From 84df75c732b11127f1888e3578a740fe56d2fb62 Mon Sep 17 00:00:00 2001 From: Alexander Motin Date: Mon, 3 Jun 2019 19:20:10 +0000 Subject: [PATCH 1/2] 9880 Race in ZFS parallel mount illumos/illumos-gate@bc4c0ff1343a311cc24933908ac6c4455af09031 Reviewed by: Jason King Reviewed by: Sebastien Roy Approved by: Joshua M. Clulow Author: Andy Fiddaman --- lib/libzfs/common/libzfs_mount.c | 34 +++++++++++++++++++++++++++++--- 1 file changed, 31 insertions(+), 3 deletions(-) diff --git a/lib/libzfs/common/libzfs_mount.c b/lib/libzfs/common/libzfs_mount.c index cf15735f3f9c..9bbf4d22338c 100644 --- a/lib/libzfs/common/libzfs_mount.c +++ b/lib/libzfs/common/libzfs_mount.c @@ -26,6 +26,7 @@ * Copyright 2016 Igor Kozhukhov * Copyright 2017 Joyent, Inc. * Copyright 2017 RackTop Systems. + * Copyright 2018 OmniOS Community Edition (OmniOSce) Association. */ /* @@ -1142,19 +1143,28 @@ zfs_iter_cb(zfs_handle_t *zhp, void *data) /* * Sort comparator that compares two mountpoint paths. We sort these paths so * that subdirectories immediately follow their parents. This means that we - * effectively treat the '/' character as the lowest value non-nul char. An - * example sorted list using this comparator would look like: + * effectively treat the '/' character as the lowest value non-nul char. + * Since filesystems from non-global zones can have the same mountpoint + * as other filesystems, the comparator sorts global zone filesystems to + * the top of the list. This means that the global zone will traverse the + * filesystem list in the correct order and can stop when it sees the + * first zoned filesystem. In a non-global zone, only the delegated + * filesystems are seen. + * + * An example sorted list using this comparator would look like: * * /foo * /foo/bar * /foo/bar/baz * /foo/baz * /foo.bar + * /foo (NGZ1) + * /foo (NGZ2) * * The mounting code depends on this ordering to deterministically iterate * over filesystems in order to spawn parallel mount tasks. */ -int +static int mountpoint_cmp(const void *arga, const void *argb) { zfs_handle_t *const *zap = arga; @@ -1166,6 +1176,14 @@ mountpoint_cmp(const void *arga, const void *argb) const char *a = mounta; const char *b = mountb; boolean_t gota, gotb; + uint64_t zoneda, zonedb; + + zoneda = zfs_prop_get_int(za, ZFS_PROP_ZONED); + zonedb = zfs_prop_get_int(zb, ZFS_PROP_ZONED); + if (zoneda && !zonedb) + return (1); + if (!zoneda && zonedb) + return (-1); gota = (zfs_get_type(za) == ZFS_TYPE_FILESYSTEM); if (gota) { @@ -1379,6 +1397,8 @@ void zfs_foreach_mountpoint(libzfs_handle_t *hdl, zfs_handle_t **handles, size_t num_handles, zfs_iter_f func, void *data, boolean_t parallel) { + zoneid_t zoneid = getzoneid(); + /* * The ZFS_SERIAL_MOUNT environment variable is an undocumented * variable that can be used as a convenience to do a/b comparison @@ -1414,6 +1434,14 @@ zfs_foreach_mountpoint(libzfs_handle_t *hdl, zfs_handle_t **handles, */ for (int i = 0; i < num_handles; i = non_descendant_idx(handles, num_handles, i)) { + /* + * Since the mountpoints have been sorted so that the zoned + * filesystems are at the end, a zoned filesystem seen from + * the global zone means that we're done. + */ + if (zoneid == GLOBAL_ZONEID && + zfs_prop_get_int(handles[i], ZFS_PROP_ZONED)) + break; zfs_dispatch_mount(hdl, handles, num_handles, i, func, data, tq); } From 5f8c7a6af38b13f38fc8cdce3dfe30e6ae757bb9 Mon Sep 17 00:00:00 2001 From: Alexander Motin Date: Mon, 3 Jun 2019 19:24:40 +0000 Subject: [PATCH 2/2] 9993 zil writes can get delayed in zio pipeline illumos/illumos-gate@2258ad0b755b24a55c6173b1e6bb6188389f72dd Reviewed by: Prakash Surya Reviewed by: Brad Lewis Reviewed by: Matt Ahrens Approved by: Dan McDonald Author: George Wilson --- uts/common/fs/zfs/zio.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/uts/common/fs/zfs/zio.c b/uts/common/fs/zfs/zio.c index b7e11f922a95..bd14ca38f2e8 100644 --- a/uts/common/fs/zfs/zio.c +++ b/uts/common/fs/zfs/zio.c @@ -1497,7 +1497,8 @@ zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) * If this is a high priority I/O, then use the high priority taskq if * available. */ - if (zio->io_priority == ZIO_PRIORITY_NOW && + if ((zio->io_priority == ZIO_PRIORITY_NOW || + zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) && spa->spa_zio_taskq[t][q + 1].stqs_count != 0) q++;