git-push failure following TCP changes

[waddlesplash]

I have a git-push that always fails about 10% (a few megabytes or so) the way through. The most recent TCP patch I just applied (#13704) didn't fix it, and the pcap dump seems to indicate that it's another window problem.

CC'ing Ayush, and also emailing him a copy of the pcap file. If anyone else would like to investigate, I can email them the pcap file also.

[taos]

In addition to git push failing, I'm also experiencing problems cloning the haiku git repository:

git clone https://git.haiku-os.org/haiku -v
Cloning into 'haiku'...
POST git-upload-pack (chunked)
error: RPC failed; HTTP 408 curl 22 The requested URL returned error: 408 Request Timeout
fatal: The remote end hung up unexpectedly

[a-star]

I have been real busy with exams and MS admissions.

I have started to look into the problem and found two possible issues. I will start to work on them in a day or two. Can expect a patch by the end of the week.

Sorry for the wait.

[waddlesplash]

That's alright. Thanks for looking into this!

[a-star]

Made some changes:

i) there was an integer promotion problem in updating the retransmission timeout : a signed int was being divided by an unsigned int. This was causing the values to overflow. Thus leading to huge values for timeout which manifested in the perception of pause in data flow.

ii) for an ack to be recognised as a duplicate ack, the advertised window must remain same. This was not taken care of in the code so I added it.

I tried my latest patch with files as large as 10mb (both git pull and git push) and everything worked fine. I hope the issue will be resolved now.

[waddlesplash]

That seems to improve things slightly: now instead of permanently stalling around 18%, it instead halts there for about 20 seconds and then jumps to 21% before stalling again. I'll send you another pcap file.

[waddlesplash]

Committed the patch in hrev51655.

[taos]

With hrev51703 the error message during git push changed for me to:

remote: fatal: early EOF

Error for git clone is still the same.

[pulkomandy]

@waddlesplash: please remember to set patches to "obsolete" if applying them and leaving the ticket open.

[taos]

It seems that - at least for me - the problems with git disappeared with (or around the time of) the migration of the repositories to the new server.

[waddlesplash]

The git push in question was to GitHub, so it hasn't changed.

[waddlesplash]

I have another git-push failure that looks extremely similar to this one. Any developer who wants the pcap file can have it...

[waddlesplash]

This specific git-push succeeded, so I don't have a way to reproduce this now.

[mmu_man]

I still have this issue here, trying to push to qtkeychains.

I did some tcpdump recording, and I believe we confuse the receiver with our Zero Window Probe, because we don't resent even though it only ACKs the previous segment, and we keep sending things after it:

• we send the last segment before window is closed, next seg = N,
• we get ACK for N, with window=0
• we send Zero Window Probe with 1 byte, next seg = N+1,
• we eventually get a window>0 ACK still at N,
• we start sending stuff again, but starting from N+1,
• receiver keeps ACKing N because it never accepted it.

It seems sending either 1 or 0 byte is valid for a ZWP, although I'm not entirely sure. At least it's easier to handle 0 than 1, and it seems to work. Wireshark shows them as duplicate ACKs, but they get the thing going.

References:

• ​RFC 793
• ​RFC 6429
• ​Wireshark wiki

[mmu_man]

See ​https://review.haiku-os.org/c/haiku/+/1188

[a-star]

If you want, I can look into the issue - it would be helpful if you could provide the pcap file as well. There might be some other issue underlying here since, as you mentioned, "we start sending stuff again, but starting from N+1" which is weird and not a desirable behavior.

[waddlesplash]

Patch merged in hrev53831.

[nielx]

Assign tickets with status=closed and resolution=fixed within the R1/beta2 development window to the R1/beta2 Milestone