Enable UDP GRO for QUIC

[tatsuhiro-t]

No description provided.

[bagder]

Nice. @tatsuhiro-t, have done any measurements to see how this changes performance?

[tatsuhiro-t]

I did some experiments on loopback. Both curl and server did not fully utilize CPU, and 4GiB transfer took ~11 seconds without GRO and ~10 seconds with GRO. So it does make a difference, but YMMV.

[bagder]

The linux-musl-llvm build gets this compiler error:

In file included from /home/runner/work/curl/curl/curl/_x64-linux-musl-bld/lib/CMakeFiles/libcurl_object.dir/Unity/unity_0_c.c:496:
/home/runner/work/curl/curl/curl/lib/vquic/vquic.c:339:47: error: cast from 'unsigned char *' to 'struct cmsghdr *' increases required alignment from 1 to 4 [-Werror,-Wcast-align]
  339 |   for(cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
      |                                               ^~~~~~~~~~~~~~~~~~~~~~
/usr/include/x86_64-linux-musl/sys/socket.h:359:8: note: expanded from macro 'CMSG_NXTHDR'
  359 |         ? 0 : (struct cmsghdr *)__CMSG_NEXT(cmsg))
      |               ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In file included from /home/runner/work/curl/curl/curl/_x64-linux-musl-bld/lib/CMakeFiles/libcurl_object.dir/Unity/unity_0_c.c:496:
/home/runner/work/curl/curl/curl/lib/vquic/vquic.c:339:47: error: comparison of integers of different signs: 'unsigned long' and 'long' [-Werror,-Wsign-compare]
  339 |   for(cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
      |                                               ^~~~~~~~~~~~~~~~~~~~~~
/usr/include/x86_64-linux-musl/sys/socket.h:358:44: note: expanded from macro 'CMSG_NXTHDR'
  358 |         __CMSG_LEN(cmsg) + sizeof(struct cmsghdr) >= __MHDR_END(mhdr) - (unsigned char *)(cmsg) \
      |         ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2 errors generated.

[tatsuhiro-t]

I think it needs (void*) casts somewhere to shut up these warnings.

[tatsuhiro-t]

Well, this is famous musl issue, and the only simple workaround is disable compiler warnings for this portion of the code: see https://git.openembedded.org/meta-openembedded/tree/meta-oe/recipes-devtools/breakpad/breakpad/0001-Turn-off-sign-compare-for-musl-libc.patch

[tatsuhiro-t]

GRO is only enabled for ngtcp2+nghttp3 and quiche. The others do their own I/O, and enabling GRO here causes issues unless they are prepared for it.

[bagder]

Basic speed test with ngtcp2 with and without this PR. Server: caddy. Running on localhost.

tldr: the parallel test doing 50 times 100MB downloads runs 39% faster.

without

$ python3 scorecard.py --caddy -d --download=100mb h3
  Server       Size  single(1x1) [cpu/rss]         serial(50x1) [cpu/rss]         parallel(50x50) [cpu/rss]                Errors       
  caddy       100MB     554 MB/s [67.1%/14MB]          574 MB/s [71.5%/16MB]          509 MB/s [80.0%/16MB]               -          

with

$ python3 scorecard.py --caddy -d --download=100mb h3
  Server       Size  single(1x1) [cpu/rss]         serial(50x1) [cpu/rss]         parallel(50x50) [cpu/rss]                Errors       
  caddy       100MB     680 MB/s [57.3%/14MB]          701 MB/s [61.9%/15MB]          710 MB/s [81.9%/17MB]               -          

[bagder]

Thanks!