On 26/02/2024 20:21, Thomas Hellström wrote:
Hi, Christian

On Fri, 2024-02-23 at 15:30 +0100, Christian König wrote:
Am 06.02.24 um 13:56 schrieb Christian König:
Am 06.02.24 um 13:53 schrieb Thomas Hellström:
Hi, Christian,

On Fri, 2024-01-26 at 15:09 +0100, Christian König wrote:
Previously we would never try to move a BO into the preferred
placements
when it ever landed in a busy placement since those were
considered
compatible.

Rework the whole handling and finally unify the idle and busy
handling.
ttm_bo_validate() is now responsible to try idle placement
first and
then
use the busy placement if that didn't worked.

Drawback is that we now always try the idle placement first for
each
validation which might cause some additional CPU overhead on
overcommit.

v2: fix kerneldoc warning and coding style
v3: take care of XE as well
v4: keep the ttm_bo_mem_space functionality as it is for now,
only
add
      new handling for ttm_bo_validate as suggested by Thomas

Signed-off-by: Christian König <christian.koe...@amd.com>
Reviewed-by: Zack Rusin <zack.ru...@broadcom.com> v3
Sending this through xe CI, will try to review asap.

Take your time. At the moment people are bombarding me with work
and I
have only two hands and one head as well :(

So I've digged myself out of that hole and would rather like to get
this
new feature into 6.9.

Any time to review it? I can also plan some time to review your LRU
changes next week.

Thanks,
Christian.

Sorry for the late response. Was planning to review but saw that there
was still an xe CI failure.

https://intel-gfx-ci.01.org/tree/intel-xe/xe-pw-129579v1/bat-atsm-2/igt@xe_evict_...@evict-overcommit-parallel-nofree-samefd.html

I haven't really had time to look into what might be causing this,
though.
Maybe in ttm_bo_alloc_resource():

@@ -772,7 +772,7 @@ static int ttm_bo_alloc_resource(struct ttm_buffer_object *bo,

                do {
                        ret = ttm_resource_alloc(bo, place, res);
-                       if (unlikely(ret != -ENOSPC))
+                       if (unlikely(ret && ret != -ENOSPC))
                                return ret;
                        if (likely(!ret) || !force_space)
                                break;

Otherwise we allocate VRAM but never correctly synchronise against the move fence, since we missed adding it to the BO. When we trigger async evictions that would explain the above test failure where we detect VRAM corruption, since someone else is still using the VRAM we allocated. What do you think?


/Thomas



Christian.


/Thomas


---
   drivers/gpu/drm/ttm/ttm_bo.c       | 231 +++++++++++++-------
-------
--
   drivers/gpu/drm/ttm/ttm_resource.c |  16 +-
   include/drm/ttm/ttm_resource.h     |   3 +-
   3 files changed, 121 insertions(+), 129 deletions(-)

diff --git a/drivers/gpu/drm/ttm/ttm_bo.c
b/drivers/gpu/drm/ttm/ttm_bo.c
index ba3f09e2d7e6..b12f435542a9 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -724,64 +724,36 @@ static int ttm_bo_add_move_fence(struct
ttm_buffer_object *bo,
       return ret;
   }
   -/*
- * Repeatedly evict memory from the LRU for @mem_type until we
create enough
- * space, or we've evicted everything and there isn't enough
space.
- */
-static int ttm_bo_mem_force_space(struct ttm_buffer_object
*bo,
-                  const struct ttm_place *place,
-                  struct ttm_resource **mem,
-                  struct ttm_operation_ctx *ctx)
-{
-    struct ttm_device *bdev = bo->bdev;
-    struct ttm_resource_manager *man;
-    struct ww_acquire_ctx *ticket;
-    int ret;
-
-    man = ttm_manager_type(bdev, place->mem_type);
-    ticket = dma_resv_locking_ctx(bo->base.resv);
-    do {
-        ret = ttm_resource_alloc(bo, place, mem);
-        if (likely(!ret))
-            break;
-        if (unlikely(ret != -ENOSPC))
-            return ret;
-        ret = ttm_mem_evict_first(bdev, man, place, ctx,
-                      ticket);
-        if (unlikely(ret != 0))
-            return ret;
-    } while (1);
-
-    return ttm_bo_add_move_fence(bo, man, *mem, ctx-
no_wait_gpu);
-}
-
   /**
- * ttm_bo_mem_space
+ * ttm_bo_alloc_resource - Allocate backing store for a BO
    *
- * @bo: Pointer to a struct ttm_buffer_object. the data of
which
- * we want to allocate space for.
- * @placement: Proposed new placement for the buffer object.
- * @mem: A struct ttm_resource.
+ * @bo: Pointer to a struct ttm_buffer_object of which we want
a
resource for
+ * @placement: Proposed new placement for the buffer object
    * @ctx: if and how to sleep, lock buffers and alloc memory
+ * @force_space: If we should evict buffers to force space
+ * @res: The resulting struct ttm_resource.
    *
- * Allocate memory space for the buffer object pointed to by
@bo,
using
- * the placement flags in @placement, potentially evicting
other
idle buffer objects.
- * This function may sleep while waiting for space to become
available.
+ * Allocates a resource for the buffer object pointed to by
@bo,
using the
+ * placement flags in @placement, potentially evicting other
buffer
objects when
+ * @force_space is true.
+ * This function may sleep while waiting for resources to
become
available.
    * Returns:
- * -EBUSY: No space available (only if no_wait == 1).
+ * -EBUSY: No space available (only if no_wait == true).
    * -ENOSPC: Could not allocate space for the buffer object,
either
due to
    * fragmentation or concurrent allocators.
    * -ERESTARTSYS: An interruptible sleep was interrupted by a
signal.
    */
-int ttm_bo_mem_space(struct ttm_buffer_object *bo,
-            struct ttm_placement *placement,
-            struct ttm_resource **mem,
-            struct ttm_operation_ctx *ctx)
+static int ttm_bo_alloc_resource(struct ttm_buffer_object *bo,
+                 struct ttm_placement *placement,
+                 struct ttm_operation_ctx *ctx,
+                 bool force_space,
+                 struct ttm_resource **res)
   {
       struct ttm_device *bdev = bo->bdev;
-    bool type_found = false;
+    struct ww_acquire_ctx *ticket;
       int i, ret;
   +    ticket = dma_resv_locking_ctx(bo->base.resv);
       ret = dma_resv_reserve_fences(bo->base.resv, 1);
       if (unlikely(ret))
           return ret;
@@ -790,98 +762,73 @@ int ttm_bo_mem_space(struct
ttm_buffer_object
*bo,
           const struct ttm_place *place = &placement-
placement[i];
           struct ttm_resource_manager *man;
   -        if (place->flags & TTM_PL_FLAG_FALLBACK)
-            continue;
-
           man = ttm_manager_type(bdev, place->mem_type);
           if (!man || !ttm_resource_manager_used(man))
               continue;
   -        type_found = true;
-        ret = ttm_resource_alloc(bo, place, mem);
-        if (ret == -ENOSPC)
+        if (place->flags & (force_space ?
TTM_PL_FLAG_DESIRED :
+                    TTM_PL_FLAG_FALLBACK))
+            continue;
+
+        do {
+            ret = ttm_resource_alloc(bo, place, res);
+            if (unlikely(ret != -ENOSPC))
+                return ret;
+            if (likely(!ret) || !force_space)
+                break;
+
+            ret = ttm_mem_evict_first(bdev, man, place,
ctx,
+                          ticket);
+            if (unlikely(ret == -EBUSY))
+                break;
+            if (unlikely(ret))
+                return ret;
+        } while (1);
+        if (ret)
               continue;
-        if (unlikely(ret))
-            goto error;
   -        ret = ttm_bo_add_move_fence(bo, man, *mem, ctx-
no_wait_gpu);
+        ret = ttm_bo_add_move_fence(bo, man, *res, ctx-
no_wait_gpu);
           if (unlikely(ret)) {
-            ttm_resource_free(bo, mem);
+            ttm_resource_free(bo, res);
               if (ret == -EBUSY)
                   continue;
   -            goto error;
+            return ret;
           }
           return 0;
       }
   -    for (i = 0; i < placement->num_placement; ++i) {
-        const struct ttm_place *place = &placement-
placement[i];
-        struct ttm_resource_manager *man;
-
-        if (place->flags & TTM_PL_FLAG_DESIRED)
-            continue;
-
-        man = ttm_manager_type(bdev, place->mem_type);
-        if (!man || !ttm_resource_manager_used(man))
-            continue;
-
-        type_found = true;
-        ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
-        if (likely(!ret))
-            return 0;
-
-        if (ret && ret != -EBUSY)
-            goto error;
-    }
-
-    ret = -ENOSPC;
-    if (!type_found) {
-        pr_err(TTM_PFX "No compatible memory type found\n");
-        ret = -EINVAL;
-    }
-
-error:
-    return ret;
+    return -ENOSPC;
   }
-EXPORT_SYMBOL(ttm_bo_mem_space);
   -static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
-                  struct ttm_placement *placement,
-                  struct ttm_operation_ctx *ctx)
+/*
+ * ttm_bo_mem_space - Wrapper around ttm_bo_alloc_resource
+ *
+ * @bo: Pointer to a struct ttm_buffer_object of which we want
a
resource for
+ * @placement: Proposed new placement for the buffer object
+ * @res: The resulting struct ttm_resource.
+ * @ctx: if and how to sleep, lock buffers and alloc memory
+ *
+ * Tries both idle allocation and forcefully eviction of
buffers.
See
+ * ttm_bo_alloc_resource for details.
+ */
+int ttm_bo_mem_space(struct ttm_buffer_object *bo,
+             struct ttm_placement *placement,
+             struct ttm_resource **res,
+             struct ttm_operation_ctx *ctx)
   {
-    struct ttm_resource *mem;
-    struct ttm_place hop;
+    bool force_space = false;
       int ret;
   -    dma_resv_assert_held(bo->base.resv);
+    do {
+        ret = ttm_bo_alloc_resource(bo, placement, ctx,
+                        force_space, res);
+        force_space = !force_space;
+    } while (ret == -ENOSPC && force_space);
   -    /*
-     * Determine where to move the buffer.
-     *
-     * If driver determines move is going to need
-     * an extra step then it will return -EMULTIHOP
-     * and the buffer will be moved to the temporary
-     * stop and the driver will be called to make
-     * the second hop.
-     */
-    ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
-    if (ret)
-        return ret;
-bounce:
-    ret = ttm_bo_handle_move_mem(bo, mem, false, ctx, &hop);
-    if (ret == -EMULTIHOP) {
-        ret = ttm_bo_bounce_temp_buffer(bo, &mem, ctx,
&hop);
-        if (ret)
-            goto out;
-        /* try and move to final place now. */
-        goto bounce;
-    }
-out:
-    if (ret)
-        ttm_resource_free(bo, &mem);
       return ret;
   }
+EXPORT_SYMBOL(ttm_bo_mem_space);
     /**
    * ttm_bo_validate
@@ -902,6 +849,9 @@ int ttm_bo_validate(struct
ttm_buffer_object *bo,
               struct ttm_placement *placement,
               struct ttm_operation_ctx *ctx)
   {
+    struct ttm_resource *res;
+    struct ttm_place hop;
+    bool force_space;
       int ret;
         dma_resv_assert_held(bo->base.resv);
@@ -912,20 +862,53 @@ int ttm_bo_validate(struct
ttm_buffer_object
*bo,
       if (!placement->num_placement)
           return ttm_bo_pipeline_gutting(bo);
   -    /* Check whether we need to move buffer. */
-    if (bo->resource && ttm_resource_compatible(bo->resource,
placement))
-        return 0;
+    force_space = false;
+    do {
+        /* Check whether we need to move buffer. */
+        if (bo->resource &&
+            ttm_resource_compatible(bo->resource, placement,
+                        force_space))
+            return 0;
   -    /* Moving of pinned BOs is forbidden */
-    if (bo->pin_count)
-        return -EINVAL;
+        /* Moving of pinned BOs is forbidden */
+        if (bo->pin_count)
+            return -EINVAL;
+
+        /*
+         * Determine where to move the buffer.
+         *
+         * If driver determines move is going to need
+         * an extra step then it will return -EMULTIHOP
+         * and the buffer will be moved to the temporary
+         * stop and the driver will be called to make
+         * the second hop.
+         */
+        ret = ttm_bo_alloc_resource(bo, placement, ctx,
force_space,
+                        &res);
+        force_space = !force_space;
+        if (ret == -ENOSPC)
+            continue;
+        if (ret)
+            return ret;
+
+bounce:
+        ret = ttm_bo_handle_move_mem(bo, res, false, ctx,
&hop);
+        if (ret == -EMULTIHOP) {
+            ret = ttm_bo_bounce_temp_buffer(bo, &res,
ctx, &hop);
+            /* try and move to final place now. */
+            if (!ret)
+                goto bounce;
+        }
+        if (ret) {
+            ttm_resource_free(bo, &res);
+            return ret;
+        }
+
+    } while (ret && force_space);
   -    ret = ttm_bo_move_buffer(bo, placement, ctx);
       /* For backward compatibility with userspace */
       if (ret == -ENOSPC)
           return -ENOMEM;
-    if (ret)
-        return ret;
         /*
        * We might need to add a TTM.
diff --git a/drivers/gpu/drm/ttm/ttm_resource.c
b/drivers/gpu/drm/ttm/ttm_resource.c
index fb14f7716cf8..65155f2013ca 100644
--- a/drivers/gpu/drm/ttm/ttm_resource.c
+++ b/drivers/gpu/drm/ttm/ttm_resource.c
@@ -295,11 +295,13 @@ bool ttm_resource_intersects(struct
ttm_device
*bdev,
    *
    * @res: the resource to check
    * @placement: the placement to check against
+ * @evicting: true if the caller is doing evictions
    *
    * Returns true if the placement is compatible.
    */
   bool ttm_resource_compatible(struct ttm_resource *res,
-                 struct ttm_placement *placement)
+                 struct ttm_placement *placement,
+                 bool evicting)
   {
       struct ttm_buffer_object *bo = res->bo;
       struct ttm_device *bdev = bo->bdev;
@@ -315,14 +317,20 @@ bool ttm_resource_compatible(struct
ttm_resource *res,
           if (res->mem_type != place->mem_type)
               continue;
   +        if (place->flags & (evicting ? TTM_PL_FLAG_DESIRED :
+                    TTM_PL_FLAG_FALLBACK))
+            continue;
+
+        if (place->flags & TTM_PL_FLAG_CONTIGUOUS &&
+            !(res->placement & TTM_PL_FLAG_CONTIGUOUS))
+            continue;
+
           man = ttm_manager_type(bdev, res->mem_type);
           if (man->func->compatible &&
               !man->func->compatible(man, res, place, bo-
base.size))
               continue;
   -        if ((!(place->flags & TTM_PL_FLAG_CONTIGUOUS) ||
-             (res->placement & TTM_PL_FLAG_CONTIGUOUS)))
-            return true;
+        return true;
       }
       return false;
   }
diff --git a/include/drm/ttm/ttm_resource.h
b/include/drm/ttm/ttm_resource.h
index 1afa13f0c22b..7561023db43d 100644
--- a/include/drm/ttm/ttm_resource.h
+++ b/include/drm/ttm/ttm_resource.h
@@ -366,7 +366,8 @@ bool ttm_resource_intersects(struct
ttm_device
*bdev,
                    const struct ttm_place *place,
                    size_t size);
   bool ttm_resource_compatible(struct ttm_resource *res,
-                 struct ttm_placement *placement);
+                 struct ttm_placement *placement,
+                 bool evicting);
   void ttm_resource_set_bo(struct ttm_resource *res,
                struct ttm_buffer_object *bo);



Reply via email to