Re: [PATCH 08/12] x86/mm: Remove pgd_list use from vmalloc_sync_all()

From: Andy Lutomirski
Date: Thu Jun 11 2015 - 12:10:47 EST


On Thu, Jun 11, 2015 at 8:49 AM, Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>
> * Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
>
>> On Thu, Jun 11, 2015 at 7:07 AM, Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>> > The vmalloc() code uses vmalloc_sync_all() to synchronize changes to
>> > the global reference kernel PGD to task PGDs.
>>
>> Does it? AFAICS the only caller is register_die_notifier, and it's
>> not really clear to me why that exists.
>
> Doh, indeed, got confused in that changelog - we are filling it in
> opportunistically via vmalloc_fault().
>
>> At some point I'd love to remove lazy kernel PGD sync from the kernel entirely
>> (or at least from x86) and just do it when we switch mms. Now that you're
>> removing all code that deletes kernel PGD entries, I think all we'd need to do
>> is to add a per-PGD or per-mm count of the number of kernel entries populated
>> and to fix it up when we switch to an mm with fewer entries populated than
>> init_mm.
>
> That would add a (cheap but nonzero) runtime check to every context switch. It's a
> relative slow path, but in comparison vmalloc() is an even slower slowpath, so why
> not do it there and just do synchronous updates and remove the vmalloc faults
> altogether?
>
> Also, on 64-bit it should not matter much: there the only change is the once in a
> blue moon case where we allocate a new pgd for a 512 GB block of address space
> that a single pgd entry covers.
>
> I'd hate to add a check to every context switch, no matter how cheap, just for a
> case that essentially never triggers...
>
> So how about this solution instead:
>
> - we add a generation counter to sync_global_pgds() so that it can detect when
> the number of pgds populated in init_mm changes.
>
> - we change vmalloc() to call sync_global_pgds(): this will be very cheap in the
> overwhelming majority of cases.
>
> - we eliminate vmalloc_fault(), on 64-bit at least. Yay! :-)

Should be good. The only real down side is that the update becomes a
walk over all tasks or over all mms, whereas if we checked on context
switches, then the update is just a single PGD entry copy.

That being said, the updates should be *really* rare. I think it's
completely impossible for it to happen more than 255 times per boot.
So I like your solution.

Then I can stop wondering what happens when we manage to take a
vmalloc fault early in entry_SYSCALL_64 or somewhere in the page_fault
prologue.

--Andy
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/