Re: [PATCH kernel v8 14/31] vfio: powerpc/spapr: powerpc/powernv/ioda2: Rework IOMMU ownership control

[Alexey Kardashevskiy <aik@ozlabs.ru>]

On 04/21/2015 07:43 PM, David Gibson wrote:
> On Mon, Apr 20, 2015 at 04:55:32PM +1000, Alexey Kardashevskiy wrote:
>> On 04/20/2015 12:44 PM, David Gibson wrote:
>>> On Fri, Apr 17, 2015 at 08:09:29PM +1000, Alexey Kardashevskiy wrote:
>>>> On 04/16/2015 04:07 PM, David Gibson wrote:
>>>>> On Fri, Apr 10, 2015 at 04:30:56PM +1000, Alexey Kardashevskiy wrote:
>>>>>> At the moment the iommu_table struct has a set_bypass() which enables/
>>>>>> disables DMA bypass on IODA2 PHB. This is exposed to POWERPC IOMMU code
>>>>>> which calls this callback when external IOMMU users such as VFIO are
>>>>>> about to get over a PHB.
>>>>>>
>>>>>> The set_bypass() callback is not really an iommu_table function but
>>>>>> IOMMU/PE function. This introduces a iommu_table_group_ops struct and
>>>>>> adds a set_ownership() callback to it which is called when an external
>>>>>> user takes control over the IOMMU.
>>>>>
>>>>> Do you really need separate ops structures at both the single table
>>>>> and table group level?  The different tables in a group will all
>>>>> belong to the same basic iommu won't they?
>>>>
>>>>
>>>> IOMMU tables exist alone in VIO. Also, the platform code uses just a table
>>>> (or it is in bypass mode) and does not care about table groups. It looked
>>>> more clean for myself to keep them separated. Should I still merge
>>>> those?
>>>
>>> Ok, that sounds like a reasonable argument for keeping them separate,
>>> at least for now.
>>>
>>>>>> This renames set_bypass() to set_ownership() as it is not necessarily
>>>>>> just enabling bypassing, it can be something else/more so let's give it
>>>>>> more generic name. The bool parameter is inverted.
>>>>>>
>>>>>> The callback is implemented for IODA2 only. Other platforms (P5IOC2,
>>>>>> IODA1) will use the old iommu_take_ownership/iommu_release_ownership API.
>>>>>>
>>>>>> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
>>>>>> ---
>>>>>>   arch/powerpc/include/asm/iommu.h          | 14 +++++++++++++-
>>>>>>   arch/powerpc/platforms/powernv/pci-ioda.c | 30 ++++++++++++++++++++++--------
>>>>>>   drivers/vfio/vfio_iommu_spapr_tce.c       | 25 +++++++++++++++++++++----
>>>>>>   3 files changed, 56 insertions(+), 13 deletions(-)
>>>>>>
>>>>>> diff --git a/arch/powerpc/include/asm/iommu.h b/arch/powerpc/include/asm/iommu.h
>>>>>> index b9e50d3..d1f8c6c 100644
>>>>>> --- a/arch/powerpc/include/asm/iommu.h
>>>>>> +++ b/arch/powerpc/include/asm/iommu.h
>>>>>> @@ -92,7 +92,6 @@ struct iommu_table {
>>>>>>   	unsigned long  it_page_shift;/* table iommu page size */
>>>>>>   	struct iommu_table_group *it_group;
>>>>>>   	struct iommu_table_ops *it_ops;
>>>>>> -	void (*set_bypass)(struct iommu_table *tbl, bool enable);
>>>>>>   };
>>>>>>
>>>>>>   /* Pure 2^n version of get_order */
>>>>>> @@ -127,11 +126,24 @@ extern struct iommu_table *iommu_init_table(struct iommu_table * tbl,
>>>>>>
>>>>>>   #define IOMMU_TABLE_GROUP_MAX_TABLES	1
>>>>>>
>>>>>> +struct iommu_table_group;
>>>>>> +
>>>>>> +struct iommu_table_group_ops {
>>>>>> +	/*
>>>>>> +	 * Switches ownership from the kernel itself to an external
>>>>>> +	 * user. While onwership is enabled, the kernel cannot use IOMMU
>>>>>> +	 * for itself.
>>>>>> +	 */
>>>>>> +	void (*set_ownership)(struct iommu_table_group *table_group,
>>>>>> +			bool enable);
>>>>>
>>>>> The meaning of "enable" in a function called "set_ownership" is
>>>>> entirely obscure.
>>>>
>>>> Suggest something better please :) I have nothing better...
>>>
>>> Well, given it's "set_ownershuip" you could have "owner" - that would
>>> want to be an enum with OWNER_KERNEL and OWNER_VFIO or something
>>> rather than a bool.
>>
>>
>> It is iommu_take_ownership() in upstream and it is assumed that the owner is
>> anything but the platform code (for now and probably for ever - VFIO). I am
>> not changing this now, just using same naming approach when adding a
>> callback with a similar name.
>
> So "enabled" is actually that non kernel ownership is enabled.  That
> is totally non-obvious.
>
>>> Or you could leave it a bool but call it "allow_bypass".
>>
>> Commented below.
>>
>>>>>> +};
>>>>>> +
>>>>>>   struct iommu_table_group {
>>>>>>   #ifdef CONFIG_IOMMU_API
>>>>>>   	struct iommu_group *group;
>>>>>>   #endif
>>>>>>   	struct iommu_table tables[IOMMU_TABLE_GROUP_MAX_TABLES];
>>>>>> +	struct iommu_table_group_ops *ops;
>>>>>>   };
>>>>>>
>>>>>>   #ifdef CONFIG_IOMMU_API
>>>>>> diff --git a/arch/powerpc/platforms/powernv/pci-ioda.c b/arch/powerpc/platforms/powernv/pci-ioda.c
>>>>>> index a964c50..9687731 100644
>>>>>> --- a/arch/powerpc/platforms/powernv/pci-ioda.c
>>>>>> +++ b/arch/powerpc/platforms/powernv/pci-ioda.c
>>>>>> @@ -1255,10 +1255,8 @@ static void pnv_pci_ioda_setup_dma_pe(struct pnv_phb *phb,
>>>>>>   		__free_pages(tce_mem, get_order(TCE32_TABLE_SIZE * segs));
>>>>>>   }
>>>>>>
>>>>>> -static void pnv_pci_ioda2_set_bypass(struct iommu_table *tbl, bool enable)
>>>>>> +static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable)
>>>>>>   {
>>>>>> -	struct pnv_ioda_pe *pe = container_of(tbl->it_group, struct pnv_ioda_pe,
>>>>>> -					      table_group);
>>>>>>   	uint16_t window_id = (pe->pe_number << 1 ) + 1;
>>>>>>   	int64_t rc;
>>>>>>
>>>>>> @@ -1286,7 +1284,8 @@ static void pnv_pci_ioda2_set_bypass(struct iommu_table *tbl, bool enable)
>>>>>>   		 * host side.
>>>>>>   		 */
>>>>>>   		if (pe->pdev)
>>>>>> -			set_iommu_table_base(&pe->pdev->dev, tbl);
>>>>>> +			set_iommu_table_base(&pe->pdev->dev,
>>>>>> +					&pe->table_group.tables[0]);
>>>>>>   		else
>>>>>>   			pnv_ioda_setup_bus_dma(pe, pe->pbus, false);
>>>>>>   	}
>>>>>> @@ -1302,13 +1301,27 @@ static void pnv_pci_ioda2_setup_bypass_pe(struct pnv_phb *phb,
>>>>>>   	/* TVE #1 is selected by PCI address bit 59 */
>>>>>>   	pe->tce_bypass_base = 1ull << 59;
>>>>>>
>>>>>> -	/* Install set_bypass callback for VFIO */
>>>>>> -	pe->table_group.tables[0].set_bypass = pnv_pci_ioda2_set_bypass;
>>>>>> -
>>>>>>   	/* Enable bypass by default */
>>>>>> -	pnv_pci_ioda2_set_bypass(&pe->table_group.tables[0], true);
>>>>>> +	pnv_pci_ioda2_set_bypass(pe, true);
>>>>>>   }
>>>>>>
>>>>>> +static void pnv_ioda2_set_ownership(struct iommu_table_group *table_group,
>>>>>> +				     bool enable)
>>>>>> +{
>>>>>> +	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
>>>>>> +						table_group);
>>>>>> +	if (enable)
>>>>>> +		iommu_take_ownership(table_group);
>>>>>> +	else
>>>>>> +		iommu_release_ownership(table_group);
>>>>>> +
>>>>>> +	pnv_pci_ioda2_set_bypass(pe, !enable);
>>>>>> +}
>>>>>> +
>>>>>> +static struct iommu_table_group_ops pnv_pci_ioda2_ops = {
>>>>>> +	.set_ownership = pnv_ioda2_set_ownership,
>>>>>> +};
>>>>>> +
>>>>>>   static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
>>>>>>   				       struct pnv_ioda_pe *pe)
>>>>>>   {
>>>>>> @@ -1376,6 +1389,7 @@ static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
>>>>>>   	}
>>>>>>   	tbl->it_ops = &pnv_iommu_ops;
>>>>>>   	iommu_init_table(tbl, phb->hose->node);
>>>>>> +	pe->table_group.ops = &pnv_pci_ioda2_ops;
>>>>>>   	iommu_register_group(&pe->table_group, phb->hose->global_number,
>>>>>>   			pe->pe_number);
>>>>>>
>>>>>> diff --git a/drivers/vfio/vfio_iommu_spapr_tce.c b/drivers/vfio/vfio_iommu_spapr_tce.c
>>>>>> index 9f38351..d5d8c50 100644
>>>>>> --- a/drivers/vfio/vfio_iommu_spapr_tce.c
>>>>>> +++ b/drivers/vfio/vfio_iommu_spapr_tce.c
>>>>>> @@ -535,9 +535,22 @@ static int tce_iommu_attach_group(void *iommu_data,
>>>>>>   		goto unlock_exit;
>>>>>>   	}
>>>>>>
>>>>>> -	ret = iommu_take_ownership(table_group);
>>>>>> -	if (!ret)
>>>>>> -		container->grp = iommu_group;
>>>>>> +	if (!table_group->ops || !table_group->ops->set_ownership) {
>>>>>> +		ret = iommu_take_ownership(table_group);
>>>>>> +	} else {
>>>>>> +		/*
>>>>>> +		 * Disable iommu bypass, otherwise the user can DMA to all of
>>>>>> +		 * our physical memory via the bypass window instead of just
>>>>>> +		 * the pages that has been explicitly mapped into the iommu
>>>>>> +		 */
>>>>>> +		table_group->ops->set_ownership(table_group, true);
>>>>>
>>>>> And here to disable bypass you call it with enable=true, so it doesn't
>>>>> even have the same meaning as it used to.
>>>>
>>>>
>>>> I do not disable bypass per se (even if it what set_ownership(true) does) as
>>>> it is IODA business and VFIO has no idea about it. I do take control over
>>>> the group. I am not following you here - what used to have the same
>>>> meaning?
>>>
>>> Well, in set_bypass, the enable parameter was whether bypass was
>>> enabled.  Here you're setting enable to true, when you want to
>>> *disable* bypass (in the existing case).  If the "enable" parameter
>>> isn't about enabling bypass, it's meaning is even more confusing than
>>> I thought.
>>
>>
>> Its meaning is "take ownership over the group". In this patch
>> set_ownership(true) means set_bypass(false).
>
> Ok.  So "take_ownership" isn't quite as clear as I'd like, but it's
> not too bad because it's implied that it's the caller that's taking
> the ownership.  *set* ownership makes no sense without saying who the
> new owner is.  "enable" has no clear meaning in that context.
>
> Calling it "kernel_owned" or "non_kernel_owned" would be ok if a bit
> clunky.


Strictly speaking VFIO and platform code are both kernel.

So which one to choose?

+struct iommu_table_group_ops {
+	void (*take_ownership)(struct iommu_table_group *table_group);
+	void (*release_ownership)(struct iommu_table_group *table_group);
+};


OR

+enum { IOMMU_TABLE_GROUP_OWNER_KERNEL,	IOMMU_TABLE_GROUP_OWNER_VFIO };
+struct iommu_table_group_ops {
+	void (*set_ownership)(struct iommu_table_group *table_group,
+			long owner);
+};


I have bad taste for names like this, need a hint here, please :)


>
>> But later (in 25/31) set_ownership(true) becomes unset(windows0) +
>> free(table0) + set_bypass(false) = clear DMA setup for the group (i.e.
>> invalidate both TVTs) so it is not just about bypass (which is TVT#1 but not
>> TVT#0) anymore.
>
> Right, I have no problem with a combined function for the operation
> here.  It's purely a naming thing "set_ownership" and "enable" are
> just not concepts that fit together sensibly.



-- 
Alexey