Re: [PATCH iwl-net v2] ice: fix Rx byte accounting for dropped XDP frames

From
Alexander Lobakin <aleksander.lobakin@intel.com>
Date
2026-11-10 15:34:51
Also in
intel-wired-lan, stable
Reply to
Dominik Dvořák, 2026-11-09
From: Dominik Dvořák Via B4 Relay
[off-list ref]
Date: Mon, 09 Nov 2026 09:24:10 +0100
From: Dominik Dvořák <redacted>

Since the Rx path was converted to page pool and libeth, ice_run_xdp()
returns the buffer itself via libeth_xdp_return_buff() on XDP_DROP,
XDP_ABORTED and on a failed XDP_TX or XDP_REDIRECT. That helper also
sets xdp->data to NULL. ice_clean_rx_irq() then calls xdp_get_buff_len()
on the same buffer to update the Rx byte counter, which evaluates
xdp->data_end - NULL and adds the low 32 bits of a kernel address to
the ring's byte counter for every such frame. For multi-buffer frames
it would also read skb_shared_info from a page that has already been
returned to the page pool.

On an E810 with an XDP blocklist attached to an 802.3ad bond, this shows
up as jumps of up to 4 GiB in the per-queue and netdev rx_bytes counters,
one per dropped packet, while the packet counters and the hardware byte
counter (rx_bytes.nic) stay correct. The bogus values are also exposed
to unprivileged users through the netdev statistics.

Count the frame length once, before running the XDP program, and use it
for the byte counter on every path; this also drops the duplicate
skb->len accounting that the XDP_PASS path did afterwards. The frame
length as received is counted, rather than the length after a possible
bpf_xdp_adjust_head()/bpf_xdp_adjust_tail().

The problem was tracked down and this fix written with the help of
Claude Opus 5.5 (Anthropic), working from netdev and ethtool counters
collected on the affected system and from the driver source.

Tested on an E810-XXV (7.0.14 kernel, same code as current master) with
an XDP blocklist attached to an 802.3ad bond: without the fix, each
dropped packet added about 2.5 GB to rx_queue_N_bytes; with the fix, 40
dropped packets produced no such jumps.

Fixes: 93f53db9f9dc ("ice: switch to Page Pool")
Cc: <redacted> # 6.18.x
Link: https://lore.kernel.org/netdev/DB7PR03MB460158B90A64D025962670E0D5772@DB7PR03MB4601.eurprd03.prod.outlook.com/ (local)
Signed-off-by: Dominik Dvořák <redacted>
Reviewed-by: Alexander Lobakin <aleksander.lobakin@intel.com>

This looks elegant and even simplifies hotpath a bit. I don't think we
need to try counting post-XDP bytes as if the XDP program added some
bytes using grow tail, those weren't received by wire either way.
---
Based on the stable v7.0.14 tree, which is what the affected box runs;
the changed code is identical in current master. Tested on that box
with the fixed module built against the 7.0.14-26-pve headers.
---
Changes in v2:
- Account for Rx bytes and packets once, before ice_run_xdp(), and drop
  the separate skb->len accounting on the XDP_PASS path (Olek).
- Add Cc: stable with the minimum affected version (Olek).
Thanks,
Olek
Thread (2 messages) 2 messages, 2 authors, last reply 6d ago
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help