Re: Refreshing index timestamps without reading content

On Mon, Jan 09, 2017 at 07:01:36AM -0800, Junio C Hamano wrote:

Duy Nguyen [off-list ref] writes:

quoted

On Thu, Jan 5, 2017 at 6:23 PM, Quentin Casasnovas
[off-list ref] wrote:

quoted

Is there any way to tell git, after the git ls-tree command above, to
refresh its stat cache information and trust us that the file content has
not changed, as to avoid any useless file read (though it will obviously
will have to stat all of them, but that's not something we can really
avoid)

I don't think there's any way to do that, unfortunately.

Lose "unfortunately".

quoted

quoted

If not, I am willing to implement a --assume-content-unchanged to the git
update-index if you guys don't see something fundamentally wrong with this
approach.

If you do that, I think you should go with either of the following options

- Extend git-update-index --index-info to take stat info as well (or
maybe make a new option instead). Then you can feed stat info directly
to git without a use-case-specific "assume-content-unchanged".

- Add "git update-index --touch" that does what "touch" does. In this
case, it blindly updates stat info to latest. But like touch, we can
also specify  mtime from command line if we need to. It's a bit less
generic than the above option, but easier to use.

Even if we assume that it is a good idea to let people muck with the
index like this, either of the above would be a usable addition,
because the cached stat information does not consist solely of
mtime.

"git update-index --index-info" was invented for the case where a
user or a script _knows_ the object ID of the blob that _would_
result if a contents of a file on the filesystem were run through
hash-object.  So from the interface's point of view, it may make
sense to teach it to take an extra/optional argument that is the
path to the file and take the stat info out of the named file when
the extra/optional argument was given.

But that assumes that it is a good idea to do this in the first
place.  It was deliberate design decision that setting the cached
stat info for the entry was protected behind actual content
comparison, and removing that protection will open the index to
abuse.

Hi Junio,

Thanks for your feedback, appreciated :)

I do understand how it would be possible for someone to shoot themselves in
the feet with such option, but it solves real life use cases and improved
build times very signficantly here.

Another use case we have is setting up very lightweight linux work trees,
by reflinking from a base work-tree.  This allows for a completely
different work-tree taking up almost no size at first, whereas using a
shared clone or the recent worktree subcommand would "waste" ~500MB*:

 # linux-2.6 is a shared clone of a bare clone residing locally
 ~ $ cp --reflink -a linux-2.6 linux-2.6-reflinked

 # At this point, the mtime inside linux-2.6-reflinked are matching the
 # mtime of the source linux-2.6 (since we used the '-a' option of 'cp)
 ~ $ diff -u <(stat linux-2.6/README) <(stat linux-2.6-reflinked/README)
 --- /proc/self/fd/11  2017-01-09 16:34:04.523438942 +0100
 +++ /proc/self/fd/12  2017-01-09 16:34:04.523438942 +0100
 @@ -1,8 +1,8 @@
 -  File: 'linux-2.6/README'
 +  File: 'linux-2.6-reflinked/README'
    Size: 18372		Blocks: 40         IO Block: 4096   regular file
 -Device: fd00h/64768d	Inode: 268467090   Links: 1
 +Device: fd00h/64768d	Inode: 805970606   Links: 1
  Access: (0644/-rw-r--r--)  Uid: ( 1000/ quentin)   Gid: ( 1000/ quentin)
  Access: 2017-01-09 12:04:15.317758718 +0100
  Modify: 2017-01-09 12:04:12.566758772 +0100
 -Change: 2017-01-09 12:04:12.566758772 +0100
 +Change: 2017-01-09 16:29:48.305444003 +0100
   Birth:

  # Now let's check how long it takes to refresh the index from the source
  # and destination..
  ~/linux-2.6 $ time git update-index --refresh
  git update-index --refresh  0.04s user 0.08s system 204% cpu 0.058 total
                                                               ~~~~~~~~~~~
  ~/linux-2.6-reflinked $ time git update-index --refresh
  git update-index --refresh  2.40s user 1.43s system 38% cpu 10.003 total
                                                              ~~~~~~~~~~~~

This is quite a high penalty when a power user knows that his lightweight
copy worktree matches the index!  That's on a single worktree on a fairly
decent SSD but our build farm would do this with hundreds of work trees in
parallel, all residing on a spinning disks - the penalty would be much
worse than 10 seconds wasted.

I might have a look at adding reflinking awareness to the worktree
subcommand (if it hasn't been implemented already) to avoid this index
force refreshing but that still would not fix our other use case where we
knowingly change the mtime of the files on our worktrees.

* That's a rather low estimate:

  git ls-tree -zr --name-only v4.9 | du -csh --files0-from=- | tail -n1
  747M  total

The userbase of Git has grown wide enough that it is harder to say "If
you lie that a file whose contents does not match the index is up to date
using this mechanism, you will lose data and all bad things happen---you
can keep both halves".  Once we release a version of Git with such a
"feature", the first bug report will be "I did not want to run
'update-index --refresh' because it takes time, and some index entries
apparently did not match what is on the filesystem, and I got a corrupt
working file after a merge.  Git should make sure that the contents match
when using the new 'path to the file' argument when updating the cached
stat info!".  I do not have a good answer to such a bug report.

The good answer would probably be that this is intended behaviour and would
explicitly be mentionned in the manual for that specific option?  :)

How about renaming the option (which is more of a sub-command) to
--i-know-this-can-eat-my-data-but-refresh-stat-only?  Then we can print a
big warning saying this is generally a VERY bad idea?

Q