netlink_broadcast_filtered() calls yield() when a slow listener causes
the buffer to fill. yield() is the wrong choice here, as pointed out by
Commit 8e3fabfde4 (sched: Update yield() docs); in some cases, its use causes
"BUG: scheduling while atomic" and, when fewer cores are available,
kernel hangs:
(note: "ccp" is a kernel module which multicasts netlink messages upon
certain TCP events)
May 17 17:33:56 ccp kernel: [ 394.493978] BUG: scheduling while
atomic: iperf/4744/0x00000101
May 17 17:33:56 ccp kernel: [ 394.493979] Modules linked in: ccp(OE)
snd_intel8x0 snd_ac97_codec ac97_bus snd_pcm crct10dif_pclmul
crc32_pclmul ghash_clmulni_intel snd_seq_midi snd_seq_midi_event
snd_rawmidi snd_seq joydev pcbc snd_seq
_device snd_timer snd vboxvideo aesni_intel aes_x86_64 i2c_piix4
input_leds crypto_simd glue_helper cryptd soundcore mac_hid vboxguest
intel_rapl_perf serio_raw ttm drm_kms_helper drm fb_sys_fops
syscopyarea sysfillrect sysimgblt parpor
t_pc ppdev lp parport ip_tables x_tables autofs4 hid_generic usbhid
hid psmouse ahci e1000 libahci pata_acpi fjes video [last unloaded:
ccp]
May 17 17:33:56 ccp kernel: [ 394.493997] CPU: 0 PID: 4744 Comm:
iperf Tainted: G W OE 4.10.0-21-generic #23-Ubuntu
May 17 17:33:56 ccp kernel: [ 394.493997] Hardware name: innotek GmbH
VirtualBox/VirtualBox, BIOS VirtualBox 12/01/2006
May 17 17:33:56 ccp kernel: [ 394.493997] Call Trace:
May 17 17:33:56 ccp kernel: [ 394.493998] <IRQ>
May 17 17:33:56 ccp kernel: [ 394.494000] dump_stack+0x63/0x81
May 17 17:33:56 ccp kernel: [ 394.494001] __schedule_bug+0x54/0x70
May 17 17:33:56 ccp kernel: [ 394.494002] __schedule+0x536/0x6f0
May 17 17:33:56 ccp kernel: [ 394.494004] schedule+0x36/0x80
May 17 17:33:56 ccp kernel: [ 394.494005] sys_sched_yield+0x4f/0x60
May 17 17:33:56 ccp kernel: [ 394.494005] yield+0x33/0x40
May 17 17:33:56 ccp kernel: [ 394.494006]
netlink_broadcast_filtered+0x29b/0x3c0
May 17 17:33:56 ccp kernel: [ 394.494007] netlink_broadcast+0x1d/0x20
...
May 17 17:33:56 ccp kernel: [ 394.494080] softirq: huh, entered
softirq 3 NET_RX ffffffff85fb5c60 with preempt_count 00000101, exited
with 00000000?
Changing this call to cond_resched() prevents this.
---
net/netlink/af_netlink.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/net/netlink/af_netlink.c b/net/netlink/af_netlink.c
index 161b628..f70716c 100644
--- a/net/netlink/af_netlink.c
+++ b/net/netlink/af_netlink.c
@@ -1441,7 +1441,7 @@ int netlink_broadcast_filtered(struct sock *ssk,
struct sk_buff *skb, u32 portid
if (info.delivered) {
if (info.congested && gfpflags_allow_blocking(allocation))
- yield();
+ cond_resched();
return 0;
}
return -ESRCH;
--
2.7.4
