Re: long fast-import errors out "failed to apply delta"

From: Dmitry Ivankov <divanorama@gmail.com>
Date: Jul 23, 2011, 20:34 UTC
Message-ID: <CA+gfSn8C-nB2hSSRTqSu1N1Z-b8ctRsVmUGAjLXTW0du_W3EQw@mail.gmail.com>
In-Reply-To: <CA+gfSn8jjptyv10iVimmfXpf6QHrR_3UpkRdd+Dv1M=KgORtGQ@mail.gmail.com>

+people contributed to the related parts of fast-import.c

Hi!

I did more tests and investigations. (one can safely jump over next 2 paragraphs) There is a problem in fast-import, triggered on long svn-fe produced imports (10k..80k commits depending on the svn repo history, in my tests at least). fast-imports sometimes writes off a tree object with wrong sha1. It goes silent unless it tries to read these objects for later revisions (via ls command) and "crashes", or unless fsck is run to find sha1 mismatches.

This seems to be a old bug. Due to the content of my fast-import streams, I need cat-blob and ls commands (once the import stream is sniffed these can be ignored) and more importantly 334fba656b50c9.. "Teach fast-import to import subtrees named by tree id" 30 Jun 2010 for "M 040000 <tree id> pathname" commands. The earliest git version I've tested so far is v1.6.0.6-7-g3d1d81e Jan 2009 + dummy cat-blob and ls + 334fba656b50c9..

So, how do I currently reproduce it. Import gcc svn repository with svn-fe up to r15507 - fine. (1.9G fast-import stream, 3min to import, 1min to fsck) Import up to r15508 - fine, but fsck finds a sha1 mismatch in a tree object. Strip r15508 to a three commands:

..commit header.. ls :14842 branches/gcc3/gcc/config M 040000 fbc83f80e9516c831918dff149058cba38a2e5f1 tags/egcs_ss_970917/gcc/config ls :15459 trunk/gcc/config/alpha M 040000 9ffe84c346eec93b523d95ce642b54d54d23109c tags/egcs_ss_970917/gcc/config/alpha ls "tags/egcs_ss_970917/gcc/config/alpha" D tags/egcs_ss_970917/gcc/config/alpha/vms-tramp.asm

So here we set a directory with one old tree, then set it's child with another old tree, and then delete a file in the child. The broken tree is the resulting tree for that child. sha1 written to the pack matches the intent (the second M + D) while the content is wrong (matches the first M/subdir + D) - the second M command is partially lost.

If a checkpoint command is added before the last commit there is no sha1 mismatch. I haven't yet found a small fast-import stream with this bug, but it is a definetely a logic bug somewhere in fast-import. I randomly changed pool sizes (didn't touch packfile settings yet), run valgrind, slightly modified the stream many times, tested a bunch of different git versions on two machines - the bug is stable.

Any ideas? How is it possible at all that the tree's sha1 and content diverge, and moreover this is unnoticed by fast-import and a broken packfile is produced?

On Thu, Jul 7, 2011 at 4:47 PM, Dmitry Ivankov <divanorama@gmail.com> wrote:

Show 11 quoted lines

> Hi,
>
> I'm getting a strange error from git-fast-import.
> Tested on v1.7.5 and v1.7.6 on two machines (gentoo amd64 8gb ram,
> 3-core amd cpu; gentoo x86 2gb ram, 1-core intel mobile cpu).
> The crash is stable - same message, instruction and deepest function
> (patch_delta) parameters contents.
>
> $ git fast-import --quiet < big_dump
> fatal: failed to apply delta
> fast-import: dumping crash report to fast_import_crash_7700

[skipped the details]