[SPARC64]: Avoid membar instructions in delay slots.

In particular, avoid membar instructions in the delay
slot of a jmpl instruction.

UltraSPARC-I, II, IIi, and IIe have a bug, documented in
the UltraSPARC-IIi User's Manual, Appendix K, Erratum 51

The long and short of it is that if the IMU unit misses
on a branch or jmpl, and there is a store buffer synchronizing
membar in the delay slot, the chip can stop fetching instructions.

If interrupts are enabled or some other trap is enabled, the
chip will unwedge itself, but performance will suffer.

We already had a workaround for this bug in a few spots, but
it's better to have the entire tree sanitized for this rule.

Signed-off-by: David S. Miller <davem@davemloft.net>

3 files changed, 21 insertions, 12 deletions

diff --git a/include/asm-sparc64/rwsem.h b/include/asm-sparc64/rwsem.h
index bf2ae90ed3df..a1cc94f95984 100644
--- a/include/asm-sparc64/rwsem.h
+++ b/include/asm-sparc64/rwsem.h
@@ -55,8 +55,9 @@ static __inline__ int rwsem_atomic_update(int delta, struct rw_semaphore *sem)
                 "add            %%g1, %1, %%g7\n\t"
                 "cas            [%2], %%g1, %%g7\n\t"
                 "cmp            %%g1, %%g7\n\t"
+                "membar         #StoreLoad | #StoreStore\n\t"
                 "bne,pn         %%icc, 1b\n\t"
-                " membar        #StoreLoad | #StoreStore\n\t"
+                " nop\n\t"
                 "mov            %%g7, %0\n\t"
                 : "=&r" (tmp)
                 : "0" (tmp), "r" (sem)
diff --git a/include/asm-sparc64/spinlock.h b/include/asm-sparc64/spinlock.h
index db7581bdb531..9cb93a5c2b4f 100644
--- a/include/asm-sparc64/spinlock.h
+++ b/include/asm-sparc64/spinlock.h
@@ -52,12 +52,14 @@ static inline void _raw_spin_lock(spinlock_t *lock)
 
         __asm__ __volatile__(
 "1:     ldstub          [%1], %0\n"
+"       membar          #StoreLoad | #StoreStore\n"
 "       brnz,pn         %0, 2f\n"
-"        membar         #StoreLoad | #StoreStore\n"
+"        nop\n"
 "       .subsection     2\n"
 "2:     ldub            [%1], %0\n"
+"       membar          #LoadLoad\n"
 "       brnz,pt         %0, 2b\n"
-"        membar         #LoadLoad\n"
+"        nop\n"
 "       ba,a,pt         %%xcc, 1b\n"
 "       .previous"
         : "=&r" (tmp)
@@ -95,16 +97,18 @@ static inline void _raw_spin_lock_flags(spinlock_t *lock, unsigned long flags)
 
         __asm__ __volatile__(
 "1:     ldstub          [%2], %0\n"
-"       brnz,pn         %0, 2f\n"
 "       membar          #StoreLoad | #StoreStore\n"
+"       brnz,pn         %0, 2f\n"
+"        nop\n"
 "       .subsection     2\n"
 "2:     rdpr            %%pil, %1\n"
 "       wrpr            %3, %%pil\n"
 "3:     ldub            [%2], %0\n"
-"       brnz,pt         %0, 3b\n"
 "       membar          #LoadLoad\n"
+"       brnz,pt         %0, 3b\n"
+"        nop\n"
 "       ba,pt           %%xcc, 1b\n"
-"       wrpr            %1, %%pil\n"
+"        wrpr           %1, %%pil\n"
 "       .previous"
         : "=&r" (tmp1), "=&r" (tmp2)
         : "r"(lock), "r"(flags)
@@ -162,12 +166,14 @@ static void inline __read_lock(rwlock_t *lock)
 "4:      add            %0, 1, %1\n"
 "       cas             [%2], %0, %1\n"
 "       cmp             %0, %1\n"
+"       membar          #StoreLoad | #StoreStore\n"
 "       bne,pn          %%icc, 1b\n"
-"        membar         #StoreLoad | #StoreStore\n"
+"        nop\n"
 "       .subsection     2\n"
 "2:     ldsw            [%2], %0\n"
+"       membar          #LoadLoad\n"
 "       brlz,pt         %0, 2b\n"
-"        membar         #LoadLoad\n"
+"        nop\n"
 "       ba,a,pt         %%xcc, 4b\n"
 "       .previous"
         : "=&r" (tmp1), "=&r" (tmp2)
@@ -204,12 +210,14 @@ static void inline __write_lock(rwlock_t *lock)
 "4:      or             %0, %3, %1\n"
 "       cas             [%2], %0, %1\n"
 "       cmp             %0, %1\n"
+"       membar          #StoreLoad | #StoreStore\n"
 "       bne,pn          %%icc, 1b\n"
-"        membar         #StoreLoad | #StoreStore\n"
+"        nop\n"
 "       .subsection     2\n"
 "2:     lduw            [%2], %0\n"
+"       membar          #LoadLoad\n"
 "       brnz,pt         %0, 2b\n"
-"        membar         #LoadLoad\n"
+"        nop\n"
 "       ba,a,pt         %%xcc, 4b\n"
 "       .previous"
         : "=&r" (tmp1), "=&r" (tmp2)
@@ -240,8 +248,9 @@ static int inline __write_trylock(rwlock_t *lock)
 "        or             %0, %4, %1\n"
 "       cas             [%3], %0, %1\n"
 "       cmp             %0, %1\n"
+"       membar          #StoreLoad | #StoreStore\n"
 "       bne,pn          %%icc, 1b\n"
-"        membar         #StoreLoad | #StoreStore\n"
+"        nop\n"
 "       mov             1, %2\n"
 "2:"
         : "=&r" (tmp1), "=&r" (tmp2), "=&r" (result)
diff --git a/include/asm-sparc64/spitfire.h b/include/asm-sparc64/spitfire.h
index 9d7613eea812..1aa932773af8 100644
--- a/include/asm-sparc64/spitfire.h
+++ b/include/asm-sparc64/spitfire.h
@@ -111,7 +111,6 @@ static __inline__ void spitfire_put_dcache_tag(unsigned long addr, unsigned long
                              "membar    #Sync"
                              : /* No outputs */
                              : "r" (tag), "r" (addr), "i" (ASI_DCACHE_TAG));
-        __asm__ __volatile__ ("membar #Sync" : : : "memory");
 }
 
 /* The instruction cache lines are flushed with this, but note that