70107787c --- /dev/null +++ b/include/sys/brt.h @@ -0,0 +1,62 @@ +/* + * CDDL HEADER START + * + * The contents of this file are subject to the terms of the + * Common Development and Distribution License (the "License"). + * You may not use this file except in compliance with the License. + * + * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE + * or http://www.opensolaris.org/os/licensing. + * See the License for the specific language governing permissions + * and limitations under the License. + * + * When distributing Covered Code, include this CDDL HEADER in each + * file and include the License file at usr/src/OPENSOLARIS.LICENSE. + * If applicable, add the following below this CDDL HEADER, with the + * fields enclosed by brackets "[]" replaced with your own identifying + * information: Portions Copyright [yyyy] [name of copyright owner] + * + * CDDL HEADER END + */ +/* + * Copyright (c) 2020, 2021, 2022 by Pawel Jakub Dawidek + */ + +#ifndef _SYS_BRT_H +#define _SYS_BRT_H + +#include +#include +#include +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +extern boolean_t brt_entry_decref(spa_t *spa, const blkptr_t *bp); + +extern uint64_t brt_get_dspace(spa_t *spa); +extern uint64_t brt_get_used(spa_t *spa); +extern uint64_t brt_get_saved(spa_t *spa); +extern uint64_t brt_get_ratio(spa_t *spa); + +extern boolean_t brt_maybe_exists(spa_t *spa, const blkptr_t *bp); +extern void brt_init(void); +extern void brt_fini(void); + +extern void brt_pending_add(spa_t *spa, const blkptr_t *bp, dmu_tx_t *tx); +extern void brt_pending_remove(spa_t *spa, const blkptr_t *bp, dmu_tx_t *tx); +extern void brt_pending_apply(spa_t *spa, uint64_t txg); + +extern void brt_create(spa_t *spa); +extern int brt_load(spa_t *spa); +extern void brt_unload(spa_t *spa); +extern void brt_sync(spa_t *spa, uint64_t txg); + +#ifdef __cplusplus +} +#endif + +#endif /* _SYS_BRT_H */ diff --git a/include/sys/dbuf.h b/include/sys/dbuf.h index a1ce76b1c763..a06316362e57 100644 --- a/include/sys/dbuf.h +++ b/include/sys/dbuf.h @@ -172,6 +172,7 @@ typedef struct dbuf_dirty_record { override_states_t dr_override_state; uint8_t dr_copies; boolean_t dr_nopwrite; + boolean_t dr_brtwrite; boolean_t dr_has_raw_params; /* diff --git a/include/sys/ddt.h b/include/sys/ddt.h index d72401dcf7a4..6378c042c705 100644 --- a/include/sys/ddt.h +++ b/include/sys/ddt.h @@ -248,6 +248,8 @@ extern int ddt_walk(spa_t *spa, ddt_bookmark_t *ddb, ddt_entry_t *dde); extern int ddt_object_update(ddt_t *ddt, enum ddt_type type, enum ddt_class clazz, ddt_entry_t *dde, dmu_tx_t *tx); +extern boolean_t ddt_addref(spa_t *spa, const blkptr_t *bp); + extern const ddt_ops_t ddt_zap_ops; #ifdef __cplusplus diff --git a/include/sys/dmu.h b/include/sys/dmu.h index 93de991ccd86..1b82ff620f27 100644 --- a/include/sys/dmu.h +++ b/include/sys/dmu.h @@ -782,6 +782,8 @@ dmu_tx_t *dmu_tx_create(objset_t *os); void dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len); void dmu_tx_hold_write_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len); +void dmu_tx_hold_clone_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, + int len); void dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len); void dmu_tx_hold_free_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, @@ -1059,6 +1061,12 @@ int dmu_sync(struct zio *zio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd); int dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off); +int dmu_read_l0_bps(objset_t *os, uint64_t object, uint64_t offset, + uint64_t length, dmu_tx_t *tx, struct blkptr *bps, size_t *nbpsp); +void dmu_brt_clone(objset_t *os, uint64_t object, uint64_t offset, + uint64_t length, dmu_tx_t *tx, const struct blkptr *bps, size_t nbps, + boolean_t replay); + /* * Initial setup and final teardown. */ diff --git a/include/sys/dmu_tx.h b/include/sys/dmu_tx.h index 81e1ef6c1477..ca8514e5d2d0 100644 --- a/include/sys/dmu_tx.h +++ b/include/sys/dmu_tx.h @@ -90,6 +90,7 @@ enum dmu_tx_hold_type { THT_ZAP, THT_SPACE, THT_SPILL, + THT_CLONE, THT_NUMTYPES }; diff --git a/include/sys/fs/zfs.h b/include/sys/fs/zfs.h index e869685c5e2d..25babd4ea8cf 100644 --- a/include/sys/fs/zfs.h +++ b/include/sys/fs/zfs.h @@ -253,6 +253,9 @@ typedef enum { ZPOOL_PROP_LOAD_GUID, ZPOOL_PROP_AUTOTRIM, ZPOOL_PROP_COMPATIBILITY, + ZPOOL_PROP_BCLONEUSED, + ZPOOL_PROP_BCLONESAVED, + ZPOOL_PROP_BCLONERATIO, ZPOOL_NUM_PROPS } zpool_prop_t; diff --git a/include/sys/spa_impl.h b/include/sys/spa_impl.h index cde08ec9b0f4..8ccd58b584ca 100644 --- a/include/sys/spa_impl.h +++ b/include/sys/spa_impl.h @@ -380,6 +380,7 @@ struct spa { uint64_t spa_dedup_dspace; /* Cache get_dedup_dspace() */ uint64_t spa_dedup_checksum; /* default dedup checksum */ uint64_t spa_dspace; /* dspace in normal class */ + struct brt *spa_brt; /* in-core BRT */ kmutex_t spa_vdev_top_lock; /* dueling offline/remove */ kmutex_t spa_proc_lock; /* protects spa_proc* */ kcondvar_t spa_proc_cv; /* spa_proc_state transitions */ diff --git a/include/sys/zfs_debug.h b/include/sys/zfs_debug.h index 481209b241aa..a1dfef1d89ff 100644 --- a/include/sys/zfs_debug.h +++ b/include/sys/zfs_debug.h @@ -57,6 +57,7 @@ extern int zfs_dbgmsg_enable; #define ZFS_DEBUG_TRIM (1 << 11) #define ZFS_DEBUG_LOG_SPACEMAP (1 << 12) #define ZFS_DEBUG_METASLAB_ALLOC (1 << 13) +#define ZFS_DEBUG_BRT (1 << 14) extern void __set_error(const char *file, const char *func, int line, int err); extern void __zfs_dbgmsg(char *buf); diff --git a/include/sys/zfs_vnops.h b/include/sys/zfs_vnops.h index edff8f681dd4..5da103f17783 100644 --- a/include/sys/zfs_vnops.h +++ b/include/sys/zfs_vnops.h @@ -31,6 +31,10 @@ extern int zfs_read(znode_t *, zfs_uio_t *, int, cred_t *); extern int zfs_write(znode_t *, zfs_uio_t *, int, cred_t *); extern int zfs_holey(znode_t *, ulong_t, loff_t *); extern int zfs_access(znode_t *, int, int, cred_t *); +extern int zfs_clone_range(znode_t *, uint64_t *, znode_t *, uint64_t *, + uint64_t *, cred_t *); +extern int zfs_clone_range_replay(znode_t *, uint64_t, uint64_t, uint64_t, + const blkptr_t *, size_t); extern int zfs_getsecattr(znode_t *, vsecattr_t *, int, cred_t *); extern int zfs_setsecattr(znode_t *, vsecattr_t *, int, cred_t *); diff --git a/include/sys/zfs_znode.h b/include/sys/zfs_znode.h index fcee55b0199d..012e7403e2a6 100644 --- a/include/sys/zfs_znode.h +++ b/include/sys/zfs_znode.h @@ -315,6 +315,9 @@ extern void zfs_log_setattr(zilog_t *zilog, dmu_tx_t *tx, int txtype, znode_t *zp, vattr_t *vap, uint_t mask_applied, zfs_fuid_info_t *fuidp); extern void zfs_log_acl(zilog_t *zilog, dmu_tx_t *tx, znode_t *zp, vsecattr_t *vsecp, zfs_fuid_info_t *fuidp); +extern void zfs_log_clone_range(zilog_t *zilog, dmu_tx_t *tx, int txtype, + znode_t *zp, uint64_t offset, uint64_t length, uint64_t blksz, + const blkptr_t *bps, size_t nbps); extern void zfs_xvattr_set(znode_t *zp, xvattr_t *xvap, dmu_tx_t *tx); extern void zfs_upgrade(zfsvfs_t *zfsvfs, dmu_tx_t *tx); extern void zfs_log_setsaxattr(zilog_t *zilog, dmu_tx_t *tx, int txtype, diff --git a/include/sys/zil.h b/include/sys/zil.h index 9ac421043377..cff8ebcad819 100644 --- a/include/sys/zil.h +++ b/include/sys/zil.h @@ -166,7 +166,8 @@ typedef enum zil_create { #define TX_SETSAXATTR 21 /* Set sa xattrs on file */ #define TX_RENAME_EXCHANGE 22 /* Atomic swap via renameat2 */ #define TX_RENAME_WHITEOUT 23 /* Atomic whiteout via renameat2 */ -#define TX_MAX_TYPE 24 /* Max transaction type */ +#define TX_CLONE_RANGE 24 /* Clone a file range */ +#define TX_MAX_TYPE 25 /* Max transaction type */ /* * The transactions for mkdir, symlink, remove, rmdir, link, and rename @@ -176,9 +177,9 @@ typedef enum zil_create { #define TX_CI ((uint64_t)0x1 << 63) /* case-insensitive behavior requested */ /* - * Transactions for write, truncate, setattr, acl_v0, and acl can be logged - * out of order. For convenience in the code, all such records must have - * lr_foid at the same offset. + * Transactions for operations below can be logged out of order. + * For convenience in the code, all such records must have lr_foid + * at the same offset. */ #define TX_OOO(txtype) \ ((txtype) == TX_WRITE || \ @@ -187,7 +188,8 @@ typedef enum zil_create { (txtype) == TX_ACL_V0 || \ (txtype) == TX_ACL || \ (txtype) == TX_WRITE2 || \ - (txtype) == TX_SETSAXATTR) + (txtype) == TX_SETSAXATTR || \ + (txtype) == TX_CLONE_RANGE) /* * The number of dnode slots consumed by the object is stored in the 8 @@ -387,6 +389,17 @@ typedef struct { /* lr_acl_bytes number of variable sized ace's follows */ } lr_acl_t; +typedef struct { + lr_t lr_common; /* common portion of log record */ + uint64_t lr_foid; /* file object to clone into */ + uint64_t lr_offset; /* offset to clone to */ + uint64_t lr_length; /* length of the blocks to clone */ + uint64_t lr_blksz; /* file's block size */ + uint64_t lr_nbps; /* number of block pointers */ + blkptr_t lr_bps[]; + /* block pointers of the blocks to clone follows */ +} lr_clone_range_t; + /* * ZIL structure definitions, interface function prototype and globals. */ @@ -574,7 +587,7 @@ extern void zil_set_sync(zilog_t *zilog, uint64_t syncval); extern void zil_set_logbias(zilog_t *zilog, uint64_t slogval); extern uint64_t zil_max_copied_data(zilog_t *zilog); -extern uint64_t zil_max_log_data(zilog_t *zilog); +extern uint64_t zil_max_log_data(zilog_t *zilog, size_t hdrsize); extern void zil_sums_init(zil_sums_t *zs); extern void zil_sums_fini(zil_sums_t *zs); diff --git a/include/sys/zio.h b/include/sys/zio.h index 28ed837d829e..78603d0ebeba 100644 --- a/include/sys/zio.h +++ b/include/sys/zio.h @@ -348,6 +348,7 @@ typedef struct zio_prop { boolean_t zp_dedup; boolean_t zp_dedup_verify; boolean_t zp_nopwrite; + boolean_t zp_brtwrite; boolean_t zp_encrypt; boolean_t zp_byteorder; uint8_t zp_salt[ZIO_DATA_SALT_LEN]; @@ -556,7 +557,7 @@ extern zio_t *zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, zio_priority_t priority, zio_flag_t flags, zbookmark_phys_t *zb); extern void zio_write_override(zio_t *zio, blkptr_t *bp, int copies, - boolean_t nopwrite); + boolean_t nopwrite, boolean_t brtwrite); extern void zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp); diff --git a/include/sys/zio_impl.h b/include/sys/zio_impl.h index 199cca291edf..29a05986cd4f 100644 --- a/include/sys/zio_impl.h +++ b/include/sys/zio_impl.h @@ -77,6 +77,12 @@ extern "C" { * and zstd. Compression occurs as part of the write pipeline and is * performed in the ZIO_STAGE_WRITE_BP_INIT stage. * + * Block cloning: + * The block cloning functionality introduces ZIO_STAGE_BRT_FREE stage which + * is called during a free pipeline. If the block is referenced in the + * Block Cloning Table (BRT) we will just decrease its reference counter + * instead of actually freeing the block. + * * Dedup: * Dedup reads are handled by the ZIO_STAGE_DDT_READ_START and * ZIO_STAGE_DDT_READ_DONE stages. These stages are added to an existing @@ -127,28 +133,30 @@ enum zio_stage { ZIO_STAGE_NOP_WRITE = 1 << 8, /* -W--- */ - ZIO_STAGE_DDT_READ_START = 1 << 9, /* R---- */ - ZIO_STAGE_DDT_READ_DONE = 1 << 10, /* R---- */ - ZIO_STAGE_DDT_WRITE = 1 << 11, /* -W--- */ - ZIO_STAGE_DDT_FREE = 1 << 12, /* --F-- */ + ZIO_STAGE_BRT_FREE = 1 << 9, /* --F-- */ - ZIO_STAGE_GANG_ASSEMBLE = 1 << 13, /* RWFC- */ - ZIO_STAGE_GANG_ISSUE = 1 << 14, /* RWFC- */ + ZIO_STAGE_DDT_READ_START = 1 << 10, /* R---- */ + ZIO_STAGE_DDT_READ_DONE = 1 << 11, /* R---- */ + ZIO_STAGE_DDT_WRITE = 1 << 12, /* -W--- */ + ZIO_STAGE_DDT_FREE = 1 << 13, /* --F-- */ - ZIO_STAGE_DVA_THROTTLE = 1 << 15, /* -W--- */ - ZIO_STAGE_DVA_ALLOCATE = 1 << 16, /* -W--- */ - ZIO_STAGE_DVA_FREE = 1 << 17, /* --F-- */ - ZIO_STAGE_DVA_CLAIM = 1 << 18, /* ---C- */ + ZIO_STAGE_GANG_ASSEMBLE = 1 << 14, /* RWFC- */ + ZIO_STAGE_GANG_ISSUE = 1 << 15, /* RWFC- */ - ZIO_STAGE_READY = 1 << 19, /* RWFCI */ + ZIO_STAGE_DVA_THROTTLE = 1 << 16, /* -W--- */ + ZIO_STAGE_DVA_ALLOCATE = 1 << 17, /* -W--- */ + ZIO_STAGE_DVA_FREE = 1 << 18, /* --F-- */ + ZIO_STAGE_DVA_CLAIM = 1 << 19, /* ---C- */ - ZIO_STAGE_VDEV_IO_START = 1 << 20, /* RW--I */ - ZIO_STAGE_VDEV_IO_DONE = 1 << 21, /* RW--I */ - ZIO_STAGE_VDEV_IO_ASSESS = 1 << 22, /* RW--I */ + ZIO_STAGE_READY = 1 << 20, /* RWFCI */ - ZIO_STAGE_CHECKSUM_VERIFY = 1 << 23, /* R---- */ + ZIO_STAGE_VDEV_IO_START = 1 << 21, /* RW--I */ + ZIO_STAGE_VDEV_IO_DONE = 1 << 22, /* RW--I */ + ZIO_STAGE_VDEV_IO_ASSESS = 1 << 23, /* RW--I */ - ZIO_STAGE_DONE = 1 << 24 /* RWFCI */ + ZIO_STAGE_CHECKSUM_VERIFY = 1 << 24, /* R---- */ + + ZIO_STAGE_DONE = 1 << 25 /* RWFCI */ }; #define ZIO_INTERLOCK_STAGES \ @@ -233,6 +241,7 @@ enum zio_stage { #define ZIO_FREE_PIPELINE \ (ZIO_INTERLOCK_STAGES | \ ZIO_STAGE_FREE_BP_INIT | \ + ZIO_STAGE_BRT_FREE | \ ZIO_STAGE_DVA_FREE) #define ZIO_DDT_FREE_PIPELINE \ diff --git a/include/zfeature_common.h b/include/zfeature_common.h index 0930bc900f82..ef915a70952e 100644 --- a/include/zfeature_common.h +++ b/include/zfeature_common.h @@ -78,6 +78,7 @@ typedef enum spa_feature { SPA_FEATURE_ZILSAXATTR, SPA_FEATURE_HEAD_ERRLOG, SPA_FEATURE_BLAKE3, + SPA_FEATURE_BLOCK_CLONING, SPA_FEATURES } spa_feature_t; diff --git a/lib/libzfs/libzfs.abi b/lib/libzfs/libzfs.abi index 16fea63f895c..79c0201678b0 100644 --- a/lib/libzfs/libzfs.abi +++ b/lib/libzfs/libzfs.abi @@ -590,7 +590,7 @@ - + @@ -3174,7 +3174,10 @@ - + + + + @@ -4850,8 +4853,8 @@ - - + + @@ -4893,7 +4896,8 @@ - + + diff --git a/lib/libzfs/libzfs_pool.c b/lib/libzfs/libzfs_pool.c index b3e12bd84a2d..82965f8b993a 100644 --- a/lib/libzfs/libzfs_pool.c +++ b/lib/libzfs/libzfs_pool.c @@ -339,6 +339,8 @@ zpool_get_prop(zpool_handle_t *zhp, zpool_prop_t prop, char *buf, case ZPOOL_PROP_ASHIFT: case ZPOOL_PROP_MAXBLOCKSIZE: case ZPOOL_PROP_MAXDNODESIZE: + case ZPOOL_PROP_BCLONESAVED: + case ZPOOL_PROP_BCLONEUSED: if (literal) (void) snprintf(buf, len, "%llu", (u_longlong_t)intval); @@ -380,6 +382,7 @@ zpool_get_prop(zpool_handle_t *zhp, zpool_prop_t prop, char *buf, } break; + case ZPOOL_PROP_BCLONERATIO: case ZPOOL_PROP_DEDUPRATIO: if (literal) (void) snprintf(buf, len, "%llu.%02llu", diff --git a/lib/libzpool/Makefile.am b/lib/libzpool/Makefile.am index 0748f1240db9..ceac2963e647 100644 --- a/lib/libzpool/Makefile.am +++ b/lib/libzpool/Makefile.am @@ -74,6 +74,7 @@ nodist_libzpool_la_SOURCES = \ module/zfs/bptree.c \ module/zfs/bqueue.c \ module/zfs/btree.c \ + module/zfs/brt.c \ module/zfs/dbuf.c \ module/zfs/dbuf_stats.c \ module/zfs/ddt.c \ diff --git a/man/man7/zpool-features.7 b/man/man7/zpool-features.7 index 3ff3d97ba70c..a4d595cd3cd9 100644 --- a/man/man7/zpool-features.7 +++ b/man/man7/zpool-features.7 @@ -347,6 +347,20 @@ BLAKE3 is a secure hash algorithm focused on high performance. .Pp .checksum-spiel blake3 . +.feature com.fudosecurity block_cloning yes +When this feature is enabled ZFS will use block cloning for operations like +.Fn copy_file_range 2 . +Block cloning allows to create multiple references to a single block. +It is much faster than copying the data (as the actual data is neither read nor +written) and takes no additional space. +Blocks can be cloned across datasets under some conditions (like disabled +encryption and equal +.Nm recordsize ) . +.Pp +This feature becomes +.Sy active +when first block is cloned. +When the last cloned block is freed, it goes back to the enabled state. .feature com.delphix bookmarks yes extensible_dataset This feature enables use of the .Nm zfs Cm bookmark diff --git a/man/man7/zpoolprops.7 b/man/man7/zpoolprops.7 index 7be0a21d980a..12b9b11903df 100644 --- a/man/man7/zpoolprops.7 +++ b/man/man7/zpoolprops.7 @@ -42,13 +42,26 @@ change the behavior of the pool. .Pp The following are read-only properties: .Bl -tag -width "unsupported@guid" -.It Cm allocated +.It Sy allocated Amount of storage used within the pool. See .Sy fragmentation and .Sy free for more information. +.It Sy bcloneratio +The ratio of the total amount of storage that would be required to store all +the cloned blocks without cloning to the actual storage used. +The +.Sy bcloneratio +property is calculated as: +.Pp +.Sy ( ( bclonesaved + bcloneused ) * 100 ) / bcloneused +.It Sy bclonesaved +The amount of additional storage that would be required if block cloning +was not used. +.It Sy bcloneused +The amount of storage used by cloned blocks. .It Sy capacity Percentage of pool space used. This property can also be referred to by its shortened column name, @@ -103,16 +116,16 @@ Over time will decrease while .Sy free increases. -.It Sy leaked -Space not released while -.Sy freeing -due to corruption, now permanently leaked into the pool. +.It Sy guid +A unique identifier for the pool. .It Sy health The current health of the pool. Health can be one of .Sy ONLINE , DEGRADED , FAULTED , OFFLINE, REMOVED , UNAVAIL . -.It Sy guid -A unique identifier for the pool. +.It Sy leaked +Space not released while +.Sy freeing +due to corruption, now permanently leaked into the pool. .It Sy load_guid A unique identifier for the pool. Unlike the diff --git a/module/Kbuild.in b/module/Kbuild.in index 21606b8cae27..8d29f56c2fb8 100644 --- a/module/Kbuild.in +++ b/module/Kbuild.in @@ -305,6 +305,7 @@ ZFS_OBJS := \ bpobj.o \ bptree.o \ bqueue.o \ + brt.o \ btree.o \ dataset_kstats.o \ dbuf.o \ diff --git a/module/Makefile.bsd b/module/Makefile.bsd index 667678796779..8ec094d4ad1c 100644 --- a/module/Makefile.bsd +++ b/module/Makefile.bsd @@ -33,11 +33,11 @@ KMOD= openzfs ${SRCDIR}/zstd/lib/decompress CFLAGS+= -I${INCDIR} +CFLAGS+= -I${SRCDIR}/icp/include CFLAGS+= -I${INCDIR}/os/freebsd CFLAGS+= -I${INCDIR}/os/freebsd/spl CFLAGS+= -I${INCDIR}/os/freebsd/zfs CFLAGS+= -I${SRCDIR}/zstd/include -CFLAGS+= -I${SRCDIR}/icp/include CFLAGS+= -include ${INCDIR}/os/freebsd/spl/sys/ccompile.h CFLAGS+= -I${.CURDIR} @@ -243,6 +243,7 @@ SRCS+= abd.c \ blkptr.c \ bplist.c \ bpobj.c \ + brt.c \ btree.c \ cityhash.c \ dbuf.c \ diff --git a/module/os/freebsd/zfs/sysctl_os.c b/module/os/freebsd/zfs/sysctl_os.c index 35edea0a2427..eccb91deff4f 100644 --- a/module/os/freebsd/zfs/sysctl_os.c +++ b/module/os/freebsd/zfs/sysctl_os.c @@ -97,6 +97,8 @@ __FBSDID("$FreeBSD$"); SYSCTL_DECL(_vfs_zfs); SYSCTL_NODE(_vfs_zfs, OID_AUTO, arc, CTLFLAG_RW, 0, "ZFS adaptive replacement cache"); +SYSCTL_NODE(_vfs_zfs, OID_AUTO, brt, CTLFLAG_RW, 0, + "ZFS Block Reference Table"); SYSCTL_NODE(_vfs_zfs, OID_AUTO, condense, CTLFLAG_RW, 0, "ZFS condense"); SYSCTL_NODE(_vfs_zfs, OID_AUTO, dbuf, CTLFLAG_RW, 0, "ZFS disk buf cache"); SYSCTL_NODE(_vfs_zfs, OID_AUTO, dbuf_cache, CTLFLAG_RW, 0, diff --git a/module/os/freebsd/zfs/zfs_vfsops.c b/module/os/freebsd/zfs/zfs_vfsops.c index 9fb2873132bf..30851f5273a2 100644 --- a/module/os/freebsd/zfs/zfs_vfsops.c +++ b/module/os/freebsd/zfs/zfs_vfsops.c @@ -153,7 +153,12 @@ struct vfsops zfs_vfsops = { .vfs_quotactl = zfs_quotactl, }; -VFS_SET(zfs_vfsops, zfs, VFCF_JAIL | VFCF_DELEGADMIN); +#ifdef VFCF_CROSS_COPY_FILE_RANGE +VFS_SET(zfs_vfsops, zfs, + VFCF_DELEGADMIN | VFCF_JAIL | VFCF_CROSS_COPY_FILE_RANGE); +#else +VFS_SET(zfs_vfsops, zfs, VFCF_DELEGADMIN | VFCF_JAIL); +#endif /* * We need to keep a count of active fs's. diff --git a/module/os/freebsd/zfs/zfs_vnops_os.c b/module/os/freebsd/zfs/zfs_vnops_os.c index 148def20ce57..9169244b1a13 100644 --- a/module/os/freebsd/zfs/zfs_vnops_os.c +++ b/module/os/freebsd/zfs/zfs_vnops_os.c @@ -30,7 +30,6 @@ /* Portions Copyright 2010 Robert Milkowski */ -#include #include #include #include @@ -84,6 +83,12 @@ #include #include #include +#include +#include +#include +#include +#include +#include #include @@ -6209,6 +6214,93 @@ zfs_deallocate(struct vop_deallocate_args *ap) } #endif +#ifndef _SYS_SYSPROTO_H_ +struct vop_copy_file_range_args { + struct vnode *a_invp; + off_t *a_inoffp; + struct vnode *a_outvp; + off_t *a_outoffp; + size_t *a_lenp; + unsigned int a_flags; + struct ucred *a_incred; + struct ucred *a_outcred; + struct thread *a_fsizetd; +} +#endif +/* + * TODO: FreeBSD will only call file system-specific copy_file_range() if both + * files resides under the same mountpoint. In case of ZFS we want to be called + * even is files are in different datasets (but on the same pools, but we need + * to check that ourselves). + */ +static int +zfs_freebsd_copy_file_range(struct vop_copy_file_range_args *ap) +{ + struct vnode *invp = ap->a_invp; + struct vnode *outvp = ap->a_outvp; + struct mount *mp; + struct uio io; + int error; + + /* + * TODO: If offset/length is not aligned to recordsize, use + * vn_generic_copy_file_range() on this fragment. + * It would be better to do this after we lock the vnodes, but then we + * need something else than vn_generic_copy_file_range(). + */ + + /* Lock both vnodes, avoiding risk of deadlock. */ + do { + mp = NULL; + error = vn_start_write(outvp, &mp, V_WAIT); + if (error == 0) { + error = vn_lock(outvp, LK_EXCLUSIVE); + if (error == 0) { + if (invp == outvp) + break; + error = vn_lock(invp, LK_SHARED | LK_NOWAIT); + if (error == 0) + break; + VOP_UNLOCK(outvp); + if (mp != NULL) + vn_finished_write(mp); + mp = NULL; + error = vn_lock(invp, LK_SHARED); + if (error == 0) + VOP_UNLOCK(invp); + } + } + if (mp != NULL) + vn_finished_write(mp); + } while (error == 0); + if (error != 0) + return (error); +#ifdef MAC + error = mac_vnode_check_write(curthread->td_ucred, ap->a_outcred, + outvp); + if (error != 0) + goto unlock; +#endif + + io.uio_offset = *ap->a_outoffp; + io.uio_resid = *ap->a_lenp; + error = vn_rlimit_fsize(outvp, &io, ap->a_fsizetd); + if (error != 0) + goto unlock; + + error = zfs_clone_range(VTOZ(invp), ap->a_inoffp, VTOZ(outvp), + ap->a_outoffp, ap->a_lenp, ap->a_fsizetd->td_ucred); + +unlock: + if (invp != outvp) + VOP_UNLOCK(invp); + VOP_UNLOCK(outvp); + if (mp != NULL) + vn_finished_write(mp); + + return (error); +} + struct vop_vector zfs_vnodeops; struct vop_vector zfs_fifoops; struct vop_vector zfs_shareops; @@ -6272,6 +6364,7 @@ struct vop_vector zfs_vnodeops = { #if __FreeBSD_version >= 1400043 .vop_add_writecount = vop_stdadd_writecount_nomsync, #endif + .vop_copy_file_range = zfs_freebsd_copy_file_range, }; VFS_VOP_VECTOR_REGISTER(zfs_vnodeops); diff --git a/module/os/freebsd/zfs/zfs_znode.c b/module/os/freebsd/zfs/zfs_znode.c index 304bc71f90db..dc1d31e1bd11 100644 --- a/module/os/freebsd/zfs/zfs_znode.c +++ b/module/os/freebsd/zfs/zfs_znode.c @@ -34,6 +34,7 @@ #include #include #include +#include #include #include #include @@ -2113,3 +2114,28 @@ zfs_znode_parent_and_name(znode_t *zp, znode_t **dzpp, char *buf) return (err); } #endif /* _KERNEL */ + +#ifdef _KERNEL +int +zfs_rlimit_fsize(off_t fsize) +{ + struct thread *td = curthread; + off_t lim; + + if (td == NULL) + return (0); + + lim = lim_cur(td, RLIMIT_FSIZE); + if (__predict_true((uoff_t)fsize <= lim)) + return (0); + + /* + * The limit is reached. + */ + PROC_LOCK(td->td_proc); + kern_psignal(td->td_proc, SIGXFSZ); + PROC_UNLOCK(td->td_proc); + + return (EFBIG); +} +#endif /* _KERNEL */ diff --git a/module/zcommon/zfeature_common.c b/module/zcommon/zfeature_common.c index ea45c9f8afa9..6fe1da8ed46f 100644 --- a/module/zcommon/zfeature_common.c +++ b/module/zcommon/zfeature_common.c @@ -725,6 +725,12 @@ zpool_feature_init(void) blake3_deps, sfeatures); } + zfeature_register(SPA_FEATURE_BLOCK_CLONING, + "com.fudosecurity:block_cloning", "block_cloning", + "Support for block cloning via Block Reference Table.", + ZFEATURE_FLAG_READONLY_COMPAT, ZFEATURE_TYPE_BOOLEAN, NULL, + sfeatures); + zfs_mod_list_supported_free(sfeatures); } diff --git a/module/zcommon/zpool_prop.c b/module/zcommon/zpool_prop.c index e99acef5a8fb..459ff62fc996 100644 --- a/module/zcommon/zpool_prop.c +++ b/module/zcommon/zpool_prop.c @@ -116,6 +116,15 @@ zpool_prop_init(void) zprop_register_number(ZPOOL_PROP_DEDUPRATIO, "dedupratio", 0, PROP_READONLY, ZFS_TYPE_POOL, "<1.00x or higher if deduped>", "DEDUP", B_FALSE, sfeatures); + zprop_register_number(ZPOOL_PROP_BCLONEUSED, "bcloneused", 0, + PROP_READONLY, ZFS_TYPE_POOL, "", + "BCLONE_USED", B_FALSE, sfeatures); + zprop_register_number(ZPOOL_PROP_BCLONESAVED, "bclonesaved", 0, + PROP_READONLY, ZFS_TYPE_POOL, "", + "BCLONE_SAVED", B_FALSE, sfeatures); + zprop_register_number(ZPOOL_PROP_BCLONERATIO, "bcloneratio", 0, + PROP_READONLY, ZFS_TYPE_POOL, "<1.00x or higher if cloned>", + "BCLONE_RATIO", B_FALSE, sfeatures); /* default number properties */ zprop_register_number(ZPOOL_PROP_VERSION, "version", SPA_VERSION, diff --git a/module/zfs/brt.c b/module/zfs/brt.c new file mode 100644 index 000000000000..ca9c4e678009 --- /dev/null +++ b/module/zfs/brt.c @@ -0,0 +1,1884 @@ +/* + * CDDL HEADER START + * + * The contents of this file are subject to the terms of the + * Common Development and Distribution License (the "License"). + * You may not use this file except in compliance with the License. + * + * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE + * or http://www.opensolaris.org/os/licensing. + * See the License for the specific language governing permissions + * and limitations under the License. + * + * When distributing Covered Code, include this CDDL HEADER in each + * file and include the License file at usr/src/OPENSOLARIS.LICENSE. + * If applicable, add the following below this CDDL HEADER, with the + * fields enclosed by brackets "[]" replaced with your own identifying + * information: Portions Copyright [yyyy] [name of copyright owner] + * + * CDDL HEADER END + */ + +/* + * Copyright (c) 2020, 2021, 2022 by Pawel Jakub Dawidek + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +/* + * Block Cloning design. + * + * Block Cloning allows to manually clone a file (or a subset of its blocks) + * into another (or the same) file by just creating additional references to + * the data blocks without copying the data itself. Those references are kept + * in the Block Reference Tables (BRTs). + * + * In many ways this is similar to the existing deduplication, but there are + * some important differences: + * + * - Deduplication is automatic and Block Cloning is not - one has to use a + * dedicated system call(s) to clone the given file/blocks. + * - Deduplication keeps all data blocks in its table, even those referenced + * just once. Block Cloning creates an entry in its tables only when there + * are at least two references to the given data block. If the block was + * never explicitly cloned or the second to last reference was dropped, + * there will be neither space nor performance overhead. + * - Deduplication needs data to work - one needs to pass real data to the + * write(2) syscall, so hash can be calculated. Block Cloning doesn't require + * data, just block pointers to the data, so it is extremely fast, as we pay + * neither the cost of reading the data, nor the cost of writing the data - + * we operate exclusively on metadata. + * - If the D (dedup) bit is not set in the block pointer, it means that + * the block is not in the dedup table (DDT) and we won't consult the DDT + * when we need to free the block. Block Cloning must be consulted on every + * free, because we cannot modify the source BP (eg. by setting something + * similar to the D bit), thus we have no hint if the block is in the + * Block Reference Table (BRT), so we need to look into the BRT. There is + * an optimization in place that allows us to eliminate the majority of BRT + * lookups which is described below in the "Minimizing free penalty" section. + * - The BRT entry is much smaller than the DDT entry - for BRT we only store + * 64bit offset and 64bit reference counter. + * - Dedup keys are cryptographic hashes, so two blocks that are close to each + * other on disk are most likely in totally different parts of the DDT. + * The BRT entry keys are offsets into a single top-level VDEV, so data blocks + * from one file should have BRT entries close to each other. + * - Scrub will only do a single pass over a block that is referenced multiple + * times in the DDT. Unfortunately it is not currently (if at all) possible + * with Block Cloning and block referenced multiple times will be scrubbed + * multiple times. The new, sorted scrub should be able to eliminate + * duplicated reads given enough memory. + * - Deduplication requires cryptographically strong hash as a checksum or + * additional data verification. Block Cloning works with any checksum + * algorithm or even with checksumming disabled. + * + * As mentioned above, the BRT entries are much smaller than the DDT entries. + * To uniquely identify a block we just need its vdev id and offset. We also + * need to maintain a reference counter. The vdev id will often repeat, as there + * is a small number of top-level VDEVs and a large number of blocks stored in + * each VDEV. We take advantage of that to reduce the BRT entry size further by + * maintaining one BRT for each top-level VDEV, so we can then have only offset + * and counter as the BRT entry. + * + * Minimizing free penalty. + * + * Block Cloning allows creating additional references to any existing block. + * When we free a block there is no hint in the block pointer whether the block + * was cloned or not, so on each free we have to check if there is a + * corresponding entry in the BRT or not. If there is, we need to decrease + * the reference counter. Doing BRT lookup on every free can potentially be + * expensive by requiring additional I/Os if the BRT doesn't fit into memory. + * This is the main problem with deduplication, so we've learned our lesson and + * try not to repeat the same mistake here. How do we do that? We divide each + * top-level VDEV into 16MB regions. For each region we maintain a counter that + * is a sum of all the BRT entries that have offsets within the region. This + * creates the entries count array of 16bit numbers for each top-level VDEV. + * The entries count array is always kept in memory and updated on disk in the + * same transaction group as the BRT updates to keep everything in-sync. We can + * keep the array in memory, because it is very small. With 16MB regions and + * 1TB VDEV the array requires only 128kB of memory (we may decide to decrease + * the region size even further in the future). Now, when we want to free + * a block, we first consult the array. If the counter for the whole region is + * zero, there is no need to look for the BRT entry, as there isn't one for + * sure. If the counter for the region is greater than zero, only then we will + * do a BRT lookup and if an entry is found we will decrease the reference + * counter in the BRT entry and in the entry counters array. + * + * The entry counters array is small, but can potentially be larger for very + * large VDEVs or smaller regions. In this case we don't want to rewrite entire + * array on every change. We then divide the array into 32kB block and keep + * a bitmap of dirty blocks within a transaction group. When we sync the + * transaction group we can only update the parts of the entry counters array + * that were modified. Note: Keeping track of the dirty parts of the entry + * counters array is implemented, but updating only parts of the array on disk + * is not yet implemented - for now we will update entire array if there was + * any change. + * + * The implementation tries to be economic: if BRT is not used, or no longer + * used, there will be no entries in the MOS and no additional memory used (eg. + * the entry counters array is only allocated if needed). + * + * Interaction between Deduplication and Block Cloning. + * + * If both functionalities are in use, we could end up with a block that is + * referenced multiple times in both DDT and BRT. When we free one of the + * references we couldn't tell where it belongs, so we would have to decide + * what table takes the precedence: do we first clear DDT references or BRT + * references? To avoid this dilemma BRT cooperates with DDT - if a given block + * is being cloned using BRT and the BP has the D (dedup) bit set, BRT will + * lookup DDT entry instead and increase the counter there. No BRT entry + * will be created for a block which has the D (dedup) bit set. + * BRT may be more efficient for manual deduplication, but if the block is + * already in the DDT, then creating additional BRT entry would be less + * efficient. This clever idea was proposed by Allan Jude. + * + * Block Cloning across datasets. + * + * Block Cloning is not limited to cloning blocks within the same dataset. + * It is possible (and very useful) to clone blocks between different datasets. + * One use case is recovering files from snapshots. By cloning the files into + * dataset we need no additional storage. Without Block Cloning we would need + * additional space for those files. + * Another interesting use case is moving the files between datasets + * (copying the file content to the new dataset and removing the source file). + * In that case Block Cloning will only be used briefly, because the BRT entries + * will be removed when the source is removed. + * Note: currently it is not possible to clone blocks between encrypted + * datasets, even if those datasets use the same encryption key (this includes + * snapshots of encrypted datasets). Cloning blocks between datasets that use + * the same keys should be possible and should be implemented in the future. + * + * Block Cloning flow through ZFS layers. + * + * Note: Block Cloning can be used both for cloning file system blocks and ZVOL + * blocks. As of this writing no interface is implemented that allows for block + * cloning within a ZVOL. + * FreeBSD and Linux provides copy_file_range(2) system call and we will use it + * for blocking cloning. + * + * ssize_t + * copy_file_range(int infd, off_t *inoffp, int outfd, off_t *outoffp, + * size_t len, unsigned int flags); + * + * Even though offsets and length represent bytes, they have to be + * block-aligned or we will return the EXDEV error so the upper layer can + * fallback to the generic mechanism that will just copy the data. + * Using copy_file_range(2) will call OS-independent zfs_clone_range() function. + * This function was implemented based on zfs_write(), but instead of writing + * the given data we first read block pointers using the new dmu_read_l0_bps() + * function from the source file. Once we have BPs from the source file we call + * the dmu_brt_clone() function on the destination file. This function + * allocates BPs for us. We iterate over all source BPs. If the given BP is + * a hole or an embedded block, we just copy BP as-is. If it points to a real + * data we place this BP on a BRT pending list using the brt_pending_add() + * function. + * + * We use this pending list to keep track of all BPs that got new references + * within this transaction group. + * + * Some special cases to consider and how we address them: + * - The block we want to clone may have been created within the same + * transaction group that we are trying to clone. Such block has no BP + * allocated yet, so cannot be immediately cloned. We return EXDEV. + * - The block we want to clone may have been modified within the same + * transaction group. We return EXDEV. + * - A block may be cloned multiple times during one transaction group (that's + * why pending list is actually a tree and not an append-only list - this + * way we can figure out faster if this block is cloned for the first time + * in this txg or consecutive time). + * - A block may be cloned and freed within the same transaction group + * (see dbuf_undirty()). + * - A block may be cloned and within the same transaction group the clone + * can be cloned again (see dmu_read_l0_bps()). + * - A file might have been deleted, but the caller still has a file descriptor + * open to this file and clones it. + * + * When we free a block we have an additional step in the ZIO pipeline where we + * call the zio_brt_free() function. We then call the brt_entry_decref() + * that loads the corresponding BRT entry (if one exists) and decreases + * reference counter. If this is not the last reference we will stop ZIO + * pipeline here. If this is the last reference or the block is not in the + * BRT, we continue the pipeline and free the block as usual. + * + * At the beginning of spa_sync() where there can be no more block cloning, + * but before issuing frees we call brt_pending_apply(). This function applies + * all the new clones to the BRT table - we load BRT entries and update + * reference counters. To sync new BRT entries to disk, we use brt_sync() + * function. This function will sync all dirty per-top-level-vdev BRTs, + * the entry counters arrays, etc. + * + * Block Cloning and ZIL. + * + * Every clone operation is divided into chunks (similar to write) and each + * chunk is cloned in a separate transaction. The chunk size is determined by + * how many BPs we can fit into a single ZIL entry. + * Replaying clone operation is different from the regular clone operation, + * as when we log clone operations we cannot use the source object - it may + * reside on a different dataset, so we log BPs we want to clone. + * The ZIL is replayed when we mount the given dataset, not when the pool is + * imported. Taking this into account it is possible that the pool is imported + * without mounting datasets and the source dataset is destroyed before the + * destination dataset is mounted and its ZIL replayed. + * To address this situation we leverage zil_claim() mechanism where ZFS will + * parse all the ZILs on pool import. When we come across TX_CLONE_RANGE + * entries, we will bump reference counters for their BPs in the BRT and then + * on mount and ZIL replay we will just attach BPs to the file without + * bumping reference counters. + * Note it is still possible that after zil_claim() we never mount the + * destination, so we never replay its ZIL and we destroy it. This way we would + * end up with leaked references in BRT. We address that too as ZFS gives us + * a chance to clean this up on dataset destroy (see zil_free_clone_range()). + */ + +/* + * BRT - Block Reference Table. + */ +#define BRT_OBJECT_VDEV_PREFIX "com.fudosecurity:brt:vdev:" + +/* + * We divide each VDEV into 16MB chunks. Each chunk is represented in memory + * by a 16bit counter, thus 1TB VDEV requires 128kB of memory: (1TB / 16MB) * 2B + * Each element in this array represents how many BRT entries do we have in this + * chunk of storage. We always load this entire array into memory and update as + * needed. By having it in memory we can quickly tell (during zio_free()) if + * there are any BRT entries that we might need to update. + * + * This value cannot be larger than 16MB, at least as long as we support + * 512 byte block sizes. With 512 byte block size we can have exactly + * 32768 blocks in 16MB. In 32MB we could have 65536 blocks, which is one too + * many for a 16bit counter. + */ +#define BRT_RANGESIZE (16 * 1024 * 1024) +_Static_assert(BRT_RANGESIZE / SPA_MINBLOCKSIZE <= UINT16_MAX, + "BRT_RANGESIZE is too large."); +/* + * We don't want to update the whole structure every time. Maintain bitmap + * of dirty blocks within the regions, so that a single bit represents a + * block size of entcounts. For example if we have a 1PB vdev then all + * entcounts take 128MB of memory ((64TB / 16MB) * 2B). We can divide this + * 128MB array of entcounts into 32kB disk blocks, as we don't want to update + * the whole 128MB on disk when we have updated only a single entcount. + * We maintain a bitmap where each 32kB disk block within 128MB entcounts array + * is represented by a single bit. This gives us 4096 bits. A set bit in the + * bitmap means that we had a change in at least one of the 16384 entcounts + * that reside on a 32kB disk block (32kB / sizeof (uint16_t)). + */ +#define BRT_BLOCKSIZE (32 * 1024) +#define BRT_RANGESIZE_TO_NBLOCKS(size) \ + (((size) - 1) / BRT_BLOCKSIZE / sizeof (uint16_t) + 1) + +#define BRT_LITTLE_ENDIAN 0 +#define BRT_BIG_ENDIAN 1 +#ifdef _ZFS_LITTLE_ENDIAN +#define BRT_NATIVE_BYTEORDER BRT_LITTLE_ENDIAN +#define BRT_NON_NATIVE_BYTEORDER BRT_BIG_ENDIAN +#else +#define BRT_NATIVE_BYTEORDER BRT_BIG_ENDIAN +#define BRT_NON_NATIVE_BYTEORDER BRT_LITTLE_ENDIAN +#endif + +typedef struct brt_vdev_phys { + uint64_t bvp_mos_entries; + uint64_t bvp_size; + uint64_t bvp_byteorder; + uint64_t bvp_totalcount; + uint64_t bvp_rangesize; + uint64_t bvp_usedspace; + uint64_t bvp_savedspace; +} brt_vdev_phys_t; + +typedef struct brt_vdev { + /* + * VDEV id. + */ + uint64_t bv_vdevid; + /* + * Is the structure initiated? + * (bv_entcount and bv_bitmap are allocated?) + */ + boolean_t bv_initiated; + /* + * Object number in the MOS for the entcount array and brt_vdev_phys. + */ + uint64_t bv_mos_brtvdev; + /* + * Object number in the MOS for the entries table. + */ + uint64_t bv_mos_entries; + /* + * Entries to sync. + */ + avl_tree_t bv_tree; + /* + * Does the bv_entcount[] array needs byte swapping? + */ + boolean_t bv_need_byteswap; + /* + * Number of entries in the bv_entcount[] array. + */ + uint64_t bv_size; + /* + * This is the array with BRT entry count per BRT_RANGESIZE. + */ + uint16_t *bv_entcount; + /* + * Sum of all bv_entcount[]s. + */ + uint64_t bv_totalcount; + /* + * Space on disk occupied by cloned blocks (without compression). + */ + uint64_t bv_usedspace; + /* + * How much additional space would be occupied without block cloning. + */ + uint64_t bv_savedspace; + /* + * brt_vdev_phys needs updating on disk. + */ + boolean_t bv_meta_dirty; + /* + * bv_entcount[] needs updating on disk. + */ + boolean_t bv_entcount_dirty; + /* + * bv_entcount[] potentially can be a bit too big to sychronize it all + * when we just changed few entcounts. The fields below allow us to + * track updates to bv_entcount[] array since the last sync. + * A single bit in the bv_bitmap represents as many entcounts as can + * fit into a single BRT_BLOCKSIZE. + * For example we have 65536 entcounts in the bv_entcount array + * (so the whole array is 128kB). We updated bv_entcount[2] and + * bv_entcount[5]. In that case only first bit in the bv_bitmap will + * be set and we will write only first BRT_BLOCKSIZE out of 128kB. + */ + ulong_t *bv_bitmap; + uint64_t bv_nblocks; +} brt_vdev_t; + +/* + * In-core brt + */ +typedef struct brt { + krwlock_t brt_lock; + spa_t *brt_spa; +#define brt_mos brt_spa->spa_meta_objset + uint64_t brt_rangesize; + uint64_t brt_usedspace; + uint64_t brt_savedspace; + avl_tree_t brt_pending_tree[TXG_SIZE]; + kmutex_t brt_pending_lock[TXG_SIZE]; + /* Sum of all entries across all bv_trees. */ + uint64_t brt_nentries; + brt_vdev_t *brt_vdevs; + uint64_t brt_nvdevs; +} brt_t; + +/* Size of bre_offset / sizeof (uint64_t). */ +#define BRT_KEY_WORDS (1) + +/* + * In-core brt entry. + * On-disk we use bre_offset as the key and bre_refcount as the value. + */ +typedef struct brt_entry { + uint64_t bre_offset; + uint64_t bre_refcount; + avl_node_t bre_node; +} brt_entry_t; + +typedef struct brt_pending_entry { + blkptr_t bpe_bp; + int bpe_count; + avl_node_t bpe_node; +} brt_pending_entry_t; + +static kmem_cache_t *brt_entry_cache; +static kmem_cache_t *brt_pending_entry_cache; + +/* + * Enable/disable prefetching of BRT entries that we are going to modify. + */ +int zfs_brt_prefetch = 1; + +#ifdef ZFS_DEBUG +#define BRT_DEBUG(...) do { \ + if ((zfs_flags & ZFS_DEBUG_BRT) != 0) { \ + __dprintf(B_TRUE, __FILE__, __func__, __LINE__, __VA_ARGS__); \ + } \ +} while (0) +#else +#define BRT_DEBUG(...) do { } while (0) +#endif + +int brt_zap_leaf_blockshift = 12; +int brt_zap_indirect_blockshift = 12; + +static kstat_t *brt_ksp; + +typedef struct brt_stats { + kstat_named_t brt_addref_entry_in_memory; + kstat_named_t brt_addref_entry_not_on_disk; + kstat_named_t brt_addref_entry_on_disk; + kstat_named_t brt_addref_entry_read_lost_race; + kstat_named_t brt_decref_entry_in_memory; + kstat_named_t brt_decref_entry_loaded_from_disk; + kstat_named_t brt_decref_entry_not_in_memory; + kstat_named_t brt_decref_entry_not_on_disk; + kstat_named_t brt_decref_entry_read_lost_race; + kstat_named_t brt_decref_entry_still_referenced; + kstat_named_t brt_decref_free_data_later; + kstat_named_t brt_decref_free_data_now; + kstat_named_t brt_decref_no_entry; +} brt_stats_t; + +static brt_stats_t brt_stats = { + { "addref_entry_in_memory", KSTAT_DATA_UINT64 }, + { "addref_entry_not_on_disk", KSTAT_DATA_UINT64 }, + { "addref_entry_on_disk", KSTAT_DATA_UINT64 }, + { "addref_entry_read_lost_race", KSTAT_DATA_UINT64 }, + { "decref_entry_in_memory", KSTAT_DATA_UINT64 }, + { "decref_entry_loaded_from_disk", KSTAT_DATA_UINT64 }, + { "decref_entry_not_in_memory", KSTAT_DATA_UINT64 }, + { "decref_entry_not_on_disk", KSTAT_DATA_UINT64 }, + { "decref_entry_read_lost_race", KSTAT_DATA_UINT64 }, + { "decref_entry_still_referenced", KSTAT_DATA_UINT64 }, + { "decref_free_data_later", KSTAT_DATA_UINT64 }, + { "decref_free_data_now", KSTAT_DATA_UINT64 }, + { "decref_no_entry", KSTAT_DATA_UINT64 } +}; + +struct { + wmsum_t brt_addref_entry_in_memory; + wmsum_t brt_addref_entry_not_on_disk; + wmsum_t brt_addref_entry_on_disk; + wmsum_t brt_addref_entry_read_lost_race; + wmsum_t brt_decref_entry_in_memory; + wmsum_t brt_decref_entry_loaded_from_disk; + wmsum_t brt_decref_entry_not_in_memory; + wmsum_t brt_decref_entry_not_on_disk; + wmsum_t brt_decref_entry_read_lost_race; + wmsum_t brt_decref_entry_still_referenced; + wmsum_t brt_decref_free_data_later; + wmsum_t brt_decref_free_data_now; + wmsum_t brt_decref_no_entry; +} brt_sums; + +#define BRTSTAT_BUMP(stat) wmsum_add(&brt_sums.stat, 1) + +static int brt_entry_compare(const void *x1, const void *x2); +static int brt_pending_entry_compare(const void *x1, const void *x2); + +static void +brt_rlock(brt_t *brt) +{ + rw_enter(&brt->brt_lock, RW_READER); +} + +static void +brt_wlock(brt_t *brt) +{ + rw_enter(&brt->brt_lock, RW_WRITER); +} + +static void +brt_unlock(brt_t *brt) +{ + rw_exit(&brt->brt_lock); +} + +static uint16_t +brt_vdev_entcount_get(const brt_vdev_t *brtvd, uint64_t idx) +{ + + ASSERT3U(idx, <, brtvd->bv_size); + + if (brtvd->bv_need_byteswap) { + return (BSWAP_16(brtvd->bv_entcount[idx])); + } else { + return (brtvd->bv_entcount[idx]); + } +} + +static void +brt_vdev_entcount_set(brt_vdev_t *brtvd, uint64_t idx, uint16_t entcnt) +{ + + ASSERT3U(idx, <, brtvd->bv_size); + + if (brtvd->bv_need_byteswap) { + brtvd->bv_entcount[idx] = BSWAP_16(entcnt); + } else { + brtvd->bv_entcount[idx] = entcnt; + } +} + +static void +brt_vdev_entcount_inc(brt_vdev_t *brtvd, uint64_t idx) +{ + uint16_t entcnt; + + ASSERT3U(idx, <, brtvd->bv_size); + + entcnt = brt_vdev_entcount_get(brtvd, idx); + ASSERT(entcnt < UINT16_MAX); + + brt_vdev_entcount_set(brtvd, idx, entcnt + 1); +} + +static void +brt_vdev_entcount_dec(brt_vdev_t *brtvd, uint64_t idx) +{ + uint16_t entcnt; + + ASSERT3U(idx, <, brtvd->bv_size); + + entcnt = brt_vdev_entcount_get(brtvd, idx); + ASSERT(entcnt > 0); + + brt_vdev_entcount_set(brtvd, idx, entcnt - 1); +} + +#ifdef ZFS_DEBUG +static void +brt_vdev_dump(brt_t *brt) +{ + brt_vdev_t *brtvd; + uint64_t vdevid; + + if ((zfs_flags & ZFS_DEBUG_BRT) == 0) { + return; + } + + if (brt->brt_nvdevs == 0) { + zfs_dbgmsg("BRT empty"); + return; + } + + zfs_dbgmsg("BRT vdev dump:"); + for (vdevid = 0; vdevid < brt->brt_nvdevs; vdevid++) { + uint64_t idx; + + brtvd = &brt->brt_vdevs[vdevid]; + zfs_dbgmsg(" vdevid=%llu/%llu meta_dirty=%d entcount_dirty=%d " + "size=%llu totalcount=%llu nblocks=%llu bitmapsize=%zu\n", + (u_longlong_t)vdevid, (u_longlong_t)brtvd->bv_vdevid, + brtvd->bv_meta_dirty, brtvd->bv_entcount_dirty, + (u_longlong_t)brtvd->bv_size, + (u_longlong_t)brtvd->bv_totalcount, + (u_longlong_t)brtvd->bv_nblocks, + (size_t)BT_SIZEOFMAP(brtvd->bv_nblocks)); + if (brtvd->bv_totalcount > 0) { + zfs_dbgmsg(" entcounts:"); + for (idx = 0; idx < brtvd->bv_size; idx++) { + if (brt_vdev_entcount_get(brtvd, idx) > 0) { + zfs_dbgmsg(" [%04llu] %hu", + (u_longlong_t)idx, + brt_vdev_entcount_get(brtvd, idx)); + } + } + } + if (brtvd->bv_entcount_dirty) { + char *bitmap; + + bitmap = kmem_alloc(brtvd->bv_nblocks + 1, KM_SLEEP); + for (idx = 0; idx < brtvd->bv_nblocks; idx++) { + bitmap[idx] = + BT_TEST(brtvd->bv_bitmap, idx) ? 'x' : '.'; + } + bitmap[idx] = '\0'; + zfs_dbgmsg(" bitmap: %s", bitmap); + kmem_free(bitmap, brtvd->bv_nblocks + 1); + } + } +} +#endif + +static brt_vdev_t * +brt_vdev(brt_t *brt, uint64_t vdevid) +{ + brt_vdev_t *brtvd; + + ASSERT(RW_LOCK_HELD(&brt->brt_lock)); + + if (vdevid < brt->brt_nvdevs) { + brtvd = &brt->brt_vdevs[vdevid]; + } else { + brtvd = NULL; + } + + return (brtvd); +} + +static void +brt_vdev_create(brt_t *brt, brt_vdev_t *brtvd, dmu_tx_t *tx) +{ + char name[64]; + + ASSERT(RW_WRITE_HELD(&brt->brt_lock)); + ASSERT0(brtvd->bv_mos_brtvdev); + ASSERT0(brtvd->bv_mos_entries); + ASSERT(brtvd->bv_entcount != NULL); + ASSERT(brtvd->bv_size > 0); + ASSERT(brtvd->bv_bitmap != NULL); + ASSERT(brtvd->bv_nblocks > 0); + + brtvd->bv_mos_entries = zap_create_flags(brt->brt_mos, 0, + ZAP_FLAG_HASH64 | ZAP_FLAG_UINT64_KEY, DMU_OTN_ZAP_METADATA, + brt_zap_leaf_blockshift, brt_zap_indirect_blockshift, DMU_OT_NONE, + 0, tx); + VERIFY(brtvd->bv_mos_entries != 0); + BRT_DEBUG("MOS entries created, object=%llu", + (u_longlong_t)brtvd->bv_mos_entries); + + /* + * We allocate DMU buffer to store the bv_entcount[] array. + * We will keep array size (bv_size) and cummulative count for all + * bv_entcount[]s (bv_totalcount) in the bonus buffer. + */ + brtvd->bv_mos_brtvdev = dmu_object_alloc(brt->brt_mos, + DMU_OTN_UINT64_METADATA, BRT_BLOCKSIZE, + DMU_OTN_UINT64_METADATA, sizeof (brt_vdev_phys_t), tx); + VERIFY(brtvd->bv_mos_brtvdev != 0); + BRT_DEBUG("MOS BRT VDEV created, object=%llu", + (u_longlong_t)brtvd->bv_mos_brtvdev); + + snprintf(name, sizeof (name), "%s%llu", BRT_OBJECT_VDEV_PREFIX, + (u_longlong_t)brtvd->bv_vdevid); + VERIFY0(zap_add(brt->brt_mos, DMU_POOL_DIRECTORY_OBJECT, name, + sizeof (uint64_t), 1, &brtvd->bv_mos_brtvdev, tx)); + BRT_DEBUG("Pool directory object created, object=%s", name); + + spa_feature_incr(brt->brt_spa, SPA_FEATURE_BLOCK_CLONING, tx); +} + +static void +brt_vdev_realloc(brt_t *brt, brt_vdev_t *brtvd) +{ + vdev_t *vd; + uint16_t *entcount; + ulong_t *bitmap; + uint64_t nblocks, size; + + ASSERT(RW_WRITE_HELD(&brt->brt_lock)); + + spa_config_enter(brt->brt_spa, SCL_VDEV, FTAG, RW_READER); + vd = vdev_lookup_top(brt->brt_spa, brtvd->bv_vdevid); + size = (vdev_get_min_asize(vd) - 1) / brt->brt_rangesize + 1; + spa_config_exit(brt->brt_spa, SCL_VDEV, FTAG); + + entcount = kmem_zalloc(sizeof (entcount[0]) * size, KM_SLEEP); + nblocks = BRT_RANGESIZE_TO_NBLOCKS(size); + bitmap = kmem_zalloc(BT_SIZEOFMAP(nblocks), KM_SLEEP); + + if (!brtvd->bv_initiated) { + ASSERT0(brtvd->bv_size); + ASSERT(brtvd->bv_entcount == NULL); + ASSERT(brtvd->bv_bitmap == NULL); + ASSERT0(brtvd->bv_nblocks); + + avl_create(&brtvd->bv_tree, brt_entry_compare, + sizeof (brt_entry_t), offsetof(brt_entry_t, bre_node)); + } else { + ASSERT(brtvd->bv_size > 0); + ASSERT(brtvd->bv_entcount != NULL); + ASSERT(brtvd->bv_bitmap != NULL); + ASSERT(brtvd->bv_nblocks > 0); + /* + * TODO: Allow vdev shrinking. We only need to implement + * shrinking the on-disk BRT VDEV object. + * dmu_free_range(brt->brt_mos, brtvd->bv_mos_brtvdev, offset, + * size, tx); + */ + ASSERT3U(brtvd->bv_size, <=, size); + + memcpy(entcount, brtvd->bv_entcount, + sizeof (entcount[0]) * MIN(size, brtvd->bv_size)); + memcpy(bitmap, brtvd->bv_bitmap, MIN(BT_SIZEOFMAP(nblocks), + BT_SIZEOFMAP(brtvd->bv_nblocks))); + kmem_free(brtvd->bv_entcount, + sizeof (entcount[0]) * brtvd->bv_size); + kmem_free(brtvd->bv_bitmap, BT_SIZEOFMAP(brtvd->bv_nblocks)); + } + + brtvd->bv_size = size; + brtvd->bv_entcount = entcount; + brtvd->bv_bitmap = bitmap; + brtvd->bv_nblocks = nblocks; + if (!brtvd->bv_initiated) { + brtvd->bv_need_byteswap = FALSE; + brtvd->bv_initiated = TRUE; + BRT_DEBUG("BRT VDEV %llu initiated.", + (u_longlong_t)brtvd->bv_vdevid); + } +} + +static void +brt_vdev_load(brt_t *brt, brt_vdev_t *brtvd) +{ + char name[64]; + dmu_buf_t *db; + brt_vdev_phys_t *bvphys; + int error; + + snprintf(name, sizeof (name), "%s%llu", BRT_OBJECT_VDEV_PREFIX, + (u_longlong_t)brtvd->bv_vdevid); + error = zap_lookup(brt->brt_mos, DMU_POOL_DIRECTORY_OBJECT, name, + sizeof (uint64_t), 1, &brtvd->bv_mos_brtvdev); + if (error != 0) + return; + ASSERT(brtvd->bv_mos_brtvdev != 0); + + error = dmu_bonus_hold(brt->brt_mos, brtvd->bv_mos_brtvdev, FTAG, &db); + ASSERT0(error); + if (error != 0) + return; + + bvphys = db->db_data; + if (brt->brt_rangesize == 0) { + brt->brt_rangesize = bvphys->bvp_rangesize; + } else { + ASSERT3U(brt->brt_rangesize, ==, bvphys->bvp_rangesize); + } + + ASSERT(!brtvd->bv_initiated); + brt_vdev_realloc(brt, brtvd); + + /* TODO: We don't support VDEV shrinking. */ + ASSERT3U(bvphys->bvp_size, <=, brtvd->bv_size); + + /* + * If VDEV grew, we will leave new bv_entcount[] entries zeroed out. + */ + error = dmu_read(brt->brt_mos, brtvd->bv_mos_brtvdev, 0, + MIN(brtvd->bv_size, bvphys->bvp_size) * sizeof (uint16_t), + brtvd->bv_entcount, DMU_READ_NO_PREFETCH); + ASSERT0(error); + + brtvd->bv_mos_entries = bvphys->bvp_mos_entries; + ASSERT(brtvd->bv_mos_entries != 0); + brtvd->bv_need_byteswap = + (bvphys->bvp_byteorder != BRT_NATIVE_BYTEORDER); + brtvd->bv_totalcount = bvphys->bvp_totalcount; + brtvd->bv_usedspace = bvphys->bvp_usedspace; + brtvd->bv_savedspace = bvphys->bvp_savedspace; + brt->brt_usedspace += brtvd->bv_usedspace; + brt->brt_savedspace += brtvd->bv_savedspace; + + dmu_buf_rele(db, FTAG); + + BRT_DEBUG("MOS BRT VDEV %s loaded: mos_brtvdev=%llu, mos_entries=%llu", + name, (u_longlong_t)brtvd->bv_mos_brtvdev, + (u_longlong_t)brtvd->bv_mos_entries); +} + +static void +brt_vdev_dealloc(brt_t *brt, brt_vdev_t *brtvd) +{ + + ASSERT(RW_WRITE_HELD(&brt->brt_lock)); + ASSERT(brtvd->bv_initiated); + + kmem_free(brtvd->bv_entcount, sizeof (uint16_t) * brtvd->bv_size); + brtvd->bv_entcount = NULL; + kmem_free(brtvd->bv_bitmap, BT_SIZEOFMAP(brtvd->bv_nblocks)); + brtvd->bv_bitmap = NULL; + ASSERT0(avl_numnodes(&brtvd->bv_tree)); + avl_destroy(&brtvd->bv_tree); + + brtvd->bv_size = 0; + brtvd->bv_nblocks = 0; + + brtvd->bv_initiated = FALSE; + BRT_DEBUG("BRT VDEV %llu deallocated.", (u_longlong_t)brtvd->bv_vdevid); +} + +static void +brt_vdev_destroy(brt_t *brt, brt_vdev_t *brtvd, dmu_tx_t *tx) +{ + char name[64]; + uint64_t count; + dmu_buf_t *db; + brt_vdev_phys_t *bvphys; + + ASSERT(RW_WRITE_HELD(&brt->brt_lock)); + ASSERT(brtvd->bv_mos_brtvdev != 0); + ASSERT(brtvd->bv_mos_entries != 0); + + VERIFY0(zap_count(brt->brt_mos, brtvd->bv_mos_entries, &count)); + VERIFY0(count); + VERIFY0(zap_destroy(brt->brt_mos, brtvd->bv_mos_entries, tx)); + BRT_DEBUG("MOS entries destroyed, object=%llu", + (u_longlong_t)brtvd->bv_mos_entries); + brtvd->bv_mos_entries = 0; + + VERIFY0(dmu_bonus_hold(brt->brt_mos, brtvd->bv_mos_brtvdev, FTAG, &db)); + bvphys = db->db_data; + ASSERT0(bvphys->bvp_totalcount); + ASSERT0(bvphys->bvp_usedspace); + ASSERT0(bvphys->bvp_savedspace); + dmu_buf_rele(db, FTAG); + + VERIFY0(dmu_object_free(brt->brt_mos, brtvd->bv_mos_brtvdev, tx)); + BRT_DEBUG("MOS BRT VDEV destroyed, object=%llu", + (u_longlong_t)brtvd->bv_mos_brtvdev); + brtvd->bv_mos_brtvdev = 0; + + snprintf(name, sizeof (name), "%s%llu", BRT_OBJECT_VDEV_PREFIX, + (u_longlong_t)brtvd->bv_vdevid); + VERIFY0(zap_remove(brt->brt_mos, DMU_POOL_DIRECTORY_OBJECT, name, tx)); + BRT_DEBUG("Pool directory object removed, object=%s", name); + + brt_vdev_dealloc(brt, brtvd); + + spa_feature_decr(brt->brt_spa, SPA_FEATURE_BLOCK_CLONING, tx); +} + +static void +brt_vdevs_expand(brt_t *brt, uint64_t nvdevs) +{ + brt_vdev_t *brtvd, *vdevs; + uint64_t vdevid; + + ASSERT(RW_WRITE_HELD(&brt->brt_lock)); + ASSERT3U(nvdevs, >, brt->brt_nvdevs); + + vdevs = kmem_zalloc(sizeof (vdevs[0]) * nvdevs, KM_SLEEP); + if (brt->brt_nvdevs > 0) { + ASSERT(brt->brt_vdevs != NULL); + + memcpy(vdevs, brt->brt_vdevs, + sizeof (brt_vdev_t) * brt->brt_nvdevs); + kmem_free(brt->brt_vdevs, + sizeof (brt_vdev_t) * brt->brt_nvdevs); + } + for (vdevid = brt->brt_nvdevs; vdevid < nvdevs; vdevid++) { + brtvd = &vdevs[vdevid]; + + brtvd->bv_vdevid = vdevid; + brtvd->bv_initiated = FALSE; + } + + BRT_DEBUG("BRT VDEVs expanded from %llu to %llu.", + (u_longlong_t)brt->brt_nvdevs, (u_longlong_t)nvdevs); + + brt->brt_vdevs = vdevs; + brt->brt_nvdevs = nvdevs; +} + +static boolean_t +brt_vdev_lookup(brt_t *brt, brt_vdev_t *brtvd, const brt_entry_t *bre) +{ + uint64_t idx; + + ASSERT(RW_LOCK_HELD(&brt->brt_lock)); + + idx = bre->bre_offset / brt->brt_rangesize; + if (brtvd->bv_entcount != NULL && idx < brtvd->bv_size) { + /* VDEV wasn't expanded. */ + return (brt_vdev_entcount_get(brtvd, idx) > 0); + } + + return (FALSE); +} + +static void +brt_vdev_addref(brt_t *brt, brt_vdev_t *brtvd, const brt_entry_t *bre, + uint64_t dsize) +{ + uint64_t idx; + + ASSERT(RW_LOCK_HELD(&brt->brt_lock)); + ASSERT(brtvd != NULL); + ASSERT(brtvd->bv_entcount != NULL); + + brt->brt_savedspace += dsize; + brtvd->bv_savedspace += dsize; + brtvd->bv_meta_dirty = TRUE; + + if (bre->bre_refcount > 1) { + return; + } + + brt->brt_usedspace += dsize; + brtvd->bv_usedspace += dsize; + + idx = bre->bre_offset / brt->brt_rangesize; + if (idx >= brtvd->bv_size) { + /* VDEV has been expanded. */ + brt_vdev_realloc(brt, brtvd); + } + + ASSERT3U(idx, <, brtvd->bv_size); + + brtvd->bv_totalcount++; + brt_vdev_entcount_inc(brtvd, idx); + brtvd->bv_entcount_dirty = TRUE; + idx = idx / BRT_BLOCKSIZE / 8; + BT_SET(brtvd->bv_bitmap, idx); + +#ifdef ZFS_DEBUG + brt_vdev_dump(brt); +#endif +} + +static void +brt_vdev_decref(brt_t *brt, brt_vdev_t *brtvd, const brt_entry_t *bre, + uint64_t dsize) +{ + uint64_t idx; + + ASSERT(RW_WRITE_HELD(&brt->brt_lock)); + ASSERT(brtvd != NULL); + ASSERT(brtvd->bv_entcount != NULL); + + brt->brt_savedspace -= dsize; + brtvd->bv_savedspace -= dsize; + brtvd->bv_meta_dirty = TRUE; + + if (bre->bre_refcount > 0) { + return; + } + + brt->brt_usedspace -= dsize; + brtvd->bv_usedspace -= dsize; + + idx = bre->bre_offset / brt->brt_rangesize; + ASSERT3U(idx, <, brtvd->bv_size); + + ASSERT(brtvd->bv_totalcount > 0); + brtvd->bv_totalcount--; + brt_vdev_entcount_dec(brtvd, idx); + brtvd->bv_entcount_dirty = TRUE; + idx = idx / BRT_BLOCKSIZE / 8; + BT_SET(brtvd->bv_bitmap, idx); + +#ifdef ZFS_DEBUG + brt_vdev_dump(brt); +#endif +} + +static void +brt_vdev_sync(brt_t *brt, brt_vdev_t *brtvd, dmu_tx_t *tx) +{ + dmu_buf_t *db; + brt_vdev_phys_t *bvphys; + + ASSERT(brtvd->bv_meta_dirty); + ASSERT(brtvd->bv_mos_brtvdev != 0); + ASSERT(dmu_tx_is_syncing(tx)); + + VERIFY0(dmu_bonus_hold(brt->brt_mos, brtvd->bv_mos_brtvdev, FTAG, &db)); + + if (brtvd->bv_entcount_dirty) { + /* + * TODO: Walk brtvd->bv_bitmap and write only the dirty blocks. + */ + dmu_write(brt->brt_mos, brtvd->bv_mos_brtvdev, 0, + brtvd->bv_size * sizeof (brtvd->bv_entcount[0]), + brtvd->bv_entcount, tx); + memset(brtvd->bv_bitmap, 0, BT_SIZEOFMAP(brtvd->bv_nblocks)); + brtvd->bv_entcount_dirty = FALSE; + } + + dmu_buf_will_dirty(db, tx); + bvphys = db->db_data; + bvphys->bvp_mos_entries = brtvd->bv_mos_entries; + bvphys->bvp_size = brtvd->bv_size; + if (brtvd->bv_need_byteswap) { + bvphys->bvp_byteorder = BRT_NON_NATIVE_BYTEORDER; + } else { + bvphys->bvp_byteorder = BRT_NATIVE_BYTEORDER; + } + bvphys->bvp_totalcount = brtvd->bv_totalcount; + bvphys->bvp_rangesize = brt->brt_rangesize; + bvphys->bvp_usedspace = brtvd->bv_usedspace; + bvphys->bvp_savedspace = brtvd->bv_savedspace; + dmu_buf_rele(db, FTAG); + + brtvd->bv_meta_dirty = FALSE; +} + +static void +brt_vdevs_alloc(brt_t *brt, boolean_t load) +{ + brt_vdev_t *brtvd; + uint64_t vdevid; + + brt_wlock(brt); + + brt_vdevs_expand(brt, brt->brt_spa->spa_root_vdev->vdev_children); + + if (load) { + for (vdevid = 0; vdevid < brt->brt_nvdevs; vdevid++) { + brtvd = &brt->brt_vdevs[vdevid]; + ASSERT(brtvd->bv_entcount == NULL); + + brt_vdev_load(brt, brtvd); + } + } + + if (brt->brt_rangesize == 0) { + brt->brt_rangesize = BRT_RANGESIZE; + } + + brt_unlock(brt); +} + +static void +brt_vdevs_free(brt_t *brt) +{ + brt_vdev_t *brtvd; + uint64_t vdevid; + + brt_wlock(brt); + + for (vdevid = 0; vdevid < brt->brt_nvdevs; vdevid++) { + brtvd = &brt->brt_vdevs[vdevid]; + if (brtvd->bv_initiated) + brt_vdev_dealloc(brt, brtvd); + } + kmem_free(brt->brt_vdevs, sizeof (brt_vdev_t) * brt->brt_nvdevs); + + brt_unlock(brt); +} + +static void +brt_entry_fill(const blkptr_t *bp, brt_entry_t *bre, uint64_t *vdevidp) +{ + + bre->bre_offset = DVA_GET_OFFSET(&bp->blk_dva[0]); + bre->bre_refcount = 0; + + *vdevidp = DVA_GET_VDEV(&bp->blk_dva[0]); +} + +static int +brt_entry_compare(const void *x1, const void *x2) +{ + const brt_entry_t *bre1 = x1; + const brt_entry_t *bre2 = x2; + + return (TREE_CMP(bre1->bre_offset, bre2->bre_offset)); +} + +static int +brt_entry_lookup(brt_t *brt, brt_vdev_t *brtvd, brt_entry_t *bre) +{ + uint64_t mos_entries; + uint64_t one, physsize; + int error; + + ASSERT(RW_LOCK_HELD(&brt->brt_lock)); + + if (!brt_vdev_lookup(brt, brtvd, bre)) + return (SET_ERROR(ENOENT)); + + /* + * Remember mos_entries object number. After we reacquire the BRT lock, + * the brtvd pointer may be invalid. + */ + mos_entries = brtvd->bv_mos_entries; + if (mos_entries == 0) + return (SET_ERROR(ENOENT)); + + brt_unlock(brt); + + error = zap_length_uint64(brt->brt_mos, mos_entries, &bre->bre_offset, + BRT_KEY_WORDS, &one, &physsize); + if (error == 0) { + ASSERT3U(one, ==, 1); + ASSERT3U(physsize, ==, sizeof (bre->bre_refcount)); + + error = zap_lookup_uint64(brt->brt_mos, mos_entries, + &bre->bre_offset, BRT_KEY_WORDS, 1, + sizeof (bre->bre_refcount), &bre->bre_refcount); + BRT_DEBUG("ZAP lookup: object=%llu vdev=%llu offset=%llu " + "count=%llu error=%d", (u_longlong_t)mos_entries, + (u_longlong_t)brtvd->bv_vdevid, + (u_longlong_t)bre->bre_offset, + error == 0 ? (u_longlong_t)bre->bre_refcount : 0, error); + } + + brt_wlock(brt); + + return (error); +} + +static void +brt_entry_prefetch(brt_t *brt, uint64_t vdevid, brt_entry_t *bre) +{ + brt_vdev_t *brtvd; + uint64_t mos_entries = 0; + + brt_rlock(brt); + brtvd = brt_vdev(brt, vdevid); + if (brtvd != NULL) + mos_entries = brtvd->bv_mos_entries; + brt_unlock(brt); + + if (mos_entries == 0) + return; + + BRT_DEBUG("ZAP prefetch: object=%llu vdev=%llu offset=%llu", + (u_longlong_t)mos_entries, (u_longlong_t)vdevid, + (u_longlong_t)bre->bre_offset); + (void) zap_prefetch_uint64(brt->brt_mos, mos_entries, + (uint64_t *)&bre->bre_offset, BRT_KEY_WORDS); +} + +static int +brt_entry_update(brt_t *brt, brt_vdev_t *brtvd, brt_entry_t *bre, dmu_tx_t *tx) +{ + int error; + + ASSERT(RW_LOCK_HELD(&brt->brt_lock)); + ASSERT(brtvd->bv_mos_entries != 0); + ASSERT(bre->bre_refcount > 0); + + error = zap_update_uint64(brt->brt_mos, brtvd->bv_mos_entries, + (uint64_t *)&bre->bre_offset, BRT_KEY_WORDS, 1, + sizeof (bre->bre_refcount), &bre->bre_refcount, tx); + BRT_DEBUG("ZAP update: object=%llu vdev=%llu offset=%llu count=%llu " + "error=%d", (u_longlong_t)brtvd->bv_mos_entries, + (u_longlong_t)brtvd->bv_vdevid, (u_longlong_t)bre->bre_offset, + (u_longlong_t)bre->bre_refcount, error); + + return (error); +} + +static int +brt_entry_remove(brt_t *brt, brt_vdev_t *brtvd, brt_entry_t *bre, dmu_tx_t *tx) +{ + int error; + + ASSERT(RW_LOCK_HELD(&brt->brt_lock)); + ASSERT(brtvd->bv_mos_entries != 0); + ASSERT0(bre->bre_refcount); + + error = zap_remove_uint64(brt->brt_mos, brtvd->bv_mos_entries, + (uint64_t *)&bre->bre_offset, BRT_KEY_WORDS, tx); + BRT_DEBUG("ZAP remove: object=%llu vdev=%llu offset=%llu count=%llu " + "error=%d", (u_longlong_t)brtvd->bv_mos_entries, + (u_longlong_t)brtvd->bv_vdevid, (u_longlong_t)bre->bre_offset, + (u_longlong_t)bre->bre_refcount, error); + + return (error); +} + +/* + * Return TRUE if we _can_ have BRT entry for this bp. It might be false + * positive, but gives us quick answer if we should look into BRT, which + * may require reads and thus will be more expensive. + */ +boolean_t +brt_maybe_exists(spa_t *spa, const blkptr_t *bp) +{ + brt_t *brt = spa->spa_brt; + brt_vdev_t *brtvd; + brt_entry_t bre_search; + boolean_t mayexists = FALSE; + uint64_t vdevid; + + brt_entry_fill(bp, &bre_search, &vdevid); + + brt_rlock(brt); + + brtvd = brt_vdev(brt, vdevid); + if (brtvd != NULL && brtvd->bv_initiated) { + if (!avl_is_empty(&brtvd->bv_tree) || + brt_vdev_lookup(brt, brtvd, &bre_search)) { + mayexists = TRUE; + } + } + + brt_unlock(brt); + + return (mayexists); +} + +uint64_t +brt_get_dspace(spa_t *spa) +{ + brt_t *brt = spa->spa_brt; + + if (brt == NULL) + return (0); + + return (brt->brt_savedspace); +} + +uint64_t +brt_get_used(spa_t *spa) +{ + brt_t *brt = spa->spa_brt; + + if (brt == NULL) + return (0); + + return (brt->brt_usedspace); +} + +uint64_t +brt_get_saved(spa_t *spa) +{ + brt_t *brt = spa->spa_brt; + + if (brt == NULL) + return (0); + + return (brt->brt_savedspace); +} + +uint64_t +brt_get_ratio(spa_t *spa) +{ + brt_t *brt = spa->spa_brt; + + if (brt->brt_usedspace == 0) + return (100); + + return ((brt->brt_usedspace + brt->brt_savedspace) * 100 / + brt->brt_usedspace); +} + +static int +brt_kstats_update(kstat_t *ksp, int rw) +{ + brt_stats_t *bs = ksp->ks_data; + + if (rw == KSTAT_WRITE) + return (EACCES); + + bs->brt_addref_entry_in_memory.value.ui64 = + wmsum_value(&brt_sums.brt_addref_entry_in_memory); + bs->brt_addref_entry_not_on_disk.value.ui64 = + wmsum_value(&brt_sums.brt_addref_entry_not_on_disk); + bs->brt_addref_entry_on_disk.value.ui64 = + wmsum_value(&brt_sums.brt_addref_entry_on_disk); + bs->brt_addref_entry_read_lost_race.value.ui64 = + wmsum_value(&brt_sums.brt_addref_entry_read_lost_race); + bs->brt_decref_entry_in_memory.value.ui64 = + wmsum_value(&brt_sums.brt_decref_entry_in_memory); + bs->brt_decref_entry_loaded_from_disk.value.ui64 = + wmsum_value(&brt_sums.brt_decref_entry_loaded_from_disk); + bs->brt_decref_entry_not_in_memory.value.ui64 = + wmsum_value(&brt_sums.brt_decref_entry_not_in_memory); + bs->brt_decref_entry_not_on_disk.value.ui64 = + wmsum_value(&brt_sums.brt_decref_entry_not_on_disk); + bs->brt_decref_entry_read_lost_race.value.ui64 = + wmsum_value(&brt_sums.brt_decref_entry_read_lost_race); + bs->brt_decref_entry_still_referenced.value.ui64 = + wmsum_value(&brt_sums.brt_decref_entry_still_referenced); + bs->brt_decref_free_data_later.value.ui64 = + wmsum_value(&brt_sums.brt_decref_free_data_later); + bs->brt_decref_free_data_now.value.ui64 = + wmsum_value(&brt_sums.brt_decref_free_data_now); + bs->brt_decref_no_entry.value.ui64 = + wmsum_value(&brt_sums.brt_decref_no_entry); + + return (0); +} + +static void +brt_stat_init(void) +{ + + wmsum_init(&brt_sums.brt_addref_entry_in_memory, 0); + wmsum_init(&brt_sums.brt_addref_entry_not_on_disk, 0); + wmsum_init(&brt_sums.brt_addref_entry_on_disk, 0); + wmsum_init(&brt_sums.brt_addref_entry_read_lost_race, 0); + wmsum_init(&brt_sums.brt_decref_entry_in_memory, 0); + wmsum_init(&brt_sums.brt_decref_entry_loaded_from_disk, 0); + wmsum_init(&brt_sums.brt_decref_entry_not_in_memory, 0); + wmsum_init(&brt_sums.brt_decref_entry_not_on_disk, 0); + wmsum_init(&brt_sums.brt_decref_entry_read_lost_race, 0); + wmsum_init(&brt_sums.brt_decref_entry_still_referenced, 0); + wmsum_init(&brt_sums.brt_decref_free_data_later, 0); + wmsum_init(&brt_sums.brt_decref_free_data_now, 0); + wmsum_init(&brt_sums.brt_decref_no_entry, 0); + + brt_ksp = kstat_create("zfs", 0, "brtstats", "misc", KSTAT_TYPE_NAMED, + sizeof (brt_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); + if (brt_ksp != NULL) { + brt_ksp->ks_data = &brt_stats; + brt_ksp->ks_update = brt_kstats_update; + kstat_install(brt_ksp); + } +} + +static void +brt_stat_fini(void) +{ + if (brt_ksp != NULL) { + kstat_delete(brt_ksp); + brt_ksp = NULL; + } + + wmsum_fini(&brt_sums.brt_addref_entry_in_memory); + wmsum_fini(&brt_sums.brt_addref_entry_not_on_disk); + wmsum_fini(&brt_sums.brt_addref_entry_on_disk); + wmsum_fini(&brt_sums.brt_addref_entry_read_lost_race); + wmsum_fini(&brt_sums.brt_decref_entry_in_memory); + wmsum_fini(&brt_sums.brt_decref_entry_loaded_from_disk); + wmsum_fini(&brt_sums.brt_decref_entry_not_in_memory); + wmsum_fini(&brt_sums.brt_decref_entry_not_on_disk); + wmsum_fini(&brt_sums.brt_decref_entry_read_lost_race); + wmsum_fini(&brt_sums.brt_decref_entry_still_referenced); + wmsum_fini(&brt_sums.brt_decref_free_data_later); + wmsum_fini(&brt_sums.brt_decref_free_data_now); + wmsum_fini(&brt_sums.brt_decref_no_entry); +} + +void +brt_init(void) +{ + brt_entry_cache = kmem_cache_create("brt_entry_cache", + sizeof (brt_entry_t), 0, NULL, NULL, NULL, NULL, NULL, 0); + brt_pending_entry_cache = kmem_cache_create("brt_pending_entry_cache", + sizeof (brt_pending_entry_t), 0, NULL, NULL, NULL, NULL, NULL, 0); + + brt_stat_init(); +} + +void +brt_fini(void) +{ + brt_stat_fini(); + + kmem_cache_destroy(brt_entry_cache); + kmem_cache_destroy(brt_pending_entry_cache); +} + +static brt_entry_t * +brt_entry_alloc(const brt_entry_t *bre_init) +{ + brt_entry_t *bre; + + bre = kmem_cache_alloc(brt_entry_cache, KM_SLEEP); + bre->bre_offset = bre_init->bre_offset; + bre->bre_refcount = bre_init->bre_refcount; + + return (bre); +} + +static void +brt_entry_free(brt_entry_t *bre) +{ + + kmem_cache_free(brt_entry_cache, bre); +} + +static void +brt_entry_addref(brt_t *brt, const blkptr_t *bp) +{ + brt_vdev_t *brtvd; + brt_entry_t *bre, *racebre; + brt_entry_t bre_search; + avl_index_t where; + uint64_t vdevid; + int error; + + ASSERT(!RW_WRITE_HELD(&brt->brt_lock)); + + brt_entry_fill(bp, &bre_search, &vdevid); + + brt_wlock(brt); + + brtvd = brt_vdev(brt, vdevid); + if (brtvd == NULL) { + ASSERT3U(vdevid, >=, brt->brt_nvdevs); + + /* New VDEV was added. */ + brt_vdevs_expand(brt, vdevid + 1); + brtvd = brt_vdev(brt, vdevid); + } + ASSERT(brtvd != NULL); + if (!brtvd->bv_initiated) + brt_vdev_realloc(brt, brtvd); + + bre = avl_find(&brtvd->bv_tree, &bre_search, NULL); + if (bre != NULL) { + BRTSTAT_BUMP(brt_addref_entry_in_memory); + } else { + /* + * brt_entry_lookup() may drop the BRT (read) lock and + * reacquire it (write). + */ + error = brt_entry_lookup(brt, brtvd, &bre_search); + /* bre_search now contains correct bre_refcount */ + ASSERT(error == 0 || error == ENOENT); + if (error == 0) + BRTSTAT_BUMP(brt_addref_entry_on_disk); + else + BRTSTAT_BUMP(brt_addref_entry_not_on_disk); + /* + * When the BRT lock was dropped, brt_vdevs[] may have been + * expanded and reallocated, we need to update brtvd's pointer. + */ + brtvd = brt_vdev(brt, vdevid); + ASSERT(brtvd != NULL); + + racebre = avl_find(&brtvd->bv_tree, &bre_search, &where); + if (racebre == NULL) { + bre = brt_entry_alloc(&bre_search); + ASSERT(RW_WRITE_HELD(&brt->brt_lock)); + avl_insert(&brtvd->bv_tree, bre, where); + brt->brt_nentries++; + } else { + /* + * The entry was added when the BRT lock was dropped in + * brt_entry_lookup(). + */ + BRTSTAT_BUMP(brt_addref_entry_read_lost_race); + bre = racebre; + } + } + bre->bre_refcount++; + brt_vdev_addref(brt, brtvd, bre, bp_get_dsize(brt->brt_spa, bp)); + + brt_unlock(brt); +} + +/* Return TRUE if block should be freed immediately. */ +boolean_t +brt_entry_decref(spa_t *spa, const blkptr_t *bp) +{ + brt_t *brt = spa->spa_brt; + brt_vdev_t *brtvd; + brt_entry_t *bre, *racebre; + brt_entry_t bre_search; + avl_index_t where; + uint64_t vdevid; + int error; + + brt_entry_fill(bp, &bre_search, &vdevid); + + brt_wlock(brt); + + brtvd = brt_vdev(brt, vdevid); + ASSERT(brtvd != NULL); + + bre = avl_find(&brtvd->bv_tree, &bre_search, NULL); + if (bre != NULL) { + BRTSTAT_BUMP(brt_decref_entry_in_memory); + goto out; + } else { + BRTSTAT_BUMP(brt_decref_entry_not_in_memory); + } + + /* + * brt_entry_lookup() may drop the BRT lock and reacquire it. + */ + error = brt_entry_lookup(brt, brtvd, &bre_search); + /* bre_search now contains correct bre_refcount */ + ASSERT(error == 0 || error == ENOENT); + /* + * When the BRT lock was dropped, brt_vdevs[] may have been expanded + * and reallocated, we need to update brtvd's pointer. + */ + brtvd = brt_vdev(brt, vdevid); + ASSERT(brtvd != NULL); + + if (error == ENOENT) { + BRTSTAT_BUMP(brt_decref_entry_not_on_disk); + bre = NULL; + goto out; + } + + racebre = avl_find(&brtvd->bv_tree, &bre_search, &where); + if (racebre != NULL) { + /* + * The entry was added when the BRT lock was dropped in + * brt_entry_lookup(). + */ + BRTSTAT_BUMP(brt_decref_entry_read_lost_race); + bre = racebre; + goto out; + } + + BRTSTAT_BUMP(brt_decref_entry_loaded_from_disk); + bre = brt_entry_alloc(&bre_search); + ASSERT(RW_WRITE_HELD(&brt->brt_lock)); + avl_insert(&brtvd->bv_tree, bre, where); + brt->brt_nentries++; + +out: + if (bre == NULL) { + /* + * This is a free of a regular (not cloned) block. + */ + brt_unlock(brt); + BRTSTAT_BUMP(brt_decref_no_entry); + return (B_TRUE); + } + if (bre->bre_refcount == 0) { + brt_unlock(brt); + BRTSTAT_BUMP(brt_decref_free_data_now); + return (B_TRUE); + } + + ASSERT(bre->bre_refcount > 0); + bre->bre_refcount--; + if (bre->bre_refcount == 0) + BRTSTAT_BUMP(brt_decref_free_data_later); + else + BRTSTAT_BUMP(brt_decref_entry_still_referenced); + brt_vdev_decref(brt, brtvd, bre, bp_get_dsize(brt->brt_spa, bp)); + + brt_unlock(brt); + + return (B_FALSE); +} + +static void +brt_prefetch(brt_t *brt, const blkptr_t *bp) +{ + brt_entry_t bre; + uint64_t vdevid; + + ASSERT(bp != NULL); + + if (!zfs_brt_prefetch) + return; + + brt_entry_fill(bp, &bre, &vdevid); + + brt_entry_prefetch(brt, vdevid, &bre); +} + +static int +brt_pending_entry_compare(const void *x1, const void *x2) +{ + const brt_pending_entry_t *bpe1 = x1, *bpe2 = x2; + const blkptr_t *bp1 = &bpe1->bpe_bp, *bp2 = &bpe2->bpe_bp; + int cmp; + + cmp = TREE_CMP(BP_PHYSICAL_BIRTH(bp1), BP_PHYSICAL_BIRTH(bp2)); + if (cmp == 0) { + cmp = TREE_CMP(DVA_GET_VDEV(&bp1->blk_dva[0]), + DVA_GET_VDEV(&bp2->blk_dva[0])); + if (cmp == 0) { + cmp = TREE_CMP(DVA_GET_OFFSET(&bp1->blk_dva[0]), + DVA_GET_OFFSET(&bp2->blk_dva[0])); + } + } + + return (cmp); +} + +void +brt_pending_add(spa_t *spa, const blkptr_t *bp, dmu_tx_t *tx) +{ + brt_t *brt; + avl_tree_t *pending_tree; + kmutex_t *pending_lock; + brt_pending_entry_t *bpe, *newbpe; + avl_index_t where; + uint64_t txg; + + brt = spa->spa_brt; + txg = dmu_tx_get_txg(tx); + ASSERT3U(txg, !=, 0); + pending_tree = &brt->brt_pending_tree[txg & TXG_MASK]; + pending_lock = &brt->brt_pending_lock[txg & TXG_MASK]; + + newbpe = kmem_cache_alloc(brt_pending_entry_cache, KM_SLEEP); + newbpe->bpe_bp = *bp; + newbpe->bpe_count = 1; + + mutex_enter(pending_lock); + + bpe = avl_find(pending_tree, newbpe, &where); + if (bpe == NULL) { + avl_insert(pending_tree, newbpe, where); + newbpe = NULL; + } else { + bpe->bpe_count++; + } + + mutex_exit(pending_lock); + + if (newbpe != NULL) { + ASSERT(bpe != NULL); + ASSERT(bpe != newbpe); + kmem_cache_free(brt_pending_entry_cache, newbpe); + } else { + ASSERT(bpe == NULL); + } + + /* Prefetch BRT entry, as we will need it in the syncing context. */ + brt_prefetch(brt, bp); +} + +void +brt_pending_remove(spa_t *spa, const blkptr_t *bp, dmu_tx_t *tx) +{ + brt_t *brt; + avl_tree_t *pending_tree; + kmutex_t *pending_lock; + brt_pending_entry_t *bpe, bpe_search; + uint64_t txg; + + brt = spa->spa_brt; + txg = dmu_tx_get_txg(tx); + ASSERT3U(txg, !=, 0); + pending_tree = &brt->brt_pending_tree[txg & TXG_MASK]; + pending_lock = &brt->brt_pending_lock[txg & TXG_MASK]; + + bpe_search.bpe_bp = *bp; + + mutex_enter(pending_lock); + + bpe = avl_find(pending_tree, &bpe_search, NULL); + /* I believe we should always find bpe when this function is called. */ + if (bpe != NULL) { + ASSERT(bpe->bpe_count > 0); + + bpe->bpe_count--; + if (bpe->bpe_count == 0) { + avl_remove(pending_tree, bpe); + kmem_cache_free(brt_pending_entry_cache, bpe); + } + } + + mutex_exit(pending_lock); +} + +void +brt_pending_apply(spa_t *spa, uint64_t txg) +{ + brt_t *brt; + brt_pending_entry_t *bpe; + avl_tree_t *pending_tree; + kmutex_t *pending_lock; + void *c; + + ASSERT3U(txg, !=, 0); + + brt = spa->spa_brt; + pending_tree = &brt->brt_pending_tree[txg & TXG_MASK]; + pending_lock = &brt->brt_pending_lock[txg & TXG_MASK]; + + mutex_enter(pending_lock); + + c = NULL; + while ((bpe = avl_destroy_nodes(pending_tree, &c)) != NULL) { + boolean_t added_to_ddt; + + mutex_exit(pending_lock); + + for (int i = 0; i < bpe->bpe_count; i++) { + /* + * If the block has DEDUP bit set, it means that it + * already exists in the DEDUP table, so we can just + * use that instead of creating new entry in + * the BRT table. + */ + if (BP_GET_DEDUP(&bpe->bpe_bp)) { + added_to_ddt = ddt_addref(spa, &bpe->bpe_bp); + } else { + added_to_ddt = B_FALSE; + } + if (!added_to_ddt) + brt_entry_addref(brt, &bpe->bpe_bp); + } + + kmem_cache_free(brt_pending_entry_cache, bpe); + mutex_enter(pending_lock); + } + + mutex_exit(pending_lock); +} + +static void +brt_sync_entry(brt_t *brt, brt_vdev_t *brtvd, brt_entry_t *bre, dmu_tx_t *tx) +{ + + ASSERT(RW_WRITE_HELD(&brt->brt_lock)); + ASSERT(brtvd->bv_mos_entries != 0); + + if (bre->bre_refcount == 0) { + int error; + + error = brt_entry_remove(brt, brtvd, bre, tx); + ASSERT(error == 0 || error == ENOENT); + /* + * If error == ENOENT then zfs_clone_range() was done from a + * removed (but opened) file (open(), unlink()). + */ + ASSERT(brt_entry_lookup(brt, brtvd, bre) == ENOENT); + } else { + VERIFY0(brt_entry_update(brt, brtvd, bre, tx)); + } +} + +static void +brt_sync_table(brt_t *brt, dmu_tx_t *tx) +{ + brt_vdev_t *brtvd; + brt_entry_t *bre; + uint64_t vdevid; + void *c; + + brt_wlock(brt); + + for (vdevid = 0; vdevid < brt->brt_nvdevs; vdevid++) { + brtvd = &brt->brt_vdevs[vdevid]; + + if (!brtvd->bv_initiated) + continue; + + if (!brtvd->bv_meta_dirty) { + ASSERT(!brtvd->bv_entcount_dirty); + ASSERT0(avl_numnodes(&brtvd->bv_tree)); + continue; + } + + ASSERT(!brtvd->bv_entcount_dirty || + avl_numnodes(&brtvd->bv_tree) != 0); + + if (brtvd->bv_mos_brtvdev == 0) + brt_vdev_create(brt, brtvd, tx); + + c = NULL; + while ((bre = avl_destroy_nodes(&brtvd->bv_tree, &c)) != NULL) { + brt_sync_entry(brt, brtvd, bre, tx); + brt_entry_free(bre); + ASSERT(brt->brt_nentries > 0); + brt->brt_nentries--; + } + + brt_vdev_sync(brt, brtvd, tx); + + if (brtvd->bv_totalcount == 0) + brt_vdev_destroy(brt, brtvd, tx); + } + + ASSERT0(brt->brt_nentries); + + brt_unlock(brt); +} + +void +brt_sync(spa_t *spa, uint64_t txg) +{ + dmu_tx_t *tx; + brt_t *brt; + + ASSERT(spa_syncing_txg(spa) == txg); + + brt = spa->spa_brt; + brt_rlock(brt); + if (brt->brt_nentries == 0) { + /* No changes. */ + brt_unlock(brt); + return; + } + brt_unlock(brt); + + tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg); + + brt_sync_table(brt, tx); + + dmu_tx_commit(tx); +} + +static void +brt_table_alloc(brt_t *brt) +{ + + for (int i = 0; i < TXG_SIZE; i++) { + avl_create(&brt->brt_pending_tree[i], + brt_pending_entry_compare, + sizeof (brt_pending_entry_t), + offsetof(brt_pending_entry_t, bpe_node)); + mutex_init(&brt->brt_pending_lock[i], NULL, MUTEX_DEFAULT, + NULL); + } +} + +static void +brt_table_free(brt_t *brt) +{ + + for (int i = 0; i < TXG_SIZE; i++) { + ASSERT(avl_is_empty(&brt->brt_pending_tree[i])); + + avl_destroy(&brt->brt_pending_tree[i]); + mutex_destroy(&brt->brt_pending_lock[i]); + } +} + +static void +brt_alloc(spa_t *spa) +{ + brt_t *brt; + + ASSERT(spa->spa_brt == NULL); + + brt = kmem_zalloc(sizeof (*brt), KM_SLEEP); + rw_init(&brt->brt_lock, NULL, RW_DEFAULT, NULL); + brt->brt_spa = spa; + brt->brt_rangesize = 0; + brt->brt_nentries = 0; + brt->brt_vdevs = NULL; + brt->brt_nvdevs = 0; + brt_table_alloc(brt); + + spa->spa_brt = brt; +} + +void +brt_create(spa_t *spa) +{ + + brt_alloc(spa); + brt_vdevs_alloc(spa->spa_brt, B_FALSE); +} + +int +brt_load(spa_t *spa) +{ + + brt_alloc(spa); + brt_vdevs_alloc(spa->spa_brt, B_TRUE); + + return (0); +} + +void +brt_unload(spa_t *spa) +{ + brt_t *brt = spa->spa_brt; + + if (brt == NULL) + return; + + brt_vdevs_free(brt); + brt_table_free(brt); + rw_destroy(&brt->brt_lock); + kmem_free(brt, sizeof (*brt)); + spa->spa_brt = NULL; +} + +/* BEGIN CSTYLED */ +ZFS_MODULE_PARAM(zfs_brt, zfs_brt_, prefetch, INT, ZMOD_RW, + "Enable prefetching of BRT entries"); +#ifdef ZFS_BRT_DEBUG +ZFS_MODULE_PARAM(zfs_brt, zfs_brt_, debug, INT, ZMOD_RW, "BRT debug"); +#endif +/* END CSTYLED */ diff --git a/module/zfs/dbuf.c b/module/zfs/dbuf.c index 191e5e043942..94c2ae9d736d 100644 --- a/module/zfs/dbuf.c +++ b/module/zfs/dbuf.c @@ -26,6 +26,7 @@ * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. * Copyright (c) 2019, Klara Inc. * Copyright (c) 2019, Allan Jude + * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek */ #include @@ -49,6 +50,7 @@ #include #include #include +#include #include #include #include @@ -1427,7 +1429,7 @@ dbuf_read_bonus(dmu_buf_impl_t *db, dnode_t *dn, uint32_t flags) } static void -dbuf_handle_indirect_hole(dmu_buf_impl_t *db, dnode_t *dn) +dbuf_handle_indirect_hole(dmu_buf_impl_t *db, dnode_t *dn, blkptr_t *dbbp) { blkptr_t *bps = db->db.db_data; uint32_t indbs = 1ULL << dn->dn_indblkshift; @@ -1436,12 +1438,12 @@ dbuf_handle_indirect_hole(dmu_buf_impl_t *db, dnode_t *dn) for (int i = 0; i < n_bps; i++) { blkptr_t *bp = &bps[i]; - ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, indbs); - BP_SET_LSIZE(bp, BP_GET_LEVEL(db->db_blkptr) == 1 ? - dn->dn_datablksz : BP_GET_LSIZE(db->db_blkptr)); - BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr)); - BP_SET_LEVEL(bp, BP_GET_LEVEL(db->db_blkptr) - 1); - BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0); + ASSERT3U(BP_GET_LSIZE(dbbp), ==, indbs); + BP_SET_LSIZE(bp, BP_GET_LEVEL(dbbp) == 1 ? + dn->dn_datablksz : BP_GET_LSIZE(dbbp)); + BP_SET_TYPE(bp, BP_GET_TYPE(dbbp)); + BP_SET_LEVEL(bp, BP_GET_LEVEL(dbbp) - 1); + BP_SET_BIRTH(bp, dbbp->blk_birth, 0); } } @@ -1451,30 +1453,27 @@ dbuf_handle_indirect_hole(dmu_buf_impl_t *db, dnode_t *dn) * was taken, ENOENT if no action was taken. */ static int -dbuf_read_hole(dmu_buf_impl_t *db, dnode_t *dn) +dbuf_read_hole(dmu_buf_impl_t *db, dnode_t *dn, blkptr_t *bp) { ASSERT(MUTEX_HELD(&db->db_mtx)); - int is_hole = db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr); + int is_hole = bp == NULL || BP_IS_HOLE(bp); /* * For level 0 blocks only, if the above check fails: * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() * processes the delete record and clears the bp while we are waiting * for the dn_mtx (resulting in a "no" from block_freed). */ - if (!is_hole && db->db_level == 0) { - is_hole = dnode_block_freed(dn, db->db_blkid) || - BP_IS_HOLE(db->db_blkptr); - } + if (!is_hole && db->db_level == 0) + is_hole = dnode_block_freed(dn, db->db_blkid) || BP_IS_HOLE(bp); if (is_hole) { dbuf_set_data(db, dbuf_alloc_arcbuf(db)); memset(db->db.db_data, 0, db->db.db_size); - if (db->db_blkptr != NULL && db->db_level > 0 && - BP_IS_HOLE(db->db_blkptr) && - db->db_blkptr->blk_birth != 0) { - dbuf_handle_indirect_hole(db, dn); + if (bp != NULL && db->db_level > 0 && BP_IS_HOLE(bp) && + bp->blk_birth != 0) { + dbuf_handle_indirect_hole(db, dn, bp); } db->db_state = DB_CACHED; DTRACE_SET_STATE(db, "hole read satisfied"); @@ -1551,12 +1550,13 @@ dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags, zbookmark_phys_t zb; uint32_t aflags = ARC_FLAG_NOWAIT; int err, zio_flags; + blkptr_t bp, *bpp; DB_DNODE_ENTER(db); dn = DB_DNODE(db); ASSERT(!zfs_refcount_is_zero(&db->db_holds)); ASSERT(MUTEX_HELD(&db->db_mtx)); - ASSERT(db->db_state == DB_UNCACHED); + ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); ASSERT(db->db_buf == NULL); ASSERT(db->db_parent == NULL || RW_LOCK_HELD(&db->db_parent->db_rwlock)); @@ -1566,16 +1566,46 @@ dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags, goto early_unlock; } - err = dbuf_read_hole(db, dn); + if (db->db_state == DB_UNCACHED) { + if (db->db_blkptr == NULL) { + bpp = NULL; + } else { + bp = *db->db_blkptr; + bpp = &bp; + } + } else { + struct dirty_leaf *dl; + dbuf_dirty_record_t *dr; + + ASSERT3S(db->db_state, ==, DB_NOFILL); + + dr = list_head(&db->db_dirty_records); + if (dr == NULL) { + err = EIO; + goto early_unlock; + } else { + dl = &dr->dt.dl; + if (!dl->dr_brtwrite) { + err = EIO; + goto early_unlock; + } + bp = dl->dr_overridden_by; + bpp = &bp; + } + } + + err = dbuf_read_hole(db, dn, bpp); if (err == 0) goto early_unlock; + ASSERT(bpp != NULL); + /* * Any attempt to read a redacted block should result in an error. This * will never happen under normal conditions, but can be useful for * debugging purposes. */ - if (BP_IS_REDACTED(db->db_blkptr)) { + if (BP_IS_REDACTED(bpp)) { ASSERT(dsl_dataset_feature_is_active( db->db_objset->os_dsl_dataset, SPA_FEATURE_REDACTED_DATASETS)); @@ -1590,7 +1620,7 @@ dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags, * All bps of an encrypted os should have the encryption bit set. * If this is not true it indicates tampering and we report an error. */ - if (db->db_objset->os_encrypted && !BP_USES_CRYPT(db->db_blkptr)) { + if (db->db_objset->os_encrypted && !BP_USES_CRYPT(bpp)) { spa_log_error(db->db_objset->os_spa, &zb); zfs_panic_recover("unencrypted block in encrypted " "object set %llu", dmu_objset_id(db->db_objset)); @@ -1621,15 +1651,14 @@ dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags, if ((flags & DB_RF_NO_DECRYPT) && BP_IS_PROTECTED(db->db_blkptr)) zio_flags |= ZIO_FLAG_RAW; /* - * The zio layer will copy the provided blkptr later, but we need to - * do this now so that we can release the parent's rwlock. We have to - * do that now so that if dbuf_read_done is called synchronously (on + * The zio layer will copy the provided blkptr later, but we have our + * own copy so that we can release the parent's rwlock. We have to + * do that so that if dbuf_read_done is called synchronously (on * an l1 cache hit) we don't acquire the db_mtx while holding the * parent's rwlock, which would be a lock ordering violation. */ - blkptr_t bp = *db->db_blkptr; dmu_buf_unlock_parent(db, dblt, tag); - (void) arc_read(zio, db->db_objset->os_spa, &bp, + (void) arc_read(zio, db->db_objset->os_spa, bpp, dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); return (err); @@ -1731,9 +1760,6 @@ dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) */ ASSERT(!zfs_refcount_is_zero(&db->db_holds)); - if (db->db_state == DB_NOFILL) - return (SET_ERROR(EIO)); - DB_DNODE_ENTER(db); dn = DB_DNODE(db); @@ -1780,13 +1806,13 @@ dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) } DB_DNODE_EXIT(db); DBUF_STAT_BUMP(hash_hits); - } else if (db->db_state == DB_UNCACHED) { + } else if (db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL) { boolean_t need_wait = B_FALSE; db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER, FTAG); - if (zio == NULL && - db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { + if (zio == NULL && (db->db_state == DB_NOFILL || + (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)))) { spa_t *spa = dn->dn_objset->os_spa; zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); need_wait = B_TRUE; @@ -1913,7 +1939,8 @@ dbuf_unoverride(dbuf_dirty_record_t *dr) * the buf thawed to save the effort of freezing & * immediately re-thawing it. */ - arc_release(dr->dt.dl.dr_data, db); + if (!dr->dt.dl.dr_brtwrite) + arc_release(dr->dt.dl.dr_data, db); } /* @@ -1996,6 +2023,11 @@ dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, db->db_blkid > dn->dn_maxblkid) dn->dn_maxblkid = db->db_blkid; dbuf_unoverride(dr); + if (dr->dt.dl.dr_brtwrite) { + ASSERT(db->db.db_data == NULL); + mutex_exit(&db->db_mtx); + continue; + } } else { /* * This dbuf is not dirty in the open context. @@ -2285,7 +2317,7 @@ dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); - if (db->db_blkid != DMU_BONUS_BLKID) { + if (db->db_blkid != DMU_BONUS_BLKID && db->db_state != DB_NOFILL) { dmu_objset_willuse_space(os, db->db.db_size, tx); } @@ -2328,8 +2360,9 @@ dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) sizeof (dbuf_dirty_record_t), offsetof(dbuf_dirty_record_t, dr_dirty_node)); } - if (db->db_blkid != DMU_BONUS_BLKID) + if (db->db_blkid != DMU_BONUS_BLKID && db->db_state != DB_NOFILL) { dr->dr_accounted = db->db.db_size; + } dr->dr_dbuf = db; dr->dr_txg = tx->tx_txg; list_insert_before(&db->db_dirty_records, dr_next, dr); @@ -2489,6 +2522,7 @@ static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) { uint64_t txg = tx->tx_txg; + boolean_t brtwrite; ASSERT(txg != 0); @@ -2513,6 +2547,16 @@ dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) return (B_FALSE); ASSERT(dr->dr_dbuf == db); + brtwrite = dr->dt.dl.dr_brtwrite; + if (brtwrite) { + /* + * We are freeing a block that we cloned in the same + * transaction group. + */ + brt_pending_remove(dmu_objset_spa(db->db_objset), + &dr->dt.dl.dr_overridden_by, tx); + } + dnode_t *dn = dr->dr_dnode; dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); @@ -2542,7 +2586,7 @@ dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) mutex_exit(&dn->dn_mtx); } - if (db->db_state != DB_NOFILL) { + if (db->db_state != DB_NOFILL && !brtwrite) { dbuf_unoverride(dr); ASSERT(db->db_buf != NULL); @@ -2557,7 +2601,8 @@ dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) db->db_dirtycnt -= 1; if (zfs_refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { - ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf)); + ASSERT(db->db_state == DB_NOFILL || brtwrite || + arc_released(db->db_buf)); dbuf_destroy(db); return (B_TRUE); } @@ -4748,8 +4793,10 @@ dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) ASSERT(db->db_blkid != DMU_BONUS_BLKID); ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); if (db->db_state != DB_NOFILL) { - if (dr->dt.dl.dr_data != db->db_buf) + if (dr->dt.dl.dr_data != NULL && + dr->dt.dl.dr_data != db->db_buf) { arc_buf_destroy(dr->dt.dl.dr_data, db); + } } } else { ASSERT(list_head(&dr->dt.di.dr_children) == NULL); @@ -5046,7 +5093,8 @@ dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) mutex_enter(&db->db_mtx); dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, - dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); + dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite, + dr->dt.dl.dr_brtwrite); mutex_exit(&db->db_mtx); } else if (db->db_state == DB_NOFILL) { ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || diff --git a/module/zfs/ddt.c b/module/zfs/ddt.c index 7880a899aeb1..33fea0ba3d3c 100644 --- a/module/zfs/ddt.c +++ b/module/zfs/ddt.c @@ -22,6 +22,7 @@ /* * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved. * Copyright (c) 2012, 2016 by Delphix. All rights reserved. + * Copyright (c) 2022 by Pawel Jakub Dawidek */ #include @@ -1180,5 +1181,59 @@ ddt_walk(spa_t *spa, ddt_bookmark_t *ddb, ddt_entry_t *dde) return (SET_ERROR(ENOENT)); } +/* + * This function is used by Block Cloning (brt.c) to increase reference + * counter for the DDT entry if the block is already in DDT. + * + * Return false if the block, despite having the D bit set, is not present + * in the DDT. Currently this is not possible but might be in the future. + * See the comment below. + */ +boolean_t +ddt_addref(spa_t *spa, const blkptr_t *bp) +{ + ddt_t *ddt; + ddt_entry_t *dde; + boolean_t result; + + spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER); + ddt = ddt_select(spa, bp); + ddt_enter(ddt); + + dde = ddt_lookup(ddt, bp, B_TRUE); + ASSERT(dde != NULL); + + if (dde->dde_type < DDT_TYPES) { + ddt_phys_t *ddp; + + ASSERT3S(dde->dde_class, <, DDT_CLASSES); + + ddp = &dde->dde_phys[BP_GET_NDVAS(bp)]; + if (ddp->ddp_refcnt == 0) { + /* This should never happen? */ + ddt_phys_fill(ddp, bp); + } + ddt_phys_addref(ddp); + result = B_TRUE; + } else { + /* + * At the time of implementating this if the block has the + * DEDUP flag set it must exist in the DEDUP table, but + * there are many advocates that want ability to remove + * entries from DDT with refcnt=1. If this will happen, + * we may have a block with the DEDUP set, but which doesn't + * have a corresponding entry in the DDT. Be ready. + */ + ASSERT3S(dde->dde_class, ==, DDT_CLASSES); + ddt_remove(ddt, dde); + result = B_FALSE; + } + + ddt_exit(ddt); + spa_config_exit(spa, SCL_ZIO, FTAG); + + return (result); +} + ZFS_MODULE_PARAM(zfs_dedup, zfs_dedup_, prefetch, INT, ZMOD_RW, "Enable prefetching dedup-ed blks"); diff --git a/module/zfs/dmu.c b/module/zfs/dmu.c index 9b8fc7e49b2d..e6bade11c859 100644 --- a/module/zfs/dmu.c +++ b/module/zfs/dmu.c @@ -29,6 +29,7 @@ * Copyright (c) 2019, Klara Inc. * Copyright (c) 2019, Allan Jude * Copyright (c) 2022 Hewlett Packard Enterprise Development LP. + * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek */ #include @@ -52,6 +53,7 @@ #include #include #include +#include #include #include #include @@ -513,7 +515,7 @@ dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, zio_t *zio = NULL; boolean_t missed = B_FALSE; - ASSERT(length <= DMU_MAX_ACCESS); + ASSERT(!read || length <= DMU_MAX_ACCESS); /* * Note: We directly notify the prefetch code of this read, so that @@ -2165,6 +2167,155 @@ dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) return (err); } +int +dmu_read_l0_bps(objset_t *os, uint64_t object, uint64_t offset, uint64_t length, + dmu_tx_t *tx, blkptr_t *bps, size_t *nbpsp) +{ + dmu_buf_t **dbp, *dbuf; + dmu_buf_impl_t *db; + blkptr_t *bp; + int error, numbufs; + + error = dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG, + &numbufs, &dbp); + if (error != 0) { + if (error == ESRCH) { + error = SET_ERROR(ENXIO); + } + return (error); + } + + ASSERT3U(numbufs, <=, *nbpsp); + + for (int i = 0; i < numbufs; i++) { + dbuf = dbp[i]; + db = (dmu_buf_impl_t *)dbuf; + bp = db->db_blkptr; + + /* + * If the block is not on the disk yet, it has no BP assigned. + * There is not much we can do... + */ + if (!list_is_empty(&db->db_dirty_records)) { + dbuf_dirty_record_t *dr; + + dr = list_head(&db->db_dirty_records); + if (dr->dt.dl.dr_brtwrite) { + /* + * This is very special case where we clone a + * block and in the same transaction group we + * read its BP (most likely to clone the clone). + */ + bp = &dr->dt.dl.dr_overridden_by; + } else { + /* + * The block was modified in the same + * transaction group. + */ + error = SET_ERROR(EAGAIN); + goto out; + } + } + if (bp == NULL) { + /* + * The block was created in this transaction group, + * so it has no BP yet. + */ + error = SET_ERROR(EAGAIN); + goto out; + } + if (dmu_buf_is_dirty(dbuf, tx)) { + error = SET_ERROR(EAGAIN); + goto out; + } + /* + * Make sure we clone only data blocks. + */ + if (BP_IS_METADATA(bp) && !BP_IS_HOLE(bp)) { + error = SET_ERROR(EINVAL); + goto out; + } + + bps[i] = *bp; + } + + *nbpsp = numbufs; +out: + dmu_buf_rele_array(dbp, numbufs, FTAG); + + return (error); +} + +void +dmu_brt_clone(objset_t *os, uint64_t object, uint64_t offset, uint64_t length, + dmu_tx_t *tx, const blkptr_t *bps, size_t nbps, boolean_t replay) +{ + spa_t *spa; + dmu_buf_t **dbp, *dbuf; + dmu_buf_impl_t *db; + struct dirty_leaf *dl; + dbuf_dirty_record_t *dr; + const blkptr_t *bp; + int numbufs; + + spa = os->os_spa; + + VERIFY0(dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG, + &numbufs, &dbp)); + ASSERT3U(nbps, ==, numbufs); + + for (int i = 0; i < numbufs; i++) { + dbuf = dbp[i]; + db = (dmu_buf_impl_t *)dbuf; + bp = &bps[i]; + + ASSERT0(db->db_level); + ASSERT(db->db_blkid != DMU_BONUS_BLKID); + ASSERT(BP_IS_HOLE(bp) || dbuf->db_size == BP_GET_LSIZE(bp)); + + if (db->db_state == DB_UNCACHED) { + /* + * XXX-PJD: If the dbuf is already cached, calling + * dmu_buf_will_not_fill() will panic on assertion + * (db->db_buf == NULL) in dbuf_clear_data(), + * which is called from dbuf_noread() in DB_NOFILL + * case. I'm not 100% sure this is the right thing + * to do, but it seems to work. + */ + dmu_buf_will_not_fill(dbuf, tx); + } + + dr = list_head(&db->db_dirty_records); + ASSERT3U(dr->dr_txg, ==, tx->tx_txg); + dl = &dr->dt.dl; + dl->dr_overridden_by = *bp; + dl->dr_brtwrite = B_TRUE; + + dl->dr_override_state = DR_OVERRIDDEN; + if (BP_IS_HOLE(bp)) { + dl->dr_overridden_by.blk_birth = 0; + dl->dr_overridden_by.blk_phys_birth = 0; + } else { + dl->dr_overridden_by.blk_birth = dr->dr_txg; + dl->dr_overridden_by.blk_phys_birth = + BP_PHYSICAL_BIRTH(bp); + } + + /* + * When data in embedded into BP there is no need to create + * BRT entry as there is no data block. Just copy the BP as + * it contains the data. + * Also, when replaying ZIL we don't want to bump references + * in the BRT as it was already done during ZIL claim. + */ + if (!replay && !BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) { + brt_pending_add(spa, bp, tx); + } + } + + dmu_buf_rele_array(dbp, numbufs, FTAG); +} + void __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) { diff --git a/module/zfs/dmu_tx.c b/module/zfs/dmu_tx.c index 815e27a6c7f7..1c5608c4541b 100644 --- a/module/zfs/dmu_tx.c +++ b/module/zfs/dmu_tx.c @@ -349,7 +349,7 @@ dmu_tx_mark_netfree(dmu_tx_t *tx) } static void -dmu_tx_hold_free_impl(dmu_tx_hold_t *txh, uint64_t off, uint64_t len) +dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len) { dmu_tx_t *tx = txh->txh_tx; dnode_t *dn = txh->txh_dnode; @@ -357,15 +357,11 @@ dmu_tx_hold_free_impl(dmu_tx_hold_t *txh, uint64_t off, uint64_t len) ASSERT(tx->tx_txg == 0); - dmu_tx_count_dnode(txh); - if (off >= (dn->dn_maxblkid + 1) * dn->dn_datablksz) return; if (len == DMU_OBJECT_END) len = (dn->dn_maxblkid + 1) * dn->dn_datablksz - off; - dmu_tx_count_dnode(txh); - /* * For i/o error checking, we read the first and last level-0 * blocks if they are not aligned, and all the level-1 blocks. @@ -445,8 +441,10 @@ dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len) txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object, THT_FREE, off, len); - if (txh != NULL) - (void) dmu_tx_hold_free_impl(txh, off, len); + if (txh != NULL) { + dmu_tx_count_dnode(txh); + dmu_tx_count_free(txh, off, len); + } } void @@ -455,8 +453,35 @@ dmu_tx_hold_free_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len) dmu_tx_hold_t *txh; txh = dmu_tx_hold_dnode_impl(tx, dn, THT_FREE, off, len); - if (txh != NULL) - (void) dmu_tx_hold_free_impl(txh, off, len); + if (txh != NULL) { + dmu_tx_count_dnode(txh); + dmu_tx_count_free(txh, off, len); + } +} + +static void +dmu_tx_count_clone(dmu_tx_hold_t *txh, uint64_t off, uint64_t len) +{ + + /* + * Reuse dmu_tx_count_free(), it does exactly what we need for clone. + */ + dmu_tx_count_free(txh, off, len); +} + +void +dmu_tx_hold_clone_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len) +{ + dmu_tx_hold_t *txh; + + ASSERT0(tx->tx_txg); + ASSERT(len == 0 || UINT64_MAX - off >= len - 1); + + txh = dmu_tx_hold_dnode_impl(tx, dn, THT_CLONE, off, len); + if (txh != NULL) { + dmu_tx_count_dnode(txh); + dmu_tx_count_clone(txh, off, len); + } } static void @@ -667,6 +692,10 @@ dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db) case THT_NEWOBJECT: match_object = TRUE; break; + case THT_CLONE: + if (blkid >= beginblk && blkid <= endblk) + match_offset = TRUE; + break; default: cmn_err(CE_PANIC, "bad txh_type %d", txh->txh_type); diff --git a/module/zfs/dsl_scan.c b/module/zfs/dsl_scan.c index 021cba68cd21..8e3fd126caa5 100644 --- a/module/zfs/dsl_scan.c +++ b/module/zfs/dsl_scan.c @@ -47,6 +47,7 @@ #include #include #include +#include #include #include #include @@ -3499,11 +3500,12 @@ dsl_process_async_destroys(dsl_pool_t *dp, dmu_tx_t *tx) scn->scn_dedup_frees_this_txg = 0; /* - * Write out changes to the DDT that may be required as a - * result of the blocks freed. This ensures that the DDT - * is clean when a scrub/resilver runs. + * Write out changes to the DDT and the BRT that may be required + * as a result of the blocks freed. This ensures that the DDT + * and the BRT are clean when a scrub/resilver runs. */ ddt_sync(spa, tx->tx_txg); + brt_sync(spa, tx->tx_txg); } if (err != 0) return (err); diff --git a/module/zfs/spa.c b/module/zfs/spa.c index 6be6fe11561a..98a302237df8 100644 --- a/module/zfs/spa.c +++ b/module/zfs/spa.c @@ -52,6 +52,7 @@ #include #include #include +#include #include #include #include @@ -341,6 +342,12 @@ spa_prop_get_config(spa_t *spa, nvlist_t **nvp) spa_prop_add_list(*nvp, ZPOOL_PROP_DEDUPRATIO, NULL, ddt_get_pool_dedup_ratio(spa), src); + spa_prop_add_list(*nvp, ZPOOL_PROP_BCLONEUSED, NULL, + brt_get_used(spa), src); + spa_prop_add_list(*nvp, ZPOOL_PROP_BCLONESAVED, NULL, + brt_get_saved(spa), src); + spa_prop_add_list(*nvp, ZPOOL_PROP_BCLONERATIO, NULL, + brt_get_ratio(spa), src); spa_prop_add_list(*nvp, ZPOOL_PROP_HEALTH, NULL, rvd->vdev_state, src); @@ -1707,6 +1714,7 @@ spa_unload(spa_t *spa) } ddt_unload(spa); + brt_unload(spa); spa_unload_log_sm_metadata(spa); /* @@ -4414,6 +4422,21 @@ spa_ld_load_dedup_tables(spa_t *spa) return (0); } +static int +spa_ld_load_brt(spa_t *spa) +{ + int error = 0; + vdev_t *rvd = spa->spa_root_vdev; + + error = brt_load(spa); + if (error != 0) { + spa_load_failed(spa, "brt_load failed [error=%d]", error); + return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); + } + + return (0); +} + static int spa_ld_verify_logs(spa_t *spa, spa_import_type_t type, const char **ereport) { @@ -4895,6 +4918,10 @@ spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport) if (error != 0) return (error); + error = spa_ld_load_brt(spa); + if (error != 0) + return (error); + /* * Verify the logs now to make sure we don't have any unexpected errors * when we claim log blocks later. @@ -5963,6 +5990,10 @@ spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props, * Create DDTs (dedup tables). */ ddt_create(spa); + /* + * Create BRT table and BRT table object. + */ + brt_create(spa); spa_update_dspace(spa); @@ -9138,6 +9169,7 @@ spa_sync_iterate_to_convergence(spa_t *spa, dmu_tx_t *tx) &spa->spa_deferred_bpobj, tx); } + brt_sync(spa, txg); ddt_sync(spa, txg); dsl_scan_sync(dp, tx); svr_sync(spa, tx); @@ -9262,6 +9294,13 @@ spa_sync(spa_t *spa, uint64_t txg) spa->spa_txg_zio[txg & TXG_MASK] = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); + /* + * Now that there can be no more cloning in this transaction group, + * but we are still before issuing frees, we can process pending BRT + * updates. + */ + brt_pending_apply(spa, txg); + /* * Lock out configuration changes. */ diff --git a/module/zfs/spa_misc.c b/module/zfs/spa_misc.c index 53763e915ca8..8466fa80e1e3 100644 --- a/module/zfs/spa_misc.c +++ b/module/zfs/spa_misc.c @@ -57,6 +57,7 @@ #include #include #include +#include #include #include #include "zfs_prop.h" @@ -1834,7 +1835,7 @@ void spa_update_dspace(spa_t *spa) { spa->spa_dspace = metaslab_class_get_dspace(spa_normal_class(spa)) + - ddt_get_dedup_dspace(spa); + ddt_get_dedup_dspace(spa) + brt_get_dspace(spa); if (spa->spa_nonallocating_dspace > 0) { /* * Subtract the space provided by all non-allocating vdevs that @@ -2410,6 +2411,7 @@ spa_init(spa_mode_t mode) unique_init(); zfs_btree_init(); metaslab_stat_init(); + brt_init(); ddt_init(); zio_init(); dmu_init(); @@ -2446,6 +2448,7 @@ spa_fini(void) dmu_fini(); zio_fini(); ddt_fini(); + brt_fini(); metaslab_stat_fini(); zfs_btree_fini(); unique_fini(); diff --git a/module/zfs/zfs_ioctl.c b/module/zfs/zfs_ioctl.c index 24ae0a00d92f..9b859adc5551 100644 --- a/module/zfs/zfs_ioctl.c +++ b/module/zfs/zfs_ioctl.c @@ -23,7 +23,7 @@ * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. * Portions Copyright 2011 Martin Matuska * Copyright 2015, OmniTI Computer Consulting, Inc. All rights reserved. - * Portions Copyright 2012 Pawel Jakub Dawidek + * Copyright (c) 2012 Pawel Jakub Dawidek * Copyright (c) 2014, 2016 Joyent, Inc. All rights reserved. * Copyright 2016 Nexenta Systems, Inc. All rights reserved. * Copyright (c) 2014, Joyent, Inc. All rights reserved. diff --git a/module/zfs/zfs_log.c b/module/zfs/zfs_log.c index 77bf9140d52d..d009c58d8644 100644 --- a/module/zfs/zfs_log.c +++ b/module/zfs/zfs_log.c @@ -21,6 +21,7 @@ /* * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. * Copyright (c) 2015, 2018 by Delphix. All rights reserved. + * Copyright (c) 2022 by Pawel Jakub Dawidek */ @@ -891,5 +892,56 @@ zfs_log_acl(zilog_t *zilog, dmu_tx_t *tx, znode_t *zp, zil_itx_assign(zilog, itx, tx); } +/* + * Handles TX_CLONE_RANGE transactions. + */ +void +zfs_log_clone_range(zilog_t *zilog, dmu_tx_t *tx, int txtype, znode_t *zp, + uint64_t off, uint64_t len, uint64_t blksz, const blkptr_t *bps, + size_t nbps) +{ + itx_t *itx; + lr_clone_range_t *lr; + uint64_t partlen, max_log_data; + size_t i, partnbps; + + VERIFY(!zil_replaying(zilog, tx)); + + if (zp->z_unlinked) + return; + + max_log_data = zil_max_log_data(zilog, sizeof (lr_clone_range_t)); + + while (nbps > 0) { + partnbps = MIN(nbps, max_log_data / sizeof (bps[0])); + partlen = 0; + for (i = 0; i < partnbps; i++) { + partlen += BP_GET_LSIZE(&bps[i]); + } + partlen = MIN(partlen, len); + + itx = zil_itx_create(txtype, + sizeof (*lr) + sizeof (bps[0]) * partnbps); + lr = (lr_clone_range_t *)&itx->itx_lr; + lr->lr_foid = zp->z_id; + lr->lr_offset = off; + lr->lr_length = partlen; + lr->lr_blksz = blksz; + lr->lr_nbps = partnbps; + memcpy(lr->lr_bps, bps, sizeof (bps[0]) * partnbps); + + itx->itx_sync = (zp->z_sync_cnt != 0); + + zil_itx_assign(zilog, itx, tx); + + bps += partnbps; + ASSERT3U(nbps, >=, partnbps); + nbps -= partnbps; + off += partlen; + ASSERT3U(len, >=, partlen); + len -= partlen; + } +} + ZFS_MODULE_PARAM(zfs, zfs_, immediate_write_sz, S64, ZMOD_RW, "Largest data block to write to zil"); diff --git a/module/zfs/zfs_quota.c b/module/zfs/zfs_quota.c index a5dc5c399b5d..9b351eefc04e 100644 --- a/module/zfs/zfs_quota.c +++ b/module/zfs/zfs_quota.c @@ -20,8 +20,7 @@ */ /* * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. - * Copyright (c) 2011 Pawel Jakub Dawidek . - * All rights reserved. + * Copyright (c) 2011 Pawel Jakub Dawidek * Copyright (c) 2012, 2015, 2018 by Delphix. All rights reserved. * Copyright (c) 2014 Integros [integros.com] * Copyright 2016 Nexenta Systems, Inc. All rights reserved. diff --git a/module/zfs/zfs_replay.c b/module/zfs/zfs_replay.c index 32be27a8ba6e..04dfda56b3f1 100644 --- a/module/zfs/zfs_replay.c +++ b/module/zfs/zfs_replay.c @@ -22,6 +22,7 @@ * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. * Copyright (c) 2012 Cyril Plisko. All rights reserved. * Copyright (c) 2013, 2017 by Delphix. All rights reserved. + * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek */ #include @@ -1162,6 +1163,34 @@ zfs_replay_acl(void *arg1, void *arg2, boolean_t byteswap) return (error); } +static int +zfs_replay_clone_range(void *arg1, void *arg2, boolean_t byteswap) +{ + zfsvfs_t *zfsvfs = arg1; + lr_clone_range_t *lr = arg2; + znode_t *zp; + int error; + + if (byteswap) + byteswap_uint64_array(lr, sizeof (*lr)); + + if ((error = zfs_zget(zfsvfs, lr->lr_foid, &zp)) != 0) { + /* + * Clones can be logged out of order, so don't be surprised if + * the file is gone - just return success. + */ + if (error == ENOENT) + error = 0; + return (error); + } + + error = zfs_clone_range_replay(zp, lr->lr_offset, lr->lr_length, + lr->lr_blksz, lr->lr_bps, lr->lr_nbps); + + zrele(zp); + return (error); +} + /* * Callback vectors for replaying records */ @@ -1190,4 +1219,5 @@ zil_replay_func_t *const zfs_replay_vector[TX_MAX_TYPE] = { zfs_replay_setsaxattr, /* TX_SETSAXATTR */ zfs_replay_rename_exchange, /* TX_RENAME_EXCHANGE */ zfs_replay_rename_whiteout, /* TX_RENAME_WHITEOUT */ + zfs_replay_clone_range, /* TX_CLONE_RANGE */ }; diff --git a/module/zfs/zfs_vnops.c b/module/zfs/zfs_vnops.c index 10677d8d9947..db80be783899 100644 --- a/module/zfs/zfs_vnops.c +++ b/module/zfs/zfs_vnops.c @@ -24,6 +24,7 @@ * Copyright (c) 2012, 2018 by Delphix. All rights reserved. * Copyright (c) 2015 by Chunwei Chen. All rights reserved. * Copyright 2017 Nexenta Systems, Inc. + * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek */ /* Portions Copyright 2007 Jeremy Teo */ @@ -50,6 +51,7 @@ #include #include #include +#include #include #include #include @@ -501,7 +503,7 @@ zfs_write(znode_t *zp, zfs_uio_t *uio, int ioflag, cred_t *cr) lr = zfs_rangelock_enter(&zp->z_rangelock, woff, n, RL_WRITER); } - if (zn_rlimit_fsize(zp, uio)) { + if (zn_rlimit_fsize_uio(zp, uio)) { zfs_rangelock_exit(lr); zfs_exit(zfsvfs, FTAG); return (SET_ERROR(EFBIG)); @@ -995,6 +997,467 @@ zfs_get_done(zgd_t *zgd, int error) kmem_free(zgd, sizeof (zgd_t)); } +static int +zfs_enter_two(zfsvfs_t *zfsvfs1, zfsvfs_t *zfsvfs2, const char *tag) +{ + int error; + + /* Swap. Not sure if the order of zfs_enter()s is important. */ + if (zfsvfs1 > zfsvfs2) { + zfsvfs_t *tmpzfsvfs; + + tmpzfsvfs = zfsvfs2; + zfsvfs2 = zfsvfs1; + zfsvfs1 = tmpzfsvfs; + } + + error = zfs_enter(zfsvfs1, tag); + if (error != 0) + return (error); + if (zfsvfs1 != zfsvfs2) { + error = zfs_enter(zfsvfs2, tag); + if (error != 0) { + zfs_exit(zfsvfs1, tag); + return (error); + } + } + + return (0); +} + +static void +zfs_exit_two(zfsvfs_t *zfsvfs1, zfsvfs_t *zfsvfs2, const char *tag) +{ + + zfs_exit(zfsvfs1, tag); + if (zfsvfs1 != zfsvfs2) + zfs_exit(zfsvfs2, tag); +} + +/* + * We split each clone request in chunks that can fit into a single ZIL + * log entry. Each ZIL log entry can fit 130816 bytes for a block cloning + * operation (see zil_max_log_data() and zfs_log_clone_range()). This gives + * us room for storing 1022 block pointers. + * + * On success, the function return the number of bytes copied in *lenp. + * Note, it doesn't return how much bytes are left to be copied. + */ +int +zfs_clone_range(znode_t *inzp, uint64_t *inoffp, znode_t *outzp, + uint64_t *outoffp, uint64_t *lenp, cred_t *cr) +{ + zfsvfs_t *inzfsvfs, *outzfsvfs; + objset_t *inos, *outos; + zfs_locked_range_t *inlr, *outlr; + dmu_buf_impl_t *db; + dmu_tx_t *tx; + zilog_t *zilog; + uint64_t inoff, outoff, len, done; + uint64_t outsize, size; + int error; + int count = 0; + sa_bulk_attr_t bulk[3]; + uint64_t mtime[2], ctime[2]; + uint64_t uid, gid, projid; + blkptr_t *bps; + size_t maxblocks, nbps; + uint_t inblksz; + uint64_t clear_setid_bits_txg = 0; + + inoff = *inoffp; + outoff = *outoffp; + len = *lenp; + done = 0; + + inzfsvfs = ZTOZSB(inzp); + outzfsvfs = ZTOZSB(outzp); + inos = inzfsvfs->z_os; + outos = outzfsvfs->z_os; + + /* + * Both source and destination have to belong to the same storage pool. + */ + if (dmu_objset_spa(inos) != dmu_objset_spa(outos)) { + zfs_exit_two(inzfsvfs, outzfsvfs, FTAG); + return (SET_ERROR(EXDEV)); + } + + /* + * We need to call zfs_enter() potentially on two different datasets, + * so we need a dedicated function for that. + */ + error = zfs_enter_two(inzfsvfs, outzfsvfs, FTAG); + if (error != 0) + return (error); + + ASSERT(!outzfsvfs->z_replay); + + error = zfs_verify_zp(inzp); + if (error == 0) + error = zfs_verify_zp(outzp); + if (error != 0) { + zfs_exit_two(inzfsvfs, outzfsvfs, FTAG); + return (error); + } + + if (!spa_feature_is_enabled(dmu_objset_spa(outos), + SPA_FEATURE_BLOCK_CLONING)) { + zfs_exit_two(inzfsvfs, outzfsvfs, FTAG); + return (SET_ERROR(EXDEV)); + } + + /* + * We don't copy source file's flags that's why we don't allow to clone + * files that are in quarantine. + */ + if (inzp->z_pflags & ZFS_AV_QUARANTINED) { + zfs_exit_two(inzfsvfs, outzfsvfs, FTAG); + return (SET_ERROR(EACCES)); + } + + if (inoff >= inzp->z_size) { + *lenp = 0; + zfs_exit_two(inzfsvfs, outzfsvfs, FTAG); + return (0); + } + if (len > inzp->z_size - inoff) { + len = inzp->z_size - inoff; + } + if (len == 0) { + *lenp = 0; + zfs_exit_two(inzfsvfs, outzfsvfs, FTAG); + return (0); + } + + /* + * Callers might not be able to detect properly that we are read-only, + * so check it explicitly here. + */ + if (zfs_is_readonly(outzfsvfs)) { + zfs_exit_two(inzfsvfs, outzfsvfs, FTAG); + return (SET_ERROR(EROFS)); + } + + /* + * If immutable or not appending then return EPERM. + * Intentionally allow ZFS_READONLY through here. + * See zfs_zaccess_common() + */ + if ((outzp->z_pflags & ZFS_IMMUTABLE) != 0) { + zfs_exit_two(inzfsvfs, outzfsvfs, FTAG); + return (SET_ERROR(EPERM)); + } + + /* + * No overlapping if we are cloning within the same file. + */ + if (inzp == outzp) { + if (inoff < outoff + len && outoff < inoff + len) { + zfs_exit_two(inzfsvfs, outzfsvfs, FTAG); + return (SET_ERROR(EINVAL)); + } + } + + /* + * Maintain predictable lock order. + */ + if (inzp < outzp || (inzp == outzp && inoff < outoff)) { + inlr = zfs_rangelock_enter(&inzp->z_rangelock, inoff, len, + RL_READER); + outlr = zfs_rangelock_enter(&outzp->z_rangelock, outoff, len, + RL_WRITER); + } else { + outlr = zfs_rangelock_enter(&outzp->z_rangelock, outoff, len, + RL_WRITER); + inlr = zfs_rangelock_enter(&inzp->z_rangelock, inoff, len, + RL_READER); + } + + inblksz = inzp->z_blksz; + + /* + * We cannot clone into files with different block size. + */ + if (inblksz != outzp->z_blksz && outzp->z_size > inblksz) { + error = SET_ERROR(EXDEV); + goto unlock; + } + + /* + * Offsets and len must be at block boundries. + */ + if ((inoff % inblksz) != 0 || (outoff % inblksz) != 0) { + error = SET_ERROR(EXDEV); + goto unlock; + } + /* + * Length must be multipe of blksz, except for the end of the file. + */ + if ((len % inblksz) != 0 && + (len < inzp->z_size - inoff || len < outzp->z_size - outoff)) { + error = SET_ERROR(EXDEV); + goto unlock; + } + + error = zn_rlimit_fsize(outoff + len); + if (error != 0) { + goto unlock; + } + + if (inoff >= MAXOFFSET_T || outoff >= MAXOFFSET_T) { + error = SET_ERROR(EFBIG); + goto unlock; + } + + SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(outzfsvfs), NULL, + &mtime, 16); + SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(outzfsvfs), NULL, + &ctime, 16); + SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_SIZE(outzfsvfs), NULL, + &outzp->z_size, 8); + + zilog = outzfsvfs->z_log; + maxblocks = zil_max_log_data(zilog, sizeof (lr_clone_range_t)) / + sizeof (bps[0]); + + uid = KUID_TO_SUID(ZTOUID(outzp)); + gid = KGID_TO_SGID(ZTOGID(outzp)); + projid = outzp->z_projid; + + bps = kmem_alloc(sizeof (bps[0]) * maxblocks, KM_SLEEP); + + /* + * Clone the file in reasonable size chunks. Each chunk is cloned + * in a separate transaction; this keeps the intent log records small + * and allows us to do more fine-grained space accounting. + */ + while (len > 0) { + size = MIN(inblksz * maxblocks, len); + + if (zfs_id_overblockquota(outzfsvfs, DMU_USERUSED_OBJECT, + uid) || + zfs_id_overblockquota(outzfsvfs, DMU_GROUPUSED_OBJECT, + gid) || + (projid != ZFS_DEFAULT_PROJID && + zfs_id_overblockquota(outzfsvfs, DMU_PROJECTUSED_OBJECT, + projid))) { + error = SET_ERROR(EDQUOT); + break; + } + + /* + * Start a transaction. + */ + tx = dmu_tx_create(outos); + + nbps = maxblocks; + error = dmu_read_l0_bps(inos, inzp->z_id, inoff, size, tx, bps, + &nbps); + if (error != 0) { + dmu_tx_abort(tx); + /* + * If we are tyring to clone a block that was created + * in the current transaction group. Return an error, + * so the caller can fallback to just copying the data. + */ + if (error == EAGAIN) { + error = SET_ERROR(EXDEV); + } + break; + } + /* + * Encrypted data is fine as long as it comes from the same + * dataset. + * TODO: We want to extend it in the future to allow cloning to + * datasets with the same keys, like clones or to be able to + * clone a file from a snapshot of an encrypted dataset into the + * dataset itself. + */ + if (BP_IS_PROTECTED(&bps[0])) { + if (inzfsvfs != outzfsvfs) { + dmu_tx_abort(tx); + error = SET_ERROR(EXDEV); + break; + } + } + + dmu_tx_hold_sa(tx, outzp->z_sa_hdl, B_FALSE); + db = (dmu_buf_impl_t *)sa_get_db(outzp->z_sa_hdl); + DB_DNODE_ENTER(db); + dmu_tx_hold_clone_by_dnode(tx, DB_DNODE(db), outoff, size); + DB_DNODE_EXIT(db); + zfs_sa_upgrade_txholds(tx, outzp); + error = dmu_tx_assign(tx, TXG_WAIT); + if (error != 0) { + dmu_tx_abort(tx); + break; + } + + /* + * Copy source znode's block size. This only happens on the + * first iteration since zfs_rangelock_reduce() will shrink down + * lr_len to the appropriate size. + */ + if (outlr->lr_length == UINT64_MAX) { + zfs_grow_blocksize(outzp, inblksz, tx); + /* + * Round range lock up to the block boundary, so we + * prevent appends until we are done. + */ + zfs_rangelock_reduce(outlr, outoff, + ((len - 1) / inblksz + 1) * inblksz); + } + + dmu_brt_clone(outos, outzp->z_id, outoff, size, tx, bps, nbps, + B_FALSE); + + zfs_clear_setid_bits_if_necessary(outzfsvfs, outzp, cr, + &clear_setid_bits_txg, tx); + + zfs_tstamp_update_setup(outzp, CONTENT_MODIFIED, mtime, ctime); + + /* + * Update the file size (zp_size) if it has changed; + * account for possible concurrent updates. + */ + while ((outsize = outzp->z_size) < outoff + size) { + (void) atomic_cas_64(&outzp->z_size, outsize, + outoff + size); + } + + error = sa_bulk_update(outzp->z_sa_hdl, bulk, count, tx); + + zfs_log_clone_range(zilog, tx, TX_CLONE_RANGE, outzp, outoff, + size, inblksz, bps, nbps); + + dmu_tx_commit(tx); + + if (error != 0) + break; + + inoff += size; + outoff += size; + len -= size; + done += size; + } + + kmem_free(bps, sizeof (bps[0]) * maxblocks); + zfs_znode_update_vfs(outzp); + +unlock: + zfs_rangelock_exit(outlr); + zfs_rangelock_exit(inlr); + + if (done > 0) { + /* + * If we have made at least partial progress, reset the error. + */ + error = 0; + + ZFS_ACCESSTIME_STAMP(inzfsvfs, inzp); + + if (outos->os_sync == ZFS_SYNC_ALWAYS) { + zil_commit(zilog, outzp->z_id); + } + + *inoffp += done; + *outoffp += done; + *lenp = done; + } + + zfs_exit_two(inzfsvfs, outzfsvfs, FTAG); + + return (error); +} + +/* + * Usual pattern would be to call zfs_clone_range() from zfs_replay_clone(), + * but we cannot do that, because when replaying we don't have source znode + * available. This is why we need a dedicated replay function. + */ +int +zfs_clone_range_replay(znode_t *zp, uint64_t off, uint64_t len, uint64_t blksz, + const blkptr_t *bps, size_t nbps) +{ + zfsvfs_t *zfsvfs; + dmu_buf_impl_t *db; + dmu_tx_t *tx; + int error; + int count = 0; + sa_bulk_attr_t bulk[3]; + uint64_t mtime[2], ctime[2]; + + ASSERT3U(off, <, MAXOFFSET_T); + ASSERT3U(len, >, 0); + ASSERT3U(nbps, >, 0); + + zfsvfs = ZTOZSB(zp); + + ASSERT(spa_feature_is_enabled(dmu_objset_spa(zfsvfs->z_os), + SPA_FEATURE_BLOCK_CLONING)); + + if ((error = zfs_enter_verify_zp(zfsvfs, zp, FTAG)) != 0) + return (error); + + ASSERT(zfsvfs->z_replay); + ASSERT(!zfs_is_readonly(zfsvfs)); + + if ((off % blksz) != 0) { + zfs_exit(zfsvfs, FTAG); + return (SET_ERROR(EINVAL)); + } + + SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, &mtime, 16); + SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, 16); + SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_SIZE(zfsvfs), NULL, + &zp->z_size, 8); + + /* + * Start a transaction. + */ + tx = dmu_tx_create(zfsvfs->z_os); + + dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE); + db = (dmu_buf_impl_t *)sa_get_db(zp->z_sa_hdl); + DB_DNODE_ENTER(db); + dmu_tx_hold_clone_by_dnode(tx, DB_DNODE(db), off, len); + DB_DNODE_EXIT(db); + zfs_sa_upgrade_txholds(tx, zp); + error = dmu_tx_assign(tx, TXG_WAIT); + if (error != 0) { + dmu_tx_abort(tx); + zfs_exit(zfsvfs, FTAG); + return (error); + } + + if (zp->z_blksz < blksz) + zfs_grow_blocksize(zp, blksz, tx); + + dmu_brt_clone(zfsvfs->z_os, zp->z_id, off, len, tx, bps, nbps, B_TRUE); + + zfs_tstamp_update_setup(zp, CONTENT_MODIFIED, mtime, ctime); + + if (zp->z_size < off + len) + zp->z_size = off + len; + + error = sa_bulk_update(zp->z_sa_hdl, bulk, count, tx); + + /* + * zil_replaying() not only check if we are replaying ZIL, but also + * updates the ZIL header to record replay progress. + */ + VERIFY(zil_replaying(zfsvfs->z_log, tx)); + + dmu_tx_commit(tx); + + zfs_znode_update_vfs(zp); + + zfs_exit(zfsvfs, FTAG); + + return (error); +} + EXPORT_SYMBOL(zfs_access); EXPORT_SYMBOL(zfs_fsync); EXPORT_SYMBOL(zfs_holey); @@ -1002,6 +1465,8 @@ EXPORT_SYMBOL(zfs_read); EXPORT_SYMBOL(zfs_write); EXPORT_SYMBOL(zfs_getsecattr); EXPORT_SYMBOL(zfs_setsecattr); +EXPORT_SYMBOL(zfs_clone_range); +EXPORT_SYMBOL(zfs_clone_range_replay); ZFS_MODULE_PARAM(zfs_vnops, zfs_vnops_, read_chunk_size, U64, ZMOD_RW, "Bytes to read per chunk"); diff --git a/module/zfs/zil.c b/module/zfs/zil.c index fcf4e7357b00..fba1c1999612 100644 --- a/module/zfs/zil.c +++ b/module/zfs/zil.c @@ -43,6 +43,7 @@ #include #include #include +#include #include /* @@ -578,14 +579,12 @@ zil_claim_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx, } static int -zil_claim_log_record(zilog_t *zilog, const lr_t *lrc, void *tx, - uint64_t first_txg) +zil_claim_write(zilog_t *zilog, const lr_t *lrc, void *tx, uint64_t first_txg) { lr_write_t *lr = (lr_write_t *)lrc; int error; - if (lrc->lrc_txtype != TX_WRITE) - return (0); + ASSERT(lrc->lrc_txtype == TX_WRITE); /* * If the block is not readable, don't claim it. This can happen @@ -604,6 +603,57 @@ zil_claim_log_record(zilog_t *zilog, const lr_t *lrc, void *tx, return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); } +static int +zil_claim_clone_range(zilog_t *zilog, const lr_t *lrc, void *tx) +{ + const lr_clone_range_t *lr = (const lr_clone_range_t *)lrc; + const blkptr_t *bp; + spa_t *spa; + uint_t ii; + + ASSERT(lrc->lrc_txtype == TX_CLONE_RANGE); + + if (tx == NULL) { + return (0); + } + + /* + * XXX: Do we need to byteswap lr? + */ + + spa = zilog->zl_spa; + + for (ii = 0; ii < lr->lr_nbps; ii++) { + bp = &lr->lr_bps[ii]; + + /* + * When data in embedded into BP there is no need to create + * BRT entry as there is no data block. Just copy the BP as + * it contains the data. + */ + if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) { + brt_pending_add(spa, bp, tx); + } + } + + return (0); +} + +static int +zil_claim_log_record(zilog_t *zilog, const lr_t *lrc, void *tx, + uint64_t first_txg) +{ + + switch (lrc->lrc_txtype) { + case TX_WRITE: + return (zil_claim_write(zilog, lrc, tx, first_txg)); + case TX_CLONE_RANGE: + return (zil_claim_clone_range(zilog, lrc, tx)); + default: + return (0); + } +} + static int zil_free_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx, uint64_t claim_txg) @@ -616,23 +666,70 @@ zil_free_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx, } static int -zil_free_log_record(zilog_t *zilog, const lr_t *lrc, void *tx, - uint64_t claim_txg) +zil_free_write(zilog_t *zilog, const lr_t *lrc, void *tx, uint64_t claim_txg) { lr_write_t *lr = (lr_write_t *)lrc; blkptr_t *bp = &lr->lr_blkptr; + ASSERT(lrc->lrc_txtype == TX_WRITE); + /* * If we previously claimed it, we need to free it. */ - if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && - bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && - !BP_IS_HOLE(bp)) + if (bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && + !BP_IS_HOLE(bp)) { zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); + } return (0); } +static int +zil_free_clone_range(zilog_t *zilog, const lr_t *lrc, void *tx) +{ + const lr_clone_range_t *lr = (const lr_clone_range_t *)lrc; + const blkptr_t *bp; + spa_t *spa; + uint_t ii; + + ASSERT(lrc->lrc_txtype == TX_CLONE_RANGE); + + if (tx == NULL) { + return (0); + } + + spa = zilog->zl_spa; + + for (ii = 0; ii < lr->lr_nbps; ii++) { + bp = &lr->lr_bps[ii]; + + if (!BP_IS_HOLE(bp)) { + zio_free(spa, dmu_tx_get_txg(tx), bp); + } + } + + return (0); +} + +static int +zil_free_log_record(zilog_t *zilog, const lr_t *lrc, void *tx, + uint64_t claim_txg) +{ + + if (claim_txg == 0) { + return (0); + } + + switch (lrc->lrc_txtype) { + case TX_WRITE: + return (zil_free_write(zilog, lrc, tx, claim_txg)); + case TX_CLONE_RANGE: + return (zil_free_clone_range(zilog, lrc, tx)); + default: + return (0); + } +} + static int zil_lwb_vdev_compare(const void *x1, const void *x2) { @@ -1798,13 +1895,12 @@ zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb) } /* - * Maximum amount of write data that can be put into single log block. + * Maximum amount of data that can be put into single log block. */ uint64_t -zil_max_log_data(zilog_t *zilog) +zil_max_log_data(zilog_t *zilog, size_t hdrsize) { - return (zilog->zl_max_block_size - - sizeof (zil_chain_t) - sizeof (lr_write_t)); + return (zilog->zl_max_block_size - sizeof (zil_chain_t) - hdrsize); } /* @@ -1814,7 +1910,7 @@ zil_max_log_data(zilog_t *zilog) static inline uint64_t zil_max_waste_space(zilog_t *zilog) { - return (zil_max_log_data(zilog) / 8); + return (zil_max_log_data(zilog, sizeof (lr_write_t)) / 8); } /* @@ -1887,7 +1983,7 @@ zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) * For WR_NEED_COPY optimize layout for minimal number of chunks. */ lwb_sp = lwb->lwb_sz - lwb->lwb_nused; - max_log_data = zil_max_log_data(zilog); + max_log_data = zil_max_log_data(zilog, sizeof (lr_write_t)); if (reclen > lwb_sp || (reclen + dlen > lwb_sp && lwb_sp < zil_max_waste_space(zilog) && (dlen % max_log_data == 0 || diff --git a/module/zfs/zio.c b/module/zfs/zio.c index d17ee60dcde1..1b1a1831f333 100644 --- a/module/zfs/zio.c +++ b/module/zfs/zio.c @@ -41,6 +41,7 @@ #include #include #include +#include #include #include #include @@ -1176,12 +1177,14 @@ zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data, } void -zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) +zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite, + boolean_t brtwrite) { ASSERT(zio->io_type == ZIO_TYPE_WRITE); ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); ASSERT(zio->io_stage == ZIO_STAGE_OPEN); ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); + ASSERT(!brtwrite || !nopwrite); /* * We must reset the io_prop to match the values that existed @@ -1190,6 +1193,7 @@ zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) */ zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; zio->io_prop.zp_nopwrite = nopwrite; + zio->io_prop.zp_brtwrite = brtwrite; zio->io_prop.zp_copies = copies; zio->io_bp_override = bp; } @@ -1222,7 +1226,8 @@ zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) BP_GET_DEDUP(bp) || txg != spa->spa_syncing_txg || (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free && - !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))) { + !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) || + brt_maybe_exists(spa, bp)) { metaslab_check_free(spa, bp); bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); } else { @@ -1249,11 +1254,13 @@ zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, arc_freed(spa, bp); dsl_scan_freed(spa, bp); - if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) { + if (BP_IS_GANG(bp) || + BP_GET_DEDUP(bp) || + brt_maybe_exists(spa, bp)) { /* - * GANG and DEDUP blocks can induce a read (for the gang block - * header, or the DDT), so issue them asynchronously so that - * this thread is not tied up. + * GANG, DEDUP and BRT blocks can induce a read (for the gang + * block header, the DDT or the BRT), so issue them + * asynchronously so that this thread is not tied up. */ enum zio_stage stage = ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC; @@ -1594,11 +1601,15 @@ zio_write_bp_init(zio_t *zio) zio_prop_t *zp = &zio->io_prop; ASSERT(bp->blk_birth != zio->io_txg); - ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); *bp = *zio->io_bp_override; zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; + if (zp->zp_brtwrite) + return (zio); + + ASSERT(!BP_GET_DEDUP(zio->io_bp_override)); + if (BP_IS_EMBEDDED(bp)) return (zio); @@ -3042,6 +3053,35 @@ zio_nop_write(zio_t *zio) return (zio); } +/* + * ========================================================================== + * Block Reference Table + * ========================================================================== + */ +static zio_t * +zio_brt_free(zio_t *zio) +{ + blkptr_t *bp; + + bp = zio->io_bp; + + if (BP_GET_LEVEL(bp) > 0 || + BP_IS_METADATA(bp) || + !brt_maybe_exists(zio->io_spa, bp)) { + return (zio); + } + + if (!brt_entry_decref(zio->io_spa, bp)) { + /* + * This isn't the last reference, so we cannot free + * the data yet. + */ + zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; + } + + return (zio); +} + /* * ========================================================================== * Dedup @@ -4894,6 +4934,7 @@ static zio_pipe_stage_t *zio_pipeline[] = { zio_encrypt, zio_checksum_generate, zio_nop_write, + zio_brt_free, zio_ddt_read_start, zio_ddt_read_done, zio_ddt_write, diff --git a/module/zfs/zvol.c b/module/zfs/zvol.c index 1511f763fd77..06bc75c634a6 100644 --- a/module/zfs/zvol.c +++ b/module/zfs/zvol.c @@ -482,6 +482,60 @@ zvol_replay_write(void *arg1, void *arg2, boolean_t byteswap) return (error); } +/* + * Replay a TX_CLONE_RANGE ZIL transaction that didn't get committed + * after a system failure. + * + * TODO: For now we drop block cloning transations for ZVOLs as they are + * unsupported, but we still need to inform BRT about that as we + * claimed them during pool import. + * This situation can occur when we try to import a pool from a ZFS + * version supporting block cloning for ZVOLs into a system that + * has this ZFS version, that doesn't support block cloning for ZVOLs. + */ +static int +zvol_replay_clone_range(void *arg1, void *arg2, boolean_t byteswap) +{ + char name[ZFS_MAX_DATASET_NAME_LEN]; + zvol_state_t *zv = arg1; + objset_t *os = zv->zv_objset; + lr_clone_range_t *lr = arg2; + blkptr_t *bp; + dmu_tx_t *tx; + spa_t *spa; + uint_t ii; + int error; + + dmu_objset_name(os, name); + cmn_err(CE_WARN, "ZFS dropping block cloning transaction for %s.", + name); + + if (byteswap) + byteswap_uint64_array(lr, sizeof (*lr)); + + tx = dmu_tx_create(os); + error = dmu_tx_assign(tx, TXG_WAIT); + if (error) { + dmu_tx_abort(tx); + return (error); + } + + spa = os->os_spa; + + for (ii = 0; ii < lr->lr_nbps; ii++) { + bp = &lr->lr_bps[ii]; + + if (!BP_IS_HOLE(bp)) { + zio_free(spa, dmu_tx_get_txg(tx), bp); + } + } + + (void) zil_replaying(zv->zv_zilog, tx); + dmu_tx_commit(tx); + + return (0); +} + static int zvol_replay_err(void *arg1, void *arg2, boolean_t byteswap) { @@ -516,6 +570,7 @@ zil_replay_func_t *const zvol_replay_vector[TX_MAX_TYPE] = { zvol_replay_err, /* TX_SETSAXATTR */ zvol_replay_err, /* TX_RENAME_EXCHANGE */ zvol_replay_err, /* TX_RENAME_WHITEOUT */ + zvol_replay_clone_range /* TX_CLONE_RANGE */ }; /* diff --git a/tests/zfs-tests/tests/functional/cli_root/zpool_get/zpool_get.cfg b/tests/zfs-tests/tests/functional/cli_root/zpool_get/zpool_get.cfg index 99a70fa2c04d..097cd52e4777 100644 --- a/tests/zfs-tests/tests/functional/cli_root/zpool_get/zpool_get.cfg +++ b/tests/zfs-tests/tests/functional/cli_root/zpool_get/zpool_get.cfg @@ -58,6 +58,9 @@ typeset -a properties=( "multihost" "autotrim" "compatibility" + "bcloneused" + "bclonesaved" + "bcloneratio" "feature@async_destroy" "feature@empty_bpobj" "feature@lz4_compress" @@ -100,5 +103,6 @@ if is_linux || is_freebsd; then "feature@zilsaxattr" "feature@head_errlog" "feature@blake3" + "feature@block_cloning" ) fi