[PATCH v11 00/43] KVM: arm64: Nested Virtualization support (FEAT_NV2 only)

Ganapatrao Kulkarni gankulkarni at os.amperecomputing.com
Fri Nov 24 05:22:22 PST 2023



On 24-11-2023 06:21 pm, Marc Zyngier wrote:
> On Fri, 24 Nov 2023 12:34:41 +0000,
> Ganapatrao Kulkarni <gankulkarni at os.amperecomputing.com> wrote:
>>
>>
>>
>> On 24-11-2023 03:49 pm, Marc Zyngier wrote:
>>> On Fri, 24 Nov 2023 09:50:33 +0000,
>>> Ganapatrao Kulkarni <gankulkarni at os.amperecomputing.com> wrote:
>>>>
>>>>
>>>>
>>>> On 23-11-2023 10:14 pm, Marc Zyngier wrote:
>>>>> On Thu, 23 Nov 2023 16:21:48 +0000,
>>>>> Miguel Luis <miguel.luis at oracle.com> wrote:
>>>>>>
>>>>>> Hi Marc,
>>>>>>
>>>>>> On 21/11/2023 18:02, Marc Zyngier wrote:
>>>>>>> On Tue, 21 Nov 2023 16:49:52 +0000,
>>>>>>> Miguel Luis <miguel.luis at oracle.com> wrote:
>>>>>>>> Hi Marc,
>>>>>>>>
>>>>>>>>> On 20 Nov 2023, at 12:09, Marc Zyngier <maz at kernel.org> wrote:
>>>>>>>>>
>>>>>>>>> This is the 5th drop of NV support on arm64 for this year, and most
>>>>>>>>> probably the last one for this side of Christmas.
>>>>>>>>>
>>>>>>>>> For the previous episodes, see [1].
>>>>>>>>>
>>>>>>>>> What's changed:
>>>>>>>>>
>>>>>>>>> - Drop support for the original FEAT_NV. No existing hardware supports
>>>>>>>>>     it without FEAT_NV2, and the architecture is deprecating the former
>>>>>>>>>     entirely. This results in fewer patches, and a slightly simpler
>>>>>>>>>     model overall.
>>>>>>>>>
>>>>>>>>> - Reorganise the series to make it a bit more logical now that FEAT_NV
>>>>>>>>>     is gone.
>>>>>>>>>
>>>>>>>>> - Apply the NV idreg restrictions on VM first run rather than on each
>>>>>>>>>     access.
>>>>>>>>>
>>>>>>>>> - Make the nested vgic shadow CPU interface a per-CPU structure rather
>>>>>>>>>     than per-vcpu.
>>>>>>>>>
>>>>>>>>> - Fix the EL0 timer fastpath
>>>>>>>>>
>>>>>>>>> - Work around the architecture deficiencies when trapping WFI from a
>>>>>>>>>     L2 guest.
>>>>>>>>>
>>>>>>>>> - Fix sampling of nested vgic state (MISR, ELRSR, EISR)
>>>>>>>>>
>>>>>>>>> - Drop the patches that have already been merged (NV trap forwarding,
>>>>>>>>>     per-MMU VTCR)
>>>>>>>>>
>>>>>>>>> - Rebased on top of 6.7-rc2 + the FEAT_E2H0 support [2].
>>>>>>>>>
>>>>>>>>> The branch containing these patches (and more) is at [3]. As for the
>>>>>>>>> previous rounds, my intention is to take a prefix of this series into
>>>>>>>>> 6.8, provided that it gets enough reviewing.
>>>>>>>>>
>>>>>>>>> [1] https://lore.kernel.org/r/20230515173103.1017669-1-maz@kernel.org
>>>>>>>>> [2] https://lore.kernel.org/r/20231120123721.851738-1-maz@kernel.org
>>>>>>>>> [3] https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/log/?h=kvm-arm64/nv-6.8-nv2-only
>>>>>>>>>
>>>>>>>> While I was testing this with kvmtool for 5.16 I noted the following on dmesg:
>>>>>>>>
>>>>>>>> [  803.014258] kvm [19040]: Unsupported guest sys_reg access at: 8129fa50 [600003c9]
>>>>>>>>                    { Op0( 3), Op1( 5), CRn( 1), CRm( 0), Op2( 2), func_read },
>>>>>>>>
>>>>>>>> This is CPACR_EL12.
>>>>>>> CPACR_EL12 is redirected to VNCR[0x100]. It really shouldn't trap...
>>>>>>>
>>>>>>>> Still need yet to debug.
>>>>>>> Can you disassemble the guest around the offending PC?
>>>>>>
>>>>>> [ 1248.686350] kvm [7013]: Unsupported guest sys_reg access at: 812baa50 [600003c9]
>>>>>>                    { Op0( 3), Op1( 5), CRn( 1), CRm( 0), Op2( 2), func_read },
>>>>>>
>>>>>>     12baa00:    14000008     b    0x12baa20
>>>>>>     12baa04:    d000d501     adrp    x1, 0x2d5c000
>>>>>>     12baa08:    91154021     add    x1, x1, #0x550
>>>>>>     12baa0c:    f9400022     ldr    x2, [x1]
>>>>>>     12baa10:    f9400421     ldr    x1, [x1, #8]
>>>>>>     12baa14:    8a010042     and    x2, x2, x1
>>>>>>     12baa18:    d3441c42     ubfx    x2, x2, #4, #4
>>>>>>     12baa1c:    b4000082     cbz    x2, 0x12baa2c
>>>>>>     12baa20:    d2a175a0     mov    x0, #0xbad0000                 // #195887104
>>>>>>     12baa24:    f2994220     movk    x0, #0xca11
>>>>>>     12baa28:    d69f03e0     eret
>>>>>>     12baa2c:    d2c00080     mov    x0, #0x400000000               // #17179869184
>>>>>>     12baa30:    f2b10000     movk    x0, #0x8800, lsl #16
>>>>>>     12baa34:    f2800000     movk    x0, #0x0
>>>>>>     12baa38:    d51c1100     msr    hcr_el2, x0
>>>>>>     12baa3c:    d5033fdf     isb
> 
> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> This.
> 
>>>>>>     12baa40:    d53c4100     mrs    x0, sp_el1
>>>>>>     12baa44:    9100001f     mov    sp, x0
>>>>>>     12baa48:    d538d080     mrs    x0, tpidr_el1
>>>>>>     12baa4c:    d51cd040     msr    tpidr_el2, x0
>>>>>>     12baa50:    d53d1040     mrs    x0, cpacr_el12
>>>>>>     12baa54:    d5181040     msr    cpacr_el1, x0
>>>>>>     12baa58:    d53dc000     mrs    x0, vbar_el12
>>>>>>     12baa5c:    d518c000     msr    vbar_el1, x0
>>>>>>     12baa60:    d53c1120     mrs    x0, mdcr_el2
>>>>>>     12baa64:    9272f400     and    x0, x0, #0xffffffffffffcfff
>>>>>>     12baa68:    9266f400     and    x0, x0, #0xfffffffffcffffff
>>>>>>     12baa6c:    d51c1120     msr    mdcr_el2, x0
>>>>>>     12baa70:    d53d2040     mrs    x0, tcr_el12
>>>>>>     12baa74:    d5182040     msr    tcr_el1, x0
>>>>>>     12baa78:    d53d2000     mrs    x0, ttbr0_el12
>>>>>>     12baa7c:    d5182000     msr    ttbr0_el1, x0
>>>>>>     12baa80:    d53d2020     mrs    x0, ttbr1_el12
>>>>>>     12baa84:    d5182020     msr    ttbr1_el1, x0
>>>>>>     12baa88:    d53da200     mrs    x0, mair_el12
>>>>>>     12baa8c:    d518a200     msr    mair_el1, x0
>>>>>>     12baa90:    d5380761     mrs    x1, s3_0_c0_c7_3
>>>>>>     12baa94:    d3400c21     ubfx    x1, x1, #0, #4
>>>>>>     12baa98:    b4000141     cbz    x1, 0x12baac0
>>>>>>     12baa9c:    d53d2060     mrs    x0, s3_5_c2_c0_3
>>>>>
>>>>> OK, this is suspiciously close to the location Ganapatrao was having
>>>>> issues with. Are you running on the same hardware?
>>>>>
>>>>> In any case, we should never take a trap for this access. Can you dump
>>>>> HCR_EL2 at the point where the guest traps (in switch.c)?
>>>>>
>>>>
>>>> I have dumped HCR_EL2 before entry to L1 in both V11 and V10.
>>>> on V10 HCR_EL2=0x2743c827c263f
>>>> on V11 HCR_EL2=0x27c3c827c263f
>>>>
>>>> on V11 the function vcpu_el2_e2h_is_set(vcpu) is returning false
>>>> resulting in NV1 bit set along with NV and NV2.
>>>> AFAIK, For L1 to be in VHE, NV1 bit should be zero and NV=NV2=1.
>>>>
>>>> I could boot L1 then L2, if I hack vcpu_el2_e2h_is_set to return true.
>>>> There could be a bug in V11 or E2H0 patchset resulting in
>>>> vcpu_el2_e2h_is_set() returning false?
>>>
>>> The E2H0 series should only force vcpu_el2_e2h_is_set() to return
>>> true, but not set it to false. Can you dump the *guest's* version of
>>> HCR_EL2 at this point?
>>>
>>
>> with V11: vhcr_el2=0x100030080000000 mask=0x100af00ffffffff
> 
> How is this value possible if the write to HCR_EL2 has taken place?
> When do you sample this?

I am not sure how and where it got set. I think, whatever it is set, it 
is due to false return of vcpu_el2_e2h_is_set(). Need to understand/debug.
The vhcr_el2 value I have shared is traced along with hcr in function 
__activate_traps/__compute_hcr.

> 
>> with V10: vhcr_el2=0x488000000
>> with hack+V11: vhcr_el2=0x488000000 mask=0x100af00ffffffff
> 
> Well, of course, if you constrain the value of HCR_EL2...
> 
> 	M.
> 

Thanks,
Ganapat



More information about the linux-arm-kernel mailing list