/*
** ml_get_timebase()
**
-** Entry - %rdi contains pointer to 64 bit structure.
-**
-** Exit - 64 bit structure filled in.
+** Returns TSC in RAX
**
*/
ENTRY(ml_get_timebase)
lfence
shlq $32,%rdx
orq %rdx,%rax
- movq %rax, (%rdi)
ret
*
*/
ENTRY(tmrCvt)
+ cmpq $1,%rsi /* check for unity fastpath */
+ je 1f
movq %rdi,%rax
mulq %rsi /* result is %rdx:%rax */
shrdq $32,%rdx,%rax /* %rdx:%rax >>= 32 */
ret
+1:
+ mov %rdi,%rax
+ ret
/*
* void _rtc_nanotime_adjust(
ret
/*
- * unint64_t _rtc_nanotime_read(rtc_nanotime_t *rntp, int slow);
+ * uint64_t _rtc_nanotime_read(rtc_nanotime_t *rntp);
*
* This is the same as the commpage nanotime routine, except that it uses the
* kernel internal "rtc_nanotime_info" data instead of the commpage data.
* These two copies of data are kept in sync by rtc_clock_napped().
*
- * Warning! There is another copy of this code in osfmk/x86_64/idt64.s.
- * These are kept in sync by both using the RTC_NANOTIME_READ() macro.
+ * Warning! There are several copies of this code in the trampolines found in
+ * osfmk/x86_64/idt64.s, coming from the various TIMER macros in rtclock_asm.h.
+ * They're all kept in sync by using the RTC_NANOTIME_READ() macro.
+ *
+ * The algorithm we use is:
+ *
+ * ns = ((((rdtsc - rnt_tsc_base)<<rnt_shift)*rnt_tsc_scale) / 2**32) + rnt_ns_base;
+ *
+ * rnt_shift, a constant computed during initialization, is the smallest value for which:
*
- * There are two versions of this algorithm, for "slow" and "fast" processors.
- * The more common "fast" algorithm is:
+ * (tscFreq << rnt_shift) > SLOW_TSC_THRESHOLD
*
- * ns = (((rdtsc - rnt_tsc_base)*rnt_tsc_scale) / 2**32) + rnt_ns_base;
+ * Where SLOW_TSC_THRESHOLD is about 10e9. Since most processor's tscFreqs are greater
+ * than 1GHz, rnt_shift is usually 0. rnt_tsc_scale is also a 32-bit constant:
*
- * Of course, the divide by 2**32 is a nop. rnt_tsc_scale is a constant
- * computed during initialization:
+ * rnt_tsc_scale = (10e9 * 2**32) / (tscFreq << rnt_shift);
*
- * rnt_tsc_scale = (10e9 * 2**32) / tscFreq;
+ * On 64-bit processors this algorithm could be simplified by doing a 64x64 bit
+ * multiply of rdtsc by tscFCvtt2n:
*
- * The "slow" algorithm uses long division:
+ * ns = (((rdtsc - rnt_tsc_base) * tscFCvtt2n) / 2**32) + rnt_ns_base;
*
- * ns = (((rdtsc - rnt_tsc_base) * 10e9) / tscFreq) + rnt_ns_base;
+ * We don't do so in order to use the same algorithm in 32- and 64-bit mode.
+ * When U32 goes away, we should reconsider.
*
* Since this routine is not synchronized and can be called in any context,
* we use a generation count to guard against seeing partially updated data.
* the generation is zero.
*
* unint64_t _rtc_nanotime_read(
- * rtc_nanotime_t *rntp, // %rdi
- * int slow); // %rsi
+ * rtc_nanotime_t *rntp); // %rdi
*
*/
ENTRY(_rtc_nanotime_read)
- test %rsi,%rsi
- jnz Lslow
-
- /*
- * Processor whose TSC frequency is faster than SLOW_TSC_THRESHOLD
- */
+
PAL_RTC_NANOTIME_READ_FAST()
ret
+
+/*
+ * extern uint64_t _rtc_tsc_to_nanoseconds(
+ * uint64_t value, // %rdi
+ * pal_rtc_nanotime_t *rntp); // %rsi
+ *
+ * Converts TSC units to nanoseconds, using an abbreviated form of the above
+ * algorithm. Note that while we could have simply used tmrCvt(value,tscFCvtt2n),
+ * which would avoid the need for this asm, doing so is a bit more risky since
+ * we'd be using a different algorithm with possibly different rounding etc.
+ */
- /*
- * Processor whose TSC frequency is not faster than SLOW_TSC_THRESHOLD
- * But K64 doesn't support this...
- */
-Lslow:
- lea 1f(%rip),%rdi
- xorb %al,%al
- call EXT(panic)
- hlt
- .data
-1: String "_rtc_nanotime_read() - slow algorithm not supported"
- .text
+ENTRY(_rtc_tsc_to_nanoseconds)
+ movq %rdi,%rax /* copy value (in TSC units) to convert */
+ movl RNT_SHIFT(%rsi),%ecx
+ movl RNT_SCALE(%rsi),%edx
+ shlq %cl,%rax /* tscUnits << shift */
+ mulq %rdx /* (tscUnits << shift) * scale */
+ shrdq $32,%rdx,%rax /* %rdx:%rax >>= 32 */
+ ret
+
+
Entry(call_continuation)
movq %rdi,%rcx /* get continuation */
movq %rsi, %rsp
callq *%rdi
- /*
- * Generate a 64-bit quantity with possibly random characteristics, intended for use
- * before the kernel entropy pool is available. The processor's RNG is used if
- * available, and a value derived from the Time Stamp Counter is returned if not.
- * Multiple invocations may result in well-correlated values if sourced from the TSC.
- */
-Entry(ml_early_random)
- mov %rbx, %rsi
- mov $1, %eax
- cpuid
- mov %rsi, %rbx
- test $(1 << 30), %ecx
- jz Lnon_rdrand
- RDRAND_RAX /* RAX := 64 bits of DRBG entropy */
- jnc Lnon_rdrand
- ret
-Lnon_rdrand:
- rdtsc /* EDX:EAX := TSC */
- /* Distribute low order bits */
- mov %eax, %ecx
- xor %al, %ah
- shl $16, %rcx
- xor %rcx, %rax
- xor %eax, %edx
-
- /* Incorporate ASLR entropy, if any */
- lea (%rip), %rcx
- shr $21, %rcx
- movzbl %cl, %ecx
- shl $16, %ecx
- xor %ecx, %edx
-
- mov %ah, %cl
- ror %cl, %edx /* Right rotate EDX (TSC&0xFF ^ (TSC>>8 & 0xFF))&1F */
- shl $32, %rdx
- xor %rdx, %rax
- mov %cl, %al
- ret
-
#if CONFIG_VMX
/*
ret
#endif /* CONFIG_VMX */
+
+/*
+ * mfence -- Memory Barrier
+ * Use out-of-line assembly to get
+ * standard x86-64 ABI guarantees
+ * about what the caller's codegen
+ * has in registers vs. memory
+ */
+Entry(do_mfence)
+ mfence
+ ret
projects / apple / xnu.git / blobdiff