Re: [PATCH v4 1/2] PCI: AtomicOps: Do not enable if root-port capabilities are unknown

[Gerd Bayer <gbayer@linux.ibm.com>]

On Fri, 2026-03-13 at 17:49 +0100, Gerd Bayer wrote:
> When inspecting the config space of a Connect-X physical function in an
> s390 system after it was initialized by the mlx5_core device driver, we
> found the function to be enabled to request AtomicOps despite the
> system's root-complex lacking support for completing them:
> 
> 1ed0:00:00.1 Ethernet controller: Mellanox Technologies MT2894 Family [ConnectX-6 Lx]
> 	Subsystem: Mellanox Technologies Device 0002
>   [...]
> 	DevCtl2: Completion Timeout: 50us to 50ms, TimeoutDis-
> 		 AtomicOpsCtl: ReqEn+
> 		 IDOReq- IDOCompl- LTR- EmergencyPowerReductionReq-
> 		 10BitTagReq- OBFF Disabled, EETLPPrefixBlk-
> 
> Turns out the device driver calls pci_enable_atomic_ops_to_root() which
> defaulted to enable AtomicOps requests even if it had no information
> about the root-port that the PCIe device is attached to.
> 
> Change to logic of pci_enable_atomic_ops_to_root() to fully traverse the
> PCIe tree upwards, check that the bridge devices support delivering
> AtomicOps transactions, and finally check that there is a root-port at
> the end that does support completing AtomicOps.
> 
> Do not enable AtomicOps requests if nothing can be learned about how the
> device is attached - e.g. if it is on an "isolated" bus, as in s390.
> 
> Reported-by: Alexander Schmidt <alexs@linux.ibm.com>
> Cc: stable@vger.kernel.org
> Fixes: 430a23689dea ("PCI: Add pci_enable_atomic_ops_to_root()")
> Signed-off-by: Gerd Bayer <gbayer@linux.ibm.com>
> ---
>  drivers/pci/pci.c | 30 ++++++++++++++----------------
>  1 file changed, 14 insertions(+), 16 deletions(-)
> 
> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> index 8479c2e1f74f1044416281aba11bf071ea89488a..94e90988df86b3278b1b6abbc326abf9b4a4a962 100644
> --- a/drivers/pci/pci.c
> +++ b/drivers/pci/pci.c
> @@ -3676,7 +3676,7 @@ void pci_acs_init(struct pci_dev *dev)
>  int pci_enable_atomic_ops_to_root(struct pci_dev *dev, u32 cap_mask)
>  {
>  	struct pci_bus *bus = dev->bus;
> -	struct pci_dev *bridge;
> +	struct pci_dev *bridge = NULL;
>  	u32 cap, ctl2;
>  
>  	/*
> @@ -3714,29 +3714,27 @@ int pci_enable_atomic_ops_to_root(struct pci_dev *dev, u32 cap_mask)
>  		switch (pci_pcie_type(bridge)) {
>  		/* Ensure switch ports support AtomicOp routing */
>  		case PCI_EXP_TYPE_UPSTREAM:
> -		case PCI_EXP_TYPE_DOWNSTREAM:
> -			if (!(cap & PCI_EXP_DEVCAP2_ATOMIC_ROUTE))
> -				return -EINVAL;
> -			break;
> -
> -		/* Ensure root port supports all the sizes we care about */
> -		case PCI_EXP_TYPE_ROOT_PORT:
> -			if ((cap & cap_mask) != cap_mask)
> -				return -EINVAL;
> -			break;
> -		}
> -
> -		/* Ensure upstream ports don't block AtomicOps on egress */
> -		if (pci_pcie_type(bridge) == PCI_EXP_TYPE_UPSTREAM) {
> +			/* Upstream ports must not block AtomicOps on egress */
>  			pcie_capability_read_dword(bridge, PCI_EXP_DEVCTL2,
>  						   &ctl2);
>  			if (ctl2 & PCI_EXP_DEVCTL2_ATOMIC_EGRESS_BLOCK)
>  				return -EINVAL;
> +			fallthrough;
> +		/* All switch ports need to route AtomicOps */
> +		case PCI_EXP_TYPE_DOWNSTREAM:
> +			if (!(cap & PCI_EXP_DEVCAP2_ATOMIC_ROUTE))
> +				return -EINVAL;
> +			break;
>  		}
> -
>  		bus = bus->parent;
>  	}
>  
> +	/* Finally, last bridge must be root port and support requested sizes */
> +	if ((!bridge) ||
> +	    (pci_pcie_type(bridge) != PCI_EXP_TYPE_ROOT_PORT) ||
> +	    ((cap & cap_mask) != cap_mask))
> +		return -EINVAL;
> +

Sashiko (the new review tool) annotated that this patch may regress how
RCiEP's are treated:
https://sashiko.dev/#/patchset/20260313-fix_pciatops-v4-0-93bc70a63935%40linux.ibm.com

And it has a point: While AtomicOps Requests were allowed for RCiEP's
per default - now they're forbidden per default. Turns out that the
situation for RCiEP's on all archs is the same as for "isolated
functions" on s390: With pure PCI config ops you cannot tell, whether
the root-complex supports them or not.

I'm thinking about inverting the logic here - and adding an else clause
with an arch-specific hook to serve as discriminator.
Thoughts anybody?

>  	pcie_capability_set_word(dev, PCI_EXP_DEVCTL2,
>  				 PCI_EXP_DEVCTL2_ATOMIC_REQ);
>  	return 0;

Thanks,
Gerd