net: mana: add BQL support

MANA doesn't support byte queue limits, so each TX queue can hold 256
WQEs by default, and up to 16384, of up to 64 KB TSO packets, before the
qdisc layer sees any backpressure. This adds latency to all the flows
sharing a queue, and defeats AQM qdiscs such as fq_codel, which rely on
BQL to keep the driver queues short.

Report the bytes queued and completed to BQL, except for the XDP_TX and
XDP_REDIRECT frames. These share the SQs, but bypass the qdisc and are
sent even when the queue is stopped, so BQL can't hold them back.
Counting them would let sustained XDP traffic keep the queue stopped,
starving the stack, and the TX watchdog would then either fire while the
completions still arrive, or, if the XDP sends refreshed trans_start,
not fire when the completions stop. The XDP SKBs are marked in their
skb->cb, next to the WQE size, so that their completions are skipped as
well.

Reset the BQL state when destroying the queues, once their NAPI is
disabled, since the SKBs still pending after a drain timeout are freed
without being completed.

Lost TX completions can now stop a queue sooner: it also stops when the
bytes in flight exceed the BQL limit, not only when the SQ is full, and
the limit may never grow enough to cover the lost bytes. This includes
the completions skipped by the WARN_ON_ONCE() early returns in
mana_poll_tx_cq(), which already leak SQ space: completing the SKBs
already freed there wouldn't reliably balance BQL, since
mana_gd_poll_cq() has already consumed the CQEs of the rest of the
batch. The TX watchdog then fires, as nothing refreshes the trans_start
of a queue stopped by BQL, and its handler resets the queues along with
their BQL state. The exception, as before this patch, is XDP traffic
keeping the SQ full: each XDP send then stops the queue again,
refreshing its trans_start.

Tested on two Standard_D16ds_v6 Azure VMs in a proximity placement
group, where the throughput is capped at 10.6 Gbps, with the DUT
alternately booting net-next (c1c1f0a31712) with and without this patch
applied.

The throughput stays at the cap with 1, 8 and 64 TCP flows, and the idle
TCP_RR latency doesn't change (p50 of 47 us with vs. without the patch).
With 16 senders of 64-byte UDP packets, the DUT sends 3.7 to 4.0 Mpps
without the patch and 3.7 to 4.1 Mpps with it. With bulk TCP flows
sharing the queues, the netperf TCP_RR latency drops (medians of 6
runs):

  16 queues, 64 flows: p50 4649 -> 2182 us, p99 5621 -> 2677 us
  1 queue, 8 flows:    p50 2554 ->  418 us, p99 2723 ->  624 us

The standing queue moves from the SQs to fq_codel, which now drops
packets: about 2500 qdisc drops and 88K TCP retransmissions per 48 s run
with 64 flows, against at most 29 retransmissions before. The
retransmissions are about 0.2% of the segments sent. The TCP_RR flows
lose no packets, and the packets of the bulk flows are dropped before
they reach the wire, so the goodput changes by less than 0.01%.

Signed-off-by: Hamza Mahfooz <hamzamahfooz@linux.microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Link: https://patch.msgid.link/20261013214937.145312-1-hamzamahfooz@linux.microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

1 files changed, 2 insertions, 0 deletions

diff --git a/include/net/mana/mana.h b/include/net/mana/mana.h
index 8c3dc9d0b299..e1fa9131c490 100644
--- a/include/net/mana/mana.h
+++ b/include/net/mana/mana.h
@@ -672,6 +672,8 @@ struct mana_port_context {
 	u32 steer_cqe_coalescing;
 };
 
+netdev_tx_t __mana_start_xmit(struct sk_buff *skb, struct net_device *ndev,
+			      bool xdp);
 netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct net_device *ndev);
 int mana_config_rss(struct mana_port_context *ac, enum TRI_STATE rx,
 		    bool update_hash, bool update_tab);