* [RFC] r8169: possible TX completion ordering issue on arm64
@ 2026-11-18 14:11 Hanson Wang
0 siblings, 0 replies; only message in thread
From: Hanson Wang @ 2026-11-18 14:11 UTC (permalink / raw)
To: netdev; +Cc: Heiner Kallweit, nic_swsd
Hi,
We observed four crashes in dql_completed() on an RK3588-based arm64
board (UGREEN RK3588C DH4300Plus). The NIC is an RTL8125B, XID 641,
at PCI 0002:21:00.0, linked at 2.5Gbps. The kernel is a vendor
6.1.84 build (6.1.84 #d64bed19bdfd). We compared the cur_tx publish
and read with current mainline and that part looks the same:
rtl8169_start_xmit() still does smp_wmb() followed by WRITE_ONCE(),
and rtl_tx() still does READ_ONCE() before reading the descriptor.
The vendor tree still calls netdev_completed_queue() directly; mainline
has since switched that call to netif_subqueue_completed_wake(). We
have not reproduced this on mainline.
All four oopses stop at the same place:
kernel BUG at lib/dynamic_queue_limits.c:27!
Internal error: Oops - BUG: 00000000f2000800 [#1] SMP
pc : dql_completed+0x13c/0x14c
lr : rtl8169_poll+0x3a0/0x54c [r8169]
One of them, from idle softirq:
CPU: 2 PID: 0 Comm: swapper/2 Tainted: G W E 6.1.84 #d64bed19bdfd
Hardware name: UGREEN RK3588C DH4300Plus Board (DT)
Call trace:
dql_completed+0x13c/0x14c
__napi_poll+0x3c/0x1d0
net_rx_action+0x16c/0x3a0
__do_softirq+0x120/0x338
____do_softirq+0x14/0x20
call_on_irq_stack+0x24/0x40
do_softirq_own_stack+0x20/0x30
__irq_exit_rcu+0x98/0xc0
irq_exit_rcu+0x14/0x20
el1_interrupt+0x48/0xa0
el1h_64_irq_handler+0x14/0x20
el1h_64_irq+0x74/0x78
arch_cpu_idle+0x14/0x40
...
Kernel panic - not syncing: Oops - BUG: Fatal exception in interrupt
Another is the same pc/lr from process context (ctl_serv), entering
softirq via el0_interrupt. Two of the four were preceded by
"NETDEV WATCHDOG: eth0 (r8169): transmit queue 0 timed out" and
"ASPM disabled on Tx timeout". The other two were not.
dql_completed() is called as (dql, count). At the BUG the registers
are:
x1 = 0x00000570 x2 = 0x00000534 delta = 0x3c
x1 = 0x000007de x2 = 0x000007a2 delta = 0x3c
x1 = 0x000009d8 x2 = 0x0000099c delta = 0x3c
x1 = 0x00001091 x2 = 0x00001055 delta = 0x3c
The failing check is BUG_ON(count > num_queued - num_completed). The
branch goes straight to the BRK, so x1 is still count and x2 is still
num_queued - num_completed. In every case the completion count is
0x3c = 60 = ETH_ZLEN bytes ahead of the outstanding BQL count, which
is one short frame padded to the minimum length. RTL_GIGA_MAC_VER_63
pads such frames to ETH_ZLEN. We have not established why the counts
diverge.
While looking at the TX path I noticed that rtl8169_start_xmit()
publishes the new ring index with:
smp_wmb();
WRITE_ONCE(tp->cur_tx, tp->cur_tx + frags + 1);
and rtl_tx() then does:
while (READ_ONCE(tp->cur_tx) != dirty_tx) {
status = le32_to_cpu(READ_ONCE(tp->TxDescArray[entry].opts1));
On arm64 a control dependency does not order the later load of opts1,
or the later read of the BQL counters, against the load of cur_tx.
smp_wmb() orders the writer's stores, but the reader has no pairing
barrier. A CPU could observe the new cur_tx, still see DescOwn clear
on a descriptor that has not been handed to the hardware, and account
that skb in netdev_completed_queue() before __netdev_sent_queue() is
visible. That would trip this BUG_ON. I may be missing an ordering
guarantee that already covers this.
Would an acquire load on the reader be appropriate here?
- while (READ_ONCE(tp->cur_tx) != dirty_tx) {
+ while (smp_load_acquire(&tp->cur_tx) != dirty_tx) {
This is a question, not a fix. We have not A/B tested it, and we have
not reproduced the oops on mainline.
Thanks,
Hanson Wang
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2026-11-18 14:12 UTC | newest]
Thread overview: (only message) (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-18 14:11 [RFC] r8169: possible TX completion ordering issue on arm64 Hanson Wang
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox