Re: [PATCH v4] PCI: quirks: update Cavium ThunderX ACS quirk implementation

From: Alex Williamson
Date: Wed Sep 20 2017 - 12:31:58 EST


On Mon, 18 Sep 2017 01:48:01 -0700
Vadim Lomovtsev <Vadim.Lomovtsev@xxxxxxxxxxxxxxxxxx> wrote:

> This commit makes Cavium PCI ACS quirk applicable only to Cavium
> ThunderX (CN81/83/88XX) PCIE Root Ports which has limited PCI capabilities
> in terms of no ACS support advertisement. However, the RTL internally
> implements similar protection as if ACS had completion/request redirection,
> upstream forwarding and validation features enabled.
>
> Current quirk implementation doesn't take into account PCIERCs which
> also needs to be quirked. So the pci device id check mask is updated
> and check of device ID moved into separate function.
>
> Signed-off-by: Vadim Lomovtsev <Vadim.Lomovtsev@xxxxxxxxxxxxxxxxxx>
> ---
> v1 : put device check into separate function and extend it to all
> Cavium PCIERC/PCCBR devices;
> v1 -> v2: update match function in order to filter only ThunderX devices by device
> ids to properly filter CN8XXX devices, update subject & description with
> ACS register info (rejected by maillist due to triple X in subject);
> v2 -> v3: update subject: remove CN8XXX from subject line, replace it with ThunderX;
> v3 -> v4: update ACS mask (remove TB and TD bits), update commit message (remove
> ACS register printout);
>
> drivers/pci/quirks.c | 26 ++++++++++++++++----------
> 1 file changed, 16 insertions(+), 10 deletions(-)
>
> diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
> index a4d3361..e6b904a 100644
> --- a/drivers/pci/quirks.c
> +++ b/drivers/pci/quirks.c
> @@ -4211,20 +4211,26 @@ static int pci_quirk_amd_sb_acs(struct pci_dev *dev, u16 acs_flags)
> #endif
> }
>
> -static int pci_quirk_cavium_acs(struct pci_dev *dev, u16 acs_flags)
> +/*
> + * Cavium devices matching this quirk do not perform peer-to-peer
> + * with other functions, allowing masking out these bits as if they
> + * were unimplemented in the ACS capability.

nit, the description here still steals too much from the multifunction
quirk. Multifunction devices can often support ACS with unimplemented
capabilities, which indicate that the device does not support the
behavior described by that capability bit. However, downstream ports
are required to implement certain ACS capabilities if they implement
ACS at all. So the code is actually asserting that the hardware
implements *and* enables equivalent ACS functionality for these flags.

> + */
> +#define CAVIUM_CN8XXX_ACS_FLAGS (PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_SV | PCI_ACS_UF)
> +
> +static __inline__ bool pci_quirk_cavium_acs_match(struct pci_dev *dev)
> {
> - /*
> - * Cavium devices matching this quirk do not perform peer-to-peer
> - * with other functions, allowing masking out these bits as if they
> - * were unimplemented in the ACS capability.
> - */
> - acs_flags &= ~(PCI_ACS_SV | PCI_ACS_TB | PCI_ACS_RR |
> - PCI_ACS_CR | PCI_ACS_UF | PCI_ACS_DT);
> + return (pci_is_pcie(dev) &&
> + (pci_pcie_type(dev) == PCI_EXP_TYPE_ROOT_PORT) &&
> + ((dev->device & 0xf800) == 0xa000));

That's effectively 2k device IDs, 0xa000-0xa7ff that you and Cavium are
vouching for ACS equivalent isolation. How many of these actually
exist? The PCI IDs database gets really sparse after the first 64
entries. Internally are these device IDs allocated to programs based on
the same ASICs or is this just a slightly more restricted crystal ball
(ie. wishful thinking)? Thanks,

Alex

> +}
>
> - if (!((dev->device >= 0xa000) && (dev->device <= 0xa0ff)))
> +static int pci_quirk_cavium_acs(struct pci_dev *dev, u16 acs_flags)
> +{
> + if (!pci_quirk_cavium_acs_match(dev))
> return -ENOTTY;
>
> - return acs_flags ? 0 : 1;
> + return acs_flags & ~(CAVIUM_CN8XXX_ACS_FLAGS) ? 0 : 1;
> }
>
> static int pci_quirk_xgene_acs(struct pci_dev *dev, u16 acs_flags)