On Fri, Apr 19, 2019 at 4:24 AM Jeff King <peff@xxxxxxxx> wrote:
>
> On Thu, Apr 18, 2019 at 05:14:08PM -0400, Jeff King wrote:
>
> > Just so we don't forget about it, I wrote this fix up as a patch. And in
> > fact it led to a few other cleanups. I think the first one is definitely
> > worth doing now, even if there are other similar cases lurking in the
> > rest of the index code.
> >
> > The other two are optional, though I think they are worth it (and not
> > too hard to verify that they are doing the right thing).
> >
> > These are on top of js/untracked-cache-allocfix (though they could
> > easily be ported to a separate topic if we want).
> >
> > [1/3]: untracked-cache: be defensive about missing NULs in index
> > [2/3]: untracked-cache: simplify parsing by dropping "next"
> > [3/3]: untracked-cache: simplify parsing by dropping "len"
>
> I also wondered if we could just accept the cost of calloc() here and
> use FLEX_ALLOC to simplify things. That resulted in the patch below, but
> I didn't include it with the initial 3, because I think it's too
> subtle/gross for my tastes.
It's probably ok. If I remember correctly, reading UNTR extension (on
a huge repo) took the longest time after Ben added support for reading
the index with multiple threads. So performance is a concern, but I
don't think calloc() would be the problem. malloc() itself without the
memory pool could probably slow down more when we have lots and lots
of directories.
> -- >8 --
> Subject: untracked-cache: use FLEX_ALLOC to create internal structs
>
> The untracked_cache_dir struct has a FLEX_ARRAY in it. Let's use
> FLEX_ALLOC_MEM to allocate it, which saves us having to compute the
> length ourselves.
>
> In theory this could be slightly slower, since the FLEX_ALLOC macros use
> calloc (and we just memcpy over most of the contents anyway). But in
> practice this distinction is not generally measurable.
>
> Note that because we then fill in the pre-flex elements of the struct
> using a memcpy, we need to take care to use the exact size of that
> space and _not_ "sizeof(ud)", since the latter may include padding (or
> even an extra byte on systems where FLEX_ARRAY is 1).
>
> Signed-off-by: Jeff King <peff@xxxxxxxx>
> ---
> If we wanted to go this route, I think it would make sense to provide a
> FLEX_ALLOC macro that takes a "template" set of bytes as a ptr/len pair,
> and writes it before we fill in the flex portion.
>
> Then we could do something like:
>
> FLEX_ALLOC_COPY(untracked, &ud, sizeof(ud), name, data, eos - data);
>
> If this is the only such case, it's probably not worth it (I didn't
> really look around for more, though).
>
> dir.c | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/dir.c b/dir.c
> index 60438b2cdc..7cd4eec198 100644
> --- a/dir.c
> +++ b/dir.c
> @@ -2757,9 +2757,9 @@ static int read_one_dir(struct untracked_cache_dir **untracked_,
> if (!eos || eos == end)
> return -1;
>
> - *untracked_ = untracked = xmalloc(st_add3(sizeof(*untracked), eos - data, 1));
> - memcpy(untracked, &ud, sizeof(ud));
> - memcpy(untracked->name, data, eos - data + 1);
> + FLEX_ALLOC_MEM(untracked, name, data, eos - data);
> + memcpy(untracked, &ud, offsetof(struct untracked_cache_dir, name));
> + *untracked_ = untracked;
> data = eos + 1;
>
> for (i = 0; i < untracked->untracked_nr; i++) {
> --
> 2.21.0.1092.g8b0302e9c4
>
--
Duy
