Attempt at restoring xnack support in ROCR-Runtime and AMDGPU

--- a/src/core/runtime/isa.cpp	2024-07-19 23:55:42.477579901 -0000
+++ b/src/core/runtime/isa.cpp	2024-07-19 23:55:42.477579903 -0000
@@ -142,3 +142,4 @@ bool Isa::GetInfo(const hsa_isa_info_t &
       if (this->GetVersion() == Version(7, 0, 0) ||
-          this->GetVersion() == Version(8, 0, 1)) {
+          this->GetVersion() == Version(8, 0, 1) ||
+          this->GetVersion() == Version(8, 0, 3)) {
         profiles[1] = true;
@@ -252,3 +253,5 @@ constexpr size_t hsa_name_size = 63;
   ISAREG_ENTRY_GEN("gfx802",                 8, 0, 2,  unsupported, unsupported, 64)
-  ISAREG_ENTRY_GEN("gfx803",                 8, 0, 3,  unsupported, unsupported, 64)
+  ISAREG_ENTRY_GEN("gfx803",                 8, 0, 3,  unsupported, any,         64)
+  ISAREG_ENTRY_GEN("gfx803:xnack-",          8, 0, 3,  unsupported, disabled,    64)
+  ISAREG_ENTRY_GEN("gfx803:xnack+",          8, 0, 3,  unsupported, enabled,     64)
   ISAREG_ENTRY_GEN("gfx805",                 8, 0, 5,  unsupported, unsupported, 64)
@@ -348,3 +351,5 @@ constexpr size_t hsa_name_size = 63;
   ISAREG_ENTRY_GEN("gfx1102",                11, 0, 2, unsupported, unsupported, 32)
-  ISAREG_ENTRY_GEN("gfx1103",                11, 0, 3, unsupported, unsupported, 32)
+  ISAREG_ENTRY_GEN("gfx1103",                11, 0, 3, unsupported, any,         32)
+  ISAREG_ENTRY_GEN("gfx1103:xnack-",         11, 0, 3, unsupported, disabled,    32)
+  ISAREG_ENTRY_GEN("gfx1103:xnack+",         11, 0, 3, unsupported, enabled,     32)
   ISAREG_ENTRY_GEN("gfx1150",                11, 5, 0, unsupported, unsupported, 32)
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c	2024-07-19 23:28:10.520865455 -0000
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c	2024-07-19 23:28:10.520865456 -0000
@@ -58,8 +58,8 @@ static struct workqueue_struct *kfd_proc

-/* Ordered, single-threaded workqueue for restoring evicted
- * processes. Restoring multiple processes concurrently under memory
- * pressure can lead to processes blocking each other from validating
- * their BOs and result in a live-lock situation where processes
- * remain evicted indefinitely.
- */
+ /* Ordered, single-threaded workqueue for restoring evicted
+  * processes. Restoring multiple processes concurrently under memory
+  * pressure can lead to processes blocking each other from validating
+  * their BOs and result in a live-lock situation where processes
+  * remain evicted indefinitely.
+  */
 static struct workqueue_struct *kfd_restore_wq;
@@ -1420,8 +1420,2 @@ bool kfd_process_xnack_mode(struct kfd_p

-		/* Only consider GFXv9 and higher GPUs. Older GPUs don't
-		 * support the SVM APIs and don't need to be considered
-		 * for the XNACK mode selection.
-		 */
-		if (!KFD_IS_SOC15(dev))
-			continue;
 		/* Aldebaran can always support XNACK because it can support
@@ -1438,12 +1432,17 @@ bool kfd_process_xnack_mode(struct kfd_p

-		/* GFXv10 and later GPUs do not support shader preemption
-		 * during page faults. This can lead to poor QoS for queue
-		 * management and memory-manager-related preemptions or
-		 * even deadlocks.
-		 */
-		if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1))
-			return false;
-
-		if (dev->kfd->noretry)
+		if (KFD_GC_VERSION(dev) >= IP_VERSION(8, 0, 0)) {
+			pr_warn("AMD GPUs do not properly support shader preemption");
+			pr_warn(" during page faults.\n");
+			pr_warn("This may lead to poor QoS for queue management or");
+			pr_warn(" memory-manager-related preemptions and deadlocks.\n");
+		}
+
+		if (dev->kfd->noretry) {
+			pr_err("Xnack requires noretry be disabled!\n");
+			pr_err("Via kernel cmdline: amdgpu.noretry=0\n");
+			pr_err("During module loading: printf 'options amdgpu noretry=0'");
+			pr_err(" >>/etc/modprobe.d/amdgpu.conf\n");
+			pr_err("Manually: sudo printf 0 >/sys/module/amdgpu/parameters/noretry\n");
 			return false;
+		}
 	}