http: prefer chunked encoding over Content-Length: 0

[malyzelenyhnus]

Call http_size() before checking the request for empty body to prefer Transfer-Encoding: chunked even if Content-Length is 0.

http_size() sets k->maxdownload to -1 if both headers are present to ignore content-length so it must be called before following code:

if((k->maxdownload == 0) && (k->httpversion_sent < 20))
     k->download_done = TRUE;

Current ordering causes that with Content-Length : 0 is k->download_done set to True preventing remaining chunked data from being read.

[testclutch]

Analysis of PR #21706 at 6cebe13c:

Test 795 failed, which has NOT been flaky recently, so there could be a real issue in this PR. Note that this test has failed in 12 different CI jobs (the link just goes to one of them).

Generated by Testclutch

[vszakats]

Please rebase on master to fix the many CI errors. (it's due to a mistake I made)

[dfandrich]

Is this gated on the use of HTTP/1.1 since chunked encoding isn't supported on earlier versions?

[malyzelenyhnus]

Is this gated on the use of HTTP/1.1 since chunked encoding isn't supported on earlier versions?

This patch should only fix the conflict between chunked encoding and Content-Length headers - it's broken for the special case when Content-Length is zero. It should not change the current curl behavior in other ways.

And yes, chunked encoding should no be enabled in HTTP/1.0 and it looks like curl does not check the version. But IMO that's different issue.

[icing]

Is this gated on the use of HTTP/1.1 since chunked encoding isn't supported on earlier versions?

This patch should only fix the conflict between chunked encoding and Content-Length headers - it's broken for the special case when Content-Length is zero. It should not change the current curl behavior in other ways.

Yes, the patch looks correct. When Transfer-Encoding: chunked is present, a Content-Length: header should be ignored.

And yes, chunked encoding should no be enabled in HTTP/1.0 and it looks like curl does not check the version. But IMO that's different issue.

Well, a HTTP/1.0 response should not carry chunked, so the server is already in error. libcurl could fail the transfer in this case. Maybe worth it, not sure.

[bagder]

Thanks, can you check the conflict and force-push? Someone else stole the test number you used!

[bagder]

Thanks!