From: Brian Foster <bfoster@redhat.com>
To: "Darrick J. Wong" <djwong@kernel.org>
Cc: sandeen@sandeen.net, linux-xfs@vger.kernel.org
Subject: Re: [PATCH 1/2] xfs_db: support the needsrepair feature flag in the version command
Date: Wed, 3 Feb 2021 14:09:38 -0500 [thread overview]
Message-ID: <20210203190938.GJ3647012@bfoster> (raw)
In-Reply-To: <20210203183057.GB7190@magnolia>
On Wed, Feb 03, 2021 at 10:30:57AM -0800, Darrick J. Wong wrote:
> On Wed, Feb 03, 2021 at 08:07:21AM -0500, Brian Foster wrote:
> > On Tue, Feb 02, 2021 at 01:09:15PM -0800, Darrick J. Wong wrote:
> > > On Tue, Jan 19, 2021 at 09:37:14AM -0500, Brian Foster wrote:
> > > > On Fri, Jan 15, 2021 at 05:24:47PM -0800, Darrick J. Wong wrote:
> > > > > From: Darrick J. Wong <djwong@kernel.org>
> > > > >
> > > > > Teach the xfs_db version command about the 'needsrepair' flag, which can
> > > > > be used to force the system administrator to repair the filesystem with
> > > > > xfs_repair.
> > > > >
> > > > > Signed-off-by: Darrick J. Wong <djwong@kernel.org>
> > > > > ---
> > > > > db/check.c | 5 ++
> > > > > db/sb.c | 153 ++++++++++++++++++++++++++++++++++++++++++++++++--
> > > > > db/xfs_admin.sh | 10 ++-
> > > > > man/man8/xfs_admin.8 | 15 +++++
> > > > > man/man8/xfs_db.8 | 5 ++
> > > > > 5 files changed, 178 insertions(+), 10 deletions(-)
> > > > >
> > > > >
> > ...
> > > > > diff --git a/db/sb.c b/db/sb.c
> > > > > index d09f653d..fcc2a0ed 100644
> > > > > --- a/db/sb.c
> > > > > +++ b/db/sb.c
> > ...
> > > > > @@ -717,8 +836,23 @@ version_f(
> > > > > return 0;
> > > > > }
> > > > >
> > > > > + if (xfs_sb_version_needsrepair(&mp->m_sb)) {
> > > > > + dbprintf(_("%s: filesystem needs xfs_repair\n"),
> > > > > + progname);
> > > > > + return 0;
> > > > > + }
> > > > > +
> > > > > /* Logic here derived from the IRIX xfs_chver(1M) script. */
> > > > > - if (!strcasecmp(argv[1], "extflg")) {
> > > > > + if (!strcasecmp(argv[1], "needsrepair")) {
> > > > > + if (!xfs_sb_version_hascrc(&mp->m_sb)) {
> > > > > + dbprintf(
> > > > > + _("needsrepair flag cannot be enabled on pre-V5 filesystems\n"));
> > > > > + exitcode = 2;
> > > > > + return 1;
> > > >
> > > > Hmm.. I see that exitcode 1 means error && xfs_repair while exitcode 2
> > > > means error && !xfs_repair, but I'm still not sure I follow the high
> > > > level error semantics, particularly if we happen to fail updating
> > > > secondary supers. I wonder if it would be more straightforward to have
> > > > xfs_db only return an error when an update attempt occurs and fails and
> > > > then let xfs_admin run xfs_repair if status == 0 && NEEDSREPAIR is set.
> > >
> > > Hm. I'd be even more tempted to make it run if xfs_db just failed.
> > >
> > > > I suppose the various other ".. bit already set" or "v5 super required"
> > > > conditions don't really need to be errors and thus repair would only run
> > > > in those cases if NEEDSREPAIR was still set on the fs. Otherwise if
> > > > xfs_db fails we dump an error message and encourage the user to run
> > > > xfs_repair themselves.
> > >
> > > Yeah, that does seem more reasonable. I'll change xfs_admin to force a
> > > run through repair if the NEEDSREPAIR feature is set or if the xfs_db
> > > command failed, since that probably means something's wrong with the fs.
> > >
> >
> > Perhaps, or the storage if an I/O happens to fail or something. I'm not
> > sure we should go that route where if the version upgrade happens to
> > fail we do a "well this operation failed, but something is probably
> > wrong so let me try and repair that for you." I'd personally be kind of
> > annoyed by that if I didn't have an opportunity to analyze the problem
> > before making the decision to run a (potentially destructive) repair
> > operation. I agree it's an unlikely situation, but IMO the automatically
> > invoked repair should be isolated to the specific case that warrants it
> > and everything else should probably just bail out.
>
> Hm, good point, considering how long it can take to run repair. I'm
> sure there's some sysadmin out there who would actually prefer to nuke
> the whole fs/node if the upgrade fails.
>
I don't know if there are any likely scenarios where a user might throw
a feature upgrade attempt at a borked fs only to find it falling into
repair for some other reason, but I'd rather not hear about those cases
from a support perspective, whether it be complaints about feature
upgrade taking forever, confusion over why a feature upgrade failed but
then spit out some other repair related errors, losing the ability to
gather repair -n output, a metadump, etc. We do rarely have cases where
users are surprised a repair throws out a bunch of directory entries or
makes other big and unexpected changes to the fs because they don't
quite understand that repair != data recovery. With those types of
support cases in mind, I just think less can go wrong or down an
unexpected path if the upgrade repair logic is very explicitly tied to
feature bit upgrade success && needsrepair.
> > The user really doesn't need to know or care that a repair is involved
> > in the first place, so ISTM that either "upgrade operation succeeded" or
> > "upgrade operation failed, fs has problems" is fairly expected failure
> > handling behavior (as opposed to allowing things like "upgrade operation
> > failed, we ran repair anyways and fixed some other stuff, maybe try that
> > upgrade again since the last one basically just invoked xfs_repair?").
> > If we really wanted to be careful, we could even run a repair -n first
> > and require it succeed before attempting to touch the superblock. That
> > might be annoying in cases where repair takes forever, but at least if
> > the user bails on it it would likely be before we've modified anything.
>
> ...or make it more likely that the user ^Cs and never uses our
> scurrilous upgrader tool. :)
>
Heh. :P
> I dunno, I've a slight preference for /knowing/ that the fs isn't a
> crazy tangle of crap before we try to upgrade it. Though as I've
> mentioned before, resize2fs has weird heuristics to guesstimate if a
> filesystem is clean "enough", and ext4 even tracks the last fsck and
> mount times. I don't really want to go down that path.
>
> Perhaps we should simply document xfs_repair -n as a preparation step?
>
Yeah, it's probably a good idea to put that in writing. I don't have a
strong opinion on doing the pre-update repair thing. I was just throwing
it out there as a thought. Users may not be aware that a repair is part
of the normal procedure in the first place, so documenting this might be
helpful for those responsible enough to follow best practices.
Brian
> --D
>
> >
> > Brian
> >
> > > > There are still corner cases I guess, but that does _seem_ a bit more
> > > > elegant to me. Otherwise I suppose a comment somewhere that explains
> > > > when/why to use which error code would be helpful.
> > >
> > > <nod> I'll put them in.
> > >
> > > --D
> > >
> > > > Brian
> > > >
> > > > > + }
> > > > > +
> > > > > + v5features.incompat |= XFS_SB_FEAT_INCOMPAT_NEEDSREPAIR;
> > > > > + } else if (!strcasecmp(argv[1], "extflg")) {
> > > > > switch (XFS_SB_VERSION_NUM(&mp->m_sb)) {
> > > > > case XFS_SB_VERSION_1:
> > > > > version = 0x0004 | XFS_SB_VERSION_EXTFLGBIT;
> > > > > @@ -809,6 +943,11 @@ version_f(
> > > > > mp->m_sb.sb_versionnum = version;
> > > > > mp->m_sb.sb_features2 = features;
> > > > > }
> > > > > +
> > > > > + if (!upgrade_v5_features(mp, &v5features)) {
> > > > > + exitcode = 1;
> > > > > + return 1;
> > > > > + }
> > > > > }
> > > > >
> > > > > if (argc == 3) { /* VERSIONNUM + FEATURES2 */
> > > > > diff --git a/db/xfs_admin.sh b/db/xfs_admin.sh
> > > > > index bd325da2..0e79bbf9 100755
> > > > > --- a/db/xfs_admin.sh
> > > > > +++ b/db/xfs_admin.sh
> > > > > @@ -7,9 +7,9 @@
> > > > > status=0
> > > > > DB_OPTS=""
> > > > > REPAIR_OPTS=""
> > > > > -USAGE="Usage: xfs_admin [-efjlpuV] [-c 0|1] [-L label] [-U uuid] device [logdev]"
> > > > > +USAGE="Usage: xfs_admin [-efjlpuV] [-c 0|1] [-L label] [-U uuid] [-O v5_feature] device [logdev]"
> > > > >
> > > > > -while getopts "efjlpuc:L:U:V" c
> > > > > +while getopts "efjlpuc:L:O:U:V" c
> > > > > do
> > > > > case $c in
> > > > > c) REPAIR_OPTS=$REPAIR_OPTS" -c lazycount="$OPTARG;;
> > > > > @@ -19,6 +19,9 @@ do
> > > > > l) DB_OPTS=$DB_OPTS" -r -c label";;
> > > > > L) DB_OPTS=$DB_OPTS" -c 'label "$OPTARG"'";;
> > > > > p) DB_OPTS=$DB_OPTS" -c 'version projid32bit'";;
> > > > > + O) DB_OPTS=$DB_OPTS" -c 'version "$OPTARG"'";
> > > > > + # Force repair to run by adding a single space to REPAIR_OPTS
> > > > > + REPAIR_OPTS="$REPAIR_OPTS ";;
> > > > > u) DB_OPTS=$DB_OPTS" -r -c uuid";;
> > > > > U) DB_OPTS=$DB_OPTS" -c 'uuid "$OPTARG"'";;
> > > > > V) xfs_db -p xfs_admin -V
> > > > > @@ -34,6 +37,7 @@ set -- extra $@
> > > > > shift $OPTIND
> > > > > case $# in
> > > > > 1|2)
> > > > > + status=0
> > > > > # Pick up the log device, if present
> > > > > if [ -n "$2" ]; then
> > > > > DB_OPTS=$DB_OPTS" -l '$2'"
> > > > > @@ -46,7 +50,7 @@ case $# in
> > > > > eval xfs_db -x -p xfs_admin $DB_OPTS $1
> > > > > status=$?
> > > > > fi
> > > > > - if [ -n "$REPAIR_OPTS" ]
> > > > > + if [ -n "$REPAIR_OPTS" ] && [ $status -ne 2 ]
> > > > > then
> > > > > # Hide normal repair output which is sent to stderr
> > > > > # assuming the filesystem is fine when a user is
> > > > > diff --git a/man/man8/xfs_admin.8 b/man/man8/xfs_admin.8
> > > > > index 8afc873f..b423981d 100644
> > > > > --- a/man/man8/xfs_admin.8
> > > > > +++ b/man/man8/xfs_admin.8
> > > > > @@ -6,6 +6,8 @@ xfs_admin \- change parameters of an XFS filesystem
> > > > > [
> > > > > .B \-eflpu
> > > > > ] [
> > > > > +.BI \-O " feature"
> > > > > +] [
> > > > > .BR "\-c 0" | 1
> > > > > ] [
> > > > > .B \-L
> > > > > @@ -103,6 +105,19 @@ The filesystem label can be cleared using the special "\c
> > > > > " value for
> > > > > .IR label .
> > > > > .TP
> > > > > +.BI \-O " feature"
> > > > > +Add a new feature to the filesystem.
> > > > > +Only one feature can be specified at a time.
> > > > > +Features are as follows:
> > > > > +.RS 0.7i
> > > > > +.TP
> > > > > +.B needsrepair
> > > > > +If this is a V5 filesystem, flag the filesystem as needing repairs.
> > > > > +Until
> > > > > +.BR xfs_repair (8)
> > > > > +is run, the filesystem will not be mountable.
> > > > > +.RE
> > > > > +.TP
> > > > > .BI \-U " uuid"
> > > > > Set the UUID of the filesystem to
> > > > > .IR uuid .
> > > > > diff --git a/man/man8/xfs_db.8 b/man/man8/xfs_db.8
> > > > > index 58727495..7331cf19 100644
> > > > > --- a/man/man8/xfs_db.8
> > > > > +++ b/man/man8/xfs_db.8
> > > > > @@ -971,6 +971,11 @@ may toggle between
> > > > > and
> > > > > .B attr2
> > > > > at will (older kernels may not support the newer version).
> > > > > +The filesystem can be flagged as requiring a run through
> > > > > +.BR xfs_repair (8)
> > > > > +if the
> > > > > +.B needsrepair
> > > > > +option is specified and the filesystem is formatted with the V5 format.
> > > > > .IP
> > > > > If no argument is given, the current version and feature bits are printed.
> > > > > With one argument, this command will write the updated version number
> > > > >
> > > >
> > >
> >
>
next prev parent reply other threads:[~2021-02-03 19:11 UTC|newest]
Thread overview: 18+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-01-16 1:24 [PATCHSET v2 0/2] xfs: add the ability to flag a fs for repair Darrick J. Wong
2021-01-16 1:24 ` [PATCH 1/2] xfs_db: support the needsrepair feature flag in the version command Darrick J. Wong
2021-01-19 14:37 ` Brian Foster
2021-02-02 21:09 ` Darrick J. Wong
2021-02-03 13:07 ` Brian Foster
2021-02-03 18:30 ` Darrick J. Wong
2021-02-03 19:09 ` Brian Foster [this message]
2021-01-16 1:24 ` [PATCH 2/2] xfs_repair: clear the needsrepair flag Darrick J. Wong
2021-01-19 14:37 ` Brian Foster
2021-01-19 18:15 ` Darrick J. Wong
2021-01-19 19:44 ` Brian Foster
2021-01-19 20:31 ` Darrick J. Wong
2021-01-19 23:49 ` Darrick J. Wong
2021-01-20 4:31 ` [PATCH v2.1 " Darrick J. Wong
2021-01-20 17:38 ` Brian Foster
2021-02-02 22:22 ` Darrick J. Wong
2021-01-20 4:31 ` [PATCH 3/2] xfs_repair: fix unmount error message to have a newline Darrick J. Wong
2021-01-20 17:38 ` Brian Foster
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20210203190938.GJ3647012@bfoster \
--to=bfoster@redhat.com \
--cc=djwong@kernel.org \
--cc=linux-xfs@vger.kernel.org \
--cc=sandeen@sandeen.net \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).