xref: /OK3568_Linux_fs/kernel/arch/sparc/include/asm/barrier_64.h (revision 4882a59341e53eb6f0b4789bf948001014eff981)
1*4882a593Smuzhiyun /* SPDX-License-Identifier: GPL-2.0 */
2*4882a593Smuzhiyun #ifndef __SPARC64_BARRIER_H
3*4882a593Smuzhiyun #define __SPARC64_BARRIER_H
4*4882a593Smuzhiyun 
5*4882a593Smuzhiyun /* These are here in an effort to more fully work around Spitfire Errata
6*4882a593Smuzhiyun  * #51.  Essentially, if a memory barrier occurs soon after a mispredicted
7*4882a593Smuzhiyun  * branch, the chip can stop executing instructions until a trap occurs.
8*4882a593Smuzhiyun  * Therefore, if interrupts are disabled, the chip can hang forever.
9*4882a593Smuzhiyun  *
10*4882a593Smuzhiyun  * It used to be believed that the memory barrier had to be right in the
11*4882a593Smuzhiyun  * delay slot, but a case has been traced recently wherein the memory barrier
12*4882a593Smuzhiyun  * was one instruction after the branch delay slot and the chip still hung.
13*4882a593Smuzhiyun  * The offending sequence was the following in sym_wakeup_done() of the
14*4882a593Smuzhiyun  * sym53c8xx_2 driver:
15*4882a593Smuzhiyun  *
16*4882a593Smuzhiyun  *	call	sym_ccb_from_dsa, 0
17*4882a593Smuzhiyun  *	 movge	%icc, 0, %l0
18*4882a593Smuzhiyun  *	brz,pn	%o0, .LL1303
19*4882a593Smuzhiyun  *	 mov	%o0, %l2
20*4882a593Smuzhiyun  *	membar	#LoadLoad
21*4882a593Smuzhiyun  *
22*4882a593Smuzhiyun  * The branch has to be mispredicted for the bug to occur.  Therefore, we put
23*4882a593Smuzhiyun  * the memory barrier explicitly into a "branch always, predicted taken"
24*4882a593Smuzhiyun  * delay slot to avoid the problem case.
25*4882a593Smuzhiyun  */
26*4882a593Smuzhiyun #define membar_safe(type) \
27*4882a593Smuzhiyun do {	__asm__ __volatile__("ba,pt	%%xcc, 1f\n\t" \
28*4882a593Smuzhiyun 			     " membar	" type "\n" \
29*4882a593Smuzhiyun 			     "1:\n" \
30*4882a593Smuzhiyun 			     : : : "memory"); \
31*4882a593Smuzhiyun } while (0)
32*4882a593Smuzhiyun 
33*4882a593Smuzhiyun /* The kernel always executes in TSO memory model these days,
34*4882a593Smuzhiyun  * and furthermore most sparc64 chips implement more stringent
35*4882a593Smuzhiyun  * memory ordering than required by the specifications.
36*4882a593Smuzhiyun  */
37*4882a593Smuzhiyun #define mb()	membar_safe("#StoreLoad")
38*4882a593Smuzhiyun #define rmb()	__asm__ __volatile__("":::"memory")
39*4882a593Smuzhiyun #define wmb()	__asm__ __volatile__("":::"memory")
40*4882a593Smuzhiyun 
41*4882a593Smuzhiyun #define __smp_store_release(p, v)						\
42*4882a593Smuzhiyun do {									\
43*4882a593Smuzhiyun 	compiletime_assert_atomic_type(*p);				\
44*4882a593Smuzhiyun 	barrier();							\
45*4882a593Smuzhiyun 	WRITE_ONCE(*p, v);						\
46*4882a593Smuzhiyun } while (0)
47*4882a593Smuzhiyun 
48*4882a593Smuzhiyun #define __smp_load_acquire(p)						\
49*4882a593Smuzhiyun ({									\
50*4882a593Smuzhiyun 	typeof(*p) ___p1 = READ_ONCE(*p);				\
51*4882a593Smuzhiyun 	compiletime_assert_atomic_type(*p);				\
52*4882a593Smuzhiyun 	barrier();							\
53*4882a593Smuzhiyun 	___p1;								\
54*4882a593Smuzhiyun })
55*4882a593Smuzhiyun 
56*4882a593Smuzhiyun #define __smp_mb__before_atomic()	barrier()
57*4882a593Smuzhiyun #define __smp_mb__after_atomic()	barrier()
58*4882a593Smuzhiyun 
59*4882a593Smuzhiyun #include <asm-generic/barrier.h>
60*4882a593Smuzhiyun 
61*4882a593Smuzhiyun #endif /* !(__SPARC64_BARRIER_H) */
62