Re: [PATCH v2] arm64: paravirt: Use RCU read locks to guard stolen_time

From: Juergen Gross <jgross@suse.com>
To: Will Deacon <will@kernel.org>, Elliot Berman <quic_eberman@quicinc.com>
Cc: "Srivatsa S. Bhat (VMware)" <srivatsa@csail.mit.edu>,
	Alexey Makhalov <amakhalov@vmware.com>,
	Catalin Marinas <catalin.marinas@arm.com>,
	Prakruthi Deepak Heragu <quic_pheragu@quicinc.com>,
	virtualization@lists.linux-foundation.org, x86@kernel.org,
	linux-arm-kernel@lists.infradead.org,
	linux-kernel@vger.kernel.org,
	Murali Nalajala <quic_mnalajal@quicinc.com>,
	stable@vger.kernel.org
Subject: Re: [PATCH v2] arm64: paravirt: Use RCU read locks to guard stolen_time
Date: Wed, 4 May 2022 15:38:47 +0200	[thread overview]
Message-ID: <c6689e42-e87c-0c0b-c7ff-40134406e080@suse.com> (raw)
In-Reply-To: <20220504094507.GA20305@willie-the-truck>

[-- Attachment #1.1.1: Type: text/plain, Size: 4012 bytes --]

On 04.05.22 11:45, Will Deacon wrote:
> On Thu, Apr 28, 2022 at 11:35:36AM -0700, Elliot Berman wrote:
>> From: Prakruthi Deepak Heragu <quic_pheragu@quicinc.com>
>>
>> During hotplug, the stolen time data structure is unmapped and memset.
>> There is a possibility of the timer IRQ being triggered before memset
>> and stolen time is getting updated as part of this timer IRQ handler. This
>> causes the below crash in timer handler -
>>
>>    [ 3457.473139][    C5] Unable to handle kernel paging request at virtual address ffffffc03df05148
>>    ...
>>    [ 3458.154398][    C5] Call trace:
>>    [ 3458.157648][    C5]  para_steal_clock+0x30/0x50
>>    [ 3458.162319][    C5]  irqtime_account_process_tick+0x30/0x194
>>    [ 3458.168148][    C5]  account_process_tick+0x3c/0x280
>>    [ 3458.173274][    C5]  update_process_times+0x5c/0xf4
>>    [ 3458.178311][    C5]  tick_sched_timer+0x180/0x384
>>    [ 3458.183164][    C5]  __run_hrtimer+0x160/0x57c
>>    [ 3458.187744][    C5]  hrtimer_interrupt+0x258/0x684
>>    [ 3458.192698][    C5]  arch_timer_handler_virt+0x5c/0xa0
>>    [ 3458.198002][    C5]  handle_percpu_devid_irq+0xdc/0x414
>>    [ 3458.203385][    C5]  handle_domain_irq+0xa8/0x168
>>    [ 3458.208241][    C5]  gic_handle_irq.34493+0x54/0x244
>>    [ 3458.213359][    C5]  call_on_irq_stack+0x40/0x70
>>    [ 3458.218125][    C5]  do_interrupt_handler+0x60/0x9c
>>    [ 3458.223156][    C5]  el1_interrupt+0x34/0x64
>>    [ 3458.227560][    C5]  el1h_64_irq_handler+0x1c/0x2c
>>    [ 3458.232503][    C5]  el1h_64_irq+0x7c/0x80
>>    [ 3458.236736][    C5]  free_vmap_area_noflush+0x108/0x39c
>>    [ 3458.242126][    C5]  remove_vm_area+0xbc/0x118
>>    [ 3458.246714][    C5]  vm_remove_mappings+0x48/0x2a4
>>    [ 3458.251656][    C5]  __vunmap+0x154/0x278
>>    [ 3458.255796][    C5]  stolen_time_cpu_down_prepare+0xc0/0xd8
>>    [ 3458.261542][    C5]  cpuhp_invoke_callback+0x248/0xc34
>>    [ 3458.266842][    C5]  cpuhp_thread_fun+0x1c4/0x248
>>    [ 3458.271696][    C5]  smpboot_thread_fn+0x1b0/0x400
>>    [ 3458.276638][    C5]  kthread+0x17c/0x1e0
>>    [ 3458.280691][    C5]  ret_from_fork+0x10/0x20
>>
>> As a fix, introduce rcu lock to update stolen time structure.
>>
>> Fixes: 75df529bec91 ("arm64: paravirt: Initialize steal time when cpu is online")
>> Cc: stable@vger.kernel.org
>> Signed-off-by: Prakruthi Deepak Heragu <quic_pheragu@quicinc.com>
>> Signed-off-by: Elliot Berman <quic_eberman@quicinc.com>
>> ---
>> Changes since v1: https://lore.kernel.org/all/20220420204417.155194-1-quic_eberman@quicinc.com/
>>   - Use RCU instead of disabling interrupts
>>
>>   arch/arm64/kernel/paravirt.c | 24 +++++++++++++++++++-----
>>   1 file changed, 19 insertions(+), 5 deletions(-)
>>
>> diff --git a/arch/arm64/kernel/paravirt.c b/arch/arm64/kernel/paravirt.c
>> index 75fed4460407..e724ea3d86f0 100644
>> --- a/arch/arm64/kernel/paravirt.c
>> +++ b/arch/arm64/kernel/paravirt.c
>> @@ -52,7 +52,9 @@ early_param("no-steal-acc", parse_no_stealacc);
>>   /* return stolen time in ns by asking the hypervisor */
>>   static u64 para_steal_clock(int cpu)
>>   {
>> +	struct pvclock_vcpu_stolen_time *kaddr = NULL;
>>   	struct pv_time_stolen_time_region *reg;
>> +	u64 ret = 0;
>>   
>>   	reg = per_cpu_ptr(&stolen_time_region, cpu);
>>   
>> @@ -61,28 +63,38 @@ static u64 para_steal_clock(int cpu)
>>   	 * online notification callback runs. Until the callback
>>   	 * has run we just return zero.
>>   	 */
>> -	if (!reg->kaddr)
>> +	rcu_read_lock();
>> +	kaddr = rcu_dereference(reg->kaddr);
>> +	if (!kaddr) {
>> +		rcu_read_unlock();
>>   		return 0;
>> +	}
>>   
>> -	return le64_to_cpu(READ_ONCE(reg->kaddr->stolen_time));
>> +	ret = le64_to_cpu(READ_ONCE(kaddr->stolen_time));
> 
> Is this READ_ONCE() still required now?

Yes, as it might be called for another cpu than the current one.
stolen_time might just be updated, so you want to avoid load tearing.

Juergen

[-- Attachment #1.1.2: OpenPGP public key --]
[-- Type: application/pgp-keys, Size: 3149 bytes --]

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 495 bytes --]