On Thu, Aug 30, 2018 at 08:31:59PM +0000, Vineet Gupta wrote:

On 08/30/2018 07:29 AM, Peter Zijlstra wrote:

quoted

On Thu, Aug 30, 2018 at 03:23:55PM +0100, Will Deacon wrote:

quoted

Yes, that would be worth trying. However, I also just noticed that the
fetch-ops (which are now used to implement test_and_set_bit_lock()) seem
to be missing the backwards branch in the LL/SC case. Yet another diff
below.

Will

--->8

diff --git a/arch/arc/include/asm/atomic.h b/arch/arc/include/asm/atomic.h
index 4e0072730241..f06c5ed672b3 100644
--- a/arch/arc/include/asm/atomic.h
+++ b/arch/arc/include/asm/atomic.h

@@ -84,7 +84,7 @@ static inline int atomic_fetch_##op(int i, atomic_t *v)			\
 	"1:	llock   %[orig], [%[ctr]]		\n"		\
 	"	" #asm_op " %[val], %[orig], %[i]	\n"		\
 	"	scond   %[val], [%[ctr]]		\n"		\
-	"						\n"		\
+	"	bnz	1b				\n"		\
 	: [val]	"=&r"	(val),						\
 	  [orig] "=&r" (orig)						\
 	: [ctr]	"r"	(&v->counter),					\

ACK!! sorry about that, no idea how I messed that up.

Also, once it all works, they should look at switching to _relaxed
atomics for LL/SC.

Indeed this is the mother of all issues, I tried and system is clearly hosed with
and works after.
What's amazing is the commit 4aef66c8ae9 which introduced it is from 2016 ;-)
Back then we had a retry branch with backoff stuff which I'd reverted for new
cores and the merge conflict somehow missed it.

@PeterZ I'll create a patch with you as author ? do I need any formal sign offs,
acks etc ?

Well, Will spotted it, give authorship to him, you have my ack per the
above.

So after this there are 2 other things to be addresses / looked at still while we
are still here.

1. After 84c6591103db __clear_bit_lock() implementation will be broken (or atleast
not consistent with what we had after), do we need to reinstate it.
2. Will's proposed change to remove the underlying issue, but the issue in #1
remains ?

No, like explained, for spinlock based atomics the issue _should_ not
exist, and if you look at your atomic_set() implementation for that
variant, you'll see it does the right thing by taking the lock.

Basically atomic_set() for spinlock based atomics ends up being
(void)atomic_xchg().

FWIW, also ACK on Will's patch to switch you over to asm-generic bitops
entirely.