xref: /openbmc/linux/drivers/gpu/host1x/syncpt.c (revision bc000245)
1 /*
2  * Tegra host1x Syncpoints
3  *
4  * Copyright (c) 2010-2013, NVIDIA Corporation.
5  *
6  * This program is free software; you can redistribute it and/or modify it
7  * under the terms and conditions of the GNU General Public License,
8  * version 2, as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope it will be useful, but WITHOUT
11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
13  * more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
17  */
18 
19 #include <linux/module.h>
20 #include <linux/device.h>
21 #include <linux/slab.h>
22 
23 #include <trace/events/host1x.h>
24 
25 #include "syncpt.h"
26 #include "dev.h"
27 #include "intr.h"
28 #include "debug.h"
29 
30 #define SYNCPT_CHECK_PERIOD (2 * HZ)
31 #define MAX_STUCK_CHECK_COUNT 15
32 
33 static struct host1x_syncpt_base *
34 host1x_syncpt_base_request(struct host1x *host)
35 {
36 	struct host1x_syncpt_base *bases = host->bases;
37 	unsigned int i;
38 
39 	for (i = 0; i < host->info->nb_bases; i++)
40 		if (!bases[i].requested)
41 			break;
42 
43 	if (i >= host->info->nb_bases)
44 		return NULL;
45 
46 	bases[i].requested = true;
47 	return &bases[i];
48 }
49 
50 static void host1x_syncpt_base_free(struct host1x_syncpt_base *base)
51 {
52 	if (base)
53 		base->requested = false;
54 }
55 
56 static struct host1x_syncpt *host1x_syncpt_alloc(struct host1x *host,
57 						 struct device *dev,
58 						 unsigned long flags)
59 {
60 	int i;
61 	struct host1x_syncpt *sp = host->syncpt;
62 	char *name;
63 
64 	for (i = 0; i < host->info->nb_pts && sp->name; i++, sp++)
65 		;
66 
67 	if (i >= host->info->nb_pts)
68 		return NULL;
69 
70 	if (flags & HOST1X_SYNCPT_HAS_BASE) {
71 		sp->base = host1x_syncpt_base_request(host);
72 		if (!sp->base)
73 			return NULL;
74 	}
75 
76 	name = kasprintf(GFP_KERNEL, "%02d-%s", sp->id,
77 			dev ? dev_name(dev) : NULL);
78 	if (!name)
79 		return NULL;
80 
81 	sp->dev = dev;
82 	sp->name = name;
83 
84 	if (flags & HOST1X_SYNCPT_CLIENT_MANAGED)
85 		sp->client_managed = true;
86 	else
87 		sp->client_managed = false;
88 
89 	return sp;
90 }
91 
92 u32 host1x_syncpt_id(struct host1x_syncpt *sp)
93 {
94 	return sp->id;
95 }
96 
97 /*
98  * Updates the value sent to hardware.
99  */
100 u32 host1x_syncpt_incr_max(struct host1x_syncpt *sp, u32 incrs)
101 {
102 	return (u32)atomic_add_return(incrs, &sp->max_val);
103 }
104 
105  /*
106  * Write cached syncpoint and waitbase values to hardware.
107  */
108 void host1x_syncpt_restore(struct host1x *host)
109 {
110 	struct host1x_syncpt *sp_base = host->syncpt;
111 	u32 i;
112 
113 	for (i = 0; i < host1x_syncpt_nb_pts(host); i++)
114 		host1x_hw_syncpt_restore(host, sp_base + i);
115 	for (i = 0; i < host1x_syncpt_nb_bases(host); i++)
116 		host1x_hw_syncpt_restore_wait_base(host, sp_base + i);
117 	wmb();
118 }
119 
120 /*
121  * Update the cached syncpoint and waitbase values by reading them
122  * from the registers.
123   */
124 void host1x_syncpt_save(struct host1x *host)
125 {
126 	struct host1x_syncpt *sp_base = host->syncpt;
127 	u32 i;
128 
129 	for (i = 0; i < host1x_syncpt_nb_pts(host); i++) {
130 		if (host1x_syncpt_client_managed(sp_base + i))
131 			host1x_hw_syncpt_load(host, sp_base + i);
132 		else
133 			WARN_ON(!host1x_syncpt_idle(sp_base + i));
134 	}
135 
136 	for (i = 0; i < host1x_syncpt_nb_bases(host); i++)
137 		host1x_hw_syncpt_load_wait_base(host, sp_base + i);
138 }
139 
140 /*
141  * Updates the cached syncpoint value by reading a new value from the hardware
142  * register
143  */
144 u32 host1x_syncpt_load(struct host1x_syncpt *sp)
145 {
146 	u32 val;
147 	val = host1x_hw_syncpt_load(sp->host, sp);
148 	trace_host1x_syncpt_load_min(sp->id, val);
149 
150 	return val;
151 }
152 
153 /*
154  * Get the current syncpoint base
155  */
156 u32 host1x_syncpt_load_wait_base(struct host1x_syncpt *sp)
157 {
158 	u32 val;
159 	host1x_hw_syncpt_load_wait_base(sp->host, sp);
160 	val = sp->base_val;
161 	return val;
162 }
163 
164 /*
165  * Increment syncpoint value from cpu, updating cache
166  */
167 int host1x_syncpt_incr(struct host1x_syncpt *sp)
168 {
169 	return host1x_hw_syncpt_cpu_incr(sp->host, sp);
170 }
171 
172 /*
173  * Updated sync point form hardware, and returns true if syncpoint is expired,
174  * false if we may need to wait
175  */
176 static bool syncpt_load_min_is_expired(struct host1x_syncpt *sp, u32 thresh)
177 {
178 	host1x_hw_syncpt_load(sp->host, sp);
179 	return host1x_syncpt_is_expired(sp, thresh);
180 }
181 
182 /*
183  * Main entrypoint for syncpoint value waits.
184  */
185 int host1x_syncpt_wait(struct host1x_syncpt *sp, u32 thresh, long timeout,
186 			u32 *value)
187 {
188 	DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
189 	void *ref;
190 	struct host1x_waitlist *waiter;
191 	int err = 0, check_count = 0;
192 	u32 val;
193 
194 	if (value)
195 		*value = 0;
196 
197 	/* first check cache */
198 	if (host1x_syncpt_is_expired(sp, thresh)) {
199 		if (value)
200 			*value = host1x_syncpt_load(sp);
201 		return 0;
202 	}
203 
204 	/* try to read from register */
205 	val = host1x_hw_syncpt_load(sp->host, sp);
206 	if (host1x_syncpt_is_expired(sp, thresh)) {
207 		if (value)
208 			*value = val;
209 		goto done;
210 	}
211 
212 	if (!timeout) {
213 		err = -EAGAIN;
214 		goto done;
215 	}
216 
217 	/* allocate a waiter */
218 	waiter = kzalloc(sizeof(*waiter), GFP_KERNEL);
219 	if (!waiter) {
220 		err = -ENOMEM;
221 		goto done;
222 	}
223 
224 	/* schedule a wakeup when the syncpoint value is reached */
225 	err = host1x_intr_add_action(sp->host, sp->id, thresh,
226 				     HOST1X_INTR_ACTION_WAKEUP_INTERRUPTIBLE,
227 				     &wq, waiter, &ref);
228 	if (err)
229 		goto done;
230 
231 	err = -EAGAIN;
232 	/* Caller-specified timeout may be impractically low */
233 	if (timeout < 0)
234 		timeout = LONG_MAX;
235 
236 	/* wait for the syncpoint, or timeout, or signal */
237 	while (timeout) {
238 		long check = min_t(long, SYNCPT_CHECK_PERIOD, timeout);
239 		int remain = wait_event_interruptible_timeout(wq,
240 				syncpt_load_min_is_expired(sp, thresh),
241 				check);
242 		if (remain > 0 || host1x_syncpt_is_expired(sp, thresh)) {
243 			if (value)
244 				*value = host1x_syncpt_load(sp);
245 			err = 0;
246 			break;
247 		}
248 		if (remain < 0) {
249 			err = remain;
250 			break;
251 		}
252 		timeout -= check;
253 		if (timeout && check_count <= MAX_STUCK_CHECK_COUNT) {
254 			dev_warn(sp->host->dev,
255 				"%s: syncpoint id %d (%s) stuck waiting %d, timeout=%ld\n",
256 				 current->comm, sp->id, sp->name,
257 				 thresh, timeout);
258 
259 			host1x_debug_dump_syncpts(sp->host);
260 			if (check_count == MAX_STUCK_CHECK_COUNT)
261 				host1x_debug_dump(sp->host);
262 			check_count++;
263 		}
264 	}
265 	host1x_intr_put_ref(sp->host, sp->id, ref);
266 
267 done:
268 	return err;
269 }
270 EXPORT_SYMBOL(host1x_syncpt_wait);
271 
272 /*
273  * Returns true if syncpoint is expired, false if we may need to wait
274  */
275 bool host1x_syncpt_is_expired(struct host1x_syncpt *sp, u32 thresh)
276 {
277 	u32 current_val;
278 	u32 future_val;
279 	smp_rmb();
280 	current_val = (u32)atomic_read(&sp->min_val);
281 	future_val = (u32)atomic_read(&sp->max_val);
282 
283 	/* Note the use of unsigned arithmetic here (mod 1<<32).
284 	 *
285 	 * c = current_val = min_val	= the current value of the syncpoint.
286 	 * t = thresh			= the value we are checking
287 	 * f = future_val  = max_val	= the value c will reach when all
288 	 *				  outstanding increments have completed.
289 	 *
290 	 * Note that c always chases f until it reaches f.
291 	 *
292 	 * Dtf = (f - t)
293 	 * Dtc = (c - t)
294 	 *
295 	 *  Consider all cases:
296 	 *
297 	 *	A) .....c..t..f.....	Dtf < Dtc	need to wait
298 	 *	B) .....c.....f..t..	Dtf > Dtc	expired
299 	 *	C) ..t..c.....f.....	Dtf > Dtc	expired	   (Dct very large)
300 	 *
301 	 *  Any case where f==c: always expired (for any t).	Dtf == Dcf
302 	 *  Any case where t==c: always expired (for any f).	Dtf >= Dtc (because Dtc==0)
303 	 *  Any case where t==f!=c: always wait.		Dtf <  Dtc (because Dtf==0,
304 	 *							Dtc!=0)
305 	 *
306 	 *  Other cases:
307 	 *
308 	 *	A) .....t..f..c.....	Dtf < Dtc	need to wait
309 	 *	A) .....f..c..t.....	Dtf < Dtc	need to wait
310 	 *	A) .....f..t..c.....	Dtf > Dtc	expired
311 	 *
312 	 *   So:
313 	 *	   Dtf >= Dtc implies EXPIRED	(return true)
314 	 *	   Dtf <  Dtc implies WAIT	(return false)
315 	 *
316 	 * Note: If t is expired then we *cannot* wait on it. We would wait
317 	 * forever (hang the system).
318 	 *
319 	 * Note: do NOT get clever and remove the -thresh from both sides. It
320 	 * is NOT the same.
321 	 *
322 	 * If future valueis zero, we have a client managed sync point. In that
323 	 * case we do a direct comparison.
324 	 */
325 	if (!host1x_syncpt_client_managed(sp))
326 		return future_val - thresh >= current_val - thresh;
327 	else
328 		return (s32)(current_val - thresh) >= 0;
329 }
330 
331 /* remove a wait pointed to by patch_addr */
332 int host1x_syncpt_patch_wait(struct host1x_syncpt *sp, void *patch_addr)
333 {
334 	return host1x_hw_syncpt_patch_wait(sp->host, sp, patch_addr);
335 }
336 
337 int host1x_syncpt_init(struct host1x *host)
338 {
339 	struct host1x_syncpt_base *bases;
340 	struct host1x_syncpt *syncpt;
341 	int i;
342 
343 	syncpt = devm_kzalloc(host->dev, sizeof(*syncpt) * host->info->nb_pts,
344 			      GFP_KERNEL);
345 	if (!syncpt)
346 		return -ENOMEM;
347 
348 	bases = devm_kzalloc(host->dev, sizeof(*bases) * host->info->nb_bases,
349 			     GFP_KERNEL);
350 	if (!bases)
351 		return -ENOMEM;
352 
353 	for (i = 0; i < host->info->nb_pts; i++) {
354 		syncpt[i].id = i;
355 		syncpt[i].host = host;
356 	}
357 
358 	for (i = 0; i < host->info->nb_bases; i++)
359 		bases[i].id = i;
360 
361 	host->syncpt = syncpt;
362 	host->bases = bases;
363 
364 	host1x_syncpt_restore(host);
365 
366 	/* Allocate sync point to use for clearing waits for expired fences */
367 	host->nop_sp = host1x_syncpt_alloc(host, NULL, 0);
368 	if (!host->nop_sp)
369 		return -ENOMEM;
370 
371 	return 0;
372 }
373 
374 struct host1x_syncpt *host1x_syncpt_request(struct device *dev,
375 					    unsigned long flags)
376 {
377 	struct host1x *host = dev_get_drvdata(dev->parent);
378 	return host1x_syncpt_alloc(host, dev, flags);
379 }
380 
381 void host1x_syncpt_free(struct host1x_syncpt *sp)
382 {
383 	if (!sp)
384 		return;
385 
386 	host1x_syncpt_base_free(sp->base);
387 	kfree(sp->name);
388 	sp->base = NULL;
389 	sp->dev = NULL;
390 	sp->name = NULL;
391 	sp->client_managed = false;
392 }
393 
394 void host1x_syncpt_deinit(struct host1x *host)
395 {
396 	int i;
397 	struct host1x_syncpt *sp = host->syncpt;
398 	for (i = 0; i < host->info->nb_pts; i++, sp++)
399 		kfree(sp->name);
400 }
401 
402 /*
403  * Read max. It indicates how many operations there are in queue, either in
404  * channel or in a software thread.
405  * */
406 u32 host1x_syncpt_read_max(struct host1x_syncpt *sp)
407 {
408 	smp_rmb();
409 	return (u32)atomic_read(&sp->max_val);
410 }
411 
412 /*
413  * Read min, which is a shadow of the current sync point value in hardware.
414  */
415 u32 host1x_syncpt_read_min(struct host1x_syncpt *sp)
416 {
417 	smp_rmb();
418 	return (u32)atomic_read(&sp->min_val);
419 }
420 
421 int host1x_syncpt_nb_pts(struct host1x *host)
422 {
423 	return host->info->nb_pts;
424 }
425 
426 int host1x_syncpt_nb_bases(struct host1x *host)
427 {
428 	return host->info->nb_bases;
429 }
430 
431 int host1x_syncpt_nb_mlocks(struct host1x *host)
432 {
433 	return host->info->nb_mlocks;
434 }
435 
436 struct host1x_syncpt *host1x_syncpt_get(struct host1x *host, u32 id)
437 {
438 	if (host->info->nb_pts < id)
439 		return NULL;
440 	return host->syncpt + id;
441 }
442 
443 struct host1x_syncpt_base *host1x_syncpt_get_base(struct host1x_syncpt *sp)
444 {
445 	return sp ? sp->base : NULL;
446 }
447 
448 u32 host1x_syncpt_base_id(struct host1x_syncpt_base *base)
449 {
450 	return base->id;
451 }
452