Re: [PATCH 3/3] grep: recurse in-process using 'struct repository'

[Brandon Williams <bmwill@google.com>]

On 07/11, Jonathan Nieder wrote:
> Hi,
> 
> Brandon Williams wrote:
> 
> > Convert grep to use 'struct repository' which enables recursing into
> > submodules to be handled in-process.
> 
> \o/
> 
> This will be even nicer with the changes described at
> https://public-inbox.org/git/20170706202739.6056-1-sbeller@google.com/.
> Until then, I fear it will cause a regression --- see (*) below.
> 
> [...]
> >  Documentation/git-grep.txt |   7 -
> >  builtin/grep.c             | 390 +++++++++------------------------------------
> >  cache.h                    |   1 -
> >  git.c                      |   2 +-
> >  grep.c                     |  13 --
> >  grep.h                     |   1 -
> >  setup.c                    |  12 +-
> >  7 files changed, 81 insertions(+), 345 deletions(-)
> 
> Yay, tests still pass.
> 
> [..]
> > --- a/Documentation/git-grep.txt
> > +++ b/Documentation/git-grep.txt
> > @@ -95,13 +95,6 @@ OPTIONS
> >  	<tree> option the prefix of all submodule output will be the name of
> >  	the parent project's <tree> object.
> >  
> > ---parent-basename <basename>::
> > -	For internal use only.  In order to produce uniform output with the
> > -	--recurse-submodules option, this option can be used to provide the
> > -	basename of a parent's <tree> object to a submodule so the submodule
> > -	can prefix its output with the parent's name rather than the SHA1 of
> > -	the submodule.
> 
> Being able to get rid of this is a very nice change.
> 
> [...]
> > +++ b/builtin/grep.c
> [...]
> > @@ -366,14 +349,10 @@ static int grep_file(struct grep_opt *opt, const char *filename)
> >  {
> >  	struct strbuf buf = STRBUF_INIT;
> >  
> > -	if (super_prefix)
> > -		strbuf_addstr(&buf, super_prefix);
> > -	strbuf_addstr(&buf, filename);
> > -
> >  	if (opt->relative && opt->prefix_length) {
> > -		char *name = strbuf_detach(&buf, NULL);
> > -		quote_path_relative(name, opt->prefix, &buf);
> > -		free(name);
> > +		quote_path_relative(filename, opt->prefix, &buf);
> > +	} else {
> > +		strbuf_addstr(&buf, filename);
> >  	}
> 
> style micronit: can avoid these braces since both branches are
> single-line.

Didn't realize that with all the deleted lines, I'll fix for the next
version.

> 
> [...]
> > @@ -421,284 +400,80 @@ static void run_pager(struct grep_opt *opt, const char *prefix)
> >  		exit(status);
> >  }
> >  
> > -static void compile_submodule_options(const struct grep_opt *opt,
> > -				      const char **argv,
> > -				      int cached, int untracked,
> > -				      int opt_exclude, int use_index,
> > -				      int pattern_type_arg)
> > -{
> [...]
> > -	/*
> > -	 * Limit number of threads for child process to use.
> > -	 * This is to prevent potential fork-bomb behavior of git-grep as each
> > -	 * submodule process has its own thread pool.
> > -	 */
> > -	argv_array_pushf(&submodule_options, "--threads=%d",
> > -			 (num_threads + 1) / 2);
> 
> Being able to get rid of this is another very nice change.
> 
> [...]
> > +	/* add objects to alternates */
> > +	add_to_alternates_memory(submodule.objectdir);
> 
> (*) This sets up a single in-memory object store with all the
> processed submodules.  Processed objects are never freed.
> This means that if I run a command like
> 
> 	git grep --recurse-submodules -e neverfound HEAD
> 
> in a project with many submodules then memory consumption scales in
> the same way as if the project were all one repository.  By contrast,
> without this patch, git is able to take advantage of the implicit
> free() when each child exits to limit its memory usage.
> 
> Worse, this increases the number of pack files git has to pay
> attention to the sum of the numbers of pack files in all the
> repositories processed so far.  A single object lookup can take
> O(number of packs * log(number of objects in each pack)) time.  That
> means performance is likely to suffer as the number of submodules
> increases (n^2 performance) even on systems with a lot of memory.
> 
> Once the object store is part of the repository struct and freeable,
> those problems go away and this patch becomes a no-brainer.
> 
> What should happen until then?  Should this go in "next" so we can get
> experience with it but with care not to let it graduate to "master"?

I agree that this is an issue and that we need to address by having
an object store per repository.  While that is being worked on (by
Stefan) I don't know how long it would take to have it be a reality.
So the question ends up being do we care more about the state of the
code and cleaning up a lot of 'hacks' that I introduced to get grep
working with submodules, or do we care about the performance more.  I
don't know which is the right answer but I'd personally like to see the
hacks I added to be removed sooner rather than later.  That and I think
that with the code in this sate it would make it easier to transition
once we have per-repository object-stores.

Either way I should add a NEEDSWORK comment here to indicate that it
should be removed once per-repo object-stores exist.

> 
> Aside from those two concerns, this patch looks very good from a quick
> skim, though I haven't reviewed it closely line-by-line.  Once we know
> how to go forward, I'm happy to look at it again.
> 
> Thanks,
> Jonathan

-- 
Brandon Williams