Re: [PATCH v2] x86/acpi: fix a deadlock with cpu hotplug

[Qian Cai <cai@lca.pw>]

> On Apr 3, 2020, at 5:29 AM, Rafael J. Wysocki <rjw@rjwysocki.net> wrote:
> 
> On Sunday, March 29, 2020 4:21:09 PM CEST Qian Cai wrote:
>> Similar to the commit 0266d81e9bf5 ("acpi/processor: Prevent cpu hotplug
>> deadlock") except this is for acpi_processor_ffh_cstate_probe():
>> 
>> "The problem is that the work is scheduled on the current CPU from the
>> hotplug thread associated with that CPU.
>> 
>> It's not required to invoke these functions via the workqueue because
>> the hotplug thread runs on the target CPU already.
>> 
>> Check whether current is a per cpu thread pinned on the target CPU and
>> invoke the function directly to avoid the workqueue."
>> 
>> Since CONFIG_ACPI_PROCESSOR (for cstate.c) selects
>> CONFIG_ACPI_CPU_FREQ_PSS (for processor_throttling.c) on x86, just
>> make call_on_cpu() a static inline function from processor_throttling.c
>> and use it in cstate.c.
>> 
>> WARNING: possible circular locking dependency detected
>> ------------------------------------------------------
>> cpuhp/1/15 is trying to acquire lock:
>> ffffc90003447a28 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: __flush_work+0x4c6/0x630
>> 
>> but task is already holding lock:
>> ffffffffafa1c0e8 (cpuidle_lock){+.+.}-{3:3}, at: cpuidle_pause_and_lock+0x17/0x20
>> 
>> which lock already depends on the new lock.
>> 
>> the existing dependency chain (in reverse order) is:
>> 
>> -> #1 (cpu_hotplug_lock){++++}-{0:0}:
>> cpus_read_lock+0x3e/0xc0
>> irq_calc_affinity_vectors+0x5f/0x91
>> __pci_enable_msix_range+0x10f/0x9a0
>> pci_alloc_irq_vectors_affinity+0x13e/0x1f0
>> pci_alloc_irq_vectors_affinity at drivers/pci/msi.c:1208
>> pqi_ctrl_init+0x72f/0x1618 [smartpqi]
>> pqi_pci_probe.cold.63+0x882/0x892 [smartpqi]
>> local_pci_probe+0x7a/0xc0
>> work_for_cpu_fn+0x2e/0x50
>> process_one_work+0x57e/0xb90
>> worker_thread+0x363/0x5b0
>> kthread+0x1f4/0x220
>> ret_from_fork+0x27/0x50
>> 
>> -> #0 ((work_completion)(&wfc.work)){+.+.}-{0:0}:
>> __lock_acquire+0x2244/0x32a0
>> lock_acquire+0x1a2/0x680
>> __flush_work+0x4e6/0x630
>> work_on_cpu+0x114/0x160
>> acpi_processor_ffh_cstate_probe+0x129/0x250
>> acpi_processor_evaluate_cst+0x4c8/0x580
>> acpi_processor_get_power_info+0x86/0x740
>> acpi_processor_hotplug+0xc3/0x140
>> acpi_soft_cpu_online+0x102/0x1d0
>> cpuhp_invoke_callback+0x197/0x1120
>> cpuhp_thread_fun+0x252/0x2f0
>> smpboot_thread_fn+0x255/0x440
>> kthread+0x1f4/0x220
>> ret_from_fork+0x27/0x50
>> 
>> other info that might help us debug this:
>> 
>> Chain exists of:
>> (work_completion)(&wfc.work) --> cpuhp_state-up --> cpuidle_lock
>> 
>> Possible unsafe locking scenario:
>> 
>> CPU0                    CPU1
>> ----                    ----
>> lock(cpuidle_lock);
>>                         lock(cpuhp_state-up);
>>                         lock(cpuidle_lock);
>> lock((work_completion)(&wfc.work));
>> 
>> *** DEADLOCK ***
>> 
>> 3 locks held by cpuhp/1/15:
>> #0: ffffffffaf51ab10 (cpu_hotplug_lock){++++}-{0:0}, at: cpuhp_thread_fun+0x69/0x2f0
>> #1: ffffffffaf51ad40 (cpuhp_state-up){+.+.}-{0:0}, at: cpuhp_thread_fun+0x69/0x2f0
>> #2: ffffffffafa1c0e8 (cpuidle_lock){+.+.}-{3:3}, at: cpuidle_pause_and_lock+0x17/0x20
>> 
>> Call Trace:
>> dump_stack+0xa0/0xea
>> print_circular_bug.cold.52+0x147/0x14c
>> check_noncircular+0x295/0x2d0
>> __lock_acquire+0x2244/0x32a0
>> lock_acquire+0x1a2/0x680
>> __flush_work+0x4e6/0x630
>> work_on_cpu+0x114/0x160
>> acpi_processor_ffh_cstate_probe+0x129/0x250
>> acpi_processor_evaluate_cst+0x4c8/0x580
>> acpi_processor_get_power_info+0x86/0x740
>> acpi_processor_hotplug+0xc3/0x140
>> acpi_soft_cpu_online+0x102/0x1d0
>> cpuhp_invoke_callback+0x197/0x1120
>> cpuhp_thread_fun+0x252/0x2f0
>> smpboot_thread_fn+0x255/0x440
>> kthread+0x1f4/0x220
>> ret_from_fork+0x27/0x50
>> 
>> Signed-off-by: Qian Cai <cai@lca.pw>
>> ---
>> 
>> v2:
>> Make call_on_cpu() a static inline function to avoid a compilation
>> error when ACPI_PROCESSOR=m thanks to lkp@intel.com.
>> 
>> arch/x86/kernel/acpi/cstate.c       |  3 ++-
>> drivers/acpi/processor_throttling.c |  7 -------
>> include/acpi/processor.h            | 10 ++++++++++
>> 3 files changed, 12 insertions(+), 8 deletions(-)
>> 
>> diff --git a/arch/x86/kernel/acpi/cstate.c b/arch/x86/kernel/acpi/cstate.c
>> index caf2edccbad2..49ae4e1ac9cd 100644
>> --- a/arch/x86/kernel/acpi/cstate.c
>> +++ b/arch/x86/kernel/acpi/cstate.c
>> @@ -161,7 +161,8 @@ int acpi_processor_ffh_cstate_probe(unsigned int cpu,
>> 
>> 	/* Make sure we are running on right CPU */
>> 
>> -	retval = work_on_cpu(cpu, acpi_processor_ffh_cstate_probe_cpu, cx);
>> +	retval = call_on_cpu(cpu, acpi_processor_ffh_cstate_probe_cpu, cx,
>> +			     false);
>> 	if (retval == 0) {
>> 		/* Use the hint in CST */
>> 		percpu_entry->states[cx->index].eax = cx->address;
>> diff --git a/drivers/acpi/processor_throttling.c b/drivers/acpi/processor_throttling.c
>> index 532a1ae3595a..a0bd56ece3ff 100644
>> --- a/drivers/acpi/processor_throttling.c
>> +++ b/drivers/acpi/processor_throttling.c
>> @@ -897,13 +897,6 @@ static long __acpi_processor_get_throttling(void *data)
>> 	return pr->throttling.acpi_processor_get_throttling(pr);
>> }
>> 
>> -static int call_on_cpu(int cpu, long (*fn)(void *), void *arg, bool direct)
>> -{
>> -	if (direct || (is_percpu_thread() && cpu == smp_processor_id()))
>> -		return fn(arg);
>> -	return work_on_cpu(cpu, fn, arg);
>> -}
>> -
>> static int acpi_processor_get_throttling(struct acpi_processor *pr)
>> {
>> 	if (!pr)
>> diff --git a/include/acpi/processor.h b/include/acpi/processor.h
>> index 47805172e73d..770d226b22f2 100644
>> --- a/include/acpi/processor.h
>> +++ b/include/acpi/processor.h
>> @@ -297,6 +297,16 @@ static inline void acpi_processor_ffh_cstate_enter(struct acpi_processor_cx
>> }
>> #endif
>> 
>> +#ifdef CONFIG_ACPI_CPU_FREQ_PSS
> 
> Why does this depend on CONFIG_ACPI_CPU_FREQ_PSS?

call_on_cpu() was only used in processor_throttling.c which has,

processor-$(CONFIG_ACPI_CPU_FREQ_PSS)   += processor_throttling.o

after this patch, it will also be used in cstate.c which has,

ifneq ($(CONFIG_ACPI_PROCESSOR),)
obj-y                           += cstate.o
endif

i.e.,

config ACPI_PROCESSOR
        tristate "Processor"
        depends on X86 || IA64 || ARM64
        select ACPI_PROCESSOR_IDLE
        select ACPI_CPU_FREQ_PSS if X86 || IA64

Therefore, call_on_cpu() is only used when CONFIG_ACPI_CPU_FREQ_PSS=y.

The #ifdef is rather a safe net that in the future, if we decided to make call_on_cpu()
a  non-inline function, it will prevent triggering an compilation warning for unused
function when CONFIG_ACPI_CPU_FREQ_PSS=n.

It may also serve as a documentation purpose to indicate that function is only used
with  CONFIG_ACPI_CPU_FREQ_PSS=y.

> 
>> +static inline int call_on_cpu(int cpu, long (*fn)(void *), void *arg,
>> +			      bool direct)
>> +{
>> +	if (direct || (is_percpu_thread() && cpu == smp_processor_id()))
>> +		return fn(arg);
>> +	return work_on_cpu(cpu, fn, arg);
>> +}
>> +#endif
>> +
>> /* in processor_perflib.c */
>> 
>> #ifdef CONFIG_CPU_FREQ