ROCT SVM地址空间管理一、什么是canonical address二、x86_64 Linux 地址空间布局三、ROCT如何管理SVM

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
//init_svm_apertures会调用的函数
init_mmap_apertures
{
	//1.初始化svm.apertures[SVM_DEFAULT]结构
	svm.apertures[SVM_DEFAULT].base  = (void *)base;
	svm.apertures[SVM_DEFAULT].limit = (void *)limit;
	svm.apertures[SVM_DEFAULT].align = align;
	svm.apertures[SVM_DEFAULT].guard_pages = guard_pages;
	svm.apertures[SVM_DEFAULT].is_cpu_accessible = true;
	svm.apertures[SVM_DEFAULT].ops = &mmap_aperture_ops;
	
	/* 
	 * 2.用mmap匿名映射的方式分配一页内存，如果分配成功这里就会直接返回(为了测试SVM区域可用)
	 * Try to allocate one page. If it fails, we'll fall back to
	 * managing our own reserved address range.
	 */
	addr = aperture_allocate_area(&svm.apertures[SVM_DEFAULT], NULL, PAGE_SIZE);
		return app->ops->allocate_area_aligned(app, address, MemorySizeInBytes, app->align);
			mmap_aperture_allocate_aligned
				//从这里可以看出，这里并没有强制让CPU addr等于SVM中的地址，原因是整个进程地址空间都是SVM的，当我们用
				//mmap匿名映射的方式分配出一段进程虚拟地址，这个进程虚拟地址肯定是落在SVM aperture中的。然后我们再将
				//这个进程虚拟地址告诉KFD，同样能达到CPUVA和GPUVA相等的效果。
				addr = mmap(0, aligned_padded_size, PROT_NONE, MAP_ANONYMOUS | MAP_NORESERVE | MAP_PRIVATE, -1, 0);
}


//fmm_init_process_apertures会调用的函数
init_svm_apertures
{
	/* If the limit is greater or equal 47-bits of address space,
	 * it means we have GFXv9 or later GPUs only. We don't need
	 * apertures to determine the MTYPE and the virtual address
	 * space of the GPUs covers the full CPU address range (on
	 * x86_64) or at least mmap is unlikely to run out of
	 * addresses the GPUs can handle.
	 * 该if条件成立，说明是进程地址空间全SVM
	 */
	if (limit >= (1ULL << 47) - 1 && !svm.reserve_svm) {
		//1.初始化svm.apertures[SVM_DEFAULT]结构
		//2.用mmap匿名映射的方式分配一页内存，如果分配成功这里就会直接返回(为了测试SVM区域可用)
		HSAKMT_STATUS status = init_mmap_apertures(base, limit, align,
							   guard_pages);

		if (status == HSAKMT_STATUS_SUCCESS)
			return status;
		/* fall through: fall back to reserved address space */
	}
	
	/* Try to reserve address space for SVM.
	 *
	 * Inner loop: try start addresses in huge-page increments up
	 * to half the VM size we're trying to reserve
	 *
	 * Outer loop: reduce size of the allocation by factor 2 at a
	 * time and print a warning for every reduction
	 * 如果非进程地址空间全SVM，那么就是在CPU进程地址空间划分一段出来作为SVM
	 */
	for (len = limit - base + 1; !found && len >= SVM_MIN_VM_SIZE; len = (len + 1) >> 1) {
		for (addr = (void *)base, ret_addr = NULL;(HSAuint64)addr + ((len + 1) >> 1) - 1 <= limit; addr = (void *)((HSAuint64)addr + ADDR_INC)) {
			//在进程地址空间预留SVM区域
			ret_addr = reserve_address(addr, map_size);
				//预留的CPU addr等于SVM aperture中的addr	
				ret_addr = mmap(addr, len, PROT_NONE, MAP_ANONYMOUS | MAP_NORESERVE | MAP_PRIVATE, -1, 0);
		}
	}

	//初始化svm.apertures[SVM_DEFAULT]结构
	svm.apertures[SVM_DEFAULT].base  = dgpu_shared_aperture_base  = ret_addr;
	svm.apertures[SVM_DEFAULT].limit = dgpu_shared_aperture_limit = (void *)limit;
	svm.apertures[SVM_DEFAULT].align = align;
	svm.apertures[SVM_DEFAULT].guard_pages = guard_pages;
	svm.apertures[SVM_DEFAULT].is_cpu_accessible = true;
	svm.apertures[SVM_DEFAULT].ops = &reserved_aperture_ops;
}


fmm_init_process_apertures
{
	//获取所有GPU node的GPUVM资源
	get_process_apertures(process_apertures, &num_of_sysfs_nodes);
	
	//遍历每一个GPU node的GPUVM资源
	for (i = 0 ; i < num_of_sysfs_nodes ; i++) {						
	
		//如果是canonical address，取所有GPUVM的交集作为SVM
		if (IS_CANONICAL_ADDR(process_apertures[i].gpuvm_limit)) {
			if (process_apertures[i].gpuvm_base > svm_base)
				svm_base = process_apertures[i].gpuvm_base;
			if (process_apertures[i].gpuvm_limit < svm_limit || svm_limit == 0)
				svm_limit = process_apertures[i].gpuvm_limit;
		} else {														
			//非canonical address暂且不管
		}
		
		//KFD保留每个GPUVM
		ret = acquire_vm(gpu_mem[gpu_mem_id].gpu_id, gpu_mem[gpu_mem_id].drm_render_fd);
	}
	
	//针对SVM空间的处理
	if (svm_limit) {
		//At least one GPU uses GPUVM in canonical address
		//space. Set up SVM apertures shared by all such GPUs
		ret = init_svm_apertures(svm_base, svm_limit, svm_alignment, guardPages);
	}
}