2 * Copyright © 2014-2017 Broadcom
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:
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
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
25 * @file v3d_simulator.c
27 * Implements VC5 simulation on top of a non-VC5 GEM fd.
29 * This file's goal is to emulate the VC5 ioctls' behavior in the kernel on
30 * top of the simpenrose software simulator. Generally, VC5 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.
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).
42 * Another limitation is that BO import doesn't work unless the underlying
43 * window system's BO size matches what VC5 is going to use, which of course
44 * doesn't work out in practice. This means that for now, only DRI3 (VC5
45 * makes the winsys BOs) is supported, not DRI2 (window system makes the winys
49 #ifdef USE_V3D_SIMULATOR
52 #include "util/hash_table.h"
53 #include "util/ralloc.h"
55 #include "util/u_memory.h"
56 #include "util/u_mm.h"
57 #include "drm-uapi/i915_drm.h"
58 #include "v3d_simulator_wrapper.h"
60 #include "v3d_screen.h"
61 #include "v3d_context.h"
63 /** Global (across GEM fds) state for the simulator */
64 static struct v3d_simulator_state
{
70 /* Base virtual address of the heap. */
72 /* Base hardware address of the heap. */
74 /* Size of the heap. */
77 struct mem_block
*heap
;
78 struct mem_block
*overflow
;
80 /** Mapping from GEM fd to struct v3d_simulator_file * */
81 struct hash_table
*fd_map
;
85 .mutex
= _MTX_INITIALIZER_NP
,
88 /** Per-GEM-fd state for the simulator. */
89 struct v3d_simulator_file
{
92 /** Mapping from GEM handle to struct v3d_simulator_bo * */
93 struct hash_table
*bo_map
;
95 struct mem_block
*gmp
;
98 /** Actual GEM fd is i915, so we should use their create ioctl. */
102 /** Wrapper for drm_v3d_bo tracking the simulator-specific state. */
103 struct v3d_simulator_bo
{
104 struct v3d_simulator_file
*file
;
106 /** Area for this BO within sim_state->mem */
107 struct mem_block
*block
;
109 uint64_t mmap_offset
;
119 return (void *)(uintptr_t)key
;
122 static struct v3d_simulator_file
*
123 v3d_get_simulator_file_for_fd(int fd
)
125 struct hash_entry
*entry
= _mesa_hash_table_search(sim_state
.fd_map
,
127 return entry
? entry
->data
: NULL
;
130 /* A marker placed just after each BO, then checked after rendering to make
131 * sure it's still there.
133 #define BO_SENTINEL 0xfedcba98
136 #define GMP_ALIGN2 17
139 * Sets the range of GPU virtual address space to have the given GMP
140 * permissions (bit 0 = read, bit 1 = write, write-only forbidden).
143 set_gmp_flags(struct v3d_simulator_file
*file
,
144 uint32_t offset
, uint32_t size
, uint32_t flag
)
146 assert((offset
& ((1 << GMP_ALIGN2
) - 1)) == 0);
147 int gmp_offset
= offset
>> GMP_ALIGN2
;
148 int gmp_count
= align(size
, 1 << GMP_ALIGN2
) >> GMP_ALIGN2
;
149 uint32_t *gmp
= file
->gmp_vaddr
;
153 for (int i
= gmp_offset
; i
< gmp_offset
+ gmp_count
; i
++) {
154 int32_t bitshift
= (i
% 16) * 2;
155 gmp
[i
/ 16] &= ~(0x3 << bitshift
);
156 gmp
[i
/ 16] |= flag
<< bitshift
;
161 * Allocates space in simulator memory and returns a tracking struct for it
162 * that also contains the drm_gem_cma_object struct.
164 static struct v3d_simulator_bo
*
165 v3d_create_simulator_bo(int fd
, int handle
, unsigned size
)
167 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
168 struct v3d_simulator_bo
*sim_bo
= rzalloc(file
,
169 struct v3d_simulator_bo
);
170 size
= align(size
, 4096);
173 sim_bo
->handle
= handle
;
175 mtx_lock(&sim_state
.mutex
);
176 sim_bo
->block
= u_mmAllocMem(sim_state
.heap
, size
+ 4, GMP_ALIGN2
, 0);
177 mtx_unlock(&sim_state
.mutex
);
178 assert(sim_bo
->block
);
180 set_gmp_flags(file
, sim_bo
->block
->ofs
, size
, 0x3);
184 /* Allocate space for the buffer in simulator memory. */
185 sim_bo
->sim_vaddr
= sim_state
.mem
+ sim_bo
->block
->ofs
- sim_state
.mem_base
;
186 memset(sim_bo
->sim_vaddr
, 0xd0, size
);
188 *(uint32_t *)(sim_bo
->sim_vaddr
+ sim_bo
->size
) = BO_SENTINEL
;
190 /* Map the GEM buffer for copy in/out to the simulator. i915 blocks
191 * dumb mmap on render nodes, so use their ioctl directly if we're on
196 struct drm_i915_gem_mmap_gtt map
= {
200 /* We could potentially use non-gtt (cached) for LLC systems,
201 * but the copy-in/out won't be the limiting factor on
204 ret
= drmIoctl(fd
, DRM_IOCTL_I915_GEM_MMAP_GTT
, &map
);
205 sim_bo
->mmap_offset
= map
.offset
;
207 struct drm_mode_map_dumb map
= {
211 ret
= drmIoctl(fd
, DRM_IOCTL_MODE_MAP_DUMB
, &map
);
212 sim_bo
->mmap_offset
= map
.offset
;
215 fprintf(stderr
, "Failed to get MMAP offset: %d\n", ret
);
219 sim_bo
->gem_vaddr
= mmap(NULL
, sim_bo
->size
,
220 PROT_READ
| PROT_WRITE
, MAP_SHARED
,
221 fd
, sim_bo
->mmap_offset
);
222 if (sim_bo
->gem_vaddr
== MAP_FAILED
) {
223 fprintf(stderr
, "mmap of bo %d (offset 0x%016llx, size %d) failed\n",
224 handle
, (long long)sim_bo
->mmap_offset
, sim_bo
->size
);
228 /* A handle of 0 is used for v3d_gem.c internal allocations that
229 * don't need to go in the lookup table.
232 mtx_lock(&sim_state
.mutex
);
233 _mesa_hash_table_insert(file
->bo_map
, int_to_key(handle
),
235 mtx_unlock(&sim_state
.mutex
);
242 v3d_free_simulator_bo(struct v3d_simulator_bo
*sim_bo
)
244 struct v3d_simulator_file
*sim_file
= sim_bo
->file
;
246 set_gmp_flags(sim_file
, sim_bo
->block
->ofs
, sim_bo
->size
, 0x0);
248 if (sim_bo
->gem_vaddr
)
249 munmap(sim_bo
->gem_vaddr
, sim_bo
->size
);
251 mtx_lock(&sim_state
.mutex
);
252 u_mmFreeMem(sim_bo
->block
);
253 if (sim_bo
->handle
) {
254 _mesa_hash_table_remove_key(sim_file
->bo_map
,
255 int_to_key(sim_bo
->handle
));
257 mtx_unlock(&sim_state
.mutex
);
261 static struct v3d_simulator_bo
*
262 v3d_get_simulator_bo(struct v3d_simulator_file
*file
, int gem_handle
)
264 mtx_lock(&sim_state
.mutex
);
265 struct hash_entry
*entry
=
266 _mesa_hash_table_search(file
->bo_map
, int_to_key(gem_handle
));
267 mtx_unlock(&sim_state
.mutex
);
269 return entry
? entry
->data
: NULL
;
273 v3d_simulator_copy_in_handle(struct v3d_simulator_file
*file
, int handle
)
275 struct v3d_simulator_bo
*sim_bo
= v3d_get_simulator_bo(file
, handle
);
280 memcpy(sim_bo
->sim_vaddr
, sim_bo
->gem_vaddr
, sim_bo
->size
);
284 v3d_simulator_copy_out_handle(struct v3d_simulator_file
*file
, int handle
)
286 struct v3d_simulator_bo
*sim_bo
= v3d_get_simulator_bo(file
, handle
);
291 memcpy(sim_bo
->gem_vaddr
, sim_bo
->sim_vaddr
, sim_bo
->size
);
293 if (*(uint32_t *)(sim_bo
->sim_vaddr
+
294 sim_bo
->size
) != BO_SENTINEL
) {
295 fprintf(stderr
, "Buffer overflow in handle %d\n",
301 v3d_simulator_pin_bos(struct v3d_simulator_file
*file
,
302 struct drm_v3d_submit_cl
*submit
)
304 uint32_t *bo_handles
= (uint32_t *)(uintptr_t)submit
->bo_handles
;
306 for (int i
= 0; i
< submit
->bo_handle_count
; i
++)
307 v3d_simulator_copy_in_handle(file
, bo_handles
[i
]);
313 v3d_simulator_unpin_bos(struct v3d_simulator_file
*file
,
314 struct drm_v3d_submit_cl
*submit
)
316 uint32_t *bo_handles
= (uint32_t *)(uintptr_t)submit
->bo_handles
;
318 for (int i
= 0; i
< submit
->bo_handle_count
; i
++)
319 v3d_simulator_copy_out_handle(file
, bo_handles
[i
]);
325 v3d_simulator_submit_cl_ioctl(int fd
, struct drm_v3d_submit_cl
*submit
)
327 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
330 ret
= v3d_simulator_pin_bos(file
, submit
);
334 if (sim_state
.ver
>= 41)
335 v3d41_simulator_submit_cl_ioctl(sim_state
.v3d
, submit
, file
->gmp
->ofs
);
337 v3d33_simulator_submit_cl_ioctl(sim_state
.v3d
, submit
, file
->gmp
->ofs
);
339 ret
= v3d_simulator_unpin_bos(file
, submit
);
347 * Do fixups after a BO has been opened from a handle.
349 * This could be done at DRM_IOCTL_GEM_OPEN/DRM_IOCTL_GEM_PRIME_FD_TO_HANDLE
350 * time, but we're still using drmPrimeFDToHandle() so we have this helper to
351 * be called afterward instead.
353 void v3d_simulator_open_from_handle(int fd
, int handle
, uint32_t size
)
355 v3d_create_simulator_bo(fd
, handle
, size
);
359 * Simulated ioctl(fd, DRM_VC5_CREATE_BO) implementation.
361 * Making a VC5 BO is just a matter of making a corresponding BO on the host.
364 v3d_simulator_create_bo_ioctl(int fd
, struct drm_v3d_create_bo
*args
)
366 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
368 /* i915 bans dumb create on render nodes, so we have to use their
369 * native ioctl in case we're on a render node.
373 struct drm_i915_gem_create create
= {
376 ret
= drmIoctl(fd
, DRM_IOCTL_I915_GEM_CREATE
, &create
);
378 args
->handle
= create
.handle
;
380 struct drm_mode_create_dumb create
= {
383 .height
= (args
->size
+ 127) / 128,
386 ret
= drmIoctl(fd
, DRM_IOCTL_MODE_CREATE_DUMB
, &create
);
387 assert(ret
!= 0 || create
.size
>= args
->size
);
389 args
->handle
= create
.handle
;
393 struct v3d_simulator_bo
*sim_bo
=
394 v3d_create_simulator_bo(fd
, args
->handle
, args
->size
);
396 args
->offset
= sim_bo
->block
->ofs
;
403 * Simulated ioctl(fd, DRM_VC5_MMAP_BO) implementation.
405 * We've already grabbed the mmap offset when we created the sim bo, so just
409 v3d_simulator_mmap_bo_ioctl(int fd
, struct drm_v3d_mmap_bo
*args
)
411 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
412 struct v3d_simulator_bo
*sim_bo
= v3d_get_simulator_bo(file
,
415 args
->offset
= sim_bo
->mmap_offset
;
421 v3d_simulator_get_bo_offset_ioctl(int fd
, struct drm_v3d_get_bo_offset
*args
)
423 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
424 struct v3d_simulator_bo
*sim_bo
= v3d_get_simulator_bo(file
,
427 args
->offset
= sim_bo
->block
->ofs
;
433 v3d_simulator_gem_close_ioctl(int fd
, struct drm_gem_close
*args
)
435 /* Free the simulator's internal tracking. */
436 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
437 struct v3d_simulator_bo
*sim_bo
= v3d_get_simulator_bo(file
,
440 v3d_free_simulator_bo(sim_bo
);
442 /* Pass the call on down. */
443 return drmIoctl(fd
, DRM_IOCTL_GEM_CLOSE
, args
);
447 v3d_simulator_get_param_ioctl(int fd
, struct drm_v3d_get_param
*args
)
449 if (sim_state
.ver
>= 41)
450 return v3d41_simulator_get_param_ioctl(sim_state
.v3d
, args
);
452 return v3d33_simulator_get_param_ioctl(sim_state
.v3d
, args
);
456 v3d_simulator_submit_tfu_ioctl(int fd
, struct drm_v3d_submit_tfu
*args
)
458 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
461 v3d_simulator_copy_in_handle(file
, args
->bo_handles
[0]);
462 v3d_simulator_copy_in_handle(file
, args
->bo_handles
[1]);
463 v3d_simulator_copy_in_handle(file
, args
->bo_handles
[2]);
464 v3d_simulator_copy_in_handle(file
, args
->bo_handles
[3]);
466 if (sim_state
.ver
>= 41)
467 ret
= v3d41_simulator_submit_tfu_ioctl(sim_state
.v3d
, args
);
469 ret
= v3d33_simulator_submit_tfu_ioctl(sim_state
.v3d
, args
);
471 v3d_simulator_copy_out_handle(file
, args
->bo_handles
[0]);
477 v3d_simulator_ioctl(int fd
, unsigned long request
, void *args
)
480 case DRM_IOCTL_V3D_SUBMIT_CL
:
481 return v3d_simulator_submit_cl_ioctl(fd
, args
);
482 case DRM_IOCTL_V3D_CREATE_BO
:
483 return v3d_simulator_create_bo_ioctl(fd
, args
);
484 case DRM_IOCTL_V3D_MMAP_BO
:
485 return v3d_simulator_mmap_bo_ioctl(fd
, args
);
486 case DRM_IOCTL_V3D_GET_BO_OFFSET
:
487 return v3d_simulator_get_bo_offset_ioctl(fd
, args
);
489 case DRM_IOCTL_V3D_WAIT_BO
:
490 /* We do all of the v3d rendering synchronously, so we just
491 * return immediately on the wait ioctls. This ignores any
492 * native rendering to the host BO, so it does mean we race on
493 * front buffer rendering.
497 case DRM_IOCTL_V3D_GET_PARAM
:
498 return v3d_simulator_get_param_ioctl(fd
, args
);
500 case DRM_IOCTL_GEM_CLOSE
:
501 return v3d_simulator_gem_close_ioctl(fd
, args
);
503 case DRM_IOCTL_V3D_SUBMIT_TFU
:
504 return v3d_simulator_submit_tfu_ioctl(fd
, args
);
506 case DRM_IOCTL_GEM_OPEN
:
507 case DRM_IOCTL_GEM_FLINK
:
508 return drmIoctl(fd
, request
, args
);
510 fprintf(stderr
, "Unknown ioctl 0x%08x\n", (int)request
);
516 v3d_simulator_init_global(const struct v3d_device_info
*devinfo
)
518 mtx_lock(&sim_state
.mutex
);
519 if (sim_state
.refcount
++) {
520 mtx_unlock(&sim_state
.mutex
);
524 sim_state
.v3d
= v3d_hw_auto_new(NULL
);
525 v3d_hw_alloc_mem(sim_state
.v3d
, 1024 * 1024 * 1024);
527 v3d_hw_get_mem(sim_state
.v3d
, &sim_state
.mem_size
,
530 /* Allocate from anywhere from 4096 up. We don't allocate at 0,
531 * because for OQs and some other addresses in the HW, 0 means
534 sim_state
.heap
= u_mmInit(4096, sim_state
.mem_size
- 4096);
536 /* Make a block of 0xd0 at address 0 to make sure we don't screw up
539 struct mem_block
*b
= u_mmAllocMem(sim_state
.heap
, 4096, GMP_ALIGN2
, 0);
540 memset(sim_state
.mem
+ b
->ofs
- sim_state
.mem_base
, 0xd0, 4096);
542 sim_state
.ver
= v3d_hw_get_version(sim_state
.v3d
);
544 mtx_unlock(&sim_state
.mutex
);
547 _mesa_hash_table_create(NULL
,
549 _mesa_key_pointer_equal
);
551 if (sim_state
.ver
>= 41)
552 v3d41_simulator_init_regs(sim_state
.v3d
);
554 v3d33_simulator_init_regs(sim_state
.v3d
);
558 v3d_simulator_init(struct v3d_screen
*screen
)
560 v3d_simulator_init_global(&screen
->devinfo
);
562 screen
->sim_file
= rzalloc(screen
, struct v3d_simulator_file
);
563 struct v3d_simulator_file
*sim_file
= screen
->sim_file
;
565 drmVersionPtr version
= drmGetVersion(screen
->fd
);
566 if (version
&& strncmp(version
->name
, "i915", version
->name_len
) == 0)
567 sim_file
->is_i915
= true;
568 drmFreeVersion(version
);
570 screen
->sim_file
->bo_map
=
571 _mesa_hash_table_create(screen
->sim_file
,
573 _mesa_key_pointer_equal
);
575 mtx_lock(&sim_state
.mutex
);
576 _mesa_hash_table_insert(sim_state
.fd_map
, int_to_key(screen
->fd
+ 1),
578 mtx_unlock(&sim_state
.mutex
);
580 sim_file
->gmp
= u_mmAllocMem(sim_state
.heap
, 8096, GMP_ALIGN2
, 0);
581 sim_file
->gmp_vaddr
= (sim_state
.mem
+ sim_file
->gmp
->ofs
-
583 memset(sim_file
->gmp_vaddr
, 0, 8096);
587 v3d_simulator_destroy(struct v3d_screen
*screen
)
589 mtx_lock(&sim_state
.mutex
);
590 if (!--sim_state
.refcount
) {
591 _mesa_hash_table_destroy(sim_state
.fd_map
, NULL
);
592 u_mmDestroy(sim_state
.heap
);
593 /* No memsetting the struct, because it contains the mutex. */
594 sim_state
.mem
= NULL
;
596 mtx_unlock(&sim_state
.mutex
);
599 #endif /* USE_V3D_SIMULATOR */