RE: [PATCH] Increase core.packedGitLimit

[David Turner <David.Turner@twosigma.com>]

> -----Original Message-----
> From: Johannes Schindelin [mailto:Johannes.Schindelin@gmx.de]
> Sent: Thursday, April 20, 2017 5:58 PM
> To: Jeff King <peff@peff.net>
> Cc: David Turner <David.Turner@twosigma.com>; git@vger.kernel.org
> Subject: Re: [PATCH] Increase core.packedGitLimit
> 
> Hi Peff,
> 
> On Thu, 20 Apr 2017, Jeff King wrote:
> 
> > On Thu, Apr 20, 2017 at 04:41:18PM -0400, David Turner wrote:
> >
> > > When core.packedGitLimit is exceeded, git will close packs.  If
> > > there is a repack operation going on in parallel with a fetch, the
> > > fetch might open a pack, and then be forced to close it due to
> > > packedGitLimit being hit.  The repack could then delete the pack out
> > > from under the fetch, causing the fetch to fail.
> > >
> > > Increase core.packedGitLimit's default value to prevent this.
> > >
> > > On current 64-bit x86_64 machines, 48 bits of address space are
> > > available.  It appears that 64-bit ARM machines have no standard
> > > amount of address space (that is, it varies by manufacturer), and
> > > IA64 and POWER machines have the full 64 bits.  So 48 bits is the
> > > only limit that we can reasonably care about.  We reserve a few bits
> > > of the 48-bit address space for the kernel's use (this is not
> > > strictly necessary, but it's better to be safe), and use up to the
> > > remaining 45.  No git repository will be anywhere near this large
> > > any time soon, so this should prevent the failure.
> >
> > Yep, I think this is a reasonable direction.
> >
> > > ---
> > >  git-compat-util.h | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > This probably needs an update to the core.packedGitLimit section of
> > Documentation/config.txt.
> >
> > > diff --git a/git-compat-util.h b/git-compat-util.h index
> > > 8a4a3f85e7..1c5de153a5 100644
> > > --- a/git-compat-util.h
> > > +++ b/git-compat-util.h
> > > @@ -616,7 +616,7 @@ extern int git_lstat(const char *, struct stat
> > > *);  #endif
> > >
> > >  #define DEFAULT_PACKED_GIT_LIMIT \
> > > -	((1024L * 1024L) * (size_t)(sizeof(void*) >= 8 ? 8192 : 256))
> > > +	((1024L * 1024L) * (size_t)(sizeof(void*) >= 8 ? (32 * 1024L *
> > > +1024L) : 256))
> >
> > I wondered if we would run afoul of integer sizes on 64-bit systems
> > where "long" is still only 32-bits (i.e., Windows). But I think it's
> > OK, because the values before we cast to size_t are in megabytes. So
> > your
> > 32*1024*1024 needs only 25 bits to store it. And then after we cast to
> > size_t, everything is in 64-bit.
> 
> Indeed, when I patch a local Git checkout accordingly, I see that
> packed_git_limit is set to 35184372088832.
> 
> The bigger problem in this regard is that users are allowed to override this via
> core.packedgitlimit but that value is parsed as an unsigned long.

We might want to think about replacing git_config_ulong with 
git_config_size_t in nearly all cases. "long" has ceased to be 
useful.  More modern versions of C prefer uint64_t, but I
think that we'll usually want size_t because these values will
be used as memory limits of various sorts.