1/*
2 * Copyright © 2014-2017 Broadcom
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 (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24/**
25 * @file v3d_simulator.c
26 *
27 * Implements V3D simulation on top of a non-V3D GEM fd.
28 *
29 * This file's goal is to emulate the V3D ioctls' behavior in the kernel on
30 * top of the simpenrose software simulator.  Generally, V3D driver BOs have a
31 * GEM-side copy of their contents and a simulator-side memory area that the
32 * GEM contents get copied into during simulation.  Once simulation is done,
33 * the simulator's data is copied back out to the GEM BOs, so that rendering
34 * appears on the screen as if actual hardware rendering had been done.
35 *
36 * One of the limitations of this code is that we shouldn't really need a
37 * GEM-side BO for non-window-system BOs.  However, do we need unique BO
38 * handles for each of our GEM bos so that this file can look up its state
39 * from the handle passed in at submit ioctl time (also, a couple of places
40 * outside of this file still call ioctls directly on the fd).
41 *
42 * Another limitation is that BO import doesn't work unless the underlying
43 * window system's BO size matches what V3D is going to use, which of course
44 * doesn't work out in practice.  This means that for now, only DRI3 (V3D
45 * makes the winsys BOs) is supported, not DRI2 (window system makes the winys
46 * BOs).
47 */
48
49#ifdef USE_V3D_SIMULATOR
50
51#include <stdio.h>
52#include <sys/mman.h>
53#include "c11/threads.h"
54#include "util/hash_table.h"
55#include "util/ralloc.h"
56#include "util/set.h"
57#include "util/u_dynarray.h"
58#include "util/u_memory.h"
59#include "util/u_mm.h"
60#include "util/u_math.h"
61
62#include <xf86drm.h>
63#include "drm-uapi/amdgpu_drm.h"
64#include "drm-uapi/i915_drm.h"
65#include "drm-uapi/v3d_drm.h"
66
67#include "v3d_simulator.h"
68#include "v3d_simulator_wrapper.h"
69
70/** Global (across GEM fds) state for the simulator */
71static struct v3d_simulator_state {
72        mtx_t mutex;
73        mtx_t submit_lock;
74
75        struct v3d_hw *v3d;
76        int ver;
77
78        /* Base virtual address of the heap. */
79        void *mem;
80        /* Base hardware address of the heap. */
81        uint32_t mem_base;
82        /* Size of the heap. */
83        uint32_t mem_size;
84
85        struct mem_block *heap;
86        struct mem_block *overflow;
87
88        /** Mapping from GEM fd to struct v3d_simulator_file * */
89        struct hash_table *fd_map;
90
91        /** Last performance monitor ID. */
92        uint32_t last_perfid;
93
94        struct util_dynarray bin_oom;
95        int refcount;
96} sim_state = {
97        .mutex = _MTX_INITIALIZER_NP,
98};
99
100enum gem_type {
101        GEM_I915,
102        GEM_AMDGPU,
103        GEM_DUMB
104};
105
106/** Per-GEM-fd state for the simulator. */
107struct v3d_simulator_file {
108        int fd;
109
110        /** Mapping from GEM handle to struct v3d_simulator_bo * */
111        struct hash_table *bo_map;
112
113        /** Dynamic array with performance monitors */
114        struct v3d_simulator_perfmon **perfmons;
115        uint32_t perfmons_size;
116        uint32_t active_perfid;
117
118        struct mem_block *gmp;
119        void *gmp_vaddr;
120
121        /** For specific gpus, use their create ioctl. Otherwise use dumb bo. */
122        enum gem_type gem_type;
123};
124
125/** Wrapper for drm_v3d_bo tracking the simulator-specific state. */
126struct v3d_simulator_bo {
127        struct v3d_simulator_file *file;
128
129        /** Area for this BO within sim_state->mem */
130        struct mem_block *block;
131        uint32_t size;
132        uint64_t mmap_offset;
133        void *sim_vaddr;
134        void *gem_vaddr;
135
136        int handle;
137};
138
139struct v3d_simulator_perfmon {
140        uint32_t ncounters;
141        uint8_t counters[DRM_V3D_MAX_PERF_COUNTERS];
142        uint64_t values[DRM_V3D_MAX_PERF_COUNTERS];
143};
144
145static void *
146int_to_key(int key)
147{
148        return (void *)(uintptr_t)key;
149}
150
151#define PERFMONS_ALLOC_SIZE 100
152
153static uint32_t
154perfmons_next_id(struct v3d_simulator_file *sim_file) {
155        sim_state.last_perfid++;
156        if (sim_state.last_perfid > sim_file->perfmons_size) {
157                sim_file->perfmons_size += PERFMONS_ALLOC_SIZE;
158                sim_file->perfmons = reralloc(sim_file,
159                                              sim_file->perfmons,
160                                              struct v3d_simulator_perfmon *,
161                                              sim_file->perfmons_size);
162        }
163
164        return sim_state.last_perfid;
165}
166
167static struct v3d_simulator_file *
168v3d_get_simulator_file_for_fd(int fd)
169{
170        struct hash_entry *entry = _mesa_hash_table_search(sim_state.fd_map,
171                                                           int_to_key(fd + 1));
172        return entry ? entry->data : NULL;
173}
174
175/* A marker placed just after each BO, then checked after rendering to make
176 * sure it's still there.
177 */
178#define BO_SENTINEL		0xfedcba98
179
180/* 128kb */
181#define GMP_ALIGN2		17
182
183/**
184 * Sets the range of GPU virtual address space to have the given GMP
185 * permissions (bit 0 = read, bit 1 = write, write-only forbidden).
186 */
187static void
188set_gmp_flags(struct v3d_simulator_file *file,
189              uint32_t offset, uint32_t size, uint32_t flag)
190{
191        assert((offset & ((1 << GMP_ALIGN2) - 1)) == 0);
192        int gmp_offset = offset >> GMP_ALIGN2;
193        int gmp_count = align(size, 1 << GMP_ALIGN2) >> GMP_ALIGN2;
194        uint32_t *gmp = file->gmp_vaddr;
195
196        assert(flag <= 0x3);
197
198        for (int i = gmp_offset; i < gmp_offset + gmp_count; i++) {
199                int32_t bitshift = (i % 16) * 2;
200                gmp[i / 16] &= ~(0x3 << bitshift);
201                gmp[i / 16] |= flag << bitshift;
202        }
203}
204
205/**
206 * Allocates space in simulator memory and returns a tracking struct for it
207 * that also contains the drm_gem_cma_object struct.
208 */
209static struct v3d_simulator_bo *
210v3d_create_simulator_bo(int fd, unsigned size)
211{
212        struct v3d_simulator_file *file = v3d_get_simulator_file_for_fd(fd);
213        struct v3d_simulator_bo *sim_bo = rzalloc(file,
214                                                  struct v3d_simulator_bo);
215        size = align(size, 4096);
216
217        sim_bo->file = file;
218
219        mtx_lock(&sim_state.mutex);
220        sim_bo->block = u_mmAllocMem(sim_state.heap, size + 4, GMP_ALIGN2, 0);
221        mtx_unlock(&sim_state.mutex);
222        assert(sim_bo->block);
223
224        set_gmp_flags(file, sim_bo->block->ofs, size, 0x3);
225
226        sim_bo->size = size;
227
228        /* Allocate space for the buffer in simulator memory. */
229        sim_bo->sim_vaddr = sim_state.mem + sim_bo->block->ofs - sim_state.mem_base;
230        memset(sim_bo->sim_vaddr, 0xd0, size);
231
232        *(uint32_t *)(sim_bo->sim_vaddr + sim_bo->size) = BO_SENTINEL;
233
234        return sim_bo;
235}
236
237static struct v3d_simulator_bo *
238v3d_create_simulator_bo_for_gem(int fd, int handle, unsigned size)
239{
240        struct v3d_simulator_file *file = v3d_get_simulator_file_for_fd(fd);
241        struct v3d_simulator_bo *sim_bo =
242                v3d_create_simulator_bo(fd, size);
243
244        sim_bo->handle = handle;
245
246        /* Map the GEM buffer for copy in/out to the simulator.  i915 blocks
247         * dumb mmap on render nodes, so use their ioctl directly if we're on
248         * one.
249         */
250        int ret;
251        switch (file->gem_type) {
252        case GEM_I915:
253        {
254                struct drm_i915_gem_mmap_gtt map = {
255                        .handle = handle,
256                };
257
258                /* We could potentially use non-gtt (cached) for LLC systems,
259                 * but the copy-in/out won't be the limiting factor on
260                 * simulation anyway.
261                 */
262                ret = drmIoctl(fd, DRM_IOCTL_I915_GEM_MMAP_GTT, &map);
263                sim_bo->mmap_offset = map.offset;
264                break;
265        }
266        case GEM_AMDGPU:
267        {
268                union drm_amdgpu_gem_mmap map = { 0 };
269                map.in.handle = handle;
270
271                ret = drmIoctl(fd, DRM_IOCTL_AMDGPU_GEM_MMAP, &map);
272                sim_bo->mmap_offset = map.out.addr_ptr;
273                break;
274        }
275        default:
276        {
277                struct drm_mode_map_dumb map = {
278                        .handle = handle,
279                };
280                ret = drmIoctl(fd, DRM_IOCTL_MODE_MAP_DUMB, &map);
281                sim_bo->mmap_offset = map.offset;
282        }
283        }
284        if (ret) {
285                fprintf(stderr, "Failed to get MMAP offset: %d\n", ret);
286                abort();
287        }
288
289        sim_bo->gem_vaddr = mmap(NULL, sim_bo->size,
290                                 PROT_READ | PROT_WRITE, MAP_SHARED,
291                                 fd, sim_bo->mmap_offset);
292        if (sim_bo->gem_vaddr == MAP_FAILED) {
293                fprintf(stderr, "mmap of bo %d (offset 0x%016llx, size %d) failed\n",
294                        handle, (long long)sim_bo->mmap_offset, sim_bo->size);
295                abort();
296        }
297
298        /* A handle of 0 is used for v3d_gem.c internal allocations that
299         * don't need to go in the lookup table.
300         */
301        if (handle != 0) {
302                mtx_lock(&sim_state.mutex);
303                _mesa_hash_table_insert(file->bo_map, int_to_key(handle),
304                                        sim_bo);
305                mtx_unlock(&sim_state.mutex);
306        }
307
308        return sim_bo;
309}
310
311static int bin_fd;
312
313uint32_t
314v3d_simulator_get_spill(uint32_t spill_size)
315{
316        struct v3d_simulator_bo *sim_bo =
317                v3d_create_simulator_bo(bin_fd, spill_size);
318
319        util_dynarray_append(&sim_state.bin_oom, struct v3d_simulator_bo *,
320                             sim_bo);
321
322        return sim_bo->block->ofs;
323}
324
325static void
326v3d_free_simulator_bo(struct v3d_simulator_bo *sim_bo)
327{
328        struct v3d_simulator_file *sim_file = sim_bo->file;
329
330        set_gmp_flags(sim_file, sim_bo->block->ofs, sim_bo->size, 0x0);
331
332        if (sim_bo->gem_vaddr)
333                munmap(sim_bo->gem_vaddr, sim_bo->size);
334
335        mtx_lock(&sim_state.mutex);
336        u_mmFreeMem(sim_bo->block);
337        if (sim_bo->handle) {
338                _mesa_hash_table_remove_key(sim_file->bo_map,
339                                            int_to_key(sim_bo->handle));
340        }
341        mtx_unlock(&sim_state.mutex);
342        ralloc_free(sim_bo);
343}
344
345static struct v3d_simulator_bo *
346v3d_get_simulator_bo(struct v3d_simulator_file *file, int gem_handle)
347{
348        if (gem_handle == 0)
349                return NULL;
350
351        mtx_lock(&sim_state.mutex);
352        struct hash_entry *entry =
353                _mesa_hash_table_search(file->bo_map, int_to_key(gem_handle));
354        mtx_unlock(&sim_state.mutex);
355
356        return entry ? entry->data : NULL;
357}
358
359static void
360v3d_simulator_copy_in_handle(struct v3d_simulator_file *file, int handle)
361{
362        struct v3d_simulator_bo *sim_bo = v3d_get_simulator_bo(file, handle);
363
364        if (!sim_bo)
365                return;
366
367        memcpy(sim_bo->sim_vaddr, sim_bo->gem_vaddr, sim_bo->size);
368}
369
370static void
371v3d_simulator_copy_out_handle(struct v3d_simulator_file *file, int handle)
372{
373        struct v3d_simulator_bo *sim_bo = v3d_get_simulator_bo(file, handle);
374
375        if (!sim_bo)
376                return;
377
378        memcpy(sim_bo->gem_vaddr, sim_bo->sim_vaddr, sim_bo->size);
379
380        if (*(uint32_t *)(sim_bo->sim_vaddr +
381                          sim_bo->size) != BO_SENTINEL) {
382                fprintf(stderr, "Buffer overflow in handle %d\n",
383                        handle);
384        }
385}
386
387static int
388v3d_simulator_pin_bos(struct v3d_simulator_file *file,
389                      struct drm_v3d_submit_cl *submit)
390{
391        uint32_t *bo_handles = (uint32_t *)(uintptr_t)submit->bo_handles;
392
393        for (int i = 0; i < submit->bo_handle_count; i++)
394                v3d_simulator_copy_in_handle(file, bo_handles[i]);
395
396        return 0;
397}
398
399static int
400v3d_simulator_unpin_bos(struct v3d_simulator_file *file,
401                        struct drm_v3d_submit_cl *submit)
402{
403        uint32_t *bo_handles = (uint32_t *)(uintptr_t)submit->bo_handles;
404
405        for (int i = 0; i < submit->bo_handle_count; i++)
406                v3d_simulator_copy_out_handle(file, bo_handles[i]);
407
408        return 0;
409}
410
411static struct v3d_simulator_perfmon *
412v3d_get_simulator_perfmon(int fd, uint32_t perfid)
413{
414        if (!perfid || perfid > sim_state.last_perfid)
415                return NULL;
416
417        struct v3d_simulator_file *file = v3d_get_simulator_file_for_fd(fd);
418
419        mtx_lock(&sim_state.mutex);
420        assert(perfid <= file->perfmons_size);
421        struct v3d_simulator_perfmon *perfmon = file->perfmons[perfid - 1];
422        mtx_unlock(&sim_state.mutex);
423
424        return perfmon;
425}
426
427static void
428v3d_simulator_perfmon_switch(int fd, uint32_t perfid)
429{
430        struct v3d_simulator_file *file = v3d_get_simulator_file_for_fd(fd);
431        struct v3d_simulator_perfmon *perfmon;
432
433        if (perfid == file->active_perfid)
434                return;
435
436        perfmon = v3d_get_simulator_perfmon(fd, file->active_perfid);
437        if (perfmon)
438                v3d41_simulator_perfmon_stop(sim_state.v3d,
439                                             perfmon->ncounters,
440                                             perfmon->values);
441
442        perfmon = v3d_get_simulator_perfmon(fd, perfid);
443        if (perfmon)
444                v3d41_simulator_perfmon_start(sim_state.v3d,
445                                              perfmon->ncounters,
446                                              perfmon->counters);
447
448        file->active_perfid = perfid;
449}
450
451static int
452v3d_simulator_signal_syncobjs(int fd, struct drm_v3d_multi_sync *ms)
453{
454        struct drm_v3d_sem *out_syncs = (void *)(uintptr_t)ms->out_syncs;
455        int n_syncobjs = ms->out_sync_count;
456        uint32_t syncobjs[n_syncobjs];
457
458        for (int i = 0; i < n_syncobjs; i++)
459                syncobjs[i] = out_syncs[i].handle;
460        return drmSyncobjSignal(fd, (uint32_t *) &syncobjs, n_syncobjs);
461}
462
463static int
464v3d_simulator_process_post_deps(int fd, struct drm_v3d_extension *ext)
465{
466        int ret = 0;
467        while (ext && ext->id != DRM_V3D_EXT_ID_MULTI_SYNC)
468                ext = (void *)(uintptr_t) ext->next;
469
470        if (ext) {
471                struct drm_v3d_multi_sync *ms = (struct drm_v3d_multi_sync *) ext;
472                ret = v3d_simulator_signal_syncobjs(fd, ms);
473        }
474        return ret;
475}
476
477static int
478v3d_simulator_submit_cl_ioctl(int fd, struct drm_v3d_submit_cl *submit)
479{
480        struct v3d_simulator_file *file = v3d_get_simulator_file_for_fd(fd);
481        int ret;
482
483        ret = v3d_simulator_pin_bos(file, submit);
484        if (ret)
485                return ret;
486
487        mtx_lock(&sim_state.submit_lock);
488        bin_fd = fd;
489
490        v3d_simulator_perfmon_switch(fd, submit->perfmon_id);
491
492        if (sim_state.ver >= 41)
493                v3d41_simulator_submit_cl_ioctl(sim_state.v3d, submit, file->gmp->ofs);
494        else
495                v3d33_simulator_submit_cl_ioctl(sim_state.v3d, submit, file->gmp->ofs);
496
497        util_dynarray_foreach(&sim_state.bin_oom, struct v3d_simulator_bo *,
498                              sim_bo) {
499                v3d_free_simulator_bo(*sim_bo);
500        }
501        util_dynarray_clear(&sim_state.bin_oom);
502
503        mtx_unlock(&sim_state.submit_lock);
504
505        ret = v3d_simulator_unpin_bos(file, submit);
506        if (ret)
507                return ret;
508
509        if (submit->flags & DRM_V3D_SUBMIT_EXTENSION) {
510                struct drm_v3d_extension *ext = (void *)(uintptr_t)submit->extensions;
511                ret = v3d_simulator_process_post_deps(fd, ext);
512        }
513
514        return ret;
515}
516
517/**
518 * Do fixups after a BO has been opened from a handle.
519 *
520 * This could be done at DRM_IOCTL_GEM_OPEN/DRM_IOCTL_GEM_PRIME_FD_TO_HANDLE
521 * time, but we're still using drmPrimeFDToHandle() so we have this helper to
522 * be called afterward instead.
523 */
524void v3d_simulator_open_from_handle(int fd, int handle, uint32_t size)
525{
526        v3d_create_simulator_bo_for_gem(fd, handle, size);
527}
528
529/**
530 * Simulated ioctl(fd, DRM_V3D_CREATE_BO) implementation.
531 *
532 * Making a V3D BO is just a matter of making a corresponding BO on the host.
533 */
534static int
535v3d_simulator_create_bo_ioctl(int fd, struct drm_v3d_create_bo *args)
536{
537        struct v3d_simulator_file *file = v3d_get_simulator_file_for_fd(fd);
538
539        /* i915 bans dumb create on render nodes, so we have to use their
540         * native ioctl in case we're on a render node.
541         */
542        int ret;
543        switch (file->gem_type) {
544        case GEM_I915:
545        {
546                struct drm_i915_gem_create create = {
547                        .size = args->size,
548                };
549
550                ret = drmIoctl(fd, DRM_IOCTL_I915_GEM_CREATE, &create);
551
552                args->handle = create.handle;
553                break;
554        }
555        case GEM_AMDGPU:
556        {
557                union drm_amdgpu_gem_create create = { 0 };
558                create.in.bo_size = args->size;
559
560                ret = drmIoctl(fd, DRM_IOCTL_AMDGPU_GEM_CREATE, &create);
561
562                args->handle = create.out.handle;
563                break;
564        }
565        default:
566        {
567                struct drm_mode_create_dumb create = {
568                        .width = 128,
569                        .bpp = 8,
570                        .height = (args->size + 127) / 128,
571                };
572
573                ret = drmIoctl(fd, DRM_IOCTL_MODE_CREATE_DUMB, &create);
574                assert(ret != 0 || create.size >= args->size);
575
576                args->handle = create.handle;
577        }
578        }
579        if (ret == 0) {
580                struct v3d_simulator_bo *sim_bo =
581                        v3d_create_simulator_bo_for_gem(fd, args->handle,
582                                                        args->size);
583
584                args->offset = sim_bo->block->ofs;
585        }
586
587        return ret;
588}
589
590/**
591 * Simulated ioctl(fd, DRM_V3D_MMAP_BO) implementation.
592 *
593 * We've already grabbed the mmap offset when we created the sim bo, so just
594 * return it.
595 */
596static int
597v3d_simulator_mmap_bo_ioctl(int fd, struct drm_v3d_mmap_bo *args)
598{
599        struct v3d_simulator_file *file = v3d_get_simulator_file_for_fd(fd);
600        struct v3d_simulator_bo *sim_bo = v3d_get_simulator_bo(file,
601                                                               args->handle);
602
603        args->offset = sim_bo->mmap_offset;
604
605        return 0;
606}
607
608static int
609v3d_simulator_get_bo_offset_ioctl(int fd, struct drm_v3d_get_bo_offset *args)
610{
611        struct v3d_simulator_file *file = v3d_get_simulator_file_for_fd(fd);
612        struct v3d_simulator_bo *sim_bo = v3d_get_simulator_bo(file,
613                                                               args->handle);
614
615        args->offset = sim_bo->block->ofs;
616
617        return 0;
618}
619
620static int
621v3d_simulator_gem_close_ioctl(int fd, struct drm_gem_close *args)
622{
623        /* Free the simulator's internal tracking. */
624        struct v3d_simulator_file *file = v3d_get_simulator_file_for_fd(fd);
625        struct v3d_simulator_bo *sim_bo = v3d_get_simulator_bo(file,
626                                                               args->handle);
627
628        v3d_free_simulator_bo(sim_bo);
629
630        /* Pass the call on down. */
631        return drmIoctl(fd, DRM_IOCTL_GEM_CLOSE, args);
632}
633
634static int
635v3d_simulator_get_param_ioctl(int fd, struct drm_v3d_get_param *args)
636{
637        if (sim_state.ver >= 41)
638                return v3d41_simulator_get_param_ioctl(sim_state.v3d, args);
639        else
640                return v3d33_simulator_get_param_ioctl(sim_state.v3d, args);
641}
642
643static int
644v3d_simulator_submit_tfu_ioctl(int fd, struct drm_v3d_submit_tfu *args)
645{
646        struct v3d_simulator_file *file = v3d_get_simulator_file_for_fd(fd);
647        int ret;
648
649        v3d_simulator_copy_in_handle(file, args->bo_handles[0]);
650        v3d_simulator_copy_in_handle(file, args->bo_handles[1]);
651        v3d_simulator_copy_in_handle(file, args->bo_handles[2]);
652        v3d_simulator_copy_in_handle(file, args->bo_handles[3]);
653
654        if (sim_state.ver >= 41)
655                ret = v3d41_simulator_submit_tfu_ioctl(sim_state.v3d, args);
656        else
657                ret = v3d33_simulator_submit_tfu_ioctl(sim_state.v3d, args);
658
659        v3d_simulator_copy_out_handle(file, args->bo_handles[0]);
660
661        if (ret)
662                return ret;
663
664        if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
665                struct drm_v3d_extension *ext = (void *)(uintptr_t)args->extensions;
666                ret = v3d_simulator_process_post_deps(fd, ext);
667        }
668
669        return ret;
670}
671
672static int
673v3d_simulator_submit_csd_ioctl(int fd, struct drm_v3d_submit_csd *args)
674{
675        struct v3d_simulator_file *file = v3d_get_simulator_file_for_fd(fd);
676        uint32_t *bo_handles = (uint32_t *)(uintptr_t)args->bo_handles;
677        int ret;
678
679        for (int i = 0; i < args->bo_handle_count; i++)
680                v3d_simulator_copy_in_handle(file, bo_handles[i]);
681
682        v3d_simulator_perfmon_switch(fd, args->perfmon_id);
683
684        if (sim_state.ver >= 41)
685                ret = v3d41_simulator_submit_csd_ioctl(sim_state.v3d, args,
686                                                       file->gmp->ofs);
687        else
688                ret = -1;
689
690        for (int i = 0; i < args->bo_handle_count; i++)
691                v3d_simulator_copy_out_handle(file, bo_handles[i]);
692
693        if (ret < 0)
694                return ret;
695
696        if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
697                struct drm_v3d_extension *ext = (void *)(uintptr_t)args->extensions;
698                ret = v3d_simulator_process_post_deps(fd, ext);
699        }
700
701        return ret;
702}
703
704static int
705v3d_simulator_perfmon_create_ioctl(int fd, struct drm_v3d_perfmon_create *args)
706{
707        struct v3d_simulator_file *file = v3d_get_simulator_file_for_fd(fd);
708
709        if (args->ncounters == 0 ||
710            args->ncounters > DRM_V3D_MAX_PERF_COUNTERS)
711                return -EINVAL;
712
713        struct v3d_simulator_perfmon *perfmon = rzalloc(file,
714                                                        struct v3d_simulator_perfmon);
715
716        perfmon->ncounters = args->ncounters;
717        for (int i = 0; i < args->ncounters; i++) {
718                if (args->counters[i] >= V3D_PERFCNT_NUM) {
719                        ralloc_free(perfmon);
720                        return -EINVAL;
721                } else {
722                        perfmon->counters[i] = args->counters[i];
723                }
724        }
725
726        mtx_lock(&sim_state.mutex);
727        args->id = perfmons_next_id(file);
728        file->perfmons[args->id - 1] = perfmon;
729        mtx_unlock(&sim_state.mutex);
730
731        return 0;
732}
733
734static int
735v3d_simulator_perfmon_destroy_ioctl(int fd, struct drm_v3d_perfmon_destroy *args)
736{
737        struct v3d_simulator_file *file = v3d_get_simulator_file_for_fd(fd);
738        struct v3d_simulator_perfmon *perfmon =
739                v3d_get_simulator_perfmon(fd, args->id);
740
741        if (!perfmon)
742                return -EINVAL;
743
744        mtx_lock(&sim_state.mutex);
745        file->perfmons[args->id - 1] = NULL;
746        mtx_unlock(&sim_state.mutex);
747
748        ralloc_free(perfmon);
749
750        return 0;
751}
752
753static int
754v3d_simulator_perfmon_get_values_ioctl(int fd, struct drm_v3d_perfmon_get_values *args)
755{
756        struct v3d_simulator_file *file = v3d_get_simulator_file_for_fd(fd);
757
758        mtx_lock(&sim_state.submit_lock);
759
760        /* Stop the perfmon if it is still active */
761        if (args->id == file->active_perfid)
762                v3d_simulator_perfmon_switch(fd, 0);
763
764        mtx_unlock(&sim_state.submit_lock);
765
766        struct v3d_simulator_perfmon *perfmon =
767                v3d_get_simulator_perfmon(fd, args->id);
768
769        if (!perfmon)
770                return -EINVAL;
771
772        memcpy((void *)args->values_ptr, perfmon->values, perfmon->ncounters * sizeof(uint64_t));
773
774        return 0;
775}
776
777int
778v3d_simulator_ioctl(int fd, unsigned long request, void *args)
779{
780        switch (request) {
781        case DRM_IOCTL_V3D_SUBMIT_CL:
782                return v3d_simulator_submit_cl_ioctl(fd, args);
783        case DRM_IOCTL_V3D_CREATE_BO:
784                return v3d_simulator_create_bo_ioctl(fd, args);
785        case DRM_IOCTL_V3D_MMAP_BO:
786                return v3d_simulator_mmap_bo_ioctl(fd, args);
787        case DRM_IOCTL_V3D_GET_BO_OFFSET:
788                return v3d_simulator_get_bo_offset_ioctl(fd, args);
789
790        case DRM_IOCTL_V3D_WAIT_BO:
791                /* We do all of the v3d rendering synchronously, so we just
792                 * return immediately on the wait ioctls.  This ignores any
793                 * native rendering to the host BO, so it does mean we race on
794                 * front buffer rendering.
795                 */
796                return 0;
797
798        case DRM_IOCTL_V3D_GET_PARAM:
799                return v3d_simulator_get_param_ioctl(fd, args);
800
801        case DRM_IOCTL_GEM_CLOSE:
802                return v3d_simulator_gem_close_ioctl(fd, args);
803
804        case DRM_IOCTL_V3D_SUBMIT_TFU:
805                return v3d_simulator_submit_tfu_ioctl(fd, args);
806
807        case DRM_IOCTL_V3D_SUBMIT_CSD:
808                return v3d_simulator_submit_csd_ioctl(fd, args);
809
810        case DRM_IOCTL_V3D_PERFMON_CREATE:
811                return v3d_simulator_perfmon_create_ioctl(fd, args);
812
813        case DRM_IOCTL_V3D_PERFMON_DESTROY:
814                return v3d_simulator_perfmon_destroy_ioctl(fd, args);
815
816        case DRM_IOCTL_V3D_PERFMON_GET_VALUES:
817                return v3d_simulator_perfmon_get_values_ioctl(fd, args);
818
819        case DRM_IOCTL_GEM_OPEN:
820        case DRM_IOCTL_GEM_FLINK:
821                return drmIoctl(fd, request, args);
822        default:
823                fprintf(stderr, "Unknown ioctl 0x%08x\n", (int)request);
824                abort();
825        }
826}
827
828uint32_t
829v3d_simulator_get_mem_size(void)
830{
831   return sim_state.mem_size;
832}
833
834static void
835v3d_simulator_init_global()
836{
837        mtx_lock(&sim_state.mutex);
838        if (sim_state.refcount++) {
839                mtx_unlock(&sim_state.mutex);
840                return;
841        }
842
843        sim_state.v3d = v3d_hw_auto_new(NULL);
844        v3d_hw_alloc_mem(sim_state.v3d, 1024 * 1024 * 1024);
845        sim_state.mem_base =
846                v3d_hw_get_mem(sim_state.v3d, &sim_state.mem_size,
847                               &sim_state.mem);
848
849        /* Allocate from anywhere from 4096 up.  We don't allocate at 0,
850         * because for OQs and some other addresses in the HW, 0 means
851         * disabled.
852         */
853        sim_state.heap = u_mmInit(4096, sim_state.mem_size - 4096);
854
855        /* Make a block of 0xd0 at address 0 to make sure we don't screw up
856         * and land there.
857         */
858        struct mem_block *b = u_mmAllocMem(sim_state.heap, 4096, GMP_ALIGN2, 0);
859        memset(sim_state.mem + b->ofs - sim_state.mem_base, 0xd0, 4096);
860
861        sim_state.ver = v3d_hw_get_version(sim_state.v3d);
862
863        mtx_unlock(&sim_state.mutex);
864
865        sim_state.fd_map =
866                _mesa_hash_table_create(NULL,
867                                        _mesa_hash_pointer,
868                                        _mesa_key_pointer_equal);
869
870        util_dynarray_init(&sim_state.bin_oom, NULL);
871
872        if (sim_state.ver >= 41)
873                v3d41_simulator_init_regs(sim_state.v3d);
874        else
875                v3d33_simulator_init_regs(sim_state.v3d);
876}
877
878struct v3d_simulator_file *
879v3d_simulator_init(int fd)
880{
881        v3d_simulator_init_global();
882
883        struct v3d_simulator_file *sim_file = rzalloc(NULL, struct v3d_simulator_file);
884
885        drmVersionPtr version = drmGetVersion(fd);
886        if (version && strncmp(version->name, "i915", version->name_len) == 0)
887                sim_file->gem_type = GEM_I915;
888        else if (version && strncmp(version->name, "amdgpu", version->name_len) == 0)
889                sim_file->gem_type = GEM_AMDGPU;
890        else
891                sim_file->gem_type = GEM_DUMB;
892        drmFreeVersion(version);
893
894        sim_file->bo_map =
895                _mesa_hash_table_create(sim_file,
896                                        _mesa_hash_pointer,
897                                        _mesa_key_pointer_equal);
898
899        mtx_lock(&sim_state.mutex);
900        _mesa_hash_table_insert(sim_state.fd_map, int_to_key(fd + 1),
901                                sim_file);
902        mtx_unlock(&sim_state.mutex);
903
904        sim_file->gmp = u_mmAllocMem(sim_state.heap, 8096, GMP_ALIGN2, 0);
905        sim_file->gmp_vaddr = (sim_state.mem + sim_file->gmp->ofs -
906                               sim_state.mem_base);
907        memset(sim_file->gmp_vaddr, 0, 8096);
908
909        return sim_file;
910}
911
912void
913v3d_simulator_destroy(struct v3d_simulator_file *sim_file)
914{
915        mtx_lock(&sim_state.mutex);
916        if (!--sim_state.refcount) {
917                _mesa_hash_table_destroy(sim_state.fd_map, NULL);
918                util_dynarray_fini(&sim_state.bin_oom);
919                u_mmDestroy(sim_state.heap);
920                /* No memsetting the struct, because it contains the mutex. */
921                sim_state.mem = NULL;
922        }
923        mtx_unlock(&sim_state.mutex);
924        ralloc_free(sim_file);
925}
926
927#endif /* USE_V3D_SIMULATOR */
928