Re: [Qemu-devel] [PULL 08/10] target/arm: Conditionalize some asserts on aarch32 support

[Laszlo Ersek <lersek@redhat.com>]

On 07/18/19 14:30, Peter Maydell wrote:
> On Wed, 17 Jul 2019 at 16:08, Laszlo Ersek <lersek@redhat.com> wrote:
>>
>> On 07/17/19 15:46, Peter Maydell wrote:
>>> On Wed, 17 Jul 2019 at 14:36, Philippe Mathieu-Daudé <philmd@redhat.com> wrote:
>>>> I still wonder why this didn't assert on Peter's setup.
>>>
>>> My setup does not assert because my host kernel correctly
>>> provides the ID register values to QEMU. Laszlo's appears
>>> to be providing all-zeroes, which then obviously breaks
>>> assertions being made about the sanity of those ID register
>>> values...
>>
>> OK. Can you suggest a location that I should check in the host kernel?
> 
> I was about to write out the process of how we get these values
> from the kernel, but as the first step of that I read through
> QEMU's target/arm/kvm64.c:kvm_arm_get_host_cpu_features(),
> which is the function which reads these values using the
> KVM_GET_ONE_REG ioctl. It starts with an attempt to read
> ID_AA64PFR0, and has a comment for the error-handling case:
> 
>         /*
>          * Before v4.15, the kernel only exposed a limited number of system
>          * registers, not including any of the interesting AArch64 ID regs.
>          * For the most part we could leave these fields as zero with minimal
>          * effect, since this does not affect the values seen by the guest.
>          *
>          * However, it could cause problems down the line for QEMU,
>          * so provide a minimal v8.0 default.
>          *
>          * ??? Could read MIDR and use knowledge from cpu64.c.
>          * ??? Could map a page of memory into our temp guest and
>          *     run the tiniest of hand-crafted kernels to extract
>          *     the values seen by the guest.
>          * ??? Either of these sounds like too much effort just
>          *     to work around running a modern host kernel.
>          */
> 
> I have 4.15, and don't hit this assert; you have 4.14 and do,
> so I think you're going to be going through this codepath which
> currently sets only ahcf->isar.id_aa64pfr0 and none of the other
> ID register fields in the isar struct.
> 
> I'm not sure exactly which kernel commits added the ID register
> reading support. (The relevant kernel code is in
> arch/arm64/kvm/sys_regs.c I think.)

I compared that file between the downstream kernel source and upstream v4.15, and it looks like the following series (indeed released as part of v4.15) is what's missing down-stream, for this particular use case:

     1  27e64b4be4b8 regset: Add support for dynamically sized regsets
     2  94ef7ecbdf6f arm64: fpsimd: Correctly annotate exception helpers called from asm
     3  abf73988a7c2 arm64: signal: Verify extra data is user-readable in sys_rt_sigreturn
     4  93390c0a1b20 arm64: KVM: Hide unsupported AArch64 CPU features from guests
     5  b472db6cf8c6 arm64: efi: Add missing Kconfig dependency on KERNEL_MODE_NEON
     6  38b9aeb32fa7 arm64: Port deprecated instruction emulation to new sysctl interface
     7  9cf5b54fafed arm64: fpsimd: Simplify uses of {set,clear}_ti_thread_flag()
     8  672365649cca arm64/sve: System register and exception syndrome definitions
     9  1fc5dce78ad1 arm64/sve: Low-level SVE architectural state manipulation functions
    10  ddd25ad1fde8 arm64/sve: Kconfig update and conditional compilation support
    11  d0b8cd318788 arm64/sve: Signal frame and context structure definition
    12  22043a3c082a arm64/sve: Low-level CPU setup
    13  bc0ee4760364 arm64/sve: Core task context handling
    14  79ab047c75d6 arm64/sve: Support vector length resetting for new processes
    15  8cd969d28fd2 arm64/sve: Signal handling support
    16  7582e22038a2 arm64/sve: Backend logic for setting the vector length
    17  8f1eec57cdcc arm64: cpufeature: Move sys_caps_initialised declarations
    18  2e0f2478ea37 arm64/sve: Probe SVE capabilities and usable vector lengths
    19  1bd3f93641ec arm64/sve: Preserve SVE registers around kernel-mode NEON use
    20  fdfa976cae5c arm64/sve: Preserve SVE registers around EFI runtime service calls
    21  43d4da2c45b2 arm64/sve: ptrace and ELF coredump support
    22  2d2123bc7c7f arm64/sve: Add prctl controls for userspace vector length management
    23  4ffa09a939ab arm64/sve: Add sysctl to set the default vector length for new processes
    24  17eed27b02da arm64/sve: KVM: Prevent guests from using SVE
    25  aac45ffd1f8e arm64/sve: KVM: Treat guest SVE use as undefined instruction execution
    26  07d79fe7c223 arm64/sve: KVM: Hide SVE from CPU features exposed to guests
    27  43994d824e84 arm64/sve: Detect SVE and activate runtime support
    28  ce6990813f15 arm64/sve: Add documentation

The differences found by the simple "diff" that I mention above are mainly due to commit #4 (93390c0a1b20, "arm64: KVM: Hide unsupported AArch64 CPU features from guests", 2017-11-03).

I found a (likely non-final) version of the cover letter too, here: https://www.spinics.net/lists/arm-kernel/msg599528.html

I guess I should convince myself to install RHEL8 on the Mustang sometime...

Thanks!
Laszlo

> Anyway, I think we need to do at least one of:
>  * enhance the "provide a minimal v8.0 default" code in this
>    condition in kvm_arm_get_host_cpu_features() so that it
>    populates the ID registers sufficiently to avoid asserts
>    and other bad things
>  * make the asserts on ID register oddnesses be only for TCG
>    (ie where QEMU controls the values) and not for KVM
> 
> thanks
> -- PMM
>