This is useful for some setups like swiotlb or VDUSE where the DMA
operations are expensive and/or need to be performed with a write lock.
After applying this patch, fio read test goes from 1124MiB/s to 1191MiB/s.
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
---
fs/fuse/virtio_fs.c | 40 ++++++++++++++++++++++++++++++++++++++--
1 file changed, 38 insertions(+), 2 deletions(-)
diff --git a/fs/fuse/virtio_fs.c b/fs/fuse/virtio_fs.c
index 1344c5782a7c..e19c78f2480e 100644
--- a/fs/fuse/virtio_fs.c
+++ b/fs/fuse/virtio_fs.c
@@ -836,8 +836,19 @@ static void virtio_fs_requests_done_work(struct work_struct *work)
/* End requests */
list_for_each_entry_safe(req, next, &reqs, list) {
+ unsigned int total_sgs = req->out_sgs + req->in_sgs;
+
list_del_init(&req->list);
+ for (unsigned int i = 0; i < total_sgs; ++i) {
+ enum dma_data_direction dir = (i < req->out_sgs) ?
+ DMA_TO_DEVICE : DMA_FROM_DEVICE;
+ dma_unmap_page(vq->vdev->dev.parent,
+ sg_dma_address(&req->sg[i]),
+ sg_dma_len(&req->sg[i]), dir);
+
+ }
+
/* blocking async request completes in a worker context */
if (req->args->may_block) {
struct virtio_fs_req_work *w;
@@ -1426,6 +1437,24 @@ static int virtio_fs_enqueue_req(struct virtio_fs_vq *fsvq,
sgs[i] = &req->sg[i];
WARN_ON(req->out_sgs + req->in_sgs != total_sgs);
+ // TODO can we change this ptr out of the lock?
+ vq = fsvq->vq;
+ // TODO handle this and following errors
+ for (i = 0; i < total_sgs; i++) {
+ struct page *page = sg_page(&req->sg[i]);
+ enum dma_data_direction dir = (i < req->out_sgs) ?
+ DMA_TO_DEVICE : DMA_FROM_DEVICE;
+ dma_addr_t dma_addr = dma_map_page(vq->vdev->dev.parent, page,
+ req->sg[i].offset, req->sg[i].length, dir);
+
+ if (dma_mapping_error(vq->vdev->dev.parent, dma_addr)) {
+ ret = -ENOMEM;
+ goto out;
+ }
+ sg_dma_address(&req->sg[i]) = dma_addr;
+ sg_dma_len(&req->sg[i]) = req->sg[i].length;
+ }
+
spin_lock(&fsvq->lock);
if (!fsvq->connected) {
@@ -1434,8 +1463,8 @@ static int virtio_fs_enqueue_req(struct virtio_fs_vq *fsvq,
goto out;
}
- vq = fsvq->vq;
- ret = virtqueue_add_sgs(vq, sgs, req->out_sgs, req->in_sgs, req, GFP_ATOMIC);
+ ret = virtqueue_add_sgs_premapped(vq, sgs, req->out_sgs,
+ req->in_sgs, req, GFP_ATOMIC);
if (ret < 0) {
spin_unlock(&fsvq->lock);
goto out;
@@ -1460,6 +1489,13 @@ static int virtio_fs_enqueue_req(struct virtio_fs_vq *fsvq,
virtqueue_notify(vq);
out:
+ for (unsigned int j = 0; ret && j < total_sgs; ++j) {
+ enum dma_data_direction dir = (j < req->out_sgs) ?
+ DMA_TO_DEVICE : DMA_FROM_DEVICE;
+ dma_unmap_page(vq->vdev->dev.parent,
+ sg_dma_address(&req->sg[j]),
+ sg_dma_len(&req->sg[j]), dir);
+ }
if (ret < 0 && req->argbuf) {
kfree(req->argbuf);
req->argbuf = NULL;
--
2.48.1
On Sat, Feb 22, 2025 at 1:07 AM Eugenio Pérez <eperezma@redhat.com> wrote:
>
> This is useful for some setups like swiotlb or VDUSE where the DMA
> operations are expensive and/or need to be performed with a write lock.
>
> After applying this patch, fio read test goes from 1124MiB/s to 1191MiB/s.
What FIO parameter have you used? It might be worth trying different
sizes. It seems to be more obvious when using larger requests when I'm
testing similar optimization for virtio-blk.
And we also need to test without VDUSE, to make sure no regression in
classical setups.
Thanks
>
> Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
> ---
> fs/fuse/virtio_fs.c | 40 ++++++++++++++++++++++++++++++++++++++--
> 1 file changed, 38 insertions(+), 2 deletions(-)
>
> diff --git a/fs/fuse/virtio_fs.c b/fs/fuse/virtio_fs.c
> index 1344c5782a7c..e19c78f2480e 100644
> --- a/fs/fuse/virtio_fs.c
> +++ b/fs/fuse/virtio_fs.c
> @@ -836,8 +836,19 @@ static void virtio_fs_requests_done_work(struct work_struct *work)
>
> /* End requests */
> list_for_each_entry_safe(req, next, &reqs, list) {
> + unsigned int total_sgs = req->out_sgs + req->in_sgs;
> +
> list_del_init(&req->list);
>
> + for (unsigned int i = 0; i < total_sgs; ++i) {
> + enum dma_data_direction dir = (i < req->out_sgs) ?
> + DMA_TO_DEVICE : DMA_FROM_DEVICE;
> + dma_unmap_page(vq->vdev->dev.parent,
> + sg_dma_address(&req->sg[i]),
> + sg_dma_len(&req->sg[i]), dir);
> +
> + }
> +
> /* blocking async request completes in a worker context */
> if (req->args->may_block) {
> struct virtio_fs_req_work *w;
> @@ -1426,6 +1437,24 @@ static int virtio_fs_enqueue_req(struct virtio_fs_vq *fsvq,
> sgs[i] = &req->sg[i];
> WARN_ON(req->out_sgs + req->in_sgs != total_sgs);
>
> + // TODO can we change this ptr out of the lock?
> + vq = fsvq->vq;
> + // TODO handle this and following errors
> + for (i = 0; i < total_sgs; i++) {
> + struct page *page = sg_page(&req->sg[i]);
> + enum dma_data_direction dir = (i < req->out_sgs) ?
> + DMA_TO_DEVICE : DMA_FROM_DEVICE;
> + dma_addr_t dma_addr = dma_map_page(vq->vdev->dev.parent, page,
> + req->sg[i].offset, req->sg[i].length, dir);
> +
> + if (dma_mapping_error(vq->vdev->dev.parent, dma_addr)) {
> + ret = -ENOMEM;
> + goto out;
> + }
> + sg_dma_address(&req->sg[i]) = dma_addr;
> + sg_dma_len(&req->sg[i]) = req->sg[i].length;
> + }
> +
> spin_lock(&fsvq->lock);
>
> if (!fsvq->connected) {
> @@ -1434,8 +1463,8 @@ static int virtio_fs_enqueue_req(struct virtio_fs_vq *fsvq,
> goto out;
> }
>
> - vq = fsvq->vq;
> - ret = virtqueue_add_sgs(vq, sgs, req->out_sgs, req->in_sgs, req, GFP_ATOMIC);
> + ret = virtqueue_add_sgs_premapped(vq, sgs, req->out_sgs,
> + req->in_sgs, req, GFP_ATOMIC);
> if (ret < 0) {
> spin_unlock(&fsvq->lock);
> goto out;
> @@ -1460,6 +1489,13 @@ static int virtio_fs_enqueue_req(struct virtio_fs_vq *fsvq,
> virtqueue_notify(vq);
>
> out:
> + for (unsigned int j = 0; ret && j < total_sgs; ++j) {
> + enum dma_data_direction dir = (j < req->out_sgs) ?
> + DMA_TO_DEVICE : DMA_FROM_DEVICE;
> + dma_unmap_page(vq->vdev->dev.parent,
> + sg_dma_address(&req->sg[j]),
> + sg_dma_len(&req->sg[j]), dir);
> + }
> if (ret < 0 && req->argbuf) {
> kfree(req->argbuf);
> req->argbuf = NULL;
> --
> 2.48.1
>
© 2016 - 2025 Red Hat, Inc.