ARC: add compiler barrier to LLSC based cmpxchg
authorVineet Gupta <vgupta@synopsys.com>
Thu, 13 Nov 2014 10:24:01 +0000 (15:54 +0530)
committerVineet Gupta <vgupta@synopsys.com>
Thu, 25 Jun 2015 00:29:23 +0000 (05:59 +0530)
When auditing cmpxchg call sites, Chuck noted that gcc was optimizing
away some of the desired LDs.

| do {
| new = old = *ipi_data_ptr;
| new |= 1U << msg;
| } while (cmpxchg(ipi_data_ptr, old, new) != old);

was generating to below

8015cef8: ld         r2,[r4,0]  <-- First LD
8015cefc: bset       r1,r2,r1
|
8015cf00: llock      r3,[r4]  <-- atomic op
8015cf04: brne       r3,r2,8015cf10
8015cf08: scond      r1,[r4]
8015cf0c: bnz        8015cf00
|
8015cf10: brne       r3,r2,8015cf00  <-- Branch doesn't go to orig LD

Although this was fixed by adding a ACCESS_ONCE in this call site, it
seems safer (for now at least) to add compiler barrier to LLSC based
cmpxchg

Reported-by: Chuck Jordan <cjordan@synopsys,com>
Cc: <stable@vger.kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
arch/arc/include/asm/cmpxchg.h

index 03cd689..90de5c5 100644 (file)
@@ -25,10 +25,11 @@ __cmpxchg(volatile void *ptr, unsigned long expected, unsigned long new)
        "       scond   %3, [%1]        \n"
        "       bnz     1b              \n"
        "2:                             \n"
-       : "=&r"(prev)
-       : "r"(ptr), "ir"(expected),
-         "r"(new) /* can't be "ir". scond can't take limm for "b" */
-       : "cc");
+       : "=&r"(prev)   /* Early clobber, to prevent reg reuse */
+       : "r"(ptr),     /* Not "m": llock only supports reg direct addr mode */
+         "ir"(expected),
+         "r"(new)      /* can't be "ir". scond can't take LIMM for "b" */
+       : "cc", "memory"); /* so that gcc knows memory is being written here */
 
        return prev;
 }