threads / patch / 50685

patchconvert: avoid malloc of original file size

Subject: [PATCH] convert: avoid malloc of original file size

## tl;dr

3 messages between Mar 7, 2019 and Mar 8, 2019. Diffs are folded; open one to read it.

replies: 2people: 2as markdown or json

Jeff King· Mar 7, 2019, 19:56 UTC · lore

From: Joey Hess <id@joeyh.name>

We write the output of a "clean" filter into a strbuf. Rather than growing the strbuf dynamically as we read its output, we make the initial allocation as large as the original input file. This is a good guess when the filter is just tweaking a few bytes, but it's disastrous when the point of the filter is to condense a very large file into a short identifier (e.g., the way git-lfs and git-annex do). We may ask to allocate many gigabytes, causing the allocation to fail and Git to die().

Instead, let's just let strbuf do its usual growth.

When the clean filter does output something around the same size as the worktree file, the buffer will need to be reallocated until it fits, starting at 8192 and doubling in size. Benchmarking indicates that reallocation is not a significant overhead for outputs up to a few MB in size.

Signed-off-by: Joey Hess <id@joeyh.name>
Signed-off-by: Jeff King <peff@peff.net>
---
This is a resurrection of the patch from:

  https://public-inbox.org/git/20190122220714.GA6176@kitenet.net/

It got stalled on discussion of the commit message, which I've rewritten here to match the suggestions in the thread.

As discussed there, I do think this only solves half the problem, as the smudge filter has the same issue in reverse. That's more complicated to fix, and AFAIK nobody is working on it. But I don't think there's any reason not to pick up this part in the meantime.

 convert.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Show changes to convert.c +1 −1

diff --git a/convert.c b/convert.c
index 5d0307fc10..94ff837649 100644
--- a/convert.c
+++ b/convert.c
@@ -731,7 +731,7 @@ static int apply_single_file_filter(const char *path, const char *src, size_t le
 	if (start_async(&async))
 		return 0;	/* error was already reported */
 
-	if (strbuf_read(&nbuf, async.out, len) < 0) {
+	if (strbuf_read(&nbuf, async.out, 0) < 0) {
 		err = error(_("read from external filter '%s' failed"), cmd);
 	}
 	if (close(async.out)) {

-- 
2.21.0.787.g929e938557

Junio C Hamano· Mar 8, 2019, 01:26 UTC · re: Jeff King · lore

Re: [PATCH] convert: avoid malloc of original file size

Jeff King <peff@peff.net> writes:

> As discussed there, I do think this only solves half the problem, as the
> smudge filter has the same issue in reverse. That's more complicated to
> fix, and AFAIK nobody is working on it. But I don't think there's any
> reason not to pick up this part in the meantime.

Yeah, I agree that the reverse direction shares the same issue.

I am not sure 0 is a good initial value in this direction, either; I'd rather clip to min(len, core.bigfilethreshold) or something like that, to avoid regressing the more normal use cases.

But let's queue this and see what happens.

Thanks.

Show 17 quoted lines

>
>  convert.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/convert.c b/convert.c
> index 5d0307fc10..94ff837649 100644
> --- a/convert.c
> +++ b/convert.c
> @@ -731,7 +731,7 @@ static int apply_single_file_filter(const char *path, const char *src, size_t le
>  	if (start_async(&async))
>  		return 0;	/* error was already reported */
>  
> -	if (strbuf_read(&nbuf, async.out, len) < 0) {
> +	if (strbuf_read(&nbuf, async.out, 0) < 0) {
>  		err = error(_("read from external filter '%s' failed"), cmd);
>  	}
>  	if (close(async.out)) {

Jeff King· Mar 8, 2019, 02:52 UTC · re: Junio C Hamano · lore

Re: [PATCH] convert: avoid malloc of original file size

On Fri, Mar 08, 2019 at 10:26:24AM +0900, Junio C Hamano wrote:

Show 12 quoted lines

> Jeff King <peff@peff.net> writes:
> 
> > As discussed there, I do think this only solves half the problem, as the
> > smudge filter has the same issue in reverse. That's more complicated to
> > fix, and AFAIK nobody is working on it. But I don't think there's any
> > reason not to pick up this part in the meantime.
> 
> Yeah, I agree that the reverse direction shares the same issue.
> 
> I am not sure 0 is a good initial value in this direction, either;
> I'd rather clip to min(len, core.bigfilethreshold) or something like
> that, to avoid regressing the more normal use cases.

That was my initial thought, too, but Joey's benchmarks show that it doesn't seem to make a big difference either way. In his numbers it did get measurable for a 1GB file, but we'd still not use "hint == len" in that case (we'd probably do one or two doublings to get there).

I also think running a real (non-condensing) filter on a 1GB file is already a pretty unlikely corner case.

> But let's queue this and see what happens.

Sounds good to me.

-Peff