Replace mtime counter with realtime timer

Replace 'mtime' counter by retrieving time of the specified clock ID.
Additionally, a  Real-Time counter (RTC) mechanism is implemented to
prevent 'mtime' overflow.

By changing 'mtime' to use the retrieved timer, real timestamps are
visible in the booting log due to 'CONFIG_PRINTK_TIME' being enable
by default.

However, SEMU experiences a slowdown as the number of simulated harts
increases. This is due to the increased number of timer interrupts
and 'clock_gettime' system calls in this commit.

Another issue addressed is RCU CPU stall warning[1], which occurs
when the system waits more than a grace period (typically 21 seconds).

[1]: https://docs.kernel.org/RCU/stallwarn.html

Author: chiangkd

Makefile

@@ -4,6 +4,11 @@ CC ?= gcc
 CFLAGS := -O2 -g -Wall -Wextra
 CFLAGS += -include common.h
 
+# clock frequency
+CLOCK_FREQ ?= 65000000
+DT_CFLAGS := -D CLOCK_FREQ=$(CLOCK_FREQ)
+CFLAGS += $(DT_CFLAGS)
+
 OBJS_EXTRA :=
 # command line option
 OPTS :=
@@ -42,6 +47,7 @@ all: $(BIN) minimal.dtb
 OBJS := \
 	riscv.o \
 	ram.o \
+	utils.o \
 	plic.o \
 	uart.o \
 	main.o \
@@ -72,11 +78,12 @@ S := $E $E
 SMP ?= 1
 .PHONY: riscv-harts.dtsi
 riscv-harts.dtsi:
-	$(Q)python3 scripts/gen-hart-dts.py $@ $(SMP)
+	$(Q)python3 scripts/gen-hart-dts.py $@ $(SMP) $(CLOCK_FREQ)
 
 minimal.dtb: minimal.dts riscv-harts.dtsi
 	$(VECHO) " DTC\t$@\n"
 	$(Q)$(CC) -nostdinc -E -P -x assembler-with-cpp -undef \
+	    $(DT_CFLAGS) \
 	    $(subst ^,$S,$(filter -D^SEMU_FEATURE_%, $(subst -D$(S)SEMU_FEATURE,-D^SEMU_FEATURE,$(CFLAGS)))) $< \
 	    | $(DTC) - > $@
 

clint.c

@@ -5,7 +5,9 @@
 
 void clint_update_interrupts(hart_t *hart, clint_state_t *clint)
 {
-    if (clint->mtime > clint->mtimecmp[hart->mhartid])
+    uint64_t time_delta =
+        clint->mtimecmp[hart->mhartid] - semu_timer_get(&hart->time);
+    if ((int64_t) time_delta <= 0)
         hart->sip |= RV_INT_STI_BIT;
     else
         hart->sip &= ~RV_INT_STI_BIT;
@@ -31,7 +33,8 @@ static bool clint_reg_read(clint_state_t *clint, uint32_t addr, uint32_t *value)
     }
 
     if (addr < 0xBFFF) {
-        *value = clint->mtime >> (32 & -!!(addr & 0b100));
+        *value = (uint32_t) (semu_timer_get(&clint->mtime) >>
+                             (32 & -!!(addr & 0b100)));
         return true;
     }
     return false;
@@ -58,14 +61,14 @@ static bool clint_reg_write(clint_state_t *clint, uint32_t addr, uint32_t value)
     }
 
     if (addr < 0xBFFF) {
-        int32_t upper = clint->mtime >> 32;
-        int32_t lowwer = clint->mtime;
+        int32_t upper = clint->mtime.begin >> 32;
+        int32_t lower = clint->mtime.begin;
         if (addr & 0b100)
             upper = value;
         else
-            lowwer = value;
+            lower = value;
 
-        clint->mtime = (uint64_t) upper << 32 | lowwer;
+        semu_timer_rebase(&clint->mtime, (uint64_t) upper << 32 | lower);
         return true;
     }
     return false;

device.h

@@ -175,7 +175,7 @@ uint32_t *virtio_blk_init(virtio_blk_state_t *vblk, char *disk_file);
 typedef struct {
     uint32_t msip[4096];
     uint64_t mtimecmp[4095];
-    uint64_t mtime;
+    semu_timer_t mtime;
 } clint_state_t;
 
 void clint_update_interrupts(hart_t *vm, clint_state_t *clint);

main.c

@@ -81,13 +81,6 @@ static void emu_update_timer_interrupt(hart_t *hart)
     clint_update_interrupts(hart, &data->clint);
 }
 
-static void emu_update_global_timer(vm_t *vm)
-{
-    emu_state_t *data = PRIV(vm->hart[0]);
-    data->clint.mtime++;
-    return;
-}
-
 static void mem_load(hart_t *hart,
                      uint32_t addr,
                      uint8_t width,
@@ -523,6 +516,7 @@ static int semu_start(int argc, char **argv)
     /* Initialize the emulator */
     emu_state_t emu;
     memset(&emu, 0, sizeof(emu));
+    semu_timer_init(&emu.clint.mtime, CLOCK_FREQ);
 
     /* Set up RAM */
     emu.ram = mmap(NULL, RAM_SIZE, PROT_READ | PROT_WRITE,
@@ -591,8 +585,6 @@ static int semu_start(int argc, char **argv)
     /* Emulate */
     uint32_t peripheral_update_ctr = 0;
     while (!emu.stopped) {
-        emu_update_global_timer(&vm);
-
         for (uint32_t i = 0; i < vm.n_hart; i++) {
             if (peripheral_update_ctr-- == 0) {
                 peripheral_update_ctr = 64;

minimal.dts

@@ -24,7 +24,7 @@
     cpus {
         #address-cells = <1>;
         #size-cells = <0>;
-        timebase-frequency = <65000000>;
+        timebase-frequency = <CLOCK_FREQ>;
     };
 
     sram: memory@0 {

riscv.c

@@ -434,10 +434,10 @@ static void csr_read(hart_t *vm, uint16_t addr, uint32_t *value)
 {
     switch (addr) {
     case RV_CSR_TIME:
-        *value = vm->time;
+        *value = semu_timer_get(&vm->time);
         return;
     case RV_CSR_TIMEH:
-        *value = vm->time >> 32;
+        *value = semu_timer_get(&vm->time) >> 32;
         return;
     case RV_CSR_INSTRET:
         *value = vm->instret;

riscv.h

@@ -3,6 +3,8 @@
 #include <stdbool.h>
 #include <stdint.h>
 
+#include "utils.h"
+
 /* ERR_EXCEPTION indicates that the instruction has raised one of the
  * exceptions defined in the specification. If this flag is set, the
  * additional fields "exc_cause" and "exc_val" must also be set to values
@@ -70,8 +72,7 @@ struct __hart_internal {
      * resets.
      */
     uint64_t instret;
-    uint64_t time;
-
+    semu_timer_t time;
     /* Instruction execution state must be set to "NONE" for instruction
      * execution to continue. If the state is not "NONE," the vm_step()
      * function will exit.

scripts/gen-hart-dts.py

@@ -34,12 +34,12 @@ def clint_irq_format(nums):
         s += f"<&cpu{i}_intc 3 &cpu{i}_intc 7>, "
     return s[:-2]
 
-def dtsi_template (cpu_list: str, plic_list, clint_list):
+def dtsi_template (cpu_list: str, plic_list, clint_list, clock_freq):
     return f"""/{{
     cpus {{
         #address-cells = <1>;
         #size-cells = <0>;
-        timebase-frequency = <65000000>;
+        timebase-frequency = <{clock_freq}>;
         {cpu_list}
     }};
 
@@ -67,6 +67,7 @@ def dtsi_template (cpu_list: str, plic_list, clint_list):
 
 dtsi = sys.argv[1]
 harts = int(sys.argv[2])
+clock_freq = int(sys.argv[3])
 
 with open(dtsi, "w") as dts:
-    dts.write(dtsi_template(cpu_format(harts), plic_irq_format(harts), clint_irq_format(harts)))
+    dts.write(dtsi_template(cpu_format(harts), plic_irq_format(harts), clint_irq_format(harts), clock_freq))

utils.c

@@ -0,0 +1,47 @@
+#include <time.h>
+
+#include "utils.h"
+
+#if defined(__APPLE__)
+#define HAVE_MACH_TIMER
+#include <mach/mach_time.h>
+#elif !defined(_WIN32) && !defined(_WIN64)
+#define HAVE_POSIX_TIMER
+#ifdef CLOCK_MONOTONIC
+#define CLOCKID CLOCK_MONOTONIC
+#else
+#define CLOCKID CLOCK_REALTIME
+#endif
+#endif
+
+void semu_timer_init(semu_timer_t *timer, uint64_t freq)
+{
+    timer->freq = freq;
+    semu_timer_rebase(timer, 0);
+}
+
+static uint64_t semu_timer_clocksource(uint64_t freq)
+{
+#if defined(HAVE_POSIX_TIMER)
+    struct timespec t;
+    clock_gettime(CLOCKID, &t);
+    return (t.tv_sec * freq) + (t.tv_nsec * freq / 1e9);
+#elif defined(HAVE_MACH_TIMER)
+    static mach_timebase_info_data_t t;
+    if (mach_clk.denom == 0)
+        (void) mach_timebase_info(&t);
+    return mach_absolute_time() * freq / t.denom * t.numer;
+#else
+    return time(0) * freq;
+#endif
+}
+
+uint64_t semu_timer_get(semu_timer_t *timer)
+{
+    return semu_timer_clocksource(timer->freq) - timer->begin;
+}
+
+void semu_timer_rebase(semu_timer_t *timer, uint64_t time)
+{
+    timer->begin = semu_timer_clocksource(timer->freq) - time;
+}

utils.h

@@ -0,0 +1,13 @@
+#pragma once
+
+#include <stdint.h>
+
+/* TIMER */
+typedef struct {
+    uint64_t begin;
+    uint64_t freq;
+} semu_timer_t;
+
+void semu_timer_init(semu_timer_t *timer, uint64_t freq);
+uint64_t semu_timer_get(semu_timer_t *timer);
+void semu_timer_rebase(semu_timer_t *timer, uint64_t time);