Re: [PATCH RFC v1 1/3] r8169: fix the KCSAN reported data-race in rtl_tx() while reading tp->cur_tx
From: Mirsad Todorovac
Date: Fri Oct 27 2023 - 11:14:58 EST
On 10/22/2023 11:23 PM, Heiner Kallweit wrote:
On 22.10.2023 19:24, Mirsad Todorovac wrote:
On 9/27/23 21:52, Heiner Kallweit wrote:
On 27.09.2023 20:41, Mirsad Goran Todorovac wrote:
KCSAN reported the following data-race:
==================================================================
BUG: KCSAN: data-race in rtl8169_poll [r8169] / rtl8169_start_xmit [r8169]
write (marked) to 0xffff888102474b74 of 4 bytes by task 5358 on cpu 29:
rtl8169_start_xmit (drivers/net/ethernet/realtek/r8169_main.c:4254) r8169
dev_hard_start_xmit (./include/linux/netdevice.h:4889 ./include/linux/netdevice.h:4903 net/core/dev.c:3544 net/core/dev.c:3560)
sch_direct_xmit (net/sched/sch_generic.c:342)
__dev_queue_xmit (net/core/dev.c:3817 net/core/dev.c:4306)
ip_finish_output2 (./include/linux/netdevice.h:3082 ./include/net/neighbour.h:526 ./include/net/neighbour.h:540 net/ipv4/ip_output.c:233)
__ip_finish_output (net/ipv4/ip_output.c:311 net/ipv4/ip_output.c:293)
ip_finish_output (net/ipv4/ip_output.c:328)
ip_output (net/ipv4/ip_output.c:435)
ip_send_skb (./include/net/dst.h:458 net/ipv4/ip_output.c:127 net/ipv4/ip_output.c:1486)
udp_send_skb (net/ipv4/udp.c:963)
udp_sendmsg (net/ipv4/udp.c:1246)
inet_sendmsg (net/ipv4/af_inet.c:840 (discriminator 4))
sock_sendmsg (net/socket.c:730 net/socket.c:753)
__sys_sendto (net/socket.c:2177)
__x64_sys_sendto (net/socket.c:2185)
do_syscall_64 (arch/x86/entry/common.c:50 arch/x86/entry/common.c:80)
entry_SYSCALL_64_after_hwframe (arch/x86/entry/entry_64.S:120)
read to 0xffff888102474b74 of 4 bytes by interrupt on cpu 21:
rtl8169_poll (drivers/net/ethernet/realtek/r8169_main.c:4397 drivers/net/ethernet/realtek/r8169_main.c:4581) r8169
__napi_poll (net/core/dev.c:6527)
net_rx_action (net/core/dev.c:6596 net/core/dev.c:6727)
__do_softirq (kernel/softirq.c:553)
__irq_exit_rcu (kernel/softirq.c:427 kernel/softirq.c:632)
irq_exit_rcu (kernel/softirq.c:647)
common_interrupt (arch/x86/kernel/irq.c:247 (discriminator 14))
asm_common_interrupt (./arch/x86/include/asm/idtentry.h:636)
cpuidle_enter_state (drivers/cpuidle/cpuidle.c:291)
cpuidle_enter (drivers/cpuidle/cpuidle.c:390)
call_cpuidle (kernel/sched/idle.c:135)
do_idle (kernel/sched/idle.c:219 kernel/sched/idle.c:282)
cpu_startup_entry (kernel/sched/idle.c:378 (discriminator 1))
start_secondary (arch/x86/kernel/smpboot.c:210 arch/x86/kernel/smpboot.c:294)
secondary_startup_64_no_verify (arch/x86/kernel/head_64.S:433)
value changed: 0x002f4815 -> 0x002f4816
Reported by Kernel Concurrency Sanitizer on:
CPU: 21 PID: 0 Comm: swapper/21 Tainted: G L 6.6.0-rc2-kcsan-00143-gb5cbe7c00aa0 #41
Hardware name: ASRock X670E PG Lightning/X670E PG Lightning, BIOS 1.21 04/26/2023
==================================================================
The write side of drivers/net/ethernet/realtek/r8169_main.c is:
==================
4251 /* rtl_tx needs to see descriptor changes before updated tp->cur_tx */
4252 smp_wmb();
4253
→ 4254 WRITE_ONCE(tp->cur_tx, tp->cur_tx + frags + 1);
4255
4256 stop_queue = !netif_subqueue_maybe_stop(dev, 0, rtl_tx_slots_avail(tp),
4257 R8169_TX_STOP_THRS,
4258 R8169_TX_START_THRS);
The read side is the function rtl_tx():
4355 static void rtl_tx(struct net_device *dev, struct rtl8169_private *tp,
4356 int budget)
4357 {
4358 unsigned int dirty_tx, bytes_compl = 0, pkts_compl = 0;
4359 struct sk_buff *skb;
4360
4361 dirty_tx = tp->dirty_tx;
4362
4363 while (READ_ONCE(tp->cur_tx) != dirty_tx) {
4364 unsigned int entry = dirty_tx % NUM_TX_DESC;
4365 u32 status;
4366
4367 status = le32_to_cpu(tp->TxDescArray[entry].opts1);
4368 if (status & DescOwn)
4369 break;
4370
4371 skb = tp->tx_skb[entry].skb;
4372 rtl8169_unmap_tx_skb(tp, entry);
4373
4374 if (skb) {
4375 pkts_compl++;
4376 bytes_compl += skb->len;
4377 napi_consume_skb(skb, budget);
4378 }
4379 dirty_tx++;
4380 }
4381
4382 if (tp->dirty_tx != dirty_tx) {
4383 dev_sw_netstats_tx_add(dev, pkts_compl, bytes_compl);
4384 WRITE_ONCE(tp->dirty_tx, dirty_tx);
4385
4386 netif_subqueue_completed_wake(dev, 0, pkts_compl, bytes_compl,
4387 rtl_tx_slots_avail(tp),
4388 R8169_TX_START_THRS);
4389 /*
4390 * 8168 hack: TxPoll requests are lost when the Tx packets are
4391 * too close. Let's kick an extra TxPoll request when a burst
4392 * of start_xmit activity is detected (if it is not detected,
4393 * it is slow enough). -- FR
4394 * If skb is NULL then we come here again once a tx irq is
4395 * triggered after the last fragment is marked transmitted.
4396 */
→ 4397 if (tp->cur_tx != dirty_tx && skb)
4398 rtl8169_doorbell(tp);
4399 }
4400 }
Obviously from the code, an earlier detected data-race for tp->cur_tx was fixed in the
line 4363:
4363 while (READ_ONCE(tp->cur_tx) != dirty_tx) {
but the same solution is required for protecting the other access to tp->cur_tx:
→ 4397 if (READ_ONCE(tp->cur_tx) != dirty_tx && skb)
4398 rtl8169_doorbell(tp);
The write in the line 4254 is protected with WRITE_ONCE(), but the read in the line 4397
might have suffered read tearing under some compiler optimisations.
The fix eliminated the KCSAN data-race report for this bug.
It is yet to be evaluated what happens if tp->cur_tx changes between the test in line 4363
and line 4397. This test should certainly not be cached by the compiler in some register
for such a long time, while asynchronous writes to tp->cur_tx might have occurred in line
4254 in the meantime.
netif_subqueue_completed_wake() has barriers ensuring that no cached value for tp->cur_tx
is used in line 4397. I'm not aware of any reported issues with an obvious link to the
potentential issue you describe.
I don't have a strong opinion on these patches. They shouldn't hurt, and if they make
KCSAN happy, why not.
Hi, Mr. Kallweit,
Your words have been taken with great weight and serious concern on my behalf, so I took
some time to investigate (even after the patches were Ack-ed).
No need to be so formal. I'm just a hobbyist developer who once suddenly
found himself in a maintainer position ;)
Well, if "manners maketh men" as they said, then it is better to start
more conservative I suppose, than the other way around - start friendly
and end in a formal communication :-)
The question is whether we deal with a data-race or a false positive reported by KCSAN.
I can indeed independently confirm that netif_subqueue_completed_wake() eventually expands to
call to:
include/net/netdev_queues.h:
==============================================
/* Variant of netdev_tx_completed_queue() which guarantees smp_mb() if
* @bytes != 0, regardless of kernel config.
*/
static inline void
netdev_txq_completed_mb(struct netdev_queue *dev_queue,
unsigned int pkts, unsigned int bytes)
{
if (IS_ENABLED(CONFIG_BQL))
netdev_tx_completed_queue(dev_queue, pkts, bytes);
else if (bytes)
smp_mb();
}
Obviously, smp_mb() is here.
On my __x86_64__ it expands to an ugly hack:
#define smp_mb() asm volatile("lock; addl $0,-132(%%rsp)" ::: "memory", "cc")
Thus far I concur with your findings.
But I see also Mr. Elver's point:
1. With the reference to: [1] https://www.kernel.org/doc/html/v6.2/core-api/wrappers/memory-barriers.html
compiler can do some "smart" optimisations that only work on single core, single threaded use.
What comes to mind is that "struct rtl8169_private" and its members might not be on the same cache line
with the pointer struct rtl8169_private *tp.
By the exposed matter in [1], READ_ONCE() provides a way to explain to the compiler that tp->cur_tx can
be concurrently modified by the other thread/core.
To be honest, I haven't caught a network card error or a driver bug either, but only the KCSAN complaint.
It would seem prudent to protect all the variables with the concurrent accesses from several cores/threads
or devices through DMA with (READ|WRITE)_ONCE.
r8169 driver supports one queue only. NAPI softirq guarantees that calls to rtl8169_poll
are serialized. Threaded NAPI may be a different story, however I'm not aware of any bug
report related to threaded NAPI.
That sounds all Greek to me. I am not that deep into the network stack.
35M+ lines of the Source still look intimidating. :-/
Especially since I don't see where the callback added in drivers/net/ethernet/realtek/r8169_main.c:5277
is called from:
netif_napi_add(dev, &tp->napi, rtl8169_poll);
I am not trying to be an arbiter nor an authority on the rtl8169 driver, which would obviously be you.
However, I feel compelled by something to review all the accesses to tp->cur_tx, tp->cur_tx and other
structure members against concurrent access, and systematically. With your permission, of course.
I would appreciate your mentoring on the task if you are interested.
Mainly, I figured out that in:
4575 static int rtl8169_poll(struct napi_struct *napi, int budget)
4576 {
4577 struct rtl8169_private *tp = container_of(napi, struct rtl8169_private, napi);
4578 struct net_device *dev = tp->dev;
4579 int work_done;
4580
4581 rtl_tx(dev, tp, budget);
4582
4583 work_done = rtl_rx(dev, tp, budget);
4584
4585 if (work_done < budget && napi_complete_done(napi, work_done))
4586 rtl_irq_enable(tp);
4587
4588 return work_done;
4589 }
rtl_tx() and rtl_rx() are called in sequence. Is it possible to imagine doing it in parallel?
Adding a separated TX NAPI is possible, question is whether it's worth the
effort. Look for users of netif_napi_add_tx() to get an idea of separated TX NAPI.
I once experimented with it in context of r8169 and didn't really find a benefit.
I guess it might be a bit more scalable. Would completions be too slow
to handle incoming and outgoing packets?
Certainly, what comes to mind is that we are limited with DMA and what
it can do.
Can the Realtek NIC hardware support that at all and which generations can?
Well, major issue with Realtek HW is that there's no public datasheets.
Only source of information is their out-of-tree drivers (r8168, r8101, r8125).
With regard to RSS: r8125 supports it, r8168 not.
This sounds like a challenge. Currently I happen to have a r8125 and RSS
seems interesting.
One coleague of mine said that Windows have better drivers, but mostly
it is because we as the opn source community cannot afford to buy the
specs for the fancy features in the hw. :-(
Best regards,
Mirsad Todorovac
(Of course, Ethernet might be full-duplex, but PCI Express line might not be.)
[2] https://www.rambus.com/blogs/pci-express-4/ Full-duplex PCI 4.0 specs.
Thank you.
Best regards,
Mirsad Todorovac
Fixes: 94d8a98e6235c ("r8169: reduce number of workaround doorbell rings")
Cc: Heiner Kallweit <hkallweit1@xxxxxxxxx>
Cc: nic_swsd@xxxxxxxxxxx
Cc: "David S. Miller" <davem@xxxxxxxxxxxxx>
Cc: Eric Dumazet <edumazet@xxxxxxxxxx>
Cc: Jakub Kicinski <kuba@xxxxxxxxxx>
Cc: Paolo Abeni <pabeni@xxxxxxxxxx>
Cc: Marco Elver <elver@xxxxxxxxxx>
Cc: netdev@xxxxxxxxxxxxxxx
Link: https://lore.kernel.org/lkml/dc7fc8fa-4ea4-e9a9-30a6-7c83e6b53188@xxxxxxxxxxxx/
Signed-off-by: Mirsad Goran Todorovac <mirsad.todorovac@xxxxxxxxxxxx>
---
v1:
the initial patch proposal. fixes the KCSAN warning.
drivers/net/ethernet/realtek/r8169_main.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c
index 6351a2dc13bc..281aaa851847 100644
--- a/drivers/net/ethernet/realtek/r8169_main.c
+++ b/drivers/net/ethernet/realtek/r8169_main.c
@@ -4394,7 +4394,7 @@ static void rtl_tx(struct net_device *dev, struct rtl8169_private *tp,
* If skb is NULL then we come here again once a tx irq is
* triggered after the last fragment is marked transmitted.
*/
- if (tp->cur_tx != dirty_tx && skb)
+ if (READ_ONCE(tp->cur_tx) != dirty_tx && skb)
rtl8169_doorbell(tp);
}
}