Re: [PATCH] swiotlb: Fix inversed DMA_ATTR_NO_WARN test

From: "Michel Dänzer" <michel@daenzer.net>
To: Christoph Hellwig <hch@lst.de>, christian.koenig@amd.com
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	dri-devel <dri-devel@lists.freedesktop.org>,
	iommu@lists.linux-foundation.org
Subject: Re: [PATCH] swiotlb: Fix inversed DMA_ATTR_NO_WARN test
Date: Fri, 25 May 2018 10:41:29 +0200	[thread overview]
Message-ID: <0a17ab46-2d42-48ae-cf1f-492e17254767@daenzer.net> (raw)
In-Reply-To: <20180525084120.GA19063@lst.de>

On 2018-05-25 10:41 AM, Christoph Hellwig wrote:
> On Tue, May 22, 2018 at 03:13:58PM +0200, Christian König wrote:
>> Am 02.05.2018 um 18:59 schrieb Michel Dänzer:
>>> On 2018-05-02 06:21 PM, Christoph Hellwig wrote:
>>>> On Wed, May 02, 2018 at 04:31:09PM +0200, Michel Dänzer wrote:
>>>>>> No.  __GFP_NOWARN (and gfp_t flags in general) are the wrong interface
>>>>>> for dma allocations and just cause problems.  I actually plan to
>>>>>> get rid of the gfp_t argument in dma_alloc_attrs sooner, and only
>>>>>> allow either GFP_KERNEL or GFP_DMA passed in dma_alloc_coherent.
>>>>> How about GFP_TRANSHUGE_LIGHT? TTM uses that to opportunistically
>>>>> allocate huge pages (GFP_TRANSHUGE can result in unacceptably long
>>>>> delays with memory pressure).
>>>> Well, that is exactly what I don't want drivers to do - same for
>>>> __GFP_COMP in some drm code.  This very much assumes the page allocator
>>>> is used to back dma allocations, which very often it actually isn't, and
>>>> any use of magic gfp flags creates a tight coupling of consumers with a
>>>> specific implementation.
>>>>
>>>> In general I can't think of a good reason not to actually use
>>>> GFP_TRANSHUGE_LIGHT by default in the dma allocator unless
>>>> DMA_ATTR_ALLOC_SINGLE_PAGES is set.  Can you prepare a patch for that?
>>> I'm afraid I'll have to leave that to somebody else.
>>
>> Coming back to this topic once more, sorry for the delay but busy as usual 
>> :)
>>
>> What exactly do you mean with "dma allocator" here? The TTM allocator using 
>> the dma_alloc_coherent calls? Or the swiotlb implementation of the calls?
> 
> dma allocatr in this case: backends for dma_alloc_coherent/
> dma_alloc_attrs.  Most importantly dma_direct_alloc.
> 
> But while we're at it I can't actually see any GFP_TRANSHUGE_LIGHT
> usage in TTM, just plain old GFP_TRANSHUGE.

See
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=da291320baec914f0bb4e65a9dccb86bd6c728f2
.


-- 
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer