Mark Jamsek:

> The main change is that we memory map the file index to avoid lots of
> small reads that can be costly in large repos, for example, with many
> files.

I'd think that "got status" in /usr/src is dominated by the 92,300
fstatat() calls...

I tried this (cd /usr/src; got st) on an APU2, which qualifies as
slow for amd64 purposes:

0.92-current, three runs:
    0m19.97s real     0m08.82s user     0m11.16s system
    0m19.95s real     0m08.56s user     0m10.93s system
    0m19.99s real     0m08.74s user     0m10.96s system

+ patch, three runs:
    0m19.28s real     0m08.21s user     0m10.96s system
    0m19.14s real     0m08.02s user     0m10.76s system
    0m19.33s real     0m07.72s user     0m11.42s system

I think we need more measurements to be certain that there is any
effect at all. :->


I still have a code comment:

> +static const struct got_error *
> +mread_fileindex_path(char **path, struct got_hash *ctx, const uint8_t *map,
> +    size_t mapsz, size_t *offset)
> +{
> +	const uint8_t	*p, *nul;
> +	size_t		 len, pad, pathlen;
> +
> +	if (mapsz < *offset)
> +		return got_error(GOT_ERR_FILEIDX_BAD);
> +
> +	p = map + *offset;
> +
> +	nul = memchr(p, '\0', mapsz);
> +	len = nul - p;
> +	pad = 8 - len % 8;
> +
> +	pathlen = len + pad;
> +
> +	if (mapsz < *offset + pathlen)
> +		return got_error(GOT_ERR_FILEIDX_BAD);

If the file index is corrupt, the memchr() can overrun the mapped
area and trigger a segfault.  The max length should be something
like map + mapz - p.

-- 
Christian "naddy" Weisgerber                          naddy@mips.inka.de