Bananymous/banan-os
1
1
Fork 1
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Kernel/LibC: support RDTSCP based clock_gettime

Browse Source 
RDTSCP seems to be faster than LSL and it removes the need for getting
the current cpu twice to make sure TSC is read on the correct CPU
This commit is contained in:
Bananymous
2026-07-03 05:27:54 +03:00
parent 917321eb5f
commit 63b2284324
5 changed files with 86 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files

104
userspace/libraries/LibC/time.cpp
View File

@@ -1,3 +1,4 @@
#include <BAN/Atomic.h>
#include <BAN/Assert.h>
#include <BAN/Debug.h>
#include <BAN/Math.h>
@@ -28,53 +29,68 @@ int clock_gettime(clockid_t clock_id, struct timespec* tp)
if (g_shared_page == nullptr || !(g_shared_page->features & Kernel::API::SPF_GETTIME))
return syscall(SYS_CLOCK_GETTIME, clock_id, tp);
const auto get_cpu =
[]() -> uint8_t {
uint16_t limit;
asm volatile("lsl %1, %0" : "=r"(limit) : "r"(g_shared_page->gdt_cpu_offset));
return limit;
};
uint32_t mult;
int8_t shift;
uint64_t last_ns;
uint64_t last_tsc;
uint64_t curr_tsc;
for (;;)
const auto read_tsc_info = [&](volatile decltype(g_shared_page->cpus[0].gettime_local)& lgettime) {
uint32_t seq1, seq2;
do {
seq1 = BAN::atomic_load(lgettime.seq, BAN::memory_order_acquire);
mult = lgettime.mult;
shift = lgettime.shift;
last_ns = lgettime.last_ns;
last_tsc = lgettime.last_tsc;
seq2 = BAN::atomic_load(lgettime.seq, BAN::memory_order_acquire);
} while (seq1 != seq2 || (seq1 & 1));
};
read_tsc_info:
if (g_shared_page->features & Kernel::API::SPF_RDTSCP)
{
const auto cpu = get_cpu();
const auto& sgettime = g_shared_page->gettime_shared;
const auto& lgettime = g_shared_page->cpus[cpu].gettime_local;
const auto old_seq = lgettime.seq;
if (old_seq & 1)
continue;
uint64_t monotonic_ns = __builtin_ia32_rdtsc() - lgettime.last_tsc;
if (lgettime.shift >= 0)
monotonic_ns <<= lgettime.shift;
else
monotonic_ns >>= -lgettime.shift;
monotonic_ns = (monotonic_ns * lgettime.mult) >> 32;
monotonic_ns += lgettime.last_ns;
if (old_seq != lgettime.seq || cpu != get_cpu())
continue;
*tp = {
.tv_sec = static_cast<time_t>(monotonic_ns / 1'000'000'000),
.tv_nsec = static_cast<long>(monotonic_ns % 1'000'000'000)
};
if (clock_id == CLOCK_REALTIME)
{
tp->tv_sec += sgettime.realtime_s;
tp->tv_nsec += sgettime.realtime_ns;
if (tp->tv_nsec >= 1'000'000'000)
{
tp->tv_sec += tp->tv_nsec / 1'000'000'000;
tp->tv_nsec = tp->tv_nsec % 1'000'000'000;
}
}
return 0;
uint32_t cpu;
curr_tsc = __builtin_ia32_rdtscp(&cpu);
read_tsc_info(g_shared_page->cpus[cpu].gettime_local);
}
else for (;;)
{
uint16_t cpu1, cpu2;
asm volatile("lsl %1, %0" : "=r"(cpu1) : "r"(g_shared_page->gdt_cpu_offset));
curr_tsc = __builtin_ia32_rdtsc();
asm volatile("lsl %1, %0" : "=r"(cpu2) : "r"(g_shared_page->gdt_cpu_offset));
if (cpu1 != cpu2)
continue;
read_tsc_info(g_shared_page->cpus[cpu1].gettime_local);
break;
}
// NOTE: as we read TSC before getting the calibration, it is possible
// for the calibration to get updated in between.
if (curr_tsc < last_tsc)
goto read_tsc_info;
uint64_t clock_ns = curr_tsc - last_tsc;
if (shift >= 0)
clock_ns <<= shift;
else
clock_ns >>= -shift;
clock_ns = (clock_ns * mult) >> 32;
clock_ns += last_ns;
if (clock_id == CLOCK_REALTIME)
{
const auto& sgettime = g_shared_page->gettime_shared;
clock_ns += sgettime.realtime_s * 1'000'000'000 + sgettime.realtime_ns;
}
*tp = {
.tv_sec = static_cast<time_t>(clock_ns / 1'000'000'000),
.tv_nsec = static_cast<long>(clock_ns % 1'000'000'000)
};
return 0;
}
int clock_getres(clockid_t clock_id, struct timespec* res)