Linux Audio

Check our new training course

Loading...
v4.6
 
  1/*
  2 * Copyright (C) 2013 Red Hat
  3 * Author: Rob Clark <robdclark@gmail.com>
  4 *
  5 * This program is free software; you can redistribute it and/or modify it
  6 * under the terms of the GNU General Public License version 2 as published by
  7 * the Free Software Foundation.
  8 *
  9 * This program is distributed in the hope that it will be useful, but WITHOUT
 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 11 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
 12 * more details.
 13 *
 14 * You should have received a copy of the GNU General Public License along with
 15 * this program.  If not, see <http://www.gnu.org/licenses/>.
 16 */
 17
 18#include "msm_gpu.h"
 19#include "msm_gem.h"
 20#include "msm_mmu.h"
 21
 
 
 
 
 
 
 
 
 
 22
 23/*
 24 * Power Management:
 25 */
 26
 27#ifdef DOWNSTREAM_CONFIG_MSM_BUS_SCALING
 28#include <mach/board.h>
 29static void bs_init(struct msm_gpu *gpu)
 30{
 31	if (gpu->bus_scale_table) {
 32		gpu->bsc = msm_bus_scale_register_client(gpu->bus_scale_table);
 33		DBG("bus scale client: %08x", gpu->bsc);
 34	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 35}
 36
 37static void bs_fini(struct msm_gpu *gpu)
 
 38{
 39	if (gpu->bsc) {
 40		msm_bus_scale_unregister_client(gpu->bsc);
 41		gpu->bsc = 0;
 42	}
 
 
 
 
 
 
 
 
 
 
 
 43}
 44
 45static void bs_set(struct msm_gpu *gpu, int idx)
 46{
 47	if (gpu->bsc) {
 48		DBG("set bus scaling: %d", idx);
 49		msm_bus_scale_client_update_request(gpu->bsc, idx);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 50	}
 51}
 52#else
 53static void bs_init(struct msm_gpu *gpu) {}
 54static void bs_fini(struct msm_gpu *gpu) {}
 55static void bs_set(struct msm_gpu *gpu, int idx) {}
 56#endif
 57
 58static int enable_pwrrail(struct msm_gpu *gpu)
 59{
 60	struct drm_device *dev = gpu->dev;
 61	int ret = 0;
 62
 63	if (gpu->gpu_reg) {
 64		ret = regulator_enable(gpu->gpu_reg);
 65		if (ret) {
 66			dev_err(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
 67			return ret;
 68		}
 69	}
 70
 71	if (gpu->gpu_cx) {
 72		ret = regulator_enable(gpu->gpu_cx);
 73		if (ret) {
 74			dev_err(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
 75			return ret;
 76		}
 77	}
 78
 79	return 0;
 80}
 81
 82static int disable_pwrrail(struct msm_gpu *gpu)
 83{
 84	if (gpu->gpu_cx)
 85		regulator_disable(gpu->gpu_cx);
 86	if (gpu->gpu_reg)
 87		regulator_disable(gpu->gpu_reg);
 88	return 0;
 89}
 90
 91static int enable_clk(struct msm_gpu *gpu)
 92{
 93	struct clk *rate_clk = NULL;
 94	int i;
 95
 96	/* NOTE: kgsl_pwrctrl_clk() ignores grp_clks[0].. */
 97	for (i = ARRAY_SIZE(gpu->grp_clks) - 1; i > 0; i--) {
 98		if (gpu->grp_clks[i]) {
 99			clk_prepare(gpu->grp_clks[i]);
100			rate_clk = gpu->grp_clks[i];
101		}
102	}
103
104	if (rate_clk && gpu->fast_rate)
105		clk_set_rate(rate_clk, gpu->fast_rate);
106
107	for (i = ARRAY_SIZE(gpu->grp_clks) - 1; i > 0; i--)
108		if (gpu->grp_clks[i])
109			clk_enable(gpu->grp_clks[i]);
110
111	return 0;
112}
113
114static int disable_clk(struct msm_gpu *gpu)
115{
116	struct clk *rate_clk = NULL;
117	int i;
118
119	/* NOTE: kgsl_pwrctrl_clk() ignores grp_clks[0].. */
120	for (i = ARRAY_SIZE(gpu->grp_clks) - 1; i > 0; i--) {
121		if (gpu->grp_clks[i]) {
122			clk_disable(gpu->grp_clks[i]);
123			rate_clk = gpu->grp_clks[i];
124		}
125	}
126
127	if (rate_clk && gpu->slow_rate)
128		clk_set_rate(rate_clk, gpu->slow_rate);
 
 
 
 
 
129
130	for (i = ARRAY_SIZE(gpu->grp_clks) - 1; i > 0; i--)
131		if (gpu->grp_clks[i])
132			clk_unprepare(gpu->grp_clks[i]);
133
134	return 0;
135}
136
137static int enable_axi(struct msm_gpu *gpu)
138{
139	if (gpu->ebi1_clk)
140		clk_prepare_enable(gpu->ebi1_clk);
141	if (gpu->bus_freq)
142		bs_set(gpu, gpu->bus_freq);
143	return 0;
144}
145
146static int disable_axi(struct msm_gpu *gpu)
147{
148	if (gpu->ebi1_clk)
149		clk_disable_unprepare(gpu->ebi1_clk);
150	if (gpu->bus_freq)
151		bs_set(gpu, 0);
152	return 0;
153}
154
155int msm_gpu_pm_resume(struct msm_gpu *gpu)
156{
157	struct drm_device *dev = gpu->dev;
158	int ret;
159
160	DBG("%s: active_cnt=%d", gpu->name, gpu->active_cnt);
161
162	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 
163
164	if (gpu->active_cnt++ > 0)
165		return 0;
 
166
167	if (WARN_ON(gpu->active_cnt <= 0))
168		return -EINVAL;
169
170	ret = enable_pwrrail(gpu);
171	if (ret)
172		return ret;
173
174	ret = enable_clk(gpu);
175	if (ret)
176		return ret;
177
178	ret = enable_axi(gpu);
179	if (ret)
180		return ret;
181
 
 
 
 
182	return 0;
183}
184
185int msm_gpu_pm_suspend(struct msm_gpu *gpu)
186{
187	struct drm_device *dev = gpu->dev;
188	int ret;
189
190	DBG("%s: active_cnt=%d", gpu->name, gpu->active_cnt);
191
192	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
193
194	if (--gpu->active_cnt > 0)
195		return 0;
196
197	if (WARN_ON(gpu->active_cnt < 0))
198		return -EINVAL;
199
200	ret = disable_axi(gpu);
201	if (ret)
202		return ret;
203
204	ret = disable_clk(gpu);
205	if (ret)
206		return ret;
207
208	ret = disable_pwrrail(gpu);
209	if (ret)
210		return ret;
211
 
 
212	return 0;
213}
214
215/*
216 * Inactivity detection (for suspend):
217 */
 
 
 
 
 
218
219static void inactive_worker(struct work_struct *work)
 
 
 
 
 
 
 
 
 
 
220{
221	struct msm_gpu *gpu = container_of(work, struct msm_gpu, inactive_work);
222	struct drm_device *dev = gpu->dev;
223
224	if (gpu->inactive)
225		return;
 
 
226
227	DBG("%s: inactive!\n", gpu->name);
228	mutex_lock(&dev->struct_mutex);
229	if (!(msm_gpu_active(gpu) || gpu->inactive)) {
230		disable_axi(gpu);
231		disable_clk(gpu);
232		gpu->inactive = true;
233	}
234	mutex_unlock(&dev->struct_mutex);
235}
236
237static void inactive_handler(unsigned long data)
 
 
238{
239	struct msm_gpu *gpu = (struct msm_gpu *)data;
240	struct msm_drm_private *priv = gpu->dev->dev_private;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
241
242	queue_work(priv->wq, &gpu->inactive_work);
 
 
 
 
243}
244
245/* cancel inactive timer and make sure we are awake: */
246static void inactive_cancel(struct msm_gpu *gpu)
247{
248	DBG("%s", gpu->name);
249	del_timer(&gpu->inactive_timer);
250	if (gpu->inactive) {
251		enable_clk(gpu);
252		enable_axi(gpu);
253		gpu->inactive = false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
254	}
 
 
255}
256
257static void inactive_start(struct msm_gpu *gpu)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
258{
259	DBG("%s", gpu->name);
260	mod_timer(&gpu->inactive_timer,
261			round_jiffies_up(jiffies + DRM_MSM_INACTIVE_JIFFIES));
262}
 
263
264/*
265 * Hangcheck detection for locked gpu:
266 */
267
268static void retire_submits(struct msm_gpu *gpu, uint32_t fence);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
269
270static void recover_worker(struct work_struct *work)
271{
272	struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
273	struct drm_device *dev = gpu->dev;
274
275	dev_err(dev->dev, "%s: hangcheck recover!\n", gpu->name);
 
 
 
276
277	mutex_lock(&dev->struct_mutex);
278	if (msm_gpu_active(gpu)) {
279		struct msm_gem_submit *submit;
280		uint32_t fence = gpu->funcs->last_fence(gpu);
281
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
282		/* retire completed submits, plus the one that hung: */
283		retire_submits(gpu, fence + 1);
284
285		inactive_cancel(gpu);
286		gpu->funcs->recover(gpu);
 
287
288		/* replay the remaining submits after the one that hung: */
289		list_for_each_entry(submit, &gpu->submit_list, node) {
290			gpu->funcs->submit(gpu, submit, NULL);
 
 
 
 
 
 
 
 
291		}
292	}
 
293	mutex_unlock(&dev->struct_mutex);
294
295	msm_gpu_retire(gpu);
296}
297
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
298static void hangcheck_timer_reset(struct msm_gpu *gpu)
299{
300	DBG("%s", gpu->name);
301	mod_timer(&gpu->hangcheck_timer,
302			round_jiffies_up(jiffies + DRM_MSM_HANGCHECK_JIFFIES));
303}
304
305static void hangcheck_handler(unsigned long data)
306{
307	struct msm_gpu *gpu = (struct msm_gpu *)data;
308	struct drm_device *dev = gpu->dev;
309	struct msm_drm_private *priv = dev->dev_private;
310	uint32_t fence = gpu->funcs->last_fence(gpu);
311
312	if (fence != gpu->hangcheck_fence) {
313		/* some progress has been made.. ya! */
314		gpu->hangcheck_fence = fence;
315	} else if (fence < gpu->submitted_fence) {
316		/* no progress and not done.. hung! */
317		gpu->hangcheck_fence = fence;
318		dev_err(dev->dev, "%s: hangcheck detected gpu lockup!\n",
319				gpu->name);
320		dev_err(dev->dev, "%s:     completed fence: %u\n",
321				gpu->name, fence);
322		dev_err(dev->dev, "%s:     submitted fence: %u\n",
323				gpu->name, gpu->submitted_fence);
324		queue_work(priv->wq, &gpu->recover_work);
 
325	}
326
327	/* if still more pending work, reset the hangcheck timer: */
328	if (gpu->submitted_fence > gpu->hangcheck_fence)
329		hangcheck_timer_reset(gpu);
330
331	/* workaround for missing irq: */
332	queue_work(priv->wq, &gpu->retire_work);
333}
334
335/*
336 * Performance Counters:
337 */
338
339/* called under perf_lock */
340static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
341{
342	uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
343	int i, n = min(ncntrs, gpu->num_perfcntrs);
344
345	/* read current values: */
346	for (i = 0; i < gpu->num_perfcntrs; i++)
347		current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
348
349	/* update cntrs: */
350	for (i = 0; i < n; i++)
351		cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
352
353	/* save current values: */
354	for (i = 0; i < gpu->num_perfcntrs; i++)
355		gpu->last_cntrs[i] = current_cntrs[i];
356
357	return n;
358}
359
360static void update_sw_cntrs(struct msm_gpu *gpu)
361{
362	ktime_t time;
363	uint32_t elapsed;
364	unsigned long flags;
365
366	spin_lock_irqsave(&gpu->perf_lock, flags);
367	if (!gpu->perfcntr_active)
368		goto out;
369
370	time = ktime_get();
371	elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
372
373	gpu->totaltime += elapsed;
374	if (gpu->last_sample.active)
375		gpu->activetime += elapsed;
376
377	gpu->last_sample.active = msm_gpu_active(gpu);
378	gpu->last_sample.time = time;
379
380out:
381	spin_unlock_irqrestore(&gpu->perf_lock, flags);
382}
383
384void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
385{
386	unsigned long flags;
387
 
 
388	spin_lock_irqsave(&gpu->perf_lock, flags);
389	/* we could dynamically enable/disable perfcntr registers too.. */
390	gpu->last_sample.active = msm_gpu_active(gpu);
391	gpu->last_sample.time = ktime_get();
392	gpu->activetime = gpu->totaltime = 0;
393	gpu->perfcntr_active = true;
394	update_hw_cntrs(gpu, 0, NULL);
395	spin_unlock_irqrestore(&gpu->perf_lock, flags);
396}
397
398void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
399{
400	gpu->perfcntr_active = false;
 
401}
402
403/* returns -errno or # of cntrs sampled */
404int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
405		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
406{
407	unsigned long flags;
408	int ret;
409
410	spin_lock_irqsave(&gpu->perf_lock, flags);
411
412	if (!gpu->perfcntr_active) {
413		ret = -EINVAL;
414		goto out;
415	}
416
417	*activetime = gpu->activetime;
418	*totaltime = gpu->totaltime;
419
420	gpu->activetime = gpu->totaltime = 0;
421
422	ret = update_hw_cntrs(gpu, ncntrs, cntrs);
423
424out:
425	spin_unlock_irqrestore(&gpu->perf_lock, flags);
426
427	return ret;
428}
429
430/*
431 * Cmdstream submission/retirement:
432 */
433
434static void retire_submits(struct msm_gpu *gpu, uint32_t fence)
 
435{
436	struct drm_device *dev = gpu->dev;
 
 
 
437
438	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 
 
 
 
 
 
 
 
 
439
440	while (!list_empty(&gpu->submit_list)) {
441		struct msm_gem_submit *submit;
442
443		submit = list_first_entry(&gpu->submit_list,
444				struct msm_gem_submit, node);
445
446		if (submit->fence <= fence) {
447			list_del(&submit->node);
448			kfree(submit);
449		} else {
450			break;
451		}
452	}
453}
454
455static void retire_worker(struct work_struct *work)
456{
457	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
458	struct drm_device *dev = gpu->dev;
459	uint32_t fence = gpu->funcs->last_fence(gpu);
460
461	msm_update_fence(gpu->dev, fence);
 
462
463	mutex_lock(&dev->struct_mutex);
464
465	retire_submits(gpu, fence);
466
467	while (!list_empty(&gpu->active_list)) {
468		struct msm_gem_object *obj;
469
470		obj = list_first_entry(&gpu->active_list,
471				struct msm_gem_object, mm_list);
 
472
473		if ((obj->read_fence <= fence) &&
474				(obj->write_fence <= fence)) {
475			/* move to inactive: */
476			msm_gem_move_to_inactive(&obj->base);
477			msm_gem_put_iova(&obj->base, gpu->id);
478			drm_gem_object_unreference(&obj->base);
479		} else {
480			break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
481		}
482	}
 
483
484	mutex_unlock(&dev->struct_mutex);
 
 
 
 
 
 
485
486	if (!msm_gpu_active(gpu))
487		inactive_start(gpu);
488}
489
490/* call from irq handler to schedule work to retire bo's */
491void msm_gpu_retire(struct msm_gpu *gpu)
492{
493	struct msm_drm_private *priv = gpu->dev->dev_private;
494	queue_work(priv->wq, &gpu->retire_work);
495	update_sw_cntrs(gpu);
496}
497
498/* add bo's to gpu's ring, and kick gpu: */
499int msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
500		struct msm_file_private *ctx)
501{
502	struct drm_device *dev = gpu->dev;
503	struct msm_drm_private *priv = dev->dev_private;
504	int i, ret;
 
505
506	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
507
508	submit->fence = ++priv->next_fence;
509
510	gpu->submitted_fence = submit->fence;
511
512	inactive_cancel(gpu);
513
514	list_add_tail(&submit->node, &gpu->submit_list);
515
516	msm_rd_dump_submit(submit);
517
518	gpu->submitted_fence = submit->fence;
519
520	update_sw_cntrs(gpu);
521
522	for (i = 0; i < submit->nr_bos; i++) {
523		struct msm_gem_object *msm_obj = submit->bos[i].obj;
 
 
524
525		/* can't happen yet.. but when we add 2d support we'll have
526		 * to deal w/ cross-ring synchronization:
527		 */
528		WARN_ON(is_active(msm_obj) && (msm_obj->gpu != gpu));
529
530		if (!is_active(msm_obj)) {
531			uint32_t iova;
 
 
532
533			/* ring takes a reference to the bo and iova: */
534			drm_gem_object_reference(&msm_obj->base);
535			msm_gem_get_iova_locked(&msm_obj->base,
536					submit->gpu->id, &iova);
537		}
538
539		if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
540			msm_gem_move_to_active(&msm_obj->base, gpu, false, submit->fence);
 
 
 
541
542		if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
543			msm_gem_move_to_active(&msm_obj->base, gpu, true, submit->fence);
544	}
545
546	ret = gpu->funcs->submit(gpu, submit, ctx);
547	priv->lastctx = ctx;
548
549	hangcheck_timer_reset(gpu);
550
551	return ret;
552}
553
554/*
555 * Init/Cleanup:
556 */
557
558static irqreturn_t irq_handler(int irq, void *data)
559{
560	struct msm_gpu *gpu = data;
561	return gpu->funcs->irq(gpu);
562}
563
564static const char *clk_names[] = {
565		"src_clk", "core_clk", "iface_clk", "mem_clk", "mem_iface_clk",
566		"alt_mem_iface_clk",
567};
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
568
569int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
570		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
571		const char *name, const char *ioname, const char *irqname, int ringsz)
572{
573	struct iommu_domain *iommu;
574	int i, ret;
 
575
576	if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
577		gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
578
579	gpu->dev = drm;
580	gpu->funcs = funcs;
581	gpu->name = name;
582	gpu->inactive = true;
 
 
 
 
 
 
 
 
583
584	INIT_LIST_HEAD(&gpu->active_list);
585	INIT_WORK(&gpu->retire_work, retire_worker);
586	INIT_WORK(&gpu->inactive_work, inactive_worker);
587	INIT_WORK(&gpu->recover_work, recover_worker);
588
589	INIT_LIST_HEAD(&gpu->submit_list);
590
591	setup_timer(&gpu->inactive_timer, inactive_handler,
592			(unsigned long)gpu);
593	setup_timer(&gpu->hangcheck_timer, hangcheck_handler,
594			(unsigned long)gpu);
595
596	spin_lock_init(&gpu->perf_lock);
597
598	BUG_ON(ARRAY_SIZE(clk_names) != ARRAY_SIZE(gpu->grp_clks));
599
600	/* Map registers: */
601	gpu->mmio = msm_ioremap(pdev, ioname, name);
602	if (IS_ERR(gpu->mmio)) {
603		ret = PTR_ERR(gpu->mmio);
604		goto fail;
605	}
606
607	/* Get Interrupt: */
608	gpu->irq = platform_get_irq_byname(pdev, irqname);
609	if (gpu->irq < 0) {
610		ret = gpu->irq;
611		dev_err(drm->dev, "failed to get irq: %d\n", ret);
612		goto fail;
613	}
614
615	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
616			IRQF_TRIGGER_HIGH, gpu->name, gpu);
617	if (ret) {
618		dev_err(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
619		goto fail;
620	}
621
622	/* Acquire clocks: */
623	for (i = 0; i < ARRAY_SIZE(clk_names); i++) {
624		gpu->grp_clks[i] = devm_clk_get(&pdev->dev, clk_names[i]);
625		DBG("grp_clks[%s]: %p", clk_names[i], gpu->grp_clks[i]);
626		if (IS_ERR(gpu->grp_clks[i]))
627			gpu->grp_clks[i] = NULL;
628	}
629
630	gpu->ebi1_clk = devm_clk_get(&pdev->dev, "bus_clk");
631	DBG("ebi1_clk: %p", gpu->ebi1_clk);
632	if (IS_ERR(gpu->ebi1_clk))
633		gpu->ebi1_clk = NULL;
634
635	/* Acquire regulators: */
636	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
637	DBG("gpu_reg: %p", gpu->gpu_reg);
638	if (IS_ERR(gpu->gpu_reg))
639		gpu->gpu_reg = NULL;
640
641	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
642	DBG("gpu_cx: %p", gpu->gpu_cx);
643	if (IS_ERR(gpu->gpu_cx))
644		gpu->gpu_cx = NULL;
645
646	/* Setup IOMMU.. eventually we will (I think) do this once per context
647	 * and have separate page tables per context.  For now, to keep things
648	 * simple and to get something working, just use a single address space:
649	 */
650	iommu = iommu_domain_alloc(&platform_bus_type);
651	if (iommu) {
652		dev_info(drm->dev, "%s: using IOMMU\n", name);
653		gpu->mmu = msm_iommu_new(&pdev->dev, iommu);
654		if (IS_ERR(gpu->mmu)) {
655			ret = PTR_ERR(gpu->mmu);
656			dev_err(drm->dev, "failed to init iommu: %d\n", ret);
657			gpu->mmu = NULL;
658			iommu_domain_free(iommu);
659			goto fail;
660		}
661
662	} else {
663		dev_info(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
664	}
665	gpu->id = msm_register_mmu(drm, gpu->mmu);
666
 
667
668	/* Create ringbuffer: */
669	mutex_lock(&drm->struct_mutex);
670	gpu->rb = msm_ringbuffer_new(gpu, ringsz);
671	mutex_unlock(&drm->struct_mutex);
672	if (IS_ERR(gpu->rb)) {
673		ret = PTR_ERR(gpu->rb);
674		gpu->rb = NULL;
675		dev_err(drm->dev, "could not create ringbuffer: %d\n", ret);
676		goto fail;
677	}
678
679	bs_init(gpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
680
681	return 0;
682
683fail:
 
 
 
 
 
 
 
 
684	return ret;
685}
686
687void msm_gpu_cleanup(struct msm_gpu *gpu)
688{
 
 
689	DBG("%s", gpu->name);
690
691	WARN_ON(!list_empty(&gpu->active_list));
692
693	bs_fini(gpu);
 
 
 
 
 
 
 
 
 
 
694
695	if (gpu->rb) {
696		if (gpu->rb_iova)
697			msm_gem_put_iova(gpu->rb->bo, gpu->id);
698		msm_ringbuffer_destroy(gpu->rb);
699	}
700
701	if (gpu->mmu)
702		gpu->mmu->funcs->destroy(gpu->mmu);
703}
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2013 Red Hat
   4 * Author: Rob Clark <robdclark@gmail.com>
 
 
 
 
 
 
 
 
 
 
 
 
   5 */
   6
   7#include "msm_gpu.h"
   8#include "msm_gem.h"
   9#include "msm_mmu.h"
  10#include "msm_fence.h"
  11#include "msm_gpu_trace.h"
  12#include "adreno/adreno_gpu.h"
  13
  14#include <generated/utsrelease.h>
  15#include <linux/string_helpers.h>
  16#include <linux/devfreq.h>
  17#include <linux/devfreq_cooling.h>
  18#include <linux/devcoredump.h>
  19#include <linux/sched/task.h>
  20
  21/*
  22 * Power Management:
  23 */
  24
  25static int msm_devfreq_target(struct device *dev, unsigned long *freq,
  26		u32 flags)
 
  27{
  28	struct msm_gpu *gpu = dev_to_gpu(dev);
  29	struct dev_pm_opp *opp;
  30
  31	opp = devfreq_recommended_opp(dev, freq, flags);
  32
  33	if (IS_ERR(opp))
  34		return PTR_ERR(opp);
  35
  36	trace_msm_gpu_freq_change(dev_pm_opp_get_freq(opp));
  37
  38	if (gpu->funcs->gpu_set_freq)
  39		gpu->funcs->gpu_set_freq(gpu, opp);
  40	else
  41		clk_set_rate(gpu->core_clk, *freq);
  42
  43	dev_pm_opp_put(opp);
  44
  45	return 0;
  46}
  47
  48static int msm_devfreq_get_dev_status(struct device *dev,
  49		struct devfreq_dev_status *status)
  50{
  51	struct msm_gpu *gpu = dev_to_gpu(dev);
  52	ktime_t time;
  53
  54	if (gpu->funcs->gpu_get_freq)
  55		status->current_frequency = gpu->funcs->gpu_get_freq(gpu);
  56	else
  57		status->current_frequency = clk_get_rate(gpu->core_clk);
  58
  59	status->busy_time = gpu->funcs->gpu_busy(gpu);
  60
  61	time = ktime_get();
  62	status->total_time = ktime_us_delta(time, gpu->devfreq.time);
  63	gpu->devfreq.time = time;
  64
  65	return 0;
  66}
  67
  68static int msm_devfreq_get_cur_freq(struct device *dev, unsigned long *freq)
  69{
  70	struct msm_gpu *gpu = dev_to_gpu(dev);
  71
  72	if (gpu->funcs->gpu_get_freq)
  73		*freq = gpu->funcs->gpu_get_freq(gpu);
  74	else
  75		*freq = clk_get_rate(gpu->core_clk);
  76
  77	return 0;
  78}
  79
  80static struct devfreq_dev_profile msm_devfreq_profile = {
  81	.polling_ms = 10,
  82	.target = msm_devfreq_target,
  83	.get_dev_status = msm_devfreq_get_dev_status,
  84	.get_cur_freq = msm_devfreq_get_cur_freq,
  85};
  86
  87static void msm_devfreq_init(struct msm_gpu *gpu)
  88{
  89	/* We need target support to do devfreq */
  90	if (!gpu->funcs->gpu_busy)
  91		return;
  92
  93	msm_devfreq_profile.initial_freq = gpu->fast_rate;
  94
  95	/*
  96	 * Don't set the freq_table or max_state and let devfreq build the table
  97	 * from OPP
  98	 * After a deferred probe, these may have be left to non-zero values,
  99	 * so set them back to zero before creating the devfreq device
 100	 */
 101	msm_devfreq_profile.freq_table = NULL;
 102	msm_devfreq_profile.max_state = 0;
 103
 104	gpu->devfreq.devfreq = devm_devfreq_add_device(&gpu->pdev->dev,
 105			&msm_devfreq_profile, DEVFREQ_GOV_SIMPLE_ONDEMAND,
 106			NULL);
 107
 108	if (IS_ERR(gpu->devfreq.devfreq)) {
 109		DRM_DEV_ERROR(&gpu->pdev->dev, "Couldn't initialize GPU devfreq\n");
 110		gpu->devfreq.devfreq = NULL;
 111		return;
 112	}
 113
 114	devfreq_suspend_device(gpu->devfreq.devfreq);
 115
 116	gpu->cooling = of_devfreq_cooling_register(gpu->pdev->dev.of_node,
 117			gpu->devfreq.devfreq);
 118	if (IS_ERR(gpu->cooling)) {
 119		DRM_DEV_ERROR(&gpu->pdev->dev,
 120				"Couldn't register GPU cooling device\n");
 121		gpu->cooling = NULL;
 122	}
 123}
 
 
 
 
 
 124
 125static int enable_pwrrail(struct msm_gpu *gpu)
 126{
 127	struct drm_device *dev = gpu->dev;
 128	int ret = 0;
 129
 130	if (gpu->gpu_reg) {
 131		ret = regulator_enable(gpu->gpu_reg);
 132		if (ret) {
 133			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
 134			return ret;
 135		}
 136	}
 137
 138	if (gpu->gpu_cx) {
 139		ret = regulator_enable(gpu->gpu_cx);
 140		if (ret) {
 141			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
 142			return ret;
 143		}
 144	}
 145
 146	return 0;
 147}
 148
 149static int disable_pwrrail(struct msm_gpu *gpu)
 150{
 151	if (gpu->gpu_cx)
 152		regulator_disable(gpu->gpu_cx);
 153	if (gpu->gpu_reg)
 154		regulator_disable(gpu->gpu_reg);
 155	return 0;
 156}
 157
 158static int enable_clk(struct msm_gpu *gpu)
 159{
 160	if (gpu->core_clk && gpu->fast_rate)
 161		clk_set_rate(gpu->core_clk, gpu->fast_rate);
 162
 163	/* Set the RBBM timer rate to 19.2Mhz */
 164	if (gpu->rbbmtimer_clk)
 165		clk_set_rate(gpu->rbbmtimer_clk, 19200000);
 
 
 
 
 
 
 
 166
 167	return clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
 
 
 
 
 168}
 169
 170static int disable_clk(struct msm_gpu *gpu)
 171{
 172	clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
 
 
 
 
 
 
 
 
 
 173
 174	/*
 175	 * Set the clock to a deliberately low rate. On older targets the clock
 176	 * speed had to be non zero to avoid problems. On newer targets this
 177	 * will be rounded down to zero anyway so it all works out.
 178	 */
 179	if (gpu->core_clk)
 180		clk_set_rate(gpu->core_clk, 27000000);
 181
 182	if (gpu->rbbmtimer_clk)
 183		clk_set_rate(gpu->rbbmtimer_clk, 0);
 
 184
 185	return 0;
 186}
 187
 188static int enable_axi(struct msm_gpu *gpu)
 189{
 190	return clk_prepare_enable(gpu->ebi1_clk);
 
 
 
 
 191}
 192
 193static int disable_axi(struct msm_gpu *gpu)
 194{
 195	clk_disable_unprepare(gpu->ebi1_clk);
 
 
 
 196	return 0;
 197}
 198
 199void msm_gpu_resume_devfreq(struct msm_gpu *gpu)
 200{
 201	gpu->devfreq.busy_cycles = 0;
 202	gpu->devfreq.time = ktime_get();
 
 
 203
 204	devfreq_resume_device(gpu->devfreq.devfreq);
 205}
 206
 207int msm_gpu_pm_resume(struct msm_gpu *gpu)
 208{
 209	int ret;
 210
 211	DBG("%s", gpu->name);
 212	trace_msm_gpu_resume(0);
 213
 214	ret = enable_pwrrail(gpu);
 215	if (ret)
 216		return ret;
 217
 218	ret = enable_clk(gpu);
 219	if (ret)
 220		return ret;
 221
 222	ret = enable_axi(gpu);
 223	if (ret)
 224		return ret;
 225
 226	msm_gpu_resume_devfreq(gpu);
 227
 228	gpu->needs_hw_init = true;
 229
 230	return 0;
 231}
 232
 233int msm_gpu_pm_suspend(struct msm_gpu *gpu)
 234{
 
 235	int ret;
 236
 237	DBG("%s", gpu->name);
 238	trace_msm_gpu_suspend(0);
 
 
 
 
 239
 240	devfreq_suspend_device(gpu->devfreq.devfreq);
 
 241
 242	ret = disable_axi(gpu);
 243	if (ret)
 244		return ret;
 245
 246	ret = disable_clk(gpu);
 247	if (ret)
 248		return ret;
 249
 250	ret = disable_pwrrail(gpu);
 251	if (ret)
 252		return ret;
 253
 254	gpu->suspend_count++;
 255
 256	return 0;
 257}
 258
 259int msm_gpu_hw_init(struct msm_gpu *gpu)
 260{
 261	int ret;
 262
 263	WARN_ON(!mutex_is_locked(&gpu->dev->struct_mutex));
 264
 265	if (!gpu->needs_hw_init)
 266		return 0;
 267
 268	disable_irq(gpu->irq);
 269	ret = gpu->funcs->hw_init(gpu);
 270	if (!ret)
 271		gpu->needs_hw_init = false;
 272	enable_irq(gpu->irq);
 273
 274	return ret;
 275}
 276
 277static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
 278		uint32_t fence)
 279{
 280	struct msm_gem_submit *submit;
 
 281
 282	spin_lock(&ring->submit_lock);
 283	list_for_each_entry(submit, &ring->submits, node) {
 284		if (submit->seqno > fence)
 285			break;
 286
 287		msm_update_fence(submit->ring->fctx,
 288			submit->fence->seqno);
 
 
 
 
 289	}
 290	spin_unlock(&ring->submit_lock);
 291}
 292
 293#ifdef CONFIG_DEV_COREDUMP
 294static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
 295		size_t count, void *data, size_t datalen)
 296{
 297	struct msm_gpu *gpu = data;
 298	struct drm_print_iterator iter;
 299	struct drm_printer p;
 300	struct msm_gpu_state *state;
 301
 302	state = msm_gpu_crashstate_get(gpu);
 303	if (!state)
 304		return 0;
 305
 306	iter.data = buffer;
 307	iter.offset = 0;
 308	iter.start = offset;
 309	iter.remain = count;
 310
 311	p = drm_coredump_printer(&iter);
 312
 313	drm_printf(&p, "---\n");
 314	drm_printf(&p, "kernel: " UTS_RELEASE "\n");
 315	drm_printf(&p, "module: " KBUILD_MODNAME "\n");
 316	drm_printf(&p, "time: %lld.%09ld\n",
 317		state->time.tv_sec, state->time.tv_nsec);
 318	if (state->comm)
 319		drm_printf(&p, "comm: %s\n", state->comm);
 320	if (state->cmd)
 321		drm_printf(&p, "cmdline: %s\n", state->cmd);
 322
 323	gpu->funcs->show(gpu, state, &p);
 324
 325	msm_gpu_crashstate_put(gpu);
 326
 327	return count - iter.remain;
 328}
 329
 330static void msm_gpu_devcoredump_free(void *data)
 
 331{
 332	struct msm_gpu *gpu = data;
 333
 334	msm_gpu_crashstate_put(gpu);
 335}
 336
 337static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
 338		struct msm_gem_object *obj, u64 iova, u32 flags)
 339{
 340	struct msm_gpu_state_bo *state_bo = &state->bos[state->nr_bos];
 341
 342	/* Don't record write only objects */
 343	state_bo->size = obj->base.size;
 344	state_bo->iova = iova;
 345
 346	/* Only store data for non imported buffer objects marked for read */
 347	if ((flags & MSM_SUBMIT_BO_READ) && !obj->base.import_attach) {
 348		void *ptr;
 349
 350		state_bo->data = kvmalloc(obj->base.size, GFP_KERNEL);
 351		if (!state_bo->data)
 352			goto out;
 353
 354		msm_gem_lock(&obj->base);
 355		ptr = msm_gem_get_vaddr_active(&obj->base);
 356		msm_gem_unlock(&obj->base);
 357		if (IS_ERR(ptr)) {
 358			kvfree(state_bo->data);
 359			state_bo->data = NULL;
 360			goto out;
 361		}
 362
 363		memcpy(state_bo->data, ptr, obj->base.size);
 364		msm_gem_put_vaddr(&obj->base);
 365	}
 366out:
 367	state->nr_bos++;
 368}
 369
 370static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
 371		struct msm_gem_submit *submit, char *comm, char *cmd)
 372{
 373	struct msm_gpu_state *state;
 374
 375	/* Check if the target supports capturing crash state */
 376	if (!gpu->funcs->gpu_state_get)
 377		return;
 378
 379	/* Only save one crash state at a time */
 380	if (gpu->crashstate)
 381		return;
 382
 383	state = gpu->funcs->gpu_state_get(gpu);
 384	if (IS_ERR_OR_NULL(state))
 385		return;
 386
 387	/* Fill in the additional crash state information */
 388	state->comm = kstrdup(comm, GFP_KERNEL);
 389	state->cmd = kstrdup(cmd, GFP_KERNEL);
 390	state->fault_info = gpu->fault_info;
 391
 392	if (submit) {
 393		int i, nr = 0;
 394
 395		/* count # of buffers to dump: */
 396		for (i = 0; i < submit->nr_bos; i++)
 397			if (should_dump(submit, i))
 398				nr++;
 399		/* always dump cmd bo's, but don't double count them: */
 400		for (i = 0; i < submit->nr_cmds; i++)
 401			if (!should_dump(submit, submit->cmd[i].idx))
 402				nr++;
 403
 404		state->bos = kcalloc(nr,
 405			sizeof(struct msm_gpu_state_bo), GFP_KERNEL);
 406
 407		for (i = 0; i < submit->nr_bos; i++) {
 408			if (should_dump(submit, i)) {
 409				msm_gpu_crashstate_get_bo(state, submit->bos[i].obj,
 410					submit->bos[i].iova, submit->bos[i].flags);
 411			}
 412		}
 413
 414		for (i = 0; state->bos && i < submit->nr_cmds; i++) {
 415			int idx = submit->cmd[i].idx;
 416
 417			if (!should_dump(submit, submit->cmd[i].idx)) {
 418				msm_gpu_crashstate_get_bo(state, submit->bos[idx].obj,
 419					submit->bos[idx].iova, submit->bos[idx].flags);
 420			}
 421		}
 422	}
 423
 424	/* Set the active crash state to be dumped on failure */
 425	gpu->crashstate = state;
 426
 427	/* FIXME: Release the crashstate if this errors out? */
 428	dev_coredumpm(gpu->dev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
 429		msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
 430}
 431#else
 432static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
 433		struct msm_gem_submit *submit, char *comm, char *cmd)
 434{
 
 
 
 435}
 436#endif
 437
 438/*
 439 * Hangcheck detection for locked gpu:
 440 */
 441
 442static struct msm_gem_submit *
 443find_submit(struct msm_ringbuffer *ring, uint32_t fence)
 444{
 445	struct msm_gem_submit *submit;
 446
 447	spin_lock(&ring->submit_lock);
 448	list_for_each_entry(submit, &ring->submits, node) {
 449		if (submit->seqno == fence) {
 450			spin_unlock(&ring->submit_lock);
 451			return submit;
 452		}
 453	}
 454	spin_unlock(&ring->submit_lock);
 455
 456	return NULL;
 457}
 458
 459static void retire_submits(struct msm_gpu *gpu);
 460
 461static void recover_worker(struct kthread_work *work)
 462{
 463	struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
 464	struct drm_device *dev = gpu->dev;
 465	struct msm_drm_private *priv = dev->dev_private;
 466	struct msm_gem_submit *submit;
 467	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
 468	char *comm = NULL, *cmd = NULL;
 469	int i;
 470
 471	mutex_lock(&dev->struct_mutex);
 
 
 
 472
 473	DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
 474
 475	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
 476	if (submit) {
 477		struct task_struct *task;
 478
 479		/* Increment the fault counts */
 480		gpu->global_faults++;
 481		submit->queue->faults++;
 482
 483		task = get_pid_task(submit->pid, PIDTYPE_PID);
 484		if (task) {
 485			comm = kstrdup(task->comm, GFP_KERNEL);
 486			cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
 487			put_task_struct(task);
 488		}
 489
 490		/* msm_rd_dump_submit() needs bo locked to dump: */
 491		for (i = 0; i < submit->nr_bos; i++)
 492			msm_gem_lock(&submit->bos[i].obj->base);
 493
 494		if (comm && cmd) {
 495			DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
 496				gpu->name, comm, cmd);
 497
 498			msm_rd_dump_submit(priv->hangrd, submit,
 499				"offending task: %s (%s)", comm, cmd);
 500		} else {
 501			msm_rd_dump_submit(priv->hangrd, submit, NULL);
 502		}
 503
 504		for (i = 0; i < submit->nr_bos; i++)
 505			msm_gem_unlock(&submit->bos[i].obj->base);
 506	}
 507
 508	/* Record the crash state */
 509	pm_runtime_get_sync(&gpu->pdev->dev);
 510	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
 511	pm_runtime_put_sync(&gpu->pdev->dev);
 512
 513	kfree(cmd);
 514	kfree(comm);
 515
 516	/*
 517	 * Update all the rings with the latest and greatest fence.. this
 518	 * needs to happen after msm_rd_dump_submit() to ensure that the
 519	 * bo's referenced by the offending submit are still around.
 520	 */
 521	for (i = 0; i < gpu->nr_rings; i++) {
 522		struct msm_ringbuffer *ring = gpu->rb[i];
 523
 524		uint32_t fence = ring->memptrs->fence;
 525
 526		/*
 527		 * For the current (faulting?) ring/submit advance the fence by
 528		 * one more to clear the faulting submit
 529		 */
 530		if (ring == cur_ring)
 531			fence++;
 532
 533		update_fences(gpu, ring, fence);
 534	}
 535
 536	if (msm_gpu_active(gpu)) {
 537		/* retire completed submits, plus the one that hung: */
 538		retire_submits(gpu);
 539
 540		pm_runtime_get_sync(&gpu->pdev->dev);
 541		gpu->funcs->recover(gpu);
 542		pm_runtime_put_sync(&gpu->pdev->dev);
 543
 544		/*
 545		 * Replay all remaining submits starting with highest priority
 546		 * ring
 547		 */
 548		for (i = 0; i < gpu->nr_rings; i++) {
 549			struct msm_ringbuffer *ring = gpu->rb[i];
 550
 551			spin_lock(&ring->submit_lock);
 552			list_for_each_entry(submit, &ring->submits, node)
 553				gpu->funcs->submit(gpu, submit);
 554			spin_unlock(&ring->submit_lock);
 555		}
 556	}
 557
 558	mutex_unlock(&dev->struct_mutex);
 559
 560	msm_gpu_retire(gpu);
 561}
 562
 563static void fault_worker(struct kthread_work *work)
 564{
 565	struct msm_gpu *gpu = container_of(work, struct msm_gpu, fault_work);
 566	struct drm_device *dev = gpu->dev;
 567	struct msm_gem_submit *submit;
 568	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
 569	char *comm = NULL, *cmd = NULL;
 570
 571	mutex_lock(&dev->struct_mutex);
 572
 573	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
 574	if (submit && submit->fault_dumped)
 575		goto resume_smmu;
 576
 577	if (submit) {
 578		struct task_struct *task;
 579
 580		task = get_pid_task(submit->pid, PIDTYPE_PID);
 581		if (task) {
 582			comm = kstrdup(task->comm, GFP_KERNEL);
 583			cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
 584			put_task_struct(task);
 585		}
 586
 587		/*
 588		 * When we get GPU iova faults, we can get 1000s of them,
 589		 * but we really only want to log the first one.
 590		 */
 591		submit->fault_dumped = true;
 592	}
 593
 594	/* Record the crash state */
 595	pm_runtime_get_sync(&gpu->pdev->dev);
 596	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
 597	pm_runtime_put_sync(&gpu->pdev->dev);
 598
 599	kfree(cmd);
 600	kfree(comm);
 601
 602resume_smmu:
 603	memset(&gpu->fault_info, 0, sizeof(gpu->fault_info));
 604	gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu);
 605
 606	mutex_unlock(&dev->struct_mutex);
 607}
 608
 609static void hangcheck_timer_reset(struct msm_gpu *gpu)
 610{
 611	struct msm_drm_private *priv = gpu->dev->dev_private;
 612	mod_timer(&gpu->hangcheck_timer,
 613			round_jiffies_up(jiffies + msecs_to_jiffies(priv->hangcheck_period)));
 614}
 615
 616static void hangcheck_handler(struct timer_list *t)
 617{
 618	struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
 619	struct drm_device *dev = gpu->dev;
 620	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
 621	uint32_t fence = ring->memptrs->fence;
 622
 623	if (fence != ring->hangcheck_fence) {
 624		/* some progress has been made.. ya! */
 625		ring->hangcheck_fence = fence;
 626	} else if (fence < ring->seqno) {
 627		/* no progress and not done.. hung! */
 628		ring->hangcheck_fence = fence;
 629		DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
 630				gpu->name, ring->id);
 631		DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
 632				gpu->name, fence);
 633		DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
 634				gpu->name, ring->seqno);
 635
 636		kthread_queue_work(gpu->worker, &gpu->recover_work);
 637	}
 638
 639	/* if still more pending work, reset the hangcheck timer: */
 640	if (ring->seqno > ring->hangcheck_fence)
 641		hangcheck_timer_reset(gpu);
 642
 643	/* workaround for missing irq: */
 644	kthread_queue_work(gpu->worker, &gpu->retire_work);
 645}
 646
 647/*
 648 * Performance Counters:
 649 */
 650
 651/* called under perf_lock */
 652static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
 653{
 654	uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
 655	int i, n = min(ncntrs, gpu->num_perfcntrs);
 656
 657	/* read current values: */
 658	for (i = 0; i < gpu->num_perfcntrs; i++)
 659		current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
 660
 661	/* update cntrs: */
 662	for (i = 0; i < n; i++)
 663		cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
 664
 665	/* save current values: */
 666	for (i = 0; i < gpu->num_perfcntrs; i++)
 667		gpu->last_cntrs[i] = current_cntrs[i];
 668
 669	return n;
 670}
 671
 672static void update_sw_cntrs(struct msm_gpu *gpu)
 673{
 674	ktime_t time;
 675	uint32_t elapsed;
 676	unsigned long flags;
 677
 678	spin_lock_irqsave(&gpu->perf_lock, flags);
 679	if (!gpu->perfcntr_active)
 680		goto out;
 681
 682	time = ktime_get();
 683	elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
 684
 685	gpu->totaltime += elapsed;
 686	if (gpu->last_sample.active)
 687		gpu->activetime += elapsed;
 688
 689	gpu->last_sample.active = msm_gpu_active(gpu);
 690	gpu->last_sample.time = time;
 691
 692out:
 693	spin_unlock_irqrestore(&gpu->perf_lock, flags);
 694}
 695
 696void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
 697{
 698	unsigned long flags;
 699
 700	pm_runtime_get_sync(&gpu->pdev->dev);
 701
 702	spin_lock_irqsave(&gpu->perf_lock, flags);
 703	/* we could dynamically enable/disable perfcntr registers too.. */
 704	gpu->last_sample.active = msm_gpu_active(gpu);
 705	gpu->last_sample.time = ktime_get();
 706	gpu->activetime = gpu->totaltime = 0;
 707	gpu->perfcntr_active = true;
 708	update_hw_cntrs(gpu, 0, NULL);
 709	spin_unlock_irqrestore(&gpu->perf_lock, flags);
 710}
 711
 712void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
 713{
 714	gpu->perfcntr_active = false;
 715	pm_runtime_put_sync(&gpu->pdev->dev);
 716}
 717
 718/* returns -errno or # of cntrs sampled */
 719int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
 720		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
 721{
 722	unsigned long flags;
 723	int ret;
 724
 725	spin_lock_irqsave(&gpu->perf_lock, flags);
 726
 727	if (!gpu->perfcntr_active) {
 728		ret = -EINVAL;
 729		goto out;
 730	}
 731
 732	*activetime = gpu->activetime;
 733	*totaltime = gpu->totaltime;
 734
 735	gpu->activetime = gpu->totaltime = 0;
 736
 737	ret = update_hw_cntrs(gpu, ncntrs, cntrs);
 738
 739out:
 740	spin_unlock_irqrestore(&gpu->perf_lock, flags);
 741
 742	return ret;
 743}
 744
 745/*
 746 * Cmdstream submission/retirement:
 747 */
 748
 749static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
 750		struct msm_gem_submit *submit)
 751{
 752	int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
 753	volatile struct msm_gpu_submit_stats *stats;
 754	u64 elapsed, clock = 0;
 755	int i;
 756
 757	stats = &ring->memptrs->stats[index];
 758	/* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
 759	elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
 760	do_div(elapsed, 192);
 761
 762	/* Calculate the clock frequency from the number of CP cycles */
 763	if (elapsed) {
 764		clock = (stats->cpcycles_end - stats->cpcycles_start) * 1000;
 765		do_div(clock, elapsed);
 766	}
 767
 768	trace_msm_gpu_submit_retired(submit, elapsed, clock,
 769		stats->alwayson_start, stats->alwayson_end);
 770
 771	for (i = 0; i < submit->nr_bos; i++) {
 772		struct drm_gem_object *obj = &submit->bos[i].obj->base;
 773
 774		msm_gem_lock(obj);
 775		msm_gem_active_put(obj);
 776		msm_gem_unpin_iova_locked(obj, submit->aspace);
 777		msm_gem_unlock(obj);
 778		drm_gem_object_put(obj);
 
 779	}
 
 
 
 
 
 
 
 780
 781	pm_runtime_mark_last_busy(&gpu->pdev->dev);
 782	pm_runtime_put_autosuspend(&gpu->pdev->dev);
 783
 784	spin_lock(&ring->submit_lock);
 785	list_del(&submit->node);
 786	spin_unlock(&ring->submit_lock);
 787
 788	msm_gem_submit_put(submit);
 789}
 790
 791static void retire_submits(struct msm_gpu *gpu)
 792{
 793	int i;
 794
 795	/* Retire the commits starting with highest priority */
 796	for (i = 0; i < gpu->nr_rings; i++) {
 797		struct msm_ringbuffer *ring = gpu->rb[i];
 798
 799		while (true) {
 800			struct msm_gem_submit *submit = NULL;
 801
 802			spin_lock(&ring->submit_lock);
 803			submit = list_first_entry_or_null(&ring->submits,
 804					struct msm_gem_submit, node);
 805			spin_unlock(&ring->submit_lock);
 806
 807			/*
 808			 * If no submit, we are done.  If submit->fence hasn't
 809			 * been signalled, then later submits are not signalled
 810			 * either, so we are also done.
 811			 */
 812			if (submit && dma_fence_is_signaled(submit->fence)) {
 813				retire_submit(gpu, ring, submit);
 814			} else {
 815				break;
 816			}
 817		}
 818	}
 819}
 820
 821static void retire_worker(struct kthread_work *work)
 822{
 823	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
 824	int i;
 825
 826	for (i = 0; i < gpu->nr_rings; i++)
 827		update_fences(gpu, gpu->rb[i], gpu->rb[i]->memptrs->fence);
 828
 829	retire_submits(gpu);
 
 830}
 831
 832/* call from irq handler to schedule work to retire bo's */
 833void msm_gpu_retire(struct msm_gpu *gpu)
 834{
 835	kthread_queue_work(gpu->worker, &gpu->retire_work);
 
 836	update_sw_cntrs(gpu);
 837}
 838
 839/* add bo's to gpu's ring, and kick gpu: */
 840void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
 
 841{
 842	struct drm_device *dev = gpu->dev;
 843	struct msm_drm_private *priv = dev->dev_private;
 844	struct msm_ringbuffer *ring = submit->ring;
 845	int i;
 846
 847	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 848
 849	pm_runtime_get_sync(&gpu->pdev->dev);
 
 
 
 
 850
 851	msm_gpu_hw_init(gpu);
 852
 853	submit->seqno = ++ring->seqno;
 854
 855	msm_rd_dump_submit(priv->rd, submit, NULL);
 856
 857	update_sw_cntrs(gpu);
 858
 859	for (i = 0; i < submit->nr_bos; i++) {
 860		struct msm_gem_object *msm_obj = submit->bos[i].obj;
 861		struct drm_gem_object *drm_obj = &msm_obj->base;
 862		uint64_t iova;
 863
 864		/* submit takes a reference to the bo and iova until retired: */
 865		drm_gem_object_get(&msm_obj->base);
 866		msm_gem_get_and_pin_iova_locked(&msm_obj->base, submit->aspace, &iova);
 
 867
 868		if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
 869			dma_resv_add_excl_fence(drm_obj->resv, submit->fence);
 870		else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
 871			dma_resv_add_shared_fence(drm_obj->resv, submit->fence);
 872
 873		msm_gem_active_get(drm_obj, gpu);
 874	}
 
 
 
 875
 876	/*
 877	 * ring->submits holds a ref to the submit, to deal with the case
 878	 * that a submit completes before msm_ioctl_gem_submit() returns.
 879	 */
 880	msm_gem_submit_get(submit);
 881
 882	spin_lock(&ring->submit_lock);
 883	list_add_tail(&submit->node, &ring->submits);
 884	spin_unlock(&ring->submit_lock);
 885
 886	gpu->funcs->submit(gpu, submit);
 887	priv->lastctx = submit->queue->ctx;
 888
 889	hangcheck_timer_reset(gpu);
 
 
 890}
 891
 892/*
 893 * Init/Cleanup:
 894 */
 895
 896static irqreturn_t irq_handler(int irq, void *data)
 897{
 898	struct msm_gpu *gpu = data;
 899	return gpu->funcs->irq(gpu);
 900}
 901
 902static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
 903{
 904	int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
 905
 906	if (ret < 1) {
 907		gpu->nr_clocks = 0;
 908		return ret;
 909	}
 910
 911	gpu->nr_clocks = ret;
 912
 913	gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
 914		gpu->nr_clocks, "core");
 915
 916	gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
 917		gpu->nr_clocks, "rbbmtimer");
 918
 919	return 0;
 920}
 921
 922/* Return a new address space for a msm_drm_private instance */
 923struct msm_gem_address_space *
 924msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task)
 925{
 926	struct msm_gem_address_space *aspace = NULL;
 927	if (!gpu)
 928		return NULL;
 929
 930	/*
 931	 * If the target doesn't support private address spaces then return
 932	 * the global one
 933	 */
 934	if (gpu->funcs->create_private_address_space) {
 935		aspace = gpu->funcs->create_private_address_space(gpu);
 936		if (!IS_ERR(aspace))
 937			aspace->pid = get_pid(task_pid(task));
 938	}
 939
 940	if (IS_ERR_OR_NULL(aspace))
 941		aspace = msm_gem_address_space_get(gpu->aspace);
 942
 943	return aspace;
 944}
 945
 946int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
 947		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
 948		const char *name, struct msm_gpu_config *config)
 949{
 950	int i, ret, nr_rings = config->nr_rings;
 951	void *memptrs;
 952	uint64_t memptrs_iova;
 953
 954	if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
 955		gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
 956
 957	gpu->dev = drm;
 958	gpu->funcs = funcs;
 959	gpu->name = name;
 960
 961	gpu->worker = kthread_create_worker(0, "%s-worker", gpu->name);
 962	if (IS_ERR(gpu->worker)) {
 963		ret = PTR_ERR(gpu->worker);
 964		gpu->worker = NULL;
 965		goto fail;
 966	}
 967
 968	sched_set_fifo_low(gpu->worker->task);
 969
 970	INIT_LIST_HEAD(&gpu->active_list);
 971	kthread_init_work(&gpu->retire_work, retire_worker);
 972	kthread_init_work(&gpu->recover_work, recover_worker);
 973	kthread_init_work(&gpu->fault_work, fault_worker);
 974
 975	timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
 
 
 
 
 
 976
 977	spin_lock_init(&gpu->perf_lock);
 978
 
 979
 980	/* Map registers: */
 981	gpu->mmio = msm_ioremap(pdev, config->ioname, name);
 982	if (IS_ERR(gpu->mmio)) {
 983		ret = PTR_ERR(gpu->mmio);
 984		goto fail;
 985	}
 986
 987	/* Get Interrupt: */
 988	gpu->irq = platform_get_irq(pdev, 0);
 989	if (gpu->irq < 0) {
 990		ret = gpu->irq;
 991		DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret);
 992		goto fail;
 993	}
 994
 995	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
 996			IRQF_TRIGGER_HIGH, gpu->name, gpu);
 997	if (ret) {
 998		DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
 999		goto fail;
1000	}
1001
1002	ret = get_clocks(pdev, gpu);
1003	if (ret)
1004		goto fail;
 
 
 
 
1005
1006	gpu->ebi1_clk = msm_clk_get(pdev, "bus");
1007	DBG("ebi1_clk: %p", gpu->ebi1_clk);
1008	if (IS_ERR(gpu->ebi1_clk))
1009		gpu->ebi1_clk = NULL;
1010
1011	/* Acquire regulators: */
1012	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
1013	DBG("gpu_reg: %p", gpu->gpu_reg);
1014	if (IS_ERR(gpu->gpu_reg))
1015		gpu->gpu_reg = NULL;
1016
1017	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
1018	DBG("gpu_cx: %p", gpu->gpu_cx);
1019	if (IS_ERR(gpu->gpu_cx))
1020		gpu->gpu_cx = NULL;
1021
1022	gpu->pdev = pdev;
1023	platform_set_drvdata(pdev, &gpu->adreno_smmu);
1024
1025	msm_devfreq_init(gpu);
 
 
 
 
 
 
 
 
 
 
 
1026
 
 
 
 
1027
1028	gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
1029
1030	if (gpu->aspace == NULL)
1031		DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
1032	else if (IS_ERR(gpu->aspace)) {
1033		ret = PTR_ERR(gpu->aspace);
 
 
 
 
1034		goto fail;
1035	}
1036
1037	memptrs = msm_gem_kernel_new(drm,
1038		sizeof(struct msm_rbmemptrs) * nr_rings,
1039		check_apriv(gpu, MSM_BO_UNCACHED), gpu->aspace, &gpu->memptrs_bo,
1040		&memptrs_iova);
1041
1042	if (IS_ERR(memptrs)) {
1043		ret = PTR_ERR(memptrs);
1044		DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
1045		goto fail;
1046	}
1047
1048	msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
1049
1050	if (nr_rings > ARRAY_SIZE(gpu->rb)) {
1051		DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
1052			ARRAY_SIZE(gpu->rb));
1053		nr_rings = ARRAY_SIZE(gpu->rb);
1054	}
1055
1056	/* Create ringbuffer(s): */
1057	for (i = 0; i < nr_rings; i++) {
1058		gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
1059
1060		if (IS_ERR(gpu->rb[i])) {
1061			ret = PTR_ERR(gpu->rb[i]);
1062			DRM_DEV_ERROR(drm->dev,
1063				"could not create ringbuffer %d: %d\n", i, ret);
1064			goto fail;
1065		}
1066
1067		memptrs += sizeof(struct msm_rbmemptrs);
1068		memptrs_iova += sizeof(struct msm_rbmemptrs);
1069	}
1070
1071	gpu->nr_rings = nr_rings;
1072
1073	return 0;
1074
1075fail:
1076	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++)  {
1077		msm_ringbuffer_destroy(gpu->rb[i]);
1078		gpu->rb[i] = NULL;
1079	}
1080
1081	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace, false);
1082
1083	platform_set_drvdata(pdev, NULL);
1084	return ret;
1085}
1086
1087void msm_gpu_cleanup(struct msm_gpu *gpu)
1088{
1089	int i;
1090
1091	DBG("%s", gpu->name);
1092
1093	WARN_ON(!list_empty(&gpu->active_list));
1094
1095	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
1096		msm_ringbuffer_destroy(gpu->rb[i]);
1097		gpu->rb[i] = NULL;
1098	}
1099
1100	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace, false);
1101
1102	if (!IS_ERR_OR_NULL(gpu->aspace)) {
1103		gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
1104		msm_gem_address_space_put(gpu->aspace);
1105	}
1106
1107	if (gpu->worker) {
1108		kthread_destroy_worker(gpu->worker);
 
 
1109	}
1110
1111	devfreq_cooling_unregister(gpu->cooling);
 
1112}