From: René Scharfe Date: Fri, 15 May 2026 14:30:34 GMT Subject: Re: [PATCH 1/2] strbuf: use st_add3() in strbuf_grow() Message-ID: <459f5f2b-2565-4dae-9f9f-8848a5cb9d94@web.de> In-Reply-To: <20260515043606.GA83595@coredump.intra.peff.net> On 5/15/26 6:36 AM, Jeff King wrote: > On Thu, May 14, 2026 at 10:13:19PM +0200, René Scharfe wrote: > >> Hmm, alloc_nr() doesn't do any overflow checking. It should, though, >> shouldn't it? > > Yes, probably. It's a known blind spot in the overflow checking, but > I think is OK in practice because: > > 1. We are growing an existing buffer by ~3/2. So even with ordering > the multiplication first, an overflow implies that you have a > single buffer consuming ~1/3 of your address space. > > On 64-bit systems that's impractically large, and on 32-bit systems I > think you generally run into fragmentation and address-space issues > first. > > 2. If alloc_nr(alloc) is less than the desired nr, we just use that nr > directly. So even if we did overflow, I think the result is > too-slow allocation, and not a buffer overflow. > > But it would be nice to be less hand-wavy. One of the reasons I hadn't > dug into it further is that I wanted to start making use of intrinsics > to avoid slowdowns. But since you're already doing that (and finding > that the compiler was doing the fast thing anyway!) it might be a good > time to make the jump. Didn't look at __builtin_mul_overflow() in detail; its situation could be different than for __builtin_add_overflow(), which turned out to be unnecessary on x64. > That's all assuming that no overflow happens before ALLOC_GROW() gets > the values. We also tend to do unchecked computions for the "nr" field > there, but it's usually just "nr_foo + 1", so the same logic applies: > you'd have to have an existing array consuming the entire address space > minus one byte to trigger an overflow. The use in read-cache.c::do_read_index() looks odd. Has been present since commit one. Is the point that it over-allocates to have room for additions right from the start? For read-only commands this only wastes memory, no? René