From: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Date: Mon, 25 Jan 2021 11:10:43 +0800
On Fri, 22 Jan 2021 16:24:17 +0000, Alexander Lobakin [off-list ref] wrote:
quoted
From: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Date: Fri, 22 Jan 2021 23:36:29 +0800
quoted
On Fri, 22 Jan 2021 12:08:00 +0000, Alexander Lobakin [off-list ref] wrote:
quoted
From: Alexander Lobakin <redacted>
Date: Fri, 22 Jan 2021 11:55:35 +0000
quoted
From: Alexander Lobakin <redacted>
Date: Fri, 22 Jan 2021 11:47:45 +0000
quoted
From: Eric Dumazet <redacted>
Date: Thu, 21 Jan 2021 16:41:33 +0100
quoted
On 1/21/21 2:47 PM, Xuan Zhuo wrote:
quoted
This patch is used to construct skb based on page to save memory copy
overhead.
This function is implemented based on IFF_TX_SKB_NO_LINEAR. Only the
network card priv_flags supports IFF_TX_SKB_NO_LINEAR will use page to
directly construct skb. If this feature is not supported, it is still
necessary to copy data to construct skb.
---------------- Performance Testing ------------
The test environment is Aliyun ECS server.
Test cmd:
xdpsock -i eth0 -t -S -s <msg size>
Test result data:
size 64 512 1024 1500
copy 1916747 1775988 1600203 1440054
page 1974058 1953655 1945463 1904478
percent 3.0% 10.0% 21.58% 32.3%
Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Reviewed-by: Dust Li <dust.li@linux.alibaba.com>
---
net/xdp/xsk.c | 104 ++++++++++++++++++++++++++++++++++++++++++++++++----------
1 file changed, 86 insertions(+), 18 deletions(-)
diff --git a/net/xdp/xsk.c b/net/xdp/xsk.c
index 4a83117..38af7f1 100644
--- a/net/xdp/xsk.c
+++ b/net/xdp/xsk.c
@@ -430,6 +430,87 @@ static void xsk_destruct_skb(struct sk_buff *skb)
sock_wfree(skb);
}
+static struct sk_buff *xsk_build_skb_zerocopy(struct xdp_sock *xs,
+ struct xdp_desc *desc)
+{
+ u32 len, offset, copy, copied;
+ struct sk_buff *skb;
+ struct page *page;
+ void *buffer;
+ int err, i;
+ u64 addr;
+
+ skb = sock_alloc_send_skb(&xs->sk, 0, 1, &err);
Also,
maybe we should allocate it with NET_SKB_PAD so NIC drivers could
use some reserved space?
skb = sock_alloc_send_skb(&xs->sk, NET_SKB_PAD, 1, &err);
...
skb_reserve(skb, NET_SKB_PAD);
Eric, what do you think?
I think you are right. Some space should be added to continuous equipment. This
space should also be added in the copy mode below. Is LL_RESERVED_SPACE more
appropriate?
No. If you look at __netdev_alloc_skb() and __napi_alloc_skb(), they
reserve NET_SKB_PAD at the beginning of linear area. Documentation of
__build_skb() also says that driver should reserve NET_SKB_PAD before
the actual frame, so it is a standartized hardware-independent
headroom.
I understand that these scenarios are in the case of receiving packets, and the
increased space is used by the protocol stack, especially RPS. I don't know if
this also applies to the sending scenario?
quoted
Leaving that space in skb->head will allow developers to implement
IFF_TX_SKB_NO_LINEAR in a wider variety of drivers, especially when
a driver has to prepend some sort of data before the actual frame.
Since it's usually of a size of one cacheline, shouldn't be a big
deal.
I agree with this. Some network cards require some space. For example,
virtio-net needs to add a virtio_net_hdr_mrg_rxbuf before skb->data, so my
original understanding is used here. When we send the skb to the
driver, the driver may need a memory space. So I refer to the
implementation of __ip_append_data, I feel that adding
LL_RESERVED_SPACE is a suitable solution.
I feel that I may still not understand the use scene you mentioned. Can you
elaborate on what you understand this space will be used for?
LL_RESERVED_SPACE() consists of L2 header size (Ethernet for the most
cases) and dev->needed_headroom. That is not a value to count on, as:
- L2 header is already here in XSK buffer;
- not all drivers set dev->needed_headroom;
- it's aligned by 16, not L1_CACHE_SIZE.
As this path is XSK generic path, i.e. when driver-side XSK is not
present or not requested, it can be applied to every driver. Many
of them call skb_cow_head() + skb_push() on their xmit path:
- nearly all virtual drivers (to insert their specific headers);
- nearly all switch drivers (to insert switch CPU port tags);
- some enterprise NIC drivers (ChelsIO for LSO, Netronome
for TLS etc.).
skb_cow_head() + skb_push() relies on a required NET_SKB_PAD headroom.
In case where there is no enough space (and you allocate an skb with
no headroom at all), skb will be COWed, which is a huge overhead and
will cause slowdowns.
So, adding NET_SKB_PAD would save from almost all, if not all, such
reallocations.
Thanks.
quoted
[ I also had an idea of allocating an skb with a headroom of
NET_SKB_PAD + 256 bytes, so nearly all drivers could just call
pskb_pull_tail() to support such type of skbuffs without much
effort, but I think that it's better to teach drivers to support
xmitting of really headless ones. If virtio_net can do it, why
shouldn't the others ]
quoted
quoted
quoted
quoted
quoted
quoted
+ if (unlikely(!skb))
+ return ERR_PTR(err);
+
+ addr = desc->addr;
+ len = desc->len;
+
+ buffer = xsk_buff_raw_get_data(xs->pool, addr);
+ offset = offset_in_page(buffer);
+ addr = buffer - xs->pool->addrs;
+
+ for (copied = 0, i = 0; copied < len; i++) {
+ page = xs->pool->umem->pgs[addr >> PAGE_SHIFT];
+
+ get_page(page);
+
+ copy = min_t(u32, PAGE_SIZE - offset, len - copied);
+
+ skb_fill_page_desc(skb, i, page, offset, copy);
+
+ copied += copy;
+ addr += copy;
+ offset = 0;
+ }
+
+ skb->len += len;
+ skb->data_len += len;
quoted
+ skb->truesize += len;
This is not the truesize, unfortunately.
We need to account for the number of pages, not number of bytes.
The easiest solution is:
skb->truesize += PAGE_SIZE * i;
i would be equal to skb_shinfo(skb)->nr_frags after exiting the loop.
Oops, pls ignore this. I forgot that XSK buffers are not
"one per page".
We need to count the number of pages manually and then do
skb->truesize += PAGE_SIZE * npages;
Right.
quoted
quoted
quoted
+
+ refcount_add(len, &xs->sk.sk_wmem_alloc);
+
+ return skb;
+}
+
Al
Thanks,
Al
Al
Thanks,
Al