Re: [PATCH v8 22/23] PCI/P2PDMA: Evaluate the ATS path for clients with ATS enabled

[Thomas Hellström]

On Mon, 2026-09-28 at 14:19 +0300, Leon Romanovsky wrote:
> From: Leon Romanovsky <leonro@xxxxxxxxxx>
> 
> P2PDMA assumes every client issues Untranslated Requests. A device
> with
> ATS enabled may translate any address it is handed, including a bus
> address, and Translation Blocking can reject its Translated Requests
> on
> a route the Untranslated answer called usable.
> 
> Unless the client declared per-mapping ATS, take the Address Type
> from
> its ATS Enable bit. Never hand such a client bus addresses: only an
> IOVA
> survives translation, and whatever it issues untranslated goes
> through
> the host bridge, whose route must therefore be usable. Apply this
> after
> the cache, which keeps holding answers that depend only on the
> topology.
> 
> Signed-off-by: Leon Romanovsky <leonro@xxxxxxxxxx>
Reviewed-by: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx>

> ---
>  Documentation/driver-api/pci/p2pdma.rst |  8 ++++++
>  drivers/pci/p2pdma.c                    | 50
> +++++++++++++++++++++++++++++++--
>  include/linux/pci-p2pdma.h              |  3 +-
>  3 files changed, 58 insertions(+), 3 deletions(-)
> 
> diff --git a/Documentation/driver-api/pci/p2pdma.rst
> b/Documentation/driver-api/pci/p2pdma.rst
> index 7a7f29f455b0..bf9f96c07b58 100644
> --- a/Documentation/driver-api/pci/p2pdma.rst
> +++ b/Documentation/driver-api/pci/p2pdma.rst
> @@ -40,6 +40,14 @@ A client therefore describes its traffic with
> ``enum pci_p2pdma_tlp_flags``
>  and asks ``pci_p2pdma_map_type_tlp()``. ``pci_p2pdma_map_type()``
> answers for
>  the default: strictly ordered Requests carrying an Untranslated
> address.
>  
> +The PCIe ATS Enable bit covers the whole device, and most devices
> translate
> +any address they are handed once it is set. Unless a driver has
> declared with
> +``pcim_p2pdma_set_ats_per_mapping()`` that its device chooses ATS
> per mapping,
> +that bit rather than the flags decides the Address Type, and a
> client with ATS
> +enabled is never handed a bus address, which it would translate as
> though it
> +were an IOVA. Its Translated Requests may still route directly, but
> anything
> +it issues untranslated reaches the host bridge, so that route has to
> work too.
> +
>  The two directions are decided independently. Translation Blocking
> (TB),
>  Direct Translated P2P (DT), Request Redirect (RR) and Egress Control
> (EC) on
>  the client-side port decide the Request:
> diff --git a/drivers/pci/p2pdma.c b/drivers/pci/p2pdma.c
> index ae676509f9f3..ca79af2e64c7 100644
> --- a/drivers/pci/p2pdma.c
> +++ b/drivers/pci/p2pdma.c
> @@ -1595,6 +1595,41 @@ void pcim_p2pdma_set_ats_per_mapping(struct
> pci_dev *pdev)
>  }
>  EXPORT_SYMBOL_GPL(pcim_p2pdma_set_ats_per_mapping);
>  
> +static unsigned int pci_p2pdma_client_tlp_flags(struct pci_dev
> *client,
> +						unsigned int
> tlp_flags)
> +{
> +	if (client->ats_enabled)
> +		return tlp_flags | PCI_P2PDMA_TLP_TRANSLATED;
> +
> +	return tlp_flags & ~PCI_P2PDMA_TLP_TRANSLATED;
> +}
> +
> +/*
> + * A client that translates every address it is handed cannot be
> handed a bus
> + * address, which it would translate as though it were an IOVA. Its
> Translated
> + * Requests may still route directly, but a Request it issues
> without a
> + * translation carries the IOVA to the host bridge, so that route
> has to work
> + * as well.
> + */
> +static enum pci_p2pdma_map_type
> +pci_p2pdma_client_map_type(struct pci_dev *provider, struct pci_dev
> *client,
> +			   bool per_mapping, enum
> pci_p2pdma_map_type type)
> +{
> +	if (per_mapping || !client->ats_enabled ||
> +	    type != PCI_P2PDMA_MAP_BUS_ADDR)
> +		return type;
> +
> +	pci_dbg(client,
> +		"P2PDMA ACS: provider=%s bus address withheld; ATS
> is enabled for the whole client\n",
> +		 pci_name(provider));
> +
> +	if (cpu_supports_p2pdma() ||
> +	    host_bridge_whitelist(provider, client, false))
> +		return PCI_P2PDMA_MAP_THRU_HOST_BRIDGE;
> +
> +	return PCI_P2PDMA_MAP_NOT_SUPPORTED;
> +}
> +
>  /**
>   * pci_p2pdma_map_type_tlp - Determine the mapping type for P2PDMA
> transfers
>   * @provider: P2PDMA provider structure
> @@ -1609,6 +1644,11 @@
> EXPORT_SYMBOL_GPL(pcim_p2pdma_set_ats_per_mapping);
>   * ACS routes a peer-to-peer transaction by the attributes its TLPs
> carry, so
>   * the answer depends on @tlp_flags. A caller that passes flags its
> traffic
>   * does not match gets a mapping the fabric will not deliver.
> + *
> + * Only a client whose driver called
> pcim_p2pdma_set_ats_per_mapping() takes
> + * the Address Type from @tlp_flags. For any other client its ATS
> Enable bit
> + * decides, and a client that translates every address it is handed
> never gets
> + * %PCI_P2PDMA_MAP_BUS_ADDR.
>   */
>  enum pci_p2pdma_map_type
>  pci_p2pdma_map_type_tlp(struct p2pdma_provider *provider, struct
> device *dev,
> @@ -1620,6 +1660,7 @@ pci_p2pdma_map_type_tlp(struct p2pdma_provider
> *provider, struct device *dev,
>  	struct pci_p2pdma *p2pdma;
>  	struct pci_dev *client;
>  	bool provider_state;
> +	bool per_mapping;
>  	int dist;
>  
>  	if (!pdev->p2pdma) {
> @@ -1639,6 +1680,9 @@ pci_p2pdma_map_type_tlp(struct p2pdma_provider
> *provider, struct device *dev,
>  	cache_index = map_types_idx(client);
>  
>  	rcu_read_lock();
> +	/* The declaration belongs to the client, the cache to the
> provider. */
> +	p2pdma = rcu_dereference(client->p2pdma);
> +	per_mapping = p2pdma && p2pdma->ats_per_mapping;
>  	p2pdma = rcu_dereference(pdev->p2pdma);
>  
>  	if (p2pdma)
> @@ -1646,6 +1690,8 @@ pci_p2pdma_map_type_tlp(struct p2pdma_provider
> *provider, struct device *dev,
>  					     cache_index));
>  	provider_state = !!p2pdma;
>  	rcu_read_unlock();
> +	if (!per_mapping)
> +		tlp_flags = pci_p2pdma_client_tlp_flags(client,
> tlp_flags);
>  	type = pci_p2pdma_map_types_unpack(cached, tlp_flags);
>  	pci_dbg(client,
>  		 "P2PDMA ACS: map lookup provider=%s index=%#lx tlp-
> flags=%#x cached=%s(%d) provider-state=%u\n",
> @@ -1653,10 +1699,10 @@ pci_p2pdma_map_type_tlp(struct
> p2pdma_provider *provider, struct device *dev,
>  		 pci_p2pdma_map_type_name(type), type,
> provider_state);
>  
>  	if (type == PCI_P2PDMA_MAP_UNKNOWN)
> -		return calc_map_type_and_dist(pdev, client, &dist,
> tlp_flags,
> +		type = calc_map_type_and_dist(pdev, client, &dist,
> tlp_flags,
>  					      true);
>  
> -	return type;
> +	return pci_p2pdma_client_map_type(pdev, client, per_mapping,
> type);
>  }
>  
>  void __pci_p2pdma_update_state(struct pci_p2pdma_map_state *state,
> diff --git a/include/linux/pci-p2pdma.h b/include/linux/pci-p2pdma.h
> index 543a09426a11..8d96659df9df 100644
> --- a/include/linux/pci-p2pdma.h
> +++ b/include/linux/pci-p2pdma.h
> @@ -207,7 +207,8 @@ static inline struct pci_dev
> *pci_p2pmem_find(struct device *client)
>   * @dev: Client device that initiates the transfer
>   *
>   * Same as pci_p2pdma_map_type_tlp() for a client issuing strictly
> ordered
> - * Requests that carry an Untranslated address.
> + * Requests. Their Address Type is Untranslated unless the client
> enables ATS
> + * for the whole device.
>   */
>  static inline enum pci_p2pdma_map_type
>  pci_p2pdma_map_type(struct p2pdma_provider *provider, struct device
> *dev)