On 2017年05月05日 13:53, Wei Wang wrote:
On 05/05/2017 10:27 AM, Jason Wang wrote:
On 2017年05月04日 18:58, Wang, Wei W wrote:
I want to re-open the discussion left long time ago:
, and discuss the possibility of changing the hardcoded (256) TX queue
size to be configurable between 256 and 1024.
Yes, I think we probably need this.
That's great, thanks.
The reason to propose this request is that a severe issue of packet drops in
TX direction was observed with the existing hardcoded 256 queue size,
which causes performance issues for packet drop sensitive guest
applications that cannot use indirect descriptor tables. The issue goes away
with 1K queue size.
Do we need even more, what if we find 1K is even not sufficient in the future? Modern nics has size up to ~8192.
Yes. Probably, we can also set the RX queue size to 8192 (currently it's 1K) as well.
The concern mentioned in the previous discussion (please check the link
above) is that the number of chained descriptors would exceed
UIO_MAXIOV (1024) supported by the Linux.
We could try to address this limitation but probably need a new feature bit to allow more than UIO_MAXIOV sgs.
I think we should first discuss whether it would be an issue below.
From the code, I think the number of the chained descriptors is limited to
MAX_SKB_FRAGS + 2 (~18), which is much less than UIO_MAXIOV.
This is the limitation of #page frags for skb, not the iov limitation.
I think the number of page frags are filled into the same number of descriptors
in the virtio-net driver (e.g. use 10 descriptors for 10 page frags). On the other
side, the virtio-net backend uses the same number of iov for the descriptors.
Since the number of page frags is limited to 18, I think there wouldn't be more
than 18 iovs to be passed to writev, right?
This limitation assumption is incorrect for Windows. We saw cases (and strangely enough with small packets) when Windows returns scatter gather list with 32 descriptors or more.
Looks not, see skb_copy_datagram_from_iter().