| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <20211031104700-mutt-send-email-mst@kernel.org>
Date: Sun, 31 Oct 2021 10:47:34 -0400
From: "Michael S. Tsirkin" <mst@...hat.com>
To: Xuan Zhuo <xuanzhuo@...ux.alibaba.com>
Cc: Jason Wang <jasowang@...hat.com>,
virtualization <virtualization@...ts.linux-foundation.org>,
netdev <netdev@...r.kernel.org>,
"David S. Miller" <davem@...emloft.net>,
Jakub Kicinski <kuba@...nel.org>
Subject: Re: [PATCH 1/3] virtio: cache indirect desc for split
On Thu, Oct 28, 2021 at 02:16:03PM +0800, Xuan Zhuo wrote:
> On Thu, 28 Oct 2021 10:16:10 +0800, Jason Wang <jasowang@...hat.com> wrote:
> > On Thu, Oct 28, 2021 at 1:07 AM Michael S. Tsirkin <mst@...hat.com> wrote:
> > >
> > > On Wed, Oct 27, 2021 at 02:19:11PM +0800, Xuan Zhuo wrote:
> > > > In the case of using indirect, indirect desc must be allocated and
> > > > released each time, which increases a lot of cpu overhead.
> > > >
> > > > Here, a cache is added for indirect. If the number of indirect desc to be
> > > > applied for is less than VIRT_QUEUE_CACHE_DESC_NUM, the desc array with
> > > > the size of VIRT_QUEUE_CACHE_DESC_NUM is fixed and cached for reuse.
> > > >
> > > > Signed-off-by: Xuan Zhuo <xuanzhuo@...ux.alibaba.com>
> > > > ---
> > > > drivers/virtio/virtio.c | 6 ++++
> > > > drivers/virtio/virtio_ring.c | 63 ++++++++++++++++++++++++++++++------
> > > > include/linux/virtio.h | 10 ++++++
> > > > 3 files changed, 70 insertions(+), 9 deletions(-)
> > > >
> > > > diff --git a/drivers/virtio/virtio.c b/drivers/virtio/virtio.c
> > > > index 0a5b54034d4b..04bcb74e5b9a 100644
> > > > --- a/drivers/virtio/virtio.c
> > > > +++ b/drivers/virtio/virtio.c
> > > > @@ -431,6 +431,12 @@ bool is_virtio_device(struct device *dev)
> > > > }
> > > > EXPORT_SYMBOL_GPL(is_virtio_device);
> > > >
> > > > +void virtio_use_desc_cache(struct virtio_device *dev, bool val)
> > > > +{
> > > > + dev->desc_cache = val;
> > > > +}
> > > > +EXPORT_SYMBOL_GPL(virtio_use_desc_cache);
> > > > +
> > > > void unregister_virtio_device(struct virtio_device *dev)
> > > > {
> > > > int index = dev->index; /* save for after device release */
> > > > diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
> > > > index dd95dfd85e98..0b9a8544b0e8 100644
> > > > --- a/drivers/virtio/virtio_ring.c
> > > > +++ b/drivers/virtio/virtio_ring.c
> > > > @@ -117,6 +117,10 @@ struct vring_virtqueue {
> > > > /* Hint for event idx: already triggered no need to disable. */
> > > > bool event_triggered;
> > > >
> > > > + /* Is indirect cache used? */
> > > > + bool use_desc_cache;
> > > > + void *desc_cache_chain;
> > > > +
> > > > union {
> > > > /* Available for split ring */
> > > > struct {
> > > > @@ -423,12 +427,47 @@ static unsigned int vring_unmap_one_split(const struct vring_virtqueue *vq,
> > > > return extra[i].next;
> > > > }
> > > >
> > > > -static struct vring_desc *alloc_indirect_split(struct virtqueue *_vq,
> > > > +#define VIRT_QUEUE_CACHE_DESC_NUM 4
> > > > +
> > > > +static void desc_cache_chain_free_split(void *chain)
> > > > +{
> > > > + struct vring_desc *desc;
> > > > +
> > > > + while (chain) {
> > > > + desc = chain;
> > > > + chain = (void *)desc->addr;
> > > > + kfree(desc);
> > > > + }
> > > > +}
> > > > +
> > > > +static void desc_cache_put_split(struct vring_virtqueue *vq,
> > > > + struct vring_desc *desc, int n)
> > > > +{
> > > > + if (vq->use_desc_cache && n <= VIRT_QUEUE_CACHE_DESC_NUM) {
> > > > + desc->addr = (u64)vq->desc_cache_chain;
> > > > + vq->desc_cache_chain = desc;
> > > > + } else {
> > > > + kfree(desc);
> > > > + }
> > > > +}
> > > > +
> > >
> > >
> > > So I have a question here. What happens if we just do:
> > >
> > > if (n <= VIRT_QUEUE_CACHE_DESC_NUM) {
> > > return kmem_cache_alloc(VIRT_QUEUE_CACHE_DESC_NUM * sizeof desc, gfp)
> > > } else {
> > > return kmalloc_arrat(n, sizeof desc, gfp)
> > > }
> > >
> > > A small change and won't we reap most performance benefits?
> >
> > Yes, I think we need a benchmark to use private cache to see how much
> > it can help.
>
> I did a test, the code is as follows:
>
> +static void desc_cache_put_packed(struct vring_virtqueue *vq,
> + struct vring_packed_desc *desc, int n)
> + {
> + if (n <= VIRT_QUEUE_CACHE_DESC_NUM) {
> + kmem_cache_free(vq->desc_cache, desc);
> + } else {
> + kfree(desc);
> + }
>
>
> @@ -476,11 +452,14 @@ static struct vring_desc *alloc_indirect_split(struct vring_virtqueue *vq,
> */
> gfp &= ~__GFP_HIGHMEM;
>
> - desc = kmalloc_array(n, sizeof(struct vring_desc), gfp);
> + if (total_sg <= VIRT_QUEUE_CACHE_DESC_NUM)
> + desc = kmem_cache_alloc(vq->desc_cache, gfp);
> + else
> + desc = kmalloc_array(total_sg, sizeof(struct vring_desc), gfp);
> +
>
> .......
>
> + vq->desc_cache = kmem_cache_create("virio_desc",
> + 4 * sizeof(struct vring_desc),
> + 0, 0, NULL);
>
> The effect is not good, basically there is no improvement, using perf top can
> see that the overhead of kmem_cache_alloc/kmem_cache_free is also relatively
> large:
>
> 26.91% [kernel] [k] virtqueue_add
> 15.35% [kernel] [k] detach_buf_split
> 14.15% [kernel] [k] virtnet_xsk_xmit
> 13.24% [kernel] [k] virtqueue_add_outbuf
> > 9.30% [kernel] [k] __slab_free
> > 3.91% [kernel] [k] kmem_cache_alloc
> 2.85% [kernel] [k] virtqueue_get_buf_ctx
> > 2.82% [kernel] [k] kmem_cache_free
> 2.54% [kernel] [k] memset_erms
> 2.37% [kernel] [k] xsk_tx_peek_desc
> 1.20% [kernel] [k] virtnet_xsk_run
> 0.81% [kernel] [k] vring_map_one_sg
> 0.69% [kernel] [k] __free_old_xmit_ptr
> 0.69% [kernel] [k] virtqueue_kick_prepare
> 0.43% [kernel] [k] sg_init_table
> 0.41% [kernel] [k] sg_next
> 0.28% [kernel] [k] vring_unmap_one_split
> 0.25% [kernel] [k] vring_map_single.constprop.34
> 0.24% [kernel] [k] net_rx_action
>
> Thanks.
How about batching these? E.g. kmem_cache_alloc_bulk/kmem_cache_free_bulk?
> >
> > Thanks
> >
> > >
> > > --
> > > MST
> > >
> >
Powered by blists - more mailing lists