[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <87d2a5f1m4.fsf@linux.vnet.ibm.com>
Date: Tue, 07 Oct 2014 00:48:43 +0530
From: "Aneesh Kumar K.V" <aneesh.kumar@...ux.vnet.ibm.com>
To: Mel Gorman <mgorman@...e.de>,
Linus Torvalds <torvalds@...ux-foundation.org>
Cc: Hugh Dickins <hughd@...gle.com>, Dave Jones <davej@...hat.com>,
Al Viro <viro@...iv.linux.org.uk>,
Linux Kernel <linux-kernel@...r.kernel.org>,
Rik van Riel <riel@...hat.com>,
Ingo Molnar <mingo@...hat.com>,
Michel Lespinasse <walken@...gle.com>,
"Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>,
Sasha Levin <sasha.levin@...cle.com>,
Benjamin Herrenschmidt <benh@...nel.crashing.org>
Subject: Re: pipe/page fault oddness.
Mel Gorman <mgorman@...e.de> writes:
> On Wed, Oct 01, 2014 at 09:18:25AM -0700, Linus Torvalds wrote:
>> On Wed, Oct 1, 2014 at 9:01 AM, Linus Torvalds
>> <torvalds@...ux-foundation.org> wrote:
>> >
>> > We need to get rid of it, and just make it the same as pte_protnone().
>> > And then the real protnone is in the vma flags, and if you actually
>> > ever get to a pte that is marked protnone, you know it's a numa page.
>>
>> So I'd really suggest we do exactly that. Get rid of "pte_numa()"
>> entirely, get rid of "_PAGE_[BIT_]NUMA" entirely, and instead add a
>> "pte_protnone()" helper to check for the "protnone" case (which on x86
>> is testing the _PAGE_PROTNONE bit, and on most other architectures is
>> just testing that the page has no access rights).
>>
>
> Do not interpret the following as being against the idea of taking the
> pte_protnone approach. This is intended to give background.
>
> At the time the changes were made to the _PAGE_NUMA bits it was acknowledged
> that a full move to prot_none was an option but it was not the preferred
> solution at the time. It replaced one set of corner cases with another and
> the last time like this time, there was considerable time pressure. The
> VMA would be required to distinguish between a NUMA hinting fault and a
> real prot_none bit. In most cases, we have the VMA now with the exception
> of GUP. GUP would have to unconditionally go into the slow path to do the
> VMA lookup. That is not likely to be a big of a problem but it was a concern.
>
> In early implementations based on prot_none there were some VMA-based
> protection checks that had higher overhead. At the time, there were severe
> problems with overhead due to NUMA balancing and adding more was not
> desirable. This has been addressed since with changes in multiple other
> areas so it's much less of a concern now than it was. In the current shape,
> these probably is not as much a problem as long as any check on pte_numa
> was first guarded by a VMA check. One way of handling the corner cases
> where would be to pass in the VMA where available and have a VM_BUG_ON that
> fires if its a PROT_NONE VMA. That would catch problems during debugging
> without adding overhead in the !debug case.
>
> Going back to the start, the PTE bit was used as the approach due to
> concerns that a pte_protnone helper would not work on all architectures,
> ppc64 in particular. There was no PROT_NONE bit there and instead prot_none
> protections rely on PAGE_USER not being set so it's inaccessible from
> userspace. There was discussion at the time that this could conceivably be
> broken from some sub-architectures but I don't recall the details. Looking
> at the current shape and your patch, it's conceivable that the pte_protnone
> could be implemented as a _PAGE_PRESENT && !_PAGE_USER check as long
> as it was guarded by a VMA check which x86 requires anyway. Not sure
> if that would work for PMDs as I'm not familiar with with ppc64 to tell
> offhand. Alternatively, ppc64 would potentially use the bit currently used
> for _PAGE_NUMA as a _PROT_NONE bit.
Are we still looking at these options ? I could look at implementing the
first option which will also enable us to free up one pte bit.
Note: Freeing up one bit will enable us to implement soft dirty tracking
needed for CRIU.
-aneesh
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Powered by blists - more mailing lists