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
53 #include "c11/threads.h"
54 #include "util/hash_table.h"
55 #include "util/ralloc.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"
63 #include "drm-uapi/i915_drm.h"
64 #include "drm-uapi/v3d_drm.h"
66 #include "v3d_simulator.h"
67 #include "v3d_simulator_wrapper.h"
69 /** Global (across GEM fds) state for the simulator */
70 static struct v3d_simulator_state
{
77 /* Base virtual address of the heap. */
79 /* Base hardware address of the heap. */
81 /* Size of the heap. */
84 struct mem_block
*heap
;
85 struct mem_block
*overflow
;
87 /** Mapping from GEM fd to struct v3d_simulator_file * */
88 struct hash_table
*fd_map
;
90 struct util_dynarray bin_oom
;
93 .mutex
= _MTX_INITIALIZER_NP
,
96 /** Per-GEM-fd state for the simulator. */
97 struct v3d_simulator_file
{
100 /** Mapping from GEM handle to struct v3d_simulator_bo * */
101 struct hash_table
*bo_map
;
103 struct mem_block
*gmp
;
106 /** Actual GEM fd is i915, so we should use their create ioctl. */
110 /** Wrapper for drm_v3d_bo tracking the simulator-specific state. */
111 struct v3d_simulator_bo
{
112 struct v3d_simulator_file
*file
;
114 /** Area for this BO within sim_state->mem */
115 struct mem_block
*block
;
117 uint64_t mmap_offset
;
127 return (void *)(uintptr_t)key
;
130 static struct v3d_simulator_file
*
131 v3d_get_simulator_file_for_fd(int fd
)
133 struct hash_entry
*entry
= _mesa_hash_table_search(sim_state
.fd_map
,
135 return entry
? entry
->data
: NULL
;
138 /* A marker placed just after each BO, then checked after rendering to make
139 * sure it's still there.
141 #define BO_SENTINEL 0xfedcba98
144 #define GMP_ALIGN2 17
147 * Sets the range of GPU virtual address space to have the given GMP
148 * permissions (bit 0 = read, bit 1 = write, write-only forbidden).
151 set_gmp_flags(struct v3d_simulator_file
*file
,
152 uint32_t offset
, uint32_t size
, uint32_t flag
)
154 assert((offset
& ((1 << GMP_ALIGN2
) - 1)) == 0);
155 int gmp_offset
= offset
>> GMP_ALIGN2
;
156 int gmp_count
= align(size
, 1 << GMP_ALIGN2
) >> GMP_ALIGN2
;
157 uint32_t *gmp
= file
->gmp_vaddr
;
161 for (int i
= gmp_offset
; i
< gmp_offset
+ gmp_count
; i
++) {
162 int32_t bitshift
= (i
% 16) * 2;
163 gmp
[i
/ 16] &= ~(0x3 << bitshift
);
164 gmp
[i
/ 16] |= flag
<< bitshift
;
169 * Allocates space in simulator memory and returns a tracking struct for it
170 * that also contains the drm_gem_cma_object struct.
172 static struct v3d_simulator_bo
*
173 v3d_create_simulator_bo(int fd
, unsigned size
)
175 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
176 struct v3d_simulator_bo
*sim_bo
= rzalloc(file
,
177 struct v3d_simulator_bo
);
178 size
= align(size
, 4096);
182 mtx_lock(&sim_state
.mutex
);
183 sim_bo
->block
= u_mmAllocMem(sim_state
.heap
, size
+ 4, GMP_ALIGN2
, 0);
184 mtx_unlock(&sim_state
.mutex
);
185 assert(sim_bo
->block
);
187 set_gmp_flags(file
, sim_bo
->block
->ofs
, size
, 0x3);
191 /* Allocate space for the buffer in simulator memory. */
192 sim_bo
->sim_vaddr
= sim_state
.mem
+ sim_bo
->block
->ofs
- sim_state
.mem_base
;
193 memset(sim_bo
->sim_vaddr
, 0xd0, size
);
195 *(uint32_t *)(sim_bo
->sim_vaddr
+ sim_bo
->size
) = BO_SENTINEL
;
200 static struct v3d_simulator_bo
*
201 v3d_create_simulator_bo_for_gem(int fd
, int handle
, unsigned size
)
203 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
204 struct v3d_simulator_bo
*sim_bo
=
205 v3d_create_simulator_bo(fd
, size
);
207 sim_bo
->handle
= handle
;
209 /* Map the GEM buffer for copy in/out to the simulator. i915 blocks
210 * dumb mmap on render nodes, so use their ioctl directly if we're on
215 struct drm_i915_gem_mmap_gtt map
= {
219 /* We could potentially use non-gtt (cached) for LLC systems,
220 * but the copy-in/out won't be the limiting factor on
223 ret
= drmIoctl(fd
, DRM_IOCTL_I915_GEM_MMAP_GTT
, &map
);
224 sim_bo
->mmap_offset
= map
.offset
;
226 struct drm_mode_map_dumb map
= {
230 ret
= drmIoctl(fd
, DRM_IOCTL_MODE_MAP_DUMB
, &map
);
231 sim_bo
->mmap_offset
= map
.offset
;
234 fprintf(stderr
, "Failed to get MMAP offset: %d\n", ret
);
238 sim_bo
->gem_vaddr
= mmap(NULL
, sim_bo
->size
,
239 PROT_READ
| PROT_WRITE
, MAP_SHARED
,
240 fd
, sim_bo
->mmap_offset
);
241 if (sim_bo
->gem_vaddr
== MAP_FAILED
) {
242 fprintf(stderr
, "mmap of bo %d (offset 0x%016llx, size %d) failed\n",
243 handle
, (long long)sim_bo
->mmap_offset
, sim_bo
->size
);
247 /* A handle of 0 is used for v3d_gem.c internal allocations that
248 * don't need to go in the lookup table.
251 mtx_lock(&sim_state
.mutex
);
252 _mesa_hash_table_insert(file
->bo_map
, int_to_key(handle
),
254 mtx_unlock(&sim_state
.mutex
);
263 v3d_simulator_get_spill(uint32_t spill_size
)
265 struct v3d_simulator_bo
*sim_bo
=
266 v3d_create_simulator_bo(bin_fd
, spill_size
);
268 util_dynarray_append(&sim_state
.bin_oom
, struct v3d_simulator_bo
*,
271 return sim_bo
->block
->ofs
;
275 v3d_free_simulator_bo(struct v3d_simulator_bo
*sim_bo
)
277 struct v3d_simulator_file
*sim_file
= sim_bo
->file
;
279 set_gmp_flags(sim_file
, sim_bo
->block
->ofs
, sim_bo
->size
, 0x0);
281 if (sim_bo
->gem_vaddr
)
282 munmap(sim_bo
->gem_vaddr
, sim_bo
->size
);
284 mtx_lock(&sim_state
.mutex
);
285 u_mmFreeMem(sim_bo
->block
);
286 if (sim_bo
->handle
) {
287 _mesa_hash_table_remove_key(sim_file
->bo_map
,
288 int_to_key(sim_bo
->handle
));
290 mtx_unlock(&sim_state
.mutex
);
294 static struct v3d_simulator_bo
*
295 v3d_get_simulator_bo(struct v3d_simulator_file
*file
, int gem_handle
)
300 mtx_lock(&sim_state
.mutex
);
301 struct hash_entry
*entry
=
302 _mesa_hash_table_search(file
->bo_map
, int_to_key(gem_handle
));
303 mtx_unlock(&sim_state
.mutex
);
305 return entry
? entry
->data
: NULL
;
309 v3d_simulator_copy_in_handle(struct v3d_simulator_file
*file
, int handle
)
311 struct v3d_simulator_bo
*sim_bo
= v3d_get_simulator_bo(file
, handle
);
316 memcpy(sim_bo
->sim_vaddr
, sim_bo
->gem_vaddr
, sim_bo
->size
);
320 v3d_simulator_copy_out_handle(struct v3d_simulator_file
*file
, int handle
)
322 struct v3d_simulator_bo
*sim_bo
= v3d_get_simulator_bo(file
, handle
);
327 memcpy(sim_bo
->gem_vaddr
, sim_bo
->sim_vaddr
, sim_bo
->size
);
329 if (*(uint32_t *)(sim_bo
->sim_vaddr
+
330 sim_bo
->size
) != BO_SENTINEL
) {
331 fprintf(stderr
, "Buffer overflow in handle %d\n",
337 v3d_simulator_pin_bos(struct v3d_simulator_file
*file
,
338 struct drm_v3d_submit_cl
*submit
)
340 uint32_t *bo_handles
= (uint32_t *)(uintptr_t)submit
->bo_handles
;
342 for (int i
= 0; i
< submit
->bo_handle_count
; i
++)
343 v3d_simulator_copy_in_handle(file
, bo_handles
[i
]);
349 v3d_simulator_unpin_bos(struct v3d_simulator_file
*file
,
350 struct drm_v3d_submit_cl
*submit
)
352 uint32_t *bo_handles
= (uint32_t *)(uintptr_t)submit
->bo_handles
;
354 for (int i
= 0; i
< submit
->bo_handle_count
; i
++)
355 v3d_simulator_copy_out_handle(file
, bo_handles
[i
]);
361 v3d_simulator_submit_cl_ioctl(int fd
, struct drm_v3d_submit_cl
*submit
)
363 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
366 ret
= v3d_simulator_pin_bos(file
, submit
);
370 mtx_lock(&sim_state
.submit_lock
);
372 if (sim_state
.ver
>= 41)
373 v3d41_simulator_submit_cl_ioctl(sim_state
.v3d
, submit
, file
->gmp
->ofs
);
375 v3d33_simulator_submit_cl_ioctl(sim_state
.v3d
, submit
, file
->gmp
->ofs
);
377 util_dynarray_foreach(&sim_state
.bin_oom
, struct v3d_simulator_bo
*,
379 v3d_free_simulator_bo(*sim_bo
);
381 util_dynarray_clear(&sim_state
.bin_oom
);
383 mtx_unlock(&sim_state
.submit_lock
);
385 ret
= v3d_simulator_unpin_bos(file
, submit
);
393 * Do fixups after a BO has been opened from a handle.
395 * This could be done at DRM_IOCTL_GEM_OPEN/DRM_IOCTL_GEM_PRIME_FD_TO_HANDLE
396 * time, but we're still using drmPrimeFDToHandle() so we have this helper to
397 * be called afterward instead.
399 void v3d_simulator_open_from_handle(int fd
, int handle
, uint32_t size
)
401 v3d_create_simulator_bo_for_gem(fd
, handle
, size
);
405 * Simulated ioctl(fd, DRM_VC5_CREATE_BO) implementation.
407 * Making a VC5 BO is just a matter of making a corresponding BO on the host.
410 v3d_simulator_create_bo_ioctl(int fd
, struct drm_v3d_create_bo
*args
)
412 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
414 /* i915 bans dumb create on render nodes, so we have to use their
415 * native ioctl in case we're on a render node.
419 struct drm_i915_gem_create create
= {
422 ret
= drmIoctl(fd
, DRM_IOCTL_I915_GEM_CREATE
, &create
);
424 args
->handle
= create
.handle
;
426 struct drm_mode_create_dumb create
= {
429 .height
= (args
->size
+ 127) / 128,
432 ret
= drmIoctl(fd
, DRM_IOCTL_MODE_CREATE_DUMB
, &create
);
433 assert(ret
!= 0 || create
.size
>= args
->size
);
435 args
->handle
= create
.handle
;
439 struct v3d_simulator_bo
*sim_bo
=
440 v3d_create_simulator_bo_for_gem(fd
, args
->handle
,
443 args
->offset
= sim_bo
->block
->ofs
;
450 * Simulated ioctl(fd, DRM_VC5_MMAP_BO) implementation.
452 * We've already grabbed the mmap offset when we created the sim bo, so just
456 v3d_simulator_mmap_bo_ioctl(int fd
, struct drm_v3d_mmap_bo
*args
)
458 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
459 struct v3d_simulator_bo
*sim_bo
= v3d_get_simulator_bo(file
,
462 args
->offset
= sim_bo
->mmap_offset
;
468 v3d_simulator_get_bo_offset_ioctl(int fd
, struct drm_v3d_get_bo_offset
*args
)
470 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
471 struct v3d_simulator_bo
*sim_bo
= v3d_get_simulator_bo(file
,
474 args
->offset
= sim_bo
->block
->ofs
;
480 v3d_simulator_gem_close_ioctl(int fd
, struct drm_gem_close
*args
)
482 /* Free the simulator's internal tracking. */
483 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
484 struct v3d_simulator_bo
*sim_bo
= v3d_get_simulator_bo(file
,
487 v3d_free_simulator_bo(sim_bo
);
489 /* Pass the call on down. */
490 return drmIoctl(fd
, DRM_IOCTL_GEM_CLOSE
, args
);
494 v3d_simulator_get_param_ioctl(int fd
, struct drm_v3d_get_param
*args
)
496 if (sim_state
.ver
>= 41)
497 return v3d41_simulator_get_param_ioctl(sim_state
.v3d
, args
);
499 return v3d33_simulator_get_param_ioctl(sim_state
.v3d
, args
);
503 v3d_simulator_submit_tfu_ioctl(int fd
, struct drm_v3d_submit_tfu
*args
)
505 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
508 v3d_simulator_copy_in_handle(file
, args
->bo_handles
[0]);
509 v3d_simulator_copy_in_handle(file
, args
->bo_handles
[1]);
510 v3d_simulator_copy_in_handle(file
, args
->bo_handles
[2]);
511 v3d_simulator_copy_in_handle(file
, args
->bo_handles
[3]);
513 if (sim_state
.ver
>= 41)
514 ret
= v3d41_simulator_submit_tfu_ioctl(sim_state
.v3d
, args
);
516 ret
= v3d33_simulator_submit_tfu_ioctl(sim_state
.v3d
, args
);
518 v3d_simulator_copy_out_handle(file
, args
->bo_handles
[0]);
524 v3d_simulator_submit_csd_ioctl(int fd
, struct drm_v3d_submit_csd
*args
)
526 struct v3d_simulator_file
*file
= v3d_get_simulator_file_for_fd(fd
);
527 uint32_t *bo_handles
= (uint32_t *)(uintptr_t)args
->bo_handles
;
530 for (int i
= 0; i
< args
->bo_handle_count
; i
++)
531 v3d_simulator_copy_in_handle(file
, bo_handles
[i
]);
533 if (sim_state
.ver
>= 41)
534 ret
= v3d41_simulator_submit_csd_ioctl(sim_state
.v3d
, args
,
539 for (int i
= 0; i
< args
->bo_handle_count
; i
++)
540 v3d_simulator_copy_out_handle(file
, bo_handles
[i
]);
546 v3d_simulator_ioctl(int fd
, unsigned long request
, void *args
)
549 case DRM_IOCTL_V3D_SUBMIT_CL
:
550 return v3d_simulator_submit_cl_ioctl(fd
, args
);
551 case DRM_IOCTL_V3D_CREATE_BO
:
552 return v3d_simulator_create_bo_ioctl(fd
, args
);
553 case DRM_IOCTL_V3D_MMAP_BO
:
554 return v3d_simulator_mmap_bo_ioctl(fd
, args
);
555 case DRM_IOCTL_V3D_GET_BO_OFFSET
:
556 return v3d_simulator_get_bo_offset_ioctl(fd
, args
);
558 case DRM_IOCTL_V3D_WAIT_BO
:
559 /* We do all of the v3d rendering synchronously, so we just
560 * return immediately on the wait ioctls. This ignores any
561 * native rendering to the host BO, so it does mean we race on
562 * front buffer rendering.
566 case DRM_IOCTL_V3D_GET_PARAM
:
567 return v3d_simulator_get_param_ioctl(fd
, args
);
569 case DRM_IOCTL_GEM_CLOSE
:
570 return v3d_simulator_gem_close_ioctl(fd
, args
);
572 case DRM_IOCTL_V3D_SUBMIT_TFU
:
573 return v3d_simulator_submit_tfu_ioctl(fd
, args
);
575 case DRM_IOCTL_V3D_SUBMIT_CSD
:
576 return v3d_simulator_submit_csd_ioctl(fd
, args
);
578 case DRM_IOCTL_GEM_OPEN
:
579 case DRM_IOCTL_GEM_FLINK
:
580 return drmIoctl(fd
, request
, args
);
582 fprintf(stderr
, "Unknown ioctl 0x%08x\n", (int)request
);
588 v3d_simulator_init_global()
590 mtx_lock(&sim_state
.mutex
);
591 if (sim_state
.refcount
++) {
592 mtx_unlock(&sim_state
.mutex
);
596 sim_state
.v3d
= v3d_hw_auto_new(NULL
);
597 v3d_hw_alloc_mem(sim_state
.v3d
, 1024 * 1024 * 1024);
599 v3d_hw_get_mem(sim_state
.v3d
, &sim_state
.mem_size
,
602 /* Allocate from anywhere from 4096 up. We don't allocate at 0,
603 * because for OQs and some other addresses in the HW, 0 means
606 sim_state
.heap
= u_mmInit(4096, sim_state
.mem_size
- 4096);
608 /* Make a block of 0xd0 at address 0 to make sure we don't screw up
611 struct mem_block
*b
= u_mmAllocMem(sim_state
.heap
, 4096, GMP_ALIGN2
, 0);
612 memset(sim_state
.mem
+ b
->ofs
- sim_state
.mem_base
, 0xd0, 4096);
614 sim_state
.ver
= v3d_hw_get_version(sim_state
.v3d
);
616 mtx_unlock(&sim_state
.mutex
);
619 _mesa_hash_table_create(NULL
,
621 _mesa_key_pointer_equal
);
623 util_dynarray_init(&sim_state
.bin_oom
, NULL
);
625 if (sim_state
.ver
>= 41)
626 v3d41_simulator_init_regs(sim_state
.v3d
);
628 v3d33_simulator_init_regs(sim_state
.v3d
);
631 struct v3d_simulator_file
*
632 v3d_simulator_init(int fd
)
634 v3d_simulator_init_global();
636 struct v3d_simulator_file
*sim_file
= rzalloc(NULL
, struct v3d_simulator_file
);
638 drmVersionPtr version
= drmGetVersion(fd
);
639 if (version
&& strncmp(version
->name
, "i915", version
->name_len
) == 0)
640 sim_file
->is_i915
= true;
641 drmFreeVersion(version
);
644 _mesa_hash_table_create(sim_file
,
646 _mesa_key_pointer_equal
);
648 mtx_lock(&sim_state
.mutex
);
649 _mesa_hash_table_insert(sim_state
.fd_map
, int_to_key(fd
+ 1),
651 mtx_unlock(&sim_state
.mutex
);
653 sim_file
->gmp
= u_mmAllocMem(sim_state
.heap
, 8096, GMP_ALIGN2
, 0);
654 sim_file
->gmp_vaddr
= (sim_state
.mem
+ sim_file
->gmp
->ofs
-
656 memset(sim_file
->gmp_vaddr
, 0, 8096);
662 v3d_simulator_destroy(struct v3d_simulator_file
*sim_file
)
664 mtx_lock(&sim_state
.mutex
);
665 if (!--sim_state
.refcount
) {
666 _mesa_hash_table_destroy(sim_state
.fd_map
, NULL
);
667 util_dynarray_fini(&sim_state
.bin_oom
);
668 u_mmDestroy(sim_state
.heap
);
669 /* No memsetting the struct, because it contains the mutex. */
670 sim_state
.mem
= NULL
;
672 mtx_unlock(&sim_state
.mutex
);
673 ralloc_free(sim_file
);
676 #endif /* USE_V3D_SIMULATOR */