On Fri, Oct 27, 2017 at 11:50:53AM +0100, Dave P Martin wrote:
> This patch adds the core support for switching and managing the SVE
> architectural state of user tasks.
>
> Calls to the existing FPSIMD low-level save/restore functions are
> factored out as new functions task_fpsimd_{save,load}(), since SVE
> now dynamically may or may not need to be handled at these points
> depending on the kernel configuration, hardware features discovered
> at boot, and the runtime state of the task. To make these
> decisions as fast as possible, const cpucaps are used where
> feasible, via the system_supports_sve() helper.
>
> The SVE registers are only tracked for threads that have explicitly
> used SVE, indicated by the new thread flag TIF_SVE. Otherwise, the
> FPSIMD view of the architectural state is stored in
> thread.fpsimd_state as usual.
>
> When in use, the SVE registers are not stored directly in
> thread_struct due to their potentially large and variable size.
> Because the task_struct slab allocator must be configured very
> early during kernel boot, it is also tricky to configure it
> correctly to match the maximum vector length provided by the
> hardware, since this depends on examining secondary CPUs as well as
> the primary. Instead, a pointer sve_state in thread_struct points
> to a dynamically allocated buffer containing the SVE register data,
> and code is added to allocate and free this buffer at appropriate
> times.
>
> TIF_SVE is set when taking an SVE access trap from userspace, if
> suitable hardware support has been detected. This enables SVE for
> the thread: a subsequent return to userspace will disable the trap
> accordingly. If such a trap is taken without sufficient system-
> wide hardware support, SIGILL is sent to the thread instead as if
> an undefined instruction had been executed: this may happen if
> userspace tries to use SVE in a system where not all CPUs support
> it for example.
>
> The kernel will clear TIF_SVE and disable SVE for the thread
> whenever an explicit syscall is made by userspace. For backwards
> compatibility reasons and conformance with the spirit of the base
> AArch64 procedure call standard, the subset of the SVE register
> state that aliases the FPSIMD registers is still preserved across a
> syscall even if this happens. The remainder of the SVE register
> state logically becomes zero at syscall entry, though the actual
> zeroing work is currently deferred until the thread next tries to
> use SVE, causing another trap to the kernel. This implementation
> is suboptimal: in the future, the fastpath case may be optimised
> to zero the registers in-place and leave SVE enabled for the task,
> where beneficial.
>
> TIF_SVE is also cleared in the following slowpath cases, which are
> taken as reasonable hints that the task may no longer use SVE:
> * exec
> * fork and clone
>
> Code is added to sync data between thread.fpsimd_state and
> thread.sve_state whenever enabling/disabling SVE, in a manner
> consistent with the SVE architectural programmer's model.
>
> Signed-off-by: Dave Martin <Dave.Martin@xxxxxxx>
> Cc: Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx>
> Cc: Alex Bennée <alex.bennee@xxxxxxxxxx>
Reviewed-by: Catalin Marinas <catalin.marinas@xxxxxxx>
