在 2023/7/19 17:11, Michael S. Tsirkin 写道: > On Wed, Jul 19, 2023 at 03:53:45PM +0800, Hawkins Jiawei wrote: >> This patchset allows QEMU to delay polling and checking the device >> used buffer until either the SVQ is full or control commands shadow >> buffers are full, instead of polling and checking immediately after >> sending each SVQ control command, so that QEMU can send all the SVQ >> control commands in parallel, which have better performance improvement. >> >> I use vp_vdpa device to simulate vdpa device, and create 4094 VLANS in >> guest to build a test environment for sending multiple CVQ state load >> commands. This patch series can improve latency from 10023 us to >> 8697 us for about 4099 CVQ state load commands, about 0.32 us per command. > > Looks like a tiny improvement. > At the same time we have O(n^2) behaviour with memory mappings.
Hi Michael, Thanks for your review. I wonder why you say "we have O(n^2) behaviour on memory mappings" here? From my understanding, QEMU maps two page-size buffers as control commands shadow buffers at device startup. These buffers then are used to cache SVQ control commands, where QEMU fills them with multiple SVQ control commands bytes, flushes them when SVQ descriptors are full or these control commands shadow buffers reach their capacity. QEMU repeats this process until all CVQ state load commands have been sent in loading. In this loading process, only control commands shadow buffers translation should be relative to memory mappings, which should be O(log n) behaviour to my understanding(Please correct me if I am wrong). > Not saying we must not do this but I think it's worth > checking where the bottleneck is. My guess would be > vp_vdpa is not doing things in parallel. Want to try fixing that As for "vp_vdpa is not doing things in parallel.", do you mean the vp_vdpa device cannot process QEMU's SVQ control commands in parallel? In this situation, I will try to use real vdpa hardware to test the patch series performance. > to see how far it can be pushed? Currently, I am involved in the "Add virtio-net Control Virtqueue state restore support" project in Google Summer of Code now. Because I am uncertain about the time it will take to fix that problem in the vp_vdpa device, I prefer to complete the gsoc project first. Thanks! > > >> Note that this patch should be based on >> patch "Vhost-vdpa Shadow Virtqueue VLAN support" at [1]. >> >> [1]. https://lists.gnu.org/archive/html/qemu-devel/2023-07/msg03719.html >> >> TestStep >> ======== >> 1. regression testing using vp-vdpa device >> - For L0 guest, boot QEMU with two virtio-net-pci net device with >> `ctrl_vq`, `ctrl_rx`, `ctrl_rx_extra` features on, command line like: >> -device virtio-net-pci,disable-legacy=on,disable-modern=off, >> iommu_platform=on,mq=on,ctrl_vq=on,guest_announce=off, >> indirect_desc=off,queue_reset=off,ctrl_rx=on,ctrl_rx_extra=on,... >> >> - For L1 guest, apply the patch series and compile the source code, >> start QEMU with two vdpa device with svq mode on, enable the `ctrl_vq`, >> `ctrl_rx`, `ctrl_rx_extra` features on, command line like: >> -netdev type=vhost-vdpa,x-svq=true,... >> -device virtio-net-pci,mq=on,guest_announce=off,ctrl_vq=on, >> ctrl_rx=on,ctrl_rx_extra=on... >> >> - For L2 source guest, run the following bash command: >> ```bash >> #!/bin/sh >> >> for idx1 in {0..9} >> do >> for idx2 in {0..9} >> do >> for idx3 in {0..6} >> do >> ip link add macvlan$idx1$idx2$idx3 link eth0 >> address 4a:30:10:19:$idx1$idx2:1$idx3 type macvlan mode bridge >> ip link set macvlan$idx1$idx2$idx3 up >> done >> done >> done >> ``` >> - Execute the live migration in L2 source monitor >> >> - Result >> * with this series, QEMU should not trigger any error or warning. >> >> >> >> 2. perf using vp-vdpa device >> - For L0 guest, boot QEMU with two virtio-net-pci net device with >> `ctrl_vq`, `ctrl_vlan` features on, command line like: >> -device virtio-net-pci,disable-legacy=on,disable-modern=off, >> iommu_platform=on,mq=on,ctrl_vq=on,guest_announce=off, >> indirect_desc=off,queue_reset=off,ctrl_vlan=on,... >> >> - For L1 guest, apply the patch series, then apply an addtional >> patch to record the load time in microseconds as following: >> ```diff >> diff --git a/hw/net/vhost_net.c b/hw/net/vhost_net.c >> index 6b958d6363..501b510fd2 100644 >> --- a/hw/net/vhost_net.c >> +++ b/hw/net/vhost_net.c >> @@ -295,7 +295,10 @@ static int vhost_net_start_one(struct vhost_net *net, >> } >> >> if (net->nc->info->load) { >> + int64_t start_us = g_get_monotonic_time(); >> r = net->nc->info->load(net->nc); >> + error_report("vhost_vdpa_net_load() = %ld us", >> + g_get_monotonic_time() - start_us); >> if (r < 0) { >> goto fail; >> } >> ``` >> >> - For L1 guest, compile the code, and start QEMU with two vdpa device >> with svq mode on, enable the `ctrl_vq`, `ctrl_vlan` features on, >> command line like: >> -netdev type=vhost-vdpa,x-svq=true,... >> -device virtio-net-pci,mq=on,guest_announce=off,ctrl_vq=on, >> ctrl_vlan=on... >> >> - For L2 source guest, run the following bash command: >> ```bash >> #!/bin/sh >> >> for idx in {1..4094} >> do >> ip link add link eth0 name vlan$idx type vlan id $idx >> done >> ``` >> >> - wait for some time, then execute the live migration in L2 source monitor >> >> - Result >> * with this series, QEMU should not trigger any warning >> or error except something like "vhost_vdpa_net_load() = 8697 us" >> * without this series, QEMU should not trigger any warning >> or error except something like "vhost_vdpa_net_load() = 10023 us" >> >> ChangeLog >> ========= >> v3: >> - refactor vhost_svq_poll() to accept cmds_in_flight >> suggested by Jason and Eugenio >> - refactor vhost_vdpa_net_cvq_add() to make control commands buffers >> is not tied to `s->cvq_cmd_out_buffer` and `s->status`, so we can reuse >> it suggested by Eugenio >> - poll and check when SVQ is full or control commands shadow buffers is >> full >> >> v2: https://lore.kernel.org/all/cover.1683371965.git.yin31...@gmail.com/ >> - recover accidentally deleted rows >> - remove extra newline >> - refactor `need_poll_len` to `cmds_in_flight` >> - return -EINVAL when vhost_svq_poll() return 0 or check >> on buffers written by device fails >> - change the type of `in_cursor`, and refactor the >> code for updating cursor >> - return directly when vhost_vdpa_net_load_{mac,mq}() >> returns a failure in vhost_vdpa_net_load() >> >> v1: https://lore.kernel.org/all/cover.1681732982.git.yin31...@gmail.com/ >> >> Hawkins Jiawei (8): >> vhost: Add argument to vhost_svq_poll() >> vdpa: Use iovec for vhost_vdpa_net_cvq_add() >> vhost: Expose vhost_svq_available_slots() >> vdpa: Avoid using vhost_vdpa_net_load_*() outside >> vhost_vdpa_net_load() >> vdpa: Check device ack in vhost_vdpa_net_load_rx_mode() >> vdpa: Move vhost_svq_poll() to the caller of vhost_vdpa_net_cvq_add() >> vdpa: Add cursors to vhost_vdpa_net_loadx() >> vdpa: Send cvq state load commands in parallel >> >> hw/virtio/vhost-shadow-virtqueue.c | 38 ++-- >> hw/virtio/vhost-shadow-virtqueue.h | 3 +- >> net/vhost-vdpa.c | 354 ++++++++++++++++++----------- >> 3 files changed, 249 insertions(+), 146 deletions(-) >> >> -- >> 2.25.1 >
