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#include "radeon_drm_bo.h"
29#include "radeon_drm_cs.h"
30#include "radeon_drm_public.h"
31
32#include "util/os_file.h"
33#include "util/u_cpu_detect.h"
34#include "util/u_memory.h"
35#include "util/u_hash_table.h"
36#include "util/u_pointer.h"
37
38#include <xf86drm.h>
39#include <stdio.h>
40#include <sys/types.h>
41#include <sys/stat.h>
42#include <unistd.h>
43#include <fcntl.h>
44#include <radeon_surface.h>
45
46static struct hash_table *fd_tab = NULL;
47static mtx_t fd_tab_mutex = _MTX_INITIALIZER_NP;
48
49/* Enable/disable feature access for one command stream.
50 * If enable == true, return true on success.
51 * Otherwise, return false.
52 *
53 * We basically do the same thing kernel does, because we have to deal
54 * with multiple contexts (here command streams) backed by one winsys. */
55static bool radeon_set_fd_access(struct radeon_drm_cs *applier,
56                                 struct radeon_drm_cs **owner,
57                                 mtx_t *mutex,
58                                 unsigned request, const char *request_name,
59                                 bool enable)
60{
61   struct drm_radeon_info info;
62   unsigned value = enable ? 1 : 0;
63
64   memset(&info, 0, sizeof(info));
65
66   mtx_lock(&*mutex);
67
68   /* Early exit if we are sure the request will fail. */
69   if (enable) {
70      if (*owner) {
71         mtx_unlock(&*mutex);
72         return false;
73      }
74   } else {
75      if (*owner != applier) {
76         mtx_unlock(&*mutex);
77         return false;
78      }
79   }
80
81   /* Pass through the request to the kernel. */
82   info.value = (unsigned long)&value;
83   info.request = request;
84   if (drmCommandWriteRead(applier->ws->fd, DRM_RADEON_INFO,
85                           &info, sizeof(info)) != 0) {
86      mtx_unlock(&*mutex);
87      return false;
88   }
89
90   /* Update the rights in the winsys. */
91   if (enable) {
92      if (value) {
93         *owner = applier;
94         mtx_unlock(&*mutex);
95         return true;
96      }
97   } else {
98      *owner = NULL;
99   }
100
101   mtx_unlock(&*mutex);
102   return false;
103}
104
105static bool radeon_get_drm_value(int fd, unsigned request,
106                                 const char *errname, uint32_t *out)
107{
108   struct drm_radeon_info info;
109   int retval;
110
111   memset(&info, 0, sizeof(info));
112
113   info.value = (unsigned long)out;
114   info.request = request;
115
116   retval = drmCommandWriteRead(fd, DRM_RADEON_INFO, &info, sizeof(info));
117   if (retval) {
118      if (errname) {
119         fprintf(stderr, "radeon: Failed to get %s, error number %d\n",
120                 errname, retval);
121      }
122      return false;
123   }
124   return true;
125}
126
127/* Helper function to do the ioctls needed for setup and init. */
128static bool do_winsys_init(struct radeon_drm_winsys *ws)
129{
130   struct drm_radeon_gem_info gem_info;
131   int retval;
132   drmVersionPtr version;
133
134   memset(&gem_info, 0, sizeof(gem_info));
135
136   /* We do things in a specific order here.
137    *
138    * DRM version first. We need to be sure we're running on a KMS chipset.
139    * This is also for some features.
140    *
141    * Then, the PCI ID. This is essential and should return usable numbers
142    * for all Radeons. If this fails, we probably got handed an FD for some
143    * non-Radeon card.
144    *
145    * The GEM info is actually bogus on the kernel side, as well as our side
146    * (see radeon_gem_info_ioctl in radeon_gem.c) but that's alright because
147    * we don't actually use the info for anything yet.
148    *
149    * The GB and Z pipe requests should always succeed, but they might not
150    * return sensical values for all chipsets, but that's alright because
151    * the pipe drivers already know that.
152    */
153
154   /* Get DRM version. */
155   version = drmGetVersion(ws->fd);
156   if (version->version_major != 2 ||
157       version->version_minor < 12) {
158      fprintf(stderr, "%s: DRM version is %d.%d.%d but this driver is "
159                      "only compatible with 2.12.0 (kernel 3.2) or later.\n",
160              __FUNCTION__,
161              version->version_major,
162              version->version_minor,
163              version->version_patchlevel);
164      drmFreeVersion(version);
165      return false;
166   }
167
168   ws->info.drm_major = version->version_major;
169   ws->info.drm_minor = version->version_minor;
170   ws->info.drm_patchlevel = version->version_patchlevel;
171   ws->info.is_amdgpu = false;
172   drmFreeVersion(version);
173
174   /* Get PCI ID. */
175   if (!radeon_get_drm_value(ws->fd, RADEON_INFO_DEVICE_ID, "PCI ID",
176                             &ws->info.pci_id))
177      return false;
178
179   /* Check PCI ID. */
180   switch (ws->info.pci_id) {
181#define CHIPSET(pci_id, name, cfamily) case pci_id: ws->info.family = CHIP_##cfamily; ws->gen = DRV_R300; break;
182#include "pci_ids/r300_pci_ids.h"
183#undef CHIPSET
184
185#define CHIPSET(pci_id, name, cfamily) case pci_id: ws->info.family = CHIP_##cfamily; ws->gen = DRV_R600; break;
186#include "pci_ids/r600_pci_ids.h"
187#undef CHIPSET
188
189#define CHIPSET(pci_id, cfamily) \
190   case pci_id: \
191   ws->info.family = CHIP_##cfamily; \
192   ws->info.name = #cfamily; \
193   ws->gen = DRV_SI; \
194   break;
195#include "pci_ids/radeonsi_pci_ids.h"
196#undef CHIPSET
197
198   default:
199      fprintf(stderr, "radeon: Invalid PCI ID.\n");
200      return false;
201   }
202
203   switch (ws->info.family) {
204   default:
205   case CHIP_UNKNOWN:
206      fprintf(stderr, "radeon: Unknown family.\n");
207      return false;
208   case CHIP_R300:
209   case CHIP_R350:
210   case CHIP_RV350:
211   case CHIP_RV370:
212   case CHIP_RV380:
213   case CHIP_RS400:
214   case CHIP_RC410:
215   case CHIP_RS480:
216      ws->info.chip_class = R300;
217      break;
218   case CHIP_R420:     /* R4xx-based cores. */
219   case CHIP_R423:
220   case CHIP_R430:
221   case CHIP_R480:
222   case CHIP_R481:
223   case CHIP_RV410:
224   case CHIP_RS600:
225   case CHIP_RS690:
226   case CHIP_RS740:
227      ws->info.chip_class = R400;
228      break;
229   case CHIP_RV515:    /* R5xx-based cores. */
230   case CHIP_R520:
231   case CHIP_RV530:
232   case CHIP_R580:
233   case CHIP_RV560:
234   case CHIP_RV570:
235      ws->info.chip_class = R500;
236      break;
237   case CHIP_R600:
238   case CHIP_RV610:
239   case CHIP_RV630:
240   case CHIP_RV670:
241   case CHIP_RV620:
242   case CHIP_RV635:
243   case CHIP_RS780:
244   case CHIP_RS880:
245      ws->info.chip_class = R600;
246      break;
247   case CHIP_RV770:
248   case CHIP_RV730:
249   case CHIP_RV710:
250   case CHIP_RV740:
251      ws->info.chip_class = R700;
252      break;
253   case CHIP_CEDAR:
254   case CHIP_REDWOOD:
255   case CHIP_JUNIPER:
256   case CHIP_CYPRESS:
257   case CHIP_HEMLOCK:
258   case CHIP_PALM:
259   case CHIP_SUMO:
260   case CHIP_SUMO2:
261   case CHIP_BARTS:
262   case CHIP_TURKS:
263   case CHIP_CAICOS:
264      ws->info.chip_class = EVERGREEN;
265      break;
266   case CHIP_CAYMAN:
267   case CHIP_ARUBA:
268      ws->info.chip_class = CAYMAN;
269      break;
270   case CHIP_TAHITI:
271   case CHIP_PITCAIRN:
272   case CHIP_VERDE:
273   case CHIP_OLAND:
274   case CHIP_HAINAN:
275      ws->info.chip_class = GFX6;
276      break;
277   case CHIP_BONAIRE:
278   case CHIP_KAVERI:
279   case CHIP_KABINI:
280   case CHIP_HAWAII:
281      ws->info.chip_class = GFX7;
282      break;
283   }
284
285   /* Set which chips don't have dedicated VRAM. */
286   switch (ws->info.family) {
287   case CHIP_RS400:
288   case CHIP_RC410:
289   case CHIP_RS480:
290   case CHIP_RS600:
291   case CHIP_RS690:
292   case CHIP_RS740:
293   case CHIP_RS780:
294   case CHIP_RS880:
295   case CHIP_PALM:
296   case CHIP_SUMO:
297   case CHIP_SUMO2:
298   case CHIP_ARUBA:
299   case CHIP_KAVERI:
300   case CHIP_KABINI:
301      ws->info.has_dedicated_vram = false;
302      break;
303
304   default:
305      ws->info.has_dedicated_vram = true;
306   }
307
308   ws->info.num_rings[RING_GFX] = 1;
309   /* Check for dma */
310   ws->info.num_rings[RING_DMA] = 0;
311   /* DMA is disabled on R700. There is IB corruption and hangs. */
312   if (ws->info.chip_class >= EVERGREEN && ws->info.drm_minor >= 27) {
313      ws->info.num_rings[RING_DMA] = 1;
314   }
315
316   /* Check for UVD and VCE */
317   ws->info.has_video_hw.uvd_decode = false;
318   ws->info.has_video_hw.vce_encode = false;
319   ws->info.vce_fw_version = 0x00000000;
320   if (ws->info.drm_minor >= 32) {
321      uint32_t value = RADEON_CS_RING_UVD;
322      if (radeon_get_drm_value(ws->fd, RADEON_INFO_RING_WORKING,
323                               "UVD Ring working", &value)) {
324         ws->info.has_video_hw.uvd_decode = value;
325         ws->info.num_rings[RING_UVD] = 1;
326      }
327
328      value = RADEON_CS_RING_VCE;
329      if (radeon_get_drm_value(ws->fd, RADEON_INFO_RING_WORKING,
330                               NULL, &value) && value) {
331
332         if (radeon_get_drm_value(ws->fd, RADEON_INFO_VCE_FW_VERSION,
333                                  "VCE FW version", &value)) {
334            ws->info.vce_fw_version = value;
335            ws->info.num_rings[RING_VCE] = 1;
336            ws->info.has_video_hw.vce_encode = true;
337         }
338      }
339   }
340
341   /* Check for userptr support. */
342   {
343      struct drm_radeon_gem_userptr args = {0};
344
345      /* If the ioctl doesn't exist, -EINVAL is returned.
346       *
347       * If the ioctl exists, it should return -EACCES
348       * if RADEON_GEM_USERPTR_READONLY or RADEON_GEM_USERPTR_REGISTER
349       * aren't set.
350       */
351      ws->info.has_userptr =
352            drmCommandWriteRead(ws->fd, DRM_RADEON_GEM_USERPTR,
353                                &args, sizeof(args)) == -EACCES;
354   }
355
356   /* Get GEM info. */
357   retval = drmCommandWriteRead(ws->fd, DRM_RADEON_GEM_INFO,
358                                &gem_info, sizeof(gem_info));
359   if (retval) {
360      fprintf(stderr, "radeon: Failed to get MM info, error number %d\n",
361              retval);
362      return false;
363   }
364   ws->info.gart_size = gem_info.gart_size;
365   ws->info.vram_size = gem_info.vram_size;
366   ws->info.vram_vis_size = gem_info.vram_visible;
367   /* Older versions of the kernel driver reported incorrect values, and
368    * didn't support more than 256MB of visible VRAM anyway
369    */
370   if (ws->info.drm_minor < 49)
371      ws->info.vram_vis_size = MIN2(ws->info.vram_vis_size, 256*1024*1024);
372
373   ws->info.gart_size_kb = DIV_ROUND_UP(ws->info.gart_size, 1024);
374   ws->info.vram_size_kb = DIV_ROUND_UP(ws->info.vram_size, 1024);
375
376   /* Radeon allocates all buffers contiguously, which makes large allocations
377    * unlikely to succeed. */
378   if (ws->info.has_dedicated_vram)
379      ws->info.max_alloc_size = ws->info.vram_size * 0.7;
380   else
381      ws->info.max_alloc_size = ws->info.gart_size * 0.7;
382
383   if (ws->info.drm_minor < 40)
384      ws->info.max_alloc_size = MIN2(ws->info.max_alloc_size, 256*1024*1024);
385   /* Both 32-bit and 64-bit address spaces only have 4GB. */
386   ws->info.max_alloc_size = MIN2(ws->info.max_alloc_size, 3ull*1024*1024*1024);
387
388   /* Get max clock frequency info and convert it to MHz */
389   radeon_get_drm_value(ws->fd, RADEON_INFO_MAX_SCLK, NULL,
390                        &ws->info.max_shader_clock);
391   ws->info.max_shader_clock /= 1000;
392
393   ws->num_cpus = sysconf(_SC_NPROCESSORS_ONLN);
394
395   /* Generation-specific queries. */
396   if (ws->gen == DRV_R300) {
397      if (!radeon_get_drm_value(ws->fd, RADEON_INFO_NUM_GB_PIPES,
398                                "GB pipe count",
399                                &ws->info.r300_num_gb_pipes))
400         return false;
401
402      if (!radeon_get_drm_value(ws->fd, RADEON_INFO_NUM_Z_PIPES,
403                                "Z pipe count",
404                                &ws->info.r300_num_z_pipes))
405         return false;
406   }
407   else if (ws->gen >= DRV_R600) {
408      uint32_t tiling_config = 0;
409
410      if (!radeon_get_drm_value(ws->fd, RADEON_INFO_NUM_BACKENDS,
411                                "num backends",
412                                &ws->info.max_render_backends))
413         return false;
414
415      /* get the GPU counter frequency, failure is not fatal */
416      radeon_get_drm_value(ws->fd, RADEON_INFO_CLOCK_CRYSTAL_FREQ, NULL,
417                           &ws->info.clock_crystal_freq);
418
419      radeon_get_drm_value(ws->fd, RADEON_INFO_TILING_CONFIG, NULL,
420                           &tiling_config);
421
422      ws->info.r600_num_banks =
423            ws->info.chip_class >= EVERGREEN ?
424               4 << ((tiling_config & 0xf0) >> 4) :
425                    4 << ((tiling_config & 0x30) >> 4);
426
427      ws->info.pipe_interleave_bytes =
428            ws->info.chip_class >= EVERGREEN ?
429               256 << ((tiling_config & 0xf00) >> 8) :
430                      256 << ((tiling_config & 0xc0) >> 6);
431
432      if (!ws->info.pipe_interleave_bytes)
433         ws->info.pipe_interleave_bytes =
434               ws->info.chip_class >= EVERGREEN ? 512 : 256;
435
436      radeon_get_drm_value(ws->fd, RADEON_INFO_NUM_TILE_PIPES, NULL,
437                           &ws->info.num_tile_pipes);
438
439      /* "num_tiles_pipes" must be equal to the number of pipes (Px) in the
440       * pipe config field of the GB_TILE_MODE array. Only one card (Tahiti)
441       * reports a different value (12). Fix it by setting what's in the
442       * GB_TILE_MODE array (8).
443       */
444      if (ws->gen == DRV_SI && ws->info.num_tile_pipes == 12)
445         ws->info.num_tile_pipes = 8;
446
447      if (radeon_get_drm_value(ws->fd, RADEON_INFO_BACKEND_MAP, NULL,
448                               &ws->info.r600_gb_backend_map))
449         ws->info.r600_gb_backend_map_valid = true;
450
451      /* Default value. */
452      ws->info.enabled_rb_mask = u_bit_consecutive(0, ws->info.max_render_backends);
453      /*
454       * This fails (silently) on non-GCN or older kernels, overwriting the
455       * default enabled_rb_mask with the result of the last query.
456       */
457      if (ws->gen >= DRV_SI)
458         radeon_get_drm_value(ws->fd, RADEON_INFO_SI_BACKEND_ENABLED_MASK, NULL,
459                              &ws->info.enabled_rb_mask);
460
461      ws->info.r600_has_virtual_memory = false;
462      if (ws->info.drm_minor >= 13) {
463         uint32_t ib_vm_max_size;
464
465         ws->info.r600_has_virtual_memory = true;
466         if (!radeon_get_drm_value(ws->fd, RADEON_INFO_VA_START, NULL,
467                                   &ws->va_start))
468            ws->info.r600_has_virtual_memory = false;
469         if (!radeon_get_drm_value(ws->fd, RADEON_INFO_IB_VM_MAX_SIZE, NULL,
470                                   &ib_vm_max_size))
471            ws->info.r600_has_virtual_memory = false;
472         radeon_get_drm_value(ws->fd, RADEON_INFO_VA_UNMAP_WORKING, NULL,
473                              &ws->va_unmap_working);
474      }
475      if (ws->gen == DRV_R600 && !debug_get_bool_option("RADEON_VA", false))
476         ws->info.r600_has_virtual_memory = false;
477   }
478
479   /* Get max pipes, this is only needed for compute shaders.  All evergreen+
480    * chips have at least 2 pipes, so we use 2 as a default. */
481   ws->info.r600_max_quad_pipes = 2;
482   radeon_get_drm_value(ws->fd, RADEON_INFO_MAX_PIPES, NULL,
483                        &ws->info.r600_max_quad_pipes);
484
485   /* All GPUs have at least one compute unit */
486   ws->info.num_good_compute_units = 1;
487   radeon_get_drm_value(ws->fd, RADEON_INFO_ACTIVE_CU_COUNT, NULL,
488                        &ws->info.num_good_compute_units);
489
490   radeon_get_drm_value(ws->fd, RADEON_INFO_MAX_SE, NULL,
491                        &ws->info.max_se);
492
493   switch (ws->info.family) {
494   case CHIP_HAINAN:
495   case CHIP_KABINI:
496      ws->info.max_tcc_blocks = 2;
497      break;
498   case CHIP_VERDE:
499   case CHIP_OLAND:
500   case CHIP_BONAIRE:
501   case CHIP_KAVERI:
502      ws->info.max_tcc_blocks = 4;
503      break;
504   case CHIP_PITCAIRN:
505      ws->info.max_tcc_blocks = 8;
506      break;
507   case CHIP_TAHITI:
508      ws->info.max_tcc_blocks = 12;
509      break;
510   case CHIP_HAWAII:
511      ws->info.max_tcc_blocks = 16;
512      break;
513   default:
514      ws->info.max_tcc_blocks = 0;
515      break;
516   }
517
518   if (!ws->info.max_se) {
519      switch (ws->info.family) {
520      default:
521         ws->info.max_se = 1;
522         break;
523      case CHIP_CYPRESS:
524      case CHIP_HEMLOCK:
525      case CHIP_BARTS:
526      case CHIP_CAYMAN:
527      case CHIP_TAHITI:
528      case CHIP_PITCAIRN:
529      case CHIP_BONAIRE:
530         ws->info.max_se = 2;
531         break;
532      case CHIP_HAWAII:
533         ws->info.max_se = 4;
534         break;
535      }
536   }
537
538   ws->info.num_se = ws->info.max_se;
539
540   radeon_get_drm_value(ws->fd, RADEON_INFO_MAX_SH_PER_SE, NULL,
541                        &ws->info.max_sa_per_se);
542   if (ws->gen == DRV_SI) {
543      ws->info.max_good_cu_per_sa =
544      ws->info.min_good_cu_per_sa = ws->info.num_good_compute_units /
545                                    (ws->info.max_se * ws->info.max_sa_per_se);
546   }
547
548   radeon_get_drm_value(ws->fd, RADEON_INFO_ACCEL_WORKING2, NULL,
549                        &ws->accel_working2);
550   if (ws->info.family == CHIP_HAWAII && ws->accel_working2 < 2) {
551      fprintf(stderr, "radeon: GPU acceleration for Hawaii disabled, "
552                      "returned accel_working2 value %u is smaller than 2. "
553                      "Please install a newer kernel.\n",
554              ws->accel_working2);
555      return false;
556   }
557
558   if (ws->info.chip_class == GFX7) {
559      if (!radeon_get_drm_value(ws->fd, RADEON_INFO_CIK_MACROTILE_MODE_ARRAY, NULL,
560                                ws->info.cik_macrotile_mode_array)) {
561         fprintf(stderr, "radeon: Kernel 3.13 is required for Sea Islands support.\n");
562         return false;
563      }
564   }
565
566   if (ws->info.chip_class >= GFX6) {
567      if (!radeon_get_drm_value(ws->fd, RADEON_INFO_SI_TILE_MODE_ARRAY, NULL,
568                                ws->info.si_tile_mode_array)) {
569         fprintf(stderr, "radeon: Kernel 3.10 is required for Southern Islands support.\n");
570         return false;
571      }
572   }
573
574   /* Hawaii with old firmware needs type2 nop packet.
575    * accel_working2 with value 3 indicates the new firmware.
576    */
577   ws->info.gfx_ib_pad_with_type2 = ws->info.chip_class <= GFX6 ||
578                                    (ws->info.family == CHIP_HAWAII &&
579                                     ws->accel_working2 < 3);
580   ws->info.tcc_cache_line_size = 64; /* TC L2 line size on GCN */
581   ws->info.ib_alignment = 4096;
582   ws->info.kernel_flushes_hdp_before_ib = ws->info.drm_minor >= 40;
583   /* HTILE is broken with 1D tiling on old kernels and GFX7. */
584   ws->info.htile_cmask_support_1d_tiling = ws->info.chip_class != GFX7 ||
585                                                                   ws->info.drm_minor >= 38;
586   ws->info.si_TA_CS_BC_BASE_ADDR_allowed = ws->info.drm_minor >= 48;
587   ws->info.has_bo_metadata = false;
588   ws->info.has_gpu_reset_status_query = ws->info.drm_minor >= 43;
589   ws->info.has_eqaa_surface_allocator = false;
590   ws->info.has_format_bc1_through_bc7 = ws->info.drm_minor >= 31;
591   ws->info.kernel_flushes_tc_l2_after_ib = true;
592   /* Old kernels disallowed register writes via COPY_DATA
593    * that are used for indirect compute dispatches. */
594   ws->info.has_indirect_compute_dispatch = ws->info.chip_class == GFX7 ||
595                                            (ws->info.chip_class == GFX6 &&
596                                             ws->info.drm_minor >= 45);
597   /* GFX6 doesn't support unaligned loads. */
598   ws->info.has_unaligned_shader_loads = ws->info.chip_class == GFX7 &&
599                                         ws->info.drm_minor >= 50;
600   ws->info.has_sparse_vm_mappings = false;
601   /* 2D tiling on GFX7 is supported since DRM 2.35.0 */
602   ws->info.has_2d_tiling = ws->info.chip_class <= GFX6 || ws->info.drm_minor >= 35;
603   ws->info.has_read_registers_query = ws->info.drm_minor >= 42;
604   ws->info.max_alignment = 1024*1024;
605   ws->info.has_graphics = true;
606   ws->info.cpdma_prefetch_writes_memory = true;
607   ws->info.max_wave64_per_simd = 10;
608   ws->info.num_physical_sgprs_per_simd = 512;
609   ws->info.num_physical_wave64_vgprs_per_simd = 256;
610   ws->info.has_3d_cube_border_color_mipmap = true;
611
612   ws->check_vm = strstr(debug_get_option("R600_DEBUG", ""), "check_vm") != NULL ||
613                                                                            strstr(debug_get_option("AMD_DEBUG", ""), "check_vm") != NULL;
614   ws->noop_cs = debug_get_bool_option("RADEON_NOOP", false);
615
616   return true;
617}
618
619static void radeon_winsys_destroy(struct radeon_winsys *rws)
620{
621   struct radeon_drm_winsys *ws = (struct radeon_drm_winsys*)rws;
622
623   if (util_queue_is_initialized(&ws->cs_queue))
624      util_queue_destroy(&ws->cs_queue);
625
626   mtx_destroy(&ws->hyperz_owner_mutex);
627   mtx_destroy(&ws->cmask_owner_mutex);
628
629   if (ws->info.r600_has_virtual_memory)
630      pb_slabs_deinit(&ws->bo_slabs);
631   pb_cache_deinit(&ws->bo_cache);
632
633   if (ws->gen >= DRV_R600) {
634      radeon_surface_manager_free(ws->surf_man);
635   }
636
637   _mesa_hash_table_destroy(ws->bo_names, NULL);
638   _mesa_hash_table_destroy(ws->bo_handles, NULL);
639   _mesa_hash_table_u64_destroy(ws->bo_vas);
640   mtx_destroy(&ws->bo_handles_mutex);
641   mtx_destroy(&ws->vm32.mutex);
642   mtx_destroy(&ws->vm64.mutex);
643   mtx_destroy(&ws->bo_fence_lock);
644
645   if (ws->fd >= 0)
646      close(ws->fd);
647
648   FREE(rws);
649}
650
651static void radeon_query_info(struct radeon_winsys *rws,
652                              struct radeon_info *info,
653                              bool enable_smart_access_memory,
654                              bool disable_smart_access_memory)
655{
656   *info = ((struct radeon_drm_winsys *)rws)->info;
657}
658
659static bool radeon_cs_request_feature(struct radeon_cmdbuf *rcs,
660                                      enum radeon_feature_id fid,
661                                      bool enable)
662{
663   struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
664
665   switch (fid) {
666   case RADEON_FID_R300_HYPERZ_ACCESS:
667      return radeon_set_fd_access(cs, &cs->ws->hyperz_owner,
668                                  &cs->ws->hyperz_owner_mutex,
669                                  RADEON_INFO_WANT_HYPERZ, "Hyper-Z",
670                                  enable);
671
672   case RADEON_FID_R300_CMASK_ACCESS:
673      return radeon_set_fd_access(cs, &cs->ws->cmask_owner,
674                                  &cs->ws->cmask_owner_mutex,
675                                  RADEON_INFO_WANT_CMASK, "AA optimizations",
676                                  enable);
677   }
678   return false;
679}
680
681uint32_t radeon_drm_get_gpu_reset_counter(struct radeon_drm_winsys *ws)
682{
683   uint64_t retval = 0;
684
685   if (!ws->info.has_gpu_reset_status_query)
686      return 0;
687
688   radeon_get_drm_value(ws->fd, RADEON_INFO_GPU_RESET_COUNTER,
689                        "gpu-reset-counter", (uint32_t*)&retval);
690   return retval;
691}
692
693static uint64_t radeon_query_value(struct radeon_winsys *rws,
694                                   enum radeon_value_id value)
695{
696   struct radeon_drm_winsys *ws = (struct radeon_drm_winsys*)rws;
697   uint64_t retval = 0;
698
699   switch (value) {
700   case RADEON_REQUESTED_VRAM_MEMORY:
701      return ws->allocated_vram;
702   case RADEON_REQUESTED_GTT_MEMORY:
703      return ws->allocated_gtt;
704   case RADEON_MAPPED_VRAM:
705      return ws->mapped_vram;
706   case RADEON_MAPPED_GTT:
707      return ws->mapped_gtt;
708   case RADEON_BUFFER_WAIT_TIME_NS:
709      return ws->buffer_wait_time;
710   case RADEON_NUM_MAPPED_BUFFERS:
711      return ws->num_mapped_buffers;
712   case RADEON_TIMESTAMP:
713      if (ws->info.drm_minor < 20 || ws->gen < DRV_R600) {
714         assert(0);
715         return 0;
716      }
717
718      radeon_get_drm_value(ws->fd, RADEON_INFO_TIMESTAMP, "timestamp",
719                           (uint32_t*)&retval);
720      return retval;
721   case RADEON_NUM_GFX_IBS:
722      return ws->num_gfx_IBs;
723   case RADEON_NUM_SDMA_IBS:
724      return ws->num_sdma_IBs;
725   case RADEON_NUM_BYTES_MOVED:
726      radeon_get_drm_value(ws->fd, RADEON_INFO_NUM_BYTES_MOVED,
727                           "num-bytes-moved", (uint32_t*)&retval);
728      return retval;
729   case RADEON_NUM_EVICTIONS:
730   case RADEON_NUM_VRAM_CPU_PAGE_FAULTS:
731   case RADEON_VRAM_VIS_USAGE:
732   case RADEON_GFX_BO_LIST_COUNTER:
733   case RADEON_GFX_IB_SIZE_COUNTER:
734   case RADEON_SLAB_WASTED_VRAM:
735   case RADEON_SLAB_WASTED_GTT:
736      return 0; /* unimplemented */
737   case RADEON_VRAM_USAGE:
738      radeon_get_drm_value(ws->fd, RADEON_INFO_VRAM_USAGE,
739                           "vram-usage", (uint32_t*)&retval);
740      return retval;
741   case RADEON_GTT_USAGE:
742      radeon_get_drm_value(ws->fd, RADEON_INFO_GTT_USAGE,
743                           "gtt-usage", (uint32_t*)&retval);
744      return retval;
745   case RADEON_GPU_TEMPERATURE:
746      radeon_get_drm_value(ws->fd, RADEON_INFO_CURRENT_GPU_TEMP,
747                           "gpu-temp", (uint32_t*)&retval);
748      return retval;
749   case RADEON_CURRENT_SCLK:
750      radeon_get_drm_value(ws->fd, RADEON_INFO_CURRENT_GPU_SCLK,
751                           "current-gpu-sclk", (uint32_t*)&retval);
752      return retval;
753   case RADEON_CURRENT_MCLK:
754      radeon_get_drm_value(ws->fd, RADEON_INFO_CURRENT_GPU_MCLK,
755                           "current-gpu-mclk", (uint32_t*)&retval);
756      return retval;
757   case RADEON_CS_THREAD_TIME:
758      return util_queue_get_thread_time_nano(&ws->cs_queue, 0);
759   }
760   return 0;
761}
762
763static bool radeon_read_registers(struct radeon_winsys *rws,
764                                  unsigned reg_offset,
765                                  unsigned num_registers, uint32_t *out)
766{
767   struct radeon_drm_winsys *ws = (struct radeon_drm_winsys*)rws;
768   unsigned i;
769
770   for (i = 0; i < num_registers; i++) {
771      uint32_t reg = reg_offset + i*4;
772
773      if (!radeon_get_drm_value(ws->fd, RADEON_INFO_READ_REG, NULL, &reg))
774         return false;
775      out[i] = reg;
776   }
777   return true;
778}
779
780DEBUG_GET_ONCE_BOOL_OPTION(thread, "RADEON_THREAD", true)
781
782static bool radeon_winsys_unref(struct radeon_winsys *ws)
783{
784   struct radeon_drm_winsys *rws = (struct radeon_drm_winsys*)ws;
785   bool destroy;
786
787   /* When the reference counter drops to zero, remove the fd from the table.
788    * This must happen while the mutex is locked, so that
789    * radeon_drm_winsys_create in another thread doesn't get the winsys
790    * from the table when the counter drops to 0. */
791   mtx_lock(&fd_tab_mutex);
792
793   destroy = pipe_reference(&rws->reference, NULL);
794   if (destroy && fd_tab) {
795      _mesa_hash_table_remove_key(fd_tab, intptr_to_pointer(rws->fd));
796      if (_mesa_hash_table_num_entries(fd_tab) == 0) {
797         _mesa_hash_table_destroy(fd_tab, NULL);
798         fd_tab = NULL;
799      }
800   }
801
802   mtx_unlock(&fd_tab_mutex);
803   return destroy;
804}
805
806static void radeon_pin_threads_to_L3_cache(struct radeon_winsys *ws,
807                                           unsigned cache)
808{
809   struct radeon_drm_winsys *rws = (struct radeon_drm_winsys*)ws;
810
811   if (util_queue_is_initialized(&rws->cs_queue)) {
812      util_set_thread_affinity(rws->cs_queue.threads[0],
813                               util_get_cpu_caps()->L3_affinity_mask[cache],
814                               NULL, util_get_cpu_caps()->num_cpu_mask_bits);
815   }
816}
817
818static bool radeon_cs_is_secure(struct radeon_cmdbuf* cs)
819{
820    return false;
821}
822
823PUBLIC struct radeon_winsys *
824radeon_drm_winsys_create(int fd, const struct pipe_screen_config *config,
825                         radeon_screen_create_t screen_create)
826{
827   struct radeon_drm_winsys *ws;
828
829   mtx_lock(&fd_tab_mutex);
830   if (!fd_tab) {
831      fd_tab = util_hash_table_create_fd_keys();
832   }
833
834   ws = util_hash_table_get(fd_tab, intptr_to_pointer(fd));
835   if (ws) {
836      pipe_reference(NULL, &ws->reference);
837      mtx_unlock(&fd_tab_mutex);
838      return &ws->base;
839   }
840
841   ws = CALLOC_STRUCT(radeon_drm_winsys);
842   if (!ws) {
843      mtx_unlock(&fd_tab_mutex);
844      return NULL;
845   }
846
847   ws->fd = os_dupfd_cloexec(fd);
848
849   if (!do_winsys_init(ws))
850      goto fail1;
851
852   pb_cache_init(&ws->bo_cache, RADEON_MAX_CACHED_HEAPS,
853                 500000, ws->check_vm ? 1.0f : 2.0f, 0,
854                 MIN2(ws->info.vram_size, ws->info.gart_size), NULL,
855                 radeon_bo_destroy,
856                 radeon_bo_can_reclaim);
857
858   if (ws->info.r600_has_virtual_memory) {
859      /* There is no fundamental obstacle to using slab buffer allocation
860       * without GPUVM, but enabling it requires making sure that the drivers
861       * honor the address offset.
862       */
863      if (!pb_slabs_init(&ws->bo_slabs,
864                         RADEON_SLAB_MIN_SIZE_LOG2, RADEON_SLAB_MAX_SIZE_LOG2,
865                         RADEON_MAX_SLAB_HEAPS, false,
866                         ws,
867                         radeon_bo_can_reclaim_slab,
868                         radeon_bo_slab_alloc,
869                         radeon_bo_slab_free))
870         goto fail_cache;
871
872      ws->info.min_alloc_size = 1 << RADEON_SLAB_MIN_SIZE_LOG2;
873   } else {
874      ws->info.min_alloc_size = ws->info.gart_page_size;
875   }
876
877   if (ws->gen >= DRV_R600) {
878      ws->surf_man = radeon_surface_manager_new(ws->fd);
879      if (!ws->surf_man)
880         goto fail_slab;
881   }
882
883   /* init reference */
884   pipe_reference_init(&ws->reference, 1);
885
886   /* Set functions. */
887   ws->base.unref = radeon_winsys_unref;
888   ws->base.destroy = radeon_winsys_destroy;
889   ws->base.query_info = radeon_query_info;
890   ws->base.pin_threads_to_L3_cache = radeon_pin_threads_to_L3_cache;
891   ws->base.cs_request_feature = radeon_cs_request_feature;
892   ws->base.query_value = radeon_query_value;
893   ws->base.read_registers = radeon_read_registers;
894   ws->base.cs_is_secure = radeon_cs_is_secure;
895
896   radeon_drm_bo_init_functions(ws);
897   radeon_drm_cs_init_functions(ws);
898   radeon_surface_init_functions(ws);
899
900   (void) mtx_init(&ws->hyperz_owner_mutex, mtx_plain);
901   (void) mtx_init(&ws->cmask_owner_mutex, mtx_plain);
902
903   ws->bo_names = util_hash_table_create_ptr_keys();
904   ws->bo_handles = util_hash_table_create_ptr_keys();
905   ws->bo_vas = _mesa_hash_table_u64_create(NULL);
906   (void) mtx_init(&ws->bo_handles_mutex, mtx_plain);
907   (void) mtx_init(&ws->vm32.mutex, mtx_plain);
908   (void) mtx_init(&ws->vm64.mutex, mtx_plain);
909   (void) mtx_init(&ws->bo_fence_lock, mtx_plain);
910   list_inithead(&ws->vm32.holes);
911   list_inithead(&ws->vm64.holes);
912
913   /* The kernel currently returns 8MB. Make sure this doesn't change. */
914   if (ws->va_start > 8 * 1024 * 1024) {
915      /* Not enough 32-bit address space. */
916      radeon_winsys_destroy(&ws->base);
917      mtx_unlock(&fd_tab_mutex);
918      return NULL;
919   }
920
921   ws->vm32.start = ws->va_start;
922   ws->vm32.end = 1ull << 32;
923
924   /* The maximum is 8GB of virtual address space limited by the kernel.
925    * It's obviously not enough for bigger cards, like Hawaiis with 4GB
926    * and 8GB of physical memory and 4GB of GART.
927    *
928    * Older kernels set the limit to 4GB, which is even worse, so they only
929    * have 32-bit address space.
930    */
931   if (ws->info.drm_minor >= 41) {
932      ws->vm64.start = 1ull << 32;
933      ws->vm64.end = 1ull << 33;
934   }
935
936   /* TTM aligns the BO size to the CPU page size */
937   ws->info.gart_page_size = sysconf(_SC_PAGESIZE);
938   ws->info.pte_fragment_size = 64 * 1024; /* GPUVM page size */
939
940   if (ws->num_cpus > 1 && debug_get_option_thread())
941      util_queue_init(&ws->cs_queue, "rcs", 8, 1, 0, NULL);
942
943   /* Create the screen at the end. The winsys must be initialized
944    * completely.
945    *
946    * Alternatively, we could create the screen based on "ws->gen"
947    * and link all drivers into one binary blob. */
948   ws->base.screen = screen_create(&ws->base, config);
949   if (!ws->base.screen) {
950      radeon_winsys_destroy(&ws->base);
951      mtx_unlock(&fd_tab_mutex);
952      return NULL;
953   }
954
955   _mesa_hash_table_insert(fd_tab, intptr_to_pointer(ws->fd), ws);
956
957   /* We must unlock the mutex once the winsys is fully initialized, so that
958    * other threads attempting to create the winsys from the same fd will
959    * get a fully initialized winsys and not just half-way initialized. */
960   mtx_unlock(&fd_tab_mutex);
961
962   return &ws->base;
963
964fail_slab:
965   if (ws->info.r600_has_virtual_memory)
966      pb_slabs_deinit(&ws->bo_slabs);
967fail_cache:
968   pb_cache_deinit(&ws->bo_cache);
969fail1:
970   mtx_unlock(&fd_tab_mutex);
971   if (ws->surf_man)
972      radeon_surface_manager_free(ws->surf_man);
973   if (ws->fd >= 0)
974      close(ws->fd);
975
976   FREE(ws);
977   return NULL;
978}
979