Message-ID: <CAGHCxhcad=zx7ihX5zoDB=ZOLGGuZty=Bck6zSoMQ-9S3ZJo7w@mail.gmail.com>
Date:	Mon, 15 Oct 2012 22:20:54 +0200
From:	Dmitry Fleytman <dmitry@...nix.com>
To:	Alexander Duyck <alexander.h.duyck@...el.com>
Cc:	linux-kernel@...r.kernel.org, e1000-devel@...ts.sourceforge.net,
	Chris Webb <chris.webb@...stichosts.com>,
	Richard Davies <richard.davies@...stichosts.com>
Subject: Re: [PATCH] e1000 driver RX race condition fixed

Hello, Alex

Many thanks for clarification. I think your assumption is
correct and this is exactly what needs to be fixed in QEMU.

Is there any publicly available specification for Intel devices that
explains their operation in such a great details?

Dmitry.


On Mon, Oct 15, 2012 at 10:03 PM, Alexander Duyck
<alexander.h.duyck@...el.com> wrote:
> Hello Dmitry,
>
> My concern is that on many of our parts the behavior is to initialize
> both the head and tail to 0, enable Rx for either the ring or device
> depending on the queue configuration, and then allocate buffers and bump
> tail to indicate that the new buffers are present.  The reason behind
> enabling Rx and bumping tail is because that signals the DMA engine to
> start fetching buffers.  In my experience most of our hardware will
> ignore the tail bump if it is done first and the Rx is enabled.
>
> With both head and tail at the same value it should not be possible for
> any of the devices to start a DMA.  This is probably what you should be
> checking for in fixing QEMU/KVM as it may be incorrectly assuming it can
> fetch the descriptor pointed to by tail.
>
> We have your patch in our queue and can test to verify my assumptions
> are correct.  If they are we will let you know and reject the patch.
>
> Thanks,
>
> Alex
>
>
>
> On 10/15/2012 12:44 PM, Dmitry Fleytman wrote:
>> Hello, Alex
>>
>> Originally this bug was reported for virtual machines running on top
>> of QEMU/KVM.
>> After patch preparation I've tested it on physical e1000 card and it
>> worked fine.
>>
>> However, it could be I've missed something, as I see now other Intel
>> drivers (e1000e, ixgb, etc.)
>> use the same sequence (RX enable and then ring allocate), so I'm
>> starting to suspect that this is
>> the correct behavior.
>>
>> If you confirm this is the way HW works, the this patch should be
>> ignored. This is pure QEMU bug and we'll fix it there.
>>
>> Thanks,
>> Dmitry.
>>
>> On Mon, Oct 15, 2012 at 8:53 PM, Alexander Duyck
>> <alexander.h.duyck@...el.com> wrote:
>>> On 10/14/2012 10:19 AM, Dmitry Fleytman wrote:
>>>> There is a race condition in e1000 driver.
>>>> It enables HW receive before RX rings initalization.
>>>> In case of specific timing this may lead to host memory corruption
>>>> due to DMA write to arbitrary memory location.
>>>> Following patch fixes this issue by reordering initialization steps.
>>>>
>>>> Other Intel network drivers does not seem to have this issue.
>>>>
>>>> Dmitry Fleytman (1):
>>>>   RX initialization sequence fixed - enable RX after corresponding ring
>>>>     initialization only
>>>>
>>>>  drivers/net/ethernet/intel/e1000/e1000_ethtool.c |    9 +++++----
>>>>  drivers/net/ethernet/intel/e1000/e1000_main.c    |   18 ++++++++++++++++--
>>>>  2 files changed, 21 insertions(+), 6 deletions(-)
>>>>
>>> What device was it you saw this issue with?  The reason why I ask is
>>> because I suspect this change should cause most of our e1000 hardware to
>>> lock up since normally if you allocate buffers and then enable Rx it
>>> will mean the ring was not updated and it will treat it as if there are
>>> no buffers available.
>>>
>>> Thanks,
>>>
>>> Alex
>



-- 
Dmitry Fleytman
Technology Expert and Consultant,

Daynix Computing Ltd.

Cell: +972-54-2819481

Skype: dmitry.fleytman
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists