nouveau: synchronize "scratch runout" destruction with the command stream

When nvc0_push_vbo calls nouveau_scratch_done it does not mean
scratch buffers can be freed immediately. It means "when hardware
advances to this place in the command stream the scratch buffers
can be freed".

To fix it, just postpone scratch runout destruction after current
fence is signalled.

The bug existed for a very long time. Nobody noticed, because
"scratch runout" code path is rarely executed.

Fixes hang at the very beginning of first mission in "Serious Sam 3"
on nve7/gk107. It manifested as:

nouveau E[   PFIFO][0000:01:00.0] read fault at 0x000a9e0000 [PTE] from GR/GPC0/PE_2 on channel 0x007f853000 [Sam3[17056]]

Cc: "10.4 10.5" <[email protected]>
Reviewed-by: Ilia Mirkin <[email protected]>

1 files changed, 32 insertions, 16 deletions

diff --git a/src/gallium/drivers/nouveau/nouveau_buffer.c b/src/gallium/drivers/nouveau/nouveau_buffer.c
index 49ff100c4ec..32fa65c8a51 100644
--- a/src/gallium/drivers/nouveau/nouveau_buffer.c
+++ b/src/gallium/drivers/nouveau/nouveau_buffer.c
@@ -846,17 +846,28 @@ nouveau_scratch_bo_alloc(struct nouveau_context *nv, struct nouveau_bo **pbo,
                          4096, size, NULL, pbo);
 }
 
+static void
+nouveau_scratch_unref_bos(void *d)
+{
+   struct runout *b = d;
+   int i;
+
+   for (i = 0; i < b->nr; ++i)
+      nouveau_bo_ref(NULL, &b->bo[i]);
+
+   FREE(b);
+}
+
 void
 nouveau_scratch_runout_release(struct nouveau_context *nv)
 {
-   if (!nv->scratch.nr_runout)
+   if (!nv->scratch.runout)
+      return;
+
+   if (!nouveau_fence_work(nv->screen->fence.current, nouveau_scratch_unref_bos,
+         nv->scratch.runout))
       return;
-   do {
-      --nv->scratch.nr_runout;
-      nouveau_bo_ref(NULL, &nv->scratch.runout[nv->scratch.nr_runout]);
-   } while (nv->scratch.nr_runout);
 
-   FREE(nv->scratch.runout);
    nv->scratch.end = 0;
    nv->scratch.runout = NULL;
 }
@@ -868,21 +879,26 @@ static INLINE boolean
 nouveau_scratch_runout(struct nouveau_context *nv, unsigned size)
 {
    int ret;
-   const unsigned n = nv->scratch.nr_runout++;
+   unsigned n;
 
-   nv->scratch.runout = REALLOC(nv->scratch.runout,
-                                (n + 0) * sizeof(*nv->scratch.runout),
-                                (n + 1) * sizeof(*nv->scratch.runout));
-   nv->scratch.runout[n] = NULL;
-
-   ret = nouveau_scratch_bo_alloc(nv, &nv->scratch.runout[n], size);
+   if (nv->scratch.runout)
+      n = nv->scratch.runout->nr;
+   else
+      n = 0;
+   nv->scratch.runout = REALLOC(nv->scratch.runout, n == 0 ? 0 :
+                                (sizeof(*nv->scratch.runout) + (n + 0) * sizeof(void *)),
+                                 sizeof(*nv->scratch.runout) + (n + 1) * sizeof(void *));
+   nv->scratch.runout->nr = n + 1;
+   nv->scratch.runout->bo[n] = NULL;
+
+   ret = nouveau_scratch_bo_alloc(nv, &nv->scratch.runout->bo[n], size);
    if (!ret) {
-      ret = nouveau_bo_map(nv->scratch.runout[n], 0, NULL);
+      ret = nouveau_bo_map(nv->scratch.runout->bo[n], 0, NULL);
       if (ret)
-         nouveau_bo_ref(NULL, &nv->scratch.runout[--nv->scratch.nr_runout]);
+         nouveau_bo_ref(NULL, &nv->scratch.runout->bo[--nv->scratch.runout->nr]);
    }
    if (!ret) {
-      nv->scratch.current = nv->scratch.runout[n];
+      nv->scratch.current = nv->scratch.runout->bo[n];
       nv->scratch.offset = 0;
       nv->scratch.end = size;
       nv->scratch.map = nv->scratch.current->map;