MIME-Version: 1.0
In-Reply-To: <20160408065626.GA25580@e106921-lin.trondheim.arm.com>
References: <1460028599-22356-1-git-send-email-john.reitan@foss.arm.com>
	<5706B70E.4040004@redhat.com>
	<20160408065626.GA25580@e106921-lin.trondheim.arm.com>
Date: Fri, 8 Apr 2016 16:11:01 -0700
Message-ID: <CAMbhsRTOpdjG8yShpFUvC3=aswLbFY8p7Q2Cab7pBh+2smY=+Q@mail.gmail.com>
Subject: Re: [PATCH] ion: scatterlist offset not used for buffer map
From: Colin Cross <ccross@google.com>
To: Laura Abbott <labbott@redhat.com>, Greg KH <gregkh@linuxfoundation.org>,
        =?UTF-8?B?QXJ2ZSBIasO4bm5ldsOlZw==?= <arve@android.com>,
        "devel@driverdev.osuosl.org" <devel@driverdev.osuosl.org>,
        lkml <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset=UTF-8
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1380
Lines: 28

On Thu, Apr 7, 2016 at 11:56 PM, John Einar Reitan
<john.reitan@foss.arm.com> wrote:
> On Thu, Apr 07, 2016 at 12:37:50PM -0700, Laura Abbott wrote:
>> On 04/07/2016 04:29 AM, John Einar Reitan wrote:
>> > ion's default user/kernel page mapping code don't honor the offset
>> > option for scatterlists. It uses sg_page and expect the whole page to be
>> > mapped, while the offset could dictate an offset within a large page.
>> >
>> > sg_phys correctly accounts for the offset, so should be used instead.
>> >
>>
>> Can you be more specific about which heap and which allocation pattern
>> is exposing this bug?
>
> The heap that exposed the bug is one I'm developing and will be posting
> as a RFC soon. It uses compound pages and an sub-divides it into surface
> buffers. The ion buffers are configured to hold sgl's with the compound
> page and the correct offset of the buffer, via
> sg_set_page(.., compound_page, .., offset_of_logical_buffer);

I don't think this is right.  A compound_page still has a page struct
for every page, you should be passing the page struct where your data
starts.  Using an offset > PAGE_SIZE is going to break lots of places,
for example anywhere that uses kmap(sg_page(sg)).

> sg_phys/sg_virt  includes this offset, but if you poke the sg and extract
> the page with sg_page yourself you must include this offset in your
> calculations too.