Re: [PATCH v1] drm/ttm: Refcount allocated tail pages

From: "Christian König" <christian.koenig@amd.com>
To: Dmitry Osipenko <dmitry.osipenko@collabora.com>,
	David Airlie <airlied@linux.ie>, Huang Rui <ray.huang@amd.com>,
	Daniel Vetter <daniel@ffwll.ch>,
	Trigger Huang <Trigger.Huang@gmail.com>,
	Gert Wollny <gert.wollny@collabora.com>,
	Antonio Caggiano <antonio.caggiano@collabora.com>
Cc: dri-devel@lists.freedesktop.org, linux-kernel@vger.kernel.org,
	Dmitry Osipenko <digetx@gmail.com>,
	kvm@vger.kernel.org, kernel@collabora.com,
	virtualization@lists.linux-foundation.org
Subject: Re: [PATCH v1] drm/ttm: Refcount allocated tail pages
Date: Mon, 15 Aug 2022 12:42:22 +0200	[thread overview]
Message-ID: <2e64c037-ee5a-0c60-5dba-2e48bb437014@amd.com> (raw)
In-Reply-To: <14be3b22-1d60-732b-c695-ddacc6b21055@collabora.com>

Am 15.08.22 um 12:18 schrieb Dmitry Osipenko:
> On 8/15/22 13:14, Christian König wrote:
>> Am 15.08.22 um 12:11 schrieb Christian König:
>>> Am 15.08.22 um 12:09 schrieb Dmitry Osipenko:
>>>> On 8/15/22 13:05, Christian König wrote:
>>>>> Am 15.08.22 um 11:54 schrieb Dmitry Osipenko:
>>>>>> Higher order pages allocated using alloc_pages() aren't refcounted and
>>>>>> they
>>>>>> need to be refcounted, otherwise it's impossible to map them by
>>>>>> KVM. This
>>>>>> patch sets the refcount of the tail pages and fixes the KVM memory
>>>>>> mapping
>>>>>> faults.
>>>>>>
>>>>>> Without this change guest virgl driver can't map host buffers into
>>>>>> guest
>>>>>> and can't provide OpenGL 4.5 profile support to the guest. The host
>>>>>> mappings are also needed for enabling the Venus driver using host GPU
>>>>>> drivers that are utilizing TTM.
>>>>>>
>>>>>> Based on a patch proposed by Trigger Huang.
>>>>> Well I can't count how often I have repeated this: This is an
>>>>> absolutely
>>>>> clear NAK!
>>>>>
>>>>> TTM pages are not reference counted in the first place and because of
>>>>> this giving them to virgl is illegal.
>>>> A? The first page is refcounted when allocated, the tail pages are not.
>>> No they aren't. The first page is just by coincident initialized with
>>> a refcount of 1. This refcount is completely ignored and not used at all.
>>>
>>> Incrementing the reference count and by this mapping the page into
>>> some other address space is illegal and corrupts the internal state
>>> tracking of TTM.
>> See this comment in the source code as well:
>>
>>          /* Don't set the __GFP_COMP flag for higher order allocations.
>>           * Mapping pages directly into an userspace process and calling
>>           * put_page() on a TTM allocated page is illegal.
>>           */
>>
>> I have absolutely no idea how somebody had the idea he could do this.
> I saw this comment, but it doesn't make sense because it doesn't explain
> why it's illegal. Hence it looks like a bogus comment since the
> refcouting certainly works, at least to a some degree because I haven't
> noticed any problems in practice, maybe by luck :)

Well exactly that's the problem. It does not work, you are just lucky :)

I will provide a patch to set the reference count to zero even for 
non-compound pages. Maybe that will yield more backtrace to abusers of 
this interface.

Regards,
Christian.

>
> I'll try to dig out the older discussions, thank you for the quick reply!
>