Re: [PATCH 2/4] mm: Add arch hooks for saving/restoring tags

From: Steven Price <steven.price@arm.com>
To: Catalin Marinas <catalin.marinas@arm.com>,
	Dave Hansen <dave.hansen@intel.com>
Cc: linux-arm-kernel@lists.infradead.org, linux-mm@kvack.org,
	linux-arch@vger.kernel.org,
	Andrew Morton <akpm@linux-foundation.org>,
	Arnd Bergmann <arnd@arndb.de>, Hugh Dickins <hughd@google.com>,
	Vincenzo Frascino <vincenzo.frascino@arm.com>,
	Will Deacon <will@kernel.org>
Subject: Re: [PATCH 2/4] mm: Add arch hooks for saving/restoring tags
Date: Thu, 23 Apr 2020 13:37:21 +0100	[thread overview]
Message-ID: <380bdf0b-aa7c-399c-d16d-b2aecd3f363c@arm.com> (raw)
In-Reply-To: <20200423090919.GA4963@gaia>

On 23/04/2020 10:09, Catalin Marinas wrote:
> On Wed, Apr 22, 2020 at 11:08:10AM -0700, Dave Hansen wrote:
>> On 4/22/20 7:25 AM, Steven Price wrote:
>>> Three new hooks are added to the swap code:
>>>   * arch_prepare_to_swap() and
>>>   * arch_swap_invalidate_page() / arch_swap_invalidate_area().
>>> One new hook is added to shmem:
>>>   * arch_swap_restore_tags()
>>
>> How do the tags get restored outside of the shmem path?  I was expecting
>> to see more arch_swap_restore_tags() sites.
> 
> The restoring is done via set_pte_at() -> mte_sync_tags() ->
> mte_restore_tags() in the arch code (see patch 3).
> arch_swap_restore_tags() just calls mte_restore_tags() directly.
> 
> shmem is slightly problematic as it moves the page from the swap cache
> to the shmem one and I think arch_swap_invalidate_page() would have
> already been called by the time we get to set_pte_at() (Steven can
> correct me if I got this wrong).

That's correct - shmem can pull in pages (into it's own cache) and 
invalidate the swap entries without any process having a PTE restored. 
So we need to hook shmem to restore the tags even though there's no PTE 
restored yet.

The set_pte_at() 'trick' enables delaying the restoring of the tags (in 
the usual case) until the I/O for the page has completed, which might be 
necessary in some cases if the I/O can clobber the tags in memory. I 
couldn't find a better way of hooking this.

Steve