On Mon, Feb 25, 2019 at 03:49:01PM -0800, Jonathan Tan wrote: > Thanks, Peff, for noticing this. It's because the client sometimes sends > "0000" as a single request (that is, it flushes, and then before it > sends any data, it flushes again). And post_rpc() assumes that it can > always read something - which is usually correct, but not in this case; > we read in stateless_connect() first, and if we read "0000", we need to > tell post_rpc() to not read at all. > > This is a fixup on the tip of jt/http-auth-proto-v2-fix that fixes that. Thanks, I can confirm that this makes the problem go away (and your explanation makes perfect sense to me). > As for why the client sends a lone "0000", I'm not sure, but that's > outside the scope of this patch set, I think. Yeah, that does seem odd. I noticed it on noop fetches. So after we've done "ls-refs", would "fetch" need to send a flush to say "I don't want anything?" I guess not, since we're stateless, and it is literally making a new HTTP request just to say nothing. It does seem unique to protocol v2. -Peff
