On Thu, Feb 11, 2021 at 04:17:31PM -0800, Darrick J. Wong wrote:
> On Thu, Feb 11, 2021 at 05:29:05PM -0600, Eric Sandeen wrote:
> > On 2/11/21 4:59 PM, Darrick J. Wong wrote:
> > > From: Darrick J. Wong <djwong@xxxxxxxxxx>
> > >
> > > Quietly set up the ability to tell xfs_repair to set NEEDSREPAIR at
> > > program start and (presumably) clear it by the end of the run. This
> > > code isn't terribly useful to users; it's mainly here so that fstests
> > > can exercise the functionality. We don't document this flag in the
> > > manual pages at all because repair clears needsrepair at exit, which
> > > means the knobs only exist for fstests to exercise the functionality.
> > >
> > > Note that we can't do any of these upgrades until we've at least done a
> > > preliminary scan of the primary super and the log.
> > >
> > > Signed-off-by: Darrick J. Wong <djwong@xxxxxxxxxx>
> > > Reviewed-by: Christoph Hellwig <hch@xxxxxx>
> > > Reviewed-by: Brian Foster <bfoster@xxxxxxxxxx>
> >
> >
> > I'm still a little on the fence about the cmdline option for crashing
> > repair at a certain point from the POV that Brian kind of pointed out
> > that this doesn't exactly scale as we need more hooks.
>
> (That's in the next patch.)
>
> > but
> >
> > ehhhh it's a test-only undocumented option and I guess we could change
> > it later if desired
> >
> > we do have other debug options on the commandline already as well....
>
> I don't mind moving the debugging hooks to be seekrit environment
> variables or something, but I don't think I've quite addressed some of
> Brian's comments from last time:
>
> [paste in stuff Brian said]
>
> > But is it worth maintaining test specific debug logic in an
> > application just to confirm that particular feature bit upgrades
> > actually set the bit?
>
> I argue that yes, this is important enough to burn a debugging knob.
> The sequence that I think we should prevent through testing is the one
> where we've set the new feature on the primary super but we haven't
> finished generating whatever new metadata is needed to complete the
> upgrade, the system crashes, and on remount the verifiers explode.
>
> Chances are pretty good that we'll get an angry bug report on the
> mailing list: "I upgraded my fs, the power went down, and the kernel
> sprayed corruption everywhere!" If we get a customer escalation like
> this, I'd /much/ rather it be about not being able to mount right after
> the reboot than a latent corruption that grows unseen until somebody's
> filesystem loses data.
>
> If a future patch to repair accidentally breaks the behavior where we
> set NEEDSREPAIR at the same time as we set the new feature and flush the
> super to disk, we cannot tell that there's been a regression in this
> safety mechanism just by looking at the output of an otherwise
> successful xfs_repair run...
>
So I think what urks me most about this is how specific it is to the
particular test. IMO, it would be _nice_ to be able to induce xfs_repair
aborts at random purely via external mechanism, but I don't view that as
a hard requirement and so don't necessarily oppose an injection
mechanism in general. I also don't think this particular mechanism is as
robust as suggested because it tests for one very particular failure
scenario (i.e. failure to set the bit) over and over. If somebody was so
misguided as to rewrite the superblock sometime later in repair without
the bit set (somehow and for who knows what reason), this test wouldn't
catch it.
Here are some handwavy random thoughts on approaches for inducing
failures that I think would be more preferable, yet wouldn't preclude
the specific test this mechanism intends to support:
- Define a custom signal handler to trigger an do_abort() and invoke it
randomly via test (or just kill -9 randomly). Con: this might require
a non-trivial test fs and some looping to provide adequate coverage.
- Rework the current hook into somewhere more generic that allows either
a random or generally more configurable trigger:
- I.e., randomly abort in the buffer I/O completion path based
on a percentage passed by the user.
- Refactor the per-phase timestamp() calls into a helper and
wire in a per-phase injection point, then let the test produce
explicit failures at the end of each phase, 1-7. This is not
quite as random, but certainly more thorough than a single
specific failure point.
These would probably still require some command line option to enable,
but it becomes less of a "test that nobody screws up these few lines of
code we just added" regression test. IMO, those tests tend to fail more
rarely than the randomized stress/failure tests that have at least some
capability to produce unforeseen failure scenarios.
> > It seems sufficient to me to test that needsrepair functionality works
> > as expected and that individual feature upgrade works as well.
>
> ...so in other words, we need some point to inject an error to make sure
> that the upgrade interlock is correct.
>
> > Given the discussion on patch 7, perhaps it makes more sense to at
> > least defer this sort of injection mechanism until we have a scheme
> > for generic needsrepair usage worked out for xfs_repair?
>
> I'm in the midst of prototyping what I said in the last thread --
> hooking the buffe cache so that repair can catch the first time we
> actually write anything to the filesystem, and using that to set
> NEEDSREPAIR. I've not run it through full fstests yet, but AFAICT I can
> keep using the same tests and the same injection knobs I already wrote.
>
> > I am wondering if there's a way to make repair fail without requiring
> > additional code, but if not and we do require some sort of injection
> > mode, I suspect we might end up better served by something more
> > generic (i.e. capable of failures at random points) rather than
> > defining a command line option specifically for a particular fstest..
>
> Probably yes, but ... uh I don't want this to drag on into building a
> generic error injection framework for userspace.
>
That's certainly fair. That's partly why I suggested to kick this can
down the road just a bit. At the same time I don't see the suggestions
above as necessarily more complex or more involved than this patch. It
may require around the same amount of code either way, just with a bit
more generic of an implementation.
Brian
> I would /really/ like to get inobtcount/bigtime tests into the kernel
> without a giant detour they have nearly zero test coverage from the
> wider community.
>
> --D
>
> >
> > > ---
> > > repair/globals.c | 2 ++
> > > repair/globals.h | 2 ++
> > > repair/phase2.c | 63 +++++++++++++++++++++++++++++++++++++++++++++++++++
> > > repair/xfs_repair.c | 9 +++++++
> > > 4 files changed, 76 insertions(+)
> > >
> > >
> > > diff --git a/repair/globals.c b/repair/globals.c
> > > index 110d98b6..699a96ee 100644
> > > --- a/repair/globals.c
> > > +++ b/repair/globals.c
> > > @@ -49,6 +49,8 @@ int rt_spec; /* Realtime dev specified as option */
> > > int convert_lazy_count; /* Convert lazy-count mode on/off */
> > > int lazy_count; /* What to set if to if converting */
> > >
> > > +bool add_needsrepair; /* forcibly set needsrepair while repairing */
> > > +
> > > /* misc status variables */
> > >
> > > int primary_sb_modified;
> > > diff --git a/repair/globals.h b/repair/globals.h
> > > index 1d397b35..043b3e8e 100644
> > > --- a/repair/globals.h
> > > +++ b/repair/globals.h
> > > @@ -90,6 +90,8 @@ extern int rt_spec; /* Realtime dev specified as option */
> > > extern int convert_lazy_count; /* Convert lazy-count mode on/off */
> > > extern int lazy_count; /* What to set if to if converting */
> > >
> > > +extern bool add_needsrepair;
> > > +
> > > /* misc status variables */
> > >
> > > extern int primary_sb_modified;
> > > diff --git a/repair/phase2.c b/repair/phase2.c
> > > index 952ac4a5..9a8d42e1 100644
> > > --- a/repair/phase2.c
> > > +++ b/repair/phase2.c
> > > @@ -131,6 +131,63 @@ zero_log(
> > > libxfs_max_lsn = log->l_last_sync_lsn;
> > > }
> > >
> > > +static bool
> > > +set_needsrepair(
> > > + struct xfs_mount *mp)
> > > +{
> > > + if (!xfs_sb_version_hascrc(&mp->m_sb)) {
> > > + printf(
> > > + _("needsrepair flag only supported on V5 filesystems.\n"));
> > > + exit(0);
> > > + }
> > > +
> > > + if (xfs_sb_version_needsrepair(&mp->m_sb)) {
> > > + printf(_("Filesystem already marked as needing repair.\n"));
> > > + exit(0);
> > > + }
> > > +
> > > + printf(_("Marking filesystem in need of repair.\n"));
> > > + mp->m_sb.sb_features_incompat |= XFS_SB_FEAT_INCOMPAT_NEEDSREPAIR;
> > > + return true;
> > > +}
> > > +
> > > +/* Perform the user's requested upgrades on filesystem. */
> > > +static void
> > > +upgrade_filesystem(
> > > + struct xfs_mount *mp)
> > > +{
> > > + struct xfs_buf *bp;
> > > + bool dirty = false;
> > > + int error;
> > > +
> > > + if (add_needsrepair)
> > > + dirty |= set_needsrepair(mp);
> > > +
> > > + if (no_modify || !dirty)
> > > + return;
> > > +
> > > + bp = libxfs_getsb(mp);
> > > + if (!bp || bp->b_error) {
> > > + do_error(
> > > + _("couldn't get superblock for feature upgrade, err=%d\n"),
> > > + bp ? bp->b_error : ENOMEM);
> > > + } else {
> > > + libxfs_sb_to_disk(bp->b_addr, &mp->m_sb);
> > > +
> > > + /*
> > > + * Write the primary super to disk immediately so that
> > > + * needsrepair will be set if repair doesn't complete.
> > > + */
> > > + error = -libxfs_bwrite(bp);
> > > + if (error)
> > > + do_error(
> > > + _("filesystem feature upgrade failed, err=%d\n"),
> > > + error);
> > > + }
> > > + if (bp)
> > > + libxfs_buf_relse(bp);
> > > +}
> > > +
> > > /*
> > > * ok, at this point, the fs is mounted but the root inode may be
> > > * trashed and the ag headers haven't been checked. So we have
> > > @@ -235,4 +292,10 @@ phase2(
> > > do_warn(_("would correct\n"));
> > > }
> > > }
> > > +
> > > + /*
> > > + * Upgrade the filesystem now that we've done a preliminary check of
> > > + * the superblocks, the AGs, the log, and the metadata inodes.
> > > + */
> > > + upgrade_filesystem(mp);
> > > }
> > > diff --git a/repair/xfs_repair.c b/repair/xfs_repair.c
> > > index 90d1a95a..a613505f 100644
> > > --- a/repair/xfs_repair.c
> > > +++ b/repair/xfs_repair.c
> > > @@ -65,11 +65,13 @@ static char *o_opts[] = {
> > > */
> > > enum c_opt_nums {
> > > CONVERT_LAZY_COUNT = 0,
> > > + CONVERT_NEEDSREPAIR,
> > > C_MAX_OPTS,
> > > };
> > >
> > > static char *c_opts[] = {
> > > [CONVERT_LAZY_COUNT] = "lazycount",
> > > + [CONVERT_NEEDSREPAIR] = "needsrepair",
> > > [C_MAX_OPTS] = NULL,
> > > };
> > >
> > > @@ -302,6 +304,13 @@ process_args(int argc, char **argv)
> > > lazy_count = (int)strtol(val, NULL, 0);
> > > convert_lazy_count = 1;
> > > break;
> > > + case CONVERT_NEEDSREPAIR:
> > > + if (!val)
> > > + do_abort(
> > > + _("-c needsrepair requires a parameter\n"));
> > > + if (strtol(val, NULL, 0) == 1)
> > > + add_needsrepair = true;
> > > + break;
> > > default:
> > > unknown('c', val);
> > > break;
> > >
>
