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}
704