revlog: add a mechanism to verify expected file position before appending
If someone uses `hg debuglocks`, or some non-hg process writes to the .hg
directory without respecting the locks, or if the repo's on a networked
filesystem, it's possible for the revlog code to write out corrupted data.
The form of this corruption can vary depending on what data was written and how
that happened. We are in the "networked filesystem" case (though I've had users
also do this to themselves with the "`hg debuglocks`" scenario), and most often
see this with the changelog. What ends up happening is we produce two items
(let's call them rev1 and rev2) in the .i file that have the same linkrev,
baserev, and offset into the .d file, while the data in the .d file is appended
properly. rev2's compressed_size is accurate for rev2, but when we go to
decompress the data in the .d file, we use the offset that's recorded in the
index file, which is the same as rev1, and attempt to decompress
rev2.compressed_size bytes of rev1's data. This usually does not succeed. :)
When using inline data, this also fails, though I haven't investigated why too
closely. This shows up as a "patch decode" error. I believe what's happening
there is that we're basically ignoring the offset field, getting the data
properly, but since baserev != rev, it thinks this is a delta based on rev
(instead of a full text) and can't actually apply it as such.
For now, I'm going to make this an optional component and default it to entirely
off. I may increase the default severity of this in the future, once I've
enabled it for my users and we gain more experience with it. Luckily, most of my
users have a versioned filesystem and can roll back to before the corruption has
been written, it's just a hassle to do so and not everyone knows how (so it's a
support burden). Users on other filesystems will not have that luxury, and this
can cause them to have a corrupted repository that they are unlikely to know how
to resolve, and they'll see this as a data-loss event. Refusing to create the
corruption is a much better user experience.
This mechanism is not perfect. There may be false-negatives (racy writes that
are not detected). There should not be any false-positives (non-racy writes that
are detected as such). This is not a mechanism that makes putting a repo on a
networked filesystem "safe" or "supported", just *less* likely to cause
corruption.
Differential Revision: https://phab.mercurial-scm.org/D9952
#testcases skip-detection fail-if-detected
Test situations that "should" only be reproducible:
- on networked filesystems, or
- user using `hg debuglocks` to eliminate the lock file, or
- something (that doesn't respect the lock file) writing to the .hg directory
while we're running
$ hg init a
$ cd a
$ cat > "$TESTTMP/waitlock_editor.sh" <<EOF
> [ -n "\${WAITLOCK_ANNOUNCE:-}" ] && touch "\${WAITLOCK_ANNOUNCE}"
> f="\${WAITLOCK_FILE}"
> start=\`date +%s\`
> timeout=5
> while [ \\( ! -f \$f \\) -a \\( ! -L \$f \\) ]; do
> now=\`date +%s\`
> if [ "\`expr \$now - \$start\`" -gt \$timeout ]; then
> echo "timeout: \$f was not created in \$timeout seconds (it is now \$(date +%s))"
> exit 1
> fi
> sleep 0.1
> done
> if [ \$# -gt 1 ]; then
> cat "\$@"
> fi
> EOF
$ chmod +x "$TESTTMP/waitlock_editor.sh"
Things behave differently if we don't already have a 00changelog.i file when
this all starts, so let's make one.
$ echo r0 > r0
$ hg commit -qAm 'r0'
Start an hg commit that will take a while
$ EDITOR_STARTED="$(pwd)/.editor_started"
$ MISCHIEF_MANAGED="$(pwd)/.mischief_managed"
$ JOBS_FINISHED="$(pwd)/.jobs_finished"
#if fail-if-detected
$ cat >> .hg/hgrc << EOF
> [debug]
> revlog.verifyposition.changelog = fail
> EOF
#endif
$ echo foo > foo
$ (WAITLOCK_ANNOUNCE="${EDITOR_STARTED}" \
> WAITLOCK_FILE="${MISCHIEF_MANAGED}" \
> HGEDITOR="$TESTTMP/waitlock_editor.sh" \
> hg commit -qAm 'r1 (foo)' --edit foo > .foo_commit_out 2>&1 ; touch "${JOBS_FINISHED}") &
Wait for the "editor" to actually start
$ WAITLOCK_FILE="${EDITOR_STARTED}" "$TESTTMP/waitlock_editor.sh"
Break the locks, and make another commit.
$ hg debuglocks -LW
$ echo bar > bar
$ hg commit -qAm 'r2 (bar)' bar
$ hg debugrevlogindex -c
rev linkrev nodeid p1 p2
0 0 222799e2f90b 000000000000 000000000000
1 1 6f124f6007a0 222799e2f90b 000000000000
Awaken the editor from that first commit
$ touch "${MISCHIEF_MANAGED}"
And wait for it to finish
$ WAITLOCK_FILE="${JOBS_FINISHED}" "$TESTTMP/waitlock_editor.sh"
#if skip-detection
(Ensure there was no output)
$ cat .foo_commit_out
And observe a corrupted repository -- rev 2's linkrev is 1, which should never
happen for the changelog (the linkrev should always refer to itself).
$ hg debugrevlogindex -c
rev linkrev nodeid p1 p2
0 0 222799e2f90b 000000000000 000000000000
1 1 6f124f6007a0 222799e2f90b 000000000000
2 1 ac80e6205bb2 222799e2f90b 000000000000
#endif
#if fail-if-detected
$ cat .foo_commit_out
transaction abort!
rollback completed
note: commit message saved in .hg/last-message.txt
note: use 'hg commit --logfile .hg/last-message.txt --edit' to reuse it
abort: 00changelog.i: file cursor at position 249, expected 121
And no corruption in the changelog.
$ hg debugrevlogindex -c
rev linkrev nodeid p1 p2
0 0 222799e2f90b 000000000000 000000000000
1 1 6f124f6007a0 222799e2f90b 000000000000
And, because of transactions, there's none in the manifestlog either.
$ hg debugrevlogindex -m
rev linkrev nodeid p1 p2
0 0 7b7020262a56 000000000000 000000000000
1 1 ad3fe36d86d9 7b7020262a56 000000000000
#endif