• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2009 Corbin Simpson
3  * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
4  * All Rights Reserved.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining
7  * a copy of this software and associated documentation files (the
8  * "Software"), to deal in the Software without restriction, including
9  * without limitation the rights to use, copy, modify, merge, publish,
10  * distribute, sub license, and/or sell copies of the Software, and to
11  * permit persons to whom the Software is furnished to do so, subject to
12  * the following conditions:
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
15  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
16  * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
17  * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
18  * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
22  *
23  * The above copyright notice and this permission notice (including the
24  * next paragraph) shall be included in all copies or substantial portions
25  * of the Software.
26  */
27 /*
28  * Authors:
29  *      Corbin Simpson <MostAwesomeDude@gmail.com>
30  *      Joakim Sindholt <opensource@zhasha.com>
31  *      Marek Olšák <maraeo@gmail.com>
32  */
33 
34 #include "radeon_drm_bo.h"
35 #include "radeon_drm_cs.h"
36 #include "radeon_drm_public.h"
37 
38 #include "util/u_memory.h"
39 #include "util/u_hash_table.h"
40 
41 #include <xf86drm.h>
42 #include <stdio.h>
43 #include <sys/types.h>
44 #include <sys/stat.h>
45 #include <unistd.h>
46 #include <fcntl.h>
47 #include <radeon_surface.h>
48 
49 #ifndef RADEON_INFO_ACTIVE_CU_COUNT
50 #define RADEON_INFO_ACTIVE_CU_COUNT 0x20
51 #endif
52 
53 #ifndef RADEON_INFO_CURRENT_GPU_TEMP
54 #define RADEON_INFO_CURRENT_GPU_TEMP	0x21
55 #define RADEON_INFO_CURRENT_GPU_SCLK	0x22
56 #define RADEON_INFO_CURRENT_GPU_MCLK	0x23
57 #define RADEON_INFO_READ_REG		0x24
58 #endif
59 
60 #define RADEON_INFO_VA_UNMAP_WORKING	0x25
61 
62 #ifndef RADEON_INFO_GPU_RESET_COUNTER
63 #define RADEON_INFO_GPU_RESET_COUNTER   0x26
64 #endif
65 
66 static struct util_hash_table *fd_tab = NULL;
67 pipe_static_mutex(fd_tab_mutex);
68 
69 /* Enable/disable feature access for one command stream.
70  * If enable == true, return true on success.
71  * Otherwise, return false.
72  *
73  * We basically do the same thing kernel does, because we have to deal
74  * with multiple contexts (here command streams) backed by one winsys. */
radeon_set_fd_access(struct radeon_drm_cs * applier,struct radeon_drm_cs ** owner,pipe_mutex * mutex,unsigned request,const char * request_name,bool enable)75 static bool radeon_set_fd_access(struct radeon_drm_cs *applier,
76                                  struct radeon_drm_cs **owner,
77                                  pipe_mutex *mutex,
78                                  unsigned request, const char *request_name,
79                                  bool enable)
80 {
81     struct drm_radeon_info info;
82     unsigned value = enable ? 1 : 0;
83 
84     memset(&info, 0, sizeof(info));
85 
86     pipe_mutex_lock(*mutex);
87 
88     /* Early exit if we are sure the request will fail. */
89     if (enable) {
90         if (*owner) {
91             pipe_mutex_unlock(*mutex);
92             return false;
93         }
94     } else {
95         if (*owner != applier) {
96             pipe_mutex_unlock(*mutex);
97             return false;
98         }
99     }
100 
101     /* Pass through the request to the kernel. */
102     info.value = (unsigned long)&value;
103     info.request = request;
104     if (drmCommandWriteRead(applier->ws->fd, DRM_RADEON_INFO,
105                             &info, sizeof(info)) != 0) {
106         pipe_mutex_unlock(*mutex);
107         return false;
108     }
109 
110     /* Update the rights in the winsys. */
111     if (enable) {
112         if (value) {
113             *owner = applier;
114             pipe_mutex_unlock(*mutex);
115             return true;
116         }
117     } else {
118         *owner = NULL;
119     }
120 
121     pipe_mutex_unlock(*mutex);
122     return false;
123 }
124 
radeon_get_drm_value(int fd,unsigned request,const char * errname,uint32_t * out)125 static bool radeon_get_drm_value(int fd, unsigned request,
126                                  const char *errname, uint32_t *out)
127 {
128     struct drm_radeon_info info;
129     int retval;
130 
131     memset(&info, 0, sizeof(info));
132 
133     info.value = (unsigned long)out;
134     info.request = request;
135 
136     retval = drmCommandWriteRead(fd, DRM_RADEON_INFO, &info, sizeof(info));
137     if (retval) {
138         if (errname) {
139             fprintf(stderr, "radeon: Failed to get %s, error number %d\n",
140                     errname, retval);
141         }
142         return false;
143     }
144     return true;
145 }
146 
147 /* Helper function to do the ioctls needed for setup and init. */
do_winsys_init(struct radeon_drm_winsys * ws)148 static bool do_winsys_init(struct radeon_drm_winsys *ws)
149 {
150     struct drm_radeon_gem_info gem_info;
151     int retval;
152     drmVersionPtr version;
153 
154     memset(&gem_info, 0, sizeof(gem_info));
155 
156     /* We do things in a specific order here.
157      *
158      * DRM version first. We need to be sure we're running on a KMS chipset.
159      * This is also for some features.
160      *
161      * Then, the PCI ID. This is essential and should return usable numbers
162      * for all Radeons. If this fails, we probably got handed an FD for some
163      * non-Radeon card.
164      *
165      * The GEM info is actually bogus on the kernel side, as well as our side
166      * (see radeon_gem_info_ioctl in radeon_gem.c) but that's alright because
167      * we don't actually use the info for anything yet.
168      *
169      * The GB and Z pipe requests should always succeed, but they might not
170      * return sensical values for all chipsets, but that's alright because
171      * the pipe drivers already know that.
172      */
173 
174     /* Get DRM version. */
175     version = drmGetVersion(ws->fd);
176     if (version->version_major != 2 ||
177         version->version_minor < 12) {
178         fprintf(stderr, "%s: DRM version is %d.%d.%d but this driver is "
179                 "only compatible with 2.12.0 (kernel 3.2) or later.\n",
180                 __FUNCTION__,
181                 version->version_major,
182                 version->version_minor,
183                 version->version_patchlevel);
184         drmFreeVersion(version);
185         return false;
186     }
187 
188     ws->info.drm_major = version->version_major;
189     ws->info.drm_minor = version->version_minor;
190     ws->info.drm_patchlevel = version->version_patchlevel;
191     drmFreeVersion(version);
192 
193     /* Get PCI ID. */
194     if (!radeon_get_drm_value(ws->fd, RADEON_INFO_DEVICE_ID, "PCI ID",
195                               &ws->info.pci_id))
196         return false;
197 
198     /* Check PCI ID. */
199     switch (ws->info.pci_id) {
200 #define CHIPSET(pci_id, name, cfamily) case pci_id: ws->info.family = CHIP_##cfamily; ws->gen = DRV_R300; break;
201 #include "pci_ids/r300_pci_ids.h"
202 #undef CHIPSET
203 
204 #define CHIPSET(pci_id, name, cfamily) case pci_id: ws->info.family = CHIP_##cfamily; ws->gen = DRV_R600; break;
205 #include "pci_ids/r600_pci_ids.h"
206 #undef CHIPSET
207 
208 #define CHIPSET(pci_id, name, cfamily) case pci_id: ws->info.family = CHIP_##cfamily; ws->gen = DRV_SI; break;
209 #include "pci_ids/radeonsi_pci_ids.h"
210 #undef CHIPSET
211 
212     default:
213         fprintf(stderr, "radeon: Invalid PCI ID.\n");
214         return false;
215     }
216 
217     switch (ws->info.family) {
218     default:
219     case CHIP_UNKNOWN:
220         fprintf(stderr, "radeon: Unknown family.\n");
221         return false;
222     case CHIP_R300:
223     case CHIP_R350:
224     case CHIP_RV350:
225     case CHIP_RV370:
226     case CHIP_RV380:
227     case CHIP_RS400:
228     case CHIP_RC410:
229     case CHIP_RS480:
230         ws->info.chip_class = R300;
231         break;
232     case CHIP_R420:     /* R4xx-based cores. */
233     case CHIP_R423:
234     case CHIP_R430:
235     case CHIP_R480:
236     case CHIP_R481:
237     case CHIP_RV410:
238     case CHIP_RS600:
239     case CHIP_RS690:
240     case CHIP_RS740:
241         ws->info.chip_class = R400;
242         break;
243     case CHIP_RV515:    /* R5xx-based cores. */
244     case CHIP_R520:
245     case CHIP_RV530:
246     case CHIP_R580:
247     case CHIP_RV560:
248     case CHIP_RV570:
249         ws->info.chip_class = R500;
250         break;
251     case CHIP_R600:
252     case CHIP_RV610:
253     case CHIP_RV630:
254     case CHIP_RV670:
255     case CHIP_RV620:
256     case CHIP_RV635:
257     case CHIP_RS780:
258     case CHIP_RS880:
259         ws->info.chip_class = R600;
260         break;
261     case CHIP_RV770:
262     case CHIP_RV730:
263     case CHIP_RV710:
264     case CHIP_RV740:
265         ws->info.chip_class = R700;
266         break;
267     case CHIP_CEDAR:
268     case CHIP_REDWOOD:
269     case CHIP_JUNIPER:
270     case CHIP_CYPRESS:
271     case CHIP_HEMLOCK:
272     case CHIP_PALM:
273     case CHIP_SUMO:
274     case CHIP_SUMO2:
275     case CHIP_BARTS:
276     case CHIP_TURKS:
277     case CHIP_CAICOS:
278         ws->info.chip_class = EVERGREEN;
279         break;
280     case CHIP_CAYMAN:
281     case CHIP_ARUBA:
282         ws->info.chip_class = CAYMAN;
283         break;
284     case CHIP_TAHITI:
285     case CHIP_PITCAIRN:
286     case CHIP_VERDE:
287     case CHIP_OLAND:
288     case CHIP_HAINAN:
289         ws->info.chip_class = SI;
290         break;
291     case CHIP_BONAIRE:
292     case CHIP_KAVERI:
293     case CHIP_KABINI:
294     case CHIP_HAWAII:
295     case CHIP_MULLINS:
296         ws->info.chip_class = CIK;
297         break;
298     }
299 
300     /* Set which chips don't have dedicated VRAM. */
301     switch (ws->info.family) {
302     case CHIP_RS400:
303     case CHIP_RC410:
304     case CHIP_RS480:
305     case CHIP_RS600:
306     case CHIP_RS690:
307     case CHIP_RS740:
308     case CHIP_RS780:
309     case CHIP_RS880:
310     case CHIP_PALM:
311     case CHIP_SUMO:
312     case CHIP_SUMO2:
313     case CHIP_ARUBA:
314     case CHIP_KAVERI:
315     case CHIP_KABINI:
316     case CHIP_MULLINS:
317        ws->info.has_dedicated_vram = false;
318        break;
319 
320     default:
321        ws->info.has_dedicated_vram = true;
322     }
323 
324     /* Check for dma */
325     ws->info.has_sdma = false;
326     /* DMA is disabled on R700. There is IB corruption and hangs. */
327     if (ws->info.chip_class >= EVERGREEN && ws->info.drm_minor >= 27) {
328         ws->info.has_sdma = true;
329     }
330 
331     /* Check for UVD and VCE */
332     ws->info.has_uvd = false;
333     ws->info.vce_fw_version = 0x00000000;
334     if (ws->info.drm_minor >= 32) {
335 	uint32_t value = RADEON_CS_RING_UVD;
336         if (radeon_get_drm_value(ws->fd, RADEON_INFO_RING_WORKING,
337                                  "UVD Ring working", &value))
338             ws->info.has_uvd = value;
339 
340         value = RADEON_CS_RING_VCE;
341         if (radeon_get_drm_value(ws->fd, RADEON_INFO_RING_WORKING,
342                                  NULL, &value) && value) {
343 
344             if (radeon_get_drm_value(ws->fd, RADEON_INFO_VCE_FW_VERSION,
345                                      "VCE FW version", &value))
346                 ws->info.vce_fw_version = value;
347 	}
348     }
349 
350     /* Check for userptr support. */
351     {
352         struct drm_radeon_gem_userptr args = {0};
353 
354         /* If the ioctl doesn't exist, -EINVAL is returned.
355          *
356          * If the ioctl exists, it should return -EACCES
357          * if RADEON_GEM_USERPTR_READONLY or RADEON_GEM_USERPTR_REGISTER
358          * aren't set.
359          */
360         ws->info.has_userptr =
361             drmCommandWriteRead(ws->fd, DRM_RADEON_GEM_USERPTR,
362                                 &args, sizeof(args)) == -EACCES;
363     }
364 
365     /* Get GEM info. */
366     retval = drmCommandWriteRead(ws->fd, DRM_RADEON_GEM_INFO,
367             &gem_info, sizeof(gem_info));
368     if (retval) {
369         fprintf(stderr, "radeon: Failed to get MM info, error number %d\n",
370                 retval);
371         return false;
372     }
373     ws->info.gart_size = gem_info.gart_size;
374     ws->info.vram_size = gem_info.vram_size;
375 
376     /* Radeon allocates all buffers as contigous, which makes large allocations
377      * unlikely to succeed. */
378     ws->info.max_alloc_size = MAX2(ws->info.vram_size, ws->info.gart_size) * 0.7;
379     if (ws->info.drm_minor < 40)
380         ws->info.max_alloc_size = MIN2(ws->info.max_alloc_size, 256*1024*1024);
381 
382     /* Get max clock frequency info and convert it to MHz */
383     radeon_get_drm_value(ws->fd, RADEON_INFO_MAX_SCLK, NULL,
384                          &ws->info.max_shader_clock);
385     ws->info.max_shader_clock /= 1000;
386 
387     radeon_get_drm_value(ws->fd, RADEON_INFO_SI_BACKEND_ENABLED_MASK, NULL,
388                          &ws->info.enabled_rb_mask);
389 
390     ws->num_cpus = sysconf(_SC_NPROCESSORS_ONLN);
391 
392     /* Generation-specific queries. */
393     if (ws->gen == DRV_R300) {
394         if (!radeon_get_drm_value(ws->fd, RADEON_INFO_NUM_GB_PIPES,
395                                   "GB pipe count",
396                                   &ws->info.r300_num_gb_pipes))
397             return false;
398 
399         if (!radeon_get_drm_value(ws->fd, RADEON_INFO_NUM_Z_PIPES,
400                                   "Z pipe count",
401                                   &ws->info.r300_num_z_pipes))
402             return false;
403     }
404     else if (ws->gen >= DRV_R600) {
405         uint32_t tiling_config = 0;
406 
407         if (!radeon_get_drm_value(ws->fd, RADEON_INFO_NUM_BACKENDS,
408                                   "num backends",
409                                   &ws->info.num_render_backends))
410             return false;
411 
412         /* get the GPU counter frequency, failure is not fatal */
413         radeon_get_drm_value(ws->fd, RADEON_INFO_CLOCK_CRYSTAL_FREQ, NULL,
414                              &ws->info.clock_crystal_freq);
415 
416         radeon_get_drm_value(ws->fd, RADEON_INFO_TILING_CONFIG, NULL,
417                              &tiling_config);
418 
419         ws->info.r600_num_banks =
420             ws->info.chip_class >= EVERGREEN ?
421                 4 << ((tiling_config & 0xf0) >> 4) :
422                 4 << ((tiling_config & 0x30) >> 4);
423 
424         ws->info.pipe_interleave_bytes =
425             ws->info.chip_class >= EVERGREEN ?
426                 256 << ((tiling_config & 0xf00) >> 8) :
427                 256 << ((tiling_config & 0xc0) >> 6);
428 
429         if (!ws->info.pipe_interleave_bytes)
430             ws->info.pipe_interleave_bytes =
431                 ws->info.chip_class >= EVERGREEN ? 512 : 256;
432 
433         radeon_get_drm_value(ws->fd, RADEON_INFO_NUM_TILE_PIPES, NULL,
434                              &ws->info.num_tile_pipes);
435 
436         /* "num_tiles_pipes" must be equal to the number of pipes (Px) in the
437          * pipe config field of the GB_TILE_MODE array. Only one card (Tahiti)
438          * reports a different value (12). Fix it by setting what's in the
439          * GB_TILE_MODE array (8).
440          */
441         if (ws->gen == DRV_SI && ws->info.num_tile_pipes == 12)
442             ws->info.num_tile_pipes = 8;
443 
444         if (radeon_get_drm_value(ws->fd, RADEON_INFO_BACKEND_MAP, NULL,
445                                   &ws->info.r600_gb_backend_map))
446             ws->info.r600_gb_backend_map_valid = true;
447 
448         ws->info.has_virtual_memory = false;
449         if (ws->info.drm_minor >= 13) {
450             uint32_t ib_vm_max_size;
451 
452             ws->info.has_virtual_memory = true;
453             if (!radeon_get_drm_value(ws->fd, RADEON_INFO_VA_START, NULL,
454                                       &ws->va_start))
455                 ws->info.has_virtual_memory = false;
456             if (!radeon_get_drm_value(ws->fd, RADEON_INFO_IB_VM_MAX_SIZE, NULL,
457                                       &ib_vm_max_size))
458                 ws->info.has_virtual_memory = false;
459             radeon_get_drm_value(ws->fd, RADEON_INFO_VA_UNMAP_WORKING, NULL,
460                                  &ws->va_unmap_working);
461         }
462 	if (ws->gen == DRV_R600 && !debug_get_bool_option("RADEON_VA", false))
463 		ws->info.has_virtual_memory = false;
464     }
465 
466     /* Get max pipes, this is only needed for compute shaders.  All evergreen+
467      * chips have at least 2 pipes, so we use 2 as a default. */
468     ws->info.r600_max_quad_pipes = 2;
469     radeon_get_drm_value(ws->fd, RADEON_INFO_MAX_PIPES, NULL,
470                          &ws->info.r600_max_quad_pipes);
471 
472     /* All GPUs have at least one compute unit */
473     ws->info.num_good_compute_units = 1;
474     radeon_get_drm_value(ws->fd, RADEON_INFO_ACTIVE_CU_COUNT, NULL,
475                          &ws->info.num_good_compute_units);
476 
477     radeon_get_drm_value(ws->fd, RADEON_INFO_MAX_SE, NULL,
478                          &ws->info.max_se);
479 
480     if (!ws->info.max_se) {
481         switch (ws->info.family) {
482         default:
483             ws->info.max_se = 1;
484             break;
485         case CHIP_CYPRESS:
486         case CHIP_HEMLOCK:
487         case CHIP_BARTS:
488         case CHIP_CAYMAN:
489         case CHIP_TAHITI:
490         case CHIP_PITCAIRN:
491         case CHIP_BONAIRE:
492             ws->info.max_se = 2;
493             break;
494         case CHIP_HAWAII:
495             ws->info.max_se = 4;
496             break;
497         }
498     }
499 
500     radeon_get_drm_value(ws->fd, RADEON_INFO_MAX_SH_PER_SE, NULL,
501                          &ws->info.max_sh_per_se);
502 
503     radeon_get_drm_value(ws->fd, RADEON_INFO_ACCEL_WORKING2, NULL,
504                          &ws->accel_working2);
505     if (ws->info.family == CHIP_HAWAII && ws->accel_working2 < 2) {
506         fprintf(stderr, "radeon: GPU acceleration for Hawaii disabled, "
507                 "returned accel_working2 value %u is smaller than 2. "
508                 "Please install a newer kernel.\n",
509                 ws->accel_working2);
510         return false;
511     }
512 
513     if (ws->info.chip_class == CIK) {
514         if (!radeon_get_drm_value(ws->fd, RADEON_INFO_CIK_MACROTILE_MODE_ARRAY, NULL,
515                                   ws->info.cik_macrotile_mode_array)) {
516             fprintf(stderr, "radeon: Kernel 3.13 is required for CIK support.\n");
517             return false;
518         }
519     }
520 
521     if (ws->info.chip_class >= SI) {
522         if (!radeon_get_drm_value(ws->fd, RADEON_INFO_SI_TILE_MODE_ARRAY, NULL,
523                                   ws->info.si_tile_mode_array)) {
524             fprintf(stderr, "radeon: Kernel 3.10 is required for SI support.\n");
525             return false;
526         }
527     }
528 
529     /* Hawaii with old firmware needs type2 nop packet.
530      * accel_working2 with value 3 indicates the new firmware.
531      */
532     ws->info.gfx_ib_pad_with_type2 = ws->info.chip_class <= SI ||
533 				     (ws->info.family == CHIP_HAWAII &&
534 				      ws->accel_working2 < 3);
535 
536     ws->check_vm = strstr(debug_get_option("R600_DEBUG", ""), "check_vm") != NULL;
537 
538     return true;
539 }
540 
radeon_winsys_destroy(struct radeon_winsys * rws)541 static void radeon_winsys_destroy(struct radeon_winsys *rws)
542 {
543     struct radeon_drm_winsys *ws = (struct radeon_drm_winsys*)rws;
544 
545     if (util_queue_is_initialized(&ws->cs_queue))
546         util_queue_destroy(&ws->cs_queue);
547 
548     pipe_mutex_destroy(ws->hyperz_owner_mutex);
549     pipe_mutex_destroy(ws->cmask_owner_mutex);
550 
551     if (ws->info.has_virtual_memory)
552         pb_slabs_deinit(&ws->bo_slabs);
553     pb_cache_deinit(&ws->bo_cache);
554 
555     if (ws->gen >= DRV_R600) {
556         radeon_surface_manager_free(ws->surf_man);
557     }
558 
559     util_hash_table_destroy(ws->bo_names);
560     util_hash_table_destroy(ws->bo_handles);
561     util_hash_table_destroy(ws->bo_vas);
562     pipe_mutex_destroy(ws->bo_handles_mutex);
563     pipe_mutex_destroy(ws->bo_va_mutex);
564     pipe_mutex_destroy(ws->bo_fence_lock);
565 
566     if (ws->fd >= 0)
567         close(ws->fd);
568 
569     FREE(rws);
570 }
571 
radeon_query_info(struct radeon_winsys * rws,struct radeon_info * info)572 static void radeon_query_info(struct radeon_winsys *rws,
573                               struct radeon_info *info)
574 {
575     *info = ((struct radeon_drm_winsys *)rws)->info;
576 }
577 
radeon_cs_request_feature(struct radeon_winsys_cs * rcs,enum radeon_feature_id fid,bool enable)578 static bool radeon_cs_request_feature(struct radeon_winsys_cs *rcs,
579                                       enum radeon_feature_id fid,
580                                       bool enable)
581 {
582     struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
583 
584     switch (fid) {
585     case RADEON_FID_R300_HYPERZ_ACCESS:
586         return radeon_set_fd_access(cs, &cs->ws->hyperz_owner,
587                                     &cs->ws->hyperz_owner_mutex,
588                                     RADEON_INFO_WANT_HYPERZ, "Hyper-Z",
589                                     enable);
590 
591     case RADEON_FID_R300_CMASK_ACCESS:
592         return radeon_set_fd_access(cs, &cs->ws->cmask_owner,
593                                     &cs->ws->cmask_owner_mutex,
594                                     RADEON_INFO_WANT_CMASK, "AA optimizations",
595                                     enable);
596     }
597     return false;
598 }
599 
radeon_query_value(struct radeon_winsys * rws,enum radeon_value_id value)600 static uint64_t radeon_query_value(struct radeon_winsys *rws,
601                                    enum radeon_value_id value)
602 {
603     struct radeon_drm_winsys *ws = (struct radeon_drm_winsys*)rws;
604     uint64_t retval = 0;
605 
606     switch (value) {
607     case RADEON_REQUESTED_VRAM_MEMORY:
608         return ws->allocated_vram;
609     case RADEON_REQUESTED_GTT_MEMORY:
610         return ws->allocated_gtt;
611     case RADEON_MAPPED_VRAM:
612        return ws->mapped_vram;
613     case RADEON_MAPPED_GTT:
614        return ws->mapped_gtt;
615     case RADEON_BUFFER_WAIT_TIME_NS:
616         return ws->buffer_wait_time;
617     case RADEON_TIMESTAMP:
618         if (ws->info.drm_minor < 20 || ws->gen < DRV_R600) {
619             assert(0);
620             return 0;
621         }
622 
623         radeon_get_drm_value(ws->fd, RADEON_INFO_TIMESTAMP, "timestamp",
624                              (uint32_t*)&retval);
625         return retval;
626     case RADEON_NUM_GFX_IBS:
627         return ws->num_gfx_IBs;
628     case RADEON_NUM_SDMA_IBS:
629         return ws->num_sdma_IBs;
630     case RADEON_NUM_BYTES_MOVED:
631         radeon_get_drm_value(ws->fd, RADEON_INFO_NUM_BYTES_MOVED,
632                              "num-bytes-moved", (uint32_t*)&retval);
633         return retval;
634     case RADEON_NUM_EVICTIONS:
635         return 0; /* unimplemented */
636     case RADEON_VRAM_USAGE:
637         radeon_get_drm_value(ws->fd, RADEON_INFO_VRAM_USAGE,
638                              "vram-usage", (uint32_t*)&retval);
639         return retval;
640     case RADEON_GTT_USAGE:
641         radeon_get_drm_value(ws->fd, RADEON_INFO_GTT_USAGE,
642                              "gtt-usage", (uint32_t*)&retval);
643         return retval;
644     case RADEON_GPU_TEMPERATURE:
645         radeon_get_drm_value(ws->fd, RADEON_INFO_CURRENT_GPU_TEMP,
646                              "gpu-temp", (uint32_t*)&retval);
647         return retval;
648     case RADEON_CURRENT_SCLK:
649         radeon_get_drm_value(ws->fd, RADEON_INFO_CURRENT_GPU_SCLK,
650                              "current-gpu-sclk", (uint32_t*)&retval);
651         return retval;
652     case RADEON_CURRENT_MCLK:
653         radeon_get_drm_value(ws->fd, RADEON_INFO_CURRENT_GPU_MCLK,
654                              "current-gpu-mclk", (uint32_t*)&retval);
655         return retval;
656     case RADEON_GPU_RESET_COUNTER:
657         radeon_get_drm_value(ws->fd, RADEON_INFO_GPU_RESET_COUNTER,
658                              "gpu-reset-counter", (uint32_t*)&retval);
659         return retval;
660     }
661     return 0;
662 }
663 
radeon_read_registers(struct radeon_winsys * rws,unsigned reg_offset,unsigned num_registers,uint32_t * out)664 static bool radeon_read_registers(struct radeon_winsys *rws,
665                                   unsigned reg_offset,
666                                   unsigned num_registers, uint32_t *out)
667 {
668     struct radeon_drm_winsys *ws = (struct radeon_drm_winsys*)rws;
669     unsigned i;
670 
671     for (i = 0; i < num_registers; i++) {
672         uint32_t reg = reg_offset + i*4;
673 
674         if (!radeon_get_drm_value(ws->fd, RADEON_INFO_READ_REG, NULL, &reg))
675             return false;
676         out[i] = reg;
677     }
678     return true;
679 }
680 
hash_fd(void * key)681 static unsigned hash_fd(void *key)
682 {
683     int fd = pointer_to_intptr(key);
684     struct stat stat;
685     fstat(fd, &stat);
686 
687     return stat.st_dev ^ stat.st_ino ^ stat.st_rdev;
688 }
689 
compare_fd(void * key1,void * key2)690 static int compare_fd(void *key1, void *key2)
691 {
692     int fd1 = pointer_to_intptr(key1);
693     int fd2 = pointer_to_intptr(key2);
694     struct stat stat1, stat2;
695     fstat(fd1, &stat1);
696     fstat(fd2, &stat2);
697 
698     return stat1.st_dev != stat2.st_dev ||
699            stat1.st_ino != stat2.st_ino ||
700            stat1.st_rdev != stat2.st_rdev;
701 }
702 
703 DEBUG_GET_ONCE_BOOL_OPTION(thread, "RADEON_THREAD", true)
704 
radeon_winsys_unref(struct radeon_winsys * ws)705 static bool radeon_winsys_unref(struct radeon_winsys *ws)
706 {
707     struct radeon_drm_winsys *rws = (struct radeon_drm_winsys*)ws;
708     bool destroy;
709 
710     /* When the reference counter drops to zero, remove the fd from the table.
711      * This must happen while the mutex is locked, so that
712      * radeon_drm_winsys_create in another thread doesn't get the winsys
713      * from the table when the counter drops to 0. */
714     pipe_mutex_lock(fd_tab_mutex);
715 
716     destroy = pipe_reference(&rws->reference, NULL);
717     if (destroy && fd_tab)
718         util_hash_table_remove(fd_tab, intptr_to_pointer(rws->fd));
719 
720     pipe_mutex_unlock(fd_tab_mutex);
721     return destroy;
722 }
723 
724 #define PTR_TO_UINT(x) ((unsigned)((intptr_t)(x)))
725 
handle_hash(void * key)726 static unsigned handle_hash(void *key)
727 {
728     return PTR_TO_UINT(key);
729 }
730 
handle_compare(void * key1,void * key2)731 static int handle_compare(void *key1, void *key2)
732 {
733     return PTR_TO_UINT(key1) != PTR_TO_UINT(key2);
734 }
735 
736 PUBLIC struct radeon_winsys *
radeon_drm_winsys_create(int fd,radeon_screen_create_t screen_create)737 radeon_drm_winsys_create(int fd, radeon_screen_create_t screen_create)
738 {
739     struct radeon_drm_winsys *ws;
740 
741     pipe_mutex_lock(fd_tab_mutex);
742     if (!fd_tab) {
743         fd_tab = util_hash_table_create(hash_fd, compare_fd);
744     }
745 
746     ws = util_hash_table_get(fd_tab, intptr_to_pointer(fd));
747     if (ws) {
748         pipe_reference(NULL, &ws->reference);
749         pipe_mutex_unlock(fd_tab_mutex);
750         return &ws->base;
751     }
752 
753     ws = CALLOC_STRUCT(radeon_drm_winsys);
754     if (!ws) {
755         pipe_mutex_unlock(fd_tab_mutex);
756         return NULL;
757     }
758 
759     ws->fd = fcntl(fd, F_DUPFD_CLOEXEC, 3);
760 
761     if (!do_winsys_init(ws))
762         goto fail1;
763 
764     pb_cache_init(&ws->bo_cache, 500000, ws->check_vm ? 1.0f : 2.0f, 0,
765                   MIN2(ws->info.vram_size, ws->info.gart_size),
766                   radeon_bo_destroy,
767                   radeon_bo_can_reclaim);
768 
769     if (ws->info.has_virtual_memory) {
770         /* There is no fundamental obstacle to using slab buffer allocation
771          * without GPUVM, but enabling it requires making sure that the drivers
772          * honor the address offset.
773          */
774         if (!pb_slabs_init(&ws->bo_slabs,
775                            RADEON_SLAB_MIN_SIZE_LOG2, RADEON_SLAB_MAX_SIZE_LOG2,
776                            12,
777                            ws,
778                            radeon_bo_can_reclaim_slab,
779                            radeon_bo_slab_alloc,
780                            radeon_bo_slab_free))
781             goto fail_cache;
782 
783         ws->info.min_alloc_size = 1 << RADEON_SLAB_MIN_SIZE_LOG2;
784     } else {
785         ws->info.min_alloc_size = ws->info.gart_page_size;
786     }
787 
788     if (ws->gen >= DRV_R600) {
789         ws->surf_man = radeon_surface_manager_new(ws->fd);
790         if (!ws->surf_man)
791             goto fail_slab;
792     }
793 
794     /* init reference */
795     pipe_reference_init(&ws->reference, 1);
796 
797     /* Set functions. */
798     ws->base.unref = radeon_winsys_unref;
799     ws->base.destroy = radeon_winsys_destroy;
800     ws->base.query_info = radeon_query_info;
801     ws->base.cs_request_feature = radeon_cs_request_feature;
802     ws->base.query_value = radeon_query_value;
803     ws->base.read_registers = radeon_read_registers;
804 
805     radeon_drm_bo_init_functions(ws);
806     radeon_drm_cs_init_functions(ws);
807     radeon_surface_init_functions(ws);
808 
809     pipe_mutex_init(ws->hyperz_owner_mutex);
810     pipe_mutex_init(ws->cmask_owner_mutex);
811 
812     ws->bo_names = util_hash_table_create(handle_hash, handle_compare);
813     ws->bo_handles = util_hash_table_create(handle_hash, handle_compare);
814     ws->bo_vas = util_hash_table_create(handle_hash, handle_compare);
815     pipe_mutex_init(ws->bo_handles_mutex);
816     pipe_mutex_init(ws->bo_va_mutex);
817     pipe_mutex_init(ws->bo_fence_lock);
818     ws->va_offset = ws->va_start;
819     list_inithead(&ws->va_holes);
820 
821     /* TTM aligns the BO size to the CPU page size */
822     ws->info.gart_page_size = sysconf(_SC_PAGESIZE);
823 
824     if (ws->num_cpus > 1 && debug_get_option_thread())
825         util_queue_init(&ws->cs_queue, "radeon_cs", 8, 1);
826 
827     /* Create the screen at the end. The winsys must be initialized
828      * completely.
829      *
830      * Alternatively, we could create the screen based on "ws->gen"
831      * and link all drivers into one binary blob. */
832     ws->base.screen = screen_create(&ws->base);
833     if (!ws->base.screen) {
834         radeon_winsys_destroy(&ws->base);
835         pipe_mutex_unlock(fd_tab_mutex);
836         return NULL;
837     }
838 
839     util_hash_table_set(fd_tab, intptr_to_pointer(ws->fd), ws);
840 
841     /* We must unlock the mutex once the winsys is fully initialized, so that
842      * other threads attempting to create the winsys from the same fd will
843      * get a fully initialized winsys and not just half-way initialized. */
844     pipe_mutex_unlock(fd_tab_mutex);
845 
846     return &ws->base;
847 
848 fail_slab:
849     if (ws->info.has_virtual_memory)
850         pb_slabs_deinit(&ws->bo_slabs);
851 fail_cache:
852     pb_cache_deinit(&ws->bo_cache);
853 fail1:
854     pipe_mutex_unlock(fd_tab_mutex);
855     if (ws->surf_man)
856         radeon_surface_manager_free(ws->surf_man);
857     if (ws->fd >= 0)
858         close(ws->fd);
859 
860     FREE(ws);
861     return NULL;
862 }
863