570.124.04

Author: niv

README.md

@@ -1,7 +1,7 @@
 # NVIDIA Linux Open GPU Kernel Module Source
 
 This is the source release of the NVIDIA Linux open GPU kernel modules,
-version 570.86.16.
+version 570.124.04.
 
 
 ## How to Build
@@ -17,7 +17,7 @@ as root:
 
 Note that the kernel modules built here must be used with GSP
 firmware and user-space NVIDIA GPU driver components from a corresponding
-570.86.16 driver release.  This can be achieved by installing
+570.124.04 driver release.  This can be achieved by installing
 the NVIDIA GPU driver from the .run file using the `--no-kernel-modules`
 option.  E.g.,
 
@@ -185,7 +185,7 @@ table below).
 For details on feature support and limitations, see the NVIDIA GPU driver
 end user README here:
 
-https://us.download.nvidia.com/XFree86/Linux-x86_64/570.86.16/README/kernel_open.html
+https://us.download.nvidia.com/XFree86/Linux-x86_64/570.124.04/README/kernel_open.html
 
 For vGPU support, please refer to the README.vgpu packaged in the vGPU Host
 Package for more details.
@@ -938,10 +938,17 @@ Subsystem Device ID.
 | NVIDIA RTX 500 Ada Generation Laptop GPU        | 28BB           |
 | NVIDIA GeForce RTX 4060 Laptop GPU              | 28E0           |
 | NVIDIA GeForce RTX 4050 Laptop GPU              | 28E1           |
+| NVIDIA GeForce RTX 3050 A Laptop GPU            | 28E3           |
 | NVIDIA RTX 2000 Ada Generation Embedded GPU     | 28F8           |
 | NVIDIA B200                                     | 2901 10DE 1999 |
 | NVIDIA B200                                     | 2901 10DE 199B |
 | NVIDIA B200                                     | 2901 10DE 20DA |
+| HGX GB200                                       | 2941 10DE 2046 |
+| HGX GB200                                       | 2941 10DE 20CA |
+| HGX GB200                                       | 2941 10DE 20D5 |
+| HGX GB200                                       | 2941 10DE 21C9 |
+| HGX GB200                                       | 2941 10DE 21CA |
 | NVIDIA GeForce RTX 5090                         | 2B85           |
 | NVIDIA GeForce RTX 5090 D                       | 2B87           |
 | NVIDIA GeForce RTX 5080                         | 2C02           |
+| NVIDIA GeForce RTX 5070 Ti                      | 2C05           |

kernel-open/Kbuild

@@ -86,7 +86,7 @@ EXTRA_CFLAGS += -I$(src)/common/inc
 EXTRA_CFLAGS += -I$(src)
 EXTRA_CFLAGS += -Wall $(DEFINES) $(INCLUDES) -Wno-cast-qual -Wno-format-extra-args
 EXTRA_CFLAGS += -D__KERNEL__ -DMODULE -DNVRM
-EXTRA_CFLAGS += -DNV_VERSION_STRING=\"570.86.16\"
+EXTRA_CFLAGS += -DNV_VERSION_STRING=\"570.124.04\"
 
 ifneq ($(SYSSRCHOST1X),)
  EXTRA_CFLAGS += -I$(SYSSRCHOST1X)

kernel-open/common/inc/nv_uvm_types.h

@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2014-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 2014-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  * SPDX-License-Identifier: MIT
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
@@ -756,6 +756,8 @@ typedef struct UvmGpuFbInfo_tag
     NvBool bStaticBar1Enabled; // Static BAR1 mode is enabled
     NvU64  staticBar1StartOffset;  // The start offset of the the static mapping
     NvU64  staticBar1Size;     // The size of the static mapping
+    NvU32  heapStart;          // The start offset of heap in KB, helpful for MIG
+                               // systems
 } UvmGpuFbInfo;
 
 typedef struct UvmGpuEccInfo_tag

kernel-open/conftest.sh

@@ -6307,6 +6307,32 @@ compile_test() {
             compile_check_conftest "$CODE" "NV_NUM_REGISTERED_FB_PRESENT" "" "types"
         ;;
 
+        acpi_video_register_backlight)
+            #
+            # Determine if acpi_video_register_backlight() function is present
+            #
+            # acpi_video_register_backlight was added by commit 3dbc80a3e4c55c
+            # (ACPI: video: Make backlight class device registration a separate
+            # step (v2)) for v6.0 (2022-09-02).
+            # Note: the include directive for <linux/types> in this conftest is
+            # necessary in order to support kernels between commit 0b9f7d93ca61
+            # ("ACPI / i915: ignore firmware requests backlight change") for
+            # v3.16 (2014-07-07) and commit 3bd6bce369f5 ("ACPI / video: Port
+            # to new backlight interface selection API") for v4.2 (2015-07-16).
+            # Kernels within this range use the 'bool' type and the related
+            # 'false' value in <acpi/video.h> without first including the
+            # definitions of that type and value.
+            #
+            CODE="
+            #include <linux/types.h>
+            #include <acpi/video.h>
+            void conftest_acpi_video_register_backlight(void) {
+                acpi_video_register_backlight(0);
+            }"
+
+            compile_check_conftest "$CODE" "NV_ACPI_VIDEO_REGISTER_BACKLIGHT" "" "functions"
+        ;;
+
         acpi_video_backlight_use_native)
             #
             # Determine if acpi_video_backlight_use_native() function is present
@@ -6690,13 +6716,18 @@ compile_test() {
             #
             # Determine whether drm_client_setup is present.
             #
-            # Added by commit d07fdf922592 ("drm/fbdev-ttm:
-            # Convert to client-setup") in v6.13.
+            # Added by commit d07fdf922592 ("drm/fbdev-ttm: Convert to
+            # client-setup") in v6.13 in drm/drm_client_setup.h, but then moved
+            # to drm/clients/drm_client_setup.h by commit b86711c6d6e2
+            # ("drm/client: Move public client header to clients/ subdirectory")
+            # in linux-next b86711c6d6e2.
             #
             CODE="
             #include <drm/drm_fb_helper.h>
             #if defined(NV_DRM_DRM_CLIENT_SETUP_H_PRESENT)
             #include <drm/drm_client_setup.h>
+            #elif defined(NV_DRM_CLIENTS_DRM_CLIENT_SETUP_H_PRESENT)
+            #include <drm/clients/drm_client_setup.h>
             #endif
             void conftest_drm_client_setup(void) {
                 drm_client_setup();
@@ -7509,6 +7540,31 @@ compile_test() {
             compile_check_conftest "$CODE" "NV_MODULE_IMPORT_NS_TAKES_CONSTANT" "" "generic"
         ;;
 
+
+        drm_driver_has_date)
+            #
+            # Determine if the 'drm_driver' structure has a 'date' field.
+            #
+            # Removed by commit cb2e1c2136f7 ("drm: remove driver date from
+            # struct drm_driver and all drivers") in linux-next, expected in
+            # v6.14.
+            #
+            CODE="
+            #if defined(NV_DRM_DRMP_H_PRESENT)
+            #include <drm/drmP.h>
+            #endif
+
+            #if defined(NV_DRM_DRM_DRV_H_PRESENT)
+            #include <drm/drm_drv.h>
+            #endif
+
+            int conftest_drm_driver_has_date(void) {
+                return offsetof(struct drm_driver, date);
+            }"
+
+            compile_check_conftest "$CODE" "NV_DRM_DRIVER_HAS_DATE" "" "types"
+        ;;
+
         # When adding a new conftest entry, please use the correct format for
         # specifying the relevant upstream Linux kernel commit.  Please
         # avoid specifying -rc kernels, and only use SHAs that actually exist

kernel-open/header-presence-tests.mk

@@ -31,6 +31,7 @@ NV_HEADER_PRESENCE_TESTS = \
   drm/drm_mode_config.h \
   drm/drm_modeset_lock.h \
   drm/drm_property.h \
+  drm/clients/drm_client_setup.h \
   dt-bindings/interconnect/tegra_icc_id.h \
   generated/autoconf.h \
   generated/compile.h \

kernel-open/nvidia-drm/nvidia-drm-conftest.h

@@ -65,9 +65,13 @@
 #if defined(NV_DRM_CLIENT_SETUP_PRESENT) &&                                    \
     (defined(NV_DRM_APERTURE_REMOVE_CONFLICTING_PCI_FRAMEBUFFERS_PRESENT) ||   \
      defined(NV_APERTURE_REMOVE_CONFLICTING_PCI_DEVICES_PRESENT))
+// XXX remove dependency on DRM_TTM_HELPER by implementing nvidia-drm's own
+// .fbdev_probe callback that uses NVKMS kapi
+#if IS_ENABLED(CONFIG_DRM_TTM_HELPER)
 #define NV_DRM_FBDEV_AVAILABLE
 #define NV_DRM_CLIENT_AVAILABLE
 #endif
+#endif
 
 /*
  * We can support color management if either drm_helper_crtc_enable_color_mgmt()

kernel-open/nvidia-drm/nvidia-drm-drv.c

@@ -78,6 +78,8 @@
 
 #if defined(NV_DRM_DRM_CLIENT_SETUP_H_PRESENT)
 #include <drm/drm_client_setup.h>
+#elif defined(NV_DRM_CLIENTS_DRM_CLIENT_SETUP_H_PRESENT)
+#include <drm/clients/drm_client_setup.h>
 #endif
 
 #if defined(NV_DRM_DRM_FBDEV_TTM_H_PRESENT)
@@ -1915,14 +1917,18 @@ static struct drm_driver nv_drm_driver = {
     .name                   = "nvidia-drm",
 
     .desc                   = "NVIDIA DRM driver",
+
+#if defined(NV_DRM_DRIVER_HAS_DATE)
     .date                   = "20160202",
+#endif
 
 #if defined(NV_DRM_DRIVER_HAS_DEVICE_LIST)
     .device_list            = LIST_HEAD_INIT(nv_drm_driver.device_list),
 #elif defined(NV_DRM_DRIVER_HAS_LEGACY_DEV_LIST)
     .legacy_dev_list        = LIST_HEAD_INIT(nv_drm_driver.legacy_dev_list),
 #endif
-#if defined(DRM_FBDEV_TTM_DRIVER_OPS)
+// XXX implement nvidia-drm's own .fbdev_probe callback that uses NVKMS kapi directly
+#if defined(NV_DRM_FBDEV_AVAILABLE) && defined(DRM_FBDEV_TTM_DRIVER_OPS)
     DRM_FBDEV_TTM_DRIVER_OPS,
 #endif
 };

kernel-open/nvidia-drm/nvidia-drm-sources.mk

@@ -143,4 +143,5 @@ NV_CONFTEST_TYPE_COMPILE_TESTS += drm_color_lut
 NV_CONFTEST_TYPE_COMPILE_TESTS += drm_property_blob_put
 NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_has_gem_prime_mmap
 NV_CONFTEST_TYPE_COMPILE_TESTS += drm_output_poll_changed
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_has_date
 NV_CONFTEST_TYPE_COMPILE_TESTS += file_operations_fop_unsigned_offset_present

kernel-open/nvidia-modeset/nvidia-modeset-linux.c

@@ -1050,6 +1050,11 @@ nvkms_register_backlight(NvU32 gpu_id, NvU32 display_id, void *drv_priv,
 
 #if defined(NV_ACPI_VIDEO_BACKLIGHT_USE_NATIVE)
     if (!acpi_video_backlight_use_native()) {
+#if defined(NV_ACPI_VIDEO_REGISTER_BACKLIGHT)
+        nvkms_log(NVKMS_LOG_LEVEL_INFO, NVKMS_LOG_PREFIX,
+                  "ACPI reported no NVIDIA native backlight available; attempting to use ACPI backlight.");
+        acpi_video_register_backlight();
+#endif
         return NULL;
     }
 #endif

kernel-open/nvidia-modeset/nvidia-modeset.Kbuild

@@ -102,4 +102,5 @@ NV_CONFTEST_FUNCTION_COMPILE_TESTS += list_is_first
 NV_CONFTEST_FUNCTION_COMPILE_TESTS += ktime_get_real_ts64
 NV_CONFTEST_FUNCTION_COMPILE_TESTS += ktime_get_raw_ts64
 NV_CONFTEST_FUNCTION_COMPILE_TESTS += acpi_video_backlight_use_native
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += acpi_video_register_backlight
 NV_CONFTEST_FUNCTION_COMPILE_TESTS += kernel_read_has_pointer_pos_arg

kernel-open/nvidia-uvm/uvm_ats_faults.c

@@ -29,6 +29,7 @@
 #include <linux/nodemask.h>
 #include <linux/mempolicy.h>
 #include <linux/mmu_notifier.h>
+#include <linux/topology.h>
 
 #if UVM_HMM_RANGE_FAULT_SUPPORTED()
 #include <linux/hmm.h>
@@ -291,6 +292,27 @@ static const struct mmu_interval_notifier_ops uvm_ats_notifier_ops =
 
 #endif
 
+static bool resident_policy_match(struct vm_area_struct *vma, int dst_nid, int src_nid)
+{
+#if defined(NV_MEMPOLICY_HAS_UNIFIED_NODES)
+    struct mempolicy *vma_policy = vma_policy(vma);
+
+    // TODO: Bug 4981209: When migrations between CPU numa nodes are supported,
+    // add (dst_nid != closest_cpu_numa_node) to allow migrations between CPU
+    // NUMA nodes when destination is the closest_cpu_numa_node.
+    if (vma_policy &&
+        node_isset(src_nid, vma_policy->nodes) &&
+        node_isset(dst_nid, vma_policy->nodes) &&
+        !cpumask_empty(cpumask_of_node(src_nid)) &&
+        !cpumask_empty(cpumask_of_node(dst_nid))) {
+
+        return true;
+    }
+#endif
+
+    return false;
+}
+
 static NV_STATUS ats_compute_residency_mask(uvm_gpu_va_space_t *gpu_va_space,
                                             struct vm_area_struct *vma,
                                             NvU64 base,
@@ -370,9 +392,23 @@ static NV_STATUS ats_compute_residency_mask(uvm_gpu_va_space_t *gpu_va_space,
 
             if (pfn & HMM_PFN_VALID) {
                 struct page *page = hmm_pfn_to_page(pfn);
-
-                if (page_to_nid(page) == ats_context->residency_node)
+                int resident_node = page_to_nid(page);
+
+                // Set the residency_mask if:
+                // - The page is already resident at the intended destination.
+                //   or
+                // - If both the source and destination nodes are CPU nodes and
+                //   source node is already in the list of preferred nodes for
+                //   the vma. On multi-CPU NUMA node architectures, this avoids
+                //   unnecessary migrations between CPU nodes. Since the
+                //   specific ats_context->residency_node selected by
+                //   ats_batch_select_residency() is just a guess among the list
+                //   of preferred nodes, paying the cost of migration across the
+                //   CPU preferred nodes in this case can't be justified.
+                if ((resident_node == ats_context->residency_node) ||
+                    resident_policy_match(vma, ats_context->residency_node, resident_node)) {
                     uvm_page_mask_set(residency_mask, page_index);
+                }
 
                 ats_context->prefetch_state.first_touch = false;
             }

kernel-open/nvidia-uvm/uvm_global.c

@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2015-2024 NVIDIA Corporation
+    Copyright (c) 2015-2025 NVIDIA Corporation
 
     Permission is hereby granted, free of charge, to any person obtaining a copy
     of this software and associated documentation files (the "Software"), to
@@ -35,6 +35,7 @@
 #include "uvm_mmu.h"
 #include "uvm_perf_heuristics.h"
 #include "uvm_pmm_sysmem.h"
+#include "uvm_pmm_gpu.h"
 #include "uvm_migrate.h"
 #include "uvm_gpu_access_counters.h"
 #include "uvm_va_space_mm.h"
@@ -90,6 +91,8 @@ NV_STATUS uvm_global_init(void)
     uvm_spin_lock_irqsave_init(&g_uvm_global.gpu_table_lock, UVM_LOCK_ORDER_LEAF);
     uvm_mutex_init(&g_uvm_global.va_spaces.lock, UVM_LOCK_ORDER_VA_SPACES_LIST);
     INIT_LIST_HEAD(&g_uvm_global.va_spaces.list);
+    uvm_mutex_init(&g_uvm_global.devmem_ranges.lock, UVM_LOCK_ORDER_LEAF);
+    INIT_LIST_HEAD(&g_uvm_global.devmem_ranges.list);
 
     status = uvm_kvmalloc_init();
     if (status != NV_OK) {
@@ -231,6 +234,7 @@ void uvm_global_exit(void)
     uvm_va_policy_exit();
     uvm_mem_global_exit();
     uvm_pmm_sysmem_exit();
+    uvm_pmm_devmem_exit();
     uvm_gpu_exit();
     uvm_processor_mask_cache_exit();
 

kernel-open/nvidia-uvm/uvm_global.h

@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2015-2024 NVIDIA Corporation
+    Copyright (c) 2015-2025 NVIDIA Corporation
 
     Permission is hereby granted, free of charge, to any person obtaining a copy
     of this software and associated documentation files (the "Software"), to
@@ -157,6 +157,12 @@ struct uvm_global_struct
     // This field is set once during global initialization (uvm_global_init),
     // and can be read afterwards without acquiring any locks.
     bool conf_computing_enabled;
+
+    // List of all devmem ranges allocted on this GPU
+    struct {
+        uvm_mutex_t lock;
+        struct list_head list;
+    } devmem_ranges;
 };
 
 // Initialize global uvm state