amdgpu: use the high VA range if possible
[platform/upstream/libdrm.git] / amdgpu / amdgpu_device.c
1 /*
2  * Copyright 2014 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  */
23
24 /**
25  * \file amdgpu_device.c
26  *
27  *  Implementation of functions for AMD GPU device
28  *
29  */
30
31 #ifdef HAVE_CONFIG_H
32 #include "config.h"
33 #endif
34
35 #include <sys/stat.h>
36 #include <errno.h>
37 #include <string.h>
38 #include <stdio.h>
39 #include <stdlib.h>
40 #include <unistd.h>
41
42 #include "xf86drm.h"
43 #include "amdgpu_drm.h"
44 #include "amdgpu_internal.h"
45 #include "util_hash_table.h"
46 #include "util_math.h"
47
48 #define PTR_TO_UINT(x) ((unsigned)((intptr_t)(x)))
49 #define UINT_TO_PTR(x) ((void *)((intptr_t)(x)))
50
51 static pthread_mutex_t fd_mutex = PTHREAD_MUTEX_INITIALIZER;
52 static struct util_hash_table *fd_tab;
53
54 static unsigned handle_hash(void *key)
55 {
56         return PTR_TO_UINT(key);
57 }
58
59 static int handle_compare(void *key1, void *key2)
60 {
61         return PTR_TO_UINT(key1) != PTR_TO_UINT(key2);
62 }
63
64 static unsigned fd_hash(void *key)
65 {
66         int fd = PTR_TO_UINT(key);
67         char *name = drmGetPrimaryDeviceNameFromFd(fd);
68         unsigned result = 0;
69         char *c;
70
71         if (name == NULL)
72                 return 0;
73
74         for (c = name; *c; ++c)
75                 result += *c;
76
77         free(name);
78
79         return result;
80 }
81
82 static int fd_compare(void *key1, void *key2)
83 {
84         int fd1 = PTR_TO_UINT(key1);
85         int fd2 = PTR_TO_UINT(key2);
86         char *name1 = drmGetPrimaryDeviceNameFromFd(fd1);
87         char *name2 = drmGetPrimaryDeviceNameFromFd(fd2);
88         int result;
89
90         if (name1 == NULL || name2 == NULL) {
91                 free(name1);
92                 free(name2);
93                 return 0;
94         }
95
96         result = strcmp(name1, name2);
97         free(name1);
98         free(name2);
99
100         return result;
101 }
102
103 /**
104 * Get the authenticated form fd,
105 *
106 * \param   fd   - \c [in]  File descriptor for AMD GPU device
107 * \param   auth - \c [out] Pointer to output the fd is authenticated or not
108 *                          A render node fd, output auth = 0
109 *                          A legacy fd, get the authenticated for compatibility root
110 *
111 * \return   0 on success\n
112 *          >0 - AMD specific error code\n
113 *          <0 - Negative POSIX Error code
114 */
115 static int amdgpu_get_auth(int fd, int *auth)
116 {
117         int r = 0;
118         drm_client_t client = {};
119
120         if (drmGetNodeTypeFromFd(fd) == DRM_NODE_RENDER)
121                 *auth = 0;
122         else {
123                 client.idx = 0;
124                 r = drmIoctl(fd, DRM_IOCTL_GET_CLIENT, &client);
125                 if (!r)
126                         *auth = client.auth;
127         }
128         return r;
129 }
130
131 static void amdgpu_device_free_internal(amdgpu_device_handle dev)
132 {
133         const struct amdgpu_asic_id *id;
134         amdgpu_vamgr_deinit(&dev->vamgr_32);
135         amdgpu_vamgr_deinit(&dev->vamgr);
136         util_hash_table_destroy(dev->bo_flink_names);
137         util_hash_table_destroy(dev->bo_handles);
138         pthread_mutex_destroy(&dev->bo_table_mutex);
139         util_hash_table_remove(fd_tab, UINT_TO_PTR(dev->fd));
140         close(dev->fd);
141         if ((dev->flink_fd >= 0) && (dev->fd != dev->flink_fd))
142                 close(dev->flink_fd);
143         if (dev->asic_ids) {
144                 for (id = dev->asic_ids; id->did; id++)
145                         free(id->marketing_name);
146
147                 free(dev->asic_ids);
148         }
149         free(dev);
150 }
151
152 /**
153  * Assignment between two amdgpu_device pointers with reference counting.
154  *
155  * Usage:
156  *    struct amdgpu_device *dst = ... , *src = ...;
157  *
158  *    dst = src;
159  *    // No reference counting. Only use this when you need to move
160  *    // a reference from one pointer to another.
161  *
162  *    amdgpu_device_reference(&dst, src);
163  *    // Reference counters are updated. dst is decremented and src is
164  *    // incremented. dst is freed if its reference counter is 0.
165  */
166 static void amdgpu_device_reference(struct amdgpu_device **dst,
167                              struct amdgpu_device *src)
168 {
169         if (update_references(&(*dst)->refcount, &src->refcount))
170                 amdgpu_device_free_internal(*dst);
171         *dst = src;
172 }
173
174 int amdgpu_device_initialize(int fd,
175                              uint32_t *major_version,
176                              uint32_t *minor_version,
177                              amdgpu_device_handle *device_handle)
178 {
179         struct amdgpu_device *dev;
180         drmVersionPtr version;
181         int r;
182         int flag_auth = 0;
183         int flag_authexist=0;
184         uint32_t accel_working = 0;
185         uint64_t start, max;
186
187         *device_handle = NULL;
188
189         pthread_mutex_lock(&fd_mutex);
190         if (!fd_tab)
191                 fd_tab = util_hash_table_create(fd_hash, fd_compare);
192         r = amdgpu_get_auth(fd, &flag_auth);
193         if (r) {
194                 fprintf(stderr, "%s: amdgpu_get_auth (1) failed (%i)\n",
195                         __func__, r);
196                 pthread_mutex_unlock(&fd_mutex);
197                 return r;
198         }
199         dev = util_hash_table_get(fd_tab, UINT_TO_PTR(fd));
200         if (dev) {
201                 r = amdgpu_get_auth(dev->fd, &flag_authexist);
202                 if (r) {
203                         fprintf(stderr, "%s: amdgpu_get_auth (2) failed (%i)\n",
204                                 __func__, r);
205                         pthread_mutex_unlock(&fd_mutex);
206                         return r;
207                 }
208                 if ((flag_auth) && (!flag_authexist)) {
209                         dev->flink_fd = dup(fd);
210                 }
211                 *major_version = dev->major_version;
212                 *minor_version = dev->minor_version;
213                 amdgpu_device_reference(device_handle, dev);
214                 pthread_mutex_unlock(&fd_mutex);
215                 return 0;
216         }
217
218         dev = calloc(1, sizeof(struct amdgpu_device));
219         if (!dev) {
220                 fprintf(stderr, "%s: calloc failed\n", __func__);
221                 pthread_mutex_unlock(&fd_mutex);
222                 return -ENOMEM;
223         }
224
225         dev->fd = -1;
226         dev->flink_fd = -1;
227
228         atomic_set(&dev->refcount, 1);
229
230         version = drmGetVersion(fd);
231         if (version->version_major != 3) {
232                 fprintf(stderr, "%s: DRM version is %d.%d.%d but this driver is "
233                         "only compatible with 3.x.x.\n",
234                         __func__,
235                         version->version_major,
236                         version->version_minor,
237                         version->version_patchlevel);
238                 drmFreeVersion(version);
239                 r = -EBADF;
240                 goto cleanup;
241         }
242
243         dev->fd = dup(fd);
244         dev->flink_fd = dev->fd;
245         dev->major_version = version->version_major;
246         dev->minor_version = version->version_minor;
247         drmFreeVersion(version);
248
249         dev->bo_flink_names = util_hash_table_create(handle_hash,
250                                                      handle_compare);
251         dev->bo_handles = util_hash_table_create(handle_hash, handle_compare);
252         pthread_mutex_init(&dev->bo_table_mutex, NULL);
253
254         /* Check if acceleration is working. */
255         r = amdgpu_query_info(dev, AMDGPU_INFO_ACCEL_WORKING, 4, &accel_working);
256         if (r) {
257                 fprintf(stderr, "%s: amdgpu_query_info(ACCEL_WORKING) failed (%i)\n",
258                         __func__, r);
259                 goto cleanup;
260         }
261         if (!accel_working) {
262                 fprintf(stderr, "%s: AMDGPU_INFO_ACCEL_WORKING = 0\n", __func__);
263                 r = -EBADF;
264                 goto cleanup;
265         }
266
267         r = amdgpu_query_gpu_info_init(dev);
268         if (r) {
269                 fprintf(stderr, "%s: amdgpu_query_gpu_info_init failed\n", __func__);
270                 goto cleanup;
271         }
272
273         start = dev->dev_info.virtual_address_offset;
274         max = MIN2(dev->dev_info.virtual_address_max, 0x100000000ULL);
275         amdgpu_vamgr_init(&dev->vamgr_32, start, max,
276                           dev->dev_info.virtual_address_alignment);
277
278         if (dev->dev_info.high_va_offset && dev->dev_info.high_va_max) {
279                 start = dev->dev_info.high_va_offset;
280                 max = dev->dev_info.high_va_max;
281         } else {
282                 start = MAX2(dev->dev_info.virtual_address_offset, 0x100000000ULL);
283                 max = MAX2(dev->dev_info.virtual_address_max, 0x100000000ULL);
284         }
285         amdgpu_vamgr_init(&dev->vamgr, start, max,
286                           dev->dev_info.virtual_address_alignment);
287
288         r = amdgpu_parse_asic_ids(&dev->asic_ids);
289         if (r) {
290                 fprintf(stderr, "%s: Cannot parse ASIC IDs, 0x%x.",
291                         __func__, r);
292         }
293
294         *major_version = dev->major_version;
295         *minor_version = dev->minor_version;
296         *device_handle = dev;
297         util_hash_table_set(fd_tab, UINT_TO_PTR(dev->fd), dev);
298         pthread_mutex_unlock(&fd_mutex);
299
300         return 0;
301
302 cleanup:
303         if (dev->fd >= 0)
304                 close(dev->fd);
305         free(dev);
306         pthread_mutex_unlock(&fd_mutex);
307         return r;
308 }
309
310 int amdgpu_device_deinitialize(amdgpu_device_handle dev)
311 {
312         amdgpu_device_reference(&dev, NULL);
313         return 0;
314 }
315
316 const char *amdgpu_get_marketing_name(amdgpu_device_handle dev)
317 {
318         const struct amdgpu_asic_id *id;
319
320         if (!dev->asic_ids)
321                 return NULL;
322
323         for (id = dev->asic_ids; id->did; id++) {
324                 if ((id->did == dev->info.asic_id) &&
325                     (id->rid == dev->info.pci_rev_id))
326                         return id->marketing_name;
327         }
328
329         return NULL;
330 }