From: Marek Olšák <marek.olsak@amd.com>
Date: Tue, 19 Jul 2016 16:28:24 +0000 (+0200)
Subject: ddebug: implement pipelined hang detection mode
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=b47727a83ad6b5386f27b2c174a5c0f110a61038;p=mesa.git

ddebug: implement pipelined hang detection mode

For good performance while being able to generate decent hang reports.
The report doesn't contain the parsed IB and the buffer list, but it
isolates the draw call and dumps shaders while not having to flush
the context.

This is for GPU hangs that are harder to reproduce and require interactive
playing for minutes or even hours.

dd_pipe.h explains some implementation details. Initializing, copying
(recording) and clearing states is most of the code.

The performance should be at least 50% of the normal performance depending
on the circumstances. (i.e. 50% is expected to be the worst case scenario,
not the best case) The majority of time is spent in
dump_debug_state(PIPE_DUMP_CURRENT_SHADERS) and that's after all
the optimizations in later patches. There is no obvious way to optimize
that further.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
---

diff --git a/src/gallium/drivers/ddebug/dd_context.c b/src/gallium/drivers/ddebug/dd_context.c
index 96fea3a80c3..4423e904dc1 100644
--- a/src/gallium/drivers/ddebug/dd_context.c
+++ b/src/gallium/drivers/ddebug/dd_context.c
@@ -27,6 +27,7 @@
 
 #include "dd_pipe.h"
 #include "tgsi/tgsi_parse.h"
+#include "util/u_inlines.h"
 #include "util/u_memory.h"
 
 
@@ -589,6 +590,19 @@ dd_context_destroy(struct pipe_context *_pipe)
    struct dd_context *dctx = dd_context(_pipe);
    struct pipe_context *pipe = dctx->pipe;
 
+   if (dctx->thread) {
+      pipe_mutex_lock(dctx->mutex);
+      dctx->kill_thread = 1;
+      pipe_mutex_unlock(dctx->mutex);
+      pipe_thread_wait(dctx->thread);
+      pipe_mutex_destroy(dctx->mutex);
+      assert(!dctx->records);
+   }
+
+   if (dctx->fence) {
+      pipe->transfer_unmap(pipe, dctx->fence_transfer);
+      pipe_resource_reference(&dctx->fence, NULL);
+   }
    pipe->destroy(pipe);
    FREE(dctx);
 }
@@ -731,10 +745,8 @@ dd_context_create(struct dd_screen *dscreen, struct pipe_context *pipe)
       return NULL;
 
    dctx = CALLOC_STRUCT(dd_context);
-   if (!dctx) {
-      pipe->destroy(pipe);
-      return NULL;
-   }
+   if (!dctx)
+      goto fail;
 
    dctx->pipe = pipe;
    dctx->base.priv = pipe->priv; /* expose wrapped priv data */
@@ -826,5 +838,40 @@ dd_context_create(struct dd_screen *dscreen, struct pipe_context *pipe)
    dd_init_draw_functions(dctx);
 
    dctx->draw_state.sample_mask = ~0;
+
+   if (dscreen->mode == DD_DETECT_HANGS_PIPELINED) {
+      dctx->fence = pipe_buffer_create(dscreen->screen, PIPE_BIND_CUSTOM,
+                                            PIPE_USAGE_STAGING, 4);
+      if (!dctx->fence)
+         goto fail;
+
+      dctx->mapped_fence = pipe_buffer_map(pipe, dctx->fence,
+                                           PIPE_TRANSFER_READ_WRITE |
+                                           PIPE_TRANSFER_PERSISTENT |
+                                           PIPE_TRANSFER_COHERENT,
+                                           &dctx->fence_transfer);
+      if (!dctx->mapped_fence)
+         goto fail;
+
+      *dctx->mapped_fence = 0;
+
+      pipe_mutex_init(dctx->mutex);
+      dctx->thread = pipe_thread_create(dd_thread_pipelined_hang_detect, dctx);
+      if (!dctx->thread) {
+         pipe_mutex_destroy(dctx->mutex);
+         goto fail;
+      }
+   }
+
    return &dctx->base;
+
+fail:
+   if (dctx) {
+      if (dctx->mapped_fence)
+         pipe_transfer_unmap(pipe, dctx->fence_transfer);
+      pipe_resource_reference(&dctx->fence, NULL);
+      FREE(dctx);
+   }
+   pipe->destroy(pipe);
+   return NULL;
 }
diff --git a/src/gallium/drivers/ddebug/dd_draw.c b/src/gallium/drivers/ddebug/dd_draw.c
index 836f812296c..a54cecb5f08 100644
--- a/src/gallium/drivers/ddebug/dd_draw.c
+++ b/src/gallium/drivers/ddebug/dd_draw.c
@@ -29,7 +29,14 @@
 
 #include "util/u_dump.h"
 #include "util/u_format.h"
+#include "util/u_framebuffer.h"
+#include "util/u_helpers.h"
+#include "util/u_inlines.h"
+#include "util/u_memory.h"
+#include "tgsi/tgsi_parse.h"
 #include "tgsi/tgsi_scan.h"
+#include "os/os_time.h"
+#include <inttypes.h>
 
 
 static FILE *
@@ -569,6 +576,444 @@ dd_flush_and_handle_hang(struct dd_context *dctx,
    }
 }
 
+static void
+dd_unreference_copy_of_call(struct dd_call *dst)
+{
+   switch (dst->type) {
+   case CALL_DRAW_VBO:
+      pipe_so_target_reference(&dst->info.draw_vbo.count_from_stream_output, NULL);
+      pipe_resource_reference(&dst->info.draw_vbo.indirect, NULL);
+      pipe_resource_reference(&dst->info.draw_vbo.indirect_params, NULL);
+      break;
+   case CALL_LAUNCH_GRID:
+      pipe_resource_reference(&dst->info.launch_grid.indirect, NULL);
+      break;
+   case CALL_RESOURCE_COPY_REGION:
+      pipe_resource_reference(&dst->info.resource_copy_region.dst, NULL);
+      pipe_resource_reference(&dst->info.resource_copy_region.src, NULL);
+      break;
+   case CALL_BLIT:
+      pipe_resource_reference(&dst->info.blit.dst.resource, NULL);
+      pipe_resource_reference(&dst->info.blit.src.resource, NULL);
+      break;
+   case CALL_FLUSH_RESOURCE:
+      pipe_resource_reference(&dst->info.flush_resource, NULL);
+      break;
+   case CALL_CLEAR:
+      break;
+   case CALL_CLEAR_BUFFER:
+      pipe_resource_reference(&dst->info.clear_buffer.res, NULL);
+      break;
+   case CALL_CLEAR_RENDER_TARGET:
+      break;
+   case CALL_CLEAR_DEPTH_STENCIL:
+      break;
+   case CALL_GENERATE_MIPMAP:
+      pipe_resource_reference(&dst->info.generate_mipmap.res, NULL);
+      break;
+   }
+}
+
+static void
+dd_copy_call(struct dd_call *dst, struct dd_call *src)
+{
+   dst->type = src->type;
+
+   switch (src->type) {
+   case CALL_DRAW_VBO:
+      pipe_so_target_reference(&dst->info.draw_vbo.count_from_stream_output,
+                               src->info.draw_vbo.count_from_stream_output);
+      pipe_resource_reference(&dst->info.draw_vbo.indirect,
+                              src->info.draw_vbo.indirect);
+      pipe_resource_reference(&dst->info.draw_vbo.indirect_params,
+                              src->info.draw_vbo.indirect_params);
+      dst->info.draw_vbo = src->info.draw_vbo;
+      break;
+   case CALL_LAUNCH_GRID:
+      pipe_resource_reference(&dst->info.launch_grid.indirect,
+                              src->info.launch_grid.indirect);
+      dst->info.launch_grid = src->info.launch_grid;
+      break;
+   case CALL_RESOURCE_COPY_REGION:
+      pipe_resource_reference(&dst->info.resource_copy_region.dst,
+                              src->info.resource_copy_region.dst);
+      pipe_resource_reference(&dst->info.resource_copy_region.src,
+                              src->info.resource_copy_region.src);
+      dst->info.resource_copy_region = src->info.resource_copy_region;
+      break;
+   case CALL_BLIT:
+      pipe_resource_reference(&dst->info.blit.dst.resource,
+                              src->info.blit.dst.resource);
+      pipe_resource_reference(&dst->info.blit.src.resource,
+                              src->info.blit.src.resource);
+      dst->info.blit = src->info.blit;
+      break;
+   case CALL_FLUSH_RESOURCE:
+      pipe_resource_reference(&dst->info.flush_resource,
+                              src->info.flush_resource);
+      break;
+   case CALL_CLEAR:
+      dst->info.clear = src->info.clear;
+      break;
+   case CALL_CLEAR_BUFFER:
+      pipe_resource_reference(&dst->info.clear_buffer.res,
+                              src->info.clear_buffer.res);
+      dst->info.clear_buffer = src->info.clear_buffer;
+      break;
+   case CALL_CLEAR_RENDER_TARGET:
+      break;
+   case CALL_CLEAR_DEPTH_STENCIL:
+      break;
+   case CALL_GENERATE_MIPMAP:
+      pipe_resource_reference(&dst->info.generate_mipmap.res,
+                              src->info.generate_mipmap.res);
+      dst->info.generate_mipmap = src->info.generate_mipmap;
+      break;
+   }
+}
+
+static void
+dd_init_copy_of_draw_state(struct dd_draw_state_copy *state)
+{
+   unsigned i,j;
+
+   /* Just clear pointers to gallium objects. Don't clear the whole structure,
+    * because it would kill performance with its size of 130 KB.
+    */
+   memset(&state->base.index_buffer, 0,
+          sizeof(state->base.index_buffer));
+   memset(state->base.vertex_buffers, 0,
+          sizeof(state->base.vertex_buffers));
+   memset(state->base.so_targets, 0,
+          sizeof(state->base.so_targets));
+   memset(state->base.constant_buffers, 0,
+          sizeof(state->base.constant_buffers));
+   memset(state->base.sampler_views, 0,
+          sizeof(state->base.sampler_views));
+   memset(state->base.shader_images, 0,
+          sizeof(state->base.shader_images));
+   memset(state->base.shader_buffers, 0,
+          sizeof(state->base.shader_buffers));
+   memset(&state->base.framebuffer_state, 0,
+          sizeof(state->base.framebuffer_state));
+
+   memset(state->shaders, 0, sizeof(state->shaders));
+
+   state->base.render_cond.query = &state->render_cond;
+
+   for (i = 0; i < PIPE_SHADER_TYPES; i++) {
+      state->base.shaders[i] = &state->shaders[i];
+      for (j = 0; j < PIPE_MAX_SAMPLERS; j++)
+         state->base.sampler_states[i][j] = &state->sampler_states[i][j];
+   }
+
+   state->base.velems = &state->velems;
+   state->base.rs = &state->rs;
+   state->base.dsa = &state->dsa;
+   state->base.blend = &state->blend;
+}
+
+static void
+dd_unreference_copy_of_draw_state(struct dd_draw_state_copy *state)
+{
+   struct dd_draw_state *dst = &state->base;
+   unsigned i,j;
+
+   util_set_index_buffer(&dst->index_buffer, NULL);
+
+   for (i = 0; i < ARRAY_SIZE(dst->vertex_buffers); i++)
+      pipe_resource_reference(&dst->vertex_buffers[i].buffer, NULL);
+   for (i = 0; i < ARRAY_SIZE(dst->so_targets); i++)
+      pipe_so_target_reference(&dst->so_targets[i], NULL);
+
+   for (i = 0; i < PIPE_SHADER_TYPES; i++) {
+      if (dst->shaders[i])
+         tgsi_free_tokens(dst->shaders[i]->state.shader.tokens);
+
+      for (j = 0; j < PIPE_MAX_CONSTANT_BUFFERS; j++)
+         pipe_resource_reference(&dst->constant_buffers[i][j].buffer, NULL);
+      for (j = 0; j < PIPE_MAX_SAMPLERS; j++)
+         pipe_sampler_view_reference(&dst->sampler_views[i][j], NULL);
+      for (j = 0; j < PIPE_MAX_SHADER_IMAGES; j++)
+         pipe_resource_reference(&dst->shader_images[i][j].resource, NULL);
+      for (j = 0; j < PIPE_MAX_SHADER_BUFFERS; j++)
+         pipe_resource_reference(&dst->shader_buffers[i][j].buffer, NULL);
+   }
+
+   util_unreference_framebuffer_state(&dst->framebuffer_state);
+}
+
+static void
+dd_copy_draw_state(struct dd_draw_state *dst, struct dd_draw_state *src)
+{
+   unsigned i,j;
+
+   if (src->render_cond.query) {
+      *dst->render_cond.query = *src->render_cond.query;
+      dst->render_cond.condition = src->render_cond.condition;
+      dst->render_cond.mode = src->render_cond.mode;
+   } else {
+      dst->render_cond.query = NULL;
+   }
+
+   util_set_index_buffer(&dst->index_buffer, &src->index_buffer);
+
+   for (i = 0; i < ARRAY_SIZE(src->vertex_buffers); i++) {
+      pipe_resource_reference(&dst->vertex_buffers[i].buffer,
+                              src->vertex_buffers[i].buffer);
+      memcpy(&dst->vertex_buffers[i], &src->vertex_buffers[i],
+             sizeof(src->vertex_buffers[i]));
+   }
+
+   dst->num_so_targets = src->num_so_targets;
+   for (i = 0; i < ARRAY_SIZE(src->so_targets); i++)
+      pipe_so_target_reference(&dst->so_targets[i], src->so_targets[i]);
+   memcpy(dst->so_offsets, src->so_offsets, sizeof(src->so_offsets));
+
+   for (i = 0; i < PIPE_SHADER_TYPES; i++) {
+      if (!src->shaders[i]) {
+         dst->shaders[i] = NULL;
+         continue;
+      }
+
+      if (src->shaders[i]) {
+         dst->shaders[i]->state.shader = src->shaders[i]->state.shader;
+         dst->shaders[i]->state.shader.tokens =
+            tgsi_dup_tokens(src->shaders[i]->state.shader.tokens);
+      } else {
+         dst->shaders[i] = NULL;
+      }
+
+      for (j = 0; j < PIPE_MAX_CONSTANT_BUFFERS; j++) {
+         pipe_resource_reference(&dst->constant_buffers[i][j].buffer,
+                                 src->constant_buffers[i][j].buffer);
+         memcpy(&dst->constant_buffers[i][j], &src->constant_buffers[i][j],
+                sizeof(src->constant_buffers[i][j]));
+      }
+
+      for (j = 0; j < PIPE_MAX_SAMPLERS; j++) {
+         pipe_sampler_view_reference(&dst->sampler_views[i][j],
+                                     src->sampler_views[i][j]);
+         if (src->sampler_states[i][j])
+            dst->sampler_states[i][j]->state.sampler =
+               src->sampler_states[i][j]->state.sampler;
+         else
+            dst->sampler_states[i][j] = NULL;
+      }
+      /* TODO: shader buffers & images */
+   }
+
+   if (src->velems)
+      dst->velems->state.velems = src->velems->state.velems;
+   else
+      dst->velems = NULL;
+
+   if (src->rs)
+      dst->rs->state.rs = src->rs->state.rs;
+   else
+      dst->rs = NULL;
+
+   if (src->dsa)
+      dst->dsa->state.dsa = src->dsa->state.dsa;
+   else
+      dst->dsa = NULL;
+
+   if (src->blend)
+      dst->blend->state.blend = src->blend->state.blend;
+   else
+      dst->blend = NULL;
+
+   dst->blend_color = src->blend_color;
+   dst->stencil_ref = src->stencil_ref;
+   dst->sample_mask = src->sample_mask;
+   dst->min_samples = src->min_samples;
+   dst->clip_state = src->clip_state;
+   util_copy_framebuffer_state(&dst->framebuffer_state, &src->framebuffer_state);
+   memcpy(dst->scissors, src->scissors, sizeof(src->scissors));
+   memcpy(dst->viewports, src->viewports, sizeof(src->viewports));
+   memcpy(dst->tess_default_levels, src->tess_default_levels,
+          sizeof(src->tess_default_levels));
+   dst->apitrace_call_number = src->apitrace_call_number;
+}
+
+static void
+dd_free_record(struct dd_draw_record **record)
+{
+   struct dd_draw_record *next = (*record)->next;
+
+   dd_unreference_copy_of_call(&(*record)->call);
+   dd_unreference_copy_of_draw_state(&(*record)->draw_state);
+   FREE((*record)->driver_state_log);
+   FREE(*record);
+   *record = next;
+}
+
+static void
+dd_dump_record(struct dd_context *dctx, struct dd_draw_record *record,
+               uint32_t hw_sequence_no, int64_t now)
+{
+   FILE *f = dd_get_file_stream(dd_screen(dctx->base.screen),
+                                record->draw_state.base.apitrace_call_number);
+   if (!f)
+      return;
+
+   fprintf(f, "Draw call sequence # = %u\n", record->sequence_no);
+   fprintf(f, "HW reached sequence # = %u\n", hw_sequence_no);
+   fprintf(f, "Elapsed time = %"PRIi64" ms\n\n",
+           (now - record->timestamp) / 1000);
+
+   dd_dump_call(f, &record->draw_state.base, &record->call);
+   fprintf(f, "%s\n", record->driver_state_log);
+
+   dctx->pipe->dump_debug_state(dctx->pipe, f,
+                                PIPE_DUMP_DEVICE_STATUS_REGISTERS);
+   dd_dump_dmesg(f);
+   fclose(f);
+}
+
+PIPE_THREAD_ROUTINE(dd_thread_pipelined_hang_detect, input)
+{
+   struct dd_context *dctx = (struct dd_context *)input;
+   struct dd_screen *dscreen = dd_screen(dctx->base.screen);
+
+   pipe_mutex_lock(dctx->mutex);
+
+   while (!dctx->kill_thread) {
+      struct dd_draw_record **record = &dctx->records;
+
+      /* Loop over all records. */
+      while (*record) {
+         int64_t now;
+
+         /* If the fence has been signalled, release the record and all older
+          * records.
+          */
+         if (*dctx->mapped_fence >= (*record)->sequence_no) {
+            while (*record)
+               dd_free_record(record);
+            break;
+         }
+
+         /* The fence hasn't been signalled. Check the timeout. */
+         now = os_time_get();
+         if (os_time_timeout((*record)->timestamp,
+                             (*record)->timestamp + dscreen->timeout_ms * 1000,
+                             now)) {
+            fprintf(stderr, "GPU hang detected.\n");
+
+            /* Get the oldest unsignalled draw call. */
+            while ((*record)->next &&
+                   *dctx->mapped_fence < (*record)->next->sequence_no)
+               record = &(*record)->next;
+
+            dd_dump_record(dctx, *record, *dctx->mapped_fence, now);
+            dd_kill_process();
+         }
+
+         record = &(*record)->next;
+      }
+
+      /* Unlock and sleep before starting all over again. */
+      pipe_mutex_unlock(dctx->mutex);
+      os_time_sleep(10000); /* 10 ms */
+      pipe_mutex_lock(dctx->mutex);
+   }
+
+   /* Thread termination. */
+   while (dctx->records)
+      dd_free_record(&dctx->records);
+
+   pipe_mutex_unlock(dctx->mutex);
+   return 0;
+}
+
+static char *
+dd_get_driver_shader_log(struct dd_context *dctx)
+{
+   FILE *f;
+   char *buf;
+   int written_bytes;
+
+   if (!dctx->max_log_buffer_size)
+      dctx->max_log_buffer_size = 16 * 1024;
+
+   /* Keep increasing the buffer size until there is enough space.
+    *
+    * open_memstream can resize automatically, but it's VERY SLOW.
+    * fmemopen is much faster.
+    */
+   while (1) {
+      buf = malloc(dctx->max_log_buffer_size);
+      buf[0] = 0;
+
+      f = fmemopen(buf, dctx->max_log_buffer_size, "a");
+      if (!f) {
+         free(buf);
+         return NULL;
+      }
+
+      dd_dump_driver_state(dctx, f, PIPE_DUMP_CURRENT_SHADERS);
+      written_bytes = ftell(f);
+      fclose(f);
+
+      /* Return if the backing buffer is large enough. */
+      if (written_bytes < dctx->max_log_buffer_size - 1)
+         break;
+
+      /* Try again. */
+      free(buf);
+      dctx->max_log_buffer_size *= 2;
+   }
+
+   return buf;
+}
+
+static void
+dd_pipelined_process_draw(struct dd_context *dctx, struct dd_call *call)
+{
+   struct pipe_context *pipe = dctx->pipe;
+   struct dd_draw_record *record;
+   char *log;
+
+   /* Make a record of the draw call. */
+   record = MALLOC_STRUCT(dd_draw_record);
+   if (!record)
+      return;
+
+   /* Create the log. */
+   log = dd_get_driver_shader_log(dctx);
+   if (!log) {
+      FREE(record);
+      return;
+   }
+
+   /* Update the fence with the GPU.
+    *
+    * radeonsi/clear_buffer waits in the command processor until shaders are
+    * idle before writing to memory. That's a necessary condition for isolating
+    * draw calls.
+    */
+   dctx->sequence_no++;
+   pipe->clear_buffer(pipe, dctx->fence, 0, 4, &dctx->sequence_no, 4);
+
+   /* Initialize the record. */
+   record->timestamp = os_time_get();
+   record->sequence_no = dctx->sequence_no;
+   record->driver_state_log = log;
+
+   memset(&record->call, 0, sizeof(record->call));
+   dd_copy_call(&record->call, call);
+
+   dd_init_copy_of_draw_state(&record->draw_state);
+   dd_copy_draw_state(&record->draw_state.base, &dctx->draw_state);
+
+   /* Add the record to the list. */
+   pipe_mutex_lock(dctx->mutex);
+   record->next = dctx->records;
+   dctx->records = record;
+   pipe_mutex_unlock(dctx->mutex);
+}
+
 static void
 dd_context_flush(struct pipe_context *_pipe,
                  struct pipe_fence_handle **fence, unsigned flags)
@@ -581,6 +1026,7 @@ dd_context_flush(struct pipe_context *_pipe,
       dd_flush_and_handle_hang(dctx, fence, flags,
                                "GPU hang detected in pipe->flush()");
       break;
+   case DD_DETECT_HANGS_PIPELINED: /* nothing to do here */
    case DD_DUMP_ALL_CALLS:
    case DD_DUMP_APITRACE_CALL:
       pipe->flush(pipe, fence, flags);
@@ -625,6 +1071,9 @@ dd_after_draw(struct dd_context *dctx, struct dd_call *call)
             dd_kill_process();
          }
          break;
+      case DD_DETECT_HANGS_PIPELINED:
+         dd_pipelined_process_draw(dctx, call);
+         break;
       case DD_DUMP_ALL_CALLS:
          if (!dscreen->no_flush)
             pipe->flush(pipe, NULL, 0);
diff --git a/src/gallium/drivers/ddebug/dd_pipe.h b/src/gallium/drivers/ddebug/dd_pipe.h
index 5ba1b7a6d75..08cd1e3b9f1 100644
--- a/src/gallium/drivers/ddebug/dd_pipe.h
+++ b/src/gallium/drivers/ddebug/dd_pipe.h
@@ -32,9 +32,11 @@
 #include "pipe/p_state.h"
 #include "pipe/p_screen.h"
 #include "dd_util.h"
+#include "os/os_thread.h"
 
 enum dd_mode {
    DD_DETECT_HANGS,
+   DD_DETECT_HANGS_PIPELINED,
    DD_DUMP_ALL_CALLS,
    DD_DUMP_APITRACE_CALL,
 };
@@ -181,6 +183,33 @@ struct dd_draw_state
    unsigned apitrace_call_number;
 };
 
+struct dd_draw_state_copy
+{
+   struct dd_draw_state base;
+
+   /* dd_draw_state_copy does not reference real CSOs. Instead, it points to
+    * these variables, which serve as storage.
+    */
+   struct dd_query render_cond;
+   struct dd_state shaders[PIPE_SHADER_TYPES];
+   struct dd_state sampler_states[PIPE_SHADER_TYPES][PIPE_MAX_SAMPLERS];
+   struct dd_state velems;
+   struct dd_state rs;
+   struct dd_state dsa;
+   struct dd_state blend;
+};
+
+struct dd_draw_record {
+   struct dd_draw_record *next;
+
+   int64_t timestamp;
+   uint32_t sequence_no;
+
+   struct dd_call call;
+   struct dd_draw_state_copy draw_state;
+   char *driver_state_log;
+};
+
 struct dd_context
 {
    struct pipe_context base;
@@ -188,6 +217,32 @@ struct dd_context
 
    struct dd_draw_state draw_state;
    unsigned num_draw_calls;
+
+   /* Pipelined hang detection.
+    *
+    * This is without unnecessary flushes and waits. There is a memory-based
+    * fence that is incremented by clear_buffer every draw call. Driver fences
+    * are not used.
+    *
+    * After each draw call, a new dd_draw_record is created that contains
+    * a copy of all states, the output of pipe_context::dump_debug_state,
+    * and it has a fence number assigned. That's done without knowing whether
+    * that draw call is problematic or not. The record is added into the list
+    * of all records.
+    *
+    * An independent, separate thread loops over the list of records and checks
+    * their fences. Records with signalled fences are freed. On fence timeout,
+    * the thread dumps the record of the oldest unsignalled fence.
+    */
+   pipe_thread thread;
+   pipe_mutex mutex;
+   int kill_thread;
+   struct pipe_resource *fence;
+   struct pipe_transfer *fence_transfer;
+   uint32_t *mapped_fence;
+   uint32_t sequence_no;
+   struct dd_draw_record *records;
+   int max_log_buffer_size;
 };
 
 
@@ -196,6 +251,7 @@ dd_context_create(struct dd_screen *dscreen, struct pipe_context *pipe);
 
 void
 dd_init_draw_functions(struct dd_context *dctx);
+PIPE_THREAD_ROUTINE(dd_thread_pipelined_hang_detect, input);
 
 
 static inline struct dd_context *
diff --git a/src/gallium/drivers/ddebug/dd_screen.c b/src/gallium/drivers/ddebug/dd_screen.c
index 46869ab63f2..412ea365ce2 100644
--- a/src/gallium/drivers/ddebug/dd_screen.c
+++ b/src/gallium/drivers/ddebug/dd_screen.c
@@ -314,6 +314,11 @@ ddebug_screen_create(struct pipe_screen *screen)
       puts("    fence timeout and dump context and driver information into");
       puts("    $HOME/"DD_DIR"/ when a hang is detected.");
       puts("");
+      puts("  GALLIUM_DDEBUG=\"pipelined [timeout in ms] [verbose]\"");
+      puts("    Detect a device hang after every draw call based on the given fence");
+      puts("    timeout without flushes and dump context and driver information into");
+      puts("    $HOME/"DD_DIR"/ when a hang is detected.");
+      puts("");
       puts("  GALLIUM_DDEBUG=\"apitrace [call#] [verbose]\"");
       puts("    Dump apitrace draw call information into $HOME/"DD_DIR"/. Implies 'noflush'.");
       puts("");
@@ -337,6 +342,11 @@ ddebug_screen_create(struct pipe_screen *screen)
 
       if (sscanf(option+8, "%u", &apitrace_dump_call) != 1)
          return screen;
+   } else if (!strncmp(option, "pipelined", 8)) {
+      mode = DD_DETECT_HANGS_PIPELINED;
+
+      if (sscanf(option+10, "%u", &timeout) != 1)
+         return screen;
    } else {
       mode = DD_DETECT_HANGS;
 
@@ -392,6 +402,7 @@ ddebug_screen_create(struct pipe_screen *screen)
       fprintf(stderr, "Gallium debugger active. Logging all calls.\n");
       break;
    case DD_DETECT_HANGS:
+   case DD_DETECT_HANGS_PIPELINED:
       fprintf(stderr, "Gallium debugger active. "
               "The hang detection timeout is %i ms.\n", timeout);
       break;