add arm64 support for clock frequency helpers

On arm64, `cpu_khz` and `tsc_khz` do not exist -- they are x86-only
globals tied to the TSC. Replace them with arch-appropriate alternatives:

- homa_clock_khz(): use `arch_timer_get_cntfrq() / 1000` on arm64,
  which returns the ARM generic timer frequency matching get_cycles().
- timetrace.c: replace tsc_khz with homa_clock_khz() in the three
  places that print "cpu_khz:" for trace output.
- homa_metrics.c: guard the tsc_khz-based cycle conversion with
  CONFIG_X86; on arm64 both homa_clock() and Linux timekeeping use
  the same arch timer so no conversion is needed.

Without this patch, building HomaModule on arm64 fails with:
  error: 'cpu_khz' undeclared (first use in this function)

Signed-off-by: Serapheim Dimitropoulos <sdimitropoulos@coreweave.com>

Author: sdimitro

homa_impl.h

@@ -46,6 +46,9 @@
 #ifndef __UPSTREAM__ /* See strip.py */
 #include "homa.h"
 #include <linux/version.h>
+#ifdef CONFIG_ARM64
+#include <clocksource/arm_arch_timer.h>
+#endif
 #include "homa_devel.h"
 #else /* See strip.py */
 #include <linux/homa.h>
@@ -842,7 +845,13 @@ static inline u64 homa_clock_khz(void)
 	return 1000000;
 #else /* __UNIT_TEST__ */
 #ifndef __UPSTREAM__ /* See strip.py */
+#ifdef CONFIG_X86
 	return cpu_khz;
+#elif defined(CONFIG_ARM64)
+	return arch_timer_get_cntfrq() / 1000;
+#else
+	return 1000000;
+#endif
 #else /* See strip.py */
 	return 1000000;
 #endif /* See strip.py */

homa_metrics.c

@@ -241,13 +241,24 @@ char *homa_metrics_print(void)
 		M("bypass_softirq_cycles", m->bypass_softirq_cycles,
 		  "Time spent in homa_softirq during bypass from GRO\n");
 
-		/* Adjust stats gathered in Linux that use rdtsc. */
+		/* Adjust stats gathered in Linux that use rdtsc.
+		 * On x86 the Linux timekeeping clock (tsc) may differ from
+		 * homa_clock(); on arm64 both use the arch timer so the
+		 * ratio is 1.
+		 */
+#ifdef CONFIG_X86
 		M("linux_softirq_cycles", m->linux_softirq_cycles *
 		  (homa_clock_khz() / 1000) / (tsc_khz / 1000),
 		  "Time spent in all Linux SoftIRQ\n");
 		M("napi_cycles", m->napi_cycles * (homa_clock_khz() / 1000) /
 		  (tsc_khz / 1000),
 		  "Time spent in NAPI-level packet handling\n");
+#else
+		M("linux_softirq_cycles", m->linux_softirq_cycles,
+		  "Time spent in all Linux SoftIRQ\n");
+		M("napi_cycles", m->napi_cycles,
+		  "Time spent in NAPI-level packet handling\n");
+#endif
 		M("linux_softirqd_actions", m->linux_softirqd_actions,
 		  "SoftIRQ actions taken in the background softirqd thread\n");
 		M("send_cycles", m->send_cycles,

timetrace.c

@@ -390,7 +390,8 @@ int tt_proc_open(struct inode *inode, struct file *file)
 
 	if (!tt_test_no_khz) {
 		pf->bytes_available = snprintf(pf->msg_storage, TT_PF_BUF_SIZE,
-					       "cpu_khz: %u\n", tsc_khz);
+					       "cpu_khz: %llu\n",
+					       homa_clock_khz());
 	}
 
 done:
@@ -631,7 +632,7 @@ void tt_print_file(char *path)
 
 	bytes_used += snprintf(buffer + bytes_used,
 			sizeof(buffer) - bytes_used,
-			"cpu_khz: %u\n", tsc_khz);
+			"cpu_khz: %llu\n", homa_clock_khz());
 
 	/* Each iteration of this loop printk's one event. */
 	while (true) {
@@ -758,7 +759,7 @@ void tt_printk(void)
 	}
 #endif
 
-	pr_err("cpu_khz: %u, start: %llu\n", tsc_khz, start_time);
+	pr_err("cpu_khz: %llu, start: %llu\n", homa_clock_khz(), start_time);
 
 	/* Each iteration of this loop printk's one event. */
 	while (true) {