overlay: Use the new i915 PMU to query GPU busyness

And so avoid having to hold forcewake indefinitely.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

1 files changed, 203 insertions, 50 deletions

diff --git a/overlay/gpu-top.c b/overlay/gpu-top.c
index 400cbc58..ec014744 100644
--- a/overlay/gpu-top.c
+++ b/overlay/gpu-top.c
@@ -1,8 +1,11 @@
+#include <linux/perf_event.h>
 #include <stdint.h>
-#include <unistd.h>
-#include <fcntl.h>
+#include <stdlib.h>
 #include <stdio.h>
 #include <string.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <errno.h>
 
 #include "igfx.h"
 #include "gpu-top.h"
@@ -14,52 +17,156 @@
 #define   RING_WAIT		(1<<11)
 #define   RING_WAIT_SEMAPHORE	(1<<10)
 
-struct ring {
+#define __I915_PERF_RING(n) (4*n)
+#define I915_PERF_RING_BUSY(n) (__I915_PERF_RING(n) + 0)
+#define I915_PERF_RING_WAIT(n) (__I915_PERF_RING(n) + 1)
+#define I915_PERF_RING_SEMA(n) (__I915_PERF_RING(n) + 2)
+
+static int
+perf_event_open(struct perf_event_attr *attr,
+		pid_t pid,
+		int cpu,
+		int group_fd,
+		unsigned long flags)
+{
+#ifndef __NR_perf_event_open
+#if defined(__i386__)
+#define __NR_perf_event_open 336
+#elif defined(__x86_64__)
+#define __NR_perf_event_open 298
+#else
+#define __NR_perf_event_open 0
+#endif
+#endif
+    attr->size = sizeof(*attr);
+    return syscall(__NR_perf_event_open, attr, pid, cpu, group_fd, flags);
+}
+
+static uint64_t i915_type_id(void)
+{
+	char buf[1024];
+	int fd, n;
+
+	fd = open("/sys/bus/event_source/devices/i915/type", 0);
+	if (fd < 0)
+		return 0;
+	n = read(fd, buf, sizeof(buf)-1);
+	close(fd);
+	if (n < 0)
+		return 0;
+
+	buf[n] = '\0';
+	return strtoull(buf, 0, 0);
+}
+
+static int perf_i915_open(int config, int group)
+{
+	struct perf_event_attr attr;
+
+	memset(&attr, 0, sizeof (attr));
+
+	attr.type = i915_type_id();
+	if (attr.type == 0)
+		return -ENOENT;
+	attr.config = config;
+
+	attr.freq = 1;
+	attr.sample_freq = 1000;
+
+	attr.read_format = PERF_FORMAT_TOTAL_TIME_ENABLED;
+	if (group == -1)
+		attr.read_format |= PERF_FORMAT_GROUP;
+
+	return perf_event_open(&attr, -1, 0, group, 0);
+}
+
+static int perf_init(struct gpu_top *gt)
+{
+	const char *names[] = {
+		"render",
+		"bitstream",
+		"bliter",
+		NULL,
+	};
+	int n;
+
+	gt->fd = perf_i915_open(I915_PERF_RING_BUSY(0), -1);
+	if (gt->fd < 0)
+		return -1;
+
+	if (perf_i915_open(I915_PERF_RING_WAIT(0), gt->fd) >= 0)
+		gt->have_wait = 1;
+
+	if (perf_i915_open(I915_PERF_RING_SEMA(0), gt->fd) >= 0)
+		gt->have_sema = 1;
+
+	gt->ring[0].name = names[0];
+	gt->num_rings = 1;
+
+	for (n = 1; names[n]; n++) {
+		if (perf_i915_open(I915_PERF_RING_BUSY(n), gt->fd) >= 0) {
+			if (gt->have_wait &&
+			    perf_i915_open(I915_PERF_RING_WAIT(n), gt->fd) < 0)
+				return -1;
+
+			if (gt->have_sema &&
+			    perf_i915_open(I915_PERF_RING_SEMA(n), gt->fd) < 0)
+				return -1;
+
+			gt->ring[gt->num_rings++].name = names[n];
+		}
+	}
+
+	return 0;
+}
+
+struct mmio_ring {
 	int id;
-	uint32_t mmio;
+	uint32_t base;
+	void *mmio;
 	int idle, wait, sema;
 };
 
-static void *mmio;
-
-static uint32_t ring_read(struct ring *ring, uint32_t reg)
+static uint32_t mmio_ring_read(struct mmio_ring *ring, uint32_t reg)
 {
-	return igfx_read(mmio, ring->mmio + reg);
+	return igfx_read(ring->mmio, reg);
 }
 
-static void ring_init(struct ring *ring)
+static void mmio_ring_init(struct mmio_ring *ring, void *mmio)
 {
 	uint32_t ctl;
 
-	ctl = ring_read(ring, RING_CTL);
+	ring->mmio = (char *)mmio + ring->base;
+
+	ctl = mmio_ring_read(ring, RING_CTL);
 	if ((ctl & 1) == 0)
 		ring->id = -1;
 }
 
-static void ring_reset(struct ring *ring)
+static void mmio_ring_reset(struct mmio_ring *ring)
 {
 	ring->idle = 0;
 	ring->wait = 0;
 	ring->sema = 0;
 }
 
-static void ring_sample(struct ring *ring)
+static void mmio_ring_sample(struct mmio_ring *ring)
 {
 	uint32_t head, tail, ctl;
 
 	if (ring->id == -1)
 		return;
 
-	head = ring_read(ring, RING_HEAD) & ADDR_MASK;
-	tail = ring_read(ring, RING_TAIL) & ADDR_MASK;
+	head = mmio_ring_read(ring, RING_HEAD) & ADDR_MASK;
+	tail = mmio_ring_read(ring, RING_TAIL) & ADDR_MASK;
 	ring->idle += head == tail;
 
-	ctl = ring_read(ring, RING_CTL);
+	ctl = mmio_ring_read(ring, RING_CTL);
 	ring->wait += !!(ctl & RING_WAIT);
 	ring->sema += !!(ctl & RING_WAIT_SEMAPHORE);
 }
 
-static void ring_emit(struct ring *ring, int samples, union gpu_top_payload *payload)
+static void mmio_ring_emit(struct mmio_ring *ring, int samples, union gpu_top_payload *payload)
 {
 	if (ring->id == -1)
 		return;
@@ -69,28 +176,26 @@ static void ring_emit(struct ring *ring, int samples, union gpu_top_payload *pay
 	payload[ring->id].u.sema = 100 * ring->sema / samples;
 }
 
-void gpu_top_init(struct gpu_top *gt)
+static void mmio_init(struct gpu_top *gt)
 {
-	struct ring render_ring = {
-		.mmio = 0x2030,
+	struct mmio_ring render_ring = {
+		.base = 0x2030,
 		.id = 0,
 	}, bsd_ring = {
-		.mmio = 0x4030,
+		.base = 0x4030,
 		.id = 1,
 	}, bsd6_ring = {
-		.mmio = 0x12030,
+		.base = 0x12030,
 		.id = 1,
 	}, blt_ring = {
-		.mmio = 0x22030,
+		.base = 0x22030,
 		.id = 2,
 	};
 	const struct igfx_info *info;
 	struct pci_device *igfx;
+	void *mmio;
 	int fd[2], i;
 
-	memset(gt, 0, sizeof(*gt));
-	gt->fd = -1;
-
 	igfx = igfx_get();
 	if (!igfx)
 		return;
@@ -105,6 +210,7 @@ void gpu_top_init(struct gpu_top *gt)
 	default:
 		 fcntl(fd[0], F_SETFL, fcntl(fd[0], F_GETFL) | O_NONBLOCK);
 		 gt->fd = fd[0];
+		 gt->type = MMIO;
 		 gt->ring[0].name = "render";
 		 gt->num_rings = 1;
 		 if (info->gen >= 040) {
@@ -124,54 +230,101 @@ void gpu_top_init(struct gpu_top *gt)
 
 	mmio = igfx_get_mmio(igfx);
 
-	ring_init(&render_ring);
+	mmio_ring_init(&render_ring, mmio);
 	if (info->gen >= 060) {
-		ring_init(&bsd6_ring);
-		ring_init(&blt_ring);
+		mmio_ring_init(&bsd6_ring, mmio);
+		mmio_ring_init(&blt_ring, mmio);
 	} else if (info->gen >= 040) {
-		ring_init(&bsd_ring);
+		mmio_ring_init(&bsd_ring, mmio);
 	}
 
 	for (;;) {
 		union gpu_top_payload payload[MAX_RINGS];
 
-		ring_reset(&render_ring);
-		ring_reset(&bsd_ring);
-		ring_reset(&bsd6_ring);
-		ring_reset(&blt_ring);
+		mmio_ring_reset(&render_ring);
+		mmio_ring_reset(&bsd_ring);
+		mmio_ring_reset(&bsd6_ring);
+		mmio_ring_reset(&blt_ring);
 
 		for (i = 0; i < 1000; i++) {
-			ring_sample(&render_ring);
-			ring_sample(&bsd_ring);
-			ring_sample(&bsd6_ring);
-			ring_sample(&blt_ring);
+			mmio_ring_sample(&render_ring);
+			mmio_ring_sample(&bsd_ring);
+			mmio_ring_sample(&bsd6_ring);
+			mmio_ring_sample(&blt_ring);
 			usleep(1000);
 		}
 
-		ring_emit(&render_ring, 1000, payload);
-		ring_emit(&bsd_ring, 1000, payload);
-		ring_emit(&bsd6_ring, 1000, payload);
-		ring_emit(&blt_ring, 1000, payload);
+		mmio_ring_emit(&render_ring, 1000, payload);
+		mmio_ring_emit(&bsd_ring, 1000, payload);
+		mmio_ring_emit(&bsd6_ring, 1000, payload);
+		mmio_ring_emit(&blt_ring, 1000, payload);
 
 		write(fd[1], payload, sizeof(payload));
 	}
 }
 
+void gpu_top_init(struct gpu_top *gt)
+{
+	memset(gt, 0, sizeof(*gt));
+	gt->fd = -1;
+
+	if (perf_init(gt) == 0)
+		return;
+
+	mmio_init(gt);
+}
+
 int gpu_top_update(struct gpu_top *gt)
 {
 	uint32_t data[1024];
-	int len, update = 0;
+	int update, len;
 
 	if (gt->fd < 0)
-		return update;
-
-	while ((len = read(gt->fd, data, sizeof(data))) > 0) {
-		uint32_t *ptr = &data[len/sizeof(uint32_t) - MAX_RINGS];
-		gt->ring[0].u.payload = ptr[0];
-		gt->ring[1].u.payload = ptr[1];
-		gt->ring[2].u.payload = ptr[2];
-		gt->ring[3].u.payload = ptr[3];
+		return 0;
+
+	if (gt->type == PERF) {
+		struct gpu_top_stat *s = &gt->stat[gt->count++&1];
+		struct gpu_top_stat *d = &gt->stat[gt->count&1];
+		uint64_t *sample, d_time;
+		int n;
+
+		len = read(gt->fd, data, sizeof(data));
+		if (len < 0)
+			return 0;
+
+		sample = (uint64_t *)data + 1;
+
+		s->time = *sample++;
+		for (n = 0; n < gt->num_rings; n++) {
+			s->busy[n] = sample[n];
+			if (gt->have_wait)
+				s->wait[n] = sample[n];
+			if (gt->have_sema)
+				s->sema[n] = sample[n];
+		}
+
+		if (gt->count == 1)
+			return 0;
+
+		d_time = s->time - d->time;
+		for (n = 0; n < gt->num_rings; n++) {
+			gt->ring[n].u.u.busy = 100 * (s->busy[n] - d->busy[n]) / d_time;
+			if (gt->have_wait)
+				gt->ring[n].u.u.wait = 100 * (s->wait[n] - d->wait[n]) / d_time;
+			if (gt->have_sema)
+				gt->ring[n].u.u.sema = 100 * (s->sema[n] - d->sema[n]) / d_time;
+		}
+
 		update = 1;
+	} else {
+		while ((len = read(gt->fd, data, sizeof(data))) > 0) {
+			uint32_t *ptr = &data[len/sizeof(uint32_t) - MAX_RINGS];
+			gt->ring[0].u.payload = ptr[0];
+			gt->ring[1].u.payload = ptr[1];
+			gt->ring[2].u.payload = ptr[2];
+			gt->ring[3].u.payload = ptr[3];
+			update = 1;
+		}
 	}
 
 	return update;