[v3] phb4: Enable PCI peer-to-peer

P9 supports PCI peer-to-peer: a PCI device can write directly to the
mmio space of another PCI device. It completely by-passes the CPU.

It requires some configuration on the PHBs involved:

1. on the initiating side, the address for the read/write operation is
in the mmio space of the target, i.e. well outside the range normally
allowed. So we disable range-checking on the TVT entry in bypass mode.

2. on the target side, we need to explicitly enable p2p by setting a
bit in a configuration register. It has the side-effect of reserving
an outbound (as seen from the CPU) store queue for p2p. Therefore we
only enable p2p on the PHBs using it, as we don't want to waste the
resource if we don't have to.

P9 supports p2p mmio writes. Reads are currently only supported if the
two devices are under the same PHB but that is expected to change in
the future, and it raises questions about intermediate switches
configuration, so we report an error for the time being.

The patch adds a new OPAL call to allow the OS to declare a p2p
(initiator, target) pair.

Signed-off-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
---

Changelog:
v3:
 - move target reference count to linux
 - disable any configuration with the 2 devices under the same PHB
v2:
 - change of API
 - allow disabling of p2p config

 core/pci-opal.c                       | 39 ++++++++++++++++++++++
 doc/opal-api/opal-pci-set-p2p-152.rst | 50 ++++++++++++++++++++++++++++
 hw/phb4.c                             | 62 +++++++++++++++++++++++++++++++++++
 include/opal-api.h                    | 13 +++++++-
 include/pci.h                         |  4 +++
 include/phb4-regs.h                   |  1 +
 6 files changed, 168 insertions(+), 1 deletion(-)
 create mode 100644 doc/opal-api/opal-pci-set-p2p-152.rst

Related linux patch:
http://patchwork.ozlabs.org/patch/789773/

Le 17/07/2017 à 18:39, Frederic Barrat a écrit :
> P9 supports PCI peer-to-peer: a PCI device can write directly to the
> mmio space of another PCI device. It completely by-passes the CPU.
> 
> It requires some configuration on the PHBs involved:
> 
> 1. on the initiating side, the address for the read/write operation is
> in the mmio space of the target, i.e. well outside the range normally
> allowed. So we disable range-checking on the TVT entry in bypass mode.
> 
> 2. on the target side, we need to explicitly enable p2p by setting a
> bit in a configuration register. It has the side-effect of reserving
> an outbound (as seen from the CPU) store queue for p2p. Therefore we
> only enable p2p on the PHBs using it, as we don't want to waste the
> resource if we don't have to.
> 
> P9 supports p2p mmio writes. Reads are currently only supported if the
> two devices are under the same PHB but that is expected to change in
> the future, and it raises questions about intermediate switches
> configuration, so we report an error for the time being.
> 
> The patch adds a new OPAL call to allow the OS to declare a p2p
> (initiator, target) pair.
> 
> Signed-off-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
> ---
> 
> Changelog:
> v3:
>   - move target reference count to linux
>   - disable any configuration with the 2 devices under the same PHB
> v2:
>   - change of API
>   - allow disabling of p2p config
> 
>   core/pci-opal.c                       | 39 ++++++++++++++++++++++
>   doc/opal-api/opal-pci-set-p2p-152.rst | 50 ++++++++++++++++++++++++++++
>   hw/phb4.c                             | 62 +++++++++++++++++++++++++++++++++++
>   include/opal-api.h                    | 13 +++++++-
>   include/pci.h                         |  4 +++
>   include/phb4-regs.h                   |  1 +
>   6 files changed, 168 insertions(+), 1 deletion(-)
>   create mode 100644 doc/opal-api/opal-pci-set-p2p-152.rst
> 
> diff --git a/core/pci-opal.c b/core/pci-opal.c
> index 5d58a88f..b8aec941 100644
> --- a/core/pci-opal.c
> +++ b/core/pci-opal.c
> @@ -977,3 +977,42 @@ static int64_t opal_pci_set_phb_capi_mode(uint64_t phb_id, uint64_t mode, uint64
>   	return rc;
>   }
>   opal_call(OPAL_PCI_SET_PHB_CAPI_MODE, opal_pci_set_phb_capi_mode, 3);
> +
> +static int64_t opal_pci_set_p2p(uint64_t phbid_init, uint64_t phbid_target,
> +				uint64_t desc, uint16_t pe_number)
> +{
> +	struct phb *phb_init = pci_get_phb(phbid_init);
> +	struct phb *phb_target = pci_get_phb(phbid_target);
> +
> +	if (!phb_init || !phb_target)
> +		return OPAL_PARAMETER;
> +	/*
> +	 * Having the 2 devices under the same PHB may require tuning
> +	 * the configuration of intermediate switch(es), more easily
> +	 * done from linux. And it shouldn't require a PHB config
> +	 * change.
> +	 * Return an error for the time being.
> +	 */
> +	if (phb_init == phb_target)
> +		return OPAL_UNSUPPORTED;
> +	if (!phb_init->ops->set_p2p || !phb_target->ops->set_p2p)
> +		return OPAL_UNSUPPORTED;
> +	/*
> +	 * Loads would be supported on p9 if the 2 devices are under
> +	 * the same PHB, but we ruled it out above.
> +	 */
> +	if (desc & OPAL_PCI_P2P_LOAD)
> +		return OPAL_UNSUPPORTED;
> +
> +	phb_lock(phb_init);
> +	phb_init->ops->set_p2p(phb_init, OPAL_PCI_P2P_INITIATOR, desc,
> +			pe_number);
> +	phb_unlock(phb_init);
> +
> +	phb_lock(phb_target);
> +	phb_target->ops->set_p2p(phb_target, OPAL_PCI_P2P_TARGET, desc,
> +				pe_number);
> +	phb_unlock(phb_target);
> +	return OPAL_SUCCESS;
> +}
> +opal_call(OPAL_PCI_SET_P2P, opal_pci_set_p2p, 4);
> diff --git a/doc/opal-api/opal-pci-set-p2p-152.rst b/doc/opal-api/opal-pci-set-p2p-152.rst
> new file mode 100644
> index 00000000..5b66dd4f
> --- /dev/null
> +++ b/doc/opal-api/opal-pci-set-p2p-152.rst
> @@ -0,0 +1,50 @@
> +OPAL_PCI_SET_P2P
> +================
> +::
> +
> +   #define OPAL_PCI_SET_P2P			152
> +
> +   int64_t opal_pci_set_p2p(uint64_t phbid_init, uint64_t phbid_target,
> +				uint64_t desc, uint16_t pe_number)
> +
> +   /* PCI p2p descriptor */
> +   #define OPAL_PCI_P2P_ENABLE		0x1
> +   #define OPAL_PCI_P2P_LOAD		0x2
> +   #define OPAL_PCI_P2P_STORE		0x4
> +
> +The host calls this function to enable PCI peer-to-peer on the PHBs.
> +
> +Parameters
> +----------
> +::
> +
> +   uint64_t phbid_init
> +   uint64_t phbid_target
> +   uint64_t desc
> +   uint16_t pe_number
> +
> +
> +``phbid_init``
> +  is the value from the PHB node ibm,opal-phbid property for the device initiating the p2p operation
> +
> +``phbid_target``
> +  is the value from the PHB node ibm,opal-phbid property for the device targeted by the p2p operation
> +
> +``desc``
> +  tells whether the p2p operation is a store (OPAL_PCI_P2P_STORE) or load (OPAL_PCI_P2P_LOAD). Can be both.
> +  OPAL_PCI_P2P_ENABLE enables/disables the setting
> +
> +``pe_number``
> +  PE number for the initiating device
> +
> +Return Values
> +-------------
> +
> +``OPAL_SUCCESS``
> +  Configuration was successful
> +
> +``OPAL_PARAMETER``
> +  Invalid PHB or mode parameter
> +
> +``OPAL_UNSUPPORTED``
> +  Not supported by hardware
> diff --git a/hw/phb4.c b/hw/phb4.c
> index 2f3af3ae..2fcee343 100644
> --- a/hw/phb4.c
> +++ b/hw/phb4.c
> @@ -3394,6 +3394,67 @@ static int64_t phb4_set_capi_mode(struct phb *phb, uint64_t mode,
>   	return OPAL_UNSUPPORTED;
>   }
>   
> +static void phb4_p2p_set_initiator(struct phb4 *p, uint16_t pe_number)
> +{
> +	uint64_t tve;
> +	uint16_t window_id = (pe_number << 1) + 1;
> +
> +	/*
> +	 * Initiator needs access to the MMIO space of the target,
> +	 * which is well beyond the 'normal' memory area. Set its TVE
> +	 * with no range checking.
> +	 */
> +	PHBDBG(p, "Setting TVE#1 for peer-to-peer for pe %d\n", pe_number);
> +	tve = PPC_BIT(51);
> +	phb4_ioda_sel(p, IODA3_TBL_TVT, window_id, false);
> +	out_be64(p->regs + PHB_IODA_DATA0, tve);
> +	p->tve_cache[window_id] = tve;
> +}
> +
> +static void phb4_p2p_set_target(struct phb4 *p, bool enable)
> +{
> +	uint64_t val;
> +
> +	/*
> +	 * Enabling p2p on a target PHB reserves an outbound (as seen
> +	 * from the CPU) store queue for p2p
> +	 */
> +	PHBDBG(p, "%s peer-to-peer\n", (enable ? "Enabling" : "Disabling"));
> +	xscom_read(p->chip_id,
> +		p->pe_stk_xscom + XPEC_NEST_STK_PBCQ_MODE, &val);
> +	if (enable)
> +		val |= XPEC_NEST_STK_PBCQ_MODE_P2P;
> +	else
> +		val &= ~XPEC_NEST_STK_PBCQ_MODE_P2P;
> +	xscom_write(p->chip_id,
> +		p->pe_stk_xscom + XPEC_NEST_STK_PBCQ_MODE, val);
> +}
> +
> +static void phb4_set_p2p(struct phb *phb, uint64_t mode, uint64_t flags,
> +			uint16_t pe_number)
> +{
> +	struct phb4 *p = phb_to_phb4(phb);
> +
> +	switch (mode) {
> +	case OPAL_PCI_P2P_INITIATOR:
> +		if (flags & OPAL_PCI_P2P_ENABLE)
> +			phb4_p2p_set_initiator(p, pe_number);
> +		/*
> +		 * When disabling p2p on the initiator, we should
> +		 * reset the TVE to its default bypass setting, but it
> +		 * is more easily done from the OS, as it knows the
> +		 * the start and end address and there's already an
> +		 * opal call for it, so let linux handle it.
> +		 */
> +		break;
> +	case OPAL_PCI_P2P_TARGET:
> +		phb4_p2p_set_target(p, !!(flags & OPAL_PCI_P2P_ENABLE));
> +		break;
> +	default:
> +		assert(0);
> +	}
> +}
> +
>   static const struct phb_ops phb4_ops = {
>   	.cfg_read8		= phb4_pcicfg_read8,
>   	.cfg_read16		= phb4_pcicfg_read16,
> @@ -3427,6 +3488,7 @@ static const struct phb_ops phb4_ops = {
>   	.get_diag_data2		= phb4_get_diag_data,
>   	.tce_kill		= phb4_tce_kill,
>   	.set_capi_mode		= phb4_set_capi_mode,
> +	.set_p2p		= phb4_set_p2p,
>   };
>   
>   static void phb4_init_ioda3(struct phb4 *p)
> diff --git a/include/opal-api.h b/include/opal-api.h
> index edae0693..58b8c8b9 100644
> --- a/include/opal-api.h
> +++ b/include/opal-api.h
> @@ -207,7 +207,8 @@
>   #define OPAL_IMC_COUNTERS_INIT			149
>   #define OPAL_IMC_COUNTERS_START			150
>   #define OPAL_IMC_COUNTERS_STOP			151
> -#define OPAL_LAST				151
> +#define OPAL_PCI_SET_P2P			152
> +#define OPAL_LAST				152
>   
>   /* Device tree flags */
>   
> @@ -1255,6 +1256,16 @@ enum {
>   };
>   
>   
> +/* PCI p2p descriptor */
> +#define OPAL_PCI_P2P_ENABLE		0x1
> +#define OPAL_PCI_P2P_LOAD		0x2
> +#define OPAL_PCI_P2P_STORE		0x4
> +
> +enum {
> +	OPAL_PCI_P2P_INITIATOR	= 0,
> +	OPAL_PCI_P2P_TARGET	= 1,
> +};
> +
>   #endif /* __ASSEMBLY__ */
>   
>   #endif /* __OPAL_API_H */
> diff --git a/include/pci.h b/include/pci.h
> index f216594c..54a62fd1 100644
> --- a/include/pci.h
> +++ b/include/pci.h
> @@ -323,6 +323,10 @@ struct phb_ops {
>   				 uint64_t pe_number);
>   
>   	int64_t (*set_capp_recovery)(struct phb *phb);
> +
> +	/* PCI peer-to-peer setup */
> +	void (*set_p2p)(struct phb *phb, uint64_t mode, uint64_t flags,
> +			uint16_t pe_number);
>   };
>   
>   enum phb_type {
> diff --git a/include/phb4-regs.h b/include/phb4-regs.h
> index 14e93511..a7842a36 100644
> --- a/include/phb4-regs.h
> +++ b/include/phb4-regs.h
> @@ -334,6 +334,7 @@
>   #define XPEC_NEST_STK_ERR_RPT1			0xb
>   #define XPEC_NEST_STK_PBCQ_STAT			0xc
>   #define XPEC_NEST_STK_PBCQ_MODE			0xd
> +#define   XPEC_NEST_STK_PBCQ_MODE_P2P		PPC_BIT(0)
>   #define XPEC_NEST_STK_MMIO_BAR0			0xe
>   #define XPEC_NEST_STK_MMIO_BAR0_MASK		0xf
>   #define XPEC_NEST_STK_MMIO_BAR1			0x10
>