Re: [PATCH 4.4 05/48] mm: fix devm_memremap_pages crash, use mem_hotplug_{begin, done}

[Dan Williams]

On Thu, Feb 9, 2017 at 7:26 AM, Ben Hutchings <ben@xxxxxxxxxxxxxxx> wrote:
> On Wed, 2017-01-18 at 11:46 +0100, Greg Kroah-Hartman wrote:
>> 4.4-stable review patch.  If anyone has any objections, please let me know.
>>
>> ------------------
>>
>> From: Dan Williams <dan.j.williams@xxxxxxxxx>
>>
>> commit f931ab479dd24cf7a2c6e2df19778406892591fb upstream.
>>
>> Both arch_add_memory() and arch_remove_memory() expect a single threaded
>> context.
> [...]
>> The result is that two threads calling devm_memremap_pages()
>> simultaneously can end up colliding on pgd initialization.  This leads
>> to crash signatures like the following where the loser of the race
>> initializes the wrong pgd entry:
> [...]
>> Hold the standard memory hotplug mutex over calls to
>> arch_{add,remove}_memory().
> [...]
>
> This is not a sufficient fix, because memory_hotplug.c still assumes
> there's only one 'writer':
>
> void put_online_mems(void)
> {
>         ...
>         if (!--mem_hotplug.refcount && unlikely(mem_hotplug.active_writer))
>                 wake_up_process(mem_hotplug.active_writer);
>         ...
> }
>
> void mem_hotplug_begin(void)
> {
>         mem_hotplug.active_writer = current;
>
>         memhp_lock_acquire();
>         for (;;) {
>                 mutex_lock(&mem_hotplug.lock);
>                 if (likely(!mem_hotplug.refcount))
>                         break;
>                 __set_current_state(TASK_UNINTERRUPTIBLE);
>                 mutex_unlock(&mem_hotplug.lock);
>                 schedule();
>         }
> }
>
> With multiple writers, one or more of them may hang or
> {get,put}_online_mems() may mess up the hotplug reference count.

You're right. We need to hold lock_device_hotplug_sysfs() before
calling mem_hotplug_begin().  I'll take a look at a follow-on fix and
also add an assert_held_device_hotplug() helper to catch this in the
future.