Commit 730ff06d3f ("net: mana: Use page pool fragments for RX buffers
instead of full pages to improve memory efficiency.") started handing
out RX buffers with zero headroom so that two buffers fit into one page
at the default MTU.
The MANA TX path, however, stores the per scatter-gather entry DMA
mappings in `struct mana_skb_head` at skb->head, and mana_start_xmit()
therefore calls skb_cow_head(skb, MANA_HEADROOM). The port advertises
this requirement as ndev->needed_headroom = MANA_HEADROOM.
As a result every packet that is received and then forwarded out of a
MANA port fails the skb_cow() in ip_forward() and gets reallocated and
copied by pskb_expand_head(). This is invisible to a plain RX or TX
workload, but it puts a full skb reallocation plus memcpy on the hot
path of every single forwarded packet, which is exactly what a
router/NVA workload does.
Restore the headroom. Note that reserving MANA_HEADROOM (232) is not
enough: ip_forward() asks for LL_RESERVED_SPACE(dev), which rounds
hard_header_len + needed_headroom up to HH_DATA_MOD and is 256 bytes on
ethernet. Use LL_RESERVED_SPACE() directly so the value keeps tracking
both constants. Also, since LL_RESERVED_SPACE() tracks MANA_HEADROOM,
it grows with MAX_SKB_FRAGS and for MAX_SKB_FRAGS >= 19 it is greater
than 256, so we have to account for that by using the headroom the RX
queue actually uses (instead of assuming XDP_PACKET_HEADROOM) and
turning MANA_XDP_MTU_MAX into MANA_XDP_MTU_MAX(ndev) (note that at the
default CONFIG_MAX_SKB_FRAGS=17 they are equivalent).
At the default MTU on a 4K page this means a buffer no longer fits twice
into a page (SKB_DATA_ALIGN(1500 + MANA_RXBUF_PAD + 256) = 2112), so the
frag-vs-single decision is now made by computing the real buffer size
instead of comparing the MTU against PAGE_SIZE / 2. The page_pool
fragment path is still used wherever at least two buffers genuinely fit,
e.g. on 16K and 64K page sizes.
Measured on an Azure VM with a MANA NIC acting as a forwarding NVA (UDP,
1400 byte payload, 4 streams, 8 Gbps offered, only the forwarding
node's kernel differs), 8 runs each, median:
forwarded pps throughput
before 272,830 3.06 Gbps
after 390,560 4.37 Gbps (+43%)
perf on the forwarding node, same workload:
memset_orig __pi_memcpy pskb_expand_head
before 10.07% 3.96% present
after 0.94% 0.64% gone
Cc: stable@vger.kernel.org
Fixes: 730ff06d3f ("net: mana: Use page pool fragments for RX buffers instead of full pages to improve memory efficiency.")
Signed-off-by: Hamza Mahfooz <hamzamahfooz@linux.microsoft.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20261003013647.2051416-1-hamzamahfooz@linux.microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
259c4a5191