Re: [PATCH 2/3] MIPS: enforce LL-SC loop enclosing with SYNC (ACQUIRE and RELEASE)

[Leonid Yegoshin <Leonid.Yegoshin@imgtec.com>]

[-- Attachment #1: Type: text/plain, Size: 2909 bytes --]

On 06/02/2015 04:39 AM, James Hogan wrote:
> Hi Leonid,
>
> On 02/06/15 01:09, Leonid Yegoshin wrote:
>
>> CPUs may occasionally have problems in accordance with HW team.
> "have problems in accordance with HW team" is a bit confusing. What do
> you mean?

I wrote about memory barriers and problems may happens without it.
Some details from internal e-mail exchange:

-------------------------------------------

yes, it is possible for the stores to be observed out of order
The SC B can complete if it has ownership and the SW A is still
waiting for ownership.  This scenario was also possible on older
cores.

(snip)

-----Original Message-----
From: Leonid Yegoshin
Subject: more SYNC issues

...

I want to know - do XXXX orders stores without SYNC?

Let's consider a scenario:

SW    $0, A (cache miss)
...
LL    $1, B (cache hit)
..
SC    $1, B (cache hit again)
B        (is not taken!)
..
SYNC  0x10

Is it possible for other core to get new value of B before new value of A between SC-B and SYNC?
-------------------------------------------------------------

Another mail, another processor:

==========================================

Hi Leonid

I looked into the LSU RTL and I do not see speculation being blocked for 
younger loads following LL.

I also ran a testcase to confirm my observation:

LL  (cacheable)Miss

LW (cacheable)Miss, different cacheline

Miss request for LW goes out before the Miss request for LL, also the 
GPR updated for LW happens before the LL GPR update.
==========================================


> Actually *true*? P5600 you mean? Same in Kconfig help text. 

Yes, thank you, it is P5600 and I5600 doesn't exist.


> It feels wrong to be giving the user this option. Can't we just select
> WEAK_REORDERING_BEYOND_LLSC automatically based on the hardware that
> needs to be supported by the kernel configuration (e.g. CPU_MIPSR6 or
> CPU_MIPS32_R2)?

No, we can't - a lot of old processors are in-order and all of that is 
still MIPS R2.

>   Those who care about mips r2 performance on hardware
> which doesn't strictly need it can always speak up / add an exception.
>
> Out of interest, are futex operations safe with weak llsc ordering, on
> the premise that they're mainly used by userland so ordering with normal
> kernel accesses just doesn't matter in practice?

I think futex is used to communicate between user threads and problem is 
theoretically still here.

> This patch does 3 logically separable things:
> 1) add smp_release/smp_acquire based on MIPS_LIGHTWEIGHT_SYNC and weaken
> smp_store_release()/smp_load_acquire() to use them.
> 2) weaken llsc barriers when MIPS_LIGHTWEIGHT_SYNC.
> 3) the MIPS_ENFORCE_WEAK_REORDERING_BEYOND_LLSC Kconfig stuff (or
> whatever method to select WEAK_REORDERING_BEYOND_LLSC more often).
>
> Any reason not to split them, and give a clear description of each?
>
>

I don't see a reason to split it.

- Leonid.

[-- Attachment #2: Type: text/html, Size: 5842 bytes --]