Commit 719c57197010 ("net: make napi_disable() symmetric with
enable") accidentally introduced a bug sometimes leading to a kernel
BUG when bringing an iface up/down under heavy traffic load.
Prior to this commit, napi_disable() was polling n->state until
none of (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC) is set and then
always flip them. Now there's a possibility to get away with the
NAPIF_STATE_SCHE unset as 'continue' drops us to the cmpxchg()
call with an unitialized variable, rather than straight to
another round of the state check.
Error path looks like:
napi_disable():
unsigned long val, new; /* new is uninitialized */
do {
val = READ_ONCE(n->state); /* NAPIF_STATE_NPSVC and/or
NAPIF_STATE_SCHED is set */
if (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { /* true */
usleep_range(20, 200);
continue; /* go straight to the condition check */
}
new = val | <...>
} while (cmpxchg(&n->state, val, new) != val); /* state == val, cmpxchg()
writes garbage */
napi_enable():
do {
val = READ_ONCE(n->state);
BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); /* 50/50 boom */
<...>
while the typical BUG splat is like:
[ 172.652461] ------------[ cut here ]------------
[ 172.652462] kernel BUG at net/core/dev.c:6937!
[ 172.656914] invalid opcode: 0000 [#1] PREEMPT SMP PTI
[ 172.661966] CPU: 36 PID: 2829 Comm: xdp_redirect_cp Tainted: G I 5.15.0 #42
[ 172.670222] Hardware name: Intel Corporation S2600WFT/S2600WFT, BIOS SE5C620.86B.02.01.0014.082620210524 08/26/2021
[ 172.680646] RIP: 0010:napi_enable+0x5a/0xd0
[ 172.684832] Code: 07 49 81 cc 00 01 00 00 4c 89 e2 48 89 d8 80 e6 fb f0 48 0f b1 55 10 48 39 c3 74 10 48 8b 5d 10 f6 c7 04 75 3d f6 c3 01 75 b4 <0f> 0b 5b 5d 41 5c c3 65 ff 05 b8 e5 61 53 48 c7 c6 c0 f3 34 ad 48
[ 172.703578] RSP: 0018:ffffa3c9497477a8 EFLAGS: 00010246
[ 172.708803] RAX: ffffa3c96615a014 RBX: 0000000000000000 RCX: ffff8a4b575301a0
< snip >
[ 172.782403] Call Trace:
[ 172.784857] <TASK>
[ 172.786963] ice_up_complete+0x6f/0x210 [ice]
[ 172.791349] ice_xdp+0x136/0x320 [ice]
[ 172.795108] ? ice_change_mtu+0x180/0x180 [ice]
[ 172.799648] dev_xdp_install+0x61/0xe0
[ 172.803401] dev_xdp_attach+0x1e0/0x550
[ 172.807240] dev_change_xdp_fd+0x1e6/0x220
[ 172.811338] do_setlink+0xee8/0x1010
[ 172.814917] rtnl_setlink+0xe5/0x170
[ 172.818499] ? bpf_lsm_binder_set_context_mgr+0x10/0x10
[ 172.823732] ? security_capable+0x36/0x50
< snip >
Fix this by replacing 'do { } while (cmpxchg())' with an "infinite"
for-loop with an explicit break.
From v1 [0]:
- just use a for-loop to simplify both the fix and the existing
code (Eric).
[0] https://lore.kernel.org/netdev/[email protected]
Fixes: 719c57197010 ("net: make napi_disable() symmetric with enable")
Suggested-by: Eric Dumazet <[email protected]> # for-loop
Signed-off-by: Alexander Lobakin <[email protected]>
Reviewed-by: Jesse Brandeburg <[email protected]>
---
net/core/dev.c | 7 +++++--
1 file changed, 5 insertions(+), 2 deletions(-)
diff --git a/net/core/dev.c b/net/core/dev.c
index edeb811c454e..15ac064b5562 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -6928,7 +6928,7 @@ void napi_disable(struct napi_struct *n)
might_sleep();
set_bit(NAPI_STATE_DISABLE, &n->state);
- do {
+ for ( ; ; ) {
val = READ_ONCE(n->state);
if (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) {
usleep_range(20, 200);
@@ -6937,7 +6937,10 @@ void napi_disable(struct napi_struct *n)
new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC;
new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL);
- } while (cmpxchg(&n->state, val, new) != val);
+
+ if (cmpxchg(&n->state, val, new) == val)
+ break;
+ }
hrtimer_cancel(&n->timer);
--
2.33.1
On Wed, Nov 10, 2021 at 11:56 AM Alexander Lobakin
<[email protected]> wrote:
>
> Commit 719c57197010 ("net: make napi_disable() symmetric with
> enable") accidentally introduced a bug sometimes leading to a kernel
> BUG when bringing an iface up/down under heavy traffic load.
>
> Prior to this commit, napi_disable() was polling n->state until
> none of (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC) is set and then
> always flip them. Now there's a possibility to get away with the
> NAPIF_STATE_SCHE unset as 'continue' drops us to the cmpxchg()
> call with an unitialized variable, rather than straight to
> another round of the state check.
>
...
>
> [0] https://lore.kernel.org/netdev/[email protected]
>
> Fixes: 719c57197010 ("net: make napi_disable() symmetric with enable")
> Suggested-by: Eric Dumazet <[email protected]> # for-loop
> Signed-off-by: Alexander Lobakin <[email protected]>
> Reviewed-by: Jesse Brandeburg <[email protected]>
> ---
Thanks a lot !
Reviewed-by: Eric Dumazet <[email protected]>
Hello:
This patch was applied to netdev/net.git (master)
by Jakub Kicinski <[email protected]>:
On Wed, 10 Nov 2021 20:56:05 +0100 you wrote:
> Commit 719c57197010 ("net: make napi_disable() symmetric with
> enable") accidentally introduced a bug sometimes leading to a kernel
> BUG when bringing an iface up/down under heavy traffic load.
>
> Prior to this commit, napi_disable() was polling n->state until
> none of (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC) is set and then
> always flip them. Now there's a possibility to get away with the
> NAPIF_STATE_SCHE unset as 'continue' drops us to the cmpxchg()
> call with an unitialized variable, rather than straight to
> another round of the state check.
>
> [...]
Here is the summary with links:
- [v2,net] net: fix premature exit from NAPI state polling in napi_disable()
https://git.kernel.org/netdev/net/c/0315a075f134
You are awesome, thank you!
--
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html
Alexander Lobakin [[email protected]] wrote:
> Commit 719c57197010 ("net: make napi_disable() symmetric with
> enable") accidentally introduced a bug sometimes leading to a kernel
> BUG when bringing an iface up/down under heavy traffic load.
>
> Prior to this commit, napi_disable() was polling n->state until
> none of (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC) is set and then
> always flip them. Now there's a possibility to get away with the
> NAPIF_STATE_SCHE unset as 'continue' drops us to the cmpxchg()
> call with an unitialized variable, rather than straight to
> another round of the state check.
Thanks. Tested v1 and it fixes the problem discussed at:
https://lore.kernel.org/netdev/[email protected]/
Sukadev