radeonsi: stop using the VM_ALWAYS_VALID flag

Allocation all the bo as ALWAYS_VALID means they must all fit in memory
(vram + gtt) at each command submission.
This causes some trouble when the total allocated memory is greater than
the available memory.

Possible solutions:
- being able to tag/untag a bo as ALWAYS_VALID: would require kernel changes
- disable VM_ALWAYS_VALID when memory usage is more than a percentage of the
  available memory
- disable VM_ALWAYS_VALID entirely

v1 of this patch implemented option 2. v2 (this version) implements option 3.

Related issues:
 - https://gitlab.freedesktop.org/drm/amd/issues/607
 - https://gitlab.freedesktop.org/mesa/mesa/issues/1257

It also helps with some piglit tests (-t maxsize -t "max[_-].*size" -t maxuniformblocksize):
instead of crashing the machine, the tests fail cleanly.

Closes: https://gitlab.freedesktop.org/mesa/mesa/issues/2190
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3430>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3430>

diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
index e6a5a7773e0798139124981a8463a49715c8caa6..3762b9d0c44b61c7eea1658d1dcf2a6ecdc0a602 100644 (file)
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
@@ -36,10 +36,6 @@
 #include <stdio.h>
 #include <inttypes.h>
 
-#ifndef AMDGPU_GEM_CREATE_VM_ALWAYS_VALID
-#define AMDGPU_GEM_CREATE_VM_ALWAYS_VALID (1 << 6)
-#endif
-
 #ifndef AMDGPU_VA_RANGE_HIGH
 #define AMDGPU_VA_RANGE_HIGH   0x2
 #endif
@@ -504,9 +500,6 @@ static struct amdgpu_winsys_bo *amdgpu_create_bo(struct amdgpu_winsys *ws,
       request.flags |= AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
    if (flags & RADEON_FLAG_GTT_WC)
       request.flags |= AMDGPU_GEM_CREATE_CPU_GTT_USWC;
-   if (flags & RADEON_FLAG_NO_INTERPROCESS_SHARING &&
-       ws->info.has_local_buffers)
-      request.flags |= AMDGPU_GEM_CREATE_VM_ALWAYS_VALID;
    if (ws->zero_all_vram_allocs &&
        (request.preferred_heap & AMDGPU_GEM_DOMAIN_VRAM))
       request.flags |= AMDGPU_GEM_CREATE_VRAM_CLEARED;
@@ -556,7 +549,6 @@ static struct amdgpu_winsys_bo *amdgpu_create_bo(struct amdgpu_winsys *ws,
    bo->u.real.va_handle = va_handle;
    bo->initial_domain = initial_domain;
    bo->unique_id = __sync_fetch_and_add(&ws->next_bo_unique_id, 1);
-   bo->is_local = !!(request.flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID);
 
    if (initial_domain & RADEON_DOMAIN_VRAM)
       ws->allocated_vram += align64(size, ws->info.gart_page_size);

diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.h b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.h
index 18ef7e1fa51d7cc27b5465ef0a7f11aa51f8881e..fa8aabbeede95e47f3cdc2152564ca114e602727 100644 (file)
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.h
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.h
@@ -93,7 +93,6 @@ struct amdgpu_winsys_bo {
    amdgpu_bo_handle bo; /* NULL for slab entries and sparse buffers */
    bool sparse;
    bool is_user_ptr;
-   bool is_local;
    uint32_t unique_id;
    uint64_t va;
    enum radeon_bo_domain initial_domain;

diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
index 47973d00d1d48f5f5c4e6db5da976a19de0f4cf4..4308cfb3f96661e99d92672a22e985220fc370bc 100644 (file)
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
@@ -1398,9 +1398,6 @@ void amdgpu_cs_submit_ib(void *job, int thread_index)
 
       simple_mtx_lock(&ws->global_bo_list_lock);
       LIST_FOR_EACH_ENTRY(bo, &ws->global_bo_list, u.real.global_list_item) {
-         if (bo->is_local)
-            continue;
-
          list[num_handles].bo_handle = bo->u.real.kms_handle;
          list[num_handles].bo_priority = 0;
          ++num_handles;
@@ -1425,10 +1422,6 @@ void amdgpu_cs_submit_ib(void *job, int thread_index)
       unsigned num_handles = 0;
       for (i = 0; i < cs->num_real_buffers; ++i) {
          struct amdgpu_cs_buffer *buffer = &cs->real_buffers[i];
-
-         if (buffer->bo->is_local)
-            continue;
-
          assert(buffer->u.real.priority_usage != 0);
 
          list[num_handles].bo_handle = buffer->bo->u.real.kms_handle;