[RFC 01/13] KVM: arm/arm64: Add vITS save/restore API documentation

[Auger Eric]

Hi Marc,

On 13/01/2017 10:46, Marc Zyngier wrote:
> On 13/01/17 09:07, Auger Eric wrote:
>> Hi Marc,
>>
>> On 12/01/2017 17:52, Marc Zyngier wrote:
>>> Hi Eric,
>>>
>>> On 12/01/17 15:56, Eric Auger wrote:
>>>> Add description for how to access vITS registers and how to flush/restore
>>>> vITS tables into/from memory
>>>>
>>>> Signed-off-by: Eric Auger <eric.auger at redhat.com>
>>>> ---
>>>>  Documentation/virtual/kvm/devices/arm-vgic-its.txt | 70 ++++++++++++++++++++++
>>>>  1 file changed, 70 insertions(+)
>>>>
>>>> diff --git a/Documentation/virtual/kvm/devices/arm-vgic-its.txt b/Documentation/virtual/kvm/devices/arm-vgic-its.txt
>>>> index 6081a5b..bd74613 100644
>>>> --- a/Documentation/virtual/kvm/devices/arm-vgic-its.txt
>>>> +++ b/Documentation/virtual/kvm/devices/arm-vgic-its.txt
>>>> @@ -36,3 +36,73 @@ Groups:
>>>>      -ENXIO:  ITS not properly configured as required prior to setting
>>>>               this attribute
>>>>      -ENOMEM: Memory shortage when allocating ITS internal data
>>>> +
>>>> +  KVM_DEV_ARM_VGIC_GRP_ITS_REGS
>>>> +  Attributes:
>>>> +      The attr field of kvm_device_attr encodes the offset of the
>>>> +      ITS register, relative to the ITS control frame base address
>>>> +      (ITS_base).
>>>> +
>>>> +      kvm_device_attr.addr points to a __u64 value whatever the width
>>>> +      of the addressed register (32/64 bits).
>>>> +
>>>> +      Writes to read-only registers are ignored by the kernel except
>>>> +      for a single register, GITS_READR. Normally this register is RO
>>>> +      but it needs to be restored otherwise commands in the queue will
>>>> +      be re-executed after CWRITER setting.
>>>> +
>>>> +      For other registers, Getting or setting a register has the same
>>>> +      effect as reading/writing the register on real hardware.
>>>> +  Errors:
>>>> +    -ENXIO: Offset does not correspond to any supported register
>>>> +    -EFAULT: Invalid user pointer for attr->addr
>>>> +    -EINVAL: Offset is not 64-bit aligned
>>>> +
>>>> +  KVM_DEV_ARM_VGIC_GRP_ITS_TABLES
>>>> +  Attributes
>>>> +       The attr field of kvm_device_attr is not used.
>>>> +
>>>> +       request the flush-save/restore of the ITS tables, namely
>>>> +       the device table, the collection table, all the ITT tables,
>>>> +       the LPI pending tables. On save, the tables are flushed
>>>> +       into guest memory at the location provisionned by the guest
>>>
>>> 					    provisioned
>>>
>>>> +       in GITS_BASER (device and collection tables), on MAPD command
>>>> +       (ITT_addr), GICR_PENDBASERs (pending tables).
>>>> +
>>>> +       This means the GIC should be restored before the ITS and all
>>>> +       ITS registers but the GITS_CTRL must be restored before
>>>> +       restoring the ITS tables.
>>>> +
>>>> +       Note the LPI configuration table is read-only for the
>>>> +       in-kernel ITS and its save/restore goes through the standard
>>>> +       RAM save/restore.
>>>> +
>>>> +       The layout of the tables in guest memory defines an ABI.
>>>> +       The entries are laid out in memory as follows;
>>>> +
>>>> +    Device Table Entry (DTE) layout: entry size = 16 bytes
>>>> +
>>>> +    bits:     | 63   ...  32  | 31 ... 6 | 5 | 4   ...   0 |
>>>> +    values:   |   DeviceID    |   Resv   | V |    Size     |
>>>> +
>>>> +    bits:     | 63 ... 44 | 43  ...  0  |
>>>> +    values:   |    Resv   |  ITT_addr   |
>>>
>>> While I appreciate this layout represents the absolute maximum an ITS
>>> could implement, I'm a bit concerned about the amount of memory we may
>>> end-up requiring here. All the ITSs implementations I know of seem to
>>> get away with 8 bytes per entry. Maybe I'm just too worried.
>>
>> OK so I would propose a 16b DeviceId and 16b eventid
>>
>> bits:     | 63   ...  48  | 47 ... 4 | 3   ...   0 |
>> values:   |   DeviceID    | ITT_addr |    Size     |
>>
>> I can use the size field as a validity indicator
> 
> Note that you are allowed to use a 0 size field. It means 1 bit of
> EventID (2 possible interrupts). So maybe using a particular address as
> a valid flag?
Is it really acceptable to encode the deviceId and eventid on 16 bits
instead of 32 bits max each?

Currently I do not use the deviceId indexing, ie. the device id is
directly encoded in the entry. The spec rather suggests device id
indexing in flat table and this is also stems from 2 stage table support.

So I have 2 strategies:
- ignore the device id indexing and store valid data at the beginning of
available buffers (pros: no sparsity, cons: shrinks device and eventid
to 16 bits). Natural in flat mode, less natural in 2 stage mode.
- implement device id indexing (pros: keep the full range for deviceid
and eventid, cons: sparsity). Then sparsity needs to be handled somehow.
Now I better understand your remark on first kB of the pending table...


> 
>>
>>>
>>> Also, please mention that ITT_addr is actually ITT_addr[51:8], as we're
>>> guaranteed to have an ITT that is 256 byte aligned.
>> sure
>>>
>>>> +
>>>> +    Collection Table Entry (CTE) layout: entry size = 8 bytes
>>>> +
>>>> +    bits:     | 63| 62 ..  52  | 51 ... 16 | 15  ...   0 |
>>>> +    values:   | V |    RES0    |  RDBase   |    ICID     |
>>>> +
>>>> +    Interrupt Translation Table Entry (ITTE) layout: entry size = 16 bytes
>>>
>>> The actual name is Interrupt Translation Entry (ITE). I have a patch
>>> renaming this all over the vgic-its.c file...
>> ok
>>>
>>>> +
>>>> +    bits:     | 63   ...  32  | 31 ... 17 | 16 | 15  ...  0 |
>>>> +    values:   |   DeviceID    |    RES0   | V  |    ICID    |
>>>> +
>>>> +    bits:     | 63 ...  32    | 31      ...        0 |
>>>> +    values:   |   pINTID      |        EventID       |
>>>
>>> Same concern here. 32bit DevID, EventID and INTID seem a bit over the
>>> top. But maybe we shouldn't be concerned about memory... ;-)
>> So I would suggest encoding
>> 16b DeviceId
>> 16b eventid
>> 16b collection ID
>> 16b pINTID
>>
>> bits:     | 63   ...  48  | 47 ... 32 | 31 ... 15 | 15  ...  0 |
>> values:   |   DeviceID    |   pINTID  |  EventId  |   ICID     |
>>
>> a null pINTID would meen the ITE is invalid.
>>
>> Does that make sense or should I instead reduce the number of bits
>> allocated to collections and keep the pINTID bit number larger?
> 
> 16bit worth of collections is quite a lot (64k CPUs?). I'd be perfectly
> fine with a smaller number, but let's see what people think.
This is useless to store the deviceId here since the deviceId is known
from the upper level device table. I will fix that in v2. But anyway if
I encode the ITE on 8 bytes I must shrink the pINTID/EventId compared to
their max size (32b). If EventId is encoded on 16b then I guess the
pINTID should be encoded on the same number of bits. ICID on 10 bits?

Thoughts?

Thanks

Eric
> 
>>
>>
>>>
>>>> +
>>>> +    LPI Pending Table layout:
>>>> +
>>>> +    As specified in the ARM Generic Interrupt Controller Architecture
>>>> +    Specification GIC Architecture version 3.0 and version 4. The first
>>>> +    1kB contains only zeros.
>>>>
>>>
>>> You definitely want to relax this. An ITS implementation is allowed (and
>>> actually encouraged) to maintain a coarse map in the first kB, and use
>>> this to quickly scan the table, which would be very useful on restore.
>> Maybe I miss something here. Currently I restore the ITEs before the
>> pending tables. So considering all the ITEs I know which LPI are defined
>> and which pending bits need to be restored. Why would I need to use a
>> coarse map for?
> 
> You could, instead of testing all the bits for which you can generate an
> LPI, look at the coarse map, which usually uses one bit to represent
> something like 64 bits of pending table, and find out what is currently
> pending. That's what HW does, but maybe there is no need to do this for
> the SW implementation, specially if we have very few LPIs.
> 
>> I understand the CPU cannot write the pending tables in our back, spec
>> says behavior would be unpredictable, right?
> 
> Absolutely. Only the ITS can touch that memory.
> 
> Thanks,
> 
> 	M.
>