On Tue, Dec 20, 2022 at 07:16:38AM +0000, Arseniy Krasnov wrote:
Patchset consists of two parts: 1) Kernel patch One patch from Bobby Eshleman. I took single patch from Bobby: https://lore.kernel.org/lkml/d81818b868216c774613dd03641fcfe63cc55a45 .1660362668.git.bobby.eshleman@xxxxxxxxxxxxx/ and use only part for af_vsock.c, as VMCI and Hyper-V parts were rejected. I used it, because for SOCK_SEQPACKET big messages handling was broken - ENOMEM was returned instead of EMSGSIZE. And anyway, current logic which always replaces any error code returned by transport to ENOMEM looks strange for me also(for example in EMSGSIZE case it was changed to ENOMEM). 2) Tool patches Since there is work on several significant updates for vsock(virtio/ vsock especially): skbuff, DGRAM, zerocopy rx/tx, so I think that this patchset will be useful. This patchset updates vsock tests and tools a little bit. First of all it updates test suite: two new tests are added. One test is reworked message bound test. Now it is more complex. Instead of sending 1 byte messages with one MSG_EOR bit, it sends messages of random length(one half of messages are smaller than page size, second half are bigger) with random number of MSG_EOR bits set. Receiver also don't know total number of messages. Message bounds control is maintained by hash sum of messages length calculation. Second test is for SOCK_SEQPACKET - it tries to send message with length more than allowed. I think both tests will be useful for DGRAM support also. Third thing that this patchset adds is small utility to test vsock performance for both rx and tx. I think this util could be useful as 'iperf'/'uperf', because: 1) It is small comparing to 'iperf' or 'uperf', so it very easy to add new mode or feature to it(especially vsock specific). 2) It allows to set SO_RCVLOWAT and SO_VM_SOCKETS_BUFFER_SIZE option. Whole throughtput depends on both parameters. 3) It is located in the kernel source tree, so it could be updated by the same patchset which changes related kernel functionality in vsock. I used this util very often to check performance of my rx zerocopy support(this tool has rx zerocopy support, but not in this patchset). Here is comparison of outputs from three utils: 'iperf', 'uperf' and 'vsock_perf'. In all three cases sender was at guest side. rx and tx buffers were always 64Kb(because by default 'uperf' uses 8K). iperf: [ ID] Interval Transfer Bitrate [ 5] 0.00-10.00 sec 12.8 GBytes 11.0 Gbits/sec sender [ 5] 0.00-10.00 sec 12.8 GBytes 11.0 Gbits/sec receiver uperf: Total 16.27GB / 11.36(s) = 12.30Gb/s 23455op/s vsock_perf: tx performance: 12.301529 Gbits/s rx performance: 12.288011 Gbits/s Results are almost same in all three cases.
Thanks for checking this!
Patchset was rebased and tested on skbuff v8 patch from Bobby Eshleman: https://lore.kernel.org/netdev/20221215043645.3545127-1-bobby.eshleman@xxxxxxxxxxxxx/
I reviewed all the patches, in the last one there is just to update the README, so I think it is ready for net-next (when it will re-open).
Thanks, Stefano _______________________________________________ Virtualization mailing list Virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linuxfoundation.org/mailman/listinfo/virtualization
