Re: [PATCH 0/7] microblaze: Define SMP safe operations

[Michal Simek]

On 12. 02. 20 17:08, Peter Zijlstra wrote:
> On Wed, Feb 12, 2020 at 04:42:22PM +0100, Michal Simek wrote:
> 
>> Microblaze has 32bit exclusive load/store instructions which should be used
>> instead of irq enable/disable. For more information take a look at
>> https://www.xilinx.com/support/documentation/sw_manuals/xilinx2019_2/ug984-vivado-microblaze-ref.pdf
>> starting from page 25.
> 
>>  arch/microblaze/include/asm/Kbuild           |   1 -
>>  arch/microblaze/include/asm/atomic.h         | 265 ++++++++++++++++++-
>>  arch/microblaze/include/asm/bitops.h         | 189 +++++++++++++
>>  arch/microblaze/include/asm/cmpxchg.h        |  87 ++++++
>>  arch/microblaze/include/asm/cpuinfo.h        |   2 +-
>>  arch/microblaze/include/asm/pgtable.h        |  19 +-
>>  arch/microblaze/include/asm/spinlock.h       | 240 +++++++++++++++++
>>  arch/microblaze/include/asm/spinlock_types.h |  25 ++
>>  arch/microblaze/kernel/cpu/cache.c           | 154 ++++++-----
>>  arch/microblaze/kernel/cpu/cpuinfo.c         |  38 ++-
>>  arch/microblaze/kernel/cpu/mb.c              | 207 ++++++++-------
>>  arch/microblaze/kernel/timer.c               |   2 +-
>>  arch/microblaze/mm/consistent.c              |   8 +-
>>  13 files changed, 1040 insertions(+), 197 deletions(-)
>>  create mode 100644 arch/microblaze/include/asm/bitops.h
>>  create mode 100644 arch/microblaze/include/asm/spinlock.h
>>  create mode 100644 arch/microblaze/include/asm/spinlock_types.h
> 
> I'm missing asm/barrier.h

This has been sent in previous patchset. Link was in this email.

> 
> Also that PDF (thanks for that!), seems light on memory ordering
> details.
> 
> Your comment:
> 
> +/*
> + * clear_bit doesn't imply a memory barrier
> + */
> 
> worries me, because that would imply your ll/sc does not impose order,
> but then you also don't have any explicit barriers in your locking
> primitives or atomics where required.

I think this is just comment which shouldn't be there.
clear_bit is calling clear_bits which is using exclusive load/store
instruction which should impose ordering.

> 
> In the PDF I only find MBAR; is that what smp_mb() ends up being?

yes. All barriers should end up with mbar.

Stefan: Please correct me if I am wrong?

Thanks,
Michal