[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <ZLQIeVQ6ITF8RMB/@arm.com>
Date: Sun, 16 Jul 2023 08:10:49 -0700
From: Catalin Marinas <catalin.marinas@....com>
To: Anshuman Khandual <anshuman.khandual@....com>
Cc: Mark Rutland <mark.rutland@....com>,
linux-arm-kernel@...ts.infradead.org,
Will Deacon <will@...nel.org>,
Ryan Roberts <ryan.roberts@....com>,
Andrew Morton <akpm@...ux-foundation.org>,
David Hildenbrand <david@...hat.com>,
Jonathan Corbet <corbet@....net>, linux-kernel@...r.kernel.org,
linux-doc@...r.kernel.org
Subject: Re: [RFC 0/4] arm64/mm: Clean up pte_dirty() state management
On Wed, Jul 12, 2023 at 09:31:39AM +0530, Anshuman Khandual wrote:
> On 7/10/23 16:55, Mark Rutland wrote:
> > On Fri, Jul 07, 2023 at 11:03:27AM +0530, Anshuman Khandual wrote:
> >> These pte_dirty() changes make things explicitly clear, while improving the
> >> code readability. This optimizes HW dirty state transfer into SW dirty bit.
> >> This also adds a new arm64 documentation explaining overall pte dirty state
> >> management in detail. This series applies on the latest mainline kernel.
> >
> > TBH, I think this is all swings and roundabouts, and I'm not sure this is
> > worthwhile. I appreciate that as-is some people find this confusing, but I
I'm pretty much on the same lines, though maybe I looked too much at
this code that I don't like any further changes to it ;).
> Current situation for pte_dirty() management is confusing when there are two
> distinct mechanisms to track PTE dirty states, but both are forced to work
> together because
>
> - HW DBM cannot track non-writable dirty state (PTE_DBM == PTE_WRITE)
> - Runtime check for HW DBM is avoided
Depending on how you look at it, we can say that any writeable PTE (as
in page table permission, PTE_RDONLY cleared) is dirty and we only have
a software mechanism for tracking the dirty state. The DBM feature is
not actually giving us a dirty bit but an automated way to make a PTE
writeable on access (for some historical reasons like the SMMU not
having such mechanism in place).
Maybe we can clean the code a bit based on the above perspective. E.g.
instead of pte_hw_dirty() just have a !pte_hw_rdonly() macro. It may
help with the confusion of having two mechanisms.
OTOH, with PIE, we can have a true dirty bit but at that point we can
eliminate the pte_sw_dirty() use entirely and allow soft-dirty using the
current PTE_DIRTY (with some static labels based on the feature).
> > don't think the end result of this series is actually better, and it adds more
> > code/documentation to maintain.
>
> Agreed, it does add more code and documentation but still trying to understand
> why it is not worthwhile. Regardless, following patch does optimize a situation
> where we dont need to call pte_mkdirty() knowing it will be cleared afterwards.
>
> [RFC 2/4] arm64/mm: Call pte_sw_mkdirty() while preserving the HW dirty state
I wonder whether the compiler eliminates much of this duplication since
there are some checks for pte_write() before. We may be able to remove
some checks. For example, does pte_hw_dirty() actually need to check
pte_write()? A !PTE_RDONLY entry is dirty automatically since we can't
trap any write access to it (prior to PIE; I need to check Joey's
patches on how it treats writeable+clean PTEs; still on holiday).
As for the fourth patch, I'd rather add documentation in the header
file, it's more likely to be looked at and updated.
--
Catalin
Powered by blists - more mailing lists