Re: [PATCH v7 18/23] iommu/smmuv3: Report non recoverable faults

[Robin Murphy <robin.murphy@arm.com>]

On 13/05/2019 13:32, Auger Eric wrote:
> Hi Robin,
> On 5/13/19 1:54 PM, Robin Murphy wrote:
>> On 13/05/2019 08:46, Auger Eric wrote:
>>> Hi Robin,
>>>
>>> On 5/8/19 7:20 PM, Robin Murphy wrote:
>>>> On 08/04/2019 13:19, Eric Auger wrote:
>>>>> When a stage 1 related fault event is read from the event queue,
>>>>> let's propagate it to potential external fault listeners, ie. users
>>>>> who registered a fault handler.
>>>>>
>>>>> Signed-off-by: Eric Auger <eric.auger@redhat.com>
>>>>>
>>>>> ---
>>>>> v4 -> v5:
>>>>> - s/IOMMU_FAULT_PERM_INST/IOMMU_FAULT_PERM_EXEC
>>>>> ---
>>>>>     drivers/iommu/arm-smmu-v3.c | 169
>>>>> +++++++++++++++++++++++++++++++++---
>>>>>     1 file changed, 158 insertions(+), 11 deletions(-)
>>>>>
>>>>> diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c
>>>>> index 8044445bc32a..1fd320788dcb 100644
>>>>> --- a/drivers/iommu/arm-smmu-v3.c
>>>>> +++ b/drivers/iommu/arm-smmu-v3.c
>>>>> @@ -167,6 +167,26 @@
>>>>>     #define ARM_SMMU_PRIQ_IRQ_CFG1        0xd8
>>>>>     #define ARM_SMMU_PRIQ_IRQ_CFG2        0xdc
>>>>>     +/* Events */
>>>>> +#define ARM_SMMU_EVT_F_UUT        0x01
>>>>> +#define ARM_SMMU_EVT_C_BAD_STREAMID    0x02
>>>>> +#define ARM_SMMU_EVT_F_STE_FETCH    0x03
>>>>> +#define ARM_SMMU_EVT_C_BAD_STE        0x04
>>>>> +#define ARM_SMMU_EVT_F_BAD_ATS_TREQ    0x05
>>>>> +#define ARM_SMMU_EVT_F_STREAM_DISABLED    0x06
>>>>> +#define ARM_SMMU_EVT_F_TRANSL_FORBIDDEN    0x07
>>>>> +#define ARM_SMMU_EVT_C_BAD_SUBSTREAMID    0x08
>>>>> +#define ARM_SMMU_EVT_F_CD_FETCH        0x09
>>>>> +#define ARM_SMMU_EVT_C_BAD_CD        0x0a
>>>>> +#define ARM_SMMU_EVT_F_WALK_EABT    0x0b
>>>>> +#define ARM_SMMU_EVT_F_TRANSLATION    0x10
>>>>> +#define ARM_SMMU_EVT_F_ADDR_SIZE    0x11
>>>>> +#define ARM_SMMU_EVT_F_ACCESS        0x12
>>>>> +#define ARM_SMMU_EVT_F_PERMISSION    0x13
>>>>> +#define ARM_SMMU_EVT_F_TLB_CONFLICT    0x20
>>>>> +#define ARM_SMMU_EVT_F_CFG_CONFLICT    0x21
>>>>> +#define ARM_SMMU_EVT_E_PAGE_REQUEST    0x24
>>>>> +
>>>>>     /* Common MSI config fields */
>>>>>     #define MSI_CFG0_ADDR_MASK        GENMASK_ULL(51, 2)
>>>>>     #define MSI_CFG2_SH            GENMASK(5, 4)
>>>>> @@ -332,6 +352,15 @@
>>>>>     #define EVTQ_MAX_SZ_SHIFT        7
>>>>>       #define EVTQ_0_ID            GENMASK_ULL(7, 0)
>>>>> +#define EVTQ_0_SSV            GENMASK_ULL(11, 11)
>>>>> +#define EVTQ_0_SUBSTREAMID        GENMASK_ULL(31, 12)
>>>>> +#define EVTQ_0_STREAMID            GENMASK_ULL(63, 32)
>>>>> +#define EVTQ_1_PNU            GENMASK_ULL(33, 33)
>>>>> +#define EVTQ_1_IND            GENMASK_ULL(34, 34)
>>>>> +#define EVTQ_1_RNW            GENMASK_ULL(35, 35)
>>>>> +#define EVTQ_1_S2            GENMASK_ULL(39, 39)
>>>>> +#define EVTQ_1_CLASS            GENMASK_ULL(40, 41)
>>>>> +#define EVTQ_3_FETCH_ADDR        GENMASK_ULL(51, 3)
>>>>>       /* PRI queue */
>>>>>     #define PRIQ_ENT_DWORDS            2
>>>>> @@ -639,6 +668,64 @@ struct arm_smmu_domain {
>>>>>         spinlock_t            devices_lock;
>>>>>     };
>>>>>     +/* fault propagation */
>>>>> +
>>>>> +#define IOMMU_FAULT_F_FIELDS    (IOMMU_FAULT_UNRECOV_PASID_VALID | \
>>>>> +                 IOMMU_FAULT_UNRECOV_PERM_VALID | \
>>>>> +                 IOMMU_FAULT_UNRECOV_ADDR_VALID)
>>>>> +
>>>>> +struct arm_smmu_fault_propagation_data {
>>>>> +    enum iommu_fault_reason reason;
>>>>> +    bool s1_check;
>>>>> +    u32 fields; /* IOMMU_FAULT_UNRECOV_*_VALID bits */
>>>>> +};
>>>>> +
>>>>> +/*
>>>>> + * Describes how SMMU faults translate into generic IOMMU faults
>>>>> + * and if they need to be reported externally
>>>>> + */
>>>>> +static const struct arm_smmu_fault_propagation_data
>>>>> fault_propagation[] = {
>>>>> +[ARM_SMMU_EVT_F_UUT]            = { },
>>>>> +[ARM_SMMU_EVT_C_BAD_STREAMID]        = { },
>>>>> +[ARM_SMMU_EVT_F_STE_FETCH]        = { },
>>>>> +[ARM_SMMU_EVT_C_BAD_STE]        = { },
>>>>> +[ARM_SMMU_EVT_F_BAD_ATS_TREQ]        = { },
>>>>> +[ARM_SMMU_EVT_F_STREAM_DISABLED]    = { },
>>>>> +[ARM_SMMU_EVT_F_TRANSL_FORBIDDEN]    = { },
>>>>> +[ARM_SMMU_EVT_C_BAD_SUBSTREAMID]    =
>>>>> {IOMMU_FAULT_REASON_PASID_INVALID,
>>>>> +                       false,
>>>>> +                       IOMMU_FAULT_UNRECOV_PASID_VALID
>>>>> +                      },
>>>>> +[ARM_SMMU_EVT_F_CD_FETCH]        = {IOMMU_FAULT_REASON_PASID_FETCH,
>>>>> +                       false,
>>>>> +                       IOMMU_FAULT_UNRECOV_PASID_VALID |
>>>>
>>>> It doesn't make sense to presume validity here, or in any of the faults
>>>> below...
>>>
>>>
>>>>
>>>>> +                       IOMMU_FAULT_UNRECOV_FETCH_ADDR_VALID
>>>>> +                      },
>>>>> +[ARM_SMMU_EVT_C_BAD_CD]            =
>>>>> {IOMMU_FAULT_REASON_BAD_PASID_ENTRY,
>>>>> +                       false,
>>>>> +                       IOMMU_FAULT_UNRECOV_PASID_VALID
>>>>> +                      },
>>>>> +[ARM_SMMU_EVT_F_WALK_EABT]        = {IOMMU_FAULT_REASON_WALK_EABT,
>>>>> true,
>>>>> +                       IOMMU_FAULT_F_FIELDS |
>>>>> +                       IOMMU_FAULT_UNRECOV_FETCH_ADDR_VALID
>>>>> +                      },
>>>>> +[ARM_SMMU_EVT_F_TRANSLATION]        = {IOMMU_FAULT_REASON_PTE_FETCH,
>>>>> true,
>>>>> +                       IOMMU_FAULT_F_FIELDS
>>>>> +                      },
>>>>> +[ARM_SMMU_EVT_F_ADDR_SIZE]        = {IOMMU_FAULT_REASON_OOR_ADDRESS,
>>>>> true,
>>>>> +                       IOMMU_FAULT_F_FIELDS
>>>>> +                      },
>>>>> +[ARM_SMMU_EVT_F_ACCESS]            = {IOMMU_FAULT_REASON_ACCESS, true,
>>>>> +                       IOMMU_FAULT_F_FIELDS
>>>>> +                      },
>>>>> +[ARM_SMMU_EVT_F_PERMISSION]        = {IOMMU_FAULT_REASON_PERMISSION,
>>>>> true,
>>>>> +                       IOMMU_FAULT_F_FIELDS
>>>>> +                      },
>>>>> +[ARM_SMMU_EVT_F_TLB_CONFLICT]        = { },
>>>>> +[ARM_SMMU_EVT_F_CFG_CONFLICT]        = { },
>>>>> +[ARM_SMMU_EVT_E_PAGE_REQUEST]        = { },
>>>>> +};
>>>>> +
>>>>>     struct arm_smmu_option_prop {
>>>>>         u32 opt;
>>>>>         const char *prop;
>>>>> @@ -1258,7 +1345,6 @@ static int arm_smmu_init_l2_strtab(struct
>>>>> arm_smmu_device *smmu, u32 sid)
>>>>>         return 0;
>>>>>     }
>>>>>     -__maybe_unused
>>>>>     static struct arm_smmu_master_data *
>>>>>     arm_smmu_find_master(struct arm_smmu_device *smmu, u32 sid)
>>>>>     {
>>>>> @@ -1284,24 +1370,85 @@ arm_smmu_find_master(struct arm_smmu_device
>>>>> *smmu, u32 sid)
>>>>>         return master;
>>>>>     }
>>>>>     +/* Populates the record fields according to the input SMMU event */
>>>>> +static bool arm_smmu_transcode_fault(u64 *evt, u8 type,
>>>>> +                     struct iommu_fault_unrecoverable *record)
>>>>> +{
>>>>> +    const struct arm_smmu_fault_propagation_data *data;
>>>>> +    u32 fields;
>>>>> +
>>>>> +    if (type >= ARRAY_SIZE(fault_propagation))
>>>>> +        return false;
>>>>> +
>>>>> +    data = &fault_propagation[type];
>>>>> +    if (!data->reason)
>>>>> +        return false;
>>>>> +
>>>>> +    fields = data->fields;
>>>>> +
>>>>> +    if (data->s1_check & FIELD_GET(EVTQ_1_S2, evt[1]))
>>>>> +        return false; /* S2 related fault, don't propagate */
>>>>> +
>>>>> +    if (fields & IOMMU_FAULT_UNRECOV_PASID_VALID) {
>>>>> +        if (FIELD_GET(EVTQ_0_SSV, evt[0]))
>>>>> +            record->pasid = FIELD_GET(EVTQ_0_SUBSTREAMID, evt[0]);
>>>>> +        else
>>>>> +            fields &= ~IOMMU_FAULT_UNRECOV_PASID_VALID;
>>>>
>>>> ...because this logic then breaks for C_BAD_SUBSTREAMID, which ends up
>>>> coming out of here *without* reporting the offending PASID.
>>> Correct.
>>>>
>>>>> +    }
>>>>> +    if (fields & IOMMU_FAULT_UNRECOV_PERM_VALID) {
>>>>> +        if (!FIELD_GET(EVTQ_1_RNW, evt[1]))
>>>>> +            record->perm |= IOMMU_FAULT_PERM_WRITE;
>>>>> +        if (FIELD_GET(EVTQ_1_PNU, evt[1]))
>>>>> +            record->perm |= IOMMU_FAULT_PERM_PRIV;
>>>>> +        if (FIELD_GET(EVTQ_1_IND, evt[1]))
>>>>> +            record->perm |= IOMMU_FAULT_PERM_EXEC;
>>>>> +    }
>>>>> +    if (fields & IOMMU_FAULT_UNRECOV_ADDR_VALID)
>>>>> +        record->addr = evt[2];
>>>>> +
>>>>> +    if (fields & IOMMU_FAULT_UNRECOV_FETCH_ADDR_VALID)
>>>>> +        record->fetch_addr = FIELD_GET(EVTQ_3_FETCH_ADDR, evt[3]);
>>>>> +
>>>>> +    record->flags = fields;
>>>>> +    return true;
>>>>> +}
>>>>> +
>>>>> +static void arm_smmu_report_event(struct arm_smmu_device *smmu, u64
>>>>> *evt)
>>>>> +{
>>>>> +    u32 sid = FIELD_GET(EVTQ_0_STREAMID, evt[0]);
>>>>> +    u8 type = FIELD_GET(EVTQ_0_ID, evt[0]);
>>>>> +    struct arm_smmu_master_data *master;
>>>>> +    struct iommu_fault_event event = {};
>>>>> +    int i;
>>>>> +
>>>>> +    master = arm_smmu_find_master(smmu, sid);
>>>>> +    if (WARN_ON(!master))
>>>>> +        return;
>>>>
>>>> NAK. If I'm getting global faults like C_BAD_STE where a device almost
>>>> certainly *isn't* configured (because hey, we would have initialised its
>>>> STEs if we knew), then I sure as hell want to see the actual faults.
>>>> Spamming a constant stream of stack traces *instead* of showing them is
>>>> worse than useless.
>>> Sure, if !master I will output the original traces.
>>>>
>>>>> +
>>>>> +    event.fault.type = IOMMU_FAULT_DMA_UNRECOV;
>>>>> +
>>>>> +    if (arm_smmu_transcode_fault(evt, type, &event.fault.event)) {
>>>>> +        iommu_report_device_fault(master->dev, &event);
>>>>> +        return;
>>>>
>>>> And again, the vast majority of the time, there won't be a fault handler
>>>> registered, so unconditionally suppressing the most common and useful
>>>> stuff like translation and permission faults is very much not OK.
>>> Going to test whether we are in nested mode before entering that path.
>>
>> I don't think this has to be exclusive to nesting - the generic
>> reporting mechanism feels like it might ultimately be extensible to
>> other things like Rob's case for generalised stalling. It's just that
>> for robustness, even when a fault handler is present, we still want the
>> driver to be able to report if it didn't actually handle a fault.
> 
> 
> Jean-Philippe pointed out in a previous review
> (https://patchwork.kernel.org/patch/10751801/#22424047) that the guest
> can flood the host log with S1 related faults. At the moment we do not
> check that a fault handler is registered in nested mode. Maybe we
> should? Even if the fault handler is registered, as it is based on a
> circular buffer, this latter can be full and lead to a log flood.

Ah, right, I didn't quite have the full picture in mind, thanks for the 
jog. I guess in that case we want some degree of special-casing for 
nested configs - anything S1-related that we expect the fault handler to 
simply inject back into the guest, we can indeed suppress on the host, 
but I do think it's worth anticipating host-related faults (and entirely 
host-owned fault-handlers in future) as well.

Robin.
_______________________________________________
kvmarm mailing list
kvmarm@lists.cs.columbia.edu
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm