Re: [PATCH] PCI: designware: fix asynchronous external abort in keystone PCIe h/w

[Murali Karicheri]

On 01/10/2017 06:06 PM, Bjorn Helgaas wrote:
> On Tue, Jan 10, 2017 at 12:18:29PM -0500, Murali Karicheri wrote:
>> On 01/10/2017 10:12 AM, Bjorn Helgaas wrote:
>>> Hi Murali,
>>>
>>> On Wed, Jan 04, 2017 at 02:32:30PM -0500, Murali Karicheri wrote:
>>>> Recent fixes for iATU unroll support introduced a bug that causes
>>>> asynchronous external abort in Keystone PCIe h/w which doesn't have
>>>> ATU port and the corresponding register. So the check should be moved
>>>> below where dw_pcie_prog_outbound_atu() is called to avoid that
>>>> being called on keystine PCIe h/w.
>>>>
>>>> Here is the backtrace
>>>>
>>>> [    0.771174] OF: PCI:   MEM 0x60000000..0x6fffffff -> 0x60000000
>>>> [    0.778118] Unhandled fault: asynchronous external abort (0x1211) at 0x00000000
>>>> [    0.785548] pgd = c0003000
>>>> [    0.788347] [00000000] *pgd=80000800004003, *pmd=00000000
>>>> [    0.793864] Internal error: : 1211 [#1] PREEMPT SMP ARM
>>>> [    0.799197] Modules linked in:
>>>> [    0.802351] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.9.0-00009-g6ff59d2-dirty #7
>>>> [    0.810130] Hardware name: Keystone
>>>> [    0.813717] task: eb878000 task.stack: eb866000
>>>> [    0.818356] PC is at dw_pcie_setup_rc+0x24/0x380
>>>> [    0.823083] LR is at ks_pcie_host_init+0x10/0x170
>>>>
>>>> Fixes: 416379f9ebde ("PCI: designware: Check for iATU unroll support after initializing host")
>>>
>>> I tentatively applied this to for-linus for v4.10, with a stable tag
>>> for v4.9+.
>>>
>>
>> Ok  Thanks!
>>
>>> The patch itself mostly makes sense in terms of the code: we only use
>>> iatu_unroll_enabled in dw_pcie_prog_outbound_atu().  We only call that
>>> when "!pp->ops->rd_other_conf" or "!pp->ops->wr_other_conf".
>>>
>>> So it makes sense that we only need to initialize iatu_unroll_enabled
>>> in those cases.  But the current patch only initializes it if
>>> "!pp->ops->rd_other_conf".
>>>
>>
>> I think the code before also should have checked for both 
>> if ((!pp->ops->rd_other_conf) && (!pp->ops->wr_other_conf))
>>
>> The assumption was if rd_other_conf is Null, the platform provides 
>> both rd_other_conf and wr_other_conf. I see I have added this API 
>> to support Keystone and the above assumption is true so far. 
>> It make sense to fix it if you agree.
>>
>>> If we had "pp->ops->rd_other_conf && !pp->ops->wr_other_conf", we
>>> would use iatu_unroll_enabled uninitialized in the
>>> dw_pcie_wr_other_conf() path.  I suppose that's an invalid
>>
>> Yes. So we need to fix the above to make the code correct instead
>> of leaving it exposed.
>>
>>> configuration, but it'd be better if we didn't have to rely on the
>>> host drivers to avoid that configuration.
>>
>> You mean to introduce the check in the designware core code above 
>> right?
> 
> We *could* make the code look like:
> 
>   if (!pp->ops->rd_other_conf || !pp->ops->rd_other_conf)
>     pp->iatu_unroll_enabled = dw_pcie_iatu_unroll_enabled(pp);
> 
> but that seems a little clunky and the connection between ATU and the
> ->*_other_conf() pointers is slightly obscure.  I guess it just ends
> up being more detailed than I personally really want to worry about,
> so I'm fine with leaving it as-is for now.
> 
>>> It's not obvious how this is connected to 416379f9ebde, though.  It
>>> *looks* like we call dw_pcie_iatu_unroll_enabled() on Keystone both
>>> before and after that commit, so it seems like the external abort
>>> should have happened even before it.
>>>
>>
>> You are right. The problem was introduced by 
>>
>> commit a0601a47053714eecec726aea5ebcd829f817497
>> Author: Joao Pinto <Joao.Pinto@xxxxxxxxxxxx>
>> Date:   Wed Aug 10 11:02:39 2016 +0100
>>
>>     PCI: designware: Add iATU Unroll feature
>>
>> which is fixed by commit 416379f9ebde and is again fixed by my 
>> commit. So probably I should have added both commits in
>> my description. 
> 
> I added that.  The reason I wanted the correct "Fixes" information is
> to figure out which stable kernels need this fix.  In this case, both:
> 
>   a0601a470537 ("PCI: designware: Add iATU Unroll feature") and
>   416379f9ebde ("PCI: designware: Check for iATU unroll support after
>     initializing host")
> 
> appeared in v4.9, so it doesn't change anything as far as the stable
> tag is concerned.
> 
> The below is what I have on my for-linus branch:
> 
> commit 6fdb996a55684016de1ce639b9316b7092fde95f
> Author: Murali Karicheri <m-karicheri2@xxxxxx>
> Date:   Wed Jan 4 14:32:30 2017 -0500
> 
>     PCI: designware: Check for iATU unroll only on platforms that use ATU
>     
>     Previously we checked for iATU unroll support by reading PCIE_ATU_VIEWPORT
>     even on platforms, e.g., Keystone, that do not have ATU ports.  This can
>     cause bad behavior such as asynchronous external aborts:
>     
>       OF: PCI:   MEM 0x60000000..0x6fffffff -> 0x60000000
>       Unhandled fault: asynchronous external abort (0x1211) at 0x00000000
>       pgd = c0003000
>       [00000000] *pgd=80000800004003, *pmd=00000000
>       Internal error: : 1211 [#1] PREEMPT SMP ARM
>       Modules linked in:
>       CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.9.0-00009-g6ff59d2-dirty #7
>       Hardware name: Keystone
>       task: eb878000 task.stack: eb866000
>       PC is at dw_pcie_setup_rc+0x24/0x380
>       LR is at ks_pcie_host_init+0x10/0x170
>     
>     Move the dw_pcie_iatu_unroll_enabled() check so we only call it on
>     platforms that do not use the ATU.  These platforms supply their own
>     ->rd_other_conf() and ->wr_other_conf() methods.
>     
>     [bhelgaas: changelog]
>     Fixes: a0601a470537 ("PCI: designware: Add iATU Unroll feature")
>     Fixes: 416379f9ebde ("PCI: designware: Check for iATU unroll support after initializing host")
>     Tested-by: Kishon Vijay Abraham I <kishon@xxxxxx>
>     Signed-off-by: Murali Karicheri <m-karicheri2@xxxxxx>
>     Signed-off-by: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
>     CC: stable@xxxxxxxxxxxxxxx      # v4.9+
> 
> diff --git a/drivers/pci/host/pcie-designware.c b/drivers/pci/host/pcie-designware.c
> index bed19994c1e9..af8f6e92e885 100644
> --- a/drivers/pci/host/pcie-designware.c
> +++ b/drivers/pci/host/pcie-designware.c
> @@ -807,11 +807,6 @@ void dw_pcie_setup_rc(struct pcie_port *pp)
>  {
>  	u32 val;
>  
> -	/* get iATU unroll support */
> -	pp->iatu_unroll_enabled = dw_pcie_iatu_unroll_enabled(pp);
> -	dev_dbg(pp->dev, "iATU unroll: %s\n",
> -		pp->iatu_unroll_enabled ? "enabled" : "disabled");
> -
>  	/* set the number of lanes */
>  	val = dw_pcie_readl_rc(pp, PCIE_PORT_LINK_CONTROL);
>  	val &= ~PORT_LINK_MODE_MASK;
> @@ -882,6 +877,11 @@ void dw_pcie_setup_rc(struct pcie_port *pp)
>  	 * we should not program the ATU here.
>  	 */
>  	if (!pp->ops->rd_other_conf) {
> +		/* get iATU unroll support */
> +		pp->iatu_unroll_enabled = dw_pcie_iatu_unroll_enabled(pp);
> +		dev_dbg(pp->dev, "iATU unroll: %s\n",
> +			pp->iatu_unroll_enabled ? "enabled" : "disabled");
> +
>  		dw_pcie_prog_outbound_atu(pp, PCIE_ATU_REGION_INDEX0,
>  					  PCIE_ATU_TYPE_MEM, pp->mem_base,
>  					  pp->mem_bus_addr, pp->mem_size);
> 
Thanks Bjorn! Looks good.

-- 
Murali Karicheri
Linux Kernel, Keystone