| Commit message (Collapse) | Author | Age |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
git cat-file -p cannot be relied on to tell when an object is corrupt.
If it fails, the fsck may not find all bad objects -- but as long as fsck
exits nonzero, it will return a failing fsckresult, and so recovery will
run.
In recovery, the objects get unpacked. This allows the improved findMissing
to find all corrupt loose objects when fsck is run again as part of the
recovery.
Removed the repack / prune-packed workaround that I added earlier to find
corrupt loose objects that fsck wasn't finding. That was slow, and we want
to keep all loose objects, so that findMissing will work. And, it's
unncessary, now that findMissing is fixed.
Also, fixed some places where unreadable files would crash recovery.
|
| |
|
|
|
|
|
|
|
| |
Some corrupt objects can cause it to say the object is several TB, which
led to OOM.
Added some fork overhead, but it shouldn't be too bad; this is only run
against objects fsck outputs, and most of the time that is only corrupt
objects, and objects that refer to them.
|
| |
|
|
|
|
|
| |
Apparently some corruption to an object can cause cat-file to say it's N
bytes long, but only output N-M bytes of data. This causes Git.CatFile
to stall waiting for the rest. To fix, add a 1 minute timeout to the
cat-file, which should be enough time to read any reasonable object.
|
| |
|
|
|
| |
Aguably, I should make cat-file only throw IO exceptions, but currently it
throws some errors too.
|
| | |
|
| |
|
|
|
|
| |
It turned out to be broken, and led to failures.
6d67245728bbbc07ad1eeaf5b3c49f64c6bbcd11 was a better fix for the problem
that code tried to fix.
|
| | |
|
| |
|
|
|
|
|
| |
Sometimes git fsck outputs no shas even with --verbose, but fails, due to
badly corrupt objects. The best thing to do in this situation is to try to
pull and rsync from remotes, hoping that the bad objects will be
overwritten.
|
| |
|
