commit 443d4d30d8700963f53180650c533b3334ccf1f1 Author: YAMAMOTO Takashi Date: Wed Mar 4 15:38:03 2026 +0000 zfs: port a fix for data corruption issue from illumos see https://www.illumos.org/issues/17734 for the details. note: i didn't bother to patch illumos/freebsd code in our tree. the original commit message: > commit f6559a18843abdfa5849b9e74f239f9bd15796d3 > Author: Andy Fiddaman > Date: Mon Nov 10 22:52:05 2025 +0000 > > 17734 ZFS fsync can trigger ZIL transaction reordering and data corruption > Portions contributed by: Alexander Motin > Reviewed by: Ryan Zezeski > Reviewed by: Toomas Soome > Approved by: Dan McDonald diff --git a/external/cddl/osnet/dist/uts/common/fs/zfs/sys/zfs_vfsops.h b/external/cddl/osnet/dist/uts/common/fs/zfs/sys/zfs_vfsops.h index 57cb61246c3a..5a324deb9821 100644 --- a/external/cddl/osnet/dist/uts/common/fs/zfs/sys/zfs_vfsops.h +++ b/external/cddl/osnet/dist/uts/common/fs/zfs/sys/zfs_vfsops.h @@ -138,7 +138,6 @@ typedef struct zfid_long { #define SHORT_FID_LEN (sizeof (zfid_short_t) - sizeof (uint16_t)) #define LONG_FID_LEN (sizeof (zfid_long_t) - sizeof (uint16_t)) -extern uint_t zfs_fsyncer_key; extern int zfs_super_owner; extern int zfs_suspend_fs(zfsvfs_t *zfsvfs); diff --git a/external/cddl/osnet/dist/uts/common/fs/zfs/sys/zfs_znode.h b/external/cddl/osnet/dist/uts/common/fs/zfs/sys/zfs_znode.h index 128309d233cc..58018728ce2d 100644 --- a/external/cddl/osnet/dist/uts/common/fs/zfs/sys/zfs_znode.h +++ b/external/cddl/osnet/dist/uts/common/fs/zfs/sys/zfs_znode.h @@ -364,7 +364,7 @@ extern void zfs_log_symlink(zilog_t *zilog, dmu_tx_t *tx, uint64_t txtype, extern void zfs_log_rename(zilog_t *zilog, dmu_tx_t *tx, uint64_t txtype, znode_t *sdzp, char *sname, znode_t *tdzp, char *dname, znode_t *szp); extern void zfs_log_write(zilog_t *zilog, dmu_tx_t *tx, int txtype, - znode_t *zp, offset_t off, ssize_t len, int ioflag); + znode_t *zp, offset_t off, ssize_t len, boolean_t commit); extern void zfs_log_truncate(zilog_t *zilog, dmu_tx_t *tx, int txtype, znode_t *zp, uint64_t off, uint64_t len); extern void zfs_log_setattr(zilog_t *zilog, dmu_tx_t *tx, int txtype, diff --git a/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_ioctl.c b/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_ioctl.c index 786be8420436..67082015c771 100644 --- a/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_ioctl.c +++ b/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_ioctl.c @@ -221,7 +221,6 @@ dev_info_t *zfs_dip = &__zfs_devinfo; #define vfs_rel(x) vfs_rele(x) #endif -uint_t zfs_fsyncer_key; extern uint_t rrw_tsd_key; static uint_t zfs_allow_log_key; extern uint_t zfs_geom_probe_vdev_key; @@ -6769,7 +6768,6 @@ _init(void) return (error); } - tsd_create(&zfs_fsyncer_key, NULL); tsd_create(&zfs_putpages_key, NULL); tsd_create(&rrw_tsd_key, rrw_tsd_destroy); tsd_create(&zfs_allow_log_key, zfs_allow_log_destroy); @@ -6802,7 +6800,6 @@ _fini(void) if (zfs_nfsshare_inited || zfs_smbshare_inited) (void) ddi_modclose(sharefs_mod); - tsd_destroy(&zfs_fsyncer_key); ldi_ident_release(zfs_li); zfs_li = NULL; mutex_destroy(&zfs_share_lock); @@ -6869,7 +6866,6 @@ zfs__init(void) zvol_init(); zfs_ioctl_init(); - tsd_create(&zfs_fsyncer_key, NULL); tsd_create(&rrw_tsd_key, rrw_tsd_destroy); tsd_create(&zfs_allow_log_key, zfs_allow_log_destroy); tsd_create(&zfs_geom_probe_vdev_key, NULL); @@ -6895,7 +6891,6 @@ zfs__fini(void) zfs_fini(); spa_fini(); - tsd_destroy(&zfs_fsyncer_key); tsd_destroy(&rrw_tsd_key); tsd_destroy(&zfs_allow_log_key); @@ -7180,7 +7175,6 @@ zfs_modcmd(modcmd_t cmd, void *arg) mutex_init(&zfs_share_lock, NULL, MUTEX_DEFAULT, NULL); mutex_init(&zfs_debug_mtx, NULL, MUTEX_DEFAULT, NULL); - tsd_create(&zfs_fsyncer_key, NULL); tsd_create(&rrw_tsd_key, rrw_tsd_destroy); tsd_create(&zfs_allow_log_key, zfs_allow_log_destroy); tsd_create(&zfs_putpage_key, NULL); @@ -7217,7 +7211,6 @@ attacherr: spa_fini(); tsd_destroy(&zfs_putpage_key); - tsd_destroy(&zfs_fsyncer_key); tsd_destroy(&rrw_tsd_key); tsd_destroy(&zfs_allow_log_key); diff --git a/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_log.c b/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_log.c index 31054bac5998..191a9d62b207 100644 --- a/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_log.c +++ b/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_log.c @@ -462,11 +462,10 @@ SYSCTL_LONG(_vfs_zfs, OID_AUTO, immediate_write_sz, CTLFLAG_RWTUN, void zfs_log_write(zilog_t *zilog, dmu_tx_t *tx, int txtype, - znode_t *zp, offset_t off, ssize_t resid, int ioflag) + znode_t *zp, offset_t off, ssize_t resid, boolean_t commit) { uint32_t blocksize = zp->z_blksz; itx_wr_state_t write_state; - uintptr_t fsync_cnt; if (zil_replaying(zilog, tx) || zp->z_unlinked) return; @@ -476,15 +475,11 @@ zfs_log_write(zilog_t *zilog, dmu_tx_t *tx, int txtype, else if (!spa_has_slogs(zilog->zl_spa) && resid >= zfs_immediate_write_sz) write_state = WR_INDIRECT; - else if (ioflag & (FSYNC | FDSYNC)) + else if (commit) write_state = WR_COPIED; else write_state = WR_NEED_COPY; - if ((fsync_cnt = (uintptr_t)tsd_get(zfs_fsyncer_key)) != 0) { - (void) tsd_set(zfs_fsyncer_key, (void *)(fsync_cnt - 1)); - } - while (resid) { itx_t *itx; lr_write_t *lr; @@ -515,10 +510,7 @@ zfs_log_write(zilog_t *zilog, dmu_tx_t *tx, int txtype, BP_ZERO(&lr->lr_blkptr); itx->itx_private = zp->z_zfsvfs; - - if (!(ioflag & (FSYNC | FDSYNC)) && (zp->z_sync_cnt == 0) && - (fsync_cnt == 0)) - itx->itx_sync = B_FALSE; + itx->itx_sync = (zp->z_sync_cnt != 0); zil_itx_assign(zilog, itx, tx); diff --git a/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_vnops.c b/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_vnops.c index d239a92afcf5..1e1af6baa64e 100644 --- a/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_vnops.c +++ b/external/cddl/osnet/dist/uts/common/fs/zfs/zfs_vnops.c @@ -218,9 +218,15 @@ zfs_open(vnode_t **vpp, int flag, cred_t *cr, caller_context_t *ct) } } - /* Keep a count of the synchronous opens in the znode */ - if (flag & (FSYNC | FDSYNC)) - atomic_inc_32(&zp->z_sync_cnt); + /* + * Keep a count of the synchronous opens in the znode. On first + * synchronous open we must convert all previous async transactions + * into sync to keep correct ordering. + */ + if (flag & (FSYNC | FDSYNC)) { + if (atomic_inc_32_nv(&zp->z_sync_cnt) == 1) + zil_async_to_sync(zfsvfs->z_log, zp->z_id); + } ZFS_EXIT(zfsvfs); return (0); @@ -1044,6 +1050,7 @@ zfs_write(vnode_t *vp, uio_t *uio, int ioflag, cred_t *cr, caller_context_t *ct) sa_bulk_attr_t bulk[4]; uint64_t mtime[2], ctime[2]; int segflg; + boolean_t commit; #ifdef __NetBSD__ segflg = VMSPACE_IS_KERNEL_P(uio->uio_vmspace) ? @@ -1181,6 +1188,9 @@ zfs_write(vnode_t *vp, uio_t *uio, int ioflag, cred_t *cr, caller_context_t *ct) end_size = MAX(zp->z_size, woff + n); + commit = ((ioflag & (FSYNC | FDSYNC)) != 0 || + zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS); + /* * Write the file in reasonable size chunks. Each chunk is written * in a separate transaction; this keeps the intent log records small @@ -1403,7 +1413,7 @@ zfs_write(vnode_t *vp, uio_t *uio, int ioflag, cred_t *cr, caller_context_t *ct) else (void) sa_bulk_update(zp->z_sa_hdl, bulk, count, tx); - zfs_log_write(zilog, tx, TX_WRITE, zp, woff, tx_bytes, ioflag); + zfs_log_write(zilog, tx, TX_WRITE, zp, woff, tx_bytes, commit); dmu_tx_commit(tx); if (error != 0) @@ -1439,8 +1449,7 @@ zfs_write(vnode_t *vp, uio_t *uio, int ioflag, cred_t *cr, caller_context_t *ct) } #endif - if (ioflag & (FSYNC | FDSYNC) || - zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) + if (commit) zil_commit(zilog, zp->z_id); ZFS_EXIT(zfsvfs); @@ -3009,16 +3018,12 @@ update: return (error); } -ulong_t zfs_fsync_sync_cnt = 4; - static int zfs_fsync(vnode_t *vp, int syncflag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; - (void) tsd_set(zfs_fsyncer_key, (void *)zfs_fsync_sync_cnt); - if (zfsvfs->z_os->os_sync != ZFS_SYNC_DISABLED) { ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); @@ -6213,7 +6218,8 @@ zfs_putapage(vnode_t *vp, page_t **pp, int count, int flags) B_TRUE); err = sa_bulk_update(zp->z_sa_hdl, bulk, count, tx); ASSERT0(err); - zfs_log_write(zfsvfs->z_log, tx, TX_WRITE, zp, off, len, 0); + zfs_log_write(zfsvfs->z_log, tx, TX_WRITE, zp, off, len, + B_FALSE); } dmu_tx_commit(tx); diff --git a/external/cddl/osnet/dist/uts/common/fs/zfs/zvol.c b/external/cddl/osnet/dist/uts/common/fs/zfs/zvol.c index 93ed199585fc..27f6e7dd76aa 100644 --- a/external/cddl/osnet/dist/uts/common/fs/zfs/zvol.c +++ b/external/cddl/osnet/dist/uts/common/fs/zfs/zvol.c @@ -1604,7 +1604,7 @@ SYSCTL_LONG(_vfs_zfs_vol, OID_AUTO, immediate_write_sz, CTLFLAG_RWTUN, static void zvol_log_write(zvol_state_t *zv, dmu_tx_t *tx, offset_t off, ssize_t resid, - boolean_t sync) + boolean_t commit) { uint32_t blocksize = zv->zv_volblocksize; zilog_t *zilog = zv->zv_zilog; @@ -1613,15 +1613,16 @@ zvol_log_write(zvol_state_t *zv, dmu_tx_t *tx, offset_t off, ssize_t resid, if (zil_replaying(zilog, tx)) return; - if (zilog->zl_logbias == ZFS_LOGBIAS_THROUGHPUT) + if (zilog->zl_logbias == ZFS_LOGBIAS_THROUGHPUT) { write_state = WR_INDIRECT; - else if (!spa_has_slogs(zilog->zl_spa) && - resid >= blocksize && blocksize > zvol_immediate_write_sz) + } else if (!spa_has_slogs(zilog->zl_spa) && + resid >= blocksize && blocksize > zvol_immediate_write_sz) { write_state = WR_INDIRECT; - else if (sync) + } else if (commit) { write_state = WR_COPIED; - else + } else { write_state = WR_NEED_COPY; + } while (resid) { itx_t *itx; @@ -1654,9 +1655,6 @@ zvol_log_write(zvol_state_t *zv, dmu_tx_t *tx, offset_t off, ssize_t resid, itx->itx_private = zv; - if (!sync && (zv->zv_sync_cnt == 0)) - itx->itx_sync = B_FALSE; - zil_itx_assign(zilog, itx, tx); off += len; @@ -1792,7 +1790,7 @@ zvol_strategy(buf_t *bp) boolean_t doread = 0; #endif boolean_t is_dumpified; - boolean_t sync; + boolean_t commit; #ifdef illumos zfs_soft_state_t *zs = NULL; @@ -1929,9 +1927,9 @@ zvol_strategy(buf_t *bp) } is_dumpified = B_FALSE; - sync = ((!(bp->b_flags & B_ASYNC) && + commit = ((!(bp->b_flags & B_ASYNC) && !(zv->zv_flags & ZVOL_WCE)) || - (zv->zv_objset->os_sync == ZFS_SYNC_ALWAYS)) && + zv->zv_objset->os_sync == ZFS_SYNC_ALWAYS) && !doread && !is_dumpified; mutex_enter(&zv->zv_dklock); @@ -1979,7 +1977,7 @@ zvol_strategy(buf_t *bp) dmu_tx_abort(tx); } else { dmu_write(os, ZVOL_OBJ, off, size, addr, tx); - zvol_log_write(zv, tx, off, size, sync); + zvol_log_write(zv, tx, off, size, commit); dmu_tx_commit(tx); } } @@ -2027,7 +2025,7 @@ out: if ((bp->b_resid = resid) == bp->b_bcount) bioerror(bp, off > volsize ? EINVAL : error); - if (sync) + if (commit) zil_commit(zv->zv_zilog, ZVOL_OBJ); mutex_enter(&zv->zv_dklock); disk_unbusy(&zv->zv_dk, bp->b_bcount - bp->b_resid, doread); @@ -2174,7 +2172,7 @@ zvol_write(struct cdev *dev, struct uio *uio, int ioflag) uint64_t volsize; rl_t *rl; int error = 0; - boolean_t sync; + boolean_t commit; #if defined(illumos) || defined(__NetBSD__) minor_t minor = getminor(dev); @@ -2205,7 +2203,7 @@ zvol_write(struct cdev *dev, struct uio *uio, int ioflag) sync = (ioflag & IO_SYNC) || #endif #ifdef __NetBSD__ - sync = 1 || + commit = 1 || #endif (zv->zv_objset->os_sync == ZFS_SYNC_ALWAYS); @@ -2233,14 +2231,14 @@ zvol_write(struct cdev *dev, struct uio *uio, int ioflag) } error = dmu_write_uio_dbuf(zv->zv_dbuf, uio, bytes, tx); if (error == 0) - zvol_log_write(zv, tx, off, bytes, sync); + zvol_log_write(zv, tx, off, bytes, commit); dmu_tx_commit(tx); if (error) break; } zfs_range_unlock(rl); - if (sync) + if (commit) zil_commit(zv->zv_zilog, ZVOL_OBJ); #ifdef __NetBSD__ mutex_enter(&zv->zv_dklock);