PPC: Try out if mftb works before using it

If this fails try out if mfspr268 works.

Use OPENSSL_ppccap=0x20 for enabling mftb,
OPENSSL_ppccap=0x40 for enabling mfspr268,
and OPENSSL_ppccap=0 for enabling neither.

Fixes #8012
Reviewed-by: NPaul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/8043)

(cherry picked from commit c8f370485c43729db44b680e41e875ddd7f3108c)

crypto/cryptlib.c

@@ -460,4 +460,14 @@ uint32_t OPENSSL_rdtsc(void)
 {
     return 0;
 }
+
+size_t OPENSSL_instrument_bus(unsigned int *out, size_t cnt)
+{
+    return 0;
+}
+
+size_t OPENSSL_instrument_bus2(unsigned int *out, size_t cnt, size_t max)
+{
+    return 0;
+}
 #endif

crypto/ppc_arch.h

@@ -22,5 +22,7 @@ extern unsigned int OPENSSL_ppccap_P;
 # define PPC_CRYPTO207   (1<<2)
 # define PPC_FPU         (1<<3)
 # define PPC_MADD300     (1<<4)
+# define PPC_MFTB        (1<<5)
+# define PPC_MFSPR268    (1<<6)
 
 #endif

crypto/ppccap.c

@@ -168,6 +168,45 @@ void OPENSSL_altivec_probe(void);
 void OPENSSL_crypto207_probe(void);
 void OPENSSL_madd300_probe(void);
 
+long OPENSSL_rdtsc_mftb(void);
+long OPENSSL_rdtsc_mfspr268(void);
+
+uint32_t OPENSSL_rdtsc(void)
+{
+    if (OPENSSL_ppccap_P & PPC_MFTB)
+        return OPENSSL_rdtsc_mftb();
+    else if (OPENSSL_ppccap_P & PPC_MFSPR268)
+        return OPENSSL_rdtsc_mfspr268();
+    else
+        return 0;
+}
+
+size_t OPENSSL_instrument_bus_mftb(unsigned int *, size_t);
+size_t OPENSSL_instrument_bus_mfspr268(unsigned int *, size_t);
+
+size_t OPENSSL_instrument_bus(unsigned int *out, size_t cnt)
+{
+    if (OPENSSL_ppccap_P & PPC_MFTB)
+        return OPENSSL_instrument_bus_mftb(out, cnt);
+    else if (OPENSSL_ppccap_P & PPC_MFSPR268)
+        return OPENSSL_instrument_bus_mfspr268(out, cnt);
+    else
+        return 0;
+}
+
+size_t OPENSSL_instrument_bus2_mftb(unsigned int *, size_t, size_t);
+size_t OPENSSL_instrument_bus2_mfspr268(unsigned int *, size_t, size_t);
+
+size_t OPENSSL_instrument_bus2(unsigned int *out, size_t cnt, size_t max)
+{
+    if (OPENSSL_ppccap_P & PPC_MFTB)
+        return OPENSSL_instrument_bus2_mftb(out, cnt, max);
+    else if (OPENSSL_ppccap_P & PPC_MFSPR268)
+        return OPENSSL_instrument_bus2_mfspr268(out, cnt, max);
+    else
+        return 0;
+}
+
 #if defined(__GLIBC__) && defined(__GLIBC_PREREQ)
 # if __GLIBC_PREREQ(2, 16)
 #  include <sys/auxv.h>
@@ -300,8 +339,6 @@ void OPENSSL_cpuid_setup(void)
         if (hwcap & HWCAP_ARCH_3_00) {
             OPENSSL_ppccap_P |= PPC_MADD300;
         }
-
-        return;
     }
 #endif
 
@@ -322,15 +359,16 @@ void OPENSSL_cpuid_setup(void)
     sigprocmask(SIG_SETMASK, &ill_act.sa_mask, &oset);
     sigaction(SIGILL, &ill_act, &ill_oact);
 
+#ifndef OSSL_IMPLEMENT_GETAUXVAL
     if (sigsetjmp(ill_jmp,1) == 0) {
         OPENSSL_fpu_probe();
         OPENSSL_ppccap_P |= PPC_FPU;
 
         if (sizeof(size_t) == 4) {
-#ifdef __linux
+# ifdef __linux
             struct utsname uts;
             if (uname(&uts) == 0 && strcmp(uts.machine, "ppc64") == 0)
-#endif
+# endif
                 if (sigsetjmp(ill_jmp, 1) == 0) {
                     OPENSSL_ppc64_probe();
                     OPENSSL_ppccap_P |= PPC_FPU64;
@@ -355,6 +393,15 @@ void OPENSSL_cpuid_setup(void)
         OPENSSL_madd300_probe();
         OPENSSL_ppccap_P |= PPC_MADD300;
     }
+#endif
+
+    if (sigsetjmp(ill_jmp, 1) == 0) {
+        OPENSSL_rdtsc_mftb();
+        OPENSSL_ppccap_P |= PPC_MFTB;
+    } else if (sigsetjmp(ill_jmp, 1) == 0) {
+        OPENSSL_rdtsc_mfspr268();
+        OPENSSL_ppccap_P |= PPC_MFSPR268;
+    }
 
     sigaction(SIGILL, &ill_oact, NULL);
     sigprocmask(SIG_SETMASK, &oset, NULL);

crypto/ppccpuid.pl

@@ -124,26 +124,23 @@ Ladd:	lwarx	r5,0,r3
 	.long	0
 .size	.OPENSSL_atomic_add,.-.OPENSSL_atomic_add
 
-.globl	.OPENSSL_rdtsc
+.globl	.OPENSSL_rdtsc_mftb
 .align	4
-.OPENSSL_rdtsc:
-___
-$code.=<<___	if ($flavour =~ /64/);
-	mftb	r3
-___
-$code.=<<___	if ($flavour !~ /64/);
-Loop_rdtsc:
-	mftbu	r5
+.OPENSSL_rdtsc_mftb:
 	mftb	r3
-	mftbu	r4
-	cmplw	r4,r5
-	bne	Loop_rdtsc
-___
-$code.=<<___;
 	blr
 	.long	0
 	.byte	0,12,0x14,0,0,0,0,0
-.size	.OPENSSL_rdtsc,.-.OPENSSL_rdtsc
+.size	.OPENSSL_rdtsc_mftb,.-.OPENSSL_rdtsc_mftb
+
+.globl	.OPENSSL_rdtsc_mfspr268
+.align	4
+.OPENSSL_rdtsc_mfspr268:
+	mfspr	r3,268
+	blr
+	.long	0
+	.byte	0,12,0x14,0,0,0,0,0
+.size	.OPENSSL_rdtsc_mfspr268,.-.OPENSSL_rdtsc_mfspr268
 
 .globl	.OPENSSL_cleanse
 .align	4
@@ -210,9 +207,9 @@ my ($tick,$lasttick)=("r6","r7");
 my ($diff,$lastdiff)=("r8","r9");
 
 $code.=<<___;
-.globl	.OPENSSL_instrument_bus
+.globl	.OPENSSL_instrument_bus_mftb
 .align	4
-.OPENSSL_instrument_bus:
+.OPENSSL_instrument_bus_mftb:
 	mtctr	$cnt
 
 	mftb	$lasttick		# collect 1st tick
@@ -240,11 +237,11 @@ Loop:	mftb	$tick
 	.long	0
 	.byte	0,12,0x14,0,0,0,2,0
 	.long	0
-.size	.OPENSSL_instrument_bus,.-.OPENSSL_instrument_bus
+.size	.OPENSSL_instrument_bus_mftb,.-.OPENSSL_instrument_bus_mftb
 
-.globl	.OPENSSL_instrument_bus2
+.globl	.OPENSSL_instrument_bus2_mftb
 .align	4
-.OPENSSL_instrument_bus2:
+.OPENSSL_instrument_bus2_mftb:
 	mr	r0,$cnt
 	slwi	$cnt,$cnt,2
 
@@ -292,7 +289,91 @@ Ldone2:
 	.long	0
 	.byte	0,12,0x14,0,0,0,3,0
 	.long	0
-.size	.OPENSSL_instrument_bus2,.-.OPENSSL_instrument_bus2
+.size	.OPENSSL_instrument_bus2_mftb,.-.OPENSSL_instrument_bus2_mftb
+
+.globl	.OPENSSL_instrument_bus_mfspr268
+.align	4
+.OPENSSL_instrument_bus_mfspr268:
+	mtctr	$cnt
+
+	mfspr	$lasttick,268		# collect 1st tick
+	li	$diff,0
+
+	dcbf	0,$out			# flush cache line
+	lwarx	$tick,0,$out		# load and lock
+	add	$tick,$tick,$diff
+	stwcx.	$tick,0,$out
+	stwx	$tick,0,$out
+
+Loop3:	mfspr	$tick,268
+	sub	$diff,$tick,$lasttick
+	mr	$lasttick,$tick
+	dcbf	0,$out			# flush cache line
+	lwarx	$tick,0,$out		# load and lock
+	add	$tick,$tick,$diff
+	stwcx.	$tick,0,$out
+	stwx	$tick,0,$out
+	addi	$out,$out,4		# ++$out
+	bdnz	Loop3
+
+	mr	r3,$cnt
+	blr
+	.long	0
+	.byte	0,12,0x14,0,0,0,2,0
+	.long	0
+.size	.OPENSSL_instrument_bus_mfspr268,.-.OPENSSL_instrument_bus_mfspr268
+
+.globl	.OPENSSL_instrument_bus2_mfspr268
+.align	4
+.OPENSSL_instrument_bus2_mfspr268:
+	mr	r0,$cnt
+	slwi	$cnt,$cnt,2
+
+	mfspr	$lasttick,268		# collect 1st tick
+	li	$diff,0
+
+	dcbf	0,$out			# flush cache line
+	lwarx	$tick,0,$out		# load and lock
+	add	$tick,$tick,$diff
+	stwcx.	$tick,0,$out
+	stwx	$tick,0,$out
+
+	mfspr	$tick,268		# collect 1st diff
+	sub	$diff,$tick,$lasttick
+	mr	$lasttick,$tick
+	mr	$lastdiff,$diff
+Loop4:
+	dcbf	0,$out			# flush cache line
+	lwarx	$tick,0,$out		# load and lock
+	add	$tick,$tick,$diff
+	stwcx.	$tick,0,$out
+	stwx	$tick,0,$out
+
+	addic.	$max,$max,-1
+	beq	Ldone4
+
+	mfspr	$tick,268
+	sub	$diff,$tick,$lasttick
+	mr	$lasttick,$tick
+	cmplw	7,$diff,$lastdiff
+	mr	$lastdiff,$diff
+
+	mfcr	$tick			# pull cr
+	not	$tick,$tick		# flip bits
+	rlwinm	$tick,$tick,1,29,29	# isolate flipped eq bit and scale
+
+	sub.	$cnt,$cnt,$tick		# conditional --$cnt
+	add	$out,$out,$tick		# conditional ++$out
+	bne	Loop4
+
+Ldone4:
+	srwi	$cnt,$cnt,2
+	sub	r3,r0,$cnt
+	blr
+	.long	0
+	.byte	0,12,0x14,0,0,0,3,0
+	.long	0
+.size	.OPENSSL_instrument_bus2_mfspr268,.-.OPENSSL_instrument_bus2_mfspr268
 ___
 }
 

include/internal/cryptlib.h

@@ -92,5 +92,7 @@ void *openssl_fopen(const char *filename, const char *mode);
 # endif
 
 uint32_t OPENSSL_rdtsc(void);
+size_t OPENSSL_instrument_bus(unsigned int *, size_t);
+size_t OPENSSL_instrument_bus2(unsigned int *, size_t, size_t);
 
 #endif