search for: virtqueue_map_sgs

..._vring_split() to prepare for subsequent support for premapped. DMA address will be saved as sg->dma_address, then virtqueue_add_vring_split() will use it directly. If it is a premapped scene, the transmitted sgs should have saved DMA address in dma_address, and in virtio core, we need to pass virtqueue_map_sgs(). Signed-off-by: Xuan Zhuo <xuanzhuo at linux.alibaba.com> --- drivers/virtio/virtio_ring.c | 116 ++++++++++++++++++++++++----------- 1 file changed, 80 insertions(+), 36 deletions(-) diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c index 3005893ecc61..17520f0d764...

...logic is separated from the virtqueue_add_split to prepare > for subsequent support for premapped. The patch seems to do more than what is described here. To simplify reviewers, I'd suggest to split this patch into three: 1) virtqueue_add_split_prepare() (could we have a better name?) 2) virtqueue_map_sgs() 3) virtqueue_add_split_vring() (Or only factor DMA parts out, I haven't gone through the reset of the patches) Thanks > > Signed-off-by: Xuan Zhuo <xuanzhuo at linux.alibaba.com> > --- > drivers/virtio/virtio_ring.c | 219 ++++++++++++++++++++++++----------- > 1 file...

## About DMA APIs Now, virtio may can not work with DMA APIs when virtio features do not have VIRTIO_F_ACCESS_PLATFORM. 1. I tried to let DMA APIs return phy address by virtio-device. But DMA APIs just work with the "real" devices. 2. I tried to let xsk support callballs to get phy address from virtio-net driver as the dma address. But the maintainers of xsk may want to use

...*_vq, struct scatterlist *sg; struct vring_desc *desc; unsigned int i, n, avail, descs_used, prev; - bool indirect; + bool indirect, dma_map_internal; int head; START_USE(vq); @@ -668,7 +672,8 @@ static inline int virtqueue_add_split(struct virtqueue *_vq, return -ENOSPC; } - if (virtqueue_map_sgs(vq, sgs, total_sg, out_sgs, in_sgs)) + dma_map_internal = !sgs[0]->dma_address; + if (dma_map_internal && virtqueue_map_sgs(vq, sgs, total_sg, out_sgs, in_sgs)) return -ENOMEM; for (n = 0; n < out_sgs; n++) { @@ -734,6 +739,7 @@ static inline int virtqueue_add_split(struct virt...

XDP socket(AF_XDP) is an excellent bypass kernel network framework. The zero copy feature of xsk (XDP socket) needs to be supported by the driver. The performance of zero copy is very good. ENV: Qemu with vhost. vhost cpu | Guest APP CPU |Guest Softirq CPU | PPS -----------------------------|---------------|------------------|------------ xmit by sockperf: 90% | 100%

XDP socket(AF_XDP) is an excellent bypass kernel network framework. The zero copy feature of xsk (XDP socket) needs to be supported by the driver. The performance of zero copy is very good. ENV: Qemu with vhost. vhost cpu | Guest APP CPU |Guest Softirq CPU | PPS -----------------------------|---------------|------------------|------------ xmit by sockperf: 90% | 100%

...esc *desc; > unsigned int i, n, avail, descs_used, prev; > - bool indirect; > + bool indirect, map_inter; > int head; > > START_USE(vq); > @@ -668,7 +672,8 @@ static inline int virtqueue_add_split(struct virtqueue *_vq, > return -ENOSPC; > } > > - if (virtqueue_map_sgs(vq, sgs, total_sg, out_sgs, in_sgs)) > + map_inter = !sgs[0]->dma_address; > + if (map_inter && virtqueue_map_sgs(vq, sgs, total_sg, out_sgs, in_sgs)) > return -ENOMEM; > > for (n = 0; n < out_sgs; n++) { > @@ -734,6 +739,7 @@ static inline int virtqueue_add_s...

XDP socket(AF_XDP) is an excellent bypass kernel network framework. The zero copy feature of xsk (XDP socket) needs to be supported by the driver. The performance of zero copy is very good. ENV: Qemu with vhost. vhost cpu | Guest APP CPU |Guest Softirq CPU | PPS -----------------------------|---------------|------------------|------------ xmit by sockperf: 90% | 100%

XDP socket(AF_XDP) is an excellent bypass kernel network framework. The zero copy feature of xsk (XDP socket) needs to be supported by the driver. The performance of zero copy is very good. ENV: Qemu with vhost. vhost cpu | Guest APP CPU |Guest Softirq CPU | PPS -----------------------------|---------------|------------------|------------ xmit by sockperf: 90% | 100%

...>length, DMA_TO_DEVICE); + } + } + + for (; n < (out_sgs + in_sgs); n++) { + for (sg = sgs[n]; sg; sg = sg_next(sg)) { + if (!sg->dma_address) + return; + + dma_unmap_page(vring_dma_dev(vq), sg->dma_address, + sg->length, DMA_FROM_DEVICE); + } + } +} + +static int virtqueue_map_sgs(struct vring_virtqueue *vq, + struct scatterlist *sgs[], + unsigned int total_sg, + unsigned int out_sgs, + unsigned int in_sgs) +{ + struct scatterlist *sg; + unsigned int n; + + if (!vq->use_dma_api) + return 0; + + for (n = 0; n < out_sgs; n++) { + for (sg = s...

...ring_packed() to prepare for subsequent support for premapped. DMA address will be saved as sg->dma_address, then virtqueue_add_vring_packed() will use it directly. If it is a premapped scene, the transmitted sgs should have saved DMA address in dma_address, and in virtio core, we need to pass virtqueue_map_sgs(). Signed-off-by: Xuan Zhuo <xuanzhuo at linux.alibaba.com> --- drivers/virtio/virtio_ring.c | 59 ++++++++---------------------------- 1 file changed, 12 insertions(+), 47 deletions(-) diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c index 6796cbee0207..81a9c0692fe...

## About DMA APIs Now, virtio may can not work with DMA APIs when virtio features do not have VIRTIO_F_ACCESS_PLATFORM. 1. I tried to let DMA APIs return phy address by virtio-device. But DMA APIs just work with the "real" devices. 2. I tried to let xsk support callballs to get phy address from virtio-net driver as the dma address. But the maintainers of xsk may want to use

...bool dma_map, > gfp_t gfp) > { > struct vring_virtqueue *vq = to_vvq(_vq); > @@ -767,9 +772,11 @@ static inline int virtqueue_add_split(struct virtqueue *_vq, > goto end; > } > > - err = virtqueue_map_sgs(vq, sgs, total_sg, out_sgs, in_sgs); > - if (err) > - goto err; > + if (dma_map) { Could we simply check sg->dma_addr in this case? Then we don't need to introduce the dma_map flag. Thanks > + err = virtqueue_map_sgs(vq, sgs, total_sg,...

...gfp_t gfp) > > { > > struct vring_virtqueue *vq = to_vvq(_vq); > > @@ -767,9 +772,11 @@ static inline int virtqueue_add_split(struct virtqueue *_vq, > > goto end; > > } > > > > - err = virtqueue_map_sgs(vq, sgs, total_sg, out_sgs, in_sgs); > > - if (err) > > - goto err; > > + if (dma_map) { > > Could we simply check sg->dma_addr in this case? Then we don't need to > introduce the dma_map flag. > > Thanks > > > > +...

XDP socket(AF_XDP) is an excellent bypass kernel network framework. The zero copy feature of xsk (XDP socket) needs to be supported by the driver. The performance of zero copy is very good. ENV: Qemu with vhost. vhost cpu | Guest APP CPU |Guest Softirq CPU | PPS -----------------------------|---------------|------------------|------------ xmit by sockperf: 90% | 100%

...>length, DMA_TO_DEVICE); + } + } + + for (; n < (out_sgs + in_sgs); n++) { + for (sg = sgs[n]; sg; sg = sg_next(sg)) { + if (!sg->dma_address) + return; + + dma_unmap_page(vring_dma_dev(vq), sg->dma_address, + sg->length, DMA_FROM_DEVICE); + } + } +} + +static int virtqueue_map_sgs(struct vring_virtqueue *vq, + struct scatterlist *sgs[], + unsigned int total_sg, + unsigned int out_sgs, + unsigned int in_sgs) +{ + struct scatterlist *sg; + unsigned int n; + + if (!vq->use_dma_api) + return 0; + + for (n = 0; n < out_sgs; n++) { + for (sg = s...

...>length, DMA_TO_DEVICE); + } + } + + for (; n < (out_sgs + in_sgs); n++) { + for (sg = sgs[n]; sg; sg = sg_next(sg)) { + if (!sg->dma_address) + return; + + dma_unmap_page(vring_dma_dev(vq), sg->dma_address, + sg->length, DMA_FROM_DEVICE); + } + } +} + +static int virtqueue_map_sgs(struct vring_virtqueue *vq, + struct scatterlist *sgs[], + unsigned int total_sg, + unsigned int out_sgs, + unsigned int in_sgs) +{ + struct scatterlist *sg; + unsigned int n; + + if (!vq->use_dma_api) + return 0; + + for (n = 0; n < out_sgs; n++) { + for (sg = s...

I figured out the problem. There is zerocopy IO is being done via DMA to a buffer allocated with valloc(). Right now, I am running a hack-fix locally so I can get some other stuff done first. I will propose a proper fix to the list in a few days. On 11/25/2013 10:49 AM, Richard Yao wrote: > I booted a Gentoo Linux installation in QEMU with a 9P rootfs as follows: > > sudo qemu-kvm -cpu

...- avail_used_flags = vq->packed.avail_used_flags; WARN_ON_ONCE(total_sg > vq->packed.vring.num && !vq->indirect); @@ -1479,15 +1478,13 @@ static inline int virtqueue_add_packed(struct virtqueue *_vq, id = vq->free_head; BUG_ON(id == vq->packed.vring.num); + if (virtqueue_map_sgs(vq, sgs, total_sg, out_sgs, in_sgs)) + return -ENOMEM; + curr = id; c = 0; for (n = 0; n < out_sgs + in_sgs; n++) { for (sg = sgs[n]; sg; sg = sg_next(sg)) { - dma_addr_t addr = vring_map_one_sg(vq, sg, n < out_sgs ? - DMA_TO_DEVICE : DMA_FROM_DEVICE); - if (vring_mapping_er...

As said before, please don't try to do weird runtime checks based on the scatterlist. What you have works for now, but there are plans to repalce the page + offset tuple in the scatterlist with just a phys_addr_t. And with that your "clever" scheme will break instantly.