On Mon, Sep 18, 2023 at 01:13:34PM +0200, Christian Weisgerber via curl-library wrote: > Dan Fandrich: > > I wanted to try this patch on NetBSD to see if it's related > > to Nagle's algorithm, but couldn't get to the point where I could try it: > > > -http://%HOSTIP:%HTTPPORT/we/want/%TESTNUMBER -T > > %LOGDIR/test%TESTNUMBER.txt --limit-rate 64K --expect100-timeout 0.001 > > +http://%HOSTIP:%HTTPPORT/we/want/%TESTNUMBER -T > > %LOGDIR/test%TESTNUMBER.txt --limit-rate 64K --expect100-timeout 0.001 > > --tcp-nodelay > > This makes no difference, the test fails the same way.
After much pain I was finally able to install OpenBSD under KVM and reproduce this problem. I verified that curl was correctly sending the first 65536 bytes of data to the socket in a call to send(2), but the OS reports that only 32716 bytes were actually sent. The test assumes that the OS will send all 65536 bytes, so the test fails. I tried adjusting all the relevant sysctl parameters I could find, and couldn't change this behaviour. Typically, pattern is the first three 64 KiB sends only actually send 32716 bytes each, then the next couple send 65432 bytes, then finally the kernel sends all 65536 bytes, as requested (with the occasional 65432 one). I also tried using write(2) instead of send(2) with no effect. The OS is free to do this of course, but the test depends on the OS sending all the data at once in order to set up the specific conditions needed to make the test work. I don't see any reasonable alternative to just disabling this test on NetBSD and OpenBSD. Dan -- Unsubscribe: https://lists.haxx.se/mailman/listinfo/curl-library Etiquette: https://curl.se/mail/etiquette.html
