Jason Wang
2022-Jan-10 07:41 UTC
[PATCH 2/6] virtio: split: alloc indirect desc with extra
On Mon, Jan 10, 2022 at 3:24 PM Xuan Zhuo <xuanzhuo at linux.alibaba.com> wrote:> > On Mon, 10 Jan 2022 14:43:39 +0800, Jason Wang <jasowang at redhat.com> wrote: > > > > ? 2022/1/7 ??2:33, Xuan Zhuo ??: > > > In the scenario where indirect is not used, each desc corresponds to an > > > extra, which is used to record information such as dma, flags, and > > > next. > > > > > > In the scenario of using indirect, the assigned desc does not have the > > > corresponding extra record dma information, and the dma information must > > > be obtained from the desc when unmap. > > > > > > This patch allocates the corresponding extra array when indirect desc is > > > allocated. This has these advantages: > > > 1. Record the dma information of desc, no need to read desc when unmap > > > 2. It will be more convenient and unified in processing > > > 3. Some additional information can be recorded in extra, which will be > > > used in subsequent patches. > > > > > > Two questions: > > > > 1) Is there any performance number for this change? I guess it gives > > more stress on the cache. > > I will add performance test data in the next version. > > > 2) Is there a requirement to mix the pre mapped sg with unmapped sg? If > > not, a per virtqueue flag looks sufficient > > There is this requirement. For example, in the case of AF_XDP, a patcket > contains two parts, one is virtio_net_hdr, and the other is the actual data > packet from AF_XDP. The former is unmapped sg, and the latter is pre mapped sg.Any chance to map virtio_net_hdr() manually by AF_XDP routine in this case? Thanks> > Thanks. > > > > > Thanks > > > > >
On Mon, 10 Jan 2022 15:41:27 +0800, Jason Wang <jasowang at redhat.com> wrote:> On Mon, Jan 10, 2022 at 3:24 PM Xuan Zhuo <xuanzhuo at linux.alibaba.com> wrote: > > > > On Mon, 10 Jan 2022 14:43:39 +0800, Jason Wang <jasowang at redhat.com> wrote: > > > > > > ? 2022/1/7 ??2:33, Xuan Zhuo ??: > > > > In the scenario where indirect is not used, each desc corresponds to an > > > > extra, which is used to record information such as dma, flags, and > > > > next. > > > > > > > > In the scenario of using indirect, the assigned desc does not have the > > > > corresponding extra record dma information, and the dma information must > > > > be obtained from the desc when unmap. > > > > > > > > This patch allocates the corresponding extra array when indirect desc is > > > > allocated. This has these advantages: > > > > 1. Record the dma information of desc, no need to read desc when unmap > > > > 2. It will be more convenient and unified in processing > > > > 3. Some additional information can be recorded in extra, which will be > > > > used in subsequent patches. > > > > > > > > > Two questions: > > > > > > 1) Is there any performance number for this change? I guess it gives > > > more stress on the cache. > > > > I will add performance test data in the next version. > > > > > 2) Is there a requirement to mix the pre mapped sg with unmapped sg? If > > > not, a per virtqueue flag looks sufficient > > > > There is this requirement. For example, in the case of AF_XDP, a patcket > > contains two parts, one is virtio_net_hdr, and the other is the actual data > > packet from AF_XDP. The former is unmapped sg, and the latter is pre mapped sg. > > Any chance to map virtio_net_hdr() manually by AF_XDP routine in this case?Well, it is indeed possible to do so. In the indirect scenario, we can record it in vring->split.desc_extra[head].flags Then we have to agree that there can be no mixed situation. Thanks.> > Thanks > > > > > Thanks. > > > > > > > > Thanks > > > > > > > > >