580.159.03

Author: mmaneetsingh

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 580.142.
+version 580.159.03.
 
 
 ## 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
-580.142 driver release.  This can be achieved by installing
+580.159.03 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_64/580.142/README/kernel_open.html
+https://us.download.nvidia.com/XFree86/Linux-x86_64/580.159.03/README/kernel_open.html
 
 For vGPU support, please refer to the README.vgpu packaged in the vGPU Host
 Package for more details.
@@ -974,7 +974,6 @@ 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 GeForce RTX 5080                                 | 2C02           |
 | NVIDIA GeForce RTX 5070 Ti                              | 2C05           |
 | NVIDIA GeForce RTX 5090 Laptop GPU                      | 2C18           |
@@ -993,7 +992,6 @@ 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           |

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=\"580.142\"
+ccflags-y += -DNV_VERSION_STRING=\"580.159.03\"
 
 # Include and link Tegra out-of-tree modules.
 ifneq ($(wildcard /usr/src/nvidia/nvidia-oot),)

kernel-open/Makefile

@@ -71,31 +71,6 @@ else
   CC ?= cc
   LD ?= ld
   OBJDUMP ?= objdump
-  AWK ?= awk
-  # Bake the following awk program in a string. The program is needed to add C++
-  # to the languages excluded from BTF generation.
-  #
-  # Also, unconditionally return success (0) from the awk program, rather than
-  # propagating pahole's return status (with 'exit system(pahole_cmd)'), to
-  # workaround an DW_TAG_rvalue_reference_type error in
-  # kernel/nvidia-modeset.ko.
-  #
-  # BEGIN {
-  #     pahole_cmd = "pahole"
-  #     for (i = 1; i < ARGC; i++) {
-  #         if (ARGV[i] ~ /--lang_exclude=/) {
-  #             pahole_cmd = pahole_cmd sprintf(" %s,c++", ARGV[i])
-  #         } else {
-  #             pahole_cmd = pahole_cmd sprintf(" %s", ARGV[i])
-  #         }
-  #     }
-  #     system(pahole_cmd)
-  # }
-  PAHOLE_AWK_PROGRAM = BEGIN { pahole_cmd = \"pahole\"; for (i = 1; i < ARGC; i++) { if (ARGV[i] ~ /--lang_exclude=/) { pahole_cmd = pahole_cmd sprintf(\" %s,c++\", ARGV[i]); } else { pahole_cmd = pahole_cmd sprintf(\" %s\", ARGV[i]); } } system(pahole_cmd); }
-  # If scripts/pahole-flags.sh is not present in the kernel tree, add PAHOLE and
-  # PAHOLE_AWK_PROGRAM assignments to PAHOLE_VARIABLES; otherwise assign the
-  # empty string to PAHOLE_VARIABLES.
-  PAHOLE_VARIABLES=$(if $(wildcard $(KERNEL_SOURCES)/scripts/pahole-flags.sh),,"PAHOLE=$(AWK) '$(PAHOLE_AWK_PROGRAM)'")
 
   ifndef ARCH
     ARCH := $(shell uname -m | sed -e 's/i.86/i386/' \
@@ -133,10 +108,12 @@ else
   KBUILD_PARAMS += INSTALL_MOD_DIR="$(INSTALL_MOD_DIR)"
   KBUILD_PARAMS += NV_SPECTRE_V2=$(SPECTRE_V2_RETPOLINE)
 
+  export ORIG_PAHOLE := $(PAHOLE)
+
   .PHONY: modules module clean clean_conftest modules_install
   modules clean modules_install:
 	@$(MAKE) "LD=$(LD)" "CC=$(CC)" "OBJDUMP=$(OBJDUMP)" \
-	  $(PAHOLE_VARIABLES) $(KBUILD_PARAMS) $@
+	  PAHOLE=$(CURDIR)/pahole.sh $(KBUILD_PARAMS) $@
 	@if [ "$@" = "modules" ]; then \
 	  for module in $(NV_KERNEL_MODULES); do \
 	    if [ -x split-object-file.sh ]; then \

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

@@ -64,3 +64,4 @@ module_exit(nv_linux_drm_exit);
 
 MODULE_INFO(supported, "external");
 MODULE_VERSION(NV_VERSION_STRING);
+MODULE_DESCRIPTION("NVIDIA DRM kernel module");

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

@@ -2221,3 +2221,4 @@ module_exit(nvkms_exit);
 
 MODULE_INFO(supported, "external");
 MODULE_VERSION(NV_VERSION_STRING);
+MODULE_DESCRIPTION("NVIDIA modeset kernel module");

kernel-open/nvidia-uvm/uvm.c

@@ -1232,3 +1232,4 @@ module_exit(uvm_exit_entry);
 MODULE_LICENSE("Dual MIT/GPL");
 MODULE_INFO(supported, "external");
 MODULE_VERSION(NV_VERSION_STRING);
+MODULE_DESCRIPTION("NVIDIA Unified Virtual Memory kernel module");

kernel-open/nvidia-uvm/uvm_hmm.c

@@ -1583,6 +1583,31 @@ uvm_va_block_region_t uvm_hmm_get_prefetch_region(uvm_va_block_t *va_block,
     return uvm_va_block_region_from_start_end(va_block, start, end);
 }
 
+uvm_prot_t uvm_hmm_compute_mapping_prot(uvm_va_block_t *va_block,
+                                        uvm_processor_id_t processor_id,
+                                        uvm_page_index_t page_index)
+{
+    if (!uvm_processor_mask_test(&va_block->mapped, UVM_ID_CPU))
+        return UVM_PROT_NONE;
+
+    if (uvm_page_mask_test(&va_block->cpu.pte_bits[UVM_PTE_BITS_CPU_WRITE], page_index)) {
+        if (uvm_processor_mask_test(&va_block->hmm.va_space->has_native_atomics[uvm_id_value(UVM_ID_CPU)],
+                                    processor_id))
+            // If the CPU has write access it also has atomic access, so it's
+            // fine for any GPU with HW support to do atomic accesses.
+            return UVM_PROT_READ_WRITE_ATOMIC;
+        else
+            // Otherwise the GPU needs to fault on atomic access to ensure the
+            // CPU is unmapped.
+            return UVM_PROT_READ_WRITE;
+    }
+
+    if (uvm_page_mask_test(&va_block->cpu.pte_bits[UVM_PTE_BITS_CPU_READ], page_index))
+        return UVM_PROT_READ_ONLY;
+
+    return UVM_PROT_NONE;
+}
+
 uvm_prot_t uvm_hmm_compute_logical_prot(uvm_va_block_t *va_block,
                                         struct vm_area_struct *vma,
                                         NvU64 addr)

kernel-open/nvidia-uvm/uvm_hmm.h

@@ -258,6 +258,14 @@ typedef struct
                                                       const uvm_va_policy_t *policy,
                                                       NvU64 address);
 
+    // Return the actual permissions allowed when mapping a page within a
+    // va_block on the given processor_id. This may differ from the logical
+    // permission if for example the kernel has the CPU pages mapped read-only
+    // to do copy-on-write.
+    uvm_prot_t uvm_hmm_compute_mapping_prot(uvm_va_block_t *va_block,
+                                            uvm_processor_id_t processor_id,
+                                            uvm_page_index_t page_index);
+
     // Return the logical protection allowed of a HMM va_block for the page at
     // the given address within the vma which must be valid. This is usually
     // obtained from uvm_hmm_va_block_find_create()).
@@ -561,6 +569,13 @@ typedef struct
         return (uvm_va_block_region_t){};
     }
 
+    static uvm_prot_t uvm_hmm_compute_mapping_prot(uvm_va_block_t *va_block,
+                                                   uvm_processor_id_t processor_id,
+                                                   uvm_page_index_t page_index)
+    {
+        return UVM_PROT_NONE;
+    }
+
     static uvm_prot_t uvm_hmm_compute_logical_prot(uvm_va_block_t *va_block,
                                                    struct vm_area_struct *vma,
                                                    NvU64 addr)

kernel-open/nvidia-uvm/uvm_va_block.c

@@ -11374,6 +11374,13 @@ uvm_prot_t uvm_va_block_page_compute_highest_permission(uvm_va_block_t *va_block
     uvm_processor_mask_t *resident_processors = &va_block_context->scratch_processor_mask;
     NvU32 resident_processors_count;
 
+    // TODO: Bug 5841902
+    // There are several calls to uvm_va_block_is_hmm() which need to be removed
+    if (uvm_va_block_is_hmm(va_block))
+        return uvm_hmm_compute_mapping_prot(va_block,
+                                            processor_id,
+                                            page_index);
+
     if (uvm_processor_mask_test(block_get_uvm_lite_gpus(va_block), processor_id))
         return UVM_PROT_READ_WRITE_ATOMIC;
 

kernel-open/nvidia-uvm/uvm_va_range.c

@@ -947,8 +947,12 @@ void uvm_va_range_remove_gpu_va_space(uvm_va_range_t *va_range,
                                                         gpu_va_space->gpu);
             break;
         case UVM_VA_RANGE_TYPE_DEVICE_P2P:
-            unmap_mapping_range(va_range->va_space->mapping, va_range->node.start, uvm_va_range_size(va_range), 1);
-            uvm_va_range_deinit_device_p2p(uvm_va_range_to_device_p2p(va_range), deferred_free_list);
+            // Device P2P ranges are associated with a specific GPU so destroy
+            // the range entirely if unregistering the associated GPU.
+            if (uvm_va_range_to_device_p2p(va_range)->gpu == gpu_va_space->gpu) {
+                unmap_mapping_range(va_range->va_space->mapping, va_range->node.start, uvm_va_range_size(va_range), 1);
+                uvm_va_range_deinit_device_p2p(uvm_va_range_to_device_p2p(va_range), deferred_free_list);
+            }
             break;
         default:
             UVM_ASSERT_MSG(0, "[0x%llx, 0x%llx] has type %d\n",
@@ -1223,9 +1227,10 @@ void uvm_va_range_unregister_gpu(uvm_va_range_t *va_range,
             va_range_unregister_gpu_semaphore_pool(uvm_va_range_to_semaphore_pool(va_range), gpu);
             break;
         case UVM_VA_RANGE_TYPE_DEVICE_P2P:
-            // All ranges should have been deinited by GPU VA space unregister,
-            // which should have already happened.
-            UVM_ASSERT(!uvm_va_range_to_device_p2p(va_range)->p2p_mem);
+            // All ranges for this GPU should have been deinited by GPU VA space
+            // unregister, which should have already happened.
+            if (uvm_va_range_to_device_p2p(va_range)->p2p_mem != NULL)
+                UVM_ASSERT(uvm_va_range_to_device_p2p(va_range)->gpu != gpu);
             break;
         default:
             UVM_ASSERT_MSG(0, "[0x%llx, 0x%llx] has type %d\n",