1 /*
2 * Copyright © 2012-2016 Inria. All rights reserved.
3 * See COPYING in top-level directory.
4 */
5
6 /** \file
7 * \brief Macros to help interaction between hwloc and the NVIDIA Management Library.
8 *
9 * Applications that use both hwloc and the NVIDIA Management Library may want to
10 * include this file so as to get topology information for NVML devices.
11 */
12
13 #ifndef HWLOC_NVML_H
14 #define HWLOC_NVML_H
15
16 #include <hwloc.h>
17 #include <hwloc/autogen/config.h>
18 #include <hwloc/helper.h>
19 #ifdef HWLOC_LINUX_SYS
20 #include <hwloc/linux.h>
21 #endif
22
23 #include <nvml.h>
24
25
26 #ifdef __cplusplus
27 extern "C" {
28 #endif
29
30
31 /** \defgroup hwlocality_nvml Interoperability with the NVIDIA Management Library
32 *
33 * This interface offers ways to retrieve topology information about
34 * devices managed by the NVIDIA Management Library (NVML).
35 *
36 * @{
37 */
38
39 /** \brief Get the CPU set of logical processors that are physically
40 * close to NVML device \p device.
41 *
42 * Return the CPU set describing the locality of the NVML device \p device.
43 *
44 * Topology \p topology and device \p device must match the local machine.
45 * I/O devices detection and the NVML component are not needed in the topology.
46 *
47 * The function only returns the locality of the device.
48 * If more information about the device is needed, OS objects should
49 * be used instead, see hwloc_nvml_get_device_osdev()
50 * and hwloc_nvml_get_device_osdev_by_index().
51 *
52 * This function is currently only implemented in a meaningful way for
53 * Linux; other systems will simply get a full cpuset.
54 */
55 static __hwloc_inline int
56 hwloc_nvml_get_device_cpuset(hwloc_topology_t topology __hwloc_attribute_unused,
57 nvmlDevice_t device, hwloc_cpuset_t set)
58 {
59 #ifdef HWLOC_LINUX_SYS
60 /* If we're on Linux, use the sysfs mechanism to get the local cpus */
61 #define HWLOC_NVML_DEVICE_SYSFS_PATH_MAX 128
62 char path[HWLOC_NVML_DEVICE_SYSFS_PATH_MAX];
63 nvmlReturn_t nvres;
64 nvmlPciInfo_t pci;
65
66 if (!hwloc_topology_is_thissystem(topology)) {
67 errno = EINVAL;
68 return -1;
69 }
70
71 nvres = nvmlDeviceGetPciInfo(device, &pci);
72 if (NVML_SUCCESS != nvres) {
73 errno = EINVAL;
74 return -1;
75 }
76
77 sprintf(path, "/sys/bus/pci/devices/%04x:%02x:%02x.0/local_cpus", pci.domain, pci.bus, pci.device);
78 if (hwloc_linux_read_path_as_cpumask(path, set) < 0
79 || hwloc_bitmap_iszero(set))
80 hwloc_bitmap_copy(set, hwloc_topology_get_complete_cpuset(topology));
81 #else
82 /* Non-Linux systems simply get a full cpuset */
83 hwloc_bitmap_copy(set, hwloc_topology_get_complete_cpuset(topology));
84 #endif
85 return 0;
86 }
87
88 /** \brief Get the hwloc OS device object corresponding to the
89 * NVML device whose index is \p idx.
90 *
91 * Return the OS device object describing the NVML device whose
92 * index is \p idx. Returns NULL if there is none.
93 *
94 * The topology \p topology does not necessarily have to match the current
95 * machine. For instance the topology may be an XML import of a remote host.
96 * I/O devices detection and the NVML component must be enabled in the topology.
97 *
98 * \note The corresponding PCI device object can be obtained by looking
99 * at the OS device parent object (unless PCI devices are filtered out).
100 */
101 static __hwloc_inline hwloc_obj_t
102 hwloc_nvml_get_device_osdev_by_index(hwloc_topology_t topology, unsigned idx)
103 {
104 hwloc_obj_t osdev = NULL;
105 while ((osdev = hwloc_get_next_osdev(topology, osdev)) != NULL) {
106 if (HWLOC_OBJ_OSDEV_GPU == osdev->attr->osdev.type
107 && osdev->name
108 && !strncmp("nvml", osdev->name, 4)
109 && atoi(osdev->name + 4) == (int) idx)
110 return osdev;
111 }
112 return NULL;
113 }
114
115 /** \brief Get the hwloc OS device object corresponding to NVML device \p device.
116 *
117 * Return the hwloc OS device object that describes the given
118 * NVML device \p device. Return NULL if there is none.
119 *
120 * Topology \p topology and device \p device must match the local machine.
121 * I/O devices detection and the NVML component must be enabled in the topology.
122 * If not, the locality of the object may still be found using
123 * hwloc_nvml_get_device_cpuset().
124 *
125 * \note The corresponding hwloc PCI device may be found by looking
126 * at the result parent pointer (unless PCI devices are filtered out).
127 */
128 static __hwloc_inline hwloc_obj_t
129 hwloc_nvml_get_device_osdev(hwloc_topology_t topology, nvmlDevice_t device)
130 {
131 hwloc_obj_t osdev;
132 nvmlReturn_t nvres;
133 nvmlPciInfo_t pci;
134 char uuid[64];
135
136 if (!hwloc_topology_is_thissystem(topology)) {
137 errno = EINVAL;
138 return NULL;
139 }
140
141 nvres = nvmlDeviceGetPciInfo(device, &pci);
142 if (NVML_SUCCESS != nvres)
143 return NULL;
144
145 nvres = nvmlDeviceGetUUID(device, uuid, sizeof(uuid));
146 if (NVML_SUCCESS != nvres)
147 uuid[0] = '\0';
148
149 osdev = NULL;
150 while ((osdev = hwloc_get_next_osdev(topology, osdev)) != NULL) {
151 hwloc_obj_t pcidev = osdev->parent;
152 const char *info;
153
154 if (strncmp(osdev->name, "nvml", 4))
155 continue;
156
157 if (pcidev
158 && pcidev->type == HWLOC_OBJ_PCI_DEVICE
159 && pcidev->attr->pcidev.domain == pci.domain
160 && pcidev->attr->pcidev.bus == pci.bus
161 && pcidev->attr->pcidev.dev == pci.device
162 && pcidev->attr->pcidev.func == 0)
163 return osdev;
164
165 info = hwloc_obj_get_info_by_name(osdev, "NVIDIAUUID");
166 if (info && !strcmp(info, uuid))
167 return osdev;
168 }
169
170 return NULL;
171 }
172
173 /** @} */
174
175
176 #ifdef __cplusplus
177 } /* extern "C" */
178 #endif
179
180
181 #endif /* HWLOC_NVML_H */