595.58.03

Author: aritger

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 595.45.04.
+version 595.58.03.
 
 
 ## 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
-595.45.04 driver release.  This can be achieved by installing
+595.58.03 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_64/595.45.04/README/kernel_open.html
+https://us.download.nvidia.com/XFree86/Linux-x86_64/595.58.03/README/kernel_open.html
 
 For vGPU support, please refer to the README.vgpu packaged in the vGPU Host
 Package for more details.
@@ -946,6 +946,7 @@ Subsystem Device ID.
 | NVIDIA B200                                             | 2901 10DE 1999 |
 | NVIDIA B200                                             | 2901 10DE 199B |
 | NVIDIA B200                                             | 2901 10DE 20DA |
+| NVIDIA B200                                             | 2909 10DE 22EB |
 | NVIDIA GB200                                            | 2941 10DE 2046 |
 | NVIDIA GB200                                            | 2941 10DE 20CA |
 | NVIDIA GB200                                            | 2941 10DE 20D5 |
@@ -974,6 +975,8 @@ Subsystem Device ID.
 | NVIDIA RTX PRO 6000 Blackwell Server Edition            | 2BB5 10DE 204E |
 | NVIDIA RTX PRO 6000 Blackwell Server Edition            | 2BB5 10DE 220B |
 | NVIDIA RTX 6000D                                        | 2BB9 10DE 2091 |
+| NVIDIA RTX 6000D                                        | 2BB9 10DE 2092 |
+| NVIDIA RTX 6000D                                        | 2BB9 10DE 2279 |
 | NVIDIA GeForce RTX 5080                                 | 2C02           |
 | NVIDIA GeForce RTX 5070 Ti                              | 2C05           |
 | NVIDIA GeForce RTX 5090 Laptop GPU                      | 2C18           |
@@ -992,6 +995,7 @@ Subsystem Device ID.
 | NVIDIA RTX PRO 4000 Blackwell                           | 2C34 17AA 2052 |
 | NVIDIA RTX PRO 5000 Blackwell Generation Laptop GPU     | 2C38           |
 | NVIDIA RTX PRO 4000 Blackwell Generation Laptop GPU     | 2C39           |
+| NVIDIA RTX PRO 4500 Blackwell Server Edition            | 2C3A 10DE 21F4 |
 | NVIDIA GeForce RTX 5090 Laptop GPU                      | 2C58           |
 | NVIDIA GeForce RTX 5080 Laptop GPU                      | 2C59           |
 | NVIDIA RTX PRO 5000 Blackwell Embedded GPU              | 2C77           |
@@ -1021,3 +1025,4 @@ Subsystem Device ID.
 | NVIDIA GeForce RTX 5070 Ti Laptop GPU                   | 2F58           |
 | NVIDIA B300 SXM6 AC                                     | 3182 10DE 20E6 |
 | NVIDIA GB300                                            | 31C2 10DE 21F1 |
+| NVIDIA GB300                                            | 31C3 10DE 22F8 |

kernel-open/Kbuild

@@ -79,7 +79,7 @@ ccflags-y += -I$(src)/common/inc
 ccflags-y += -I$(src)
 ccflags-y += -Wall $(DEFINES) $(INCLUDES) -Wno-cast-qual -Wno-format-extra-args
 ccflags-y += -D__KERNEL__ -DMODULE -DNVRM
-ccflags-y += -DNV_VERSION_STRING=\"595.45.04\"
+ccflags-y += -DNV_VERSION_STRING=\"595.58.03\"
 
 # Include and link Tegra out-of-tree modules.
 ifneq ($(wildcard /usr/src/nvidia/nvidia-public),)

kernel-open/common/inc/nv-linux.h

@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2001-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 2001-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  * SPDX-License-Identifier: MIT
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
@@ -1351,7 +1351,6 @@ typedef struct nv_linux_state_s {
     struct reset_control *nvdisplay_reset;
     struct reset_control *dsi_core_reset;
     struct reset_control *mipi_cal_reset;
-    struct reset_control *hdacodec_reset;
 
     /*
      * nv_imp_icc_path represents the interconnect path across which display

kernel-open/common/inc/nv-mm.h

@@ -216,6 +216,16 @@ static inline void nv_vma_flags_clear_word(struct vm_area_struct *vma, unsigned
     ACCESS_PRIVATE(vma, __vm_flags) &= ~flags;
 #endif
 }
+
+static inline int nv_is_vma_write_locked(struct vm_area_struct *vma, unsigned int *mm_lock_seq)
+{
+#if defined(NV_IS_VMA_WRITE_LOCKED_HAS_MM_LOCK_SEQ_ARG)
+    return __is_vma_write_locked(vma, mm_lock_seq);
+#else
+    *mm_lock_seq = __vma_raw_mm_seqnum(vma);
+    return __is_vma_write_locked(vma);
+#endif
+}
 #endif // !NV_CAN_CALL_VMA_START_WRITE
 
 static inline void nv_vm_flags_set(struct vm_area_struct *vma, vm_flags_t flags)

kernel-open/common/inc/nvkms-kapi.h

@@ -604,7 +604,17 @@ struct NvKmsKapiGpuInfo {
     MIGDeviceId   migDevice;
 };
 
+/*
+ * Linux kernel options CONFIG_RANDSTRUCT_* randomize structs that are composed
+ * entirely of function pointers, but can only control struct layout for sources
+ * built by kbuild. NvKmsKapiCallbacks is shared between kbuild-built
+ * nvidia-drm.ko, and the "OS-agnostic" portions of nvidia-modeset.ko (not built
+ * by kbuild). Add a _padding member to disable struct randomization.
+ *
+ * Refer to https://github.com/NVIDIA/open-gpu-kernel-modules/issues/1033
+ */
 struct NvKmsKapiCallbacks {
+    int  _padding;
     void (*suspendResume)(NvBool suspend);
     void (*remove)(NvU32 gpuId);
     void (*probe)(const struct NvKmsKapiGpuInfo *gpu_info);

kernel-open/common/inc/nvtypes.h

@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 1993-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 1993-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  * SPDX-License-Identifier: MIT
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
@@ -514,6 +514,16 @@ typedef struct
 #define NV_DECLARE_ALIGNED(TYPE_VAR, ALIGN) __declspec(align(ALIGN)) TYPE_VAR
 #elif defined(__arm)
 #define NV_DECLARE_ALIGNED(TYPE_VAR, ALIGN) __align(ALIGN) TYPE_VAR
+#endif
+
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)
+#define NV_ALIGNOF(type) _Alignof(type)
+#elif defined(__cplusplus) && (__cplusplus >= 201103L)
+#define NV_ALIGNOF(type) alignof(type)
+#elif defined(__GNUC__) || defined(__clang__)
+#define NV_ALIGNOF(type) __alignof__(type)
+#elif defined(_MSC_VER)
+#define NV_ALIGNOF(type) __alignof(type)
 #endif
 
  /***************************************************************************\

kernel-open/conftest.sh

@@ -5117,6 +5117,26 @@ compile_test() {
             compile_check_conftest "$CODE" "NV_DRM_CRTC_FUNCS_HAS_GET_VBLANK_TIMESTAMP" "" "types"
         ;;
 
+        is_vma_write_locked_has_mm_lock_seq_arg)
+            #
+            # Determine if __is_vma_write_locked() takes only a single
+            # 'struct vm_area_struct *' argument.
+            #
+            # Commit 22f7639f2f03 ("mm/vma: improve and document
+            # __is_vma_write_locked()") removed the 'unsigned int *mm_lock_seq'
+            # parameter in v7.0-rc1.
+            #
+            CODE="
+            #include <linux/mm.h>
+            #include <linux/mmap_lock.h>
+            int conftest_is_vma_write_locked_has_mm_lock_seq_arg(struct vm_area_struct *vma) {
+                unsigned int mm_lock_seq;
+                return __is_vma_write_locked(vma, &mm_lock_seq);
+            }"
+
+            compile_check_conftest "$CODE" "NV_IS_VMA_WRITE_LOCKED_HAS_MM_LOCK_SEQ_ARG" "" "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/nvidia-drm/nvidia-drm-gem-nvkms-memory.c

@@ -585,6 +585,14 @@ int nv_drm_gem_alloc_nvkms_memory_ioctl(struct drm_device *dev,
     allocParams.compressible = &p->compressible;
 
     pMemory = nvKms->allocateMemory(nv_dev->pDevice, &allocParams);
+    if (pMemory == NULL && (p->flags & NV_GEM_ALLOC_NO_SCANOUT)) {
+        NV_DRM_DEV_LOG_INFO(
+                nv_dev,
+                "Failed to allocate NVKMS video memory for GEM object, trying to fall back to sysmem");
+        allocParams.useVideoMemory = false;
+        pMemory = nvKms->allocateMemory(nv_dev->pDevice, &allocParams);
+    }
+
     if (pMemory == NULL) {
         ret = -EINVAL;
         NV_DRM_DEV_LOG_ERR(nv_dev,

kernel-open/nvidia-uvm/uvm_gpu.c

@@ -2921,7 +2921,7 @@ static NV_STATUS add_gpu(const NvProcessorUuid *gpu_uuid,
 #if UVM_IS_CONFIG_HMM()
     // HMM was disabled when first initialising the parent so we can't support
     // it now. Tell the caller to retry with it disabled.
-    else if (!parent_gpu->devmem && enable_hmm) {
+    else if (!parent_gpu->devmem && enable_hmm && uvm_hmm_is_enabled_system_wide()) {
         status = NV_ERR_BUSY_RETRY;
         goto error;
     }
@@ -3222,10 +3222,25 @@ uvm_gpu_phys_address_t uvm_gpu_peer_phys_address(uvm_gpu_t *owning_gpu, NvU64 ad
 uvm_gpu_address_t uvm_gpu_peer_copy_address(uvm_gpu_t *owning_gpu, NvU64 address, uvm_gpu_t *accessing_gpu)
 {
     uvm_gpu_identity_mapping_t *gpu_peer_mapping;
+    const bool mig_peers_use_phys = uvm_gpus_are_smc_peers(owning_gpu, accessing_gpu) &&
+        accessing_gpu->parent->ce_phys_vidmem_write_supported;
 
-    if (accessing_gpu->parent->peer_copy_mode == UVM_GPU_PEER_COPY_MODE_PHYSICAL)
+    // MIG peers do not create peer vidmem mappings like other peers. They do
+    // create their vidmem identity mappings to cover all possible physical
+    // addresses, even those of other MIG peers.
+    // Use vidmem this identity mapping if CEs need to use virtual addresses.
+    if (uvm_gpus_are_smc_peers(owning_gpu, accessing_gpu) && !mig_peers_use_phys) {
+        uvm_gpu_phys_address_t phys_address = uvm_gpu_peer_phys_address(owning_gpu, address, accessing_gpu);
+        return uvm_gpu_address_virtual_from_vidmem_phys(accessing_gpu, phys_address.address);
+    }
+
+    // Use physical addresses for MIGs peers if CE allows it. Irespective of
+    // the peer copy mode.
+    if (accessing_gpu->parent->peer_copy_mode == UVM_GPU_PEER_COPY_MODE_PHYSICAL || mig_peers_use_phys)
         return uvm_gpu_address_from_phys(uvm_gpu_peer_phys_address(owning_gpu, address, accessing_gpu));
 
+    // MIG peers do not create peer GPU mappings so it should never reach here.
+    UVM_ASSERT(!uvm_gpus_are_smc_peers(owning_gpu, accessing_gpu));
     UVM_ASSERT(accessing_gpu->parent->peer_copy_mode == UVM_GPU_PEER_COPY_MODE_VIRTUAL);
     gpu_peer_mapping = uvm_gpu_get_peer_mapping(accessing_gpu, owning_gpu->id);
 

kernel-open/nvidia/nv-mmap.c

@@ -957,15 +957,22 @@ void NV_API_CALL nv_set_safe_to_mmap_locked(
 }
 
 #if !NV_CAN_CALL_VMA_START_WRITE
+
+#if defined(VM_REFCNT_EXCLUDE_READERS_FLAG)
+#define NV_VMA_LOCK_OFFSET VM_REFCNT_EXCLUDE_READERS_FLAG
+#else
+#define NV_VMA_LOCK_OFFSET VMA_LOCK_OFFSET
+#endif
+
 static NvBool nv_vma_enter_locked(struct vm_area_struct *vma, NvBool detaching)
 {
-    NvU32 tgt_refcnt = VMA_LOCK_OFFSET;
+    NvU32 tgt_refcnt = NV_VMA_LOCK_OFFSET;
     NvBool interrupted = NV_FALSE;
     if (!detaching)
     {
         tgt_refcnt++;
     }
-    if (!refcount_add_not_zero(VMA_LOCK_OFFSET, &vma->vm_refcnt))
+    if (!refcount_add_not_zero(NV_VMA_LOCK_OFFSET, &vma->vm_refcnt))
     {
         return NV_FALSE;
     }
@@ -995,7 +1002,7 @@ static NvBool nv_vma_enter_locked(struct vm_area_struct *vma, NvBool detaching)
     if (interrupted)
     {
         // Clean up on error: release refcount and dep_map
-        refcount_sub_and_test(VMA_LOCK_OFFSET, &vma->vm_refcnt);
+        refcount_sub_and_test(NV_VMA_LOCK_OFFSET, &vma->vm_refcnt);
         rwsem_release(&vma->vmlock_dep_map, _RET_IP_);
         return NV_FALSE;
     }
@@ -1011,7 +1018,7 @@ void nv_vma_start_write(struct vm_area_struct *vma)
 {
     NvU32 mm_lock_seq;
     NvBool locked;
-    if (__is_vma_write_locked(vma, &mm_lock_seq))
+    if (nv_is_vma_write_locked(vma, &mm_lock_seq))
         return;
 
     locked = nv_vma_enter_locked(vma, NV_FALSE);
@@ -1020,7 +1027,7 @@ void nv_vma_start_write(struct vm_area_struct *vma)
     if (locked)
     {
         NvBool detached;
-        detached = refcount_sub_and_test(VMA_LOCK_OFFSET, &vma->vm_refcnt);
+        detached = refcount_sub_and_test(NV_VMA_LOCK_OFFSET, &vma->vm_refcnt);
         rwsem_release(&vma->vmlock_dep_map, _RET_IP_);
         WARN_ON_ONCE(detached);
     }