xref: /linux/drivers/gpu/drm/imagination/pvr_job.c (revision be239684b18e1cdcafcf8c7face4a2f562c745ad)
1 // SPDX-License-Identifier: GPL-2.0-only OR MIT
2 /* Copyright (c) 2023 Imagination Technologies Ltd. */
3 
4 #include "pvr_context.h"
5 #include "pvr_device.h"
6 #include "pvr_drv.h"
7 #include "pvr_gem.h"
8 #include "pvr_hwrt.h"
9 #include "pvr_job.h"
10 #include "pvr_mmu.h"
11 #include "pvr_power.h"
12 #include "pvr_rogue_fwif.h"
13 #include "pvr_rogue_fwif_client.h"
14 #include "pvr_stream.h"
15 #include "pvr_stream_defs.h"
16 #include "pvr_sync.h"
17 
18 #include <drm/drm_exec.h>
19 #include <drm/drm_gem.h>
20 #include <linux/types.h>
21 #include <uapi/drm/pvr_drm.h>
22 
23 static void pvr_job_release(struct kref *kref)
24 {
25 	struct pvr_job *job = container_of(kref, struct pvr_job, ref_count);
26 
27 	xa_erase(&job->pvr_dev->job_ids, job->id);
28 
29 	pvr_hwrt_data_put(job->hwrt);
30 	pvr_context_put(job->ctx);
31 
32 	WARN_ON(job->paired_job);
33 
34 	pvr_queue_job_cleanup(job);
35 	pvr_job_release_pm_ref(job);
36 
37 	kfree(job->cmd);
38 	kfree(job);
39 }
40 
41 /**
42  * pvr_job_put() - Release reference on job
43  * @job: Target job.
44  */
45 void
46 pvr_job_put(struct pvr_job *job)
47 {
48 	if (job)
49 		kref_put(&job->ref_count, pvr_job_release);
50 }
51 
52 /**
53  * pvr_job_process_stream() - Build job FW structure from stream
54  * @pvr_dev: Device pointer.
55  * @cmd_defs: Stream definition.
56  * @stream: Pointer to command stream.
57  * @stream_size: Size of command stream, in bytes.
58  * @job: Pointer to job.
59  *
60  * Caller is responsible for freeing the output structure.
61  *
62  * Returns:
63  *  * 0 on success,
64  *  * -%ENOMEM on out of memory, or
65  *  * -%EINVAL on malformed stream.
66  */
67 static int
68 pvr_job_process_stream(struct pvr_device *pvr_dev, const struct pvr_stream_cmd_defs *cmd_defs,
69 		       void *stream, u32 stream_size, struct pvr_job *job)
70 {
71 	int err;
72 
73 	job->cmd = kzalloc(cmd_defs->dest_size, GFP_KERNEL);
74 	if (!job->cmd)
75 		return -ENOMEM;
76 
77 	job->cmd_len = cmd_defs->dest_size;
78 
79 	err = pvr_stream_process(pvr_dev, cmd_defs, stream, stream_size, job->cmd);
80 	if (err)
81 		kfree(job->cmd);
82 
83 	return err;
84 }
85 
86 static int pvr_fw_cmd_init(struct pvr_device *pvr_dev, struct pvr_job *job,
87 			   const struct pvr_stream_cmd_defs *stream_def,
88 			   u64 stream_userptr, u32 stream_len)
89 {
90 	void *stream;
91 	int err;
92 
93 	stream = kzalloc(stream_len, GFP_KERNEL);
94 	if (!stream)
95 		return -ENOMEM;
96 
97 	if (copy_from_user(stream, u64_to_user_ptr(stream_userptr), stream_len)) {
98 		err = -EFAULT;
99 		goto err_free_stream;
100 	}
101 
102 	err = pvr_job_process_stream(pvr_dev, stream_def, stream, stream_len, job);
103 
104 err_free_stream:
105 	kfree(stream);
106 
107 	return err;
108 }
109 
110 static u32
111 convert_geom_flags(u32 in_flags)
112 {
113 	u32 out_flags = 0;
114 
115 	if (in_flags & DRM_PVR_SUBMIT_JOB_GEOM_CMD_FIRST)
116 		out_flags |= ROGUE_GEOM_FLAGS_FIRSTKICK;
117 	if (in_flags & DRM_PVR_SUBMIT_JOB_GEOM_CMD_LAST)
118 		out_flags |= ROGUE_GEOM_FLAGS_LASTKICK;
119 	if (in_flags & DRM_PVR_SUBMIT_JOB_GEOM_CMD_SINGLE_CORE)
120 		out_flags |= ROGUE_GEOM_FLAGS_SINGLE_CORE;
121 
122 	return out_flags;
123 }
124 
125 static u32
126 convert_frag_flags(u32 in_flags)
127 {
128 	u32 out_flags = 0;
129 
130 	if (in_flags & DRM_PVR_SUBMIT_JOB_FRAG_CMD_SINGLE_CORE)
131 		out_flags |= ROGUE_FRAG_FLAGS_SINGLE_CORE;
132 	if (in_flags & DRM_PVR_SUBMIT_JOB_FRAG_CMD_DEPTHBUFFER)
133 		out_flags |= ROGUE_FRAG_FLAGS_DEPTHBUFFER;
134 	if (in_flags & DRM_PVR_SUBMIT_JOB_FRAG_CMD_STENCILBUFFER)
135 		out_flags |= ROGUE_FRAG_FLAGS_STENCILBUFFER;
136 	if (in_flags & DRM_PVR_SUBMIT_JOB_FRAG_CMD_PREVENT_CDM_OVERLAP)
137 		out_flags |= ROGUE_FRAG_FLAGS_PREVENT_CDM_OVERLAP;
138 	if (in_flags & DRM_PVR_SUBMIT_JOB_FRAG_CMD_SCRATCHBUFFER)
139 		out_flags |= ROGUE_FRAG_FLAGS_SCRATCHBUFFER;
140 	if (in_flags & DRM_PVR_SUBMIT_JOB_FRAG_CMD_GET_VIS_RESULTS)
141 		out_flags |= ROGUE_FRAG_FLAGS_GET_VIS_RESULTS;
142 	if (in_flags & DRM_PVR_SUBMIT_JOB_FRAG_CMD_DISABLE_PIXELMERGE)
143 		out_flags |= ROGUE_FRAG_FLAGS_DISABLE_PIXELMERGE;
144 
145 	return out_flags;
146 }
147 
148 static int
149 pvr_geom_job_fw_cmd_init(struct pvr_job *job,
150 			 struct drm_pvr_job *args)
151 {
152 	struct rogue_fwif_cmd_geom *cmd;
153 	int err;
154 
155 	if (args->flags & ~DRM_PVR_SUBMIT_JOB_GEOM_CMD_FLAGS_MASK)
156 		return -EINVAL;
157 
158 	if (job->ctx->type != DRM_PVR_CTX_TYPE_RENDER)
159 		return -EINVAL;
160 
161 	if (!job->hwrt)
162 		return -EINVAL;
163 
164 	job->fw_ccb_cmd_type = ROGUE_FWIF_CCB_CMD_TYPE_GEOM;
165 	err = pvr_fw_cmd_init(job->pvr_dev, job, &pvr_cmd_geom_stream,
166 			      args->cmd_stream, args->cmd_stream_len);
167 	if (err)
168 		return err;
169 
170 	cmd = job->cmd;
171 	cmd->cmd_shared.cmn.frame_num = 0;
172 	cmd->flags = convert_geom_flags(args->flags);
173 	pvr_fw_object_get_fw_addr(job->hwrt->fw_obj, &cmd->cmd_shared.hwrt_data_fw_addr);
174 	return 0;
175 }
176 
177 static int
178 pvr_frag_job_fw_cmd_init(struct pvr_job *job,
179 			 struct drm_pvr_job *args)
180 {
181 	struct rogue_fwif_cmd_frag *cmd;
182 	int err;
183 
184 	if (args->flags & ~DRM_PVR_SUBMIT_JOB_FRAG_CMD_FLAGS_MASK)
185 		return -EINVAL;
186 
187 	if (job->ctx->type != DRM_PVR_CTX_TYPE_RENDER)
188 		return -EINVAL;
189 
190 	if (!job->hwrt)
191 		return -EINVAL;
192 
193 	job->fw_ccb_cmd_type = (args->flags & DRM_PVR_SUBMIT_JOB_FRAG_CMD_PARTIAL_RENDER) ?
194 			       ROGUE_FWIF_CCB_CMD_TYPE_FRAG_PR :
195 			       ROGUE_FWIF_CCB_CMD_TYPE_FRAG;
196 	err = pvr_fw_cmd_init(job->pvr_dev, job, &pvr_cmd_frag_stream,
197 			      args->cmd_stream, args->cmd_stream_len);
198 	if (err)
199 		return err;
200 
201 	cmd = job->cmd;
202 	cmd->cmd_shared.cmn.frame_num = 0;
203 	cmd->flags = convert_frag_flags(args->flags);
204 	pvr_fw_object_get_fw_addr(job->hwrt->fw_obj, &cmd->cmd_shared.hwrt_data_fw_addr);
205 	return 0;
206 }
207 
208 static u32
209 convert_compute_flags(u32 in_flags)
210 {
211 	u32 out_flags = 0;
212 
213 	if (in_flags & DRM_PVR_SUBMIT_JOB_COMPUTE_CMD_PREVENT_ALL_OVERLAP)
214 		out_flags |= ROGUE_COMPUTE_FLAG_PREVENT_ALL_OVERLAP;
215 	if (in_flags & DRM_PVR_SUBMIT_JOB_COMPUTE_CMD_SINGLE_CORE)
216 		out_flags |= ROGUE_COMPUTE_FLAG_SINGLE_CORE;
217 
218 	return out_flags;
219 }
220 
221 static int
222 pvr_compute_job_fw_cmd_init(struct pvr_job *job,
223 			    struct drm_pvr_job *args)
224 {
225 	struct rogue_fwif_cmd_compute *cmd;
226 	int err;
227 
228 	if (args->flags & ~DRM_PVR_SUBMIT_JOB_COMPUTE_CMD_FLAGS_MASK)
229 		return -EINVAL;
230 
231 	if (job->ctx->type != DRM_PVR_CTX_TYPE_COMPUTE)
232 		return -EINVAL;
233 
234 	job->fw_ccb_cmd_type = ROGUE_FWIF_CCB_CMD_TYPE_CDM;
235 	err = pvr_fw_cmd_init(job->pvr_dev, job, &pvr_cmd_compute_stream,
236 			      args->cmd_stream, args->cmd_stream_len);
237 	if (err)
238 		return err;
239 
240 	cmd = job->cmd;
241 	cmd->common.frame_num = 0;
242 	cmd->flags = convert_compute_flags(args->flags);
243 	return 0;
244 }
245 
246 static u32
247 convert_transfer_flags(u32 in_flags)
248 {
249 	u32 out_flags = 0;
250 
251 	if (in_flags & DRM_PVR_SUBMIT_JOB_TRANSFER_CMD_SINGLE_CORE)
252 		out_flags |= ROGUE_TRANSFER_FLAGS_SINGLE_CORE;
253 
254 	return out_flags;
255 }
256 
257 static int
258 pvr_transfer_job_fw_cmd_init(struct pvr_job *job,
259 			     struct drm_pvr_job *args)
260 {
261 	struct rogue_fwif_cmd_transfer *cmd;
262 	int err;
263 
264 	if (args->flags & ~DRM_PVR_SUBMIT_JOB_TRANSFER_CMD_FLAGS_MASK)
265 		return -EINVAL;
266 
267 	if (job->ctx->type != DRM_PVR_CTX_TYPE_TRANSFER_FRAG)
268 		return -EINVAL;
269 
270 	job->fw_ccb_cmd_type = ROGUE_FWIF_CCB_CMD_TYPE_TQ_3D;
271 	err = pvr_fw_cmd_init(job->pvr_dev, job, &pvr_cmd_transfer_stream,
272 			      args->cmd_stream, args->cmd_stream_len);
273 	if (err)
274 		return err;
275 
276 	cmd = job->cmd;
277 	cmd->common.frame_num = 0;
278 	cmd->flags = convert_transfer_flags(args->flags);
279 	return 0;
280 }
281 
282 static int
283 pvr_job_fw_cmd_init(struct pvr_job *job,
284 		    struct drm_pvr_job *args)
285 {
286 	switch (args->type) {
287 	case DRM_PVR_JOB_TYPE_GEOMETRY:
288 		return pvr_geom_job_fw_cmd_init(job, args);
289 
290 	case DRM_PVR_JOB_TYPE_FRAGMENT:
291 		return pvr_frag_job_fw_cmd_init(job, args);
292 
293 	case DRM_PVR_JOB_TYPE_COMPUTE:
294 		return pvr_compute_job_fw_cmd_init(job, args);
295 
296 	case DRM_PVR_JOB_TYPE_TRANSFER_FRAG:
297 		return pvr_transfer_job_fw_cmd_init(job, args);
298 
299 	default:
300 		return -EINVAL;
301 	}
302 }
303 
304 /**
305  * struct pvr_job_data - Helper container for pairing jobs with the
306  * sync_ops supplied for them by the user.
307  */
308 struct pvr_job_data {
309 	/** @job: Pointer to the job. */
310 	struct pvr_job *job;
311 
312 	/** @sync_ops: Pointer to the sync_ops associated with @job. */
313 	struct drm_pvr_sync_op *sync_ops;
314 
315 	/** @sync_op_count: Number of members of @sync_ops. */
316 	u32 sync_op_count;
317 };
318 
319 /**
320  * prepare_job_syncs() - Prepare all sync objects for a single job.
321  * @pvr_file: PowerVR file.
322  * @job_data: Precreated job and sync_ops array.
323  * @signal_array: xarray to receive signal sync objects.
324  *
325  * Returns:
326  *  * 0 on success, or
327  *  * Any error code returned by pvr_sync_signal_array_collect_ops(),
328  *    pvr_sync_add_deps_to_job(), drm_sched_job_add_resv_dependencies() or
329  *    pvr_sync_signal_array_update_fences().
330  */
331 static int
332 prepare_job_syncs(struct pvr_file *pvr_file,
333 		  struct pvr_job_data *job_data,
334 		  struct xarray *signal_array)
335 {
336 	struct dma_fence *done_fence;
337 	int err = pvr_sync_signal_array_collect_ops(signal_array,
338 						    from_pvr_file(pvr_file),
339 						    job_data->sync_op_count,
340 						    job_data->sync_ops);
341 
342 	if (err)
343 		return err;
344 
345 	err = pvr_sync_add_deps_to_job(pvr_file, &job_data->job->base,
346 				       job_data->sync_op_count,
347 				       job_data->sync_ops, signal_array);
348 	if (err)
349 		return err;
350 
351 	if (job_data->job->hwrt) {
352 		/* The geometry job writes the HWRT region headers, which are
353 		 * then read by the fragment job.
354 		 */
355 		struct drm_gem_object *obj =
356 			gem_from_pvr_gem(job_data->job->hwrt->fw_obj->gem);
357 		enum dma_resv_usage usage =
358 			dma_resv_usage_rw(job_data->job->type ==
359 					  DRM_PVR_JOB_TYPE_GEOMETRY);
360 
361 		dma_resv_lock(obj->resv, NULL);
362 		err = drm_sched_job_add_resv_dependencies(&job_data->job->base,
363 							  obj->resv, usage);
364 		dma_resv_unlock(obj->resv);
365 		if (err)
366 			return err;
367 	}
368 
369 	/* We need to arm the job to get the job done fence. */
370 	done_fence = pvr_queue_job_arm(job_data->job);
371 
372 	err = pvr_sync_signal_array_update_fences(signal_array,
373 						  job_data->sync_op_count,
374 						  job_data->sync_ops,
375 						  done_fence);
376 	return err;
377 }
378 
379 /**
380  * prepare_job_syncs_for_each() - Prepare all sync objects for an array of jobs.
381  * @pvr_file: PowerVR file.
382  * @job_data: Array of precreated jobs and their sync_ops.
383  * @job_count: Number of jobs.
384  * @signal_array: xarray to receive signal sync objects.
385  *
386  * Returns:
387  *  * 0 on success, or
388  *  * Any error code returned by pvr_vm_bind_job_prepare_syncs().
389  */
390 static int
391 prepare_job_syncs_for_each(struct pvr_file *pvr_file,
392 			   struct pvr_job_data *job_data,
393 			   u32 *job_count,
394 			   struct xarray *signal_array)
395 {
396 	for (u32 i = 0; i < *job_count; i++) {
397 		int err = prepare_job_syncs(pvr_file, &job_data[i],
398 					    signal_array);
399 
400 		if (err) {
401 			*job_count = i;
402 			return err;
403 		}
404 	}
405 
406 	return 0;
407 }
408 
409 static struct pvr_job *
410 create_job(struct pvr_device *pvr_dev,
411 	   struct pvr_file *pvr_file,
412 	   struct drm_pvr_job *args)
413 {
414 	struct pvr_job *job = NULL;
415 	int err;
416 
417 	if (!args->cmd_stream || !args->cmd_stream_len)
418 		return ERR_PTR(-EINVAL);
419 
420 	if (args->type != DRM_PVR_JOB_TYPE_GEOMETRY &&
421 	    args->type != DRM_PVR_JOB_TYPE_FRAGMENT &&
422 	    (args->hwrt.set_handle || args->hwrt.data_index))
423 		return ERR_PTR(-EINVAL);
424 
425 	job = kzalloc(sizeof(*job), GFP_KERNEL);
426 	if (!job)
427 		return ERR_PTR(-ENOMEM);
428 
429 	kref_init(&job->ref_count);
430 	job->type = args->type;
431 	job->pvr_dev = pvr_dev;
432 
433 	err = xa_alloc(&pvr_dev->job_ids, &job->id, job, xa_limit_32b, GFP_KERNEL);
434 	if (err)
435 		goto err_put_job;
436 
437 	job->ctx = pvr_context_lookup(pvr_file, args->context_handle);
438 	if (!job->ctx) {
439 		err = -EINVAL;
440 		goto err_put_job;
441 	}
442 
443 	if (args->hwrt.set_handle) {
444 		job->hwrt = pvr_hwrt_data_lookup(pvr_file, args->hwrt.set_handle,
445 						 args->hwrt.data_index);
446 		if (!job->hwrt) {
447 			err = -EINVAL;
448 			goto err_put_job;
449 		}
450 	}
451 
452 	err = pvr_job_fw_cmd_init(job, args);
453 	if (err)
454 		goto err_put_job;
455 
456 	err = pvr_queue_job_init(job);
457 	if (err)
458 		goto err_put_job;
459 
460 	return job;
461 
462 err_put_job:
463 	pvr_job_put(job);
464 	return ERR_PTR(err);
465 }
466 
467 /**
468  * pvr_job_data_fini() - Cleanup all allocs used to set up job submission.
469  * @job_data: Job data array.
470  * @job_count: Number of members of @job_data.
471  */
472 static void
473 pvr_job_data_fini(struct pvr_job_data *job_data, u32 job_count)
474 {
475 	for (u32 i = 0; i < job_count; i++) {
476 		pvr_job_put(job_data[i].job);
477 		kvfree(job_data[i].sync_ops);
478 	}
479 }
480 
481 /**
482  * pvr_job_data_init() - Init an array of created jobs, associating them with
483  * the appropriate sync_ops args, which will be copied in.
484  * @pvr_dev: Target PowerVR device.
485  * @pvr_file: Pointer to PowerVR file structure.
486  * @job_args: Job args array copied from user.
487  * @job_count: Number of members of @job_args.
488  * @job_data_out: Job data array.
489  */
490 static int pvr_job_data_init(struct pvr_device *pvr_dev,
491 			     struct pvr_file *pvr_file,
492 			     struct drm_pvr_job *job_args,
493 			     u32 *job_count,
494 			     struct pvr_job_data *job_data_out)
495 {
496 	int err = 0, i = 0;
497 
498 	for (; i < *job_count; i++) {
499 		job_data_out[i].job =
500 			create_job(pvr_dev, pvr_file, &job_args[i]);
501 		err = PTR_ERR_OR_ZERO(job_data_out[i].job);
502 
503 		if (err) {
504 			*job_count = i;
505 			job_data_out[i].job = NULL;
506 			goto err_cleanup;
507 		}
508 
509 		err = PVR_UOBJ_GET_ARRAY(job_data_out[i].sync_ops,
510 					 &job_args[i].sync_ops);
511 		if (err) {
512 			*job_count = i;
513 
514 			/* Ensure the job created above is also cleaned up. */
515 			i++;
516 			goto err_cleanup;
517 		}
518 
519 		job_data_out[i].sync_op_count = job_args[i].sync_ops.count;
520 	}
521 
522 	return 0;
523 
524 err_cleanup:
525 	pvr_job_data_fini(job_data_out, i);
526 
527 	return err;
528 }
529 
530 static void
531 push_jobs(struct pvr_job_data *job_data, u32 job_count)
532 {
533 	for (u32 i = 0; i < job_count; i++)
534 		pvr_queue_job_push(job_data[i].job);
535 }
536 
537 static int
538 prepare_fw_obj_resv(struct drm_exec *exec, struct pvr_fw_object *fw_obj)
539 {
540 	return drm_exec_prepare_obj(exec, gem_from_pvr_gem(fw_obj->gem), 1);
541 }
542 
543 static int
544 jobs_lock_all_objs(struct drm_exec *exec, struct pvr_job_data *job_data,
545 		   u32 job_count)
546 {
547 	for (u32 i = 0; i < job_count; i++) {
548 		struct pvr_job *job = job_data[i].job;
549 
550 		/* Grab a lock on a the context, to guard against
551 		 * concurrent submission to the same queue.
552 		 */
553 		int err = drm_exec_lock_obj(exec,
554 					    gem_from_pvr_gem(job->ctx->fw_obj->gem));
555 
556 		if (err)
557 			return err;
558 
559 		if (job->hwrt) {
560 			err = prepare_fw_obj_resv(exec,
561 						  job->hwrt->fw_obj);
562 			if (err)
563 				return err;
564 		}
565 	}
566 
567 	return 0;
568 }
569 
570 static int
571 prepare_job_resvs_for_each(struct drm_exec *exec, struct pvr_job_data *job_data,
572 			   u32 job_count)
573 {
574 	drm_exec_until_all_locked(exec) {
575 		int err = jobs_lock_all_objs(exec, job_data, job_count);
576 
577 		drm_exec_retry_on_contention(exec);
578 		if (err)
579 			return err;
580 	}
581 
582 	return 0;
583 }
584 
585 static void
586 update_job_resvs(struct pvr_job *job)
587 {
588 	if (job->hwrt) {
589 		enum dma_resv_usage usage = job->type == DRM_PVR_JOB_TYPE_GEOMETRY ?
590 					    DMA_RESV_USAGE_WRITE : DMA_RESV_USAGE_READ;
591 		struct drm_gem_object *obj = gem_from_pvr_gem(job->hwrt->fw_obj->gem);
592 
593 		dma_resv_add_fence(obj->resv, &job->base.s_fence->finished, usage);
594 	}
595 }
596 
597 static void
598 update_job_resvs_for_each(struct pvr_job_data *job_data, u32 job_count)
599 {
600 	for (u32 i = 0; i < job_count; i++)
601 		update_job_resvs(job_data[i].job);
602 }
603 
604 static bool can_combine_jobs(struct pvr_job *a, struct pvr_job *b)
605 {
606 	struct pvr_job *geom_job = a, *frag_job = b;
607 	struct dma_fence *fence;
608 	unsigned long index;
609 
610 	/* Geometry and fragment jobs can be combined if they are queued to the
611 	 * same context and targeting the same HWRT.
612 	 */
613 	if (a->type != DRM_PVR_JOB_TYPE_GEOMETRY ||
614 	    b->type != DRM_PVR_JOB_TYPE_FRAGMENT ||
615 	    a->ctx != b->ctx ||
616 	    a->hwrt != b->hwrt)
617 		return false;
618 
619 	xa_for_each(&frag_job->base.dependencies, index, fence) {
620 		/* We combine when we see an explicit geom -> frag dep. */
621 		if (&geom_job->base.s_fence->scheduled == fence)
622 			return true;
623 	}
624 
625 	return false;
626 }
627 
628 static struct dma_fence *
629 get_last_queued_job_scheduled_fence(struct pvr_queue *queue,
630 				    struct pvr_job_data *job_data,
631 				    u32 cur_job_pos)
632 {
633 	/* We iterate over the current job array in reverse order to grab the
634 	 * last to-be-queued job targeting the same queue.
635 	 */
636 	for (u32 i = cur_job_pos; i > 0; i--) {
637 		struct pvr_job *job = job_data[i - 1].job;
638 
639 		if (job->ctx == queue->ctx && job->type == queue->type)
640 			return dma_fence_get(&job->base.s_fence->scheduled);
641 	}
642 
643 	/* If we didn't find any, we just return the last queued job scheduled
644 	 * fence attached to the queue.
645 	 */
646 	return dma_fence_get(queue->last_queued_job_scheduled_fence);
647 }
648 
649 static int
650 pvr_jobs_link_geom_frag(struct pvr_job_data *job_data, u32 *job_count)
651 {
652 	for (u32 i = 0; i < *job_count - 1; i++) {
653 		struct pvr_job *geom_job = job_data[i].job;
654 		struct pvr_job *frag_job = job_data[i + 1].job;
655 		struct pvr_queue *frag_queue;
656 		struct dma_fence *f;
657 
658 		if (!can_combine_jobs(job_data[i].job, job_data[i + 1].job))
659 			continue;
660 
661 		/* The fragment job will be submitted by the geometry queue. We
662 		 * need to make sure it comes after all the other fragment jobs
663 		 * queued before it.
664 		 */
665 		frag_queue = pvr_context_get_queue_for_job(frag_job->ctx,
666 							   frag_job->type);
667 		f = get_last_queued_job_scheduled_fence(frag_queue, job_data,
668 							i);
669 		if (f) {
670 			int err = drm_sched_job_add_dependency(&geom_job->base,
671 							       f);
672 			if (err) {
673 				*job_count = i;
674 				return err;
675 			}
676 		}
677 
678 		/* The KCCB slot will be reserved by the geometry job, so we can
679 		 * drop the KCCB fence on the fragment job.
680 		 */
681 		pvr_kccb_fence_put(frag_job->kccb_fence);
682 		frag_job->kccb_fence = NULL;
683 
684 		geom_job->paired_job = frag_job;
685 		frag_job->paired_job = geom_job;
686 
687 		/* Skip the fragment job we just paired to the geometry job. */
688 		i++;
689 	}
690 
691 	return 0;
692 }
693 
694 /**
695  * pvr_submit_jobs() - Submit jobs to the GPU
696  * @pvr_dev: Target PowerVR device.
697  * @pvr_file: Pointer to PowerVR file structure.
698  * @args: Ioctl args.
699  *
700  * This initial implementation is entirely synchronous; on return the GPU will
701  * be idle. This will not be the case for future implementations.
702  *
703  * Returns:
704  *  * 0 on success,
705  *  * -%EFAULT if arguments can not be copied from user space, or
706  *  * -%EINVAL on invalid arguments, or
707  *  * Any other error.
708  */
709 int
710 pvr_submit_jobs(struct pvr_device *pvr_dev, struct pvr_file *pvr_file,
711 		struct drm_pvr_ioctl_submit_jobs_args *args)
712 {
713 	struct pvr_job_data *job_data = NULL;
714 	struct drm_pvr_job *job_args;
715 	struct xarray signal_array;
716 	u32 jobs_alloced = 0;
717 	struct drm_exec exec;
718 	int err;
719 
720 	if (!args->jobs.count)
721 		return -EINVAL;
722 
723 	err = PVR_UOBJ_GET_ARRAY(job_args, &args->jobs);
724 	if (err)
725 		return err;
726 
727 	job_data = kvmalloc_array(args->jobs.count, sizeof(*job_data),
728 				  GFP_KERNEL | __GFP_ZERO);
729 	if (!job_data) {
730 		err = -ENOMEM;
731 		goto out_free;
732 	}
733 
734 	err = pvr_job_data_init(pvr_dev, pvr_file, job_args, &args->jobs.count,
735 				job_data);
736 	if (err)
737 		goto out_free;
738 
739 	jobs_alloced = args->jobs.count;
740 
741 	/*
742 	 * Flush MMU if needed - this has been deferred until now to avoid
743 	 * overuse of this expensive operation.
744 	 */
745 	err = pvr_mmu_flush_exec(pvr_dev, false);
746 	if (err)
747 		goto out_job_data_cleanup;
748 
749 	drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT | DRM_EXEC_IGNORE_DUPLICATES, 0);
750 
751 	xa_init_flags(&signal_array, XA_FLAGS_ALLOC);
752 
753 	err = prepare_job_syncs_for_each(pvr_file, job_data, &args->jobs.count,
754 					 &signal_array);
755 	if (err)
756 		goto out_exec_fini;
757 
758 	err = prepare_job_resvs_for_each(&exec, job_data, args->jobs.count);
759 	if (err)
760 		goto out_exec_fini;
761 
762 	err = pvr_jobs_link_geom_frag(job_data, &args->jobs.count);
763 	if (err)
764 		goto out_exec_fini;
765 
766 	/* Anything after that point must succeed because we start exposing job
767 	 * finished fences to the outside world.
768 	 */
769 	update_job_resvs_for_each(job_data, args->jobs.count);
770 	push_jobs(job_data, args->jobs.count);
771 	pvr_sync_signal_array_push_fences(&signal_array);
772 	err = 0;
773 
774 out_exec_fini:
775 	drm_exec_fini(&exec);
776 	pvr_sync_signal_array_cleanup(&signal_array);
777 
778 out_job_data_cleanup:
779 	pvr_job_data_fini(job_data, jobs_alloced);
780 
781 out_free:
782 	kvfree(job_data);
783 	kvfree(job_args);
784 
785 	return err;
786 }
787