Re: git archive generates tar with malformed pax extended attribute

From: René Scharfe <l.s.r@web.de>
Date: Jun 2, 2019, 16:58 UTC
Message-ID: <c961d89d-db0b-597f-c183-81aa791c0987@web.de>
In-Reply-To: <20190530115554.GB31607@sigill.intra.peff.net>

Am 30.05.19 um 13:55 schrieb Jeff King:

Show 19 quoted lines

> On Wed, May 29, 2019 at 07:54:44PM +0200, René Scharfe wrote:
>
>> Am 29.05.19 um 03:17 schrieb Jeff King:
>>> But here the problem is in the tree, not the blob. So we're not finding
>>> suspect blobs, but rather re-checking each tree. And no matter what we
>>> do (whether it's visiting the object again, or creating a set or mapping
>>> with the object names) is going to be linear there. And a repository
>>> with a symlink in the root tree is going to revisit or put in our
>>> mapping every single root tree.
>>
>> That's true, potentially it needs remember and/or reprocess all trees,
>> meaning this check may double the run time of fsck in the worst case.
>> Example from the wild: The kernel repo currently has 36 symlinks and
>> 6+ million objects are checked in total, and the symlink check processes
>> 18943 trees_with_symlinks entries there.
>
> That sounds about right. It's basically every version of every tree that
> has a symlink. Did it make a noticeable difference in timing? Indexing
> the whole kernel history is already a horribly slow process. :)

Right, I didn't notice a difference -- no patience for watching that thing to the end. But here are some numbers for v2.21.0 vs. master with the patch:

Benchmark #1: git fsck
  Time (mean ± σ):     307.775 s ±  9.054 s    [User: 307.173 s, System: 0.448 s]
  Range (min … max):   294.052 s … 322.931 s    10 runs

Benchmark #2: ~/src/git/git fsck
  Time (mean ± σ):     319.754 s ±  2.255 s    [User: 318.927 s, System: 0.671 s]
  Range (min … max):   316.376 s … 323.747 s    10 runs

Summary
  'git fsck' ran
    1.04 ± 0.03 times faster than '~/src/git/git fsck'

Seeing only a single CPU core being stressed for that long is a bit sad to see. Checking individual objects should be relatively easy to parallelize, shouldn't it?

René