sparc64: Store per-cpu offset in trap_block[]

Surprisingly this actually makes LOAD_PER_CPU_BASE() a little
more efficient.

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

5 files changed, 21 insertions, 44 deletions

diff --git a/arch/sparc/include/asm/percpu_64.h b/arch/sparc/include/asm/percpu_64.h
index bee64593023e..c0ab102d11f6 100644
--- a/arch/sparc/include/asm/percpu_64.h
+++ b/arch/sparc/include/asm/percpu_64.h
@@ -7,12 +7,12 @@ register unsigned long __local_per_cpu_offset asm("g5");
 
 #ifdef CONFIG_SMP
 
+#include <asm/trap_block.h>
+
 extern void real_setup_per_cpu_areas(void);
 
-extern unsigned long __per_cpu_base;
-extern unsigned long __per_cpu_shift;
 #define __per_cpu_offset(__cpu) \
-	(__per_cpu_base + ((unsigned long)(__cpu) << __per_cpu_shift))
+	(trap_block[(__cpu)].__per_cpu_base)
 #define per_cpu_offset(x) (__per_cpu_offset(x))
 
 #define __my_cpu_offset __local_per_cpu_offset
diff --git a/arch/sparc/include/asm/trap_block.h b/arch/sparc/include/asm/trap_block.h
index 68fd9ee3e8ae..7e26b2db6211 100644
--- a/arch/sparc/include/asm/trap_block.h
+++ b/arch/sparc/include/asm/trap_block.h
@@ -48,7 +48,7 @@ struct trap_per_cpu {
 	unsigned int		dev_mondo_qmask;
 	unsigned int		resum_qmask;
 	unsigned int		nonresum_qmask;
-	unsigned long		__unused;
+	unsigned long		__per_cpu_base;
 } __attribute__((aligned(64)));
 extern struct trap_per_cpu trap_block[NR_CPUS];
 extern void init_cur_cpu_trap(struct thread_info *);
@@ -101,6 +101,7 @@ extern struct sun4v_2insn_patch_entry __sun4v_2insn_patch,
 #define TRAP_PER_CPU_DEV_MONDO_QMASK	0xec
 #define TRAP_PER_CPU_RESUM_QMASK	0xf0
 #define TRAP_PER_CPU_NONRESUM_QMASK	0xf4
+#define TRAP_PER_CPU_PER_CPU_BASE	0xf8
 
 #define TRAP_BLOCK_SZ_SHIFT		8
 
@@ -172,12 +173,11 @@ extern struct sun4v_2insn_patch_entry __sun4v_2insn_patch,
  */
 #define LOAD_PER_CPU_BASE(DEST, THR, REG1, REG2, REG3)	\
 	lduh	[THR + TI_CPU], REG1;			\
-	sethi	%hi(__per_cpu_shift), REG3;		\
-	sethi	%hi(__per_cpu_base), REG2;		\
-	ldx	[REG3 + %lo(__per_cpu_shift)], REG3;	\
-	ldx	[REG2 + %lo(__per_cpu_base)], REG2;	\
-	sllx	REG1, REG3, REG3;			\
-	add	REG3, REG2, DEST;
+	sethi	%hi(trap_block), REG2;			\
+	sllx	REG1, TRAP_BLOCK_SZ_SHIFT, REG1;	\
+	or	REG2, %lo(trap_block), REG2;		\
+	add	REG2, REG1, REG2;			\
+	ldx	[REG2 + TRAP_PER_CPU_PER_CPU_BASE], DEST;
 
 #else
 
diff --git a/arch/sparc/kernel/head_64.S b/arch/sparc/kernel/head_64.S
index 91bf4c7f79b9..f8f21050448b 100644
--- a/arch/sparc/kernel/head_64.S
+++ b/arch/sparc/kernel/head_64.S
@@ -641,28 +641,6 @@ tlb_fixup_done:
 	/* Not reached... */
 
 1:
-	/* If we boot on a non-zero cpu, all of the per-cpu
-	 * variable references we make before setting up the
-	 * per-cpu areas will use a bogus offset.  Put a
-	 * compensating factor into __per_cpu_base to handle
-	 * this cleanly.
-	 *
-	 * What the per-cpu code calculates is:
-	 *
-	 *	__per_cpu_base + (cpu << __per_cpu_shift)
-	 *
-	 * These two variables are zero initially, so to
-	 * make it all cancel out to zero we need to put
-	 * "0 - (cpu << 0)" into __per_cpu_base so that the
-	 * above formula evaluates to zero.
-	 *
-	 * We cannot even perform a printk() until this stuff
-	 * is setup as that calls cpu_clock() which uses
-	 * per-cpu variables.
-	 */
-	sub	%g0, %o0, %o1
-	sethi	%hi(__per_cpu_base), %o2
-	stx	%o1, [%o2 + %lo(__per_cpu_base)]
 #else
 	mov	0, %o0
 #endif
diff --git a/arch/sparc/kernel/smp_64.c b/arch/sparc/kernel/smp_64.c
index 4226d0ebaea5..b20f253857b7 100644
--- a/arch/sparc/kernel/smp_64.c
+++ b/arch/sparc/kernel/smp_64.c
@@ -1371,23 +1371,17 @@ void smp_send_stop(void)
 {
 }
 
-unsigned long __per_cpu_base __read_mostly;
-unsigned long __per_cpu_shift __read_mostly;
-
-EXPORT_SYMBOL(__per_cpu_base);
-EXPORT_SYMBOL(__per_cpu_shift);
-
 void __init real_setup_per_cpu_areas(void)
 {
-	unsigned long paddr, goal, size, i;
+	unsigned long base, shift, paddr, goal, size, i;
 	char *ptr;
 
 	/* Copy section for each CPU (we discard the original) */
 	goal = PERCPU_ENOUGH_ROOM;
 
-	__per_cpu_shift = PAGE_SHIFT;
+	shift = PAGE_SHIFT;
 	for (size = PAGE_SIZE; size < goal; size <<= 1UL)
-		__per_cpu_shift++;
+		shift++;
 
 	paddr = lmb_alloc(size * NR_CPUS, PAGE_SIZE);
 	if (!paddr) {
@@ -1396,10 +1390,12 @@ void __init real_setup_per_cpu_areas(void)
 	}
 
 	ptr = __va(paddr);
-	__per_cpu_base = ptr - __per_cpu_start;
+	base = ptr - __per_cpu_start;
 
-	for (i = 0; i < NR_CPUS; i++, ptr += size)
+	for (i = 0; i < NR_CPUS; i++, ptr += size) {
+		__per_cpu_offset(i) = base + (i * size);
 		memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
+	}
 
 	/* Setup %g5 for the boot cpu.  */
 	__local_per_cpu_offset = __per_cpu_offset(smp_processor_id());
diff --git a/arch/sparc/kernel/traps_64.c b/arch/sparc/kernel/traps_64.c
index d809c4ebb48f..d073aabf65ed 100644
--- a/arch/sparc/kernel/traps_64.c
+++ b/arch/sparc/kernel/traps_64.c
@@ -2509,6 +2509,7 @@ void do_getpsr(struct pt_regs *regs)
 }
 
 struct trap_per_cpu trap_block[NR_CPUS];
+EXPORT_SYMBOL(trap_block);
 
 /* This can get invoked before sched_init() so play it super safe
  * and use hard_smp_processor_id().
@@ -2592,7 +2593,9 @@ void __init trap_init(void)
 	    (TRAP_PER_CPU_RESUM_QMASK !=
 	     offsetof(struct trap_per_cpu, resum_qmask)) ||
 	    (TRAP_PER_CPU_NONRESUM_QMASK !=
-	     offsetof(struct trap_per_cpu, nonresum_qmask)))
+	     offsetof(struct trap_per_cpu, nonresum_qmask)) ||
+	    (TRAP_PER_CPU_PER_CPU_BASE !=
+	     offsetof(struct trap_per_cpu, __per_cpu_base)))
 		trap_per_cpu_offsets_are_bolixed_dave();
 
 	if ((TSB_CONFIG_TSB !=