xref: /linux/drivers/gpu/drm/imagination/pvr_hwrt.c (revision 021bc4b9)
1 // SPDX-License-Identifier: GPL-2.0-only OR MIT
2 /* Copyright (c) 2023 Imagination Technologies Ltd. */
3 
4 #include "pvr_free_list.h"
5 #include "pvr_hwrt.h"
6 #include "pvr_gem.h"
7 #include "pvr_rogue_cr_defs_client.h"
8 #include "pvr_rogue_fwif.h"
9 
10 #include <drm/drm_gem.h>
11 #include <linux/bitops.h>
12 #include <linux/math.h>
13 #include <linux/slab.h>
14 #include <linux/xarray.h>
15 #include <uapi/drm/pvr_drm.h>
16 
17 static_assert(ROGUE_FWIF_NUM_RTDATAS == 2);
18 static_assert(ROGUE_FWIF_NUM_GEOMDATAS == 1);
19 static_assert(ROGUE_FWIF_NUM_RTDATA_FREELISTS == 2);
20 
21 /*
22  * struct pvr_rt_mtile_info - Render target macrotile information
23  */
24 struct pvr_rt_mtile_info {
25 	u32 mtile_x[3];
26 	u32 mtile_y[3];
27 	u32 tile_max_x;
28 	u32 tile_max_y;
29 	u32 tile_size_x;
30 	u32 tile_size_y;
31 	u32 num_tiles_x;
32 	u32 num_tiles_y;
33 };
34 
35 /* Size of Shadow Render Target Cache entry */
36 #define SRTC_ENTRY_SIZE sizeof(u32)
37 /* Size of Renders Accumulation Array entry */
38 #define RAA_ENTRY_SIZE sizeof(u32)
39 
40 static int
41 hwrt_init_kernel_structure(struct pvr_file *pvr_file,
42 			   struct drm_pvr_ioctl_create_hwrt_dataset_args *args,
43 			   struct pvr_hwrt_dataset *hwrt)
44 {
45 	struct pvr_device *pvr_dev = pvr_file->pvr_dev;
46 	int err;
47 	int i;
48 
49 	hwrt->pvr_dev = pvr_dev;
50 	hwrt->max_rts = args->layers;
51 
52 	/* Get pointers to the free lists */
53 	for (i = 0; i < ARRAY_SIZE(hwrt->free_lists); i++) {
54 		hwrt->free_lists[i] = pvr_free_list_lookup(pvr_file,  args->free_list_handles[i]);
55 		if (!hwrt->free_lists[i]) {
56 			err = -EINVAL;
57 			goto err_put_free_lists;
58 		}
59 	}
60 
61 	if (hwrt->free_lists[ROGUE_FW_LOCAL_FREELIST]->current_pages <
62 	    pvr_get_free_list_min_pages(pvr_dev)) {
63 		err = -EINVAL;
64 		goto err_put_free_lists;
65 	}
66 
67 	return 0;
68 
69 err_put_free_lists:
70 	for (i = 0; i < ARRAY_SIZE(hwrt->free_lists); i++) {
71 		pvr_free_list_put(hwrt->free_lists[i]);
72 		hwrt->free_lists[i] = NULL;
73 	}
74 
75 	return err;
76 }
77 
78 static void
79 hwrt_fini_kernel_structure(struct pvr_hwrt_dataset *hwrt)
80 {
81 	int i;
82 
83 	for (i = 0; i < ARRAY_SIZE(hwrt->free_lists); i++) {
84 		pvr_free_list_put(hwrt->free_lists[i]);
85 		hwrt->free_lists[i] = NULL;
86 	}
87 }
88 
89 static void
90 hwrt_fini_common_fw_structure(struct pvr_hwrt_dataset *hwrt)
91 {
92 	pvr_fw_object_destroy(hwrt->common_fw_obj);
93 }
94 
95 static int
96 get_cr_isp_mtile_size_val(struct pvr_device *pvr_dev, u32 samples,
97 			  struct pvr_rt_mtile_info *info, u32 *value_out)
98 {
99 	u32 x = info->mtile_x[0];
100 	u32 y = info->mtile_y[0];
101 	u32 samples_per_pixel;
102 	int err;
103 
104 	err = PVR_FEATURE_VALUE(pvr_dev, isp_samples_per_pixel, &samples_per_pixel);
105 	if (err)
106 		return err;
107 
108 	if (samples_per_pixel == 1) {
109 		if (samples >= 4)
110 			x <<= 1;
111 		if (samples >= 2)
112 			y <<= 1;
113 	} else if (samples_per_pixel == 2) {
114 		if (samples >= 8)
115 			x <<= 1;
116 		if (samples >= 4)
117 			y <<= 1;
118 	} else if (samples_per_pixel == 4) {
119 		if (samples >= 8)
120 			y <<= 1;
121 	} else {
122 		WARN(true, "Unsupported ISP samples per pixel value");
123 		return -EINVAL;
124 	}
125 
126 	*value_out = ((x << ROGUE_CR_ISP_MTILE_SIZE_X_SHIFT) & ~ROGUE_CR_ISP_MTILE_SIZE_X_CLRMSK) |
127 		     ((y << ROGUE_CR_ISP_MTILE_SIZE_Y_SHIFT) & ~ROGUE_CR_ISP_MTILE_SIZE_Y_CLRMSK);
128 
129 	return 0;
130 }
131 
132 static int
133 get_cr_multisamplectl_val(u32 samples, bool y_flip, u64 *value_out)
134 {
135 	static const struct {
136 		u8 x[8];
137 		u8 y[8];
138 	} sample_positions[4] = {
139 		/* 1 sample */
140 		{
141 			.x = { 8 },
142 			.y = { 8 },
143 		},
144 		/* 2 samples */
145 		{
146 			.x = { 12, 4 },
147 			.y = { 12, 4 },
148 		},
149 		/* 4 samples */
150 		{
151 			.x = { 6, 14, 2, 10 },
152 			.y = { 2, 6, 10, 14 },
153 		},
154 		/* 8 samples */
155 		{
156 			.x = { 9, 7, 13, 5, 3, 1, 11, 15 },
157 			.y = { 5, 11, 9, 3, 13, 7, 15, 1 },
158 		},
159 	};
160 	const int idx = fls(samples) - 1;
161 	u64 value = 0;
162 
163 	if (idx < 0 || idx > 3)
164 		return -EINVAL;
165 
166 	for (u32 i = 0; i < 8; i++) {
167 		value |= ((u64)sample_positions[idx].x[i]) << (i * 8);
168 		if (y_flip)
169 			value |= (((u64)(16 - sample_positions[idx].y[i]) & 0xf)) << (i * 8 + 4);
170 		else
171 			value |= ((u64)sample_positions[idx].y[i]) << (i * 8 + 4);
172 	}
173 
174 	*value_out = value;
175 
176 	return 0;
177 }
178 
179 static int
180 get_cr_te_aa_val(struct pvr_device *pvr_dev, u32 samples, u32 *value_out)
181 {
182 	u32 samples_per_pixel;
183 	u32 value = 0;
184 	int err = 0;
185 
186 	err = PVR_FEATURE_VALUE(pvr_dev, isp_samples_per_pixel, &samples_per_pixel);
187 	if (err)
188 		return err;
189 
190 	switch (samples_per_pixel) {
191 	case 1:
192 		if (samples >= 2)
193 			value |= ROGUE_CR_TE_AA_Y_EN;
194 		if (samples >= 4)
195 			value |= ROGUE_CR_TE_AA_X_EN;
196 		break;
197 	case 2:
198 		if (samples >= 2)
199 			value |= ROGUE_CR_TE_AA_X2_EN;
200 		if (samples >= 4)
201 			value |= ROGUE_CR_TE_AA_Y_EN;
202 		if (samples >= 8)
203 			value |= ROGUE_CR_TE_AA_X_EN;
204 		break;
205 	case 4:
206 		if (samples >= 2)
207 			value |= ROGUE_CR_TE_AA_X2_EN;
208 		if (samples >= 4)
209 			value |= ROGUE_CR_TE_AA_Y2_EN;
210 		if (samples >= 8)
211 			value |= ROGUE_CR_TE_AA_Y_EN;
212 		break;
213 	default:
214 		WARN(true, "Unsupported ISP samples per pixel value");
215 		return -EINVAL;
216 	}
217 
218 	*value_out = value;
219 
220 	return 0;
221 }
222 
223 static void
224 hwrtdata_common_init(void *cpu_ptr, void *priv)
225 {
226 	struct pvr_hwrt_dataset *hwrt = priv;
227 
228 	memcpy(cpu_ptr, &hwrt->common, sizeof(hwrt->common));
229 }
230 
231 static int
232 hwrt_init_common_fw_structure(struct pvr_file *pvr_file,
233 			      struct drm_pvr_ioctl_create_hwrt_dataset_args *args,
234 			      struct pvr_hwrt_dataset *hwrt)
235 {
236 	struct drm_pvr_create_hwrt_geom_data_args *geom_data_args = &args->geom_data_args;
237 	struct pvr_device *pvr_dev = pvr_file->pvr_dev;
238 	struct pvr_rt_mtile_info info;
239 	int err;
240 
241 	err = PVR_FEATURE_VALUE(pvr_dev, tile_size_x, &info.tile_size_x);
242 	if (WARN_ON(err))
243 		return err;
244 
245 	err = PVR_FEATURE_VALUE(pvr_dev, tile_size_y, &info.tile_size_y);
246 	if (WARN_ON(err))
247 		return err;
248 
249 	info.num_tiles_x = DIV_ROUND_UP(args->width, info.tile_size_x);
250 	info.num_tiles_y = DIV_ROUND_UP(args->height, info.tile_size_y);
251 
252 	if (PVR_HAS_FEATURE(pvr_dev, simple_parameter_format_version)) {
253 		u32 parameter_format;
254 
255 		err = PVR_FEATURE_VALUE(pvr_dev, simple_parameter_format_version,
256 					&parameter_format);
257 		if (WARN_ON(err))
258 			return err;
259 
260 		WARN_ON(parameter_format != 2);
261 
262 		/*
263 		 * Set up 16 macrotiles with a multiple of 2x2 tiles per macrotile, which is
264 		 * aligned to a tile group.
265 		 */
266 		info.mtile_x[0] = DIV_ROUND_UP(info.num_tiles_x, 8) * 2;
267 		info.mtile_y[0] = DIV_ROUND_UP(info.num_tiles_y, 8) * 2;
268 		info.mtile_x[1] = 0;
269 		info.mtile_y[1] = 0;
270 		info.mtile_x[2] = 0;
271 		info.mtile_y[2] = 0;
272 		info.tile_max_x = round_up(info.num_tiles_x, 2) - 1;
273 		info.tile_max_y = round_up(info.num_tiles_y, 2) - 1;
274 	} else {
275 		/* Set up 16 macrotiles with a multiple of 4x4 tiles per macrotile. */
276 		info.mtile_x[0] = round_up(DIV_ROUND_UP(info.num_tiles_x, 4), 4);
277 		info.mtile_y[0] = round_up(DIV_ROUND_UP(info.num_tiles_y, 4), 4);
278 		info.mtile_x[1] = info.mtile_x[0] * 2;
279 		info.mtile_y[1] = info.mtile_y[0] * 2;
280 		info.mtile_x[2] = info.mtile_x[0] * 3;
281 		info.mtile_y[2] = info.mtile_y[0] * 3;
282 		info.tile_max_x = info.num_tiles_x - 1;
283 		info.tile_max_y = info.num_tiles_y - 1;
284 	}
285 
286 	hwrt->common.geom_caches_need_zeroing = false;
287 
288 	hwrt->common.isp_merge_lower_x = args->isp_merge_lower_x;
289 	hwrt->common.isp_merge_lower_y = args->isp_merge_lower_y;
290 	hwrt->common.isp_merge_upper_x = args->isp_merge_upper_x;
291 	hwrt->common.isp_merge_upper_y = args->isp_merge_upper_y;
292 	hwrt->common.isp_merge_scale_x = args->isp_merge_scale_x;
293 	hwrt->common.isp_merge_scale_y = args->isp_merge_scale_y;
294 
295 	err = get_cr_multisamplectl_val(args->samples, false,
296 					&hwrt->common.multi_sample_ctl);
297 	if (err)
298 		return err;
299 
300 	err = get_cr_multisamplectl_val(args->samples, true,
301 					&hwrt->common.flipped_multi_sample_ctl);
302 	if (err)
303 		return err;
304 
305 	hwrt->common.mtile_stride = info.mtile_x[0] * info.mtile_y[0];
306 
307 	err = get_cr_te_aa_val(pvr_dev, args->samples, &hwrt->common.teaa);
308 	if (err)
309 		return err;
310 
311 	hwrt->common.screen_pixel_max =
312 		(((args->width - 1) << ROGUE_CR_PPP_SCREEN_PIXXMAX_SHIFT) &
313 		 ~ROGUE_CR_PPP_SCREEN_PIXXMAX_CLRMSK) |
314 		(((args->height - 1) << ROGUE_CR_PPP_SCREEN_PIXYMAX_SHIFT) &
315 		 ~ROGUE_CR_PPP_SCREEN_PIXYMAX_CLRMSK);
316 
317 	hwrt->common.te_screen =
318 		((info.tile_max_x << ROGUE_CR_TE_SCREEN_XMAX_SHIFT) &
319 		 ~ROGUE_CR_TE_SCREEN_XMAX_CLRMSK) |
320 		((info.tile_max_y << ROGUE_CR_TE_SCREEN_YMAX_SHIFT) &
321 		 ~ROGUE_CR_TE_SCREEN_YMAX_CLRMSK);
322 	hwrt->common.te_mtile1 =
323 		((info.mtile_x[0] << ROGUE_CR_TE_MTILE1_X1_SHIFT) & ~ROGUE_CR_TE_MTILE1_X1_CLRMSK) |
324 		((info.mtile_x[1] << ROGUE_CR_TE_MTILE1_X2_SHIFT) & ~ROGUE_CR_TE_MTILE1_X2_CLRMSK) |
325 		((info.mtile_x[2] << ROGUE_CR_TE_MTILE1_X3_SHIFT) & ~ROGUE_CR_TE_MTILE1_X3_CLRMSK);
326 	hwrt->common.te_mtile2 =
327 		((info.mtile_y[0] << ROGUE_CR_TE_MTILE2_Y1_SHIFT) & ~ROGUE_CR_TE_MTILE2_Y1_CLRMSK) |
328 		((info.mtile_y[1] << ROGUE_CR_TE_MTILE2_Y2_SHIFT) & ~ROGUE_CR_TE_MTILE2_Y2_CLRMSK) |
329 		((info.mtile_y[2] << ROGUE_CR_TE_MTILE2_Y3_SHIFT) & ~ROGUE_CR_TE_MTILE2_Y3_CLRMSK);
330 
331 	err = get_cr_isp_mtile_size_val(pvr_dev, args->samples, &info,
332 					&hwrt->common.isp_mtile_size);
333 	if (err)
334 		return err;
335 
336 	hwrt->common.tpc_stride = geom_data_args->tpc_stride;
337 	hwrt->common.tpc_size = geom_data_args->tpc_size;
338 
339 	hwrt->common.rgn_header_size = args->region_header_size;
340 
341 	err = pvr_fw_object_create(pvr_dev, sizeof(struct rogue_fwif_hwrtdata_common),
342 				   PVR_BO_FW_FLAGS_DEVICE_UNCACHED, hwrtdata_common_init, hwrt,
343 				   &hwrt->common_fw_obj);
344 
345 	return err;
346 }
347 
348 static void
349 hwrt_fw_data_init(void *cpu_ptr, void *priv)
350 {
351 	struct pvr_hwrt_data *hwrt_data = priv;
352 
353 	memcpy(cpu_ptr, &hwrt_data->data, sizeof(hwrt_data->data));
354 }
355 
356 static int
357 hwrt_data_init_fw_structure(struct pvr_file *pvr_file,
358 			    struct pvr_hwrt_dataset *hwrt,
359 			    struct drm_pvr_ioctl_create_hwrt_dataset_args *args,
360 			    struct drm_pvr_create_hwrt_rt_data_args *rt_data_args,
361 			    struct pvr_hwrt_data *hwrt_data)
362 {
363 	struct drm_pvr_create_hwrt_geom_data_args *geom_data_args = &args->geom_data_args;
364 	struct pvr_device *pvr_dev = pvr_file->pvr_dev;
365 	struct rogue_fwif_rta_ctl *rta_ctl;
366 	int free_list_i;
367 	int err;
368 
369 	pvr_fw_object_get_fw_addr(hwrt->common_fw_obj,
370 				  &hwrt_data->data.hwrt_data_common_fw_addr);
371 
372 	for (free_list_i = 0; free_list_i < ARRAY_SIZE(hwrt->free_lists); free_list_i++) {
373 		pvr_fw_object_get_fw_addr(hwrt->free_lists[free_list_i]->fw_obj,
374 					  &hwrt_data->data.freelists_fw_addr[free_list_i]);
375 	}
376 
377 	hwrt_data->data.tail_ptrs_dev_addr = geom_data_args->tpc_dev_addr;
378 	hwrt_data->data.vheap_table_dev_addr = geom_data_args->vheap_table_dev_addr;
379 	hwrt_data->data.rtc_dev_addr = geom_data_args->rtc_dev_addr;
380 
381 	hwrt_data->data.pm_mlist_dev_addr = rt_data_args->pm_mlist_dev_addr;
382 	hwrt_data->data.macrotile_array_dev_addr = rt_data_args->macrotile_array_dev_addr;
383 	hwrt_data->data.rgn_header_dev_addr = rt_data_args->region_header_dev_addr;
384 
385 	rta_ctl = &hwrt_data->data.rta_ctl;
386 
387 	rta_ctl->render_target_index = 0;
388 	rta_ctl->active_render_targets = 0;
389 	rta_ctl->valid_render_targets_fw_addr = 0;
390 	rta_ctl->rta_num_partial_renders_fw_addr = 0;
391 	rta_ctl->max_rts = args->layers;
392 
393 	if (args->layers > 1) {
394 		err = pvr_fw_object_create(pvr_dev, args->layers * SRTC_ENTRY_SIZE,
395 					   PVR_BO_FW_FLAGS_DEVICE_UNCACHED,
396 					   NULL, NULL, &hwrt_data->srtc_obj);
397 		if (err)
398 			return err;
399 		pvr_fw_object_get_fw_addr(hwrt_data->srtc_obj,
400 					  &rta_ctl->valid_render_targets_fw_addr);
401 
402 		err = pvr_fw_object_create(pvr_dev, args->layers * RAA_ENTRY_SIZE,
403 					   PVR_BO_FW_FLAGS_DEVICE_UNCACHED,
404 					   NULL, NULL, &hwrt_data->raa_obj);
405 		if (err)
406 			goto err_put_shadow_rt_cache;
407 		pvr_fw_object_get_fw_addr(hwrt_data->raa_obj,
408 					  &rta_ctl->rta_num_partial_renders_fw_addr);
409 	}
410 
411 	err = pvr_fw_object_create(pvr_dev, sizeof(struct rogue_fwif_hwrtdata),
412 				   PVR_BO_FW_FLAGS_DEVICE_UNCACHED,
413 				   hwrt_fw_data_init, hwrt_data, &hwrt_data->fw_obj);
414 	if (err)
415 		goto err_put_raa_obj;
416 
417 	pvr_free_list_add_hwrt(hwrt->free_lists[0], hwrt_data);
418 
419 	return 0;
420 
421 err_put_raa_obj:
422 	if (args->layers > 1)
423 		pvr_fw_object_destroy(hwrt_data->raa_obj);
424 
425 err_put_shadow_rt_cache:
426 	if (args->layers > 1)
427 		pvr_fw_object_destroy(hwrt_data->srtc_obj);
428 
429 	return err;
430 }
431 
432 static void
433 hwrt_data_fini_fw_structure(struct pvr_hwrt_dataset *hwrt, int hwrt_nr)
434 {
435 	struct pvr_hwrt_data *hwrt_data = &hwrt->data[hwrt_nr];
436 
437 	pvr_free_list_remove_hwrt(hwrt->free_lists[0], hwrt_data);
438 
439 	if (hwrt->max_rts > 1) {
440 		pvr_fw_object_destroy(hwrt_data->raa_obj);
441 		pvr_fw_object_destroy(hwrt_data->srtc_obj);
442 	}
443 
444 	pvr_fw_object_destroy(hwrt_data->fw_obj);
445 }
446 
447 /**
448  * pvr_hwrt_dataset_create() - Create a new HWRT dataset
449  * @pvr_file: Pointer to pvr_file structure.
450  * @args: Creation arguments from userspace.
451  *
452  * Return:
453  *  * Pointer to new HWRT, or
454  *  * ERR_PTR(-%ENOMEM) on out of memory.
455  */
456 struct pvr_hwrt_dataset *
457 pvr_hwrt_dataset_create(struct pvr_file *pvr_file,
458 			struct drm_pvr_ioctl_create_hwrt_dataset_args *args)
459 {
460 	struct pvr_hwrt_dataset *hwrt;
461 	int err, i = 0;
462 
463 	/* Create and fill out the kernel structure */
464 	hwrt = kzalloc(sizeof(*hwrt), GFP_KERNEL);
465 
466 	if (!hwrt)
467 		return ERR_PTR(-ENOMEM);
468 
469 	err = hwrt_init_kernel_structure(pvr_file, args, hwrt);
470 	if (err < 0)
471 		goto err_free;
472 
473 	err = hwrt_init_common_fw_structure(pvr_file, args, hwrt);
474 	if (err < 0)
475 		goto err_fini_kernel_structure;
476 
477 	for (; i < ARRAY_SIZE(hwrt->data); i++) {
478 		err = hwrt_data_init_fw_structure(pvr_file, hwrt, args,
479 						  &args->rt_data_args[i],
480 						  &hwrt->data[i]);
481 		if (err < 0)
482 			goto err_fini_data_structures;
483 
484 		hwrt->data[i].hwrt_dataset = hwrt;
485 	}
486 
487 	kref_init(&hwrt->ref_count);
488 	return hwrt;
489 
490 err_fini_data_structures:
491 	while (--i >= 0)
492 		hwrt_data_fini_fw_structure(hwrt, i);
493 
494 err_fini_kernel_structure:
495 	hwrt_fini_kernel_structure(hwrt);
496 
497 err_free:
498 	kfree(hwrt);
499 
500 	return ERR_PTR(err);
501 }
502 
503 static void
504 pvr_hwrt_dataset_release(struct kref *ref_count)
505 {
506 	struct pvr_hwrt_dataset *hwrt =
507 		container_of(ref_count, struct pvr_hwrt_dataset, ref_count);
508 
509 	for (int i = ARRAY_SIZE(hwrt->data) - 1; i >= 0; i--) {
510 		WARN_ON(pvr_fw_structure_cleanup(hwrt->pvr_dev, ROGUE_FWIF_CLEANUP_HWRTDATA,
511 						 hwrt->data[i].fw_obj, 0));
512 		hwrt_data_fini_fw_structure(hwrt, i);
513 	}
514 
515 	hwrt_fini_common_fw_structure(hwrt);
516 	hwrt_fini_kernel_structure(hwrt);
517 
518 	kfree(hwrt);
519 }
520 
521 /**
522  * pvr_destroy_hwrt_datasets_for_file: Destroy any HWRT datasets associated
523  * with the given file.
524  * @pvr_file: Pointer to pvr_file structure.
525  *
526  * Removes all HWRT datasets associated with @pvr_file from the device
527  * hwrt_dataset list and drops initial references. HWRT datasets will then be
528  * destroyed once all outstanding references are dropped.
529  */
530 void pvr_destroy_hwrt_datasets_for_file(struct pvr_file *pvr_file)
531 {
532 	struct pvr_hwrt_dataset *hwrt;
533 	unsigned long handle;
534 
535 	xa_for_each(&pvr_file->hwrt_handles, handle, hwrt) {
536 		(void)hwrt;
537 		pvr_hwrt_dataset_put(xa_erase(&pvr_file->hwrt_handles, handle));
538 	}
539 }
540 
541 /**
542  * pvr_hwrt_dataset_put() - Release reference on HWRT dataset
543  * @hwrt: Pointer to HWRT dataset to release reference on
544  */
545 void
546 pvr_hwrt_dataset_put(struct pvr_hwrt_dataset *hwrt)
547 {
548 	if (hwrt)
549 		kref_put(&hwrt->ref_count, pvr_hwrt_dataset_release);
550 }
551