If we free the bo, then the PTE may get deallocated immediately. We have
to make sure that the submission includes a ref to the old bo so that it
remains mapped for the duration of the command execution.
Signed-off-by: Ilia Mirkin <imir...@alum.mit.edu>
---
v1 -> v2:
Ben pointed out that not deleting the BO is not enough -- we have to
actually ref it to the pushbuf, otherwise it could still get kicked
out of VRAM.
I happened to notice that the TLS logic suffers from a similar issue,
although in practice it doesn't matter since we never resize TLS.
src/gallium/drivers/nouveau/nvc0/nvc0_screen.c | 13 +++++++++++++
1 file changed, 13 insertions(+)
diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c
b/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c
index 02562dd4eb2..cac4bb89271 100644
--- a/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c
@@ -741,6 +741,13 @@ nvc0_screen_resize_tls_area(struct nvc0_screen *screen,
NULL, &bo);
if (ret)
return ret;
+
+ /* Make sure that the pushbuf has acquired a reference to the old tls
+ * segment, as it may have commands that will reference it.
+ */
+ if (screen->tls)
+ PUSH_REFN(screen->base.pushbuf, screen->tls,
+ NV_VRAM_DOMAIN(&screen->base) | NOUVEAU_BO_RDWR);
nouveau_bo_ref(NULL, &screen->tls);
screen->tls = bo;
return 0;
@@ -758,6 +765,12 @@ nvc0_screen_resize_text_area(struct nvc0_screen *screen,
uint64_t size)
if (ret)
return ret;
+ /* Make sure that the pushbuf has acquired a reference to the old text
+ * segment, as it may have commands that will reference it.
+ */
+ if (screen->text)
+ PUSH_REFN(push, screen->text,
+ NV_VRAM_DOMAIN(&screen->base) | NOUVEAU_BO_RD);
nouveau_bo_ref(NULL, &screen->text);
screen->text = bo;
--
2.13.6
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
