The following patches apply over linus's tree and the user_worker patchset here: https://lore.kernel.org/virtualization/20211129194707.5863-1-michael.christie@xxxxxxxxxx/T/#t which allows us to check the vhost owner thread's RLIMITs, and they are built over Andrey's flush cleanups: https://lore.kernel.org/virtualization/20211207024510.23292-1-michael.christie@xxxxxxxxxx/T/#t The patches allow us to support multiple vhost workers per device. The design is a modified version of Stefan's original idea where userspace has the kernel create a worker and we pass back the pid. In this version, V4, instead of passing the pid between user/kernel space we use a worker_id which is just an integer managed by the vhost driver and we allow userspace to create and free workers and then attach them to virtqueues at setup time or while IO is running. All review comments from the past reviews should be handled. If I didn't reply to a review comment, I agreed with the comment and should have handled it in this posting. Let me know if I missed one. Results: -------- fio jobs 1 2 4 8 12 16 ---------------------------------------------------------- 1 worker 84k 492k 510k - - - worker per vq 184k 380k 744k 1422k 2256k 2434k Notes: 0. This used a simple fio command: fio --filename=/dev/sdb --direct=1 --rw=randrw --bs=4k \ --ioengine=libaio --iodepth=128 --numjobs=$JOBS_ABOVE and I used a VM with 16 vCPUs and 16 virtqueues. 1. The patches were tested with emulate_pr=0 and these patches: https://lore.kernel.org/all/yq1tuhge4bg.fsf@xxxxxxxxxxxxxxxxxxxx/t/ which are in mkp's scsi branches for the next kernel. They fix the perf issues where IOPs dropped at 12 vqs/jobs. 2. Because we have a hard limit of 1024 cmds, if the num jobs * iodepth was greater than 1024, I would decrease iodepth. So 12 jobs used 85 cmds, and 16 used 64. 3. The perf issue above at 2 jobs is because when we only have 1 worker we execute more cmds per vhost_work due to all vqs funneling to one worker. This results in less context switches and better batching without having to tweak any settings. I'm working on patches to add back batching during lio completion and do polling on the submission side. We will still want the threading patches, because if we batch at the fio level plus use the vhost theading patches, we can see a big boost like below. So hopefully doing it at the kernel will allow apps to just work without having to be smart like fio. fio using io_uring and batching with the iodepth_batch* settings: fio jobs 1 2 4 8 12 16 ------------------------------------------------------------- 1 worker 494k 520k - - - - worker per vq 496k 878k 1542k 2436k 2304k 2590k V5: - Rebase against user_worker patchset. - Rebase against flush patchset. - Redo vhost-scsi tmf flush handling so it doesn't access vq->worker. V4: - fix vhost-sock VSOCK_VQ_RX use. - name functions called directly by ioctl cmd's to match the ioctl cmd. - break up VHOST_SET_VRING_WORKER into a new, free and attach cmd. - document worker lifetime, and cgroup, namespace, mm, rlimit inheritance, make it clear we currently only support sharing within the device. - add support to attach workers while IO is running. - instead of passing a pid_t of the kernel thread, pass a int allocated by the vhost layer with an idr. V3: - fully convert vhost code to use vq based APIs instead of leaving it half per dev and half per vq. - rebase against kernel worker API. - Drop delayed worker creation. We always create the default worker at VHOST_SET_OWNER time. Userspace can create and bind workers after that. V2: - change loop that we take a refcount to the worker in - replaced pid == -1 with define. - fixed tabbing/spacing coding style issue - use hash instead of list to lookup workers. - I dropped the patch that added an ioctl cmd to get a vq's worker's pid. I saw we might do a generic netlink interface instead.