[RFC PATCH 30/30] softirq: Tasklet/net-rx fixup
From: Frederic Weisbecker
Tasklets and net-rx vectors don't quite get along. If one is interrupted
by another, we may run into a nasty spin_lock recursion:
[ 135.427198] Call Trace:
[ 135.429650]
[ 135.431690] dump_stack+0x67/0x95
[ 135.435024] spin_bug+0x95/0xf0
[ 135.438187] do_raw_spin_lock+0x77/0xa0
[ 135.442079] _raw_spin_lock_nested+0x40/0x50
[ 135.446439] ? tcp_v4_rcv+0x9da/0xb10
[ 135.450131] tcp_v4_rcv+0x9da/0xb10
[ 135.453650] ? ip_local_deliver+0x78/0x260
[ 135.457758] ip_local_deliver+0xdf/0x260
[ 135.461728] ip_rcv+0x4e/0x80
[ 135.464716] __netif_receive_skb_one_core+0x55/0x80
[ 135.469623] __netif_receive_skb+0x1b/0x70
[ 135.473757] netif_receive_skb_internal+0x92/0x390
[ 135.478574] napi_gro_receive+0xdf/0x1a0
[ 135.482545] rtl8169_poll+0x2b8/0x670
[ 135.486211] net_rx_action+0x1f8/0x3e0
[ 135.489989] __do_softirq+0x1a0/0x63c
[ 135.493691] irq_exit+0x10f/0x120
[ 135.497033] do_IRQ+0x71/0x130
[ 135.500137] common_interrupt+0xf/0xf
[ 135.503839] RIP: 0010:_raw_spin_unlock_irqrestore+0x59/0x70
[ 135.509471] Code: 75 21 53 9d e8 e8 1d a5 ff bf 01 00 00 00 e8 8e f6
97 ff 65 8b 05 ef 06 8d 7e 85 c0 74 0e 5b 41 5c 5d c3 e8 c9 20 a5 ff 53 9d
dd e8 90 d4 8b ff 5b 41 5c 5d c3 66 66 2e 0f 1f 84 00 00 00 00
[ 135.528347] RSP: 0018:88021fb03d28 EFLAGS: 0246 ORIG_RAX:
ffde
[ 135.535989] RAX: 880217762480 RBX: 0246 RCX:
0002
[ 135.543201] RDX: RSI: 880217762c70 RDI:
880217762480
[ 135.550332] RBP: 88021fb03d38 R08: 0001 R09:
[ 135.557519] R10: R11: R12:
88021fa59b40
[ 135.564719] R13: 88021fa59b40 R14: fffd78b5 R15:
0e01
[ 135.571905] ? common_interrupt+0xa/0xf
[ 135.575783] mod_timer+0x196/0x440
[ 135.579221] sk_reset_timer+0x18/0x30
[ 135.582940] tcp_schedule_loss_probe+0xe9/0x120
[ 135.587515] tcp_write_xmit+0x2c4/0x1240
[ 135.591468] tcp_tsq_write.part.46+0x5e/0xb0
[ 135.595756] tcp_tsq_handler+0xa3/0xb0
[ 135.599534] tcp_tasklet_func+0xdc/0x120
[ 135.603488] tasklet_action_common.isra.17+0xa3/0xb0
[ 135.608471] tasklet_action+0x2d/0x30
[ 135.612161] __do_softirq+0x1a0/0x63c
[ 135.615847] irq_exit+0x10f/0x120
[ 135.619173] do_IRQ+0x71/0x130
[ 135.622251] common_interrupt+0xf/0xf
[ 135.625949]
This is an ugly workaround until we find a proper solution.
Signed-off-by: Frederic Weisbecker
Cc: Ingo Molnar
Cc: Sebastian Andrzej Siewior
Cc: Thomas Gleixner
Cc: Peter Zijlstra
Cc: Linus Torvalds
Cc: David S. Miller
Cc: Mauro Carvalho Chehab
Cc: Paul E. McKenney
---
kernel/softirq.c | 13 +
1 file changed, 13 insertions(+)
diff --git a/kernel/softirq.c b/kernel/softirq.c
index f4cb1ea..d95295f 100644
--- a/kernel/softirq.c
+++ b/kernel/softirq.c
@@ -320,6 +320,7 @@ asmlinkage __visible void __softirq_entry __do_softirq(void)
unsigned long old_flags = current->flags;
int max_restart = MAX_SOFTIRQ_RESTART;
struct softirq_action *h;
+ bool tasklet_enabled = false, net_rx_enabled = false;
bool in_hardirq;
__u32 pending;
int softirq_bit;
@@ -338,6 +339,10 @@ asmlinkage __visible void __softirq_entry
__do_softirq(void)
in_hardirq = lockdep_softirq_start();
restart:
+ if (local_softirq_enabled() & TASKLET_SOFTIRQ)
+ tasklet_enabled = true;
+ if (local_softirq_enabled() & NET_RX_SOFTIRQ)
+ net_rx_enabled = true;
/* Reset the pending bitmask before enabling irqs */
softirq_pending_nand(pending);
@@ -358,8 +363,16 @@ asmlinkage __visible void __softirq_entry
__do_softirq(void)
trace_softirq_entry(vec_nr);
softirq_enabled_nand(BIT(vec_nr));
+ if (vec_nr == NET_RX_SOFTIRQ && tasklet_enabled)
+ softirq_enabled_nand(BIT(TASKLET_SOFTIRQ));
+ if (vec_nr == TASKLET_SOFTIRQ && net_rx_enabled)
+ softirq_enabled_nand(BIT(NET_RX_SOFTIRQ));
barrier();
h->action(h);
+ if (vec_nr == TASKLET_SOFTIRQ && net_rx_enabled)
+ softirq_enabled_or(BIT(NET_RX_SOFTIRQ));
+ if (vec_nr == NET_RX_SOFTIRQ && tasklet_enabled)
+ softirq_enabled_or(BIT(TASKLET_SOFTIRQ));
softirq_enabled_or(BIT(vec_nr));
trace_softirq_exit(vec_nr);
if (unlikely(prev_count != preempt_count())) {
--
2.7.4
[RFC PATCH 30/30] softirq: Tasklet/net-rx fixup
From: Frederic Weisbecker
Tasklets and net-rx vectors don't quite get along. If one is interrupted
by another, we may run into a nasty spin_lock recursion:
[ 135.427198] Call Trace:
[ 135.429650]
[ 135.431690] dump_stack+0x67/0x95
[ 135.435024] spin_bug+0x95/0xf0
[ 135.438187] do_raw_spin_lock+0x77/0xa0
[ 135.442079] _raw_spin_lock_nested+0x40/0x50
[ 135.446439] ? tcp_v4_rcv+0x9da/0xb10
[ 135.450131] tcp_v4_rcv+0x9da/0xb10
[ 135.453650] ? ip_local_deliver+0x78/0x260
[ 135.457758] ip_local_deliver+0xdf/0x260
[ 135.461728] ip_rcv+0x4e/0x80
[ 135.464716] __netif_receive_skb_one_core+0x55/0x80
[ 135.469623] __netif_receive_skb+0x1b/0x70
[ 135.473757] netif_receive_skb_internal+0x92/0x390
[ 135.478574] napi_gro_receive+0xdf/0x1a0
[ 135.482545] rtl8169_poll+0x2b8/0x670
[ 135.486211] net_rx_action+0x1f8/0x3e0
[ 135.489989] __do_softirq+0x1a0/0x63c
[ 135.493691] irq_exit+0x10f/0x120
[ 135.497033] do_IRQ+0x71/0x130
[ 135.500137] common_interrupt+0xf/0xf
[ 135.503839] RIP: 0010:_raw_spin_unlock_irqrestore+0x59/0x70
[ 135.509471] Code: 75 21 53 9d e8 e8 1d a5 ff bf 01 00 00 00 e8 8e f6
97 ff 65 8b 05 ef 06 8d 7e 85 c0 74 0e 5b 41 5c 5d c3 e8 c9 20 a5 ff 53 9d
dd e8 90 d4 8b ff 5b 41 5c 5d c3 66 66 2e 0f 1f 84 00 00 00 00
[ 135.528347] RSP: 0018:88021fb03d28 EFLAGS: 0246 ORIG_RAX:
ffde
[ 135.535989] RAX: 880217762480 RBX: 0246 RCX:
0002
[ 135.543201] RDX: RSI: 880217762c70 RDI:
880217762480
[ 135.550332] RBP: 88021fb03d38 R08: 0001 R09:
[ 135.557519] R10: R11: R12:
88021fa59b40
[ 135.564719] R13: 88021fa59b40 R14: fffd78b5 R15:
0e01
[ 135.571905] ? common_interrupt+0xa/0xf
[ 135.575783] mod_timer+0x196/0x440
[ 135.579221] sk_reset_timer+0x18/0x30
[ 135.582940] tcp_schedule_loss_probe+0xe9/0x120
[ 135.587515] tcp_write_xmit+0x2c4/0x1240
[ 135.591468] tcp_tsq_write.part.46+0x5e/0xb0
[ 135.595756] tcp_tsq_handler+0xa3/0xb0
[ 135.599534] tcp_tasklet_func+0xdc/0x120
[ 135.603488] tasklet_action_common.isra.17+0xa3/0xb0
[ 135.608471] tasklet_action+0x2d/0x30
[ 135.612161] __do_softirq+0x1a0/0x63c
[ 135.615847] irq_exit+0x10f/0x120
[ 135.619173] do_IRQ+0x71/0x130
[ 135.622251] common_interrupt+0xf/0xf
[ 135.625949]
This is an ugly workaround until we find a proper solution.
Signed-off-by: Frederic Weisbecker
Cc: Ingo Molnar
Cc: Sebastian Andrzej Siewior
Cc: Thomas Gleixner
Cc: Peter Zijlstra
Cc: Linus Torvalds
Cc: David S. Miller
Cc: Mauro Carvalho Chehab
Cc: Paul E. McKenney
---
kernel/softirq.c | 13 +
1 file changed, 13 insertions(+)
diff --git a/kernel/softirq.c b/kernel/softirq.c
index f4cb1ea..d95295f 100644
--- a/kernel/softirq.c
+++ b/kernel/softirq.c
@@ -320,6 +320,7 @@ asmlinkage __visible void __softirq_entry __do_softirq(void)
unsigned long old_flags = current->flags;
int max_restart = MAX_SOFTIRQ_RESTART;
struct softirq_action *h;
+ bool tasklet_enabled = false, net_rx_enabled = false;
bool in_hardirq;
__u32 pending;
int softirq_bit;
@@ -338,6 +339,10 @@ asmlinkage __visible void __softirq_entry
__do_softirq(void)
in_hardirq = lockdep_softirq_start();
restart:
+ if (local_softirq_enabled() & TASKLET_SOFTIRQ)
+ tasklet_enabled = true;
+ if (local_softirq_enabled() & NET_RX_SOFTIRQ)
+ net_rx_enabled = true;
/* Reset the pending bitmask before enabling irqs */
softirq_pending_nand(pending);
@@ -358,8 +363,16 @@ asmlinkage __visible void __softirq_entry
__do_softirq(void)
trace_softirq_entry(vec_nr);
softirq_enabled_nand(BIT(vec_nr));
+ if (vec_nr == NET_RX_SOFTIRQ && tasklet_enabled)
+ softirq_enabled_nand(BIT(TASKLET_SOFTIRQ));
+ if (vec_nr == TASKLET_SOFTIRQ && net_rx_enabled)
+ softirq_enabled_nand(BIT(NET_RX_SOFTIRQ));
barrier();
h->action(h);
+ if (vec_nr == TASKLET_SOFTIRQ && net_rx_enabled)
+ softirq_enabled_or(BIT(NET_RX_SOFTIRQ));
+ if (vec_nr == NET_RX_SOFTIRQ && tasklet_enabled)
+ softirq_enabled_or(BIT(TASKLET_SOFTIRQ));
softirq_enabled_or(BIT(vec_nr));
trace_softirq_exit(vec_nr);
if (unlikely(prev_count != preempt_count())) {
--
2.7.4
