Skip to contentSkip to conversation

ACPI: Fix wrong atomicity check in preemption point

PATCH In review Frederic Weisbecker 3 participants 3 messages Aug 7, 2010 – Aug 12, 2010

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ACPI: Fix wrong atomicity check in preemption point

The acpi preemption point checks the atomicity of the context using in_atomic_preempt_off(). This helper must be used only to check the atomicity before a prior call to preempt_disable(), which is not what we want here.

What we want is to simply check if we are in an atomic section. This helper is actually only used by the scheduler for particular needs and shouldn't be used outside.

The check made here is then always wrong. We will schedule only if preemption has been disabled once. It never has been a problem during the boot because premption is disabled and moreover the BKL is held, so we increase twice the preempt count. But now that we drop the bkl from the boot, the preempt count is only increased once, and then we schedule in the acpi preemption point while we shouldn't.

In fact using such in_atomic*() like helpers is quite fragile to guess if we can schedule, but still, in_atomic() is less buggy than what was there before.

This fixes:

[    0.008086] BUG: scheduling while atomic: swapper/0/0x10000002
[    0.008167] no locks held by swapper/0.
[    0.008243] Modules linked in:
[    0.008356] Pid: 0, comm: swapper Not tainted 2.6.35+ #793
[    0.008437] Call Trace:
[    0.008519]  [<ffffffff8106eab3>] ? __debug_show_held_locks+0x13/0x30
[    0.008605]  [<ffffffff81039a65>] __schedule_bug+0x85/0x90
[    0.008690]  [<ffffffff815edf20>] schedule+0x670/0x840
[    0.008775]  [<ffffffff8129ff88>] ? acpi_os_release_object+0x9/0xd
[    0.008860]  [<ffffffff812beca0>] ? acpi_ps_free_op+0x22/0x24
[    0.008944]  [<ffffffff8103ccd5>] __cond_resched+0x25/0x40
[    0.009008]  [<ffffffff815ee1ed>] _cond_resched+0x2d/0x40
[    0.009091]  [<ffffffff812bdf4a>] acpi_ps_complete_op+0x292/0x2a8
[    0.009174]  [<ffffffff812be7b6>] acpi_ps_parse_loop+0x856/0x9ac
[    0.010008]  [<ffffffff812bd81d>] acpi_ps_parse_aml+0x9a/0x2b9
[    0.010092]  [<ffffffff812bc048>] acpi_ns_one_complete_parse+0xfc/0x117
[    0.010176]  [<ffffffff812bc07f>] acpi_ns_parse_table+0x1c/0x35
[    0.010259]  [<ffffffff812b9606>] acpi_ns_load_table+0x4a/0x8c
[    0.010343]  [<ffffffff812c075f>] acpi_load_tables+0xa0/0x164
[    0.010429]  [<ffffffff819751e1>] ? acpi_initialize_subsystem+0x69/0x91
[    0.010513]  [<ffffffff819740df>] acpi_early_init+0x6c/0xf7
[    0.010598]  [<ffffffff8194fd68>] start_kernel+0x3b3/0x3fb
[    0.010681]  [<ffffffff8194f26d>] x86_64_start_reservations+0x7d/0x89
[    0.010765]  [<ffffffff8194f359>] x86_64_start_kernel+0xe0/0xf2
---
 include/acpi/platform/aclinux.h |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
include/acpi/platform/aclinux.h +1 -1
index e5039a2..8da1e8c 100644
@@ -152,7 +152,7 @@ static inline void *acpi_os_acquire_object(acpi_cache_t * cache)
 #include <linux/hardirq.h>
 #define ACPI_PREEMPTION_POINT() \
 	do { \
-		if (!in_atomic_preempt_off() && !irqs_disabled()) \
+		if (!in_atomic() && !irqs_disabled()) \
 			cond_resched(); \
 	} while (0)
 
Signature · 1 line
1.6.2.3
#1 lore raw

I'll be happy to include this in the aclinux.h file if the day ever comes when it is stable.

Quoted 78 lines

-----Original Message-----
From: Frederic Weisbecker [mailto:[email protected]]
Sent: Friday, August 06, 2010 8:39 PM
To: Brown, Len
Cc: LKML; Frederic Weisbecker; Moore, Robert
Subject: [PATCH] ACPI: Fix wrong atomicity check in preemption point

The acpi preemption point checks the atomicity of the context using in_atomic_preempt_off(). This helper must be used only to check the atomicity before a prior call to preempt_disable(), which is not what we want here.

What we want is to simply check if we are in an atomic section. This helper is actually only used by the scheduler for particular needs and shouldn't be used outside.

The check made here is then always wrong. We will schedule only if preemption has been disabled once. It never has been a problem during the boot because premption is disabled and moreover the BKL is held, so we increase twice the preempt count. But now that we drop the bkl from the boot, the preempt count is only increased once, and then we schedule in the acpi preemption point while we shouldn't.

In fact using such in_atomic*() like helpers is quite fragile to guess if we can schedule, but still, in_atomic() is less buggy than what was there before.

This fixes:

[    0.008086] BUG: scheduling while atomic: swapper/0/0x10000002
[    0.008167] no locks held by swapper/0.
[    0.008243] Modules linked in:
[    0.008356] Pid: 0, comm: swapper Not tainted 2.6.35+ #793
[    0.008437] Call Trace:
[    0.008519]  [<ffffffff8106eab3>] ? __debug_show_held_locks+0x13/0x30
[    0.008605]  [<ffffffff81039a65>] __schedule_bug+0x85/0x90
[    0.008690]  [<ffffffff815edf20>] schedule+0x670/0x840
[    0.008775]  [<ffffffff8129ff88>] ? acpi_os_release_object+0x9/0xd
[    0.008860]  [<ffffffff812beca0>] ? acpi_ps_free_op+0x22/0x24
[    0.008944]  [<ffffffff8103ccd5>] __cond_resched+0x25/0x40
[    0.009008]  [<ffffffff815ee1ed>] _cond_resched+0x2d/0x40
[    0.009091]  [<ffffffff812bdf4a>] acpi_ps_complete_op+0x292/0x2a8
[    0.009174]  [<ffffffff812be7b6>] acpi_ps_parse_loop+0x856/0x9ac
[    0.010008]  [<ffffffff812bd81d>] acpi_ps_parse_aml+0x9a/0x2b9
[    0.010092]  [<ffffffff812bc048>] acpi_ns_one_complete_parse+0xfc/0x117
[    0.010176]  [<ffffffff812bc07f>] acpi_ns_parse_table+0x1c/0x35
[    0.010259]  [<ffffffff812b9606>] acpi_ns_load_table+0x4a/0x8c
[    0.010343]  [<ffffffff812c075f>] acpi_load_tables+0xa0/0x164
[    0.010429]  [<ffffffff819751e1>] ? acpi_initialize_subsystem+0x69/0x91
[    0.010513]  [<ffffffff819740df>] acpi_early_init+0x6c/0xf7
[    0.010598]  [<ffffffff8194fd68>] start_kernel+0x3b3/0x3fb
[    0.010681]  [<ffffffff8194f26d>] x86_64_start_reservations+0x7d/0x89
[    0.010765]  [<ffffffff8194f359>] x86_64_start_kernel+0xe0/0xf2

---
include/acpi/platform/aclinux.h | 2 +-

1 files changed, 1 insertions(+), 1 deletions(-)
include/acpi/platform/aclinux.h +1 -2
index e5039a2..8da1e8c 100644
@@ -152,7 +152,7 @@ static inline void *acpi_os_acquire_object(acpi_cache_t
 * cache)
 #include <linux/hardirq.h>
 #define ACPI_PREEMPTION_POINT() \
 	do { \
-		if (!in_atomic_preempt_off() && !irqs_disabled()) \
+		if (!in_atomic() && !irqs_disabled()) \
 			cond_resched(); \
 	} while (0)
 
--

1.6.2.3

#2 lore raw
Quoted 56 lines · Robert
Quoted 46 lines

-----Original Message-----
From: Frederic Weisbecker [mailto:[email protected]]
Sent: Friday, August 06, 2010 8:39 PM
To: Brown, Len
Cc: LKML; Frederic Weisbecker; Moore, Robert
Subject: [PATCH] ACPI: Fix wrong atomicity check in preemption point

The acpi preemption point checks the atomicity of the context using in_atomic_preempt_off(). This helper must be used only to check the atomicity before a prior call to preempt_disable(), which is not what we want here.

What we want is to simply check if we are in an atomic section. This helper is actually only used by the scheduler for particular needs and shouldn't be used outside.

The check made here is then always wrong. We will schedule only if preemption has been disabled once. It never has been a problem during the boot because premption is disabled and moreover the BKL is held, so we increase twice the preempt count. But now that we drop the bkl from the boot, the preempt count is only increased once, and then we schedule in the acpi preemption point while we shouldn't.

In fact using such in_atomic*() like helpers is quite fragile to guess if we can schedule, but still, in_atomic() is less buggy than what was there before.

This fixes:

[    0.008086] BUG: scheduling while atomic: swapper/0/0x10000002
[    0.008167] no locks held by swapper/0.
[    0.008243] Modules linked in:
[    0.008356] Pid: 0, comm: swapper Not tainted 2.6.35+ #793
[    0.008437] Call Trace:
[    0.008519]  [<ffffffff8106eab3>] ? __debug_show_held_locks+0x13/0x30
[    0.008605]  [<ffffffff81039a65>] __schedule_bug+0x85/0x90
[    0.008690]  [<ffffffff815edf20>] schedule+0x670/0x840
[    0.008775]  [<ffffffff8129ff88>] ? acpi_os_release_object+0x9/0xd
[    0.008860]  [<ffffffff812beca0>] ? acpi_ps_free_op+0x22/0x24
[    0.008944]  [<ffffffff8103ccd5>] __cond_resched+0x25/0x40
[    0.009008]  [<ffffffff815ee1ed>] _cond_resched+0x2d/0x40
[    0.009091]  [<ffffffff812bdf4a>] acpi_ps_complete_op+0x292/0x2a8
[    0.009174]  [<ffffffff812be7b6>] acpi_ps_parse_loop+0x856/0x9ac
[    0.010008]  [<ffffffff812bd81d>] acpi_ps_parse_aml+0x9a/0x2b9
[    0.010092]  [<ffffffff812bc048>] acpi_ns_one_complete_parse+0xfc/0x117
Quoted 9 lines
[    0.010176]  [<ffffffff812bc07f>] acpi_ns_parse_table+0x1c/0x35
[    0.010259]  [<ffffffff812b9606>] acpi_ns_load_table+0x4a/0x8c
[    0.010343]  [<ffffffff812c075f>] acpi_load_tables+0xa0/0x164
[    0.010429]  [<ffffffff819751e1>] ? acpi_initialize_subsystem+0x69/0x91
[    0.010513]  [<ffffffff819740df>] acpi_early_init+0x6c/0xf7
[    0.010598]  [<ffffffff8194fd68>] start_kernel+0x3b3/0x3fb
[    0.010681]  [<ffffffff8194f26d>] x86_64_start_reservations+0x7d/0x89
[    0.010765]  [<ffffffff8194f359>] x86_64_start_kernel+0xe0/0xf2
Quoted 3 lines

I'll be happy to include this in the aclinux.h file if the day ever comes when it is stable.

(top-posting repaired. Please don't.)

Linus merged the below fix yesterday. It's tagged for -stable backporting.

From: Thomas Gleixner <[email protected]>

The ACPI_PREEMPTION_POINT() logic was introduced in commit 8bd108d (ACPICA: add preemption point after each opcode parse). The follow up commits abe1dfab6, 138d15692, c084ca70 tried to fix the preemption logic back and forth, but nobody noticed that the usage of in_atomic_preempt_off() in that context is wrong.

The check which guards the call of cond_resched() is:

    if (!in_atomic_preempt_off() && !irqs_disabled())

in_atomic_preempt_off() is not intended for general use as the comment above the macro definition clearly says:

* Check whether we were atomic before we did preempt_disable():
* (used by the scheduler, *after* releasing the kernel lock)

On a CONFIG_PREEMPT=n kernel the usage of in_atomic_preempt_off() works by accident, but with CONFIG_PREEMPT=y it's just broken.

The whole purpose of the ACPI_PREEMPTION_POINT() is to reduce the latency on a CONFIG_PREEMPT=n kernel, so make ACPI_PREEMPTION_POINT() depend on CONFIG_PREEMPT=n and remove the in_atomic_preempt_off() check.

Addresses https://bugzilla.kernel.org/show_bug.cgi?id=16210

---

 drivers/acpi/apei/erst.c        |    1 +
 include/acpi/platform/aclinux.h |   10 +++++++---
 2 files changed, 8 insertions(+), 3 deletions(-)
drivers/acpi/apei/erst.c renamed from drivers/acpi/apei/erst.c~acpi-fix-bogus-preemption-logic +1 -0
diff -puN drivers/acpi/apei/erst.c~acpi-fix-bogus-preemption-logic drivers/acpi/apei/erst.c
@@ -33,6 +33,7 @@
 #include <linux/uaccess.h>
 #include <linux/cper.h>
 #include <linux/nmi.h>
+#include <linux/hardirq.h>
 #include <acpi/apei.h>
 
 #include "apei-internal.h"
include/acpi/platform/aclinux.h renamed from include/acpi/platform/aclinux.h~acpi-fix-bogus-preemption-logic +7 -3
diff -puN include/acpi/platform/aclinux.h~acpi-fix-bogus-preemption-logic include/acpi/platform/aclinux.h
@@ -148,13 +148,17 @@ static inline void *acpi_os_acquire_obje
 #define ACPI_ALLOCATE_ZEROED(a) acpi_os_allocate_zeroed(a)
 #define ACPI_FREE(a)            kfree(a)
 
-/* Used within ACPICA to show where it is safe to preempt execution */
-#include <linux/hardirq.h>
+#ifndef CONFIG_PREEMPT
+/*
+ * Used within ACPICA to show where it is safe to preempt execution
+ * when CONFIG_PREEMPT=n
+ */
 #define ACPI_PREEMPTION_POINT() \
 	do { \
-		if (!in_atomic_preempt_off() && !irqs_disabled()) \
+		if (!irqs_disabled()) \
 			cond_resched(); \
 	} while (0)
+#endif
 
 #endif /* __KERNEL__ */
 

_