[SPARC64]: More fully work around Spitfire Errata 51.

It appears that a memory barrier soon after a mispredicted
branch, not just in the delay slot, can cause the hang
condition of this cpu errata.

So move them out-of-line, and explicitly put them into
a "branch always, predict taken" delay slot which should
fully kill this problem.

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

3 files changed, 82 insertions, 9 deletions

diff --git a/arch/sparc64/lib/Makefile b/arch/sparc64/lib/Makefile
index 40dbeec..6201f10 100644
--- a/arch/sparc64/lib/Makefile
+++ b/arch/sparc64/lib/Makefile
@@ -12,7 +12,7 @@ lib-y := PeeCeeI.o copy_page.o clear_page.o strlen.o strncmp.o \
 	 U1memcpy.o U1copy_from_user.o U1copy_to_user.o \
 	 U3memcpy.o U3copy_from_user.o U3copy_to_user.o U3patch.o \
 	 copy_in_user.o user_fixup.o memmove.o \
-	 mcount.o ipcsum.o rwsem.o xor.o find_bit.o delay.o
+	 mcount.o ipcsum.o rwsem.o xor.o find_bit.o delay.o mb.o
 
 lib-$(CONFIG_DEBUG_SPINLOCK) += debuglocks.o
 lib-$(CONFIG_HAVE_DEC_LOCK) += dec_and_lock.o
diff --git a/arch/sparc64/lib/debuglocks.c b/arch/sparc64/lib/debuglocks.c
index 7f6ccc4..f5f0b55 100644
--- a/arch/sparc64/lib/debuglocks.c
+++ b/arch/sparc64/lib/debuglocks.c
@@ -61,7 +61,7 @@ again:
 			     : "=r" (val)
 			     : "r" (&(lock->lock))
 			     : "memory");
-	membar("#StoreLoad | #StoreStore");
+	membar_storeload_storestore();
 	if (val) {
 		while (lock->lock) {
 			if (!--stuck) {
@@ -69,7 +69,7 @@ again:
 					show(str, lock, caller);
 				stuck = INIT_STUCK;
 			}
-			membar("#LoadLoad");
+			rmb();
 		}
 		goto again;
 	}
@@ -90,7 +90,7 @@ int _do_spin_trylock(spinlock_t *lock, unsigned long caller)
 			     : "=r" (val)
 			     : "r" (&(lock->lock))
 			     : "memory");
-	membar("#StoreLoad | #StoreStore");
+	membar_storeload_storestore();
 	if (!val) {
 		lock->owner_pc = ((unsigned int)caller);
 		lock->owner_cpu = cpu;
@@ -107,7 +107,7 @@ void _do_spin_unlock(spinlock_t *lock)
 {
 	lock->owner_pc = 0;
 	lock->owner_cpu = NO_PROC_ID;
-	membar("#StoreStore | #LoadStore");
+	membar_storestore_loadstore();
 	lock->lock = 0;
 	current->thread.smp_lock_count--;
 }
@@ -129,7 +129,7 @@ wlock_again:
 				show_read(str, rw, caller);
 			stuck = INIT_STUCK;
 		}
-		membar("#LoadLoad");
+		rmb();
 	}
 	/* Try once to increment the counter.  */
 	__asm__ __volatile__(
@@ -142,7 +142,7 @@ wlock_again:
 "2:"	: "=r" (val)
 	: "0" (&(rw->lock))
 	: "g1", "g7", "memory");
-	membar("#StoreLoad | #StoreStore");
+	membar_storeload_storestore();
 	if (val)
 		goto wlock_again;
 	rw->reader_pc[cpu] = ((unsigned int)caller);
@@ -201,7 +201,7 @@ wlock_again:
 				show_write(str, rw, caller);
 			stuck = INIT_STUCK;
 		}
-		membar("#LoadLoad");
+		rmb();
 	}
 
 	/* Try to acuire the write bit.  */
@@ -256,7 +256,7 @@ wlock_again:
 					show_write(str, rw, caller);
 				stuck = INIT_STUCK;
 			}
-			membar("#LoadLoad");
+			rmb();
 		}
 		goto wlock_again;
 	}
diff --git a/arch/sparc64/lib/mb.S b/arch/sparc64/lib/mb.S
new file mode 100644
index 0000000..4004f74
--- /dev/null
+++ b/arch/sparc64/lib/mb.S
@@ -0,0 +1,73 @@
+/* mb.S: Out of line memory barriers.
+ *
+ * Copyright (C) 2005 David S. Miller (davem@davemloft.net)
+ */
+
+	/* These are here in an effort to more fully work around
+	 * Spitfire Errata #51.  Essentially, if a memory barrier
+	 * occurs soon after a mispredicted branch, the chip can stop
+	 * executing instructions until a trap occurs.  Therefore, if
+	 * interrupts are disabled, the chip can hang forever.
+	 *
+	 * It used to be believed that the memory barrier had to be
+	 * right in the delay slot, but a case has been traced
+	 * recently wherein the memory barrier was one instruction
+	 * after the branch delay slot and the chip still hung.  The
+	 * offending sequence was the following in sym_wakeup_done()
+	 * of the sym53c8xx_2 driver:
+	 *
+	 *	call	sym_ccb_from_dsa, 0
+	 *	 movge	%icc, 0, %l0
+	 *	brz,pn	%o0, .LL1303
+	 *	 mov	%o0, %l2
+	 *	membar	#LoadLoad
+	 *
+	 * The branch has to be mispredicted for the bug to occur.
+	 * Therefore, we put the memory barrier explicitly into a
+	 * "branch always, predicted taken" delay slot to avoid the
+	 * problem case.
+	 */
+
+	.text
+
+99:	retl
+	 nop
+
+	.globl	mb
+mb:	ba,pt	%xcc, 99b
+	 membar	#LoadLoad | #LoadStore | #StoreStore | #StoreLoad
+	.size	mb, .-mb
+
+	.globl	rmb
+rmb:	ba,pt	%xcc, 99b
+	 membar	#LoadLoad
+	.size	rmb, .-rmb
+
+	.globl	wmb
+wmb:	ba,pt	%xcc, 99b
+	 membar	#StoreStore
+	.size	wmb, .-wmb
+
+	.globl	membar_storeload
+membar_storeload:
+	ba,pt	%xcc, 99b
+	 membar	#StoreLoad
+	.size	membar_storeload, .-membar_storeload
+
+	.globl	membar_storeload_storestore
+membar_storeload_storestore:
+	ba,pt	%xcc, 99b
+	 membar	#StoreLoad | #StoreStore
+	.size	membar_storeload_storestore, .-membar_storeload_storestore
+
+	.globl	membar_storeload_loadload
+membar_storeload_loadload:
+	ba,pt	%xcc, 99b
+	 membar	#StoreLoad | #LoadLoad
+	.size	membar_storeload_loadload, .-membar_storeload_loadload
+
+	.globl	membar_storestore_loadstore
+membar_storestore_loadstore:
+	ba,pt	%xcc, 99b
+	 membar	#StoreStore | #LoadStore
+	.size	membar_storestore_loadstore, .-membar_storestore_loadstore