tree 3dfd9273b27ced9ce7dca065d92ad3918f11171a
parent 54b8c7cbc57c1ce21f4e35101f2609092c4af49a
author Julien Grall <julien.grall@arm.com> 1558459299 +0100
committer Catalin Marinas <catalin.marinas@arm.com> 1559650657 +0100

arm64/fpsimd: Don't disable softirq when touching FPSIMD/SVE state

When the kernel is compiled with CONFIG_KERNEL_MODE_NEON, some part of
the kernel may be able to use FPSIMD/SVE. This is for instance the case
for crypto code.

Any use of FPSIMD/SVE in the kernel are clearly marked by using the
function kernel_neon_{begin, end}. Furthermore, this can only be used
when may_use_simd() returns true.

The current implementation of may_use_simd() allows softirq to use
FPSIMD/SVE unless it is currently in use (i.e kernel_neon_busy is true).
When in use, softirqs usually fall back to a software method.

At the moment, as a softirq may use FPSIMD/SVE, softirqs are disabled
when touching the FPSIMD/SVE context. This has the drawback to disable
all softirqs even if they are not using FPSIMD/SVE.

Since a softirq is supposed to check may_use_simd() anyway before
attempting to use FPSIMD/SVE, there is limited reason to keep softirq
disabled when touching the FPSIMD/SVE context. Instead, we can simply
disable preemption and mark the FPSIMD/SVE context as in use by setting
CPU's fpsimd_context_busy flag.

Two new helpers {get, put}_cpu_fpsimd_context are introduced to mark
the area using FPSIMD/SVE context and they are used to replace
local_bh_{disable, enable}. The functions kernel_neon_{begin, end} are
also re-implemented to use the new helpers.

Additionally, double-underscored versions of the helpers are provided to
called when preemption is already disabled. These are only relevant on
paths where irqs are disabled anyway, so they are not needed for
correctness in the current code. Let's use them anyway though: this
marks critical sections clearly and will help to avoid mistakes during
future maintenance.

The change has been benchmarked on Linux 5.1-rc4 with defconfig.

On Juno2:
    * hackbench 100 process 1000 (10 times)
    * .7% quicker

On ThunderX 2:
    * hackbench 1000 process 1000 (20 times)
    * 3.4% quicker

Reviewed-by: Dave Martin <dave.martin@arm.com>
Acked-by: Marc Zyngier <marc.zyngier@arm.com>
Signed-off-by: Julien Grall <julien.grall@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>