xref: /linux/drivers/accel/ethosu/ethosu_gem.c (revision f6e8dc9edf963dbc99085e54f6ced6da9daa6100)
1 // SPDX-License-Identifier: GPL-2.0-only or MIT
2 /* Copyright 2025 Arm, Ltd. */
3 
4 #include <linux/err.h>
5 #include <linux/slab.h>
6 
7 #include <drm/ethosu_accel.h>
8 
9 #include "ethosu_device.h"
10 #include "ethosu_gem.h"
11 
12 static void ethosu_gem_free_object(struct drm_gem_object *obj)
13 {
14 	struct ethosu_gem_object *bo = to_ethosu_bo(obj);
15 
16 	kfree(bo->info);
17 	drm_gem_free_mmap_offset(&bo->base.base);
18 	drm_gem_dma_free(&bo->base);
19 }
20 
21 static int ethosu_gem_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
22 {
23 	struct ethosu_gem_object *bo = to_ethosu_bo(obj);
24 
25 	/* Don't allow mmap on objects that have the NO_MMAP flag set. */
26 	if (bo->flags & DRM_ETHOSU_BO_NO_MMAP)
27 		return -EINVAL;
28 
29 	return drm_gem_dma_object_mmap(obj, vma);
30 }
31 
32 static const struct drm_gem_object_funcs ethosu_gem_funcs = {
33 	.free = ethosu_gem_free_object,
34 	.print_info = drm_gem_dma_object_print_info,
35 	.get_sg_table = drm_gem_dma_object_get_sg_table,
36 	.vmap = drm_gem_dma_object_vmap,
37 	.mmap = ethosu_gem_mmap,
38 	.vm_ops = &drm_gem_dma_vm_ops,
39 };
40 
41 /**
42  * ethosu_gem_create_object - Implementation of driver->gem_create_object.
43  * @ddev: DRM device
44  * @size: Size in bytes of the memory the object will reference
45  *
46  * This lets the GEM helpers allocate object structs for us, and keep
47  * our BO stats correct.
48  */
49 struct drm_gem_object *ethosu_gem_create_object(struct drm_device *ddev, size_t size)
50 {
51 	struct ethosu_gem_object *obj;
52 
53 	obj = kzalloc(sizeof(*obj), GFP_KERNEL);
54 	if (!obj)
55 		return ERR_PTR(-ENOMEM);
56 
57 	obj->base.base.funcs = &ethosu_gem_funcs;
58 	return &obj->base.base;
59 }
60 
61 /**
62  * ethosu_gem_create_with_handle() - Create a GEM object and attach it to a handle.
63  * @file: DRM file.
64  * @ddev: DRM device.
65  * @size: Size of the GEM object to allocate.
66  * @flags: Combination of drm_ethosu_bo_flags flags.
67  * @handle: Pointer holding the handle pointing to the new GEM object.
68  *
69  * Return: Zero on success
70  */
71 int ethosu_gem_create_with_handle(struct drm_file *file,
72 				  struct drm_device *ddev,
73 				  u64 *size, u32 flags, u32 *handle)
74 {
75 	struct drm_gem_dma_object *mem;
76 	struct ethosu_gem_object *bo;
77 	int ret;
78 
79 	mem = drm_gem_dma_create(ddev, *size);
80 	if (IS_ERR(mem))
81 		return PTR_ERR(mem);
82 
83 	bo = to_ethosu_bo(&mem->base);
84 	bo->flags = flags;
85 
86 	/*
87 	 * Allocate an id of idr table where the obj is registered
88 	 * and handle has the id what user can see.
89 	 */
90 	ret = drm_gem_handle_create(file, &mem->base, handle);
91 	if (!ret)
92 		*size = bo->base.base.size;
93 
94 	/* drop reference from allocate - handle holds it now. */
95 	drm_gem_object_put(&mem->base);
96 
97 	return ret;
98 }
99 
100 struct dma {
101 	s8 region;
102 	u64 len;
103 	u64 offset;
104 	s64 stride[2];
105 };
106 
107 struct dma_state {
108 	u16 size0;
109 	u16 size1;
110 	s8 mode;
111 	struct dma src;
112 	struct dma dst;
113 };
114 
115 struct buffer {
116 	u64 base;
117 	u32 length;
118 	s8 region;
119 };
120 
121 struct feat_matrix {
122 	u64 base[4];
123 	s64 stride_x;
124 	s64 stride_y;
125 	s64 stride_c;
126 	s8 region;
127 	u8 broadcast;
128 	u16 stride_kernel;
129 	u16 precision;
130 	u16 depth;
131 	u16 width;
132 	u16 width0;
133 	u16 height[3];
134 	u8 pad_top;
135 	u8 pad_left;
136 	u8 pad_bottom;
137 	u8 pad_right;
138 };
139 
140 struct cmd_state {
141 	struct dma_state dma;
142 	struct buffer scale[2];
143 	struct buffer weight[4];
144 	struct feat_matrix ofm;
145 	struct feat_matrix ifm;
146 	struct feat_matrix ifm2;
147 };
148 
149 static void cmd_state_init(struct cmd_state *st)
150 {
151 	/* Initialize to all 1s to detect missing setup */
152 	memset(st, 0xff, sizeof(*st));
153 }
154 
155 static u64 cmd_to_addr(u32 *cmd)
156 {
157 	return ((u64)((cmd[0] & 0xff0000) << 16)) | cmd[1];
158 }
159 
160 static u64 dma_length(struct ethosu_validated_cmdstream_info *info,
161 		      struct dma_state *dma_st, struct dma *dma)
162 {
163 	s8 mode = dma_st->mode;
164 	u64 len = dma->len;
165 
166 	if (mode >= 1) {
167 		len += dma->stride[0];
168 		len *= dma_st->size0;
169 	}
170 	if (mode == 2) {
171 		len += dma->stride[1];
172 		len *= dma_st->size1;
173 	}
174 	if (dma->region >= 0)
175 		info->region_size[dma->region] = max(info->region_size[dma->region],
176 						     len + dma->offset);
177 
178 	return len;
179 }
180 
181 static u64 feat_matrix_length(struct ethosu_validated_cmdstream_info *info,
182 			      struct feat_matrix *fm,
183 			      u32 x, u32 y, u32 c)
184 {
185 	u32 element_size, storage = fm->precision >> 14;
186 	int tile = 0;
187 	u64 addr;
188 
189 	if (fm->region < 0)
190 		return U64_MAX;
191 
192 	switch (storage) {
193 	case 0:
194 		if (x >= fm->width0 + 1) {
195 			x -= fm->width0 + 1;
196 			tile += 1;
197 		}
198 		if (y >= fm->height[tile] + 1) {
199 			y -= fm->height[tile] + 1;
200 			tile += 2;
201 		}
202 		break;
203 	case 1:
204 		if (y >= fm->height[1] + 1) {
205 			y -= fm->height[1] + 1;
206 			tile = 2;
207 		} else if (y >= fm->height[0] + 1) {
208 			y -= fm->height[0] + 1;
209 			tile = 1;
210 		}
211 		break;
212 	}
213 	if (fm->base[tile] == U64_MAX)
214 		return U64_MAX;
215 
216 	addr = fm->base[tile] + y * fm->stride_y;
217 
218 	switch ((fm->precision >> 6) & 0x3) { // format
219 	case 0: //nhwc:
220 		addr += x * fm->stride_x + c;
221 		break;
222 	case 1: //nhcwb16:
223 		element_size = BIT((fm->precision >> 1) & 0x3);
224 
225 		addr += (c / 16) * fm->stride_c + (16 * x + (c & 0xf)) * element_size;
226 		break;
227 	}
228 
229 	info->region_size[fm->region] = max(info->region_size[fm->region], addr + 1);
230 
231 	return addr;
232 }
233 
234 static int calc_sizes(struct drm_device *ddev,
235 		      struct ethosu_validated_cmdstream_info *info,
236 		      u16 op, struct cmd_state *st,
237 		      bool ifm, bool ifm2, bool weight, bool scale)
238 {
239 	u64 len;
240 
241 	if (ifm) {
242 		if (st->ifm.stride_kernel == U16_MAX)
243 			return -EINVAL;
244 		u32 stride_y = ((st->ifm.stride_kernel >> 8) & 0x2) +
245 			((st->ifm.stride_kernel >> 1) & 0x1) + 1;
246 		u32 stride_x = ((st->ifm.stride_kernel >> 5) & 0x2) +
247 			(st->ifm.stride_kernel & 0x1) + 1;
248 		u32 ifm_height = st->ofm.height[2] * stride_y +
249 			st->ifm.height[2] - (st->ifm.pad_top + st->ifm.pad_bottom);
250 		u32 ifm_width  = st->ofm.width * stride_x +
251 			st->ifm.width - (st->ifm.pad_left + st->ifm.pad_right);
252 
253 		len = feat_matrix_length(info, &st->ifm, ifm_width,
254 					 ifm_height, st->ifm.depth);
255 		dev_dbg(ddev->dev, "op %d: IFM:%d:0x%llx-0x%llx\n",
256 			op, st->ifm.region, st->ifm.base[0], len);
257 		if (len == U64_MAX)
258 			return -EINVAL;
259 	}
260 
261 	if (ifm2) {
262 		len = feat_matrix_length(info, &st->ifm2, st->ifm.depth,
263 					 0, st->ofm.depth);
264 		dev_dbg(ddev->dev, "op %d: IFM2:%d:0x%llx-0x%llx\n",
265 			op, st->ifm2.region, st->ifm2.base[0], len);
266 		if (len == U64_MAX)
267 			return -EINVAL;
268 	}
269 
270 	if (weight) {
271 		dev_dbg(ddev->dev, "op %d: W:%d:0x%llx-0x%llx\n",
272 			op, st->weight[0].region, st->weight[0].base,
273 			st->weight[0].base + st->weight[0].length - 1);
274 		if (st->weight[0].region < 0 || st->weight[0].base == U64_MAX ||
275 		    st->weight[0].length == U32_MAX)
276 			return -EINVAL;
277 		info->region_size[st->weight[0].region] =
278 			max(info->region_size[st->weight[0].region],
279 			    st->weight[0].base + st->weight[0].length);
280 	}
281 
282 	if (scale) {
283 		dev_dbg(ddev->dev, "op %d: S:%d:0x%llx-0x%llx\n",
284 			op, st->scale[0].region, st->scale[0].base,
285 			st->scale[0].base + st->scale[0].length - 1);
286 		if (st->scale[0].region < 0 || st->scale[0].base == U64_MAX ||
287 		    st->scale[0].length == U32_MAX)
288 			return -EINVAL;
289 		info->region_size[st->scale[0].region] =
290 			max(info->region_size[st->scale[0].region],
291 			    st->scale[0].base + st->scale[0].length);
292 	}
293 
294 	len = feat_matrix_length(info, &st->ofm, st->ofm.width,
295 				 st->ofm.height[2], st->ofm.depth);
296 	dev_dbg(ddev->dev, "op %d: OFM:%d:0x%llx-0x%llx\n",
297 		op, st->ofm.region, st->ofm.base[0], len);
298 	if (len == U64_MAX)
299 		return -EINVAL;
300 	info->output_region[st->ofm.region] = true;
301 
302 	return 0;
303 }
304 
305 static int calc_sizes_elemwise(struct drm_device *ddev,
306 			       struct ethosu_validated_cmdstream_info *info,
307 			       u16 op, struct cmd_state *st,
308 			       bool ifm, bool ifm2)
309 {
310 	u32 height, width, depth;
311 	u64 len;
312 
313 	if (ifm) {
314 		height = st->ifm.broadcast & 0x1 ? 0 : st->ofm.height[2];
315 		width = st->ifm.broadcast & 0x2 ? 0 : st->ofm.width;
316 		depth = st->ifm.broadcast & 0x4 ? 0 : st->ofm.depth;
317 
318 		len = feat_matrix_length(info, &st->ifm, width,
319 					 height, depth);
320 		dev_dbg(ddev->dev, "op %d: IFM:%d:0x%llx-0x%llx\n",
321 			op, st->ifm.region, st->ifm.base[0], len);
322 		if (len == U64_MAX)
323 			return -EINVAL;
324 	}
325 
326 	if (ifm2) {
327 		height = st->ifm2.broadcast & 0x1 ? 0 : st->ofm.height[2];
328 		width = st->ifm2.broadcast & 0x2 ? 0 : st->ofm.width;
329 		depth = st->ifm2.broadcast & 0x4 ? 0 : st->ofm.depth;
330 
331 		len = feat_matrix_length(info, &st->ifm2, width,
332 					 height, depth);
333 		dev_dbg(ddev->dev, "op %d: IFM2:%d:0x%llx-0x%llx\n",
334 			op, st->ifm2.region, st->ifm2.base[0], len);
335 		if (len == U64_MAX)
336 			return -EINVAL;
337 	}
338 
339 	len = feat_matrix_length(info, &st->ofm, st->ofm.width,
340 				 st->ofm.height[2], st->ofm.depth);
341 	dev_dbg(ddev->dev, "op %d: OFM:%d:0x%llx-0x%llx\n",
342 		op, st->ofm.region, st->ofm.base[0], len);
343 	if (len == U64_MAX)
344 		return -EINVAL;
345 	info->output_region[st->ofm.region] = true;
346 
347 	return 0;
348 }
349 
350 static int ethosu_gem_cmdstream_copy_and_validate(struct drm_device *ddev,
351 						  u32 __user *ucmds,
352 						  struct ethosu_gem_object *bo,
353 						  u32 size)
354 {
355 	struct ethosu_validated_cmdstream_info __free(kfree) *info = kzalloc(sizeof(*info), GFP_KERNEL);
356 	struct ethosu_device *edev = to_ethosu_device(ddev);
357 	u32 *bocmds = bo->base.vaddr;
358 	struct cmd_state st;
359 	int i, ret;
360 
361 	if (!info)
362 		return -ENOMEM;
363 	info->cmd_size = size;
364 
365 	cmd_state_init(&st);
366 
367 	for (i = 0; i < size / 4; i++) {
368 		bool use_ifm, use_ifm2, use_scale;
369 		u64 dstlen, srclen;
370 		u16 cmd, param;
371 		u32 cmds[2];
372 		u64 addr;
373 
374 		if (get_user(cmds[0], ucmds++))
375 			return -EFAULT;
376 
377 		bocmds[i] = cmds[0];
378 
379 		cmd = cmds[0];
380 		param = cmds[0] >> 16;
381 
382 		if (cmd & 0x4000) {
383 			if (get_user(cmds[1], ucmds++))
384 				return -EFAULT;
385 
386 			i++;
387 			bocmds[i] = cmds[1];
388 			addr = cmd_to_addr(cmds);
389 		}
390 
391 		switch (cmd) {
392 		case NPU_OP_DMA_START:
393 			srclen = dma_length(info, &st.dma, &st.dma.src);
394 			dstlen = dma_length(info, &st.dma, &st.dma.dst);
395 
396 			if (st.dma.dst.region >= 0)
397 				info->output_region[st.dma.dst.region] = true;
398 			dev_dbg(ddev->dev, "cmd: DMA SRC:%d:0x%llx+0x%llx DST:%d:0x%llx+0x%llx\n",
399 				st.dma.src.region, st.dma.src.offset, srclen,
400 				st.dma.dst.region, st.dma.dst.offset, dstlen);
401 			break;
402 		case NPU_OP_CONV:
403 		case NPU_OP_DEPTHWISE:
404 			use_ifm2 = param & 0x1;  // weights_ifm2
405 			use_scale = !(st.ofm.precision & 0x100);
406 			ret = calc_sizes(ddev, info, cmd, &st, true, use_ifm2,
407 					 !use_ifm2, use_scale);
408 			if (ret)
409 				return ret;
410 			break;
411 		case NPU_OP_POOL:
412 			use_ifm = param != 0x4;  // pooling mode
413 			use_scale = !(st.ofm.precision & 0x100);
414 			ret = calc_sizes(ddev, info, cmd, &st, use_ifm, false,
415 					 false, use_scale);
416 			if (ret)
417 				return ret;
418 			break;
419 		case NPU_OP_ELEMENTWISE:
420 			use_ifm2 = !((st.ifm2.broadcast == 8) || (param == 5) ||
421 				(param == 6) || (param == 7) || (param == 0x24));
422 			use_ifm = st.ifm.broadcast != 8;
423 			ret = calc_sizes_elemwise(ddev, info, cmd, &st, use_ifm, use_ifm2);
424 			if (ret)
425 				return ret;
426 			break;
427 		case NPU_OP_RESIZE: // U85 only
428 			WARN_ON(1); // TODO
429 			break;
430 		case NPU_SET_KERNEL_WIDTH_M1:
431 			st.ifm.width = param;
432 			break;
433 		case NPU_SET_KERNEL_HEIGHT_M1:
434 			st.ifm.height[2] = param;
435 			break;
436 		case NPU_SET_KERNEL_STRIDE:
437 			st.ifm.stride_kernel = param;
438 			break;
439 		case NPU_SET_IFM_PAD_TOP:
440 			st.ifm.pad_top = param & 0x7f;
441 			break;
442 		case NPU_SET_IFM_PAD_LEFT:
443 			st.ifm.pad_left = param & 0x7f;
444 			break;
445 		case NPU_SET_IFM_PAD_RIGHT:
446 			st.ifm.pad_right = param & 0xff;
447 			break;
448 		case NPU_SET_IFM_PAD_BOTTOM:
449 			st.ifm.pad_bottom = param & 0xff;
450 			break;
451 		case NPU_SET_IFM_DEPTH_M1:
452 			st.ifm.depth = param;
453 			break;
454 		case NPU_SET_IFM_PRECISION:
455 			st.ifm.precision = param;
456 			break;
457 		case NPU_SET_IFM_BROADCAST:
458 			st.ifm.broadcast = param;
459 			break;
460 		case NPU_SET_IFM_REGION:
461 			st.ifm.region = param & 0x7f;
462 			break;
463 		case NPU_SET_IFM_WIDTH0_M1:
464 			st.ifm.width0 = param;
465 			break;
466 		case NPU_SET_IFM_HEIGHT0_M1:
467 			st.ifm.height[0] = param;
468 			break;
469 		case NPU_SET_IFM_HEIGHT1_M1:
470 			st.ifm.height[1] = param;
471 			break;
472 		case NPU_SET_IFM_BASE0:
473 		case NPU_SET_IFM_BASE1:
474 		case NPU_SET_IFM_BASE2:
475 		case NPU_SET_IFM_BASE3:
476 			st.ifm.base[cmd & 0x3] = addr;
477 			break;
478 		case NPU_SET_IFM_STRIDE_X:
479 			st.ifm.stride_x = addr;
480 			break;
481 		case NPU_SET_IFM_STRIDE_Y:
482 			st.ifm.stride_y = addr;
483 			break;
484 		case NPU_SET_IFM_STRIDE_C:
485 			st.ifm.stride_c = addr;
486 			break;
487 
488 		case NPU_SET_OFM_WIDTH_M1:
489 			st.ofm.width = param;
490 			break;
491 		case NPU_SET_OFM_HEIGHT_M1:
492 			st.ofm.height[2] = param;
493 			break;
494 		case NPU_SET_OFM_DEPTH_M1:
495 			st.ofm.depth = param;
496 			break;
497 		case NPU_SET_OFM_PRECISION:
498 			st.ofm.precision = param;
499 			break;
500 		case NPU_SET_OFM_REGION:
501 			st.ofm.region = param & 0x7;
502 			break;
503 		case NPU_SET_OFM_WIDTH0_M1:
504 			st.ofm.width0 = param;
505 			break;
506 		case NPU_SET_OFM_HEIGHT0_M1:
507 			st.ofm.height[0] = param;
508 			break;
509 		case NPU_SET_OFM_HEIGHT1_M1:
510 			st.ofm.height[1] = param;
511 			break;
512 		case NPU_SET_OFM_BASE0:
513 		case NPU_SET_OFM_BASE1:
514 		case NPU_SET_OFM_BASE2:
515 		case NPU_SET_OFM_BASE3:
516 			st.ofm.base[cmd & 0x3] = addr;
517 			break;
518 		case NPU_SET_OFM_STRIDE_X:
519 			st.ofm.stride_x = addr;
520 			break;
521 		case NPU_SET_OFM_STRIDE_Y:
522 			st.ofm.stride_y = addr;
523 			break;
524 		case NPU_SET_OFM_STRIDE_C:
525 			st.ofm.stride_c = addr;
526 			break;
527 
528 		case NPU_SET_IFM2_BROADCAST:
529 			st.ifm2.broadcast = param;
530 			break;
531 		case NPU_SET_IFM2_PRECISION:
532 			st.ifm2.precision = param;
533 			break;
534 		case NPU_SET_IFM2_REGION:
535 			st.ifm2.region = param & 0x7;
536 			break;
537 		case NPU_SET_IFM2_WIDTH0_M1:
538 			st.ifm2.width0 = param;
539 			break;
540 		case NPU_SET_IFM2_HEIGHT0_M1:
541 			st.ifm2.height[0] = param;
542 			break;
543 		case NPU_SET_IFM2_HEIGHT1_M1:
544 			st.ifm2.height[1] = param;
545 			break;
546 		case NPU_SET_IFM2_BASE0:
547 		case NPU_SET_IFM2_BASE1:
548 		case NPU_SET_IFM2_BASE2:
549 		case NPU_SET_IFM2_BASE3:
550 			st.ifm2.base[cmd & 0x3] = addr;
551 			break;
552 		case NPU_SET_IFM2_STRIDE_X:
553 			st.ifm2.stride_x = addr;
554 			break;
555 		case NPU_SET_IFM2_STRIDE_Y:
556 			st.ifm2.stride_y = addr;
557 			break;
558 		case NPU_SET_IFM2_STRIDE_C:
559 			st.ifm2.stride_c = addr;
560 			break;
561 
562 		case NPU_SET_WEIGHT_REGION:
563 			st.weight[0].region = param & 0x7;
564 			break;
565 		case NPU_SET_SCALE_REGION:
566 			st.scale[0].region = param & 0x7;
567 			break;
568 		case NPU_SET_WEIGHT_BASE:
569 			st.weight[0].base = addr;
570 			break;
571 		case NPU_SET_WEIGHT_LENGTH:
572 			st.weight[0].length = cmds[1];
573 			break;
574 		case NPU_SET_SCALE_BASE:
575 			st.scale[0].base = addr;
576 			break;
577 		case NPU_SET_SCALE_LENGTH:
578 			st.scale[0].length = cmds[1];
579 			break;
580 		case NPU_SET_WEIGHT1_BASE:
581 			st.weight[1].base = addr;
582 			break;
583 		case NPU_SET_WEIGHT1_LENGTH:
584 			st.weight[1].length = cmds[1];
585 			break;
586 		case NPU_SET_SCALE1_BASE: // NPU_SET_WEIGHT2_BASE (U85)
587 			if (ethosu_is_u65(edev))
588 				st.scale[1].base = addr;
589 			else
590 				st.weight[2].base = addr;
591 			break;
592 		case NPU_SET_SCALE1_LENGTH: // NPU_SET_WEIGHT2_LENGTH (U85)
593 			if (ethosu_is_u65(edev))
594 				st.scale[1].length = cmds[1];
595 			else
596 				st.weight[1].length = cmds[1];
597 			break;
598 		case NPU_SET_WEIGHT3_BASE:
599 			st.weight[3].base = addr;
600 			break;
601 		case NPU_SET_WEIGHT3_LENGTH:
602 			st.weight[3].length = cmds[1];
603 			break;
604 
605 		case NPU_SET_DMA0_SRC_REGION:
606 			if (param & 0x100)
607 				st.dma.src.region = -1;
608 			else
609 				st.dma.src.region = param & 0x7;
610 			st.dma.mode = (param >> 9) & 0x3;
611 			break;
612 		case NPU_SET_DMA0_DST_REGION:
613 			if (param & 0x100)
614 				st.dma.dst.region = -1;
615 			else
616 				st.dma.dst.region = param & 0x7;
617 			break;
618 		case NPU_SET_DMA0_SIZE0:
619 			st.dma.size0 = param;
620 			break;
621 		case NPU_SET_DMA0_SIZE1:
622 			st.dma.size1 = param;
623 			break;
624 		case NPU_SET_DMA0_SRC_STRIDE0:
625 			st.dma.src.stride[0] = ((s64)addr << 24) >> 24;
626 			break;
627 		case NPU_SET_DMA0_SRC_STRIDE1:
628 			st.dma.src.stride[1] = ((s64)addr << 24) >> 24;
629 			break;
630 		case NPU_SET_DMA0_DST_STRIDE0:
631 			st.dma.dst.stride[0] = ((s64)addr << 24) >> 24;
632 			break;
633 		case NPU_SET_DMA0_DST_STRIDE1:
634 			st.dma.dst.stride[1] = ((s64)addr << 24) >> 24;
635 			break;
636 		case NPU_SET_DMA0_SRC:
637 			st.dma.src.offset = addr;
638 			break;
639 		case NPU_SET_DMA0_DST:
640 			st.dma.dst.offset = addr;
641 			break;
642 		case NPU_SET_DMA0_LEN:
643 			st.dma.src.len = st.dma.dst.len = addr;
644 			break;
645 		default:
646 			break;
647 		}
648 	}
649 
650 	for (i = 0; i < NPU_BASEP_REGION_MAX; i++) {
651 		if (!info->region_size[i])
652 			continue;
653 		dev_dbg(ddev->dev, "region %d max size: 0x%llx\n",
654 			i, info->region_size[i]);
655 	}
656 
657 	bo->info = no_free_ptr(info);
658 	return 0;
659 }
660 
661 /**
662  * ethosu_gem_cmdstream_create() - Create a GEM object and attach it to a handle.
663  * @file: DRM file.
664  * @ddev: DRM device.
665  * @exclusive_vm: Exclusive VM. Not NULL if the GEM object can't be shared.
666  * @size: Size of the GEM object to allocate.
667  * @flags: Combination of drm_ethosu_bo_flags flags.
668  * @handle: Pointer holding the handle pointing to the new GEM object.
669  *
670  * Return: Zero on success
671  */
672 int ethosu_gem_cmdstream_create(struct drm_file *file,
673 				struct drm_device *ddev,
674 				u32 size, u64 data, u32 flags, u32 *handle)
675 {
676 	int ret;
677 	struct drm_gem_dma_object *mem;
678 	struct ethosu_gem_object *bo;
679 
680 	mem = drm_gem_dma_create(ddev, size);
681 	if (IS_ERR(mem))
682 		return PTR_ERR(mem);
683 
684 	bo = to_ethosu_bo(&mem->base);
685 	bo->flags = flags;
686 
687 	ret = ethosu_gem_cmdstream_copy_and_validate(ddev,
688 						     (void __user *)(uintptr_t)data,
689 						     bo, size);
690 	if (ret)
691 		goto fail;
692 
693 	/*
694 	 * Allocate an id of idr table where the obj is registered
695 	 * and handle has the id what user can see.
696 	 */
697 	ret = drm_gem_handle_create(file, &mem->base, handle);
698 
699 fail:
700 	/* drop reference from allocate - handle holds it now. */
701 	drm_gem_object_put(&mem->base);
702 
703 	return ret;
704 }
705