[lkml]   [2016]   [Jul]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH] Exchange the Assignments of `MEMORYs' and `CFGs/IOs' in Designware PCIe Driver
On Mon, Jul 4, 2016 at 1:06 PM, dongbo (E) <> wrote:
> On 2016/7/3 23:27, Pratyush Anand wrote:
>> On Tue, Jun 28, 2016 at 1:42 PM, dongbo (E) <> wrote:
>>> From: Dong Bo <>
>>> In designware PCIe driver, the iatu0 is used for both CFG and IO accesses.
>>> When sending CFGs to peripherals (e.g. lspci), iatu0 frequently switches
>>> between CFG and IO alternatively.
>>> A MEMORY probably be sent as an IOs by mistake. Considering the following
>>> configurations:
>>> MEMORY -> BASE_ADDR: 0xb4100000, LIMIT: 0xb4100FFF, TYPE=mem
>>> CFG -> BASE_ADDR: 0xb4000000, LIMIT: 0xb4000FFF, TYPE=cfg
>>> Suppose PCIe has just completed a CFG access, to switch back to IO, it set
>>> the BASE_ADDR to 0xFFFFFFFF, LIMIT 0xFFFFFFFE and TYPE to io. When another
>>> CFG comes, the BASE_ADDR is set to 0xb4000000 to switch to CFG. At this
>>> moment, a MEMORY access shows up, since it matches with iatu0
>>> (due to 0xb4000000 <= MEMORY BASE_ADDR <= MEMORY LIMIE <= 0xFFFFFFF), it
>>> is treated as an IO access by mistake, then sent to perpheral.
>>> This patch fixes the problem by exchanging the assignments of `MEMORYs'
>>> and `CFGs/IOs', which assigning MEMEORYs to iatu0, CFGs and IOs to iatu1.
>> Had a re-thought on it. While it will fix wrong memory access in your
>> case, it can still cause issues with IO access for some other
>> platform.
>> Can you please test [1] and check it that works for you. You will need
>> to define num-viewport in your device tree file.
>> ~Pratyush
>> [1]
>> .
> Checked, it works for us.

Thanks for testing.

> I think it would be better to exchange the assignments of MEMORYs and
> CFGs/IOs when num_viewports <= 2, cause it fixes wrong memory access.

OK.. I think that can be accommodated. I have rebased your patch on
top of mine with some change in commit log.

Will send both of the patches to the list for review.

> As you mentioned, other corner point for failure is still there while
> there are only two viewports. It seems that there is not a perfect
> solution.

Yes, unfortunately we will have to live with either remote possibility
of less frequent IO transfer corruption or we can disable IO transfer
for <=2 viewports. IMHO, former is the better way to go with.


 \ /
  Last update: 2016-07-04 11:21    [W:0.044 / U:6.260 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site