Cc. Alex Duyck + Eric, please criticize my idea below.
On 12/05/2023 15.08, Lorenzo Bianconi wrote:
In order to reduce page_pool memory footprint, rely on
page_pool_dev_alloc_frag routine and reduce buffer size
(VETH_PAGE_POOL_FRAG_SIZE) to PAGE_SIZE / 2 in order to consume one page
for two 1500B frames. Reduce VETH_XDP_PACKET_HEADROOM to 192 from 256
(XDP_PACKET_HEADROOM) to fit max_head_size in VETH_PAGE_POOL_FRAG_SIZE.
Please note, using default values (CONFIG_MAX_SKB_FRAGS=17), maximum
supported MTU is now reduced to 36350B.
Signed-off-by: Lorenzo Bianconi <lorenzo@xxxxxxxxxx>
---
drivers/net/veth.c | 39 +++++++++++++++++++++++++--------------
1 file changed, 25 insertions(+), 14 deletions(-)
diff --git a/drivers/net/veth.c b/drivers/net/veth.c
index 614f3e3efab0..0e648703cccf 100644
--- a/drivers/net/veth.c
+++ b/drivers/net/veth.c
@@ -31,9 +31,12 @@
#define DRV_NAME "veth"
#define DRV_VERSION "1.0"
-#define VETH_XDP_FLAG BIT(0)
-#define VETH_RING_SIZE 256
-#define VETH_XDP_HEADROOM (XDP_PACKET_HEADROOM + NET_IP_ALIGN)
+#define VETH_XDP_FLAG BIT(0)
+#define VETH_RING_SIZE 256
+#define VETH_XDP_PACKET_HEADROOM 192
+#define VETH_XDP_HEADROOM (VETH_XDP_PACKET_HEADROOM + \
+ NET_IP_ALIGN)
+#define VETH_PAGE_POOL_FRAG_SIZE 2048
#define VETH_XDP_TX_BULK_SIZE 16
#define VETH_XDP_BATCH 16
@@ -736,7 +739,7 @@ static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq,
if (skb_shared(skb) || skb_head_is_locked(skb) ||
skb_shinfo(skb)->nr_frags ||
skb_headroom(skb) < XDP_PACKET_HEADROOM) {
- u32 size, len, max_head_size, off;
+ u32 size, len, max_head_size, off, pp_off;
struct sk_buff *nskb;
struct page *page;
int i, head_off;
@@ -747,17 +750,20 @@ static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq,
*
* Make sure we have enough space for linear and paged area
*/
- max_head_size = SKB_WITH_OVERHEAD(PAGE_SIZE -
+ max_head_size = SKB_WITH_OVERHEAD(VETH_PAGE_POOL_FRAG_SIZE -
VETH_XDP_HEADROOM);
- if (skb->len > PAGE_SIZE * MAX_SKB_FRAGS + max_head_size)
+ if (skb->len >
+ VETH_PAGE_POOL_FRAG_SIZE * MAX_SKB_FRAGS + max_head_size)
goto drop;
/* Allocate skb head */
- page = page_pool_dev_alloc_pages(rq->page_pool);
It seems wasteful to allocate a full page PAGE_SIZE.
+ page = page_pool_dev_alloc_frag(rq->page_pool, &pp_off,
+ VETH_PAGE_POOL_FRAG_SIZE);
Allocating PAGE_SIZE/2 isn't much better.
At this point we already know the skb->len (and skb_headlen).
Why don't we allocated the size that we need?
See page_frag_alloc() system invented by Eric and Duyck.
if (!page)
goto drop;
- nskb = napi_build_skb(page_address(page), PAGE_SIZE);
+ nskb = napi_build_skb(page_address(page) + pp_off,
+ VETH_PAGE_POOL_FRAG_SIZE);
if (!nskb) {
page_pool_put_full_page(rq->page_pool, page, true);
goto drop;
@@ -782,15 +788,18 @@ static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq,
len = skb->len - off;
for (i = 0; i < MAX_SKB_FRAGS && off < skb->len; i++) {
- page = page_pool_dev_alloc_pages(rq->page_pool);
+ page = page_pool_dev_alloc_frag(rq->page_pool, &pp_off,
+ VETH_PAGE_POOL_FRAG_SIZE);
if (!page) {
consume_skb(nskb);
goto drop;
}
- size = min_t(u32, len, PAGE_SIZE);
- skb_add_rx_frag(nskb, i, page, 0, size, PAGE_SIZE);
- if (skb_copy_bits(skb, off, page_address(page),
+ size = min_t(u32, len, VETH_PAGE_POOL_FRAG_SIZE);
+ skb_add_rx_frag(nskb, i, page, pp_off, size,
+ VETH_PAGE_POOL_FRAG_SIZE);
+ if (skb_copy_bits(skb, off,
+ page_address(page) + pp_off,
size)) {
consume_skb(nskb);
goto drop;
[...]
