demo: implement perf command

4 files changed, 98 insertions, 6 deletions

diff --git a/demo/demo.h b/demo/demo.h
index f98d599..c889e1a 100644
--- a/demo/demo.h
+++ b/demo/demo.h
@@ -35,6 +35,7 @@ int strcmp(const char *s1, const char *s2);
 char *strtok_input(char **tokens);
 
 int expect_end(char **tokens);
+void unexpected_eof();
 
 int parse_hex(char **tokens, unsigned *val);
 int parse_ptr(char **tokens, void **ptr);
@@ -45,4 +46,7 @@ int parse_cpu_mask(char **tokens, unsigned *mask);
 
 void cache_debug(unsigned cpu, void *ptr);
 
+void perf_show(unsigned cpu);
+void perf_clear(unsigned cpu);
+
 #endif
diff --git a/demo/main.c b/demo/main.c
index 988ef55..89347ad 100644
--- a/demo/main.c
+++ b/demo/main.c
@@ -6,6 +6,11 @@ struct cpu *__cpus[] = { &cpu0, &cpu1, &cpu2, &cpu3 };
 
 static volatile unsigned boot_done;
 
+static void unknown_command(const char *cmd)
+{
+	print("unknown command '%s'", cmd);
+}
+
 static void cmd_run(char **tokens)
 {
 	unsigned mask;
@@ -55,6 +60,27 @@ static void cmd_cache(char **tokens)
 	cache_debug(cpu, ptr);
 }
 
+static void cmd_perf(char **tokens)
+{
+	unsigned cpu;
+	const char *cmd;
+
+	if (parse_cpu(tokens, &cpu) < 0)
+		return;
+	else if (!(cmd = strtok_input(tokens))) {
+		unexpected_eof();
+		return;
+	} else if (expect_end(tokens) < 0)
+		return;
+
+	if (!strcmp(cmd, "clear"))
+		perf_clear(cpu);
+	else if (!strcmp(cmd, "show"))
+		perf_show(cpu);
+	else
+		unknown_command(cmd);
+}
+
 static void bsp_main(void)
 {
 	boot_done = 0;
@@ -83,8 +109,10 @@ static void bsp_main(void)
 			cmd_write(&tokens);
 		else if (!strcmp(cmd, "cache"))
 			cmd_cache(&tokens);
+		else if (!strcmp(cmd, "perf"))
+			cmd_perf(&tokens);
 		else
-			print("unknown command '%s'", cmd);
+			unknown_command(cmd);
 	}
 }
 
diff --git a/demo/perf.c b/demo/perf.c
new file mode 100644
index 0000000..28a20c8
--- /dev/null
+++ b/demo/perf.c
@@ -0,0 +1,60 @@
+#include "demo.h"
+
+#define PERF_BASE(n)             (0x30150000 | (0x40 * (n)))
+#define PERF_CLEAR(n)            (*(volatile unsigned *)PERF_BASE(n))
+#define PERF_MEM_READS(n)        (*(const volatile unsigned *)PERF_BASE(n))
+#define PERF_MEM_WRITES(n)       (*(const volatile unsigned *)(PERF_BASE(n) + 4))
+#define PERF_MEM_READ_CYCLES(n)  (*(const volatile unsigned *)(PERF_BASE(n) + 8))
+#define PERF_MEM_WRITE_CYCLES(n) (*(const volatile unsigned *)(PERF_BASE(n) + 12))
+#define PERF_RING_READS(n)       (*(const volatile unsigned *)(PERF_BASE(n) + 16))
+#define PERF_RING_INVALS(n)      (*(const volatile unsigned *)(PERF_BASE(n) + 20))
+#define PERF_RING_READ_INVALS(n) (*(const volatile unsigned *)(PERF_BASE(n) + 24))
+#define PERF_RING_REPLIES(n)     (*(const volatile unsigned *)(PERF_BASE(n) + 28))
+#define PERF_RING_FORWARDS(n)    (*(const volatile unsigned *)(PERF_BASE(n) + 32))
+#define PERF_RING_CYCLES(n)      (*(const volatile unsigned *)(PERF_BASE(n) + 36))
+#define PERF_IO_READS(n)         (*(const volatile unsigned *)(PERF_BASE(n) + 40))
+#define PERF_IO_WRITES(n)        (*(const volatile unsigned *)(PERF_BASE(n) + 44))
+
+#define PERF_MIN_MASK  0x0000ffff
+#define PERF_MAX_MASK  0xffff0000
+#define PERF_MAX_SHIFT 16
+
+void perf_show(unsigned cpu)
+{
+	print("dumping performance counters for cpu%u", cpu);
+
+	unsigned mem_reads = PERF_MEM_READS(cpu);
+	unsigned mem_writes = PERF_MEM_WRITES(cpu);
+	unsigned read_cycles = PERF_MEM_READ_CYCLES(cpu);
+	unsigned write_cycles = PERF_MEM_WRITE_CYCLES(cpu);
+	unsigned ring_reads = PERF_RING_READS(cpu);
+	unsigned ring_invals = PERF_RING_INVALS(cpu);
+	unsigned ring_read_invals = PERF_RING_READ_INVALS(cpu);
+	unsigned ring_replies = PERF_RING_REPLIES(cpu);
+	unsigned ring_forwards = PERF_RING_FORWARDS(cpu);
+	unsigned ring_cycles = PERF_RING_CYCLES(cpu);
+	unsigned io_reads = PERF_IO_READS(cpu);
+	unsigned io_writes = PERF_IO_WRITES(cpu);
+
+	unsigned min_ring_cycles = ring_cycles & PERF_MIN_MASK;
+	unsigned max_ring_cycles = (ring_cycles & PERF_MAX_MASK) >> PERF_MAX_SHIFT;
+	unsigned min_read_cycles = read_cycles & PERF_MIN_MASK;
+	unsigned max_read_cycles = (read_cycles & PERF_MAX_MASK) >> PERF_MAX_SHIFT;
+	unsigned min_write_cycles = write_cycles & PERF_MIN_MASK;
+	unsigned max_write_cycles = (write_cycles & PERF_MAX_MASK) >> PERF_MAX_SHIFT;
+	unsigned ring_sends = ring_reads + ring_invals + ring_read_invals;
+
+	print("requests:         sends=%u forwards=%u replies=%u", ring_sends, ring_forwards, ring_replies);
+	print("requests sent:    read=%u inval=%u read_inval=%u", ring_reads, ring_invals, ring_read_invals);
+	print("memory:           misses=%u writebacks=%u", mem_reads, mem_writes);
+	print("uncached i/o:     reads=%u writes=%u", io_reads, io_writes);
+	print("ring cycles:      min=%u max=%u", min_ring_cycles, max_ring_cycles);
+	print("mem read cycles:  min=%u max=%u", min_read_cycles, max_read_cycles);
+	print("mem write cycles: min=%u max=%u", min_write_cycles, max_write_cycles);
+}
+
+void perf_clear(unsigned cpu)
+{
+	PERF_CLEAR(cpu) = 0;
+	print("cleared performance counters for cpu%u", cpu);
+}
diff --git a/demo/util.c b/demo/util.c
index d08ec20..d728fac 100644
--- a/demo/util.c
+++ b/demo/util.c
@@ -7,11 +7,6 @@ static void bad_input(void)
 	print("bad input");
 }
 
-static void unexpected_eof(void)
-{
-	print("unexpected end-of-input");
-}
-
 static int parse_cpu_token(const char *token, unsigned *cpu)
 {
 	if (token[0] != 'c' || token[1] != 'p' || token[2] != 'u'
@@ -24,6 +19,11 @@ static int parse_cpu_token(const char *token, unsigned *cpu)
 	return 0;
 }
 
+void unexpected_eof(void)
+{
+	print("unexpected end-of-input");
+}
+
 int strcmp(const char *s1, const char *s2)
 {
     while (*s1 && *s1 == *s2)