Re: [REPOST PATCH v3 5/5] amd_iommu: report x2APIC support to the operating system

[Bui Quang Minh <minhquangbui99@gmail.com>]

On 5/14/23 15:55, Bui Quang Minh wrote:
> On 5/12/23 21:39, Michael S. Tsirkin wrote:
>> On Tue, Apr 11, 2023 at 09:24:40PM +0700, Bui Quang Minh wrote:
>>> This commit adds XTSup configuration to let user choose to whether 
>>> enable
>>> this feature or not. When XTSup is enabled, additional bytes in IRTE 
>>> with
>>> enabled guest virtual VAPIC are used to support 32-bit destination id.
>>>
>>> Additionally, this commit changes to use IVHD type 0x11 in ACPI table 
>>> for
>>> feature report to operating system. This is because Linux does not use
>>> XTSup in IOMMU Feature Reporting field of IVHD type 0x10 but only use 
>>> XTSup
>>> bit in EFR Register Image of IVHD 0x11 to indicate x2APIC support (see
>>> init_iommu_one in linux/drivers/iommu/amd/init.c)
>>>
>>> Signed-off-by: Bui Quang Minh <minhquangbui99@gmail.com>
>>
>> I'm concerned that switching to type 11 will break some older guests.
>> It would be better if we could export both type 10 and type 11
>> ivhd. A question however would be how does this interact
>> with older guests. For example:
>> https://lists.linuxfoundation.org/pipermail/iommu/2016-January/015310.html
>> it looks like linux before 2016 only expected one ivhd entry?
> 
> Export both type 0x10 and 0x11 looks reasonable to me. Before the above 
> commit, I see that Linux still loops through multiple ivhd but only 
> handles one with type 0x10. On newer kernel, it will choose to handle 
> the type that appears last corresponding the first devid, which is weird 
> in my opinion.
> 
> +static u8 get_highest_supported_ivhd_type(struct acpi_table_header *ivrs)
> +{
> +    u8 *base = (u8 *)ivrs;
> +    struct ivhd_header *ivhd = (struct ivhd_header *)
> +                    (base + IVRS_HEADER_LENGTH);
> +    u8 last_type = ivhd->type;
> +    u16 devid = ivhd->devid;
> +
> +    while (((u8 *)ivhd - base < ivrs->length) &&
> +           (ivhd->type <= ACPI_IVHD_TYPE_MAX_SUPPORTED)) {
> +        u8 *p = (u8 *) ivhd;
> +
> +        if (ivhd->devid == devid)
> +            last_type = ivhd->type;
> +        ivhd = (struct ivhd_header *)(p + ivhd->length);
> +    }
> +
> +    return last_type;
> +}
> 
> So when exposing type 0x10 following by 0x11, old kernel only parses 
> 0x10 and does not support x2APIC while new kernel parse 0x11 and support 
> x2APIC. I will expose both types in the next version.
> 
>> Some research and testing here would be benefitial.
>> Similarly for windows guests.
>>
>> Thanks!
>>
>>
>>
>>> ---
>>>   hw/i386/acpi-build.c | 28 ++++++++++++++--------------
>>>   hw/i386/amd_iommu.c  | 21 +++++++++++++++++++--
>>>   hw/i386/amd_iommu.h  | 16 +++++++++++-----
>>>   3 files changed, 44 insertions(+), 21 deletions(-)
>>>
>>> diff --git a/hw/i386/acpi-build.c b/hw/i386/acpi-build.c
>>> index ec857a117e..72d6bb2892 100644
>>> --- a/hw/i386/acpi-build.c
>>> +++ b/hw/i386/acpi-build.c
>>> @@ -2339,7 +2339,7 @@ static void
>>>   build_amd_iommu(GArray *table_data, BIOSLinker *linker, const char 
>>> *oem_id,
>>>                   const char *oem_table_id)
>>>   {
>>> -    int ivhd_table_len = 24;
>>> +    int ivhd_table_len = 40;
>>>       AMDVIState *s = AMD_IOMMU_DEVICE(x86_iommu_get_default());
>>>       GArray *ivhd_blob = g_array_new(false, true, 1);
>>>       AcpiTable table = { .sig = "IVRS", .rev = 1, .oem_id = oem_id,
>>> @@ -2349,18 +2349,19 @@ build_amd_iommu(GArray *table_data, 
>>> BIOSLinker *linker, const char *oem_id,
>>>       /* IVinfo - IO virtualization information common to all
>>>        * IOMMU units in a system
>>>        */
>>> -    build_append_int_noprefix(table_data, 40UL << 8/* PASize */, 4);
>>> +    build_append_int_noprefix(table_data,
>>> +                             (1UL << 0) | /* EFRSup */
>>> +                             (40UL << 8), /* PASize */
>>> +                             4);
>>>       /* reserved */
>>>       build_append_int_noprefix(table_data, 0, 8);
>>> -    /* IVHD definition - type 10h */
>>> -    build_append_int_noprefix(table_data, 0x10, 1);
>>> +    /* IVHD definition - type 11h */
>>> +    build_append_int_noprefix(table_data, 0x11, 1);
>>>       /* virtualization flags */
>>>       build_append_int_noprefix(table_data,
>>>                                (1UL << 0) | /* HtTunEn      */
>>> -                             (1UL << 4) | /* iotblSup     */
>>
>> btw this should have been iotlbsup?
>>
>>> -                             (1UL << 6) | /* PrefSup      */
>>> -                             (1UL << 7),  /* PPRSup       */
>>> +                             (1UL << 4),  /* iotblSup     */
>>>                                1);
>>>       /*
>>
>> hmm why are you removing these other flags?
> 
> According to the AMD IOMMU specification, the bit 6, 7 are reserved in 
> type 0x11 which are PerfSup, PPRSup respectively in type 0x10 so I 
> remove those flags when changing to type 0x11. In type 0x11, these 
> feature are reported via the below EFR Register Image I believe.
> 
>>
>>> @@ -2404,13 +2405,12 @@ build_amd_iommu(GArray *table_data, 
>>> BIOSLinker *linker, const char *oem_id,
>>>       build_append_int_noprefix(table_data, 0, 2);
>>>       /* IOMMU info */
>>>       build_append_int_noprefix(table_data, 0, 2);
>>> -    /* IOMMU Feature Reporting */
>>> -    build_append_int_noprefix(table_data,
>>> -                             (48UL << 30) | /* HATS   */
>>> -                             (48UL << 28) | /* GATS   */
>>> -                             (1UL << 2)   | /* GTSup  */
>>> -                             (1UL << 6),    /* GASup  */
>>> -                             4);
>>> +    /* IOMMU Attributes */
>>> +    build_append_int_noprefix(table_data, 0, 4);
>>> +    /* EFR Register Image */
>>> +    build_append_int_noprefix(table_data, s->efr_reg, 8);
>>> +    /* EFR Register Image 2 */
>>> +    build_append_int_noprefix(table_data, 0, 8);
>>
>>
>> here too. what's going on?
> 
> Same as the above, the structure of type 0x11 is different from 0x10. At 
> offset 20 it is not IOMMU feature reporting but IOMMU attributes. And 
> there are 2 more fields: EFR Register Image, EFR Register Image 2 before 
> IVHD device entries.
> 
>>>       /* IVHD entries as found above */
>>>       g_array_append_vals(table_data, ivhd_blob->data, ivhd_blob->len);
>>> diff --git a/hw/i386/amd_iommu.c b/hw/i386/amd_iommu.c
>>> index bcd016f5c5..5dfa93d945 100644
>>> --- a/hw/i386/amd_iommu.c
>>> +++ b/hw/i386/amd_iommu.c
>>> @@ -31,6 +31,7 @@
>>>   #include "hw/i386/apic_internal.h"
>>>   #include "trace.h"
>>>   #include "hw/i386/apic-msidef.h"
>>> +#include "hw/qdev-properties.h"
>>>   /* used AMD-Vi MMIO registers */
>>>   const char *amdvi_mmio_low[] = {
>>> @@ -1155,7 +1156,12 @@ static int amdvi_int_remap_ga(AMDVIState *iommu,
>>>       irq->vector = irte.hi.fields.vector;
>>>       irq->dest_mode = irte.lo.fields_remap.dm;
>>>       irq->redir_hint = irte.lo.fields_remap.rq_eoi;
>>> -    irq->dest = irte.lo.fields_remap.destination;
>>> +    if (!iommu->xtsup) {
>>> +        irq->dest = irte.lo.fields_remap.destination & 0xff;
>>> +    } else {
>>> +        irq->dest = irte.lo.fields_remap.destination |
>>> +                    (irte.hi.fields.destination_hi << 24);
>>> +    }
>>
>> Weird way to put it. Why not if (iommu->xtsup) ... ?
> 
> Okay, I'll fix this in next version.
> 
>>>       return 0;
>>>   }
>>> @@ -1503,10 +1509,15 @@ static void amdvi_init(AMDVIState *s)
>>>       s->enabled = false;
>>>       s->ats_enabled = false;
>>>       s->cmdbuf_enabled = false;
>>> +    s->efr_reg = AMDVI_DEFAULT_EXT_FEATURES;
>>> +
>>> +    if (s->xtsup) {
>>> +        s->efr_reg |= AMDVI_FEATURE_XT;
>>> +    }
>>>       /* reset MMIO */
>>>       memset(s->mmior, 0, AMDVI_MMIO_SIZE);
>>> -    amdvi_set_quad(s, AMDVI_MMIO_EXT_FEATURES, AMDVI_EXT_FEATURES,
>>> +    amdvi_set_quad(s, AMDVI_MMIO_EXT_FEATURES, s->efr_reg,
>>>               0xffffffffffffffef, 0);
>>>       amdvi_set_quad(s, AMDVI_MMIO_STATUS, 0, 0x98, 0x67);
>>> @@ -1586,6 +1597,11 @@ static void amdvi_sysbus_realize(DeviceState 
>>> *dev, Error **errp)
>>>       amdvi_init(s);
>>>   }
>>> +static Property amdvi_properties[] = {
>>> +    DEFINE_PROP_BOOL("xtsup", AMDVIState, xtsup, false),
>>> +    DEFINE_PROP_END_OF_LIST(),
>>> +};
>>> +
>>>   static const VMStateDescription vmstate_amdvi_sysbus = {
>>>       .name = "amd-iommu",
>>>       .unmigratable = 1
>>> @@ -1612,6 +1628,7 @@ static void amdvi_sysbus_class_init(ObjectClass 
>>> *klass, void *data)
>>>       dc->user_creatable = true;
>>>       set_bit(DEVICE_CATEGORY_MISC, dc->categories);
>>>       dc->desc = "AMD IOMMU (AMD-Vi) DMA Remapping device";
>>> +    device_class_set_props(dc, amdvi_properties);
>>>   }
>>>   static const TypeInfo amdvi_sysbus = {
>>> diff --git a/hw/i386/amd_iommu.h b/hw/i386/amd_iommu.h
>>> index 79d38a3e41..96df7b0400 100644
>>> --- a/hw/i386/amd_iommu.h
>>> +++ b/hw/i386/amd_iommu.h
>>> @@ -154,6 +154,7 @@
>>>   #define AMDVI_FEATURE_PREFETCH            (1ULL << 0) /* page 
>>> prefetch       */
>>>   #define AMDVI_FEATURE_PPR                 (1ULL << 1) /* PPR 
>>> Support         */
>>> +#define AMDVI_FEATURE_XT                  (1ULL << 2) /* x2APIC 
>>> Support      */
>>>   #define AMDVI_FEATURE_GT                  (1ULL << 4) /* Guest 
>>> Translation   */
>>>   #define AMDVI_FEATURE_IA                  (1ULL << 6) /* inval all 
>>> support   */
>>>   #define AMDVI_FEATURE_GA                  (1ULL << 7) /* guest 
>>> VAPIC support */
>>> @@ -173,8 +174,9 @@
>>>   #define AMDVI_IOTLB_MAX_SIZE 1024
>>>   #define AMDVI_DEVID_SHIFT    36
>>> -/* extended feature support */
>>> -#define AMDVI_EXT_FEATURES (AMDVI_FEATURE_PREFETCH | 
>>> AMDVI_FEATURE_PPR | \
>>> +/* default extended feature */
>>> +#define AMDVI_DEFAULT_EXT_FEATURES \
>>> +        (AMDVI_FEATURE_PREFETCH | AMDVI_FEATURE_PPR | \
>>>           AMDVI_FEATURE_IA | AMDVI_FEATURE_GT | AMDVI_FEATURE_HE | \
>>>           AMDVI_GATS_MODE | AMDVI_HATS_MODE | AMDVI_FEATURE_GA)
>>> @@ -278,8 +280,8 @@ union irte_ga_lo {
>>>                   dm:1,
>>>                   /* ------ */
>>>                   guest_mode:1,
>>> -                destination:8,
>>> -                rsvd_1:48;
>>> +                destination:24,
>>> +                rsvd_1:32;
>>>     } fields_remap;
>>>   };
>>> @@ -287,7 +289,8 @@ union irte_ga_hi {
>>>     uint64_t val;
>>>     struct {
>>>         uint64_t  vector:8,
>>> -                rsvd_2:56;
>>> +                rsvd_2:48,
>>> +                destination_hi:8;
>>>     } fields;
>>>   };
>>> @@ -367,6 +370,9 @@ struct AMDVIState {
>>>       /* Interrupt remapping */
>>>       bool ga_enabled;
>>> +    bool xtsup;
>>> +
>>> +    uint64_t efr_reg;            /* extended feature register */
>>>   };
>>>   #endif
>>> -- 
>>> 2.25.1

I have posted version 4 to export both IVHD types and flip the 
iommu->xtsup for readability.

I have tested with old Linux before this commit 
8c7142f56fedfc6824b5bca56fee1f443e01746b (iommu/amd: Use the most 
comprehensive IVHD type that the driver can support) (Linux 4.6-rc2) 
that does not support IVHD type 0x11, it can still detect AMD IOMMU 
device via IVHD 0x10 but cannot detect x2APIC support. With newer Linux, 
it successfully detects x2APIC support.

With Windows, I don't have much experience, I have tried to get the 
driver for reverse engineer but looks like the driver can be installed 
via the installer only. And I don't have the device so installer aborts 
the driver installation.

Thanks,
Quang Minh.