Merge pull request #295 from ska-sa/send-gso

Speed up sending with generic segmentation offload

.ci/all-builds.sh

@@ -17,4 +17,11 @@ for python in true false; do
             done
         done
     done
+    # sendmmsg and gso don't interact with ibv, so don't test them jointly
+    for sendmmsg in auto disabled; do
+        for gso in auto disabled; do
+            meson configure -Dpython=$python -Dsendmmsg=$sendmmsg -Dgso=$gso
+            meson compile
+        done
+    done
 done

doc/dev-send-gso.rst

@@ -0,0 +1,51 @@
+Generic segmentation offload
+============================
+Linux supports a mechanism called :dfn:`generic segmentation offload` (GSO) to
+reduce packet overheads when transmitting UDP data through the kernel
+networking stack. A good overview can be found on `Cloudflare's blog`_, but
+the basic idea is this:
+
+.. _Cloudflare's blog: https://blog.cloudflare.com/accelerating-udp-packet-transmission-for-quic/
+
+1. Userspace concatenates multiple smaller packets into one mega-packet for
+   submission to the kernel.
+2. Most of the networking stack operates on the mega-packet.
+3. As late as possible (and possibly on the NIC) the mega-packet is
+   re-segmented into the original packets.
+
+The re-segmentation uses a user-supplied parameter (socket option) indicating
+the size of the original packets. This imposes a limitation that the original
+packets were all the same size, except perhaps for the last one in the
+mega-packet.
+
+The support for this in spead2 is dependent on the :manpage:`sendmmsg(2)`
+support. While there is no fundamental reason GSO can't be used without
+:manpage:`sendmmsg(2)`, supporting it would complicate the code significantly,
+and GSO is a much more recent feature so it is unlikely that this combination
+would ever be needed.
+
+Run-time detection of support is unfortunately rather complicated. The simple
+part is that an older kernel will not support the socket option. If that
+occurs, we simply disable GSO for the stream. A more tricky problem is that
+actually sending the message may fail for several reasons:
+
+- Fragmentation doesn't seem to be supported, so if the segment size is bigger
+  than the MTU, it will fail.
+- If hardware checksumming is disabled (or presumably if it is not supported),
+  it will fail.
+
+To cope with this complication, a state machine is used. It has four possible
+states:
+
+- **active**: the socket option is set to a positive value
+- **inactive**: the socket option is set to zero, but we may still transition
+  to active
+- **probe**: the last send in active state failed; the socket option is now
+  set to zero and we're retrying
+- **disabled**: the socket option is set to zero, and we will never try to set
+  it again.
+
+If send fails while in state **active**, we switch to state **probe** and try
+again (without GSO). If that succeeds, we conclude that GSO is non-functional
+for this stream and permanently go to **disabled**. If that also fails, we
+conclude that the problem was unrelated to GSO and return to **inactive**.

doc/developer.rst

@@ -27,4 +27,5 @@ the C++ classes to extend functionality.
    dev-recv-destruction
    dev-recv-chunk-group
    dev-send-rate-limit
+   dev-send-gso
    dev-ibverbs-linking

include/spead2/common_features.h.in

@@ -33,6 +33,7 @@
 #define SPEAD2_USE_MLX5DV @SPEAD2_USE_MLX5DV@
 #define SPEAD2_USE_RECVMMSG @SPEAD2_USE_RECVMMSG@
 #define SPEAD2_USE_SENDMMSG @SPEAD2_USE_SENDMMSG@
+#define SPEAD2_USE_GSO @SPEAD2_USE_GSO@
 #define SPEAD2_USE_EVENTFD @SPEAD2_USE_EVENTFD@
 #define SPEAD2_USE_PTHREAD_SETAFFINITY_NP @SPEAD2_USE_PTHREAD_SETAFFINITY_NP@
 #define SPEAD2_USE_FMV @SPEAD2_USE_FMV@

meson.build

@@ -160,6 +160,13 @@ use_sendmmsg = get_option('sendmmsg').require(
     prefix : '#include <sys/socket.h>'
   )
 ).allowed()
+use_gso = get_option('gso').require(
+  compiler.get_define(
+    'UDP_SEGMENT',
+    args : '-D_GNU_SOURCE',
+    prefix : '#include <netinet/udp.h>'
+  ) != ''
+).allowed()
 use_eventfd = get_option('eventfd').require(
   compiler.has_function(
     'eventfd',
@@ -260,6 +267,7 @@ conf.set10('SPEAD2_USE_IBV_HW_RATE_LIMIT', use_ibv_hw_rate_limit)
 conf.set10('SPEAD2_USE_MLX5DV', mlx5_dep.found())
 conf.set10('SPEAD2_USE_RECVMMSG', use_recvmmsg)
 conf.set10('SPEAD2_USE_SENDMMSG', use_sendmmsg)
+conf.set10('SPEAD2_USE_GSO', use_gso)
 conf.set10('SPEAD2_USE_EVENTFD', use_eventfd)
 conf.set10('SPEAD2_USE_POSIX_SEMAPHORES', use_posix_semaphores)
 conf.set10('SPEAD2_USE_PTHREAD_SETAFFINITY_NP', use_pthread_setaffinity_np)

meson.options

@@ -21,6 +21,7 @@ option('pcap', type : 'feature', description : 'Support reading from pcap files'
 option('cap', type : 'feature', description : 'Use libcap')
 option('recvmmsg', type : 'feature', description : 'Use recvmmsg system call')
 option('sendmmsg', type : 'feature', description : 'Use sendmmsg system call')
+option('gso', type : 'feature', description : 'Use generic segmentation offload')
 option('eventfd', type : 'feature', description : 'Use eventfd system call for semaphores')
 option('posix_semaphores', type : 'feature', description : 'Use POSIX semaphores')
 option('pthread_setaffinity_np', type : 'feature', description : 'Use pthread_setaffinity_np to set thread affinity')