habanalabs: add MMU DRAM default page mapping

This patch provides a workaround for a H/W bug in Goya, where access to
RAZWI from TPC can cause PCI completion timeout.

The WA is to use the device MMU to map any unmapped DRAM memory to a
default page in the DRAM. That way, the TPC will never reach RAZWI upon
accessing a bad address in the DRAM.

When a DRAM page is mapped by the user, its default mapping is
overwritten. Once that page is unmapped, the MMU driver will map that page
to the default page.

To help debugging, the driver will set the default page area to 0x99 on
device initialization.

Signed-off-by: Omer Shpigelman <oshpigelman@habana.ai>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

1 files changed, 7 insertions, 5 deletions

diff --git a/drivers/misc/habanalabs/memory.c b/drivers/misc/habanalabs/memory.c
index 660cf67258fd..3a12fd1a5274 100644
--- a/drivers/misc/habanalabs/memory.c
+++ b/drivers/misc/habanalabs/memory.c
@@ -925,8 +925,7 @@ static int map_device_va(struct hl_ctx *ctx, struct hl_mem_in *args,
 		goto map_err;
 	}
 
-	hdev->asic_funcs->mmu_invalidate_cache_range(hdev, false, ctx->asid,
-			ret_vaddr, phys_pg_pack->total_size);
+	hdev->asic_funcs->mmu_invalidate_cache(hdev, false);
 
 	mutex_unlock(&ctx->mmu_lock);
 
@@ -1050,8 +1049,7 @@ static int unmap_device_va(struct hl_ctx *ctx, u64 vaddr)
 			dev_warn_ratelimited(hdev->dev,
 				"unmap failed for vaddr: 0x%llx\n", next_vaddr);
 
-	hdev->asic_funcs->mmu_invalidate_cache_range(hdev, true, ctx->asid,
-			vaddr, phys_pg_pack->total_size);
+	hdev->asic_funcs->mmu_invalidate_cache(hdev, true);
 
 	mutex_unlock(&ctx->mmu_lock);
 
@@ -1455,7 +1453,11 @@ static int hl_vm_ctx_init_with_ranges(struct hl_ctx *ctx, u64 host_range_start,
 	struct hl_device *hdev = ctx->hdev;
 	int rc;
 
-	hl_mmu_ctx_init(ctx);
+	rc = hl_mmu_ctx_init(ctx);
+	if (rc) {
+		dev_err(hdev->dev, "failed to init context %d\n", ctx->asid);
+		return rc;
+	}
 
 	mutex_init(&ctx->mem_hash_lock);
 	hash_init(ctx->mem_hash);