git/list[1] front-page[2] threads[3] people[4] search[5] about
 

Re: [PATCH] wrapper: properly handle MAX_IO_SIZE in `write_in_full()`

From
Patrick Steinhardt <ps@pks.im>
Date
Apr 10, 2026, 05:19 UTC
Message-ID
<adiIfQjobtK3MDPW@pks.im>
In-Reply-To
<xmqqzf3brgbt.fsf@gitster.g>
On Thu, Apr 09, 2026 at 02:09:42PM -0700, Junio C Hamano wrote:
Show 47 quoted lines
> Jeff King <peff@peff.net> writes:
> 
> > On Thu, Apr 09, 2026 at 01:40:36PM -0700, Junio C Hamano wrote:
> >
> >> >> As the width of ssize_t in bits can be a lot smaller than size_t,
> >> >> the above "unsigned_add_overflows() triggers way too late for the
> >> >> check to matter, no?
> >> >
> >> > I think it is correct as-is.
> >> >
> >> > The real check against ssize_t is later, when we compare total_length to
> >> > MAX_IO_SIZE (which is clamped to SSIZE_MAX). So this is just making sure
> >> > we do not overflow size_t when counting up the total (and if we do, we
> >> > _know_ we are going to overflow ssize_t, which must be smaller).
> >> 
> >> But then what happens after it breaks out of the loop?  We cannot be
> >> at i==0, so let's say we have a reasonably small iov[0] and iov[1]
> >> that is so large and makes size_t wraparound.  We break out here,
> >> and then send the iov[0] with writev().  But have we checked if
> >> iov[0] is under MAX_IO_SIZE in that case before calling writev()?
> >
> > I think so. Either:
> >
> >   - We completed the first iteration of the loop successfully (and i >=
> >     1), in which case we added iov[0].iov_len to total_length, and then
> >     compared total_length against MAX_IO_SIZE, but did not break out of
> >     the loop. So we know iov[0] is within the limits.
> >
> >   - We bailed at i==0 either because of addition overflow, or because of
> >     the MAX_IO_SIZE check. Either way, we will bail to xwrite() because
> >     i is 0.
> 
> Yup, you're right.
> 
> There is no addition overflow at i==0, but I do not think we can
> construct a case where the sum is not checked against MAX_IO_SIZE
> before the vector is passed to underlying writev().
> 
> iov[0].iov_len that is slightly smaller than MAX_IO_SIZE would allow
> us to keep looping to i==1 at which time iov[1].iov_len is so big
> that we may trigger unsigned_add_overflows() check, but then what we
> send to writev() is the first segment, which is smaller than
> MAX_IO_SIZE, so we are OK.
> 
> iov[0].iov_len that is slightly larger than MAX_IO_SIZE would stop
> us moving to i==1 at the end of the loop, and directly punt to
> xwrite(), so we are OK, too.

Let's drop this patch for now. I'll pick it up again in the next release cycle when reintroducing writev(3p). Thanks, all!

Patrick
Previous: Junio C Hamano
Message 8 of 8 in “wrapper: properly handle MAX_IO_SIZE in `write_in_full()`”
  1. wrapper: properly handle MAX_IO_SIZE in `write_in_full()`Patrick Steinhardt, Apr 9, 2026
  2. Ben KnobleApr 9, 2026
  3. Junio C HamanoApr 9, 2026
  4. Jeff KingApr 9, 2026
  5. Junio C HamanoApr 9, 2026
  6. Jeff KingApr 9, 2026
  7. Junio C HamanoApr 9, 2026
  8. Patrick SteinhardtApr 10, 2026

Read the whole thread, see it on lore, or plain text.

$ cat FOOTERMessages come from the public archive at lore.kernel.org/git, fetched every hour. The front page is chosen and written each morning by an AI editor and can be wrong; the threads themselves are the record. About and API. For agents: an MCP server at https://gitlist.dev/mcp, and any thread, story or person page as Markdown by adding .md to its URL (or sending Accept: text/markdown). Details in /llms.txt.