xref: /linux/drivers/gpu/drm/panthor/panthor_drv.c (revision 663269cebc7157e487400c4aeee1f765546a9c98)
1 // SPDX-License-Identifier: GPL-2.0 or MIT
2 /* Copyright 2018 Marty E. Plummer <hanetzer@startmail.com> */
3 /* Copyright 2019 Linaro, Ltd., Rob Herring <robh@kernel.org> */
4 /* Copyright 2019 Collabora ltd. */
5 
6 #ifdef CONFIG_ARM_ARCH_TIMER
7 #include <asm/arch_timer.h>
8 #endif
9 
10 #include <linux/list.h>
11 #include <linux/module.h>
12 #include <linux/of_platform.h>
13 #include <linux/pagemap.h>
14 #include <linux/platform_device.h>
15 #include <linux/pm_runtime.h>
16 
17 #include <drm/drm_auth.h>
18 #include <drm/drm_debugfs.h>
19 #include <drm/drm_drv.h>
20 #include <drm/drm_exec.h>
21 #include <drm/drm_ioctl.h>
22 #include <drm/drm_syncobj.h>
23 #include <drm/drm_utils.h>
24 #include <drm/gpu_scheduler.h>
25 #include <drm/panthor_drm.h>
26 
27 #include "panthor_device.h"
28 #include "panthor_fw.h"
29 #include "panthor_gem.h"
30 #include "panthor_gpu.h"
31 #include "panthor_heap.h"
32 #include "panthor_mmu.h"
33 #include "panthor_regs.h"
34 #include "panthor_sched.h"
35 
36 /**
37  * DOC: user <-> kernel object copy helpers.
38  */
39 
40 /**
41  * panthor_set_uobj() - Copy kernel object to user object.
42  * @usr_ptr: Users pointer.
43  * @usr_size: Size of the user object.
44  * @min_size: Minimum size for this object.
45  * @kern_size: Size of the kernel object.
46  * @in: Address of the kernel object to copy.
47  *
48  * Helper automating kernel -> user object copies.
49  *
50  * Don't use this function directly, use PANTHOR_UOBJ_SET() instead.
51  *
52  * Return: 0 on success, a negative error code otherwise.
53  */
54 static int
55 panthor_set_uobj(u64 usr_ptr, u32 usr_size, u32 min_size, u32 kern_size, const void *in)
56 {
57 	/* User size shouldn't be smaller than the minimal object size. */
58 	if (usr_size < min_size)
59 		return -EINVAL;
60 
61 	if (copy_to_user(u64_to_user_ptr(usr_ptr), in, min_t(u32, usr_size, kern_size)))
62 		return -EFAULT;
63 
64 	/* When the kernel object is smaller than the user object, we fill the gap with
65 	 * zeros.
66 	 */
67 	if (usr_size > kern_size &&
68 	    clear_user(u64_to_user_ptr(usr_ptr + kern_size), usr_size - kern_size)) {
69 		return -EFAULT;
70 	}
71 
72 	return 0;
73 }
74 
75 /**
76  * panthor_get_uobj_array() - Copy a user object array into a kernel accessible object array.
77  * @in: The object array to copy.
78  * @min_stride: Minimum array stride.
79  * @obj_size: Kernel object size.
80  *
81  * Helper automating user -> kernel object copies.
82  *
83  * Don't use this function directly, use PANTHOR_UOBJ_GET_ARRAY() instead.
84  *
85  * Return: newly allocated object array or an ERR_PTR on error.
86  */
87 static void *
88 panthor_get_uobj_array(const struct drm_panthor_obj_array *in, u32 min_stride,
89 		       u32 obj_size)
90 {
91 	int ret = 0;
92 	void *out_alloc;
93 
94 	if (!in->count)
95 		return NULL;
96 
97 	/* User stride must be at least the minimum object size, otherwise it might
98 	 * lack useful information.
99 	 */
100 	if (in->stride < min_stride)
101 		return ERR_PTR(-EINVAL);
102 
103 	out_alloc = kvmalloc_array(in->count, obj_size, GFP_KERNEL);
104 	if (!out_alloc)
105 		return ERR_PTR(-ENOMEM);
106 
107 	if (obj_size == in->stride) {
108 		/* Fast path when user/kernel have the same uAPI header version. */
109 		if (copy_from_user(out_alloc, u64_to_user_ptr(in->array),
110 				   (unsigned long)obj_size * in->count))
111 			ret = -EFAULT;
112 	} else {
113 		void __user *in_ptr = u64_to_user_ptr(in->array);
114 		void *out_ptr = out_alloc;
115 
116 		/* If the sizes differ, we need to copy elements one by one. */
117 		for (u32 i = 0; i < in->count; i++) {
118 			ret = copy_struct_from_user(out_ptr, obj_size, in_ptr, in->stride);
119 			if (ret)
120 				break;
121 
122 			out_ptr += obj_size;
123 			in_ptr += in->stride;
124 		}
125 	}
126 
127 	if (ret) {
128 		kvfree(out_alloc);
129 		return ERR_PTR(ret);
130 	}
131 
132 	return out_alloc;
133 }
134 
135 /**
136  * PANTHOR_UOBJ_MIN_SIZE_INTERNAL() - Get the minimum user object size
137  * @_typename: Object type.
138  * @_last_mandatory_field: Last mandatory field.
139  *
140  * Get the minimum user object size based on the last mandatory field name,
141  * A.K.A, the name of the last field of the structure at the time this
142  * structure was added to the uAPI.
143  *
144  * Don't use directly, use PANTHOR_UOBJ_DECL() instead.
145  */
146 #define PANTHOR_UOBJ_MIN_SIZE_INTERNAL(_typename, _last_mandatory_field) \
147 	(offsetof(_typename, _last_mandatory_field) + \
148 	 sizeof(((_typename *)NULL)->_last_mandatory_field))
149 
150 /**
151  * PANTHOR_UOBJ_DECL() - Declare a new uAPI object whose subject to
152  * evolutions.
153  * @_typename: Object type.
154  * @_last_mandatory_field: Last mandatory field.
155  *
156  * Should be used to extend the PANTHOR_UOBJ_MIN_SIZE() list.
157  */
158 #define PANTHOR_UOBJ_DECL(_typename, _last_mandatory_field) \
159 	_typename : PANTHOR_UOBJ_MIN_SIZE_INTERNAL(_typename, _last_mandatory_field)
160 
161 /**
162  * PANTHOR_UOBJ_MIN_SIZE() - Get the minimum size of a given uAPI object
163  * @_obj_name: Object to get the minimum size of.
164  *
165  * Don't use this macro directly, it's automatically called by
166  * PANTHOR_UOBJ_{SET,GET_ARRAY}().
167  */
168 #define PANTHOR_UOBJ_MIN_SIZE(_obj_name) \
169 	_Generic(_obj_name, \
170 		 PANTHOR_UOBJ_DECL(struct drm_panthor_gpu_info, tiler_present), \
171 		 PANTHOR_UOBJ_DECL(struct drm_panthor_csif_info, pad), \
172 		 PANTHOR_UOBJ_DECL(struct drm_panthor_timestamp_info, current_timestamp), \
173 		 PANTHOR_UOBJ_DECL(struct drm_panthor_group_priorities_info, pad), \
174 		 PANTHOR_UOBJ_DECL(struct drm_panthor_sync_op, timeline_value), \
175 		 PANTHOR_UOBJ_DECL(struct drm_panthor_queue_submit, syncs), \
176 		 PANTHOR_UOBJ_DECL(struct drm_panthor_queue_create, ringbuf_size), \
177 		 PANTHOR_UOBJ_DECL(struct drm_panthor_vm_bind_op, syncs))
178 
179 /**
180  * PANTHOR_UOBJ_SET() - Copy a kernel object to a user object.
181  * @_dest_usr_ptr: User pointer to copy to.
182  * @_usr_size: Size of the user object.
183  * @_src_obj: Kernel object to copy (not a pointer).
184  *
185  * Return: 0 on success, a negative error code otherwise.
186  */
187 #define PANTHOR_UOBJ_SET(_dest_usr_ptr, _usr_size, _src_obj) \
188 	panthor_set_uobj(_dest_usr_ptr, _usr_size, \
189 			 PANTHOR_UOBJ_MIN_SIZE(_src_obj), \
190 			 sizeof(_src_obj), &(_src_obj))
191 
192 /**
193  * PANTHOR_UOBJ_GET_ARRAY() - Copy a user object array to a kernel accessible
194  * object array.
195  * @_dest_array: Local variable that will hold the newly allocated kernel
196  * object array.
197  * @_uobj_array: The drm_panthor_obj_array object describing the user object
198  * array.
199  *
200  * Return: 0 on success, a negative error code otherwise.
201  */
202 #define PANTHOR_UOBJ_GET_ARRAY(_dest_array, _uobj_array) \
203 	({ \
204 		typeof(_dest_array) _tmp; \
205 		_tmp = panthor_get_uobj_array(_uobj_array, \
206 					      PANTHOR_UOBJ_MIN_SIZE((_dest_array)[0]), \
207 					      sizeof((_dest_array)[0])); \
208 		if (!IS_ERR(_tmp)) \
209 			_dest_array = _tmp; \
210 		PTR_ERR_OR_ZERO(_tmp); \
211 	})
212 
213 /**
214  * struct panthor_sync_signal - Represent a synchronization object point to attach
215  * our job fence to.
216  *
217  * This structure is here to keep track of fences that are currently bound to
218  * a specific syncobj point.
219  *
220  * At the beginning of a job submission, the fence
221  * is retrieved from the syncobj itself, and can be NULL if no fence was attached
222  * to this point.
223  *
224  * At the end, it points to the fence of the last job that had a
225  * %DRM_PANTHOR_SYNC_OP_SIGNAL on this syncobj.
226  *
227  * With jobs being submitted in batches, the fence might change several times during
228  * the process, allowing one job to wait on a job that's part of the same submission
229  * but appears earlier in the drm_panthor_group_submit::queue_submits array.
230  */
231 struct panthor_sync_signal {
232 	/** @node: list_head to track signal ops within a submit operation */
233 	struct list_head node;
234 
235 	/** @handle: The syncobj handle. */
236 	u32 handle;
237 
238 	/**
239 	 * @point: The syncobj point.
240 	 *
241 	 * Zero for regular syncobjs, and non-zero for timeline syncobjs.
242 	 */
243 	u64 point;
244 
245 	/**
246 	 * @syncobj: The sync object pointed by @handle.
247 	 */
248 	struct drm_syncobj *syncobj;
249 
250 	/**
251 	 * @chain: Chain object used to link the new fence to an existing
252 	 * timeline syncobj.
253 	 *
254 	 * NULL for regular syncobj, non-NULL for timeline syncobjs.
255 	 */
256 	struct dma_fence_chain *chain;
257 
258 	/**
259 	 * @fence: The fence to assign to the syncobj or syncobj-point.
260 	 */
261 	struct dma_fence *fence;
262 };
263 
264 /**
265  * struct panthor_job_ctx - Job context
266  */
267 struct panthor_job_ctx {
268 	/** @job: The job that is about to be submitted to drm_sched. */
269 	struct drm_sched_job *job;
270 
271 	/** @syncops: Array of sync operations. */
272 	struct drm_panthor_sync_op *syncops;
273 
274 	/** @syncop_count: Number of sync operations. */
275 	u32 syncop_count;
276 };
277 
278 /**
279  * struct panthor_submit_ctx - Submission context
280  *
281  * Anything that's related to a submission (%DRM_IOCTL_PANTHOR_VM_BIND or
282  * %DRM_IOCTL_PANTHOR_GROUP_SUBMIT) is kept here, so we can automate the
283  * initialization and cleanup steps.
284  */
285 struct panthor_submit_ctx {
286 	/** @file: DRM file this submission happens on. */
287 	struct drm_file *file;
288 
289 	/**
290 	 * @signals: List of struct panthor_sync_signal.
291 	 *
292 	 * %DRM_PANTHOR_SYNC_OP_SIGNAL operations will be recorded here,
293 	 * and %DRM_PANTHOR_SYNC_OP_WAIT will first check if an entry
294 	 * matching the syncobj+point exists before calling
295 	 * drm_syncobj_find_fence(). This allows us to describe dependencies
296 	 * existing between jobs that are part of the same batch.
297 	 */
298 	struct list_head signals;
299 
300 	/** @jobs: Array of jobs. */
301 	struct panthor_job_ctx *jobs;
302 
303 	/** @job_count: Number of entries in the @jobs array. */
304 	u32 job_count;
305 
306 	/** @exec: drm_exec context used to acquire and prepare resv objects. */
307 	struct drm_exec exec;
308 };
309 
310 #define PANTHOR_SYNC_OP_FLAGS_MASK \
311 	(DRM_PANTHOR_SYNC_OP_HANDLE_TYPE_MASK | DRM_PANTHOR_SYNC_OP_SIGNAL)
312 
313 static bool sync_op_is_signal(const struct drm_panthor_sync_op *sync_op)
314 {
315 	return !!(sync_op->flags & DRM_PANTHOR_SYNC_OP_SIGNAL);
316 }
317 
318 static bool sync_op_is_wait(const struct drm_panthor_sync_op *sync_op)
319 {
320 	/* Note that DRM_PANTHOR_SYNC_OP_WAIT == 0 */
321 	return !(sync_op->flags & DRM_PANTHOR_SYNC_OP_SIGNAL);
322 }
323 
324 /**
325  * panthor_check_sync_op() - Check drm_panthor_sync_op fields
326  * @sync_op: The sync operation to check.
327  *
328  * Return: 0 on success, -EINVAL otherwise.
329  */
330 static int
331 panthor_check_sync_op(const struct drm_panthor_sync_op *sync_op)
332 {
333 	u8 handle_type;
334 
335 	if (sync_op->flags & ~PANTHOR_SYNC_OP_FLAGS_MASK)
336 		return -EINVAL;
337 
338 	handle_type = sync_op->flags & DRM_PANTHOR_SYNC_OP_HANDLE_TYPE_MASK;
339 	if (handle_type != DRM_PANTHOR_SYNC_OP_HANDLE_TYPE_SYNCOBJ &&
340 	    handle_type != DRM_PANTHOR_SYNC_OP_HANDLE_TYPE_TIMELINE_SYNCOBJ)
341 		return -EINVAL;
342 
343 	if (handle_type == DRM_PANTHOR_SYNC_OP_HANDLE_TYPE_SYNCOBJ &&
344 	    sync_op->timeline_value != 0)
345 		return -EINVAL;
346 
347 	return 0;
348 }
349 
350 /**
351  * panthor_sync_signal_free() - Release resources and free a panthor_sync_signal object
352  * @sig_sync: Signal object to free.
353  */
354 static void
355 panthor_sync_signal_free(struct panthor_sync_signal *sig_sync)
356 {
357 	if (!sig_sync)
358 		return;
359 
360 	drm_syncobj_put(sig_sync->syncobj);
361 	dma_fence_chain_free(sig_sync->chain);
362 	dma_fence_put(sig_sync->fence);
363 	kfree(sig_sync);
364 }
365 
366 /**
367  * panthor_submit_ctx_add_sync_signal() - Add a signal operation to a submit context
368  * @ctx: Context to add the signal operation to.
369  * @handle: Syncobj handle.
370  * @point: Syncobj point.
371  *
372  * Return: 0 on success, otherwise negative error value.
373  */
374 static int
375 panthor_submit_ctx_add_sync_signal(struct panthor_submit_ctx *ctx, u32 handle, u64 point)
376 {
377 	struct panthor_sync_signal *sig_sync;
378 	struct dma_fence *cur_fence;
379 	int ret;
380 
381 	sig_sync = kzalloc(sizeof(*sig_sync), GFP_KERNEL);
382 	if (!sig_sync)
383 		return -ENOMEM;
384 
385 	sig_sync->handle = handle;
386 	sig_sync->point = point;
387 
388 	if (point > 0) {
389 		sig_sync->chain = dma_fence_chain_alloc();
390 		if (!sig_sync->chain) {
391 			ret = -ENOMEM;
392 			goto err_free_sig_sync;
393 		}
394 	}
395 
396 	sig_sync->syncobj = drm_syncobj_find(ctx->file, handle);
397 	if (!sig_sync->syncobj) {
398 		ret = -EINVAL;
399 		goto err_free_sig_sync;
400 	}
401 
402 	/* Retrieve the current fence attached to that point. It's
403 	 * perfectly fine to get a NULL fence here, it just means there's
404 	 * no fence attached to that point yet.
405 	 */
406 	if (!drm_syncobj_find_fence(ctx->file, handle, point, 0, &cur_fence))
407 		sig_sync->fence = cur_fence;
408 
409 	list_add_tail(&sig_sync->node, &ctx->signals);
410 
411 	return 0;
412 
413 err_free_sig_sync:
414 	panthor_sync_signal_free(sig_sync);
415 	return ret;
416 }
417 
418 /**
419  * panthor_submit_ctx_search_sync_signal() - Search an existing signal operation in a
420  * submit context.
421  * @ctx: Context to search the signal operation in.
422  * @handle: Syncobj handle.
423  * @point: Syncobj point.
424  *
425  * Return: A valid panthor_sync_signal object if found, NULL otherwise.
426  */
427 static struct panthor_sync_signal *
428 panthor_submit_ctx_search_sync_signal(struct panthor_submit_ctx *ctx, u32 handle, u64 point)
429 {
430 	struct panthor_sync_signal *sig_sync;
431 
432 	list_for_each_entry(sig_sync, &ctx->signals, node) {
433 		if (handle == sig_sync->handle && point == sig_sync->point)
434 			return sig_sync;
435 	}
436 
437 	return NULL;
438 }
439 
440 /**
441  * panthor_submit_ctx_add_job() - Add a job to a submit context
442  * @ctx: Context to search the signal operation in.
443  * @idx: Index of the job in the context.
444  * @job: Job to add.
445  * @syncs: Sync operations provided by userspace.
446  *
447  * Return: 0 on success, a negative error code otherwise.
448  */
449 static int
450 panthor_submit_ctx_add_job(struct panthor_submit_ctx *ctx, u32 idx,
451 			   struct drm_sched_job *job,
452 			   const struct drm_panthor_obj_array *syncs)
453 {
454 	int ret;
455 
456 	ctx->jobs[idx].job = job;
457 
458 	ret = PANTHOR_UOBJ_GET_ARRAY(ctx->jobs[idx].syncops, syncs);
459 	if (ret)
460 		return ret;
461 
462 	ctx->jobs[idx].syncop_count = syncs->count;
463 	return 0;
464 }
465 
466 /**
467  * panthor_submit_ctx_get_sync_signal() - Search signal operation and add one if none was found.
468  * @ctx: Context to search the signal operation in.
469  * @handle: Syncobj handle.
470  * @point: Syncobj point.
471  *
472  * Return: 0 on success, a negative error code otherwise.
473  */
474 static int
475 panthor_submit_ctx_get_sync_signal(struct panthor_submit_ctx *ctx, u32 handle, u64 point)
476 {
477 	struct panthor_sync_signal *sig_sync;
478 
479 	sig_sync = panthor_submit_ctx_search_sync_signal(ctx, handle, point);
480 	if (sig_sync)
481 		return 0;
482 
483 	return panthor_submit_ctx_add_sync_signal(ctx, handle, point);
484 }
485 
486 /**
487  * panthor_submit_ctx_update_job_sync_signal_fences() - Update fences
488  * on the signal operations specified by a job.
489  * @ctx: Context to search the signal operation in.
490  * @job_idx: Index of the job to operate on.
491  *
492  * Return: 0 on success, a negative error code otherwise.
493  */
494 static int
495 panthor_submit_ctx_update_job_sync_signal_fences(struct panthor_submit_ctx *ctx,
496 						 u32 job_idx)
497 {
498 	struct panthor_device *ptdev = container_of(ctx->file->minor->dev,
499 						    struct panthor_device,
500 						    base);
501 	struct dma_fence *done_fence = &ctx->jobs[job_idx].job->s_fence->finished;
502 	const struct drm_panthor_sync_op *sync_ops = ctx->jobs[job_idx].syncops;
503 	u32 sync_op_count = ctx->jobs[job_idx].syncop_count;
504 
505 	for (u32 i = 0; i < sync_op_count; i++) {
506 		struct dma_fence *old_fence;
507 		struct panthor_sync_signal *sig_sync;
508 
509 		if (!sync_op_is_signal(&sync_ops[i]))
510 			continue;
511 
512 		sig_sync = panthor_submit_ctx_search_sync_signal(ctx, sync_ops[i].handle,
513 								 sync_ops[i].timeline_value);
514 		if (drm_WARN_ON(&ptdev->base, !sig_sync))
515 			return -EINVAL;
516 
517 		old_fence = sig_sync->fence;
518 		sig_sync->fence = dma_fence_get(done_fence);
519 		dma_fence_put(old_fence);
520 
521 		if (drm_WARN_ON(&ptdev->base, !sig_sync->fence))
522 			return -EINVAL;
523 	}
524 
525 	return 0;
526 }
527 
528 /**
529  * panthor_submit_ctx_collect_job_signal_ops() - Iterate over all job signal operations
530  * and add them to the context.
531  * @ctx: Context to search the signal operation in.
532  * @job_idx: Index of the job to operate on.
533  *
534  * Return: 0 on success, a negative error code otherwise.
535  */
536 static int
537 panthor_submit_ctx_collect_job_signal_ops(struct panthor_submit_ctx *ctx,
538 					  u32 job_idx)
539 {
540 	const struct drm_panthor_sync_op *sync_ops = ctx->jobs[job_idx].syncops;
541 	u32 sync_op_count = ctx->jobs[job_idx].syncop_count;
542 
543 	for (u32 i = 0; i < sync_op_count; i++) {
544 		int ret;
545 
546 		if (!sync_op_is_signal(&sync_ops[i]))
547 			continue;
548 
549 		ret = panthor_check_sync_op(&sync_ops[i]);
550 		if (ret)
551 			return ret;
552 
553 		ret = panthor_submit_ctx_get_sync_signal(ctx,
554 							 sync_ops[i].handle,
555 							 sync_ops[i].timeline_value);
556 		if (ret)
557 			return ret;
558 	}
559 
560 	return 0;
561 }
562 
563 /**
564  * panthor_submit_ctx_push_fences() - Iterate over the signal array, and for each entry, push
565  * the currently assigned fence to the associated syncobj.
566  * @ctx: Context to push fences on.
567  *
568  * This is the last step of a submission procedure, and is done once we know the submission
569  * is effective and job fences are guaranteed to be signaled in finite time.
570  */
571 static void
572 panthor_submit_ctx_push_fences(struct panthor_submit_ctx *ctx)
573 {
574 	struct panthor_sync_signal *sig_sync;
575 
576 	list_for_each_entry(sig_sync, &ctx->signals, node) {
577 		if (sig_sync->chain) {
578 			drm_syncobj_add_point(sig_sync->syncobj, sig_sync->chain,
579 					      sig_sync->fence, sig_sync->point);
580 			sig_sync->chain = NULL;
581 		} else {
582 			drm_syncobj_replace_fence(sig_sync->syncobj, sig_sync->fence);
583 		}
584 	}
585 }
586 
587 /**
588  * panthor_submit_ctx_add_sync_deps_to_job() - Add sync wait operations as
589  * job dependencies.
590  * @ctx: Submit context.
591  * @job_idx: Index of the job to operate on.
592  *
593  * Return: 0 on success, a negative error code otherwise.
594  */
595 static int
596 panthor_submit_ctx_add_sync_deps_to_job(struct panthor_submit_ctx *ctx,
597 					u32 job_idx)
598 {
599 	struct panthor_device *ptdev = container_of(ctx->file->minor->dev,
600 						    struct panthor_device,
601 						    base);
602 	const struct drm_panthor_sync_op *sync_ops = ctx->jobs[job_idx].syncops;
603 	struct drm_sched_job *job = ctx->jobs[job_idx].job;
604 	u32 sync_op_count = ctx->jobs[job_idx].syncop_count;
605 	int ret = 0;
606 
607 	for (u32 i = 0; i < sync_op_count; i++) {
608 		struct panthor_sync_signal *sig_sync;
609 		struct dma_fence *fence;
610 
611 		if (!sync_op_is_wait(&sync_ops[i]))
612 			continue;
613 
614 		ret = panthor_check_sync_op(&sync_ops[i]);
615 		if (ret)
616 			return ret;
617 
618 		sig_sync = panthor_submit_ctx_search_sync_signal(ctx, sync_ops[i].handle,
619 								 sync_ops[i].timeline_value);
620 		if (sig_sync) {
621 			if (drm_WARN_ON(&ptdev->base, !sig_sync->fence))
622 				return -EINVAL;
623 
624 			fence = dma_fence_get(sig_sync->fence);
625 		} else {
626 			ret = drm_syncobj_find_fence(ctx->file, sync_ops[i].handle,
627 						     sync_ops[i].timeline_value,
628 						     0, &fence);
629 			if (ret)
630 				return ret;
631 		}
632 
633 		ret = drm_sched_job_add_dependency(job, fence);
634 		if (ret)
635 			return ret;
636 	}
637 
638 	return 0;
639 }
640 
641 /**
642  * panthor_submit_ctx_collect_jobs_signal_ops() - Collect all signal operations
643  * and add them to the submit context.
644  * @ctx: Submit context.
645  *
646  * Return: 0 on success, a negative error code otherwise.
647  */
648 static int
649 panthor_submit_ctx_collect_jobs_signal_ops(struct panthor_submit_ctx *ctx)
650 {
651 	for (u32 i = 0; i < ctx->job_count; i++) {
652 		int ret;
653 
654 		ret = panthor_submit_ctx_collect_job_signal_ops(ctx, i);
655 		if (ret)
656 			return ret;
657 	}
658 
659 	return 0;
660 }
661 
662 /**
663  * panthor_submit_ctx_add_deps_and_arm_jobs() - Add jobs dependencies and arm jobs
664  * @ctx: Submit context.
665  *
666  * Must be called after the resv preparation has been taken care of.
667  *
668  * Return: 0 on success, a negative error code otherwise.
669  */
670 static int
671 panthor_submit_ctx_add_deps_and_arm_jobs(struct panthor_submit_ctx *ctx)
672 {
673 	for (u32 i = 0; i < ctx->job_count; i++) {
674 		int ret;
675 
676 		ret = panthor_submit_ctx_add_sync_deps_to_job(ctx, i);
677 		if (ret)
678 			return ret;
679 
680 		drm_sched_job_arm(ctx->jobs[i].job);
681 
682 		ret = panthor_submit_ctx_update_job_sync_signal_fences(ctx, i);
683 		if (ret)
684 			return ret;
685 	}
686 
687 	return 0;
688 }
689 
690 /**
691  * panthor_submit_ctx_push_jobs() - Push jobs to their scheduling entities.
692  * @ctx: Submit context.
693  * @upd_resvs: Callback used to update reservation objects that were previously
694  * preapred.
695  */
696 static void
697 panthor_submit_ctx_push_jobs(struct panthor_submit_ctx *ctx,
698 			     void (*upd_resvs)(struct drm_exec *, struct drm_sched_job *))
699 {
700 	for (u32 i = 0; i < ctx->job_count; i++) {
701 		upd_resvs(&ctx->exec, ctx->jobs[i].job);
702 		drm_sched_entity_push_job(ctx->jobs[i].job);
703 
704 		/* Job is owned by the scheduler now. */
705 		ctx->jobs[i].job = NULL;
706 	}
707 
708 	panthor_submit_ctx_push_fences(ctx);
709 }
710 
711 /**
712  * panthor_submit_ctx_init() - Initializes a submission context
713  * @ctx: Submit context to initialize.
714  * @file: drm_file this submission happens on.
715  * @job_count: Number of jobs that will be submitted.
716  *
717  * Return: 0 on success, a negative error code otherwise.
718  */
719 static int panthor_submit_ctx_init(struct panthor_submit_ctx *ctx,
720 				   struct drm_file *file, u32 job_count)
721 {
722 	ctx->jobs = kvmalloc_array(job_count, sizeof(*ctx->jobs),
723 				   GFP_KERNEL | __GFP_ZERO);
724 	if (!ctx->jobs)
725 		return -ENOMEM;
726 
727 	ctx->file = file;
728 	ctx->job_count = job_count;
729 	INIT_LIST_HEAD(&ctx->signals);
730 	drm_exec_init(&ctx->exec,
731 		      DRM_EXEC_INTERRUPTIBLE_WAIT | DRM_EXEC_IGNORE_DUPLICATES,
732 		      0);
733 	return 0;
734 }
735 
736 /**
737  * panthor_submit_ctx_cleanup() - Cleanup a submission context
738  * @ctx: Submit context to cleanup.
739  * @job_put: Job put callback.
740  */
741 static void panthor_submit_ctx_cleanup(struct panthor_submit_ctx *ctx,
742 				       void (*job_put)(struct drm_sched_job *))
743 {
744 	struct panthor_sync_signal *sig_sync, *tmp;
745 	unsigned long i;
746 
747 	drm_exec_fini(&ctx->exec);
748 
749 	list_for_each_entry_safe(sig_sync, tmp, &ctx->signals, node)
750 		panthor_sync_signal_free(sig_sync);
751 
752 	for (i = 0; i < ctx->job_count; i++) {
753 		job_put(ctx->jobs[i].job);
754 		kvfree(ctx->jobs[i].syncops);
755 	}
756 
757 	kvfree(ctx->jobs);
758 }
759 
760 static int panthor_query_timestamp_info(struct panthor_device *ptdev,
761 					struct drm_panthor_timestamp_info *arg)
762 {
763 	int ret;
764 
765 	ret = pm_runtime_resume_and_get(ptdev->base.dev);
766 	if (ret)
767 		return ret;
768 
769 #ifdef CONFIG_ARM_ARCH_TIMER
770 	arg->timestamp_frequency = arch_timer_get_cntfrq();
771 #else
772 	arg->timestamp_frequency = 0;
773 #endif
774 	arg->current_timestamp = panthor_gpu_read_timestamp(ptdev);
775 	arg->timestamp_offset = panthor_gpu_read_timestamp_offset(ptdev);
776 
777 	pm_runtime_put(ptdev->base.dev);
778 	return 0;
779 }
780 
781 static int group_priority_permit(struct drm_file *file,
782 				 u8 priority)
783 {
784 	/* Ensure that priority is valid */
785 	if (priority > PANTHOR_GROUP_PRIORITY_REALTIME)
786 		return -EINVAL;
787 
788 	/* Medium priority and below are always allowed */
789 	if (priority <= PANTHOR_GROUP_PRIORITY_MEDIUM)
790 		return 0;
791 
792 	/* Higher priorities require CAP_SYS_NICE or DRM_MASTER */
793 	if (capable(CAP_SYS_NICE) || drm_is_current_master(file))
794 		return 0;
795 
796 	return -EACCES;
797 }
798 
799 static void panthor_query_group_priorities_info(struct drm_file *file,
800 						struct drm_panthor_group_priorities_info *arg)
801 {
802 	int prio;
803 
804 	for (prio = PANTHOR_GROUP_PRIORITY_REALTIME; prio >= 0; prio--) {
805 		if (!group_priority_permit(file, prio))
806 			arg->allowed_mask |= BIT(prio);
807 	}
808 }
809 
810 static int panthor_ioctl_dev_query(struct drm_device *ddev, void *data, struct drm_file *file)
811 {
812 	struct panthor_device *ptdev = container_of(ddev, struct panthor_device, base);
813 	struct drm_panthor_dev_query *args = data;
814 	struct drm_panthor_timestamp_info timestamp_info;
815 	struct drm_panthor_group_priorities_info priorities_info;
816 	int ret;
817 
818 	if (!args->pointer) {
819 		switch (args->type) {
820 		case DRM_PANTHOR_DEV_QUERY_GPU_INFO:
821 			args->size = sizeof(ptdev->gpu_info);
822 			return 0;
823 
824 		case DRM_PANTHOR_DEV_QUERY_CSIF_INFO:
825 			args->size = sizeof(ptdev->csif_info);
826 			return 0;
827 
828 		case DRM_PANTHOR_DEV_QUERY_TIMESTAMP_INFO:
829 			args->size = sizeof(timestamp_info);
830 			return 0;
831 
832 		case DRM_PANTHOR_DEV_QUERY_GROUP_PRIORITIES_INFO:
833 			args->size = sizeof(priorities_info);
834 			return 0;
835 
836 		default:
837 			return -EINVAL;
838 		}
839 	}
840 
841 	switch (args->type) {
842 	case DRM_PANTHOR_DEV_QUERY_GPU_INFO:
843 		return PANTHOR_UOBJ_SET(args->pointer, args->size, ptdev->gpu_info);
844 
845 	case DRM_PANTHOR_DEV_QUERY_CSIF_INFO:
846 		return PANTHOR_UOBJ_SET(args->pointer, args->size, ptdev->csif_info);
847 
848 	case DRM_PANTHOR_DEV_QUERY_TIMESTAMP_INFO:
849 		ret = panthor_query_timestamp_info(ptdev, &timestamp_info);
850 
851 		if (ret)
852 			return ret;
853 
854 		return PANTHOR_UOBJ_SET(args->pointer, args->size, timestamp_info);
855 
856 	case DRM_PANTHOR_DEV_QUERY_GROUP_PRIORITIES_INFO:
857 		panthor_query_group_priorities_info(file, &priorities_info);
858 		return PANTHOR_UOBJ_SET(args->pointer, args->size, priorities_info);
859 
860 	default:
861 		return -EINVAL;
862 	}
863 }
864 
865 #define PANTHOR_VM_CREATE_FLAGS			0
866 
867 static int panthor_ioctl_vm_create(struct drm_device *ddev, void *data,
868 				   struct drm_file *file)
869 {
870 	struct panthor_device *ptdev = container_of(ddev, struct panthor_device, base);
871 	struct panthor_file *pfile = file->driver_priv;
872 	struct drm_panthor_vm_create *args = data;
873 	int cookie, ret;
874 
875 	if (!drm_dev_enter(ddev, &cookie))
876 		return -ENODEV;
877 
878 	ret = panthor_vm_pool_create_vm(ptdev, pfile->vms,  args);
879 	if (ret >= 0) {
880 		args->id = ret;
881 		ret = 0;
882 	}
883 
884 	drm_dev_exit(cookie);
885 	return ret;
886 }
887 
888 static int panthor_ioctl_vm_destroy(struct drm_device *ddev, void *data,
889 				    struct drm_file *file)
890 {
891 	struct panthor_file *pfile = file->driver_priv;
892 	struct drm_panthor_vm_destroy *args = data;
893 
894 	if (args->pad)
895 		return -EINVAL;
896 
897 	return panthor_vm_pool_destroy_vm(pfile->vms, args->id);
898 }
899 
900 #define PANTHOR_BO_FLAGS		DRM_PANTHOR_BO_NO_MMAP
901 
902 static int panthor_ioctl_bo_create(struct drm_device *ddev, void *data,
903 				   struct drm_file *file)
904 {
905 	struct panthor_file *pfile = file->driver_priv;
906 	struct drm_panthor_bo_create *args = data;
907 	struct panthor_vm *vm = NULL;
908 	int cookie, ret;
909 
910 	if (!drm_dev_enter(ddev, &cookie))
911 		return -ENODEV;
912 
913 	if (!args->size || args->pad ||
914 	    (args->flags & ~PANTHOR_BO_FLAGS)) {
915 		ret = -EINVAL;
916 		goto out_dev_exit;
917 	}
918 
919 	if (args->exclusive_vm_id) {
920 		vm = panthor_vm_pool_get_vm(pfile->vms, args->exclusive_vm_id);
921 		if (!vm) {
922 			ret = -EINVAL;
923 			goto out_dev_exit;
924 		}
925 	}
926 
927 	ret = panthor_gem_create_with_handle(file, ddev, vm, &args->size,
928 					     args->flags, &args->handle);
929 
930 	panthor_vm_put(vm);
931 
932 out_dev_exit:
933 	drm_dev_exit(cookie);
934 	return ret;
935 }
936 
937 static int panthor_ioctl_bo_mmap_offset(struct drm_device *ddev, void *data,
938 					struct drm_file *file)
939 {
940 	struct drm_panthor_bo_mmap_offset *args = data;
941 	struct drm_gem_object *obj;
942 	int ret;
943 
944 	if (args->pad)
945 		return -EINVAL;
946 
947 	obj = drm_gem_object_lookup(file, args->handle);
948 	if (!obj)
949 		return -ENOENT;
950 
951 	ret = drm_gem_create_mmap_offset(obj);
952 	if (ret)
953 		goto out;
954 
955 	args->offset = drm_vma_node_offset_addr(&obj->vma_node);
956 
957 out:
958 	drm_gem_object_put(obj);
959 	return ret;
960 }
961 
962 static int panthor_ioctl_group_submit(struct drm_device *ddev, void *data,
963 				      struct drm_file *file)
964 {
965 	struct panthor_file *pfile = file->driver_priv;
966 	struct drm_panthor_group_submit *args = data;
967 	struct drm_panthor_queue_submit *jobs_args;
968 	struct panthor_submit_ctx ctx;
969 	int ret = 0, cookie;
970 
971 	if (args->pad)
972 		return -EINVAL;
973 
974 	if (!drm_dev_enter(ddev, &cookie))
975 		return -ENODEV;
976 
977 	ret = PANTHOR_UOBJ_GET_ARRAY(jobs_args, &args->queue_submits);
978 	if (ret)
979 		goto out_dev_exit;
980 
981 	ret = panthor_submit_ctx_init(&ctx, file, args->queue_submits.count);
982 	if (ret)
983 		goto out_free_jobs_args;
984 
985 	/* Create jobs and attach sync operations */
986 	for (u32 i = 0; i < args->queue_submits.count; i++) {
987 		const struct drm_panthor_queue_submit *qsubmit = &jobs_args[i];
988 		struct drm_sched_job *job;
989 
990 		job = panthor_job_create(pfile, args->group_handle, qsubmit);
991 		if (IS_ERR(job)) {
992 			ret = PTR_ERR(job);
993 			goto out_cleanup_submit_ctx;
994 		}
995 
996 		ret = panthor_submit_ctx_add_job(&ctx, i, job, &qsubmit->syncs);
997 		if (ret)
998 			goto out_cleanup_submit_ctx;
999 	}
1000 
1001 	/*
1002 	 * Collect signal operations on all jobs, such that each job can pick
1003 	 * from it for its dependencies and update the fence to signal when the
1004 	 * job is submitted.
1005 	 */
1006 	ret = panthor_submit_ctx_collect_jobs_signal_ops(&ctx);
1007 	if (ret)
1008 		goto out_cleanup_submit_ctx;
1009 
1010 	/*
1011 	 * We acquire/prepare revs on all jobs before proceeding with the
1012 	 * dependency registration.
1013 	 *
1014 	 * This is solving two problems:
1015 	 * 1. drm_sched_job_arm() and drm_sched_entity_push_job() must be
1016 	 *    protected by a lock to make sure no concurrent access to the same
1017 	 *    entity get interleaved, which would mess up with the fence seqno
1018 	 *    ordering. Luckily, one of the resv being acquired is the VM resv,
1019 	 *    and a scheduling entity is only bound to a single VM. As soon as
1020 	 *    we acquire the VM resv, we should be safe.
1021 	 * 2. Jobs might depend on fences that were issued by previous jobs in
1022 	 *    the same batch, so we can't add dependencies on all jobs before
1023 	 *    arming previous jobs and registering the fence to the signal
1024 	 *    array, otherwise we might miss dependencies, or point to an
1025 	 *    outdated fence.
1026 	 */
1027 	if (args->queue_submits.count > 0) {
1028 		/* All jobs target the same group, so they also point to the same VM. */
1029 		struct panthor_vm *vm = panthor_job_vm(ctx.jobs[0].job);
1030 
1031 		drm_exec_until_all_locked(&ctx.exec) {
1032 			ret = panthor_vm_prepare_mapped_bos_resvs(&ctx.exec, vm,
1033 								  args->queue_submits.count);
1034 		}
1035 
1036 		if (ret)
1037 			goto out_cleanup_submit_ctx;
1038 	}
1039 
1040 	/*
1041 	 * Now that resvs are locked/prepared, we can iterate over each job to
1042 	 * add the dependencies, arm the job fence, register the job fence to
1043 	 * the signal array.
1044 	 */
1045 	ret = panthor_submit_ctx_add_deps_and_arm_jobs(&ctx);
1046 	if (ret)
1047 		goto out_cleanup_submit_ctx;
1048 
1049 	/* Nothing can fail after that point, so we can make our job fences
1050 	 * visible to the outside world. Push jobs and set the job fences to
1051 	 * the resv slots we reserved.  This also pushes the fences to the
1052 	 * syncobjs that are part of the signal array.
1053 	 */
1054 	panthor_submit_ctx_push_jobs(&ctx, panthor_job_update_resvs);
1055 
1056 out_cleanup_submit_ctx:
1057 	panthor_submit_ctx_cleanup(&ctx, panthor_job_put);
1058 
1059 out_free_jobs_args:
1060 	kvfree(jobs_args);
1061 
1062 out_dev_exit:
1063 	drm_dev_exit(cookie);
1064 	return ret;
1065 }
1066 
1067 static int panthor_ioctl_group_destroy(struct drm_device *ddev, void *data,
1068 				       struct drm_file *file)
1069 {
1070 	struct panthor_file *pfile = file->driver_priv;
1071 	struct drm_panthor_group_destroy *args = data;
1072 
1073 	if (args->pad)
1074 		return -EINVAL;
1075 
1076 	return panthor_group_destroy(pfile, args->group_handle);
1077 }
1078 
1079 static int panthor_ioctl_group_create(struct drm_device *ddev, void *data,
1080 				      struct drm_file *file)
1081 {
1082 	struct panthor_file *pfile = file->driver_priv;
1083 	struct drm_panthor_group_create *args = data;
1084 	struct drm_panthor_queue_create *queue_args;
1085 	int ret;
1086 
1087 	if (!args->queues.count)
1088 		return -EINVAL;
1089 
1090 	ret = PANTHOR_UOBJ_GET_ARRAY(queue_args, &args->queues);
1091 	if (ret)
1092 		return ret;
1093 
1094 	ret = group_priority_permit(file, args->priority);
1095 	if (ret)
1096 		return ret;
1097 
1098 	ret = panthor_group_create(pfile, args, queue_args);
1099 	if (ret >= 0) {
1100 		args->group_handle = ret;
1101 		ret = 0;
1102 	}
1103 
1104 	kvfree(queue_args);
1105 	return ret;
1106 }
1107 
1108 static int panthor_ioctl_group_get_state(struct drm_device *ddev, void *data,
1109 					 struct drm_file *file)
1110 {
1111 	struct panthor_file *pfile = file->driver_priv;
1112 	struct drm_panthor_group_get_state *args = data;
1113 
1114 	return panthor_group_get_state(pfile, args);
1115 }
1116 
1117 static int panthor_ioctl_tiler_heap_create(struct drm_device *ddev, void *data,
1118 					   struct drm_file *file)
1119 {
1120 	struct panthor_file *pfile = file->driver_priv;
1121 	struct drm_panthor_tiler_heap_create *args = data;
1122 	struct panthor_heap_pool *pool;
1123 	struct panthor_vm *vm;
1124 	int ret;
1125 
1126 	vm = panthor_vm_pool_get_vm(pfile->vms, args->vm_id);
1127 	if (!vm)
1128 		return -EINVAL;
1129 
1130 	pool = panthor_vm_get_heap_pool(vm, true);
1131 	if (IS_ERR(pool)) {
1132 		ret = PTR_ERR(pool);
1133 		goto out_put_vm;
1134 	}
1135 
1136 	ret = panthor_heap_create(pool,
1137 				  args->initial_chunk_count,
1138 				  args->chunk_size,
1139 				  args->max_chunks,
1140 				  args->target_in_flight,
1141 				  &args->tiler_heap_ctx_gpu_va,
1142 				  &args->first_heap_chunk_gpu_va);
1143 	if (ret < 0)
1144 		goto out_put_heap_pool;
1145 
1146 	/* Heap pools are per-VM. We combine the VM and HEAP id to make
1147 	 * a unique heap handle.
1148 	 */
1149 	args->handle = (args->vm_id << 16) | ret;
1150 	ret = 0;
1151 
1152 out_put_heap_pool:
1153 	panthor_heap_pool_put(pool);
1154 
1155 out_put_vm:
1156 	panthor_vm_put(vm);
1157 	return ret;
1158 }
1159 
1160 static int panthor_ioctl_tiler_heap_destroy(struct drm_device *ddev, void *data,
1161 					    struct drm_file *file)
1162 {
1163 	struct panthor_file *pfile = file->driver_priv;
1164 	struct drm_panthor_tiler_heap_destroy *args = data;
1165 	struct panthor_heap_pool *pool;
1166 	struct panthor_vm *vm;
1167 	int ret;
1168 
1169 	if (args->pad)
1170 		return -EINVAL;
1171 
1172 	vm = panthor_vm_pool_get_vm(pfile->vms, args->handle >> 16);
1173 	if (!vm)
1174 		return -EINVAL;
1175 
1176 	pool = panthor_vm_get_heap_pool(vm, false);
1177 	if (IS_ERR(pool)) {
1178 		ret = PTR_ERR(pool);
1179 		goto out_put_vm;
1180 	}
1181 
1182 	ret = panthor_heap_destroy(pool, args->handle & GENMASK(15, 0));
1183 	panthor_heap_pool_put(pool);
1184 
1185 out_put_vm:
1186 	panthor_vm_put(vm);
1187 	return ret;
1188 }
1189 
1190 static int panthor_ioctl_vm_bind_async(struct drm_device *ddev,
1191 				       struct drm_panthor_vm_bind *args,
1192 				       struct drm_file *file)
1193 {
1194 	struct panthor_file *pfile = file->driver_priv;
1195 	struct drm_panthor_vm_bind_op *jobs_args;
1196 	struct panthor_submit_ctx ctx;
1197 	struct panthor_vm *vm;
1198 	int ret = 0;
1199 
1200 	vm = panthor_vm_pool_get_vm(pfile->vms, args->vm_id);
1201 	if (!vm)
1202 		return -EINVAL;
1203 
1204 	ret = PANTHOR_UOBJ_GET_ARRAY(jobs_args, &args->ops);
1205 	if (ret)
1206 		goto out_put_vm;
1207 
1208 	ret = panthor_submit_ctx_init(&ctx, file, args->ops.count);
1209 	if (ret)
1210 		goto out_free_jobs_args;
1211 
1212 	for (u32 i = 0; i < args->ops.count; i++) {
1213 		struct drm_panthor_vm_bind_op *op = &jobs_args[i];
1214 		struct drm_sched_job *job;
1215 
1216 		job = panthor_vm_bind_job_create(file, vm, op);
1217 		if (IS_ERR(job)) {
1218 			ret = PTR_ERR(job);
1219 			goto out_cleanup_submit_ctx;
1220 		}
1221 
1222 		ret = panthor_submit_ctx_add_job(&ctx, i, job, &op->syncs);
1223 		if (ret)
1224 			goto out_cleanup_submit_ctx;
1225 	}
1226 
1227 	ret = panthor_submit_ctx_collect_jobs_signal_ops(&ctx);
1228 	if (ret)
1229 		goto out_cleanup_submit_ctx;
1230 
1231 	/* Prepare reservation objects for each VM_BIND job. */
1232 	drm_exec_until_all_locked(&ctx.exec) {
1233 		for (u32 i = 0; i < ctx.job_count; i++) {
1234 			ret = panthor_vm_bind_job_prepare_resvs(&ctx.exec, ctx.jobs[i].job);
1235 			drm_exec_retry_on_contention(&ctx.exec);
1236 			if (ret)
1237 				goto out_cleanup_submit_ctx;
1238 		}
1239 	}
1240 
1241 	ret = panthor_submit_ctx_add_deps_and_arm_jobs(&ctx);
1242 	if (ret)
1243 		goto out_cleanup_submit_ctx;
1244 
1245 	/* Nothing can fail after that point. */
1246 	panthor_submit_ctx_push_jobs(&ctx, panthor_vm_bind_job_update_resvs);
1247 
1248 out_cleanup_submit_ctx:
1249 	panthor_submit_ctx_cleanup(&ctx, panthor_vm_bind_job_put);
1250 
1251 out_free_jobs_args:
1252 	kvfree(jobs_args);
1253 
1254 out_put_vm:
1255 	panthor_vm_put(vm);
1256 	return ret;
1257 }
1258 
1259 static int panthor_ioctl_vm_bind_sync(struct drm_device *ddev,
1260 				      struct drm_panthor_vm_bind *args,
1261 				      struct drm_file *file)
1262 {
1263 	struct panthor_file *pfile = file->driver_priv;
1264 	struct drm_panthor_vm_bind_op *jobs_args;
1265 	struct panthor_vm *vm;
1266 	int ret;
1267 
1268 	vm = panthor_vm_pool_get_vm(pfile->vms, args->vm_id);
1269 	if (!vm)
1270 		return -EINVAL;
1271 
1272 	ret = PANTHOR_UOBJ_GET_ARRAY(jobs_args, &args->ops);
1273 	if (ret)
1274 		goto out_put_vm;
1275 
1276 	for (u32 i = 0; i < args->ops.count; i++) {
1277 		ret = panthor_vm_bind_exec_sync_op(file, vm, &jobs_args[i]);
1278 		if (ret) {
1279 			/* Update ops.count so the user knows where things failed. */
1280 			args->ops.count = i;
1281 			break;
1282 		}
1283 	}
1284 
1285 	kvfree(jobs_args);
1286 
1287 out_put_vm:
1288 	panthor_vm_put(vm);
1289 	return ret;
1290 }
1291 
1292 #define PANTHOR_VM_BIND_FLAGS DRM_PANTHOR_VM_BIND_ASYNC
1293 
1294 static int panthor_ioctl_vm_bind(struct drm_device *ddev, void *data,
1295 				 struct drm_file *file)
1296 {
1297 	struct drm_panthor_vm_bind *args = data;
1298 	int cookie, ret;
1299 
1300 	if (!drm_dev_enter(ddev, &cookie))
1301 		return -ENODEV;
1302 
1303 	if (args->flags & DRM_PANTHOR_VM_BIND_ASYNC)
1304 		ret = panthor_ioctl_vm_bind_async(ddev, args, file);
1305 	else
1306 		ret = panthor_ioctl_vm_bind_sync(ddev, args, file);
1307 
1308 	drm_dev_exit(cookie);
1309 	return ret;
1310 }
1311 
1312 static int panthor_ioctl_vm_get_state(struct drm_device *ddev, void *data,
1313 				      struct drm_file *file)
1314 {
1315 	struct panthor_file *pfile = file->driver_priv;
1316 	struct drm_panthor_vm_get_state *args = data;
1317 	struct panthor_vm *vm;
1318 
1319 	vm = panthor_vm_pool_get_vm(pfile->vms, args->vm_id);
1320 	if (!vm)
1321 		return -EINVAL;
1322 
1323 	if (panthor_vm_is_unusable(vm))
1324 		args->state = DRM_PANTHOR_VM_STATE_UNUSABLE;
1325 	else
1326 		args->state = DRM_PANTHOR_VM_STATE_USABLE;
1327 
1328 	panthor_vm_put(vm);
1329 	return 0;
1330 }
1331 
1332 static int
1333 panthor_open(struct drm_device *ddev, struct drm_file *file)
1334 {
1335 	struct panthor_device *ptdev = container_of(ddev, struct panthor_device, base);
1336 	struct panthor_file *pfile;
1337 	int ret;
1338 
1339 	if (!try_module_get(THIS_MODULE))
1340 		return -EINVAL;
1341 
1342 	pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
1343 	if (!pfile) {
1344 		ret = -ENOMEM;
1345 		goto err_put_mod;
1346 	}
1347 
1348 	pfile->ptdev = ptdev;
1349 
1350 	ret = panthor_vm_pool_create(pfile);
1351 	if (ret)
1352 		goto err_free_file;
1353 
1354 	ret = panthor_group_pool_create(pfile);
1355 	if (ret)
1356 		goto err_destroy_vm_pool;
1357 
1358 	file->driver_priv = pfile;
1359 	return 0;
1360 
1361 err_destroy_vm_pool:
1362 	panthor_vm_pool_destroy(pfile);
1363 
1364 err_free_file:
1365 	kfree(pfile);
1366 
1367 err_put_mod:
1368 	module_put(THIS_MODULE);
1369 	return ret;
1370 }
1371 
1372 static void
1373 panthor_postclose(struct drm_device *ddev, struct drm_file *file)
1374 {
1375 	struct panthor_file *pfile = file->driver_priv;
1376 
1377 	panthor_group_pool_destroy(pfile);
1378 	panthor_vm_pool_destroy(pfile);
1379 
1380 	kfree(pfile);
1381 	module_put(THIS_MODULE);
1382 }
1383 
1384 static const struct drm_ioctl_desc panthor_drm_driver_ioctls[] = {
1385 #define PANTHOR_IOCTL(n, func, flags) \
1386 	DRM_IOCTL_DEF_DRV(PANTHOR_##n, panthor_ioctl_##func, flags)
1387 
1388 	PANTHOR_IOCTL(DEV_QUERY, dev_query, DRM_RENDER_ALLOW),
1389 	PANTHOR_IOCTL(VM_CREATE, vm_create, DRM_RENDER_ALLOW),
1390 	PANTHOR_IOCTL(VM_DESTROY, vm_destroy, DRM_RENDER_ALLOW),
1391 	PANTHOR_IOCTL(VM_BIND, vm_bind, DRM_RENDER_ALLOW),
1392 	PANTHOR_IOCTL(VM_GET_STATE, vm_get_state, DRM_RENDER_ALLOW),
1393 	PANTHOR_IOCTL(BO_CREATE, bo_create, DRM_RENDER_ALLOW),
1394 	PANTHOR_IOCTL(BO_MMAP_OFFSET, bo_mmap_offset, DRM_RENDER_ALLOW),
1395 	PANTHOR_IOCTL(GROUP_CREATE, group_create, DRM_RENDER_ALLOW),
1396 	PANTHOR_IOCTL(GROUP_DESTROY, group_destroy, DRM_RENDER_ALLOW),
1397 	PANTHOR_IOCTL(GROUP_GET_STATE, group_get_state, DRM_RENDER_ALLOW),
1398 	PANTHOR_IOCTL(TILER_HEAP_CREATE, tiler_heap_create, DRM_RENDER_ALLOW),
1399 	PANTHOR_IOCTL(TILER_HEAP_DESTROY, tiler_heap_destroy, DRM_RENDER_ALLOW),
1400 	PANTHOR_IOCTL(GROUP_SUBMIT, group_submit, DRM_RENDER_ALLOW),
1401 };
1402 
1403 static int panthor_mmap(struct file *filp, struct vm_area_struct *vma)
1404 {
1405 	struct drm_file *file = filp->private_data;
1406 	struct panthor_file *pfile = file->driver_priv;
1407 	struct panthor_device *ptdev = pfile->ptdev;
1408 	u64 offset = (u64)vma->vm_pgoff << PAGE_SHIFT;
1409 	int ret, cookie;
1410 
1411 	if (!drm_dev_enter(file->minor->dev, &cookie))
1412 		return -ENODEV;
1413 
1414 #ifdef CONFIG_ARM64
1415 	/*
1416 	 * With 32-bit systems being limited by the 32-bit representation of
1417 	 * mmap2's pgoffset field, we need to make the MMIO offset arch
1418 	 * specific. This converts a user MMIO offset into something the kernel
1419 	 * driver understands.
1420 	 */
1421 	if (test_tsk_thread_flag(current, TIF_32BIT) &&
1422 	    offset >= DRM_PANTHOR_USER_MMIO_OFFSET_32BIT) {
1423 		offset += DRM_PANTHOR_USER_MMIO_OFFSET_64BIT -
1424 			  DRM_PANTHOR_USER_MMIO_OFFSET_32BIT;
1425 		vma->vm_pgoff = offset >> PAGE_SHIFT;
1426 	}
1427 #endif
1428 
1429 	if (offset >= DRM_PANTHOR_USER_MMIO_OFFSET)
1430 		ret = panthor_device_mmap_io(ptdev, vma);
1431 	else
1432 		ret = drm_gem_mmap(filp, vma);
1433 
1434 	drm_dev_exit(cookie);
1435 	return ret;
1436 }
1437 
1438 static const struct file_operations panthor_drm_driver_fops = {
1439 	.open = drm_open,
1440 	.release = drm_release,
1441 	.unlocked_ioctl = drm_ioctl,
1442 	.compat_ioctl = drm_compat_ioctl,
1443 	.poll = drm_poll,
1444 	.read = drm_read,
1445 	.llseek = noop_llseek,
1446 	.mmap = panthor_mmap,
1447 };
1448 
1449 #ifdef CONFIG_DEBUG_FS
1450 static void panthor_debugfs_init(struct drm_minor *minor)
1451 {
1452 	panthor_mmu_debugfs_init(minor);
1453 }
1454 #endif
1455 
1456 /*
1457  * PanCSF driver version:
1458  * - 1.0 - initial interface
1459  * - 1.1 - adds DEV_QUERY_TIMESTAMP_INFO query
1460  * - 1.2 - adds DEV_QUERY_GROUP_PRIORITIES_INFO query
1461  *       - adds PANTHOR_GROUP_PRIORITY_REALTIME priority
1462  */
1463 static const struct drm_driver panthor_drm_driver = {
1464 	.driver_features = DRIVER_RENDER | DRIVER_GEM | DRIVER_SYNCOBJ |
1465 			   DRIVER_SYNCOBJ_TIMELINE | DRIVER_GEM_GPUVA,
1466 	.open = panthor_open,
1467 	.postclose = panthor_postclose,
1468 	.ioctls = panthor_drm_driver_ioctls,
1469 	.num_ioctls = ARRAY_SIZE(panthor_drm_driver_ioctls),
1470 	.fops = &panthor_drm_driver_fops,
1471 	.name = "panthor",
1472 	.desc = "Panthor DRM driver",
1473 	.date = "20230801",
1474 	.major = 1,
1475 	.minor = 2,
1476 
1477 	.gem_create_object = panthor_gem_create_object,
1478 	.gem_prime_import_sg_table = drm_gem_shmem_prime_import_sg_table,
1479 #ifdef CONFIG_DEBUG_FS
1480 	.debugfs_init = panthor_debugfs_init,
1481 #endif
1482 };
1483 
1484 static int panthor_probe(struct platform_device *pdev)
1485 {
1486 	struct panthor_device *ptdev;
1487 
1488 	ptdev = devm_drm_dev_alloc(&pdev->dev, &panthor_drm_driver,
1489 				   struct panthor_device, base);
1490 	if (IS_ERR(ptdev))
1491 		return -ENOMEM;
1492 
1493 	platform_set_drvdata(pdev, ptdev);
1494 
1495 	return panthor_device_init(ptdev);
1496 }
1497 
1498 static void panthor_remove(struct platform_device *pdev)
1499 {
1500 	struct panthor_device *ptdev = platform_get_drvdata(pdev);
1501 
1502 	panthor_device_unplug(ptdev);
1503 }
1504 
1505 static const struct of_device_id dt_match[] = {
1506 	{ .compatible = "rockchip,rk3588-mali" },
1507 	{ .compatible = "arm,mali-valhall-csf" },
1508 	{}
1509 };
1510 MODULE_DEVICE_TABLE(of, dt_match);
1511 
1512 static DEFINE_RUNTIME_DEV_PM_OPS(panthor_pm_ops,
1513 				 panthor_device_suspend,
1514 				 panthor_device_resume,
1515 				 NULL);
1516 
1517 static struct platform_driver panthor_driver = {
1518 	.probe = panthor_probe,
1519 	.remove_new = panthor_remove,
1520 	.driver = {
1521 		.name = "panthor",
1522 		.pm = pm_ptr(&panthor_pm_ops),
1523 		.of_match_table = dt_match,
1524 	},
1525 };
1526 
1527 /*
1528  * Workqueue used to cleanup stuff.
1529  *
1530  * We create a dedicated workqueue so we can drain on unplug and
1531  * make sure all resources are freed before the module is unloaded.
1532  */
1533 struct workqueue_struct *panthor_cleanup_wq;
1534 
1535 static int __init panthor_init(void)
1536 {
1537 	int ret;
1538 
1539 	ret = panthor_mmu_pt_cache_init();
1540 	if (ret)
1541 		return ret;
1542 
1543 	panthor_cleanup_wq = alloc_workqueue("panthor-cleanup", WQ_UNBOUND, 0);
1544 	if (!panthor_cleanup_wq) {
1545 		pr_err("panthor: Failed to allocate the workqueues");
1546 		ret = -ENOMEM;
1547 		goto err_mmu_pt_cache_fini;
1548 	}
1549 
1550 	ret = platform_driver_register(&panthor_driver);
1551 	if (ret)
1552 		goto err_destroy_cleanup_wq;
1553 
1554 	return 0;
1555 
1556 err_destroy_cleanup_wq:
1557 	destroy_workqueue(panthor_cleanup_wq);
1558 
1559 err_mmu_pt_cache_fini:
1560 	panthor_mmu_pt_cache_fini();
1561 	return ret;
1562 }
1563 module_init(panthor_init);
1564 
1565 static void __exit panthor_exit(void)
1566 {
1567 	platform_driver_unregister(&panthor_driver);
1568 	destroy_workqueue(panthor_cleanup_wq);
1569 	panthor_mmu_pt_cache_fini();
1570 }
1571 module_exit(panthor_exit);
1572 
1573 MODULE_AUTHOR("Panthor Project Developers");
1574 MODULE_DESCRIPTION("Panthor DRM Driver");
1575 MODULE_LICENSE("Dual MIT/GPL");
1576