Regression: Client keep-alive subsequent requests very slow

[nickkelsey]

Linux, Client mode, OpenSSL and Poll defines set, keep-alive enabled, comparing 0.13.3 with 0.18.2.
Server is running Apache2 delivering static files, Content-Length used.

The first GET request takes about 200ms as expected.
With 0.13.3 the second and subsequent requests take around 60ms as expected.
With 0.18.2 the second and subsequent requests take anywhere from 4000ms to 5000ms.
If I don't enable keep-alive the requests takes around 200ms as expected.

[nickkelsey]

Additional:
https, hence why the first request takes around 200ms.
Changing the read and write timeouts doesn't seem to make a difference.
Removing the poll define doesn't make a difference.

[nickkelsey]

Correction - reducing the read timeout to 3s results in subsequent requests taking 3s. Setting the timeout longer than 4-5s does not change the 4-5s behavior.

[nickkelsey]

Tracked down the exact version the problem was introduced...
0.18.1 works.
0.18.2 fails.

[yhirose]

@nickkelsey could you try this version?
#2000

Ref: #1969

[nickkelsey]

I see "std::this_thread::sleep_for(microseconds{interval_usec});" added, but no change to the problem unfortunately.

[nickkelsey]

If I revert the code in the two places where the new is_ssl_peer_could_be_closed() function is invoked everything works.

[yhirose]

Could you point the commit that starts causing the problem? Thanks!

[nickkelsey]

Commenting out the call to is_ssl_peer_could_be_closed() in ClientImpl::send_ avoids the problem.
Looks to be Commit 970b528 relating to Fix 1980.

[yhirose]

Thanks for your help! I'll look into it when I have time.