On Tue, Feb 14, 2017 at 10:14 AM, Alex Williamson
<[email protected]> wrote:
> On Sat, 30 Jan 2016 01:33:58 +0530
> Manish Jaggi <[email protected]> wrote:
>
>> 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.
>>
>> Acked-by: Tirumalesh Chalamarla <[email protected]>
>> Signed-off-by: Manish Jaggi <[email protected]>
>> ---
>> drivers/pci/quirks.c | 15 +++++++++++++++
>> 1 file changed, 15 insertions(+)
>>
>> diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
>> index 7e32730..a300fa6 100644
>> --- a/drivers/pci/quirks.c
>> +++ b/drivers/pci/quirks.c
>> @@ -3814,6 +3814,19 @@ 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.
>> + */
>> + acs_flags &= ~(PCI_ACS_SV | PCI_ACS_TB | PCI_ACS_RR |
>> + PCI_ACS_CR | PCI_ACS_UF | PCI_ACS_DT);
>> +
>> + return acs_flags ? 0 : 1;
>> +}
>> +
>> /*
>> * Many Intel PCH root ports do provide ACS-like features to disable peer
>> * transactions and validate bus numbers in requests, but do not provide an
>> @@ -3966,6 +3979,8 @@ static const struct pci_dev_acs_enabled {
>> { PCI_VENDOR_ID_INTEL, PCI_ANY_ID, pci_quirk_intel_pch_acs },
>> { 0x19a2, 0x710, pci_quirk_mf_endpoint_acs }, /* Emulex BE3-R */
>> { 0x10df, 0x720, pci_quirk_mf_endpoint_acs }, /* Emulex Skyhawk-R */
>> + /* Cavium ThunderX */
>> + { PCI_VENDOR_ID_CAVIUM, PCI_ANY_ID, pci_quirk_cavium_acs },
>> { 0 }
>> };
>>
>
> Apologies for not catching this, but what sort of crystal ball do you
> have that can predict not only current devices, but future devices will
> not support peer-to-peer features? Is there an internal design
> guidelines reference specification for Cavium that we can realistically
> expect this to remain consistent, or is this just an attempt to never
> think about ACS again at the customer's peril? What about the existing
> non-ThunderX products with Cavium vendor ID, does this really apply to
> those? I would strongly suggest taking the device ID into account.
> See examples like the pci_quirk_intel_pch_acs quirk where the initial
> filter is PCI_ANY_ID, but specific device types and ranges of device
> IDs are identified by the function for evaluation. This seems reckless
> to me and I'd advise that it be reverted. Thanks,
>
> Alex
Just a thought, even if Cavium considers to support ACS for future devices,
wouldn't it be better to add exception list inside the quirk on a need basis
rather than adding big list of devices that don't. Especially when currently
almost all Cavium PCI devices don't support ACS.
Thanks,
Sunil.
On Sun, 5 Mar 2017 12:37:31 +0530
Sunil Kovvuri <[email protected]> wrote:
> On Tue, Feb 14, 2017 at 10:14 AM, Alex Williamson
> <[email protected]> wrote:
> > On Sat, 30 Jan 2016 01:33:58 +0530
> > Manish Jaggi <[email protected]> wrote:
> >
> >> 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.
> >>
> >> Acked-by: Tirumalesh Chalamarla <[email protected]>
> >> Signed-off-by: Manish Jaggi <[email protected]>
> >> ---
> >> drivers/pci/quirks.c | 15 +++++++++++++++
> >> 1 file changed, 15 insertions(+)
> >>
> >> diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
> >> index 7e32730..a300fa6 100644
> >> --- a/drivers/pci/quirks.c
> >> +++ b/drivers/pci/quirks.c
> >> @@ -3814,6 +3814,19 @@ 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.
> >> + */
> >> + acs_flags &= ~(PCI_ACS_SV | PCI_ACS_TB | PCI_ACS_RR |
> >> + PCI_ACS_CR | PCI_ACS_UF | PCI_ACS_DT);
> >> +
> >> + return acs_flags ? 0 : 1;
> >> +}
> >> +
> >> /*
> >> * Many Intel PCH root ports do provide ACS-like features to disable peer
> >> * transactions and validate bus numbers in requests, but do not provide an
> >> @@ -3966,6 +3979,8 @@ static const struct pci_dev_acs_enabled {
> >> { PCI_VENDOR_ID_INTEL, PCI_ANY_ID, pci_quirk_intel_pch_acs },
> >> { 0x19a2, 0x710, pci_quirk_mf_endpoint_acs }, /* Emulex BE3-R */
> >> { 0x10df, 0x720, pci_quirk_mf_endpoint_acs }, /* Emulex Skyhawk-R */
> >> + /* Cavium ThunderX */
> >> + { PCI_VENDOR_ID_CAVIUM, PCI_ANY_ID, pci_quirk_cavium_acs },
> >> { 0 }
> >> };
> >>
> >
> > Apologies for not catching this, but what sort of crystal ball do you
> > have that can predict not only current devices, but future devices will
> > not support peer-to-peer features? Is there an internal design
> > guidelines reference specification for Cavium that we can realistically
> > expect this to remain consistent, or is this just an attempt to never
> > think about ACS again at the customer's peril? What about the existing
> > non-ThunderX products with Cavium vendor ID, does this really apply to
> > those? I would strongly suggest taking the device ID into account.
> > See examples like the pci_quirk_intel_pch_acs quirk where the initial
> > filter is PCI_ANY_ID, but specific device types and ranges of device
> > IDs are identified by the function for evaluation. This seems reckless
> > to me and I'd advise that it be reverted. Thanks,
> >
> > Alex
>
> Just a thought, even if Cavium considers to support ACS for future devices,
> wouldn't it be better to add exception list inside the quirk on a need basis
> rather than adding big list of devices that don't. Especially when currently
> almost all Cavium PCI devices don't support ACS.
The same argument can be made the other way, which is the more
concerning problem. What if Cavium releases a device that does not
support ACS nor has ACS equivalent isolation? This quirk would assume
isolation exists. Now think about the compatibility problem of trying
to negate this quirk for that device on any kernel that has shipped
with this open-ended quirk. What's the support situation if a user
relies on that isolation that maybe doesn't exist? Any solution that
assumes unknown future devices have isolation is dangerous. Perhaps
it's justifiable if the company has strong internal design guidelines
about this sort of thing, but even those get lost or change over time
without necessarily evaluating all the implications. IMO, the only
workable solution is to white-list *only* the devices where we know ACS
equivalent isolation is present. We're still waiting on Cavium's
patch to make that happen. Thanks,
Alex