Hi Boris,
On 21/01/2019 17:58, Borislav Petkov wrote:
> On Mon, Dec 03, 2018 at 06:06:10PM +0000, James Morse wrote:
>> memory_failure() offlines or repairs pages of memory that have been
>> discovered to be corrupt. These may be detected by an external
>> component, (e.g. the memory controller), and notified via an IRQ.
>> In this case the work is queued as not all of memory_failure()s work
>> can happen in IRQ context.
>>
>> If the error was detected as a result of user-space accessing a
>> corrupt memory location the CPU may take an abort instead. On arm64
>> this is a 'synchronous external abort', and on a firmware first
>> system it is replayed using NOTIFY_SEA.
>>
>> This notification has NMI like properties, (it can interrupt
>> IRQ-masked code), so the memory_failure() work is queued. If we
>> return to user-space before the queued memory_failure() work is
>> processed, we will take the fault again. This loop may cause platform
>> firmware to exceed some threshold and reboot when Linux could have
>> recovered from this error.
>>
>> If a ghes notification type indicates that it may be triggered again
>> when we return to user-space, use the task-work and notify-resume
>> hooks to kick the relevant memory_failure() queue before returning
>> to user-space.
>> ---
>> I assume that if NOTIFY_NMI is coming from SMM it must suffer from
>> this problem too.
>
> Good question.
>
> I'm guessing all those things should be queued on a normal struct
> work_struct queue, no?
ghes_notify_nmi() does this today with its:
| irq_work_queue(&ghes_proc_irq_work);
Once its in IRQ context, the irq_work pokes memory_failure_queue(), which
schedule_work_on()s.
Finally we schedule() in process context, and can unmap the affected memory.
The problem is between each of these steps we might return to user-space and run
the instruction that tripped all this to begin with.
My SMM comment was because the CPU must jump from user-space->SMM, which injects
an NMI into the kernel. The kernel's EIP must point into user-space, so
returning from the NMI without doing the memory_failure() work puts us back the
same position we started in.
> Now, memory_failure_queue() does that and can run from IRQ context so
> you need only an irq_work which can queue from NMI context. We do it
> this way in the MCA code:
>
(was there something missing here?)
> We queue in an irq_work in NMI context and work through the items in
> process context.
How are you getting from NMI to process context in one go?
This patch causes the IRQ->process transition.
The arch specific bit of this gives the irq work queue a kick if returning from
the NMI would unmask IRQs. This makes it look like we moved from NMI to IRQ
context without returning to user-space.
Once ghes_handle_memory_failure() runs in IRQ context, it task_work_add()s the
call to ghes_kick_memory_failure().
Finally on the way out of the kernel to user-space that task_work runs and the
memory_failure() work happens in process context.
During all this the user-space program counter can point at a poisoned location,
but we don't return there until the memory_failure() work has been done.
>> @@ -407,7 +447,22 @@ static void ghes_handle_memory_failure(struct acpi_hest_generic_data *gdata, int
>>
>> if (flags != -1)
>> memory_failure_queue(pfn, flags);
>> -#endif
>> +
>> + /*
>> + * If the notification indicates that it was the interrupted
>> + * instruction that caused the error, try to kick the
>> + * memory_failure() queue before returning to user-space.
>> + */
>> + if (ghes_is_synchronous(ghes) && current->mm != &init_mm) {
>> + callback = kzalloc(sizeof(*callback), GFP_ATOMIC);
>
> Can we avoid that GFP_ATOMIC allocation and kfree() in
> ghes_kick_memory_failure()?
>
> I mean, that struct ghes_memory_failure_work is small enough and we
> already do lockless allocation:
>
> estatus_node = (void *)gen_pool_alloc(ghes_estatus_pool, node_len);
>
> so I guess we could add that ghes_memory_failure_work struct to that
> estatus_node, hand it into ghes_do_proc() and then free it.
I forget estatus_node is a linux thing, not an ACPI-spec thing!
Hmmm, ghes_handle_memory_failure() runs for POLLED and irq error sources too,
they don't have an estatus_node. We don't care about this ret_to_user() problem
as they are all asynchronous, this is why we have ghes_is_synchronous()...
It feels like there should be a way to do this, let me have a go...
Thanks,
James
