Re: [PATCH v5 1/3] arm64/ras: support sea error recovery

[James Morse]

Hi Xie XiuQi,

On 30/01/18 19:19, James Morse wrote:
> On 26/01/18 12:31, Xie XiuQi wrote:
>> With ARM v8.2 RAS Extension, SEA are usually triggered when memory errors
>> are consumed. According to the existing process, errors occurred in the
>> kernel, leading to direct panic, if it occurred the user-space, we should
>> just kill process.
>>
>> But there is a class of error, in fact, is not necessary to kill
>> process, you can recover and continue to run the process. Such as
>> the instruction data corrupted, where the memory page might be
>> read-only, which is has not been modified, the disk might have the
>> correct data, so you can directly drop the page, ant reload it when
>> necessary.
> 
> With firmware-first support, we do all this...
> 
> 
>> So this patchset is just try to solve such problem: if the error is
>> consumed in user-space and the error occurs on a clean page, you can
>> directly drop the memory page without killing process.
>>
>> If the corrupted page is clean, just dropped it and return to user-space
>> without side effects. And if corrupted page is dirty, memory_failure()
>> will send SIGBUS with code=BUS_MCEERR_AR. While without this patchset,
>> do_sea() will just send SIGBUS, so the process was killed in the same place.
> 
> ... but this happens too. I agree its something we should fix, but I don't think
> this is the best way to do it.
> 
> This series is pulling the memory-failure-queue details back into the arch-code
> to build a second list, that gets processed as extra work when we return to
> user-space.
> 
> 
> The root of the issue is ghes_notify_sea() claims the notification as something
> APEI has dealt with, ... but it hasn't done it yet. The signals will be
> generated by something currently stuck in a queue. (Evidently x86 doesn't handle
> synchronous errors like this using firmware-first).
> 
> I think a smaller fix is to give the queues that may be holding the
> memory_failure() work a kick as part of the code that calls ghes_notify_sea().
> This means that by the time we return to do_sea() ghes_notify_sea()'s claim that
> APEI has dealt with it is true as any generated signals are pending. We can then
> skip the existing SIGBUS generation code.
> 
> 
>> Because memory_failure() may sleep, we can not call it directly in SEA
> 
> (this one is more serious, I've attempted to fix it by moving all NMI-like
> GHES-notifications to use the estatus queue).
> 
> 
>> exception context. So we saved faulting physical address associated with
>> a process in the ghes handler and set __TIF_SEA_NOTIFY. When we return
>> from SEA exception context and get into do_notify_resume() before the
>> process running, we could check it and call memory_failure() to do
>> recovery.
> 
>> It's safe, because we are in process context.
> 
> I think this is the trick. When we take a Synchronous-external-abort out of
> userspace, we're in process context too. We can add helpers to drain the
> memory_failure_queue which can be called when do_sea() when we know we're
> preemptible and interrupts-et-al are unmasked.

Something like... base on [0], in arch/arm64/kernel/acpi.c:
-----------------%<-----------------
int apei_claim_sea(struct pt_regs *regs)
{
        int cpu;
        int err = -ENOENT;
        unsigned long current_flags = arch_local_save_flags();
        unsigned long interrupted_flags = current_flags;

        if (!IS_ENABLED(CONFIG_ACPI_APEI_SEA))
                return err;

        if (regs)
                interrupted_flags = regs->pstate;

        /*
         * APEI expects an NMI-like notification to always be called
         * in NMI context.
         */
        local_daif_restore(DAIF_ERRCTX);
        nmi_enter();
        err = ghes_notify_sea();
        cpu = smp_processor_id();
        nmi_exit();

        /*
         * APEI NMI-like notifications are deferred to irq_work. Unless
         * we interrupted irqs-masked code, we can do that now.
         */
        if (!err) {
                if (!arch_irqs_disabled_flags(interrupted_flags)) {
                        local_daif_restore(DAIF_PROCCTX_NOIRQ);
                        irq_work_run();
                } else {
                        err = -EINPROGRESS;
                }
        }

        local_daif_restore(current_flags);

        if (IS_ENABLED(CONFIG_ACPI_APEI_MEMORY_FAILURE) && !err) {
                /*
                 * Memory failure work is scheduled on the local CPU.
                 * If we interrupted userspace, or are in process context
                 * we can do that now.
                 */
                if ((regs && !user_mode(regs)) || !preemptible())
                        err = -EINPROGRESS;
                else
                        memory_failure_queue_kick(cpu);
        }

        return err;
}
-----------------%<-----------------


and to mm/memory-failure.c:
-----------------%<-----------------
@@ -1355,7 +1355,7 @@ static void memory_failure_work_func(struct work_struct *w
ork)
        unsigned long proc_flags;
        int gotten;

-       mf_cpu = this_cpu_ptr(&memory_failure_cpu);
+       mf_cpu = container_of(work, struct memory_failure_cpu, work);
        for (;;) {
                spin_lock_irqsave(&mf_cpu->lock, proc_flags);
                gotten = kfifo_get(&mf_cpu->fifo, &entry);

@@ -1369,6 +1369,22 @@ static void memory_failure_work_func(struct work_struct *
work)
        }
 }

+/*
+ * Process memory_failure work queued on the specified CPU.
+ * Used to avoid return-to-userspace racing with the memory_failure workqueue.
+ */
+void memory_failure_queue_kick(int cpu)
+{
+       unsigned long flags;
+       struct memory_failure_cpu *mf_cpu;
+
+       might_sleep();
+
+       mf_cpu = &per_cpu(memory_failure_cpu, cpu);
+       cancel_work_sync(&mf_cpu->work);
+       memory_failure_work_func(&mf_cpu->work);
+}
+
 static int __init memory_failure_init(void)
 {
        struct memory_failure_cpu *mf_cpu;
-----------------%<-----------------

I've cooked up some NOTFIY_SEA-ing APEI firmware using kvmtool to test this. I
haven't yet managed to hit irq-masked code with NOTIFY_SEA. I'll try and tidy
this up and post a branch to make it easier to test...

I prefer this as it doesn't duplicate the state then come back on a TIF flag.
I'd like to move the kicking logic into ghes.c, as that is where the queueing
happened, but the 'do-this, restore these flags, do-that' is somewhat tasteless,
and it looks like on arm64 has synchronous nmi-like notifications that must be
handled before returning to user-space...



Thanks,

James

[0] https://www.spinics.net/lists/linux-acpi/msg80149.html