Re: [BUG] 4.10-rc8 - ping spinning?
From: Soheil Hassas Yeganeh
Date: Thu Feb 16 2017 - 10:50:54 EST
Thank you Vito for the report.
The patch you cited actually resolves a similar backward compatibility
problem for traceroute.
I suspect the problem here is that there's a local error queued on the
error queue after an ICMP message. ping apparently expect the
sk->sk_err to be set for the local errors as well, and hence the
error. Ideally, ping should read the error queue if there an EPOLLERR,
because local errors never sk->sk_err on their own. That is, if we
have
But as a workaround, would you mind trying the following patch to see
if it resolves the issue reported?
From: Soheil Hassas Yeganeh <soheil@xxxxxxxxxx>
Date: Thu, 16 Feb 2017 10:48:24 -0500
---
net/core/skbuff.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index 734c71468b01..2b774b564024 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -3717,7 +3717,8 @@ EXPORT_SYMBOL(sock_queue_err_skb);
static bool is_icmp_err_skb(const struct sk_buff *skb)
{
return skb && (SKB_EXT_ERR(skb)->ee.ee_origin == SO_EE_ORIGIN_ICMP ||
- SKB_EXT_ERR(skb)->ee.ee_origin == SO_EE_ORIGIN_ICMP6);
+ SKB_EXT_ERR(skb)->ee.ee_origin == SO_EE_ORIGIN_ICMP6 ||
+ SKB_EXT_ERR(skb)->ee.ee_origin == SO_EE_ORIGIN_LOCAL);
}
struct sk_buff *sock_dequeue_err_skb(struct sock *sk)
--
Thanks!
Soheil
On Thu, Feb 16, 2017 at 7:05 AM, <lkml@xxxxxxxxxxx> wrote:
> Hello netdev,
>
> Please see the forwarded message below. This was sent to linux-kernel but
> after digging a little I suspect it's specific to the network stack.
>
> Perusing the net/ changes between 4.9 and 4.10-rc8 this sounded awful related
> to what I'm observing:
>
> commit 83a1a1a70e87f676fbb6086b26b6ac7f7fdd107d
> Author: Soheil Hassas Yeganeh <soheil@xxxxxxxxxx>
> Date: Wed Nov 30 14:01:08 2016 -0500
>
> sock: reset sk_err for ICMP packets read from error queue
>
> Only when ICMP packets are enqueued onto the error queue,
> sk_err is also set. Before f5f99309fa74 (sock: do not set sk_err
> in sock_dequeue_err_skb), a subsequent error queue read
> would set sk_err to the next error on the queue, or 0 if empty.
> As no error types other than ICMP set this field, sk_err should
> not be modified upon dequeuing them.
>
> Only for ICMP errors, reset the (racy) sk_err. Some applications,
> like traceroute, rely on it and go into a futile busy POLLERR
> loop otherwise.
>
> In principle, sk_err has to be set while an ICMP error is queued.
> Testing is_icmp_err_skb(skb_next) approximates this without
> requiring a full queue walk. Applications that receive both ICMP
> and other errors cannot rely on this legacy behavior, as other
> errors do not set sk_err in the first place.
>
> Fixes: f5f99309fa74 (sock: do not set sk_err in sock_dequeue_err_skb)
> Signed-off-by: Soheil Hassas Yeganeh <soheil@xxxxxxxxxx>
> Signed-off-by: Willem de Bruijn <willemb@xxxxxxxxxx>
> Acked-by: Eric Dumazet <edumazet@xxxxxxxxxx>
> Acked-by: Maciej Åenczykowski <maze@xxxxxxxxxx>
> Signed-off-by: David S. Miller <davem@xxxxxxxxxxxxx>
>
> Cheers,
> Vito Caputo
>
>
> ----- Forwarded message from lkml@xxxxxxxxxxx -----
>
> Date: Thu, 16 Feb 2017 03:17:49 -0800
> From: lkml@xxxxxxxxxxx
> To: linux-kernel@xxxxxxxxxxxxxxx
> Subject: Re: [BUG] 4.10-rc8 - ping spinning?
> User-Agent: Mutt/1.5.23 (2014-03-12)
>
> On Thu, Feb 16, 2017 at 03:03:03AM -0800, lkml@xxxxxxxxxxx wrote:
>> Hello list,
>>
>> Some rtl8192cu bugs of old got me in the habit of running ping in a shelved (i.e. forgotten) xterm, a harmless practice which seemed to prevent the rtl8192cu device from dying.
>>
>> This evening the system started getting very slow and to my surprise I found
>> this in `top`:
>> 5115 swivel 30 10 14772 1928 1756 R 90.9 0.0 1351:41 ping
>> 9005 swivel 30 10 14772 1892 1724 R 90.9 0.0 1354:26 ping
>>
>> This is a dual core machine (X61s, core2duo 1.8Ghz), those processes are
>> burning all the free CPU in the system context. They're identical commands,
>> just plain `ping domain.com`, to the same host. It appears I accidentally
>> (fortuitously?) had two running, which made this event more interesting.
>>
>> I can assert that these did not begin spinning simultaneously - as you can see
>> by the cumulative time in `top` there's a small delta. I also use a window
>> manager with builtin continuous process monitoring, and when I noticed this was
>> happening I was able to see that one of the processes had only recently begun
>> spinning, the other was spinning long enough to have its start fall off the
>> chart (at least ~17 minutes ago).
>>
>> This hasn't occurred before AFAIK, but I haven't spent much time in 4.10 yet.
>> I'm pretty confident this didn't happen in 4.9 which I ran for quite a while.
>>
>> `strace` of one of the aforementioned processes:
>>
>> 1487241315.073568 poll([{fd=3, events=POLLIN|POLLERR}], 1, 927) = 1 ([{fd=3, revents=POLLERR}]) <0.000022>
>> 1487241315.073665 recvmsg(3, 0x7ffc8e05e260, MSG_DONTWAIT) = -1 EAGAIN (Resource temporarily unavailable) <0.000020>
>> 1487241315.073747 gettimeofday({1487241315, 73774}, NULL) = 0 <0.000021>
>> 1487241315.073829 poll([{fd=3, events=POLLIN|POLLERR}], 1, 927) = 1 ([{fd=3, revents=POLLERR}]) <0.000025>
>> 1487241315.073927 recvmsg(3, 0x7ffc8e05e260, MSG_DONTWAIT) = -1 EAGAIN (Resource temporarily unavailable) <0.000020>
>> 1487241315.074024 gettimeofday({1487241315, 74050}, NULL) = 0 <0.000256>
>> 1487241315.074352 poll([{fd=3, events=POLLIN|POLLERR}], 1, 927) = 1 ([{fd=3, revents=POLLERR}]) <0.000026>
>> 1487241315.076241 recvmsg(3, 0x7ffc8e05e260, MSG_DONTWAIT) = -1 EAGAIN (Resource temporarily unavailable) <0.000022>
>> 1487241315.076337 gettimeofday({1487241315, 76366}, NULL) = 0 <0.000020>
>> 1487241315.076422 poll([{fd=3, events=POLLIN|POLLERR}], 1, 924) = 1 ([{fd=3, revents=POLLERR}]) <0.000025>
>> 1487241315.076523 recvmsg(3, 0x7ffc8e05e260, MSG_DONTWAIT) = -1 EAGAIN (Resource temporarily unavailable) <0.000025>
>> 1487241315.079770 gettimeofday({1487241315, 79799}, NULL) = 0 <0.000019>
>> 1487241315.079855 poll([{fd=3, events=POLLIN|POLLERR}], 1, 921) = 1 ([{fd=3, revents=POLLERR}]) <0.000024>
>> 1487241315.079956 recvmsg(3, 0x7ffc8e05e260, MSG_DONTWAIT) = -1 EAGAIN (Resource temporarily unavailable) <0.000021>
>> 1487241315.080057 gettimeofday({1487241315, 80084}, NULL) = 0 <0.000020>
>> 1487241315.080140 poll([{fd=3, events=POLLIN|POLLERR}], 1, 921) = 1 ([{fd=3, revents=POLLERR}]) <0.000024>
>> 1487241315.080238 recvmsg(3, 0x7ffc8e05e260, MSG_DONTWAIT) = -1 EAGAIN (Resource temporarily unavailable) <0.000021>
>> 1487241315.080322 gettimeofday({1487241315, 80350}, NULL) = 0 <0.000020>
>> 1487241315.080406 poll([{fd=3, events=POLLIN|POLLERR}], 1, 920) = 1 ([{fd=3, revents=POLLERR}]) <0.000023>
>> 1487241315.080502 recvmsg(3, 0x7ffc8e05e260, MSG_DONTWAIT) = -1 EAGAIN (Resource temporarily unavailable) <0.000019>
>> 1487241315.080583 gettimeofday({1487241315, 80610}, NULL) = 0 <0.000018>
>> 1487241315.080663 poll([{fd=3, events=POLLIN|POLLERR}], 1, 920) = 1 ([{fd=3, revents=POLLERR}]) <0.000024>
>> 1487241315.080761 recvmsg(3, 0x7ffc8e05e260, MSG_DONTWAIT) = -1 EAGAIN (Resource temporarily unavailable) <0.000020>
>> 1487241315.080843 gettimeofday({1487241315, 80870}, NULL) = 0 <0.000020>
>> 1487241315.080925 poll([{fd=3, events=POLLIN|POLLERR}], 1, 920) = 1 ([{fd=3, revents=POLLERR}]) <0.000037>
>> 1487241315.081037 recvmsg(3, 0x7ffc8e05e260, MSG_DONTWAIT) = -1 EAGAIN (Resource temporarily unavailable) <0.000020>
>> 1487241315.081119 gettimeofday({1487241315, 81147}, NULL) = 0 <0.000020>
>>
>
> It's worth noting that ping is still otherwise functioning correctly, despite
> the POLLERR:
>
> 1487242826.169502 sendmsg(3, {msg_name(16)={sa_family=AF_INET, sin_port=htons(0), sin_addr=inet_addr("xx.xxx.xxx.xxx")}, msg_iov(1)=[{"\10\0\245G\23\373\243uJ\206\245X\0\0\0\0\352\225\2\0\0\0\0\0\20\21\22\23\24\25\26\27"..., 64}], msg_controllen=0, msg_flags=0}, MSG_CONFIRM) = 64 <0.000133>
> 1487242826.169757 recvmsg(3, {msg_name(16)={sa_family=AF_INET, sin_port=htons(0), sin_addr=inet_addr("66.240.222.126")}, msg_iov(1)=[{"E \0T\345\364\0\0002\1w\23B\360\336~\n\0\0\23\0\0\255G\23\373\243uJ\206\245X"..., 192}], msg_controllen=32, {cmsg_len=32, cmsg_level=SOL_SOCKET, cmsg_type=0x1d /* SCM_??? */, ...}, msg_flags=0}, 0) = 84 <0.028825>
> 1487242826.198697 write(1, "64 bytes from xxxxxxx.com (xx.xx"..., 79) = 79 <0.000639>
> 1487242826.199405 gettimeofday({1487242826, 199430}, NULL) = 0 <0.000023>
> 1487242826.199486 poll([{fd=3, events=POLLIN|POLLERR}], 1, 970) = 1 ([{fd=3, revents=POLLERR}]) <0.000026>
> 1487242826.199578 recvmsg(3, 0x7ffc8e05e260, MSG_DONTWAIT) = -1 EAGAIN (Resource temporarily unavailable) <0.000024>
>
> Surprisingly ping doesn't seem to be reacting to the POLLERR though it
> requested it. Maybe this is just a ping bug? Though I haven't seen this
> before rc8.
>
> Regards,
> Vito Caputo
>
> ----- End forwarded message -----