Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII;
 format=flowed
Content-Transfer-Encoding: 7bit
Date:   Fri, 01 May 2020 11:54:08 +0100
From:   Marc Zyngier <maz@kernel.org>
To:     Mark Rutland <mark.rutland@arm.com>,
        Greg KH <gregkh@linuxfoundation.org>
Cc:     Jeremy Linton <jeremy.linton@arm.com>, linux-usb@vger.kernel.org,
        stern@rowland.harvard.edu, git@thegavinli.com,
        jarkko.sakkinen@linux.intel.com, linux-kernel@vger.kernel.org,
        linux-arm-kernel@lists.infradead.org
Subject: Re: [PATCH] usb: usbfs: correct kernel->user page attribute mismatch
In-Reply-To: <20200501103712.GA51954@C02TD0UTHF1T.local>
References: <20200430211922.929165-1-jeremy.linton@arm.com>
 <20200501070500.GA887524@kroah.com>
 <20200501103712.GA51954@C02TD0UTHF1T.local>
User-Agent: Roundcube Webmail/1.4.3
Message-ID: <0aa2a8a9a5b6c00b2e1a1fcf71c121ce@kernel.org>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 2020-05-01 11:37, Mark Rutland wrote:
> On Fri, May 01, 2020 at 09:05:00AM +0200, Greg KH wrote:
>> On Thu, Apr 30, 2020 at 04:19:22PM -0500, Jeremy Linton wrote:
>> > On arm64, and possibly other architectures, requesting
>> > IO coherent memory may return Normal-NC if the underlying
>> > hardware isn't coherent. If these pages are then
>> > remapped into userspace as Normal, that defeats the
>> > purpose of getting Normal-NC, as well as resulting in
>> > mappings with differing cache attributes.
>> 
>> What is "Normal-NC"?
> 
> Arm terminology for "Normal Non-Cacheable"; it might be better to say
> something like:
> 
> On some architectures (e.g. arm64) an IO coherent mapping may use
> non-cachable attributes if the relevant device is cache coherent.

is *not* cache coherent.

> If userspace mappings are cacheable, these may not be coherent with
> non-cacheable mappings. On arm64 this is the case for Normal-NC and
> Normal (cacheable) mappings.

And to answer some of Greg's questions:

- This can show up on current HW that doesn't offer full IO coherency,
   which is likely on low-end arm/arm64 systems.

- I guess nobody noticed this before as x86 is perfectly happy with
   conflicting attributes for the same physical page, and there is
   (wild guess) probably not that much userspace making use of shared
   mappings between kernel and userspace using this interface.

Thanks,

         M.
-- 
Jazz is not dead. It just smells funny...