BLI: use explicit task isolation, no longer part of parallel operations

After looking into task isolation issues with Sergey, we couldn't find the
reason behind the deadlocks that we are getting in T87938 and a Sprite Fright
file involving motion blur renders.

There is no apparent place where we adding or waiting on tasks in a task group
from different isolation regions, which is what is known to cause problems. Yet
it still hangs. Either we do not understand some limitation of TBB isolation,
or there is a bug in TBB, but we could not figure it out.

Instead the idea is to use isolation only where we know we need it: when
holding a mutex lock and then doing some multithreaded operation within that
locked region. Three places where we do this now:
* Generated images
* Cached BVH tree building
* OpenVDB lazy grid loading

Compared to the more automatic approach previously used, there is the downside
that it is easy to miss places where we need isolation. Yet doing it more
automatically is also causing unexpected issue and bugs that we found no
solution for, so this seems better.

Patch implemented by Sergey and me.

Differential Revision: https://developer.blender.org/D11603

intern/cycles/blender/blender_python.cpp

@@ -289,11 +289,10 @@ static PyObject *render_func(PyObject * /*self*/, PyObject *args)
   RNA_pointer_create(NULL, &RNA_Depsgraph, (ID *)PyLong_AsVoidPtr(pydepsgraph), &depsgraphptr);
   BL::Depsgraph b_depsgraph(depsgraphptr);
 
-  /* Allow Blender to execute other Python scripts, and isolate TBB tasks so we
-   * don't get deadlocks with Blender threads accessing shared data like images. */
+  /* Allow Blender to execute other Python scripts. */
   python_thread_state_save(&session->python_thread_state);
 
-  tbb::this_task_arena::isolate([&] { session->render(b_depsgraph); });
+  session->render(b_depsgraph);
 
   python_thread_state_restore(&session->python_thread_state);
 
@@ -330,8 +329,7 @@ static PyObject *bake_func(PyObject * /*self*/, PyObject *args)
 
   python_thread_state_save(&session->python_thread_state);
 
-  tbb::this_task_arena::isolate(
-      [&] { session->bake(b_depsgraph, b_object, pass_type, pass_filter, width, height); });
+  session->bake(b_depsgraph, b_object, pass_type, pass_filter, width, height);
 
   python_thread_state_restore(&session->python_thread_state);
 
@@ -377,7 +375,7 @@ static PyObject *reset_func(PyObject * /*self*/, PyObject *args)
 
   python_thread_state_save(&session->python_thread_state);
 
-  tbb::this_task_arena::isolate([&] { session->reset_session(b_data, b_depsgraph); });
+  session->reset_session(b_data, b_depsgraph);
 
   python_thread_state_restore(&session->python_thread_state);
 
@@ -399,7 +397,7 @@ static PyObject *sync_func(PyObject * /*self*/, PyObject *args)
 
   python_thread_state_save(&session->python_thread_state);
 
-  tbb::this_task_arena::isolate([&] { session->synchronize(b_depsgraph); });
+  session->synchronize(b_depsgraph);
 
   python_thread_state_restore(&session->python_thread_state);
 

source/blender/blenkernel/intern/bvhutils.c

@@ -31,6 +31,7 @@
 
 #include "BLI_linklist.h"
 #include "BLI_math.h"
+#include "BLI_task.h"
 #include "BLI_threads.h"
 #include "BLI_utildefines.h"
 
@@ -160,6 +161,26 @@ void bvhcache_free(BVHCache *bvh_cache)
   MEM_freeN(bvh_cache);
 }
 
+/* BVH tree balancing inside a mutex lock must be run in isolation. Balancing
+ * is multithreaded, and we do not want the current thread to start another task
+ * that may involve acquiring the same mutex lock that it is waiting for. */
+static void bvhtree_balance_isolated(void *userdata)
+{
+  BLI_bvhtree_balance((BVHTree *)userdata);
+}
+
+static void bvhtree_balance(BVHTree *tree, const bool isolate)
+{
+  if (tree) {
+    if (isolate) {
+      BLI_task_isolate(bvhtree_balance_isolated, tree);
+    }
+    else {
+      BLI_bvhtree_balance(tree);
+    }
+  }
+}
+
 /** \} */
 /* -------------------------------------------------------------------- */
 /** \name Local Callbacks
@@ -566,7 +587,6 @@ static BVHTree *bvhtree_from_editmesh_verts_create_tree(float epsilon,
       BLI_bvhtree_insert(tree, i, eve->co, 1);
     }
     BLI_assert(BLI_bvhtree_get_len(tree) == verts_num_active);
-    BLI_bvhtree_balance(tree);
   }
 
   return tree;
@@ -600,7 +620,6 @@ static BVHTree *bvhtree_from_mesh_verts_create_tree(float epsilon,
         BLI_bvhtree_insert(tree, i, vert[i].co, 1);
       }
       BLI_assert(BLI_bvhtree_get_len(tree) == verts_num_active);
-      BLI_bvhtree_balance(tree);
     }
   }
 
@@ -649,6 +668,7 @@ BVHTree *bvhtree_from_editmesh_verts_ex(BVHTreeFromEditMesh *data,
     if (data->cached == false) {
       tree = bvhtree_from_editmesh_verts_create_tree(
           epsilon, tree_type, axis, em, verts_mask, verts_num_active);
+      bvhtree_balance(tree, true);
 
       /* Save on cache for later use */
       /* printf("BVHTree built and saved on cache\n"); */
@@ -660,6 +680,7 @@ BVHTree *bvhtree_from_editmesh_verts_ex(BVHTreeFromEditMesh *data,
   else {
     tree = bvhtree_from_editmesh_verts_create_tree(
         epsilon, tree_type, axis, em, verts_mask, verts_num_active);
+    bvhtree_balance(tree, false);
   }
 
   if (tree) {
@@ -711,6 +732,7 @@ BVHTree *bvhtree_from_mesh_verts_ex(BVHTreeFromMesh *data,
   if (in_cache == false) {
     tree = bvhtree_from_mesh_verts_create_tree(
         epsilon, tree_type, axis, vert, verts_num, verts_mask, verts_num_active);
+    bvhtree_balance(tree, bvh_cache_p != NULL);
 
     if (bvh_cache_p) {
       /* Save on cache for later use */
@@ -771,7 +793,6 @@ static BVHTree *bvhtree_from_editmesh_edges_create_tree(float epsilon,
       BLI_bvhtree_insert(tree, i, co[0], 2);
     }
     BLI_assert(BLI_bvhtree_get_len(tree) == edges_num_active);
-    BLI_bvhtree_balance(tree);
   }
 
   return tree;
@@ -809,7 +830,6 @@ static BVHTree *bvhtree_from_mesh_edges_create_tree(const MVert *vert,
 
         BLI_bvhtree_insert(tree, i, co[0], 2);
       }
-      BLI_bvhtree_balance(tree);
     }
   }
 
@@ -861,7 +881,7 @@ BVHTree *bvhtree_from_editmesh_edges_ex(BVHTreeFromEditMesh *data,
     if (data->cached == false) {
       tree = bvhtree_from_editmesh_edges_create_tree(
           epsilon, tree_type, axis, em, edges_mask, edges_num_active);
-
+      bvhtree_balance(tree, true);
       /* Save on cache for later use */
       /* printf("BVHTree built and saved on cache\n"); */
       bvhcache_insert(bvh_cache, tree, bvh_cache_type);
@@ -872,6 +892,7 @@ BVHTree *bvhtree_from_editmesh_edges_ex(BVHTreeFromEditMesh *data,
   else {
     tree = bvhtree_from_editmesh_edges_create_tree(
         epsilon, tree_type, axis, em, edges_mask, edges_num_active);
+    bvhtree_balance(tree, false);
   }
 
   if (tree) {
@@ -928,12 +949,17 @@ BVHTree *bvhtree_from_mesh_edges_ex(BVHTreeFromMesh *data,
         vert, edge, edges_num, edges_mask, edges_num_active, epsilon, tree_type, axis);
 
     if (bvh_cache_p) {
+      bvhtree_balance(tree, true);
+
       BVHCache *bvh_cache = *bvh_cache_p;
       /* Save on cache for later use */
       /* printf("BVHTree built and saved on cache\n"); */
       bvhcache_insert(bvh_cache, tree, bvh_cache_type);
       in_cache = true;
     }
+    else {
+      bvhtree_balance(tree, false);
+    }
   }
 
   if (bvh_cache_p) {
@@ -994,7 +1020,6 @@ static BVHTree *bvhtree_from_mesh_faces_create_tree(float epsilon,
         }
       }
       BLI_assert(BLI_bvhtree_get_len(tree) == faces_num_active);
-      BLI_bvhtree_balance(tree);
     }
   }
 
@@ -1057,6 +1082,7 @@ BVHTree *bvhtree_from_mesh_faces_ex(BVHTreeFromMesh *data,
   if (in_cache == false) {
     tree = bvhtree_from_mesh_faces_create_tree(
         epsilon, tree_type, axis, vert, face, numFaces, faces_mask, faces_num_active);
+    bvhtree_balance(tree, bvh_cache_p != NULL);
 
     if (bvh_cache_p) {
       /* Save on cache for later use */
@@ -1127,7 +1153,6 @@ static BVHTree *bvhtree_from_editmesh_looptri_create_tree(float epsilon,
         }
       }
       BLI_assert(BLI_bvhtree_get_len(tree) == looptri_num_active);
-      BLI_bvhtree_balance(tree);
     }
   }
 
@@ -1173,7 +1198,6 @@ static BVHTree *bvhtree_from_mesh_looptri_create_tree(float epsilon,
         }
       }
       BLI_assert(BLI_bvhtree_get_len(tree) == looptri_num_active);
-      BLI_bvhtree_balance(tree);
     }
   }
 
@@ -1229,6 +1253,7 @@ BVHTree *bvhtree_from_editmesh_looptri_ex(BVHTreeFromEditMesh *data,
     bool in_cache = bvhcache_find(
         bvh_cache_p, bvh_cache_type, &tree, &lock_started, mesh_eval_mutex);
     BVHCache *bvh_cache = *bvh_cache_p;
+    bvhtree_balance(tree, true);
 
     if (in_cache == false) {
       tree = bvhtree_from_editmesh_looptri_create_tree(
@@ -1243,6 +1268,7 @@ BVHTree *bvhtree_from_editmesh_looptri_ex(BVHTreeFromEditMesh *data,
   else {
     tree = bvhtree_from_editmesh_looptri_create_tree(
         epsilon, tree_type, axis, em, looptri_mask, looptri_num_active);
+    bvhtree_balance(tree, false);
   }
 
   if (tree) {
@@ -1303,6 +1329,8 @@ BVHTree *bvhtree_from_mesh_looptri_ex(BVHTreeFromMesh *data,
                                                  looptri_mask,
                                                  looptri_num_active);
 
+    bvhtree_balance(tree, bvh_cache_p != NULL);
+
     if (bvh_cache_p) {
       BVHCache *bvh_cache = *bvh_cache_p;
       bvhcache_insert(bvh_cache, tree, bvh_cache_type);
@@ -1742,7 +1770,7 @@ BVHTree *BKE_bvhtree_from_pointcloud_get(BVHTreeFromPointCloud *data,
     BLI_bvhtree_insert(tree, i, pointcloud->co[i], 1);
   }
   BLI_assert(BLI_bvhtree_get_len(tree) == pointcloud->totpoint);
-  BLI_bvhtree_balance(tree);
+  bvhtree_balance(tree, false);
 
   data->coords = pointcloud->co;
   data->tree = tree;

source/blender/blenkernel/intern/editmesh_tangent.c

@@ -362,7 +362,7 @@ void BKE_editmesh_loop_tangent_calc(BMEditMesh *em,
     /* Calculation */
     if (em->tottri != 0) {
       TaskPool *task_pool;
-      task_pool = BLI_task_pool_create(NULL, TASK_PRIORITY_LOW, TASK_ISOLATION_ON);
+      task_pool = BLI_task_pool_create(NULL, TASK_PRIORITY_LOW);
 
       tangent_mask_curr = 0;
       /* Calculate tangent layers */

source/blender/blenkernel/intern/image.c

@@ -68,6 +68,7 @@
 #include "BLI_math_vector.h"
 #include "BLI_mempool.h"
 #include "BLI_system.h"
+#include "BLI_task.h"
 #include "BLI_threads.h"
 #include "BLI_timecode.h" /* For stamp time-code format. */
 #include "BLI_utildefines.h"
@@ -882,6 +883,39 @@ Image *BKE_image_load_exists(Main *bmain, const char *filepath)
   return BKE_image_load_exists_ex(bmain, filepath, NULL);
 }
 
+typedef struct ImageFillData {
+  short gen_type;
+  uint width;
+  uint height;
+  unsigned char *rect;
+  float *rect_float;
+  float fill_color[4];
+} ImageFillData;
+
+static void image_buf_fill_isolated(void *usersata_v)
+{
+  ImageFillData *usersata = usersata_v;
+
+  const short gen_type = usersata->gen_type;
+  const uint width = usersata->width;
+  const uint height = usersata->height;
+
+  unsigned char *rect = usersata->rect;
+  float *rect_float = usersata->rect_float;
+
+  switch (gen_type) {
+    case IMA_GENTYPE_GRID:
+      BKE_image_buf_fill_checker(rect, rect_float, width, height);
+      break;
+    case IMA_GENTYPE_GRID_COLOR:
+      BKE_image_buf_fill_checker_color(rect, rect_float, width, height);
+      break;
+    default:
+      BKE_image_buf_fill_color(rect, rect_float, width, height, usersata->fill_color);
+      break;
+  }
+}
+
 static ImBuf *add_ibuf_size(unsigned int width,
                             unsigned int height,
                             const char *name,
@@ -944,17 +978,16 @@ static ImBuf *add_ibuf_size(unsigned int width,
 
   STRNCPY(ibuf->name, name);
 
-  switch (gen_type) {
-    case IMA_GENTYPE_GRID:
-      BKE_image_buf_fill_checker(rect, rect_float, width, height);
-      break;
-    case IMA_GENTYPE_GRID_COLOR:
-      BKE_image_buf_fill_checker_color(rect, rect_float, width, height);
-      break;
-    default:
-      BKE_image_buf_fill_color(rect, rect_float, width, height, fill_color);
-      break;
-  }
+  ImageFillData data;
+
+  data.gen_type = gen_type;
+  data.width = width;
+  data.height = height;
+  data.rect = rect;
+  data.rect_float = rect_float;
+  copy_v4_v4(data.fill_color, fill_color);
+
+  BLI_task_isolate(image_buf_fill_isolated, &data);
 
   return ibuf;
 }

source/blender/blenkernel/intern/lib_override.c

@@ -2335,8 +2335,7 @@ bool BKE_lib_override_library_main_operations_create(Main *bmain, const bool for
   }
 
   struct LibOverrideOpCreateData create_pool_data = {.bmain = bmain, .changed = false};
-  TaskPool *task_pool = BLI_task_pool_create(
-      &create_pool_data, TASK_PRIORITY_HIGH, TASK_ISOLATION_ON);
+  TaskPool *task_pool = BLI_task_pool_create(&create_pool_data, TASK_PRIORITY_HIGH);
 
   FOREACH_MAIN_ID_BEGIN (bmain, id) {
     if (!ID_IS_LINKED(id) && ID_IS_OVERRIDE_LIBRARY_REAL(id) &&

source/blender/blenkernel/intern/mesh_evaluate.c

@@ -1715,8 +1715,7 @@ void BKE_mesh_normals_loop_split(const MVert *mverts,
     loop_split_generator(NULL, &common_data);
   }
   else {
-    TaskPool *task_pool = BLI_task_pool_create(
-        &common_data, TASK_PRIORITY_HIGH, TASK_ISOLATION_ON);
+    TaskPool *task_pool = BLI_task_pool_create(&common_data, TASK_PRIORITY_HIGH);
 
     loop_split_generator(task_pool, &common_data);
 

source/blender/blenkernel/intern/mesh_tangent.c

@@ -656,7 +656,7 @@ void BKE_mesh_calc_loop_tangent_ex(const MVert *mvert,
 
     /* Calculation */
     if (looptri_len != 0) {
-      TaskPool *task_pool = BLI_task_pool_create(NULL, TASK_PRIORITY_LOW, TASK_ISOLATION_ON);
+      TaskPool *task_pool = BLI_task_pool_create(NULL, TASK_PRIORITY_LOW);
 
       tangent_mask_curr = 0;
       /* Calculate tangent layers */

source/blender/blenkernel/intern/ocean.c

@@ -663,7 +663,7 @@ void BKE_ocean_simulate(struct Ocean *o, float t, float scale, float chop_amount
   osd.scale = scale;
   osd.chop_amount = chop_amount;
 
-  pool = BLI_task_pool_create(&osd, TASK_PRIORITY_HIGH, TASK_ISOLATION_ON);
+  pool = BLI_task_pool_create(&osd, TASK_PRIORITY_HIGH);
 
   BLI_rw_mutex_lock(&o->oceanmutex, THREAD_LOCK_WRITE);
 

source/blender/blenkernel/intern/particle.c

@@ -3179,7 +3179,7 @@ void psys_cache_child_paths(ParticleSimulationData *sim,
     return;
   }
 
-  task_pool = BLI_task_pool_create(&ctx, TASK_PRIORITY_LOW, TASK_ISOLATION_ON);
+  task_pool = BLI_task_pool_create(&ctx, TASK_PRIORITY_LOW);
   totchild = ctx.totchild;
   totparent = ctx.totparent;
 

source/blender/blenkernel/intern/particle_distribute.c

@@ -1330,7 +1330,7 @@ static void distribute_particles_on_dm(ParticleSimulationData *sim, int from)
     return;
   }
 
-  task_pool = BLI_task_pool_create(&ctx, TASK_PRIORITY_LOW, TASK_ISOLATION_ON);
+  task_pool = BLI_task_pool_create(&ctx, TASK_PRIORITY_LOW);
 
   totpart = (from == PART_FROM_CHILD ? sim->psys->totchild : sim->psys->totpart);
   psys_tasks_create(&ctx, 0, totpart, &tasks, &numtasks);

source/blender/blenkernel/intern/volume.cc

@@ -324,15 +324,20 @@ struct VolumeGrid {
 
     openvdb::io::File file(filepath);
 
-    try {
-      file.setCopyMaxBytes(0);
-      file.open();
-      openvdb::GridBase::Ptr vdb_grid = file.readGrid(name());
-      entry->grid->setTree(vdb_grid->baseTreePtr());
-    }
-    catch (const openvdb::IoError &e) {
-      entry->error_msg = e.what();
-    }
+    /* Isolate file loading since that's potentially multithreaded and we are
+     * holding a mutex lock. See BLI_task_isolate. Note OpenVDB already uses
+     * TBB, so it's fine to use here without a wrapper. */
+    tbb::this_task_arena::isolate([&] {
+      try {
+        file.setCopyMaxBytes(0);
+        file.open();
+        openvdb::GridBase::Ptr vdb_grid = file.readGrid(name());
+        entry->grid->setTree(vdb_grid->baseTreePtr());
+      }
+      catch (const openvdb::IoError &e) {
+        entry->error_msg = e.what();
+      }
+    });
 
     std::atomic_thread_fence(std::memory_order_release);
     entry->is_loaded = true;

source/blender/blenlib/BLI_task.h

@@ -67,55 +67,17 @@ typedef enum TaskPriority {
   TASK_PRIORITY_HIGH,
 } TaskPriority;
 
-/**
- * Task isolation helps avoid unexpected task scheduling decisions that can lead to bugs if wrong
- * assumptions were made. Typically that happens when doing "nested threading", i.e. one thread
- * schedules a bunch of main-tasks and those spawn new sub-tasks.
- *
- * What can happen is that when a main-task waits for its sub-tasks to complete on other threads,
- * another main-task is scheduled within the already running main-task. Generally, this is good,
- * because it leads to better performance. However, sometimes code (often unintentionally) makes
- * the assumption that at most one main-task runs on a thread at a time.
- *
- * The bugs often show themselves in two ways:
- * - Deadlock, when a main-task holds a mutex while waiting for its sub-tasks to complete.
- * - Data corruption, when a main-task makes wrong assumptions about a thread-local variable.
- *
- * Task isolation can avoid these bugs by making sure that a main-task does not start executing
- * another main-task while waiting for its sub-tasks. More precisely, a function that runs in an
- * isolated region is only allowed to run sub-tasks that were spawned in the same isolated region.
- *
- * Unfortunately, incorrect use of task isolation can lead to deadlocks itself. This can happen
- * when threading primitives are used that separate spawning tasks from executing them. The problem
- * occurs when a task is spawned in one isolated region while the tasks are waited for in another
- * isolated region. In this setup, the thread that is waiting for the spawned tasks to complete
- * cannot run the tasks itself. On a single thread, that causes a deadlock already. When there are
- * multiple threads, another thread will typically run the task and avoid the deadlock. However, if
- * this situation happens on all threads at the same time, all threads will deadlock. This happened
- * in T88598.
- */
-typedef enum TaskIsolation {
-  /* Do not use task isolation. Always use this when tasks are pushed recursively. */
-  TASK_ISOLATION_OFF,
-  /* Run each task in its own isolated region. */
-  TASK_ISOLATION_ON,
-} TaskIsolation;
-
 typedef struct TaskPool TaskPool;
 typedef void (*TaskRunFunction)(TaskPool *__restrict pool, void *taskdata);
 typedef void (*TaskFreeFunction)(TaskPool *__restrict pool, void *taskdata);
 
 /* Regular task pool that immediately starts executing tasks as soon as they
  * are pushed, either on the current or another thread. */
-TaskPool *BLI_task_pool_create(void *userdata,
-                               TaskPriority priority,
-                               TaskIsolation task_isolation);
+TaskPool *BLI_task_pool_create(void *userdata, TaskPriority priority);
 
 /* Background: always run tasks in a background thread, never immediately
  * execute them. For running background jobs. */
-TaskPool *BLI_task_pool_create_background(void *userdata,
-                                          TaskPriority priority,
-                                          TaskIsolation task_isolation);
+TaskPool *BLI_task_pool_create_background(void *userdata, TaskPriority priority);
 
 /* Background Serial: run tasks one after the other in the background,
  * without parallelization between the tasks. */
@@ -125,9 +87,7 @@ TaskPool *BLI_task_pool_create_background_serial(void *userdata, TaskPriority pr
  * as threads can't immediately start working. But it can be used if the data
  * structures the threads operate on are not fully initialized until all tasks
  * are created. */
-TaskPool *BLI_task_pool_create_suspended(void *userdata,
-                                         TaskPriority priority,
-                                         TaskIsolation task_isolation);
+TaskPool *BLI_task_pool_create_suspended(void *userdata, TaskPriority priority);
 
 /* No threads: immediately executes tasks on the same thread. For debugging. */
 TaskPool *BLI_task_pool_create_no_threads(void *userdata);
@@ -365,6 +325,36 @@ struct TaskNode *BLI_task_graph_node_create(struct TaskGraph *task_graph,
 bool BLI_task_graph_node_push_work(struct TaskNode *task_node);
 void BLI_task_graph_edge_create(struct TaskNode *from_node, struct TaskNode *to_node);
 
+/* Task Isolation
+ *
+ * Task isolation helps avoid unexpected task scheduling decisions that can lead to bugs if wrong
+ * assumptions were made. Typically that happens when doing "nested threading", i.e. one thread
+ * schedules a bunch of main-tasks and those spawn new sub-tasks.
+ *
+ * What can happen is that when a main-task waits for its sub-tasks to complete on other threads,
+ * another main-task is scheduled within the already running main-task. Generally, this is good,
+ * because it leads to better performance. However, sometimes code (often unintentionally) makes
+ * the assumption that at most one main-task runs on a thread at a time.
+ *
+ * The bugs often show themselves in two ways:
+ * - Deadlock, when a main-task holds a mutex while waiting for its sub-tasks to complete.
+ * - Data corruption, when a main-task makes wrong assumptions about a thread-local variable.
+ *
+ * Task isolation can avoid these bugs by making sure that a main-task does not start executing
+ * another main-task while waiting for its sub-tasks. More precisely, a function that runs in an
+ * isolated region is only allowed to run sub-tasks that were spawned in the same isolated region.
+ *
+ * Unfortunately, incorrect use of task isolation can lead to deadlocks itself. This can happen
+ * when threading primitives are used that separate spawning tasks from executing them. The problem
+ * occurs when a task is spawned in one isolated region while the tasks are waited for in another
+ * isolated region. In this setup, the thread that is waiting for the spawned tasks to complete
+ * cannot run the tasks itself. On a single thread, that causes a deadlock already. When there are
+ * multiple threads, another thread will typically run the task and avoid the deadlock. However, if
+ * this situation happens on all threads at the same time, all threads will deadlock. This happened
+ * in T88598.
+ */
+void BLI_task_isolate(void (*func)(void *userdata), void *userdata);
+
 #ifdef __cplusplus
 }
 #endif

source/blender/blenlib/intern/task_graph.cc

@@ -91,7 +91,7 @@ struct TaskNode {
 #ifdef WITH_TBB
   tbb::flow::continue_msg run(const tbb::flow::continue_msg UNUSED(input))
   {
-    tbb::this_task_arena::isolate([this] { run_func(task_data); });
+    run_func(task_data);
     return tbb::flow::continue_msg();
   }
 #endif

source/blender/blenlib/intern/task_iterator.c

@@ -237,7 +237,7 @@ static void task_parallel_iterator_do(const TaskParallelSettings *settings,
   void *userdata_chunk_array = NULL;
   const bool use_userdata_chunk = (userdata_chunk_size != 0) && (userdata_chunk != NULL);
 
-  TaskPool *task_pool = BLI_task_pool_create(state, TASK_PRIORITY_HIGH, TASK_ISOLATION_ON);
+  TaskPool *task_pool = BLI_task_pool_create(state, TASK_PRIORITY_HIGH);
 
   if (use_userdata_chunk) {
     userdata_chunk_array = MALLOCA(userdata_chunk_size * num_tasks);
@@ -442,7 +442,7 @@ void BLI_task_parallel_mempool(BLI_mempool *mempool,
     return;
   }
 
-  task_pool = BLI_task_pool_create(&state, TASK_PRIORITY_HIGH, TASK_ISOLATION_ON);
+  task_pool = BLI_task_pool_create(&state, TASK_PRIORITY_HIGH);
   num_threads = BLI_task_scheduler_num_threads();
 
   /* The idea here is to prevent creating task for each of the loop iterations

source/blender/blenlib/intern/task_pool.cc

@@ -22,7 +22,6 @@
 
 #include <cstdlib>
 #include <memory>
-#include <thread>
 #include <utility>
 
 #include "MEM_guardedalloc.h"
@@ -156,7 +155,6 @@ enum TaskPoolType {
 struct TaskPool {
   TaskPoolType type;
   bool use_threads;
-  TaskIsolation task_isolation;
 
   ThreadMutex user_mutex;
   void *userdata;
@@ -164,8 +162,6 @@ struct TaskPool {
 #ifdef WITH_TBB
   /* TBB task pool. */
   TBBTaskGroup tbb_group;
-  /* This is used to detect a common way to accidentally create a deadlock with task isolation. */
-  std::thread::id task_pool_create_thread_id;
 #endif
   volatile bool is_suspended;
   BLI_mempool *suspended_mempool;
@@ -179,33 +175,9 @@ struct TaskPool {
 /* Execute task. */
 void Task::operator()() const
 {
-#ifdef WITH_TBB
-  if (pool->task_isolation == TASK_ISOLATION_ON) {
-    tbb::this_task_arena::isolate([this] { run(pool, taskdata); });
-    return;
-  }
-#endif
   run(pool, taskdata);
 }
 
-static void assert_on_valid_thread(TaskPool *pool)
-{
-  /* TODO: Remove this `return` to enable the check. */
-  return;
-#ifdef DEBUG
-#  ifdef WITH_TBB
-  if (pool->task_isolation == TASK_ISOLATION_ON) {
-    const std::thread::id current_id = std::this_thread::get_id();
-    /* This task pool is modified from different threads. To avoid deadlocks, `TASK_ISOLATION_OFF`
-     * has to be used. Task isolation can still be used in a more fine-grained way within the
-     * tasks, but should not be enabled for the entire task pool. */
-    BLI_assert(pool->task_pool_create_thread_id == current_id);
-  }
-#  endif
-#endif
-  UNUSED_VARS_NDEBUG(pool);
-}
-
 /* TBB Task Pool.
  *
  * Task pool using the TBB scheduler for tasks. When building without TBB
@@ -391,10 +363,7 @@ static void background_task_pool_free(TaskPool *pool)
 
 /* Task Pool */
 
-static TaskPool *task_pool_create_ex(void *userdata,
-                                     TaskPoolType type,
-                                     TaskPriority priority,
-                                     TaskIsolation task_isolation)
+static TaskPool *task_pool_create_ex(void *userdata, TaskPoolType type, TaskPriority priority)
 {
   const bool use_threads = BLI_task_scheduler_num_threads() > 1 && type != TASK_POOL_NO_THREADS;
 
@@ -410,11 +379,6 @@ static TaskPool *task_pool_create_ex(void *userdata,
 
   pool->type = type;
   pool->use_threads = use_threads;
-  pool->task_isolation = task_isolation;
-
-#ifdef WITH_TBB
-  pool->task_pool_create_thread_id = std::this_thread::get_id();
-#endif
 
   pool->userdata = userdata;
   BLI_mutex_init(&pool->user_mutex);
@@ -437,9 +401,9 @@ static TaskPool *task_pool_create_ex(void *userdata,
 /**
  * Create a normal task pool. Tasks will be executed as soon as they are added.
  */
-TaskPool *BLI_task_pool_create(void *userdata, TaskPriority priority, TaskIsolation task_isolation)
+TaskPool *BLI_task_pool_create(void *userdata, TaskPriority priority)
 {
-  return task_pool_create_ex(userdata, TASK_POOL_TBB, priority, task_isolation);
+  return task_pool_create_ex(userdata, TASK_POOL_TBB, priority);
 }
 
 /**
@@ -454,11 +418,9 @@ TaskPool *BLI_task_pool_create(void *userdata, TaskPriority priority, TaskIsolat
  * they could end never being executed, since the 'fallback' background thread is already
  * busy with parent task in single-threaded context).
  */
-TaskPool *BLI_task_pool_create_background(void *userdata,
-                                          TaskPriority priority,
-                                          TaskIsolation task_isolation)
+TaskPool *BLI_task_pool_create_background(void *userdata, TaskPriority priority)
 {
-  return task_pool_create_ex(userdata, TASK_POOL_BACKGROUND, priority, task_isolation);
+  return task_pool_create_ex(userdata, TASK_POOL_BACKGROUND, priority);
 }
 
 /**
@@ -466,11 +428,9 @@ TaskPool *BLI_task_pool_create_background(void *userdata,
  * for until BLI_task_pool_work_and_wait() is called. This helps reducing threading
  * overhead when pushing huge amount of small initial tasks from the main thread.
  */
-TaskPool *BLI_task_pool_create_suspended(void *userdata,
-                                         TaskPriority priority,
-                                         TaskIsolation task_isolation)
+TaskPool *BLI_task_pool_create_suspended(void *userdata, TaskPriority priority)
 {
-  return task_pool_create_ex(userdata, TASK_POOL_TBB_SUSPENDED, priority, task_isolation);
+  return task_pool_create_ex(userdata, TASK_POOL_TBB_SUSPENDED, priority);
 }
 
 /**
@@ -479,8 +439,7 @@ TaskPool *BLI_task_pool_create_suspended(void *userdata,
  */
 TaskPool *BLI_task_pool_create_no_threads(void *userdata)
 {
-  return task_pool_create_ex(
-      userdata, TASK_POOL_NO_THREADS, TASK_PRIORITY_HIGH, TASK_ISOLATION_ON);
+  return task_pool_create_ex(userdata, TASK_POOL_NO_THREADS, TASK_PRIORITY_HIGH);
 }
 
 /**
@@ -489,7 +448,7 @@ TaskPool *BLI_task_pool_create_no_threads(void *userdata)
  */
 TaskPool *BLI_task_pool_create_background_serial(void *userdata, TaskPriority priority)
 {
-  return task_pool_create_ex(userdata, TASK_POOL_BACKGROUND_SERIAL, priority, TASK_ISOLATION_ON);
+  return task_pool_create_ex(userdata, TASK_POOL_BACKGROUND_SERIAL, priority);
 }
 
 void BLI_task_pool_free(TaskPool *pool)
@@ -517,8 +476,6 @@ void BLI_task_pool_push(TaskPool *pool,
                         bool free_taskdata,
                         TaskFreeFunction freedata)
 {
-  assert_on_valid_thread(pool);
-
   Task task(pool, run, taskdata, free_taskdata, freedata);
 
   switch (pool->type) {
@@ -536,8 +493,6 @@ void BLI_task_pool_push(TaskPool *pool,
 
 void BLI_task_pool_work_and_wait(TaskPool *pool)
 {
-  assert_on_valid_thread(pool);
-
   switch (pool->type) {
     case TASK_POOL_TBB:
     case TASK_POOL_TBB_SUSPENDED:

source/blender/blenlib/intern/task_range.cc

@@ -90,13 +90,11 @@ struct RangeTask {
 
   void operator()(const tbb::blocked_range<int> &r) const
   {
-    tbb::this_task_arena::isolate([this, r] {
-      TaskParallelTLS tls;
-      tls.userdata_chunk = userdata_chunk;
-      for (int i = r.begin(); i != r.end(); ++i) {
-        func(userdata, i, &tls);
-      }
-    });
+    TaskParallelTLS tls;
+    tls.userdata_chunk = userdata_chunk;
+    for (int i = r.begin(); i != r.end(); ++i) {
+      func(userdata, i, &tls);
+    }
   }
 
   void join(const RangeTask &other)

source/blender/blenlib/intern/task_scheduler.cc

@@ -28,6 +28,7 @@
 #ifdef WITH_TBB
 /* Need to include at least one header to get the version define. */
 #  include <tbb/blocked_range.h>
+#  include <tbb/task_arena.h>
 #  if TBB_INTERFACE_VERSION_MAJOR >= 10
 #    include <tbb/global_control.h>
 #    define WITH_TBB_GLOBAL_CONTROL
@@ -76,3 +77,12 @@ int BLI_task_scheduler_num_threads()
 {
   return task_scheduler_num_threads;
 }
+
+void BLI_task_isolate(void (*func)(void *userdata), void *userdata)
+{
+#ifdef WITH_TBB
+  tbb::this_task_arena::isolate([&] { func(userdata); });
+#else
+  func(userdata);
+#endif
+}

source/blender/blenlib/tests/BLI_linklist_lockfree_test.cc

@@ -81,7 +81,7 @@ TEST(LockfreeLinkList, InsertMultipleConcurrent)
   LockfreeLinkList list;
   BLI_linklist_lockfree_init(&list);
   /* Initialize task scheduler and pool. */
-  TaskPool *pool = BLI_task_pool_create_suspended(&list, TASK_PRIORITY_HIGH, TASK_ISOLATION_ON);
+  TaskPool *pool = BLI_task_pool_create_suspended(&list, TASK_PRIORITY_HIGH);
   /* Push tasks to the pool. */
   for (int i = 0; i < num_nodes; ++i) {
     BLI_task_pool_push(pool, concurrent_insert, POINTER_FROM_INT(i), false, nullptr);

source/blender/compositor/intern/COM_WorkScheduler.cc

@@ -411,8 +411,7 @@ static void threading_model_task_schedule(WorkPackage *package)
 static void threading_model_task_start()
 {
   BLI_thread_local_create(g_thread_device);
-  g_work_scheduler.task.pool = BLI_task_pool_create(
-      nullptr, TASK_PRIORITY_HIGH, TASK_ISOLATION_ON);
+  g_work_scheduler.task.pool = BLI_task_pool_create(nullptr, TASK_PRIORITY_HIGH);
 }
 
 static void threading_model_task_finish()

source/blender/depsgraph/intern/eval/deg_eval.cc

@@ -353,8 +353,7 @@ static TaskPool *deg_evaluate_task_pool_create(DepsgraphEvalState *state)
     return BLI_task_pool_create_no_threads(state);
   }
 
-  /* TODO: Disable task isolation. */
-  return BLI_task_pool_create_suspended(state, TASK_PRIORITY_HIGH, TASK_ISOLATION_ON);
+  return BLI_task_pool_create_suspended(state, TASK_PRIORITY_HIGH);
 }
 
 /**

source/blender/editors/mesh/editmesh_undo.c

@@ -626,7 +626,7 @@ static void *undomesh_from_editmesh(UndoMesh *um, BMEditMesh *em, Key *key, Undo
 
 #  ifdef USE_ARRAY_STORE_THREAD
     if (um_arraystore.task_pool == NULL) {
-      um_arraystore.task_pool = BLI_task_pool_create_background(NULL, TASK_PRIORITY_LOW, true);
+      um_arraystore.task_pool = BLI_task_pool_create_background(NULL, TASK_PRIORITY_LOW);
     }
 
     struct UMArrayData *um_data = MEM_mallocN(sizeof(*um_data), __func__);

source/blender/editors/render/render_opengl.c

@@ -859,7 +859,7 @@ static bool screen_opengl_render_init(bContext *C, wmOperator *op)
       oglrender->task_pool = BLI_task_pool_create_background_serial(oglrender, TASK_PRIORITY_LOW);
     }
     else {
-      oglrender->task_pool = BLI_task_pool_create(oglrender, TASK_PRIORITY_LOW, TASK_ISOLATION_ON);
+      oglrender->task_pool = BLI_task_pool_create(oglrender, TASK_PRIORITY_LOW);
     }
     oglrender->pool_ok = true;
     BLI_spin_init(&oglrender->reports_lock);

source/blender/editors/sculpt_paint/paint_image_proj.c

@@ -5571,7 +5571,7 @@ static bool project_paint_op(void *state, const float lastpos[2], const float po
   }
 
   if (ps->thread_tot > 1) {
-    task_pool = BLI_task_pool_create_suspended(NULL, TASK_PRIORITY_HIGH, TASK_ISOLATION_ON);
+    task_pool = BLI_task_pool_create_suspended(NULL, TASK_PRIORITY_HIGH);
   }
 
   image_pool = BKE_image_pool_new();

source/blender/editors/space_clip/clip_editor.c

@@ -904,7 +904,7 @@ static void start_prefetch_threads(MovieClip *clip,
   queue.do_update = do_update;
   queue.progress = progress;
 
-  TaskPool *task_pool = BLI_task_pool_create(&queue, TASK_PRIORITY_LOW, TASK_ISOLATION_ON);
+  TaskPool *task_pool = BLI_task_pool_create(&queue, TASK_PRIORITY_LOW);
   for (int i = 0; i < tot_thread; i++) {
     BLI_task_pool_push(task_pool, prefetch_task_func, clip, false, NULL);
   }

source/blender/editors/space_clip/clip_ops.c

@@ -1429,7 +1429,7 @@ static void do_sequence_proxy(void *pjv,
   queue.do_update = do_update;
   queue.progress = progress;
 
-  TaskPool *task_pool = BLI_task_pool_create(&queue, TASK_PRIORITY_LOW, TASK_ISOLATION_ON);
+  TaskPool *task_pool = BLI_task_pool_create(&queue, TASK_PRIORITY_LOW);
   handles = MEM_callocN(sizeof(ProxyThread) * tot_thread, "proxy threaded handles");
   for (int i = 0; i < tot_thread; i++) {
     ProxyThread *handle = &handles[i];

source/blender/editors/space_file/filelist.c

@@ -1553,8 +1553,7 @@ static void filelist_cache_preview_freef(TaskPool *__restrict UNUSED(pool), void
 static void filelist_cache_preview_ensure_running(FileListEntryCache *cache)
 {
   if (!cache->previews_pool) {
-    cache->previews_pool = BLI_task_pool_create_background(
-        cache, TASK_PRIORITY_LOW, TASK_ISOLATION_ON);
+    cache->previews_pool = BLI_task_pool_create_background(cache, TASK_PRIORITY_LOW);
     cache->previews_done = BLI_thread_queue_init();
 
     IMB_thumb_locks_acquire();

source/blender/gpencil_modifiers/intern/lineart/lineart_cpu.c

@@ -482,7 +482,7 @@ static void lineart_main_occlusion_begin(LineartRenderBuffer *rb)
   rb->material.last = rb->material.first;
   rb->edge_mark.last = rb->edge_mark.first;
 
-  TaskPool *tp = BLI_task_pool_create(NULL, TASK_PRIORITY_HIGH, TASK_ISOLATION_ON);
+  TaskPool *tp = BLI_task_pool_create(NULL, TASK_PRIORITY_HIGH);
 
   for (i = 0; i < thread_count; i++) {
     rti[i].thread_id = i;
@@ -1987,7 +1987,7 @@ static void lineart_main_load_geometries(
   }
   DEG_OBJECT_ITER_END;
 
-  TaskPool *tp = BLI_task_pool_create(NULL, TASK_PRIORITY_HIGH, TASK_ISOLATION_ON);
+  TaskPool *tp = BLI_task_pool_create(NULL, TASK_PRIORITY_HIGH);
 
   for (int i = 0; i < thread_count; i++) {
     olti[i].rb = rb;

source/blender/imbuf/intern/imageprocess.c

@@ -523,7 +523,7 @@ void IMB_processor_apply_threaded(
   int total_tasks = (buffer_lines + lines_per_task - 1) / lines_per_task;
   int i, start_line;
 
-  task_pool = BLI_task_pool_create(do_thread, TASK_PRIORITY_LOW, TASK_ISOLATION_ON);
+  task_pool = BLI_task_pool_create(do_thread, TASK_PRIORITY_LOW);
 
   handles = MEM_callocN(handle_size * total_tasks, "processor apply threaded handles");
 

source/blender/modifiers/intern/MOD_nodes_evaluator.cc

@@ -380,7 +380,7 @@ class GeometryNodesEvaluator {
 
   void execute()
   {
-    task_pool_ = BLI_task_pool_create(this, TASK_PRIORITY_HIGH, TASK_ISOLATION_OFF);
+    task_pool_ = BLI_task_pool_create(this, TASK_PRIORITY_HIGH);
 
     this->create_states_for_reachable_nodes();
     this->forward_group_inputs();