On 07.02.2018 21:38, Dave Hansen wrote:
> On 02/07/2018 08:14 AM, Kirill Tkhai wrote:
>> Sometimes it is possible to meet a situation,
>> when irq stack is corrupted, while innocent
>> callback function is being executed. This may
>> happen because of crappy drivers irq handlers,
>> when they access wrong memory on the irq stack.
> 
> Can you be more clear about the actual issue?  Which drivers do this?
> How do they even find an IRQ stack pointer?

I can't say actual driver making this, because I'm still investigating the 
guilty one.
But I have couple of crash dumps with the crash inside update_sd_lb_stats() 
function,
where stack variable sg becomes corrupted. This time all scheduler-related 
not-stack
variables are in ideal state. And update_sd_lb_stats() is the function, which 
can't
corrupt its own stack. So, I thought this functionality may be useful for 
something else,
especially because of irq stack is one of the last stacks, which are not 
sanitized.
Task's stacks are already covered, as I know

[1595450.678971] Call Trace:
[1595450.683991]  <IRQ>
[1595450.684038]
[1595450.688926]  [<ffffffff81320005>] cpumask_next_and+0x35/0x50
[1595450.693984]  [<ffffffff810d91d3>] find_busiest_group+0x143/0x950
[1595450.699088]  [<ffffffff810d9b7a>] load_balance+0x19a/0xc20
[1595450.704289]  [<ffffffff810cde55>] ? sched_clock_cpu+0x85/0xc0
[1595450.709457]  [<ffffffff810c29aa>] ? update_rq_clock.part.88+0x1a/0x150
[1595450.714711]  [<ffffffff810da770>] rebalance_domains+0x170/0x2b0
[1595450.719997]  [<ffffffff810da9d2>] run_rebalance_domains+0x122/0x1e0
[1595450.725321]  [<ffffffff816bb10f>] __do_softirq+0x10f/0x2aa
[1595450.730746]  [<ffffffff816b62ac>] call_softirq+0x1c/0x30
[1595450.736169]  [<ffffffff8102d325>] do_softirq+0x65/0xa0
[1595450.741754]  [<ffffffff81093ec5>] irq_exit+0x105/0x110
[1595450.747279]  [<ffffffff816baad2>] smp_apic_timer_interrupt+0x42/0x50
[1595450.752905]  [<ffffffff816b7a62>] apic_timer_interrupt+0x232/0x240
[1595450.758519]  <EOI>
[1595450.758569]
[1595450.764100]  [<ffffffff8152f282>] ? cpuidle_enter_state+0x52/0xc0
[1595450.769652]  [<ffffffff8152f3c8>] cpuidle_idle_call+0xd8/0x210
[1595450.775198]  [<ffffffff8103540e>] arch_cpu_idle+0xe/0x30
[1595450.780813]  [<ffffffff810effba>] cpu_startup_entry+0x14a/0x1c0
[1595450.786286]  [<ffffffff810523e6>] start_secondary+0x1d6/0x250

>> This patch aims to catch such the situations
>> and adds checks of unauthorized stack access.
> 
> I think I forgot how KASAN did this.  KASAN has metadata that says which
> areas of memory are good or bad to access, right?  So, this just tags
> IRQ stacks as bad when we are not _in_ an interrupt?
> 
>> +#define KASAN_IRQ_STACK_SIZE \
>> +    (sizeof(union irq_stack_union) - \
>> +            (offsetof(union irq_stack_union, stack_canary) + 8))
> 
> Just curious, but why leave out the canary?  It shouldn't be accessed
> either.

It's touched in several more places (e.g., in __switch_to_asm()), and I'm not
sure KASAN is OK with this. Does it?

Also gs_base is touched from load_percpu_segment(), which could be called from
different cpu, and this seems it would required some synchronization between
the handlers and this primitive.

>> +#ifdef CONFIG_KASAN
>> +void __visible x86_poison_irq_stack(void)
>> +{
>> +    if (this_cpu_read(irq_count) == -1)
>> +            kasan_poison_irq_stack();
>> +}
>> +void __visible x86_unpoison_irq_stack(void)
>> +{
>> +    if (this_cpu_read(irq_count) == -1)
>> +            kasan_unpoison_irq_stack();
>> +}
>> +#endif
> 
> It might be handy to point out here that -1 means "not in an interrupt"
> and >=0 means "in an interrupt".
> 
> Otherwise, this looks pretty straightforward.  Would it be something to
> extend to the other stacks like the NMI or double-fault stacks?  Or are
> those just not worth it
I haven't met NMI stack corrupted, so I don't have ideas about this. If
we need to check them too, one more patch should be introduced on top of
this.

Kirill

Reply via email to