Date:   Tue, 23 Jun 2020 11:11:31 +1000
From:   Alexey Kardashevskiy <aik@...abs.ru>
To:     Leonardo Bras <leobras.c@...il.com>,
        Michael Ellerman <mpe@...erman.id.au>,
        Benjamin Herrenschmidt <benh@...nel.crashing.org>,
        Paul Mackerras <paulus@...ba.org>,
        Thiago Jung Bauermann <bauerman@...ux.ibm.com>,
        Ram Pai <linuxram@...ibm.com>
Cc:     linuxppc-dev@...ts.ozlabs.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH 4/4] powerpc/pseries/iommu: Remove default DMA window
 before creating DDW



On 23/06/2020 04:59, Leonardo Bras wrote:
> Hello Alexey, thanks for the feedback!
> 
> On Mon, 2020-06-22 at 20:02 +1000, Alexey Kardashevskiy wrote:
>>
>> On 19/06/2020 15:06, Leonardo Bras wrote:
>>> On LoPAR "DMA Window Manipulation Calls", it's recommended to remove the
>>> default DMA window for the device, before attempting to configure a DDW,
>>> in order to make the maximum resources available for the next DDW to be
>>> created.
>>>
>>> This is a requirement for some devices to use DDW, given they only
>>> allow one DMA window.
>>>
>>> If setting up a new DDW fails anywhere after the removal of this
>>> default DMA window, restore it using reset_dma_window.
>>
>> Nah... If we do it like this, then under pHyp we lose 32bit DMA for good
>> as pHyp can only create a single window and it has to map at
>> 0x800.0000.0000.0000. They probably do not care though.
>>
>> Under KVM, this will fail as VFIO allows creating  2 windows and it
>> starts from 0 but the existing iommu_bypass_supported_pSeriesLP() treats
>> the window address == 0 as a failure. And we want to keep both DMA
>> windows for PCI adapters with both 64bit and 32bit PCI functions (I
>> heard AMD GPU video + audio are like this) or someone could hotplug
>> 32bit DMA device on a vphb with already present 64bit DMA window so we
>> do not remove the default window.
> 
> Well, then I would suggest doing something like this:
> 	query_ddw(...);
>   	if (query.windows_available == 0){
> 		remove_dma_window(...,default_win);
> 		query_ddw(...);
> 	}
> 
> This would make sure to cover cases of windows available == 1
> and windows available > 1; 


Is "1" what pHyp returns on query? And was it always like that? Then it
is probably ok. I just never really explored the idea of removing the
default window as we did not have to.


>> The last discussed thing I remember was that there was supposed to be a
>> new bit in "ibm,architecture-vec-5" (forgot the details), we could use
>> that to decide whether to keep the default window or not, like this.
> 
> I checked on the latest LoPAR draft (soon to be published), for the
> ibm,architecture-vec 'option array 5' and this entry was the only
> recently added one that is related to this patchset:
> 
> Byte 8 - Bit 0:
> SRIOV Virtual Functions Support Dynamic DMA Windows (DDW):
> 0: SRIOV Virtual Functions do not support DDW
> 1: SRIOV Virtual Functions do support DDW
> 
> Isn't this equivalent to having a "ibm,ddw-applicable" property?

I am not sure, is there anything else to this new bit? I'd think if the
client supports it, then pHyp will create one 64bit window per a PE and
DDW API won't be needed. Thanks,


> 
> 
>>
>>> Signed-off-by: Leonardo Bras <leobras.c@...il.com>
>>> ---
>>>  arch/powerpc/platforms/pseries/iommu.c | 20 +++++++++++++++++---
>>>  1 file changed, 17 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/arch/powerpc/platforms/pseries/iommu.c b/arch/powerpc/platforms/pseries/iommu.c
>>> index de633f6ae093..68d1ea957ac7 100644
>>> --- a/arch/powerpc/platforms/pseries/iommu.c
>>> +++ b/arch/powerpc/platforms/pseries/iommu.c
>>> @@ -1074,8 +1074,9 @@ static u64 enable_ddw(struct pci_dev *dev, struct device_node *pdn)
>>>  	u64 dma_addr, max_addr;
>>>  	struct device_node *dn;
>>>  	u32 ddw_avail[3];
>>> +
>>>  	struct direct_window *window;
>>> -	struct property *win64;
>>> +	struct property *win64, *dfl_win;
>>
>> Make it "default_win" or "def_win", "dfl" hurts to read :)
> 
> Sure, no problem :)
> 
>>
>>>  	struct dynamic_dma_window_prop *ddwprop;
>>>  	struct failed_ddw_pdn *fpdn;
>>>  
>>> @@ -1110,8 +1111,19 @@ static u64 enable_ddw(struct pci_dev *dev, struct device_node *pdn)
>>>  	if (ret)
>>>  		goto out_failed;
>>>  
>>> -       /*
>>> -	 * Query if there is a second window of size to map the
>>> +	/*
>>> +	 * First step of setting up DDW is removing the default DMA window,
>>> +	 * if it's present. It will make all the resources available to the
>>> +	 * new DDW window.
>>> +	 * If anything fails after this, we need to restore it.
>>> +	 */
>>> +
>>> +	dfl_win = of_find_property(pdn, "ibm,dma-window", NULL);
>>> +	if (dfl_win)
>>> +		remove_dma_window(pdn, ddw_avail, dfl_win);
>>
>> Before doing so, you want to make sure that the "reset" is actually
>> supported. Thanks,
> 
> Good catch, I will improve that.
> 
>>
>>
>>> +
>>> +	/*
>>> +	 * Query if there is a window of size to map the
>>>  	 * whole partition.  Query returns number of windows, largest
>>>  	 * block assigned to PE (partition endpoint), and two bitmasks
>>>  	 * of page sizes: supported and supported for migrate-dma.
>>> @@ -1219,6 +1231,8 @@ static u64 enable_ddw(struct pci_dev *dev, struct device_node *pdn)
>>>  	kfree(win64);
>>>  
>>>  out_failed:
>>> +	if (dfl_win)
>>> +		reset_dma_window(dev, pdn);
>>>  
>>>  	fpdn = kzalloc(sizeof(*fpdn), GFP_KERNEL);
>>>  	if (!fpdn)
>>>
> 
> Best regards,
> Leonardo
> 

-- 
Alexey

Powered by blists - more mailing lists