overlay: Accumulate busy times

Still a little too course as we add multiple overlapping seqnos and
waits.

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

diff --git a/overlay/gpu-perf.c b/overlay/gpu-perf.c
index 653148b..56bc8b0 100644 (file)
--- a/overlay/gpu-perf.c
+++ b/overlay/gpu-perf.c
@@ -154,7 +154,7 @@ err:
        return EINVAL;
 }
 
-static char *get_comm(pid_t pid, char *comm, int len)
+static int get_comm(pid_t pid, char *comm, int len)
 {
        char filename[1024];
        int fd;
@@ -168,9 +168,10 @@ static char *get_comm(pid_t pid, char *comm, int len)
                if (len >= 0)
                        comm[len-1] = '\0';
                close(fd);
-       }
+       } else
+               len = -1;
 
-       return comm;
+       return len;
 }
 
 static struct gpu_perf_comm *
@@ -178,6 +179,9 @@ lookup_comm(struct gpu_perf *gp, pid_t pid)
 {
        struct gpu_perf_comm *comm;
 
+       if (pid == 0)
+               return NULL;
+
        for (comm = gp->comm; comm != NULL; comm = comm->next) {
                if (comm->pid == pid)
                        break;
@@ -187,10 +191,14 @@ lookup_comm(struct gpu_perf *gp, pid_t pid)
                if (comm == NULL)
                        return NULL;
 
+               if (get_comm(pid, comm->name, sizeof(comm->name)) < 0) {
+                       free(comm);
+                       return NULL;
+               }
+
+               comm->pid = pid;
                comm->next = gp->comm;
                gp->comm = comm;
-               get_comm(pid, comm->name, sizeof(comm->name));
-               comm->pid = pid;
        }
 
        return comm;
@@ -209,13 +217,44 @@ static int request_add(struct gpu_perf *gp, const void *event)
        return 1;
 }
 
-static int seqno_start(struct gpu_perf *gp, const void *event)
+static int busy_start(struct gpu_perf *gp, const void *event)
 {
+       const struct sample_event *sample = event;
+       struct gpu_perf_comm *comm;
+       struct gpu_perf_time *busy;
+
+       comm = lookup_comm(gp, sample->pid);
+       if (comm == NULL)
+               return 0;
+
+       busy = malloc(sizeof(*busy));
+       if (busy == NULL)
+               return 0;
+
+       busy->seqno = sample->raw[2];
+       busy->time = sample->time;
+       busy->comm = comm;
+       busy->next = gp->busy;
+       gp->busy = busy;
+
        return 0;
 }
 
-static int seqno_end(struct gpu_perf *gp, const void *event)
+static int busy_end(struct gpu_perf *gp, const void *event)
 {
+       const struct sample_event *sample = event;
+       struct gpu_perf_time *busy, **prev;
+
+       for (prev = &gp->busy; (busy = *prev) != NULL; prev = &busy->next) {
+               if (busy->seqno != sample->raw[2])
+                       continue;
+
+               busy->comm->busy_time += sample->time - busy->time;
+               *prev = busy->next;
+               free(busy);
+               return 1;
+       }
+
        return 0;
 }
 
@@ -229,7 +268,7 @@ static int wait_begin(struct gpu_perf *gp, const void *event)
 {
        const struct sample_event *sample = event;
        struct gpu_perf_comm *comm;
-       struct gpu_perf_wait *wait;
+       struct gpu_perf_time *wait;
 
        comm = lookup_comm(gp, sample->pid);
        if (comm == NULL)
@@ -239,10 +278,11 @@ static int wait_begin(struct gpu_perf *gp, const void *event)
        if (wait == NULL)
                return 0;
 
+       wait->comm = comm;
        wait->seqno = sample->raw[3];
        wait->time = sample->time;
-       wait->next = comm->wait;
-       comm->wait = wait;
+       wait->next = gp->wait;
+       gp->wait = wait;
 
        return 0;
 }
@@ -250,18 +290,13 @@ static int wait_begin(struct gpu_perf *gp, const void *event)
 static int wait_end(struct gpu_perf *gp, const void *event)
 {
        const struct sample_event *sample = event;
-       struct gpu_perf_comm *comm;
-       struct gpu_perf_wait *wait, **prev;
-
-       comm = lookup_comm(gp, sample->pid);
-       if (comm == NULL)
-               return 0;
+       struct gpu_perf_time *wait, **prev;
 
-       for (prev = &comm->wait; (wait = *prev) != NULL; prev = &wait->next) {
+       for (prev = &gp->wait; (wait = *prev) != NULL; prev = &wait->next) {
                if (wait->seqno != sample->raw[3])
                        continue;
 
-               comm->wait_time += sample->time - wait->time;
+               wait->comm->wait_time += sample->time - wait->time;
                *prev = wait->next;
                free(wait);
                return 1;
@@ -277,8 +312,8 @@ void gpu_perf_init(struct gpu_perf *gp, unsigned flags)
        gp->page_size = getpagesize();
 
        perf_tracepoint_open(gp, "i915", "i915_gem_request_add", request_add);
-       if (perf_tracepoint_open(gp, "i915", "i915_gem_ring_complete", seqno_end) == 0)
-               perf_tracepoint_open(gp, "i915", "i915_gem_ring_dispatch", seqno_start);
+       if (perf_tracepoint_open(gp, "i915", "i915_gem_ring_complete", busy_end) == 0)
+               perf_tracepoint_open(gp, "i915", "i915_gem_ring_dispatch", busy_start);
        if (perf_tracepoint_open(gp, "i915", "i915_gem_request_wait_begin", wait_begin) == 0)
                perf_tracepoint_open(gp, "i915", "i915_gem_request_wait_end", wait_end);
        perf_tracepoint_open(gp, "i915", "i915_flip_complete", flip_complete);

diff --git a/overlay/gpu-perf.h b/overlay/gpu-perf.h
index 476bbaa..ced9379 100644 (file)
--- a/overlay/gpu-perf.h
+++ b/overlay/gpu-perf.h
@@ -14,19 +14,20 @@ struct gpu_perf {
        int flip_complete;
        struct gpu_perf_comm {
                struct gpu_perf_comm *next;
-               struct gpu_perf_wait {
-                       struct gpu_perf_wait *next;
-                       uint32_t seqno;
-                       uint64_t time;
-               } *wait;
                char name[256];
                pid_t pid;
                int nr_requests[4];
                void *user_data;
 
-               uint64_t wait_begin;
                uint64_t wait_time;
+               uint64_t busy_time;
        } *comm;
+       struct gpu_perf_time {
+               struct gpu_perf_time *next;
+               struct gpu_perf_comm *comm;
+               uint32_t seqno;
+               uint64_t time;
+       } *wait, *busy;
 };
 
 void gpu_perf_init(struct gpu_perf *gp, unsigned flags);

diff --git a/overlay/overlay.c b/overlay/overlay.c
index bd327ab..4232711 100644 (file)
--- a/overlay/overlay.c
+++ b/overlay/overlay.c
@@ -286,6 +286,27 @@ static void show_gpu_perf(struct overlay_context *ctx, struct overlay_gpu_perf *
                        }
                        comm->wait_time = 0;
                }
+               if (comm->busy_time) {
+                       buf[0] = '\0';
+                       if (comm->busy_time > 1000*1000) {
+                               sprintf(buf, "%s %.1f ms busy",
+                                       need_comma ? "," : "",
+                                       comm->busy_time / (1000*1000.));
+                       } else if (comm->busy_time > 100) {
+                               sprintf(buf, "%s %.1f us busy",
+                                       need_comma ? "," : "",
+                                       comm->busy_time / 1000.);
+                       } else {
+                               sprintf(buf, "%s %.0f ns busy",
+                                       need_comma ? "," : "",
+                                       (double)comm->busy_time);
+                       }
+                       if (buf[0] != '\0') {
+                               cairo_show_text(ctx->cr, buf);
+                               need_comma = true;
+                       }
+                       comm->busy_time = 0;
+               }
                y += 14;
 
                memset(comm->nr_requests, 0, sizeof(comm->nr_requests));