1 /*
2  * Copyright 2009 Jerome Glisse.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sub license, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  *
25  */
26 /*
27  * Authors:
28  *    Jerome Glisse <glisse@freedesktop.org>
29  *    Dave Airlie
30  */
31 #include <linux/seq_file.h>
32 #include <linux/atomic.h>
33 #include <linux/wait.h>
34 #include <linux/list.h>
35 #include <linux/kref.h>
36 #include <linux/slab.h>
37 #include "drmP.h"
38 #include "drm.h"
39 #include "radeon_reg.h"
40 #include "radeon.h"
41 #include "radeon_trace.h"
42 
43 static void radeon_fence_write(struct radeon_device *rdev, u32 seq, int ring)
44 {
45 	if (rdev->wb.enabled) {
46 		*rdev->fence_drv[ring].cpu_addr = cpu_to_le32(seq);
47 	} else {
48 		WREG32(rdev->fence_drv[ring].scratch_reg, seq);
49 	}
50 }
51 
52 static u32 radeon_fence_read(struct radeon_device *rdev, int ring)
53 {
54 	u32 seq = 0;
55 
56 	if (rdev->wb.enabled) {
57 		seq = le32_to_cpu(*rdev->fence_drv[ring].cpu_addr);
58 	} else {
59 		seq = RREG32(rdev->fence_drv[ring].scratch_reg);
60 	}
61 	return seq;
62 }
63 
64 int radeon_fence_emit(struct radeon_device *rdev, struct radeon_fence *fence)
65 {
66 	unsigned long irq_flags;
67 
68 	write_lock_irqsave(&rdev->fence_lock, irq_flags);
69 	if (fence->emitted) {
70 		write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
71 		return 0;
72 	}
73 	fence->seq = atomic_add_return(1, &rdev->fence_drv[fence->ring].seq);
74 	if (!rdev->ring[fence->ring].ready)
75 		/* FIXME: cp is not running assume everythings is done right
76 		 * away
77 		 */
78 		radeon_fence_write(rdev, fence->seq, fence->ring);
79 	else
80 		radeon_fence_ring_emit(rdev, fence->ring, fence);
81 
82 	trace_radeon_fence_emit(rdev->ddev, fence->seq);
83 	fence->emitted = true;
84 	list_move_tail(&fence->list, &rdev->fence_drv[fence->ring].emitted);
85 	write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
86 	return 0;
87 }
88 
89 static bool radeon_fence_poll_locked(struct radeon_device *rdev, int ring)
90 {
91 	struct radeon_fence *fence;
92 	struct list_head *i, *n;
93 	uint32_t seq;
94 	bool wake = false;
95 	unsigned long cjiffies;
96 
97 	seq = radeon_fence_read(rdev, ring);
98 	if (seq != rdev->fence_drv[ring].last_seq) {
99 		rdev->fence_drv[ring].last_seq = seq;
100 		rdev->fence_drv[ring].last_jiffies = jiffies;
101 		rdev->fence_drv[ring].last_timeout = RADEON_FENCE_JIFFIES_TIMEOUT;
102 	} else {
103 		cjiffies = jiffies;
104 		if (time_after(cjiffies, rdev->fence_drv[ring].last_jiffies)) {
105 			cjiffies -= rdev->fence_drv[ring].last_jiffies;
106 			if (time_after(rdev->fence_drv[ring].last_timeout, cjiffies)) {
107 				/* update the timeout */
108 				rdev->fence_drv[ring].last_timeout -= cjiffies;
109 			} else {
110 				/* the 500ms timeout is elapsed we should test
111 				 * for GPU lockup
112 				 */
113 				rdev->fence_drv[ring].last_timeout = 1;
114 			}
115 		} else {
116 			/* wrap around update last jiffies, we will just wait
117 			 * a little longer
118 			 */
119 			rdev->fence_drv[ring].last_jiffies = cjiffies;
120 		}
121 		return false;
122 	}
123 	n = NULL;
124 	list_for_each(i, &rdev->fence_drv[ring].emitted) {
125 		fence = list_entry(i, struct radeon_fence, list);
126 		if (fence->seq == seq) {
127 			n = i;
128 			break;
129 		}
130 	}
131 	/* all fence previous to this one are considered as signaled */
132 	if (n) {
133 		i = n;
134 		do {
135 			n = i->prev;
136 			list_move_tail(i, &rdev->fence_drv[ring].signaled);
137 			fence = list_entry(i, struct radeon_fence, list);
138 			fence->signaled = true;
139 			i = n;
140 		} while (i != &rdev->fence_drv[ring].emitted);
141 		wake = true;
142 	}
143 	return wake;
144 }
145 
146 static void radeon_fence_destroy(struct kref *kref)
147 {
148 	unsigned long irq_flags;
149         struct radeon_fence *fence;
150 
151 	fence = container_of(kref, struct radeon_fence, kref);
152 	write_lock_irqsave(&fence->rdev->fence_lock, irq_flags);
153 	list_del(&fence->list);
154 	fence->emitted = false;
155 	write_unlock_irqrestore(&fence->rdev->fence_lock, irq_flags);
156 	if (fence->semaphore)
157 		radeon_semaphore_free(fence->rdev, fence->semaphore);
158 	kfree(fence);
159 }
160 
161 int radeon_fence_create(struct radeon_device *rdev,
162 			struct radeon_fence **fence,
163 			int ring)
164 {
165 	unsigned long irq_flags;
166 
167 	*fence = kmalloc(sizeof(struct radeon_fence), GFP_KERNEL);
168 	if ((*fence) == NULL) {
169 		return -ENOMEM;
170 	}
171 	kref_init(&((*fence)->kref));
172 	(*fence)->rdev = rdev;
173 	(*fence)->emitted = false;
174 	(*fence)->signaled = false;
175 	(*fence)->seq = 0;
176 	(*fence)->ring = ring;
177 	(*fence)->semaphore = NULL;
178 	INIT_LIST_HEAD(&(*fence)->list);
179 
180 	write_lock_irqsave(&rdev->fence_lock, irq_flags);
181 	list_add_tail(&(*fence)->list, &rdev->fence_drv[ring].created);
182 	write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
183 	return 0;
184 }
185 
186 bool radeon_fence_signaled(struct radeon_fence *fence)
187 {
188 	unsigned long irq_flags;
189 	bool signaled = false;
190 
191 	if (!fence)
192 		return true;
193 
194 	if (fence->rdev->gpu_lockup)
195 		return true;
196 
197 	write_lock_irqsave(&fence->rdev->fence_lock, irq_flags);
198 	signaled = fence->signaled;
199 	/* if we are shuting down report all fence as signaled */
200 	if (fence->rdev->shutdown) {
201 		signaled = true;
202 	}
203 	if (!fence->emitted) {
204 		WARN(1, "Querying an unemitted fence : %p !\n", fence);
205 		signaled = true;
206 	}
207 	if (!signaled) {
208 		radeon_fence_poll_locked(fence->rdev, fence->ring);
209 		signaled = fence->signaled;
210 	}
211 	write_unlock_irqrestore(&fence->rdev->fence_lock, irq_flags);
212 	return signaled;
213 }
214 
215 int radeon_fence_wait(struct radeon_fence *fence, bool intr)
216 {
217 	struct radeon_device *rdev;
218 	unsigned long irq_flags, timeout;
219 	u32 seq;
220 	int r;
221 
222 	if (fence == NULL) {
223 		WARN(1, "Querying an invalid fence : %p !\n", fence);
224 		return 0;
225 	}
226 	rdev = fence->rdev;
227 	if (radeon_fence_signaled(fence)) {
228 		return 0;
229 	}
230 	timeout = rdev->fence_drv[fence->ring].last_timeout;
231 retry:
232 	/* save current sequence used to check for GPU lockup */
233 	seq = rdev->fence_drv[fence->ring].last_seq;
234 	trace_radeon_fence_wait_begin(rdev->ddev, seq);
235 	if (intr) {
236 		radeon_irq_kms_sw_irq_get(rdev, fence->ring);
237 		r = wait_event_interruptible_timeout(rdev->fence_drv[fence->ring].queue,
238 				radeon_fence_signaled(fence), timeout);
239 		radeon_irq_kms_sw_irq_put(rdev, fence->ring);
240 		if (unlikely(r < 0)) {
241 			return r;
242 		}
243 	} else {
244 		radeon_irq_kms_sw_irq_get(rdev, fence->ring);
245 		r = wait_event_timeout(rdev->fence_drv[fence->ring].queue,
246 			 radeon_fence_signaled(fence), timeout);
247 		radeon_irq_kms_sw_irq_put(rdev, fence->ring);
248 	}
249 	trace_radeon_fence_wait_end(rdev->ddev, seq);
250 	if (unlikely(!radeon_fence_signaled(fence))) {
251 		/* we were interrupted for some reason and fence isn't
252 		 * isn't signaled yet, resume wait
253 		 */
254 		if (r) {
255 			timeout = r;
256 			goto retry;
257 		}
258 		/* don't protect read access to rdev->fence_drv[t].last_seq
259 		 * if we experiencing a lockup the value doesn't change
260 		 */
261 		if (seq == rdev->fence_drv[fence->ring].last_seq &&
262 		    radeon_gpu_is_lockup(rdev, &rdev->ring[fence->ring])) {
263 			/* good news we believe it's a lockup */
264 			printk(KERN_WARNING "GPU lockup (waiting for 0x%08X last fence id 0x%08X)\n",
265 			     fence->seq, seq);
266 			/* FIXME: what should we do ? marking everyone
267 			 * as signaled for now
268 			 */
269 			rdev->gpu_lockup = true;
270 			r = radeon_gpu_reset(rdev);
271 			if (r)
272 				return r;
273 			radeon_fence_write(rdev, fence->seq, fence->ring);
274 			rdev->gpu_lockup = false;
275 		}
276 		timeout = RADEON_FENCE_JIFFIES_TIMEOUT;
277 		write_lock_irqsave(&rdev->fence_lock, irq_flags);
278 		rdev->fence_drv[fence->ring].last_timeout = RADEON_FENCE_JIFFIES_TIMEOUT;
279 		rdev->fence_drv[fence->ring].last_jiffies = jiffies;
280 		write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
281 		goto retry;
282 	}
283 	return 0;
284 }
285 
286 int radeon_fence_wait_next(struct radeon_device *rdev, int ring)
287 {
288 	unsigned long irq_flags;
289 	struct radeon_fence *fence;
290 	int r;
291 
292 	if (rdev->gpu_lockup) {
293 		return 0;
294 	}
295 	write_lock_irqsave(&rdev->fence_lock, irq_flags);
296 	if (list_empty(&rdev->fence_drv[ring].emitted)) {
297 		write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
298 		return 0;
299 	}
300 	fence = list_entry(rdev->fence_drv[ring].emitted.next,
301 			   struct radeon_fence, list);
302 	radeon_fence_ref(fence);
303 	write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
304 	r = radeon_fence_wait(fence, false);
305 	radeon_fence_unref(&fence);
306 	return r;
307 }
308 
309 int radeon_fence_wait_last(struct radeon_device *rdev, int ring)
310 {
311 	unsigned long irq_flags;
312 	struct radeon_fence *fence;
313 	int r;
314 
315 	if (rdev->gpu_lockup) {
316 		return 0;
317 	}
318 	write_lock_irqsave(&rdev->fence_lock, irq_flags);
319 	if (list_empty(&rdev->fence_drv[ring].emitted)) {
320 		write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
321 		return 0;
322 	}
323 	fence = list_entry(rdev->fence_drv[ring].emitted.prev,
324 			   struct radeon_fence, list);
325 	radeon_fence_ref(fence);
326 	write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
327 	r = radeon_fence_wait(fence, false);
328 	radeon_fence_unref(&fence);
329 	return r;
330 }
331 
332 struct radeon_fence *radeon_fence_ref(struct radeon_fence *fence)
333 {
334 	kref_get(&fence->kref);
335 	return fence;
336 }
337 
338 void radeon_fence_unref(struct radeon_fence **fence)
339 {
340 	struct radeon_fence *tmp = *fence;
341 
342 	*fence = NULL;
343 	if (tmp) {
344 		kref_put(&tmp->kref, radeon_fence_destroy);
345 	}
346 }
347 
348 void radeon_fence_process(struct radeon_device *rdev, int ring)
349 {
350 	unsigned long irq_flags;
351 	bool wake;
352 
353 	write_lock_irqsave(&rdev->fence_lock, irq_flags);
354 	wake = radeon_fence_poll_locked(rdev, ring);
355 	write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
356 	if (wake) {
357 		wake_up_all(&rdev->fence_drv[ring].queue);
358 	}
359 }
360 
361 int radeon_fence_count_emitted(struct radeon_device *rdev, int ring)
362 {
363 	unsigned long irq_flags;
364 	int not_processed = 0;
365 
366 	read_lock_irqsave(&rdev->fence_lock, irq_flags);
367 	if (!rdev->fence_drv[ring].initialized) {
368 		read_unlock_irqrestore(&rdev->fence_lock, irq_flags);
369 		return 0;
370 	}
371 
372 	if (!list_empty(&rdev->fence_drv[ring].emitted)) {
373 		struct list_head *ptr;
374 		list_for_each(ptr, &rdev->fence_drv[ring].emitted) {
375 			/* count up to 3, that's enought info */
376 			if (++not_processed >= 3)
377 				break;
378 		}
379 	}
380 	read_unlock_irqrestore(&rdev->fence_lock, irq_flags);
381 	return not_processed;
382 }
383 
384 int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring)
385 {
386 	unsigned long irq_flags;
387 	uint64_t index;
388 	int r;
389 
390 	write_lock_irqsave(&rdev->fence_lock, irq_flags);
391 	radeon_scratch_free(rdev, rdev->fence_drv[ring].scratch_reg);
392 	if (rdev->wb.use_event) {
393 		rdev->fence_drv[ring].scratch_reg = 0;
394 		index = R600_WB_EVENT_OFFSET + ring * 4;
395 	} else {
396 		r = radeon_scratch_get(rdev, &rdev->fence_drv[ring].scratch_reg);
397 		if (r) {
398 			dev_err(rdev->dev, "fence failed to get scratch register\n");
399 			write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
400 			return r;
401 		}
402 		index = RADEON_WB_SCRATCH_OFFSET +
403 			rdev->fence_drv[ring].scratch_reg -
404 			rdev->scratch.reg_base;
405 	}
406 	rdev->fence_drv[ring].cpu_addr = &rdev->wb.wb[index/4];
407 	rdev->fence_drv[ring].gpu_addr = rdev->wb.gpu_addr + index;
408 	radeon_fence_write(rdev, atomic_read(&rdev->fence_drv[ring].seq), ring);
409 	rdev->fence_drv[ring].initialized = true;
410 	DRM_INFO("fence driver on ring %d use gpu addr 0x%08Lx and cpu addr 0x%p\n",
411 		 ring, rdev->fence_drv[ring].gpu_addr, rdev->fence_drv[ring].cpu_addr);
412 	write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
413 	return 0;
414 }
415 
416 static void radeon_fence_driver_init_ring(struct radeon_device *rdev, int ring)
417 {
418 	rdev->fence_drv[ring].scratch_reg = -1;
419 	rdev->fence_drv[ring].cpu_addr = NULL;
420 	rdev->fence_drv[ring].gpu_addr = 0;
421 	atomic_set(&rdev->fence_drv[ring].seq, 0);
422 	INIT_LIST_HEAD(&rdev->fence_drv[ring].created);
423 	INIT_LIST_HEAD(&rdev->fence_drv[ring].emitted);
424 	INIT_LIST_HEAD(&rdev->fence_drv[ring].signaled);
425 	init_waitqueue_head(&rdev->fence_drv[ring].queue);
426 	rdev->fence_drv[ring].initialized = false;
427 }
428 
429 int radeon_fence_driver_init(struct radeon_device *rdev)
430 {
431 	unsigned long irq_flags;
432 	int ring;
433 
434 	write_lock_irqsave(&rdev->fence_lock, irq_flags);
435 	for (ring = 0; ring < RADEON_NUM_RINGS; ring++) {
436 		radeon_fence_driver_init_ring(rdev, ring);
437 	}
438 	write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
439 	if (radeon_debugfs_fence_init(rdev)) {
440 		dev_err(rdev->dev, "fence debugfs file creation failed\n");
441 	}
442 	return 0;
443 }
444 
445 void radeon_fence_driver_fini(struct radeon_device *rdev)
446 {
447 	unsigned long irq_flags;
448 	int ring;
449 
450 	for (ring = 0; ring < RADEON_NUM_RINGS; ring++) {
451 		if (!rdev->fence_drv[ring].initialized)
452 			continue;
453 		radeon_fence_wait_last(rdev, ring);
454 		wake_up_all(&rdev->fence_drv[ring].queue);
455 		write_lock_irqsave(&rdev->fence_lock, irq_flags);
456 		radeon_scratch_free(rdev, rdev->fence_drv[ring].scratch_reg);
457 		write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
458 		rdev->fence_drv[ring].initialized = false;
459 	}
460 }
461 
462 
463 /*
464  * Fence debugfs
465  */
466 #if defined(CONFIG_DEBUG_FS)
467 static int radeon_debugfs_fence_info(struct seq_file *m, void *data)
468 {
469 	struct drm_info_node *node = (struct drm_info_node *)m->private;
470 	struct drm_device *dev = node->minor->dev;
471 	struct radeon_device *rdev = dev->dev_private;
472 	struct radeon_fence *fence;
473 	int i;
474 
475 	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
476 		if (!rdev->fence_drv[i].initialized)
477 			continue;
478 
479 		seq_printf(m, "--- ring %d ---\n", i);
480 		seq_printf(m, "Last signaled fence 0x%08X\n",
481 			   radeon_fence_read(rdev, i));
482 		if (!list_empty(&rdev->fence_drv[i].emitted)) {
483 			fence = list_entry(rdev->fence_drv[i].emitted.prev,
484 					   struct radeon_fence, list);
485 			seq_printf(m, "Last emitted fence %p with 0x%08X\n",
486 				   fence,  fence->seq);
487 		}
488 	}
489 	return 0;
490 }
491 
492 static struct drm_info_list radeon_debugfs_fence_list[] = {
493 	{"radeon_fence_info", &radeon_debugfs_fence_info, 0, NULL},
494 };
495 #endif
496 
497 int radeon_debugfs_fence_init(struct radeon_device *rdev)
498 {
499 #if defined(CONFIG_DEBUG_FS)
500 	return radeon_debugfs_add_files(rdev, radeon_debugfs_fence_list, 1);
501 #else
502 	return 0;
503 #endif
504 }
505