Re: [PATCH v2 0/7] remote-curl: fix deadlocks when remote server disconnects

[Denton Liu <liu.denton@gmail.com>]

Hi Peff,

On Mon, May 18, 2020 at 12:50:56PM -0400, Jeff King wrote:
> On Mon, May 18, 2020 at 11:47:17AM -0400, Denton Liu wrote:
> 
> > Changes since v1:
> > 
> > * Remove fallthrough in switch in favour of just extracting the common
> >   call out of the switch in patch 3
> > 
> > * Add more detail in function comment and use `const char linelen[4]` in
> >   patch 4
> > 
> > * Implement most of Peff's suggestions[0] in patch 5
> > 
> > * Only operate on stateless_connect() in patch 5
> > 
> > * Add tests in patch 5
> > 
> > * Drop "remote-curl: ensure last packet is a flush" in favour of
> >   "stateless-connect: send response end packet"
> 
> Overall this looks pretty cleanly done. I left a few minor comments
> throughout, but the real question is whether we prefer the "0002" packet
> in the last one, or if we instead insist that the response end in a
> flush.

Thanks for the prompt review!

> At first glance, the "0002" seems like it's more flexible, because we're
> making fewer assumptions about what's being transferred over the
> stateless-connect channel. But in reality it still has to be pktlines
> (because we're checking them for incomplete or invalid packets already).
> So all it really buys us is that the server response doesn't have to end
> with a flush packet.
> 
> So I dunno. The "0002" solution is slightly more flexible, but I'm not
> sure it helps in practice. And it does eat up one of our two remaining
> special packet markers.

Yeah, I was worried about consuming a special packet. One alternative
that I considered but is kind of gross is sending something like
"0028gitremote-helper: response complete\n" instead of "0002". Then,
instead of "0002" checks, we can check for that special string instead.
I don't _think_ that stateless-connect currently allows for completely
arbitrary data but I might be mistaken.

> There is another solution, which would allow arbitrary data over
> stateless-connect: adding an extra level of pktline framing between the
> helper and the parent process. But that's rather ugly (inner pktlines
> may be split across outer ones, so you have to do a bunch of buffer
> reassembly). I think that's actually how v0 http works, IIRC.
> IMHO it probably isn't worth pursuing. That extra layer wrecks the
> elegance to the v2 stateless-connect approach, and we really do expect
> only pktlines to go over it.

This was the original approach that I was working on but I found it to
be much too invasive for my liking. (Also, I never actually managed to
get it working ;) ) I think I gave up when I realised I had to insert
reframing logic into index-pack and unpack-objects.

> So I think either of your solutions (enforcing a final flush, or the
> 0002 packet) is preferable. I'm on the fence between them.

I'm mostly on the fence too. One advantage of 0002, however, is that a
malicious server can't end a request with 0002 as that's explicitly
prevented. If a malicious server closes a connection after sending a
0000, I think that they could cause a deadlock to happen if there are
multiple flush packets expected in a response. You mentioned in a
sibling email that this currently doesn't happen wrt stateless-connect
although I'm not sure if in the future, this is something that might
change. I dunno.

> -Peff