clocksource/arm_arch_timer: fix a lockdep warning
PATCH v2 1/1 clocksource/arm_arch_timer: fix a lockdep warning
Booting this Huawei TaiShan 2280 arm64 server generated this lockdep warning.
[ 0.000000] lockdep_assert_cpus_held+0x50/0x60 [ 0.000000] static_key_enable_cpuslocked+0x30/0xe8 [ 0.000000] arch_timer_check_ool_workaround+0x128/0x2d0 [ 0.000000] arch_timer_acpi_init+0x274/0x6ac [ 0.000000] acpi_table_parse+0x1ac/0x218 [ 0.000000] __acpi_probe_device_table+0x164/0x1ec [ 0.000000] timer_probe+0x1bc/0x254 [ 0.000000] time_init+0x44/0x98 [ 0.000000] start_kernel+0x4ec/0x7d4
This is due to the commit cb538267ea1e ("jump_label/lockdep: Assert we hold the hotplug lock for _cpuslocked() operations").
Since it is applying a global workaround to all CPUs here, it did not hold any CPU locks in this path.
arch_timer_acpi_init
arch_timer_check_ool_workaround(ate_match_acpi_oem_info, table)
arch_timer_enable_workaround(wa, local = false)
for_each_possible_cpu()
per_cpu()
When hot-add a CPU, it will go with a slightly different route.
arch_timer_starting_cpu
__arch_timer_setup
arch_timer_check_ool_workaround(ate_match_local_cap_id, NULL)
arch_timer_enable_workaround(wa, local = true)
__this_cpu_write()
Hence, deal with them differently.
static_branch_enable_cpuslocked())
---
Changes since v1:
* Fixed the root cause instead of a workaround.
drivers/clocksource/arm_arch_timer.c | 15 +++++++++------ 1 file changed, 9 insertions(+), 6 deletions(-)
drivers/clocksource/arm_arch_timer.c +9 -6
@@ -492,18 +492,21 @@ void arch_timer_enable_workaround(const struct arch_timer_erratum_workaround *waif (local) {__this_cpu_write (timer_unstable_counter_workaround, wa); + +/* +* Use the locked version, as we're called from the CPU +* hotplug framework. Otherwise, we end-up in +* deadlock-land. +*/ +static_branch_enable_cpuslocked (&arch_timer_read_ool_enabled); }else {for_each_possible_cpu (i)per_cpu (timer_unstable_counter_workaround, i) = wa; +/* A global workaround is not on the CPU hotplug path. */ +static_branch_enable (&arch_timer_read_ool_enabled); }/* -* Use the locked version, as we're called from the CPU -* hotplug framework. Otherwise, we end-up in deadlock-land. -*/ -static_branch_enable_cpuslocked (&arch_timer_read_ool_enabled); - -/* * Don't use the vdso fastpath if errata require using the * out-of-line counter accessor. We may change our mind pretty * late in the game (with a per-CPU erratum, for example), so
Signature · 7 lines
1.8.3.1 _______________________________________________ linux-arm-kernel mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
PATCH RESEND v2 1/1 clocksource/arm_arch_timer: fix a lockdep warning
Booting this Huawei TaiShan 2280 arm64 server generated this lockdep warning.
[ 0.000000] lockdep_assert_cpus_held+0x50/0x60 [ 0.000000] static_key_enable_cpuslocked+0x30/0xe8 [ 0.000000] arch_timer_check_ool_workaround+0x128/0x2d0 [ 0.000000] arch_timer_acpi_init+0x274/0x6ac [ 0.000000] acpi_table_parse+0x1ac/0x218 [ 0.000000] __acpi_probe_device_table+0x164/0x1ec [ 0.000000] timer_probe+0x1bc/0x254 [ 0.000000] time_init+0x44/0x98 [ 0.000000] start_kernel+0x4ec/0x7d4
This is due to the commit cb538267ea1e ("jump_label/lockdep: Assert we hold the hotplug lock for _cpuslocked() operations").
Since it is applying a global workaround to all CPUs here, it did not hold any CPU locks in this path.
arch_timer_acpi_init
arch_timer_check_ool_workaround(ate_match_acpi_oem_info, table)
arch_timer_enable_workaround(wa, local = false)
for_each_possible_cpu()
per_cpu()
There is also another path did not have any CPU lock.
time_init
clocksource_probe
arch_timer_of_init
arch_timer_check_ool_workaround(ate_match_dt, np)
arch_timer_enable_workaround(wa, local = false)
When hot-adding a CPU, it will go with a slightly different route.
arch_timer_starting_cpu
__arch_timer_setup
arch_timer_check_ool_workaround(ate_match_local_cap_id, NULL)
arch_timer_enable_workaround(wa, local = true)
__this_cpu_write()
Hence, deal with them differently.
static_branch_enable_cpuslocked())
---
v2: fix the root cause instead of a workaround.
drivers/clocksource/arm_arch_timer.c | 15 +++++++++------ 1 file changed, 9 insertions(+), 6 deletions(-)
drivers/clocksource/arm_arch_timer.c +9 -6
@@ -492,17 +492,20 @@ void arch_timer_enable_workaround(const struct arch_timer_erratum_workaround *waif (local) {__this_cpu_write (timer_unstable_counter_workaround, wa); + +/* +* Use the locked version, as we're called from the CPU +* hotplug framework. Otherwise, we end-up in +* deadlock-land. +*/ +static_branch_enable_cpuslocked (&arch_timer_read_ool_enabled); }else {for_each_possible_cpu (i)per_cpu (timer_unstable_counter_workaround, i) = wa; +/* A global workaround is not on the CPU hotplug path. */ +static_branch_enable (&arch_timer_read_ool_enabled); } -/* -* Use the locked version, as we're called from the CPU -* hotplug framework. Otherwise, we end-up in deadlock-land. -*/ -static_branch_enable_cpuslocked (&arch_timer_read_ool_enabled); -/* * Don't use the vdso fastpath if errata require using the * out-of-line counter accessor. We may change our mind pretty
Signature · 1 line
2.17.2 (Apple Git-113)
Peter Zijlstra
maintainerQuoted 12 lines · Qian Cai
Booting this Huawei TaiShan 2280 arm64 server generated this lockdep warning.
[ 0.000000] lockdep_assert_cpus_held+0x50/0x60 [ 0.000000] static_key_enable_cpuslocked+0x30/0xe8 [ 0.000000] arch_timer_check_ool_workaround+0x128/0x2d0 [ 0.000000] arch_timer_acpi_init+0x274/0x6ac [ 0.000000] acpi_table_parse+0x1ac/0x218 [ 0.000000] __acpi_probe_device_table+0x164/0x1ec [ 0.000000] timer_probe+0x1bc/0x254 [ 0.000000] time_init+0x44/0x98 [ 0.000000] start_kernel+0x4ec/0x7d4
It seems to me we want something like:
--- kernel/cpu.c | 2 ++ 1 file changed, 2 insertions(+)
kernel/cpu.c +2 -0
@@ -313,6 +313,8 @@ void cpus_write_unlock(void)void lockdep_assert_cpus_held (void ) { +if (system_state <SYSTEM_RUNNING ) +return ;percpu_rwsem_assert_held (&cpu_hotplug_lock); }
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Hi,
On 10/12/2018 14:07, Peter Zijlstra wrote:
[...]
Quoted 9 lines
--- kernel/cpu.c | 2 ++ 1 file changed, 2 insertions(+)kernel/cpu.c +0
-0
kernel/cpu.c void cpus_write_unlock(void) · line 320
Quoted 8 lines
kernel/cpu.c +2
-0@@ -313,6 +313,8 @@ void cpus_write_unlock(void)void lockdep_assert_cpus_held (void ) { +if (system_state <SYSTEM_RUNNING ) +return ;percpu_rwsem_assert_held (&cpu_hotplug_lock); }
Kinda hijacking the thread here - is that something we'd want to replace
40fa3780bac2 ("sched/core: Take the hotplug lock in sched_init_smp()")
with?
Peter Zijlstra
maintainerQuoted 24 lines · Valentin Schneider
Hi,
On 10/12/2018 14:07, Peter Zijlstra wrote:
[...]Quoted 9 lines
--- kernel/cpu.c | 2 ++ 1 file changed, 2 insertions(+)kernel/cpu.c +0
-0kernel/cpu.c void cpus_write_unlock(void) · line 320
Quoted 8 lines
kernel/cpu.c +2
-0@@ -313,6 +313,8 @@ void cpus_write_unlock(void)void lockdep_assert_cpus_held (void ) { +if (system_state <SYSTEM_RUNNING ) +return ;percpu_rwsem_assert_held (&cpu_hotplug_lock); }Kinda hijacking the thread here - is that something we'd want to replace
Quoted 4 lines
40fa3780bac2("sched/core: Take the hotplug lock insched_init_smp()")with?
Possibly; and I have vague memories of other people running into this same thing too.
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Qian Cai
Tested-byQuoted 14 lines · Peter Zijlstra
Quoted 12 lines · Qian Cai
Booting this Huawei TaiShan 2280 arm64 server generated this lockdep warning.
[ 0.000000] lockdep_assert_cpus_held+0x50/0x60 [ 0.000000] static_key_enable_cpuslocked+0x30/0xe8 [ 0.000000] arch_timer_check_ool_workaround+0x128/0x2d0 [ 0.000000] arch_timer_acpi_init+0x274/0x6ac [ 0.000000] acpi_table_parse+0x1ac/0x218 [ 0.000000] __acpi_probe_device_table+0x164/0x1ec [ 0.000000] timer_probe+0x1bc/0x254 [ 0.000000] time_init+0x44/0x98 [ 0.000000] start_kernel+0x4ec/0x7d4
Quoted 1 line
It seems to me we want something like:
Quoted 19 lines
--- kernel/cpu.c | 2 ++ 1 file changed, 2 insertions(+)kernel/cpu.c +2
-0@@ -313,6 +313,8 @@ void cpus_write_unlock(void)void lockdep_assert_cpus_held (void ) { +if (system_state <SYSTEM_RUNNING ) +return ;percpu_rwsem_assert_held (&cpu_hotplug_lock); }
Quoted 25 lines · Peter Zijlstra
Quoted 12 lines · Qian Cai
Booting this Huawei TaiShan 2280 arm64 server generated this lockdep warning.
[ 0.000000] lockdep_assert_cpus_held+0x50/0x60 [ 0.000000] static_key_enable_cpuslocked+0x30/0xe8 [ 0.000000] arch_timer_check_ool_workaround+0x128/0x2d0 [ 0.000000] arch_timer_acpi_init+0x274/0x6ac [ 0.000000] acpi_table_parse+0x1ac/0x218 [ 0.000000] __acpi_probe_device_table+0x164/0x1ec [ 0.000000] timer_probe+0x1bc/0x254 [ 0.000000] time_init+0x44/0x98 [ 0.000000] start_kernel+0x4ec/0x7d4It seems to me we want something like:
--- kernel/cpu.c | 2 ++ 1 file changed, 2 insertions(+)kernel/cpu.c +0
-0
kernel/cpu.c void cpus_write_unlock(void) · line 319
Quoted 7 lines
kernel/cpu.c +2
-0@@ -313,6 +313,8 @@ void cpus_write_unlock(void)void lockdep_assert_cpus_held (void ) { +if (system_state <SYSTEM_RUNNING ) +return ;percpu_rwsem_assert_held (&cpu_hotplug_lock); }
Hmm, no. SYSTEM_SCHEDULING is what you want because RUNNING is set way too late.
Thanks,
tglx
Quoted 27 lines · Thomas Gleixner
Quoted 26 lines · Peter Zijlstra
Quoted 12 lines · Qian Cai
Booting this Huawei TaiShan 2280 arm64 server generated this lockdep warning.
[ 0.000000] lockdep_assert_cpus_held+0x50/0x60 [ 0.000000] static_key_enable_cpuslocked+0x30/0xe8 [ 0.000000] arch_timer_check_ool_workaround+0x128/0x2d0 [ 0.000000] arch_timer_acpi_init+0x274/0x6ac [ 0.000000] acpi_table_parse+0x1ac/0x218 [ 0.000000] __acpi_probe_device_table+0x164/0x1ec [ 0.000000] timer_probe+0x1bc/0x254 [ 0.000000] time_init+0x44/0x98 [ 0.000000] start_kernel+0x4ec/0x7d4It seems to me we want something like:
--- kernel/cpu.c | 2 ++ 1 file changed, 2 insertions(+)kernel/cpu.c +0
-0@@ -313,6 +313,8 @@ void cpus_write_unlock(void)
Quoted 9 lines
kernel/cpu.c void cpus_write_unlock(void) · line 319
Quoted 6 lines
kernel/cpu.c +2
-0void lockdep_assert_cpus_held (void ) { +if (system_state <SYSTEM_RUNNING ) +return ;percpu_rwsem_assert_held (&cpu_hotplug_lock); }Hmm, no.
SYSTEM_SCHEDULINGis what you want because RUNNING is set way too late.
SYSTEM_SCHEDULING works well here too.
Quoted 58 lines · Qian Cai
Booting this Huawei TaiShan 2280 arm64 server generated this lockdep warning.
[ 0.000000] lockdep_assert_cpus_held+0x50/0x60 [ 0.000000] static_key_enable_cpuslocked+0x30/0xe8 [ 0.000000] arch_timer_check_ool_workaround+0x128/0x2d0 [ 0.000000] arch_timer_acpi_init+0x274/0x6ac [ 0.000000] acpi_table_parse+0x1ac/0x218 [ 0.000000] __acpi_probe_device_table+0x164/0x1ec [ 0.000000] timer_probe+0x1bc/0x254 [ 0.000000] time_init+0x44/0x98 [ 0.000000] start_kernel+0x4ec/0x7d4This is due to the commit
cb538267ea1e("jump_label/lockdep: Assert we hold the hotplug lock for_cpuslocked()operations").Since it is applying a global workaround to all CPUs here, it did not hold any CPU locks in this path.
arch_timer_acpi_init
arch_timer_check_ool_workaround(ate_match_acpi_oem_info, table)
arch_timer_enable_workaround(wa, local = false)
for_each_possible_cpu()
per_cpu()There is also another path did not have any CPU lock.
time_init
clocksource_probe
arch_timer_of_init
arch_timer_check_ool_workaround(ate_match_dt, np)
arch_timer_enable_workaround(wa, local = false)When hot-adding a CPU, it will go with a slightly different route.
arch_timer_starting_cpu
__arch_timer_setup
arch_timer_check_ool_workaround(ate_match_local_cap_id, NULL)
arch_timer_enable_workaround(wa, local = true)
__this_cpu_write()Hence, deal with them differently.
Fixes 450f9689f294 (clocksource/arm_arch_timer: Use
static_branch_enable_cpuslocked())Signed-off-by Qian Cai <[email protected]>---
v2: fix the root cause instead of a workaround.
drivers/clocksource/arm_arch_timer.c | 15 +++++++++------ 1 file changed, 9 insertions(+), 6 deletions(-)drivers/clocksource/arm_arch_timer.c +0
-0
drivers/clocksource/arm_arch_timer.c void arch_timer_enable_workaround(const struct arch_timer_erratum_workaround *wa · line 501
Quoted 10 lines
drivers/clocksource/arm_arch_timer.c +7
-0@@ -492,17 +492,20 @@ void arch_timer_enable_workaround(const struct arch_timer_erratum_workaround *waif (local) {__this_cpu_write (timer_unstable_counter_workaround, wa); + +/* +* Use the locked version, as we're called from the CPU +* hotplug framework. Otherwise, we end-up in +* deadlock-land. +*/ +static_branch_enable_cpuslocked (&arch_timer_read_ool_enabled);
I have the ugly feeling that it breaks the (equally ugly) big-little stuff where the boot CPU is affected. In this context, you'd end-up without the lock being taken and you'll get the same splat.
We could start testing the CPU number, but Peter's approach seem much more palatable to me.
Thanks,
M.
Signature · 1 line
Jazz is not dead. It just smells funny...