NVIDIA/open-gpu-kernel-modules
1
0
Fork 0
mirror of https://github.com/NVIDIA/open-gpu-kernel-modules.git synced 2026-02-09 17:50:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

560.28.03

Browse Source
This commit is contained in:
Gaurav Juvekar
2024-07-19 15:45:15 -07:00
parent 5fdf5032fb
commit 448d5cc656
859 changed files with 165424 additions and 91129 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
kernel-open/nvidia-uvm/uvm_va_space.h
View File

@@ -1,5 +1,5 @@
/*******************************************************************************
Copyright (c) 2015-2023 NVIDIA Corporation
Copyright (c) 2015-2024 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
@@ -249,12 +249,6 @@ struct uvm_va_space_struct
// for atomics in HW. This is a subset of accessible_from.
uvm_processor_mask_t has_native_atomics[UVM_ID_MAX_PROCESSORS];
// Pre-computed masks that contain, for each processor memory, a mask with
// the processors that are indirect peers. Indirect peers can access each
// other's memory like regular peers, but with additional latency and/or bw
// penalty.
uvm_processor_mask_t indirect_peers[UVM_ID_MAX_PROCESSORS];
// Mask of gpu_va_spaces registered with the va space
// indexed by gpu->id
uvm_processor_mask_t registered_gpu_va_spaces;
@@ -373,11 +367,7 @@ struct uvm_va_space_struct
// uvm_processor_mask_find_closest_id.
uvm_processor_mask_t mask;
// Temporary mask to hold direct_peers in
// uvm_processor_mask_find_closest_id.
uvm_processor_mask_t direct_peers;
// Protects the mask and direct_peers above.
// Protects the mask above.
uvm_mutex_t mask_mutex;
} closest_processors;
@@ -646,24 +636,9 @@ static uvm_gpu_va_space_state_t uvm_gpu_va_space_state(uvm_gpu_va_space_t *gpu_v
return gpu_va_space->state;
}
// Return the GPU VA space for the given physical GPU.
// Return the GPU VA space for the given GPU.
// Locking: the va_space lock must be held.
uvm_gpu_va_space_t *uvm_gpu_va_space_get_by_parent_gpu(uvm_va_space_t *va_space,
uvm_parent_gpu_t *parent_gpu);
static uvm_gpu_va_space_t *uvm_gpu_va_space_get(uvm_va_space_t *va_space, uvm_gpu_t *gpu)
{
uvm_gpu_va_space_t *gpu_va_space;
if (!gpu)
return NULL;
gpu_va_space = uvm_gpu_va_space_get_by_parent_gpu(va_space, gpu->parent);
if (gpu_va_space)
UVM_ASSERT(gpu_va_space->gpu == gpu);
return gpu_va_space;
}
uvm_gpu_va_space_t *uvm_gpu_va_space_get(uvm_va_space_t *va_space, uvm_gpu_t *gpu);
#define for_each_gpu_va_space(__gpu_va_space, __va_space) \
for (__gpu_va_space = \
@@ -758,7 +733,6 @@ static uvm_gpu_t *uvm_processor_mask_find_next_va_space_gpu(const uvm_processor_
// - src itself
// - Direct NVLINK GPU peers if src is CPU or GPU (1)
// - NVLINK CPU if src is GPU
// - Indirect NVLINK GPU peers if src is GPU
// - PCIe peers if src is GPU (2)
// - CPU if src is GPU
// - Deterministic selection from the pool of candidates