Re: [PATCH v14 00/39] arm64/sme: Initial support for the Scalable Matrix Extension

[Qian Cai <quic_qiancai@quicinc.com>]

On Tue, Apr 19, 2022 at 12:22:08PM +0100, Mark Brown wrote:
> This series provides initial support for the ARMv9 Scalable Matrix
> Extension (SME).  SME takes the approach used for vectors in SVE and
> extends this to provide architectural support for matrix operations.  A
> more detailed overview can be found in [1].
> 
> For the kernel SME can be thought of as a series of features which are
> intended to be used together by applications but operate mostly
> orthogonally:
> 
>  - The ZA matrix register.
>  - Streaming mode, in which ZA can be accessed and a subset of SVE
>    features are available.
>  - A second vector length, used for streaming mode SVE and ZA and
>    controlled using a similar interface to that for SVE.
>  - TPIDR2, a new userspace controllable system register intended for use
>    by the C library for storing context related to the ZA ABI.
> 
> A substantial part of the series is dedicated to refactoring the
> existing SVE support so that we don't need to duplicate code for
> handling vector lengths and the SVE registers, this involves creating an
> array of vector types and making the users take the vector type as a
> parameter.  I'm not 100% happy with this but wasn't able to come up with
> anything better, duplicating code definitely felt like a bad idea so
> this felt like the least bad thing.  If this approach makes sense to
> people it might make sense to split this off into a separate series
> and/or merge it while the rest is pending review to try to make things a
> little more digestable, the series is very large so it'd probably make
> things easier to digest if some of the preparatory refactoring could be
> merged before the rest is ready.
> 
> One feature of the architecture of particular note is that switching
> to and from streaming mode may change the size of and invalidate the
> contents of the SVE registers, and when in streaming mode the FFR is not
> accessible.  This complicates aspects of the ABI like signal handling
> and ptrace.
> 
> This initial implementation is mainly intended to get the ABI in place,
> there are several areas which will be worked on going forwards - some of
> these will be blockers, others could be handled in followup serieses:
> 
>  - SME is currently not supported for KVM guests, this will be done as a
>    followup series.  A host system can use SME and run KVM guests but
>    SME is not available in the guests.
>  - The KVM host support is done in a very simplistic way, were anyone to
>    attempt to use it in production there would be performance impacts on
>    hosts with SME support. As part of this we also add enumeration of
>    fine grained traps.
>  - There is not currently ptrace or signal support TPIDR2, this will be
>    done as a followup series.
>  - No support is currently provided for scheduler control of SME or SME
>    applications, given the size of the SME register state the context
>    switch overhead may be noticable so this may be needed especially for
>    real time applications.  Similar concerns already exist for larger
>    SVE vector lengths but are amplified for SME, particularly as the
>    vector length increases.
>  - There has been no work on optimising the performance of anything the
>    kernel does.
> 
> It is not expected that any systems will be encountered that support SME
> but not SVE, SME is an ARMv9 feature and SVE is mandatory for ARMv9.
> The code attempts to handle any such systems that are encountered but
> this hasn't been tested extensively.

Running CPU offline/online on a Neoverse-N1 server will trigger a crash.
A data point is setting CONFIG_ARM64_SVE=n could avoid it.

 kernel BUG at arch/arm64/kernel/cpufeature.c:1353!
 Internal error: Oops - BUG: 0 [#1] PREEMPT SMP
 CPU: 88 PID: 0 Comm: swapper/88 Not tainted 5.18.0-rc4-next-20220426-00006-gfea0cdfbc1de #60
 pstate: 204001c9 (nzCv dAIF +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
 pc : __read_sysreg_by_encoding
 lr : has_cpuid_feature
 sp : ffff80000a827d10
 x29: ffff80000a827d10 x28: 0000000000000000 x27: ffffbb3c708efb8a
 x26: 1ffff7678e11df71 x25: 0000000000000002 x24: 0000000000000003
 x23: ffffbb3c75870e80 x22: dfff800000000000 x21: 0000000000000029
 x20: ffffbb3c708efba0 x19: ffffbb3c708efb80 x18: ffffbb3c73eb7d1c
 x17: 000000040044ffff x16: 1fffe7fff0e51474 x15: 1fffe806c1d7b54a
 x14: 1fffe7fff0e5146c x13: 0000000000000004 x12: ffff77678e839850
 x11: 1ffff7678e83984f x10: ffff77678e83984f x9 : ffffbb3c6d7deef0
 x8 : ffffbb3c741cc27f x7 : 0000000000000001 x6 : ffff77678e83984f
 x5 : ffffbb3c741cc278 x4 : 0000000000000000 x3 : 1fffe7fff0e51359
 x2 : 1ffff7678e11df74 x1 : 0000000000180480 x0 : 00000000001804a0
 Call trace:
  __read_sysreg_by_encoding
  has_cpuid_feature
  verify_local_cpu_caps
  verify_local_cpu_capabilities
  check_local_cpu_capabilities
  secondary_start_kernel
  __secondary_switched
 Code: 17ffff34 d5380234 17ffff32 f90013f5 (d4210000)
 ---[ end trace 0000000000000000 ]---
 Kernel panic - not syncing: Oops - BUG: Fatal exception
 SMP: stopping secondary CPUs
 Kernel Offset: 0x3b3c657a0000 from 0xffff800008000000
 PHYS_OFFSET: 0x80000000
 CPU features: 0x000,0021700d,19801c82
 Memory Limit: none
 ---[ end Kernel panic - not syncing: Oops - BUG: Fatal exception ]---

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel