On Wed, May 01, 2024 at 05:38:15PM +0200, Peter-Jan Gootzen wrote: > In this patch set we are adding multi-queue support in the virtio-fs > driver to enhance performance. > While multi-queue functionality has been in the virtio-fs specification > since its inception, the current implementation in the Linux kernel > creates multiple queues, but only enqueues requests on the first request > queue. > > The first commit in this patch set limits the number of request queues > to the number of CPUs in the system, to reduce unnecessary resource > consumption on the device-side. > The second commit adds a multi queue mapping to the virtio-fs device > driver that is created using MSI-X affinity mappings or, in case of no > existing mappings, by evenly grouping the CPUs and dividing the queues > over these groups. > > Future work in this direction could include: > - CPU hotplug support It looks like this series is safe to merge as-is. Hotplugged CPUs default to the first request queue. It's safe because mq_map[] is sized to nr_cpu_ids (max number of possible CPUs) and mq_map[] is initialized to 0 (first request queue). Hotplug can be added later. > - In case of no MSI-X, using the created multi-queue mapping to set > interrupt affinities > > See the commit message of the second patch for performance results. > > Peter-Jan Gootzen (2): > virtio-fs: limit number of request queues > virtio-fs: add multi-queue support > > fs/fuse/virtio_fs.c | 73 ++++++++++++++++++++++++++++++++++++++++----- > 1 file changed, 65 insertions(+), 8 deletions(-) Reviewed-by: Stefan Hajnoczi <stefanha@xxxxxxxxxx>
Attachment:
signature.asc
Description: PGP signature
