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)
44 {
45 	if (rdev->wb.enabled) {
46 		u32 scratch_index;
47 		if (rdev->wb.use_event)
48 			scratch_index = R600_WB_EVENT_OFFSET + rdev->fence_drv.scratch_reg - rdev->scratch.reg_base;
49 		else
50 			scratch_index = RADEON_WB_SCRATCH_OFFSET + rdev->fence_drv.scratch_reg - rdev->scratch.reg_base;
51 		rdev->wb.wb[scratch_index/4] = cpu_to_le32(seq);;
52 	} else
53 		WREG32(rdev->fence_drv.scratch_reg, seq);
54 }
55 
56 static u32 radeon_fence_read(struct radeon_device *rdev)
57 {
58 	u32 seq;
59 
60 	if (rdev->wb.enabled) {
61 		u32 scratch_index;
62 		if (rdev->wb.use_event)
63 			scratch_index = R600_WB_EVENT_OFFSET + rdev->fence_drv.scratch_reg - rdev->scratch.reg_base;
64 		else
65 			scratch_index = RADEON_WB_SCRATCH_OFFSET + rdev->fence_drv.scratch_reg - rdev->scratch.reg_base;
66 		seq = le32_to_cpu(rdev->wb.wb[scratch_index/4]);
67 	} else
68 		seq = RREG32(rdev->fence_drv.scratch_reg);
69 	return seq;
70 }
71 
72 int radeon_fence_emit(struct radeon_device *rdev, struct radeon_fence *fence)
73 {
74 	unsigned long irq_flags;
75 
76 	write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
77 	if (fence->emited) {
78 		write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
79 		return 0;
80 	}
81 	fence->seq = atomic_add_return(1, &rdev->fence_drv.seq);
82 	if (!rdev->cp.ready)
83 		/* FIXME: cp is not running assume everythings is done right
84 		 * away
85 		 */
86 		radeon_fence_write(rdev, fence->seq);
87 	else
88 		radeon_fence_ring_emit(rdev, fence);
89 
90 	trace_radeon_fence_emit(rdev->ddev, fence->seq);
91 	fence->emited = true;
92 	list_move_tail(&fence->list, &rdev->fence_drv.emited);
93 	write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
94 	return 0;
95 }
96 
97 static bool radeon_fence_poll_locked(struct radeon_device *rdev)
98 {
99 	struct radeon_fence *fence;
100 	struct list_head *i, *n;
101 	uint32_t seq;
102 	bool wake = false;
103 	unsigned long cjiffies;
104 
105 	seq = radeon_fence_read(rdev);
106 	if (seq != rdev->fence_drv.last_seq) {
107 		rdev->fence_drv.last_seq = seq;
108 		rdev->fence_drv.last_jiffies = jiffies;
109 		rdev->fence_drv.last_timeout = RADEON_FENCE_JIFFIES_TIMEOUT;
110 	} else {
111 		cjiffies = jiffies;
112 		if (time_after(cjiffies, rdev->fence_drv.last_jiffies)) {
113 			cjiffies -= rdev->fence_drv.last_jiffies;
114 			if (time_after(rdev->fence_drv.last_timeout, cjiffies)) {
115 				/* update the timeout */
116 				rdev->fence_drv.last_timeout -= cjiffies;
117 			} else {
118 				/* the 500ms timeout is elapsed we should test
119 				 * for GPU lockup
120 				 */
121 				rdev->fence_drv.last_timeout = 1;
122 			}
123 		} else {
124 			/* wrap around update last jiffies, we will just wait
125 			 * a little longer
126 			 */
127 			rdev->fence_drv.last_jiffies = cjiffies;
128 		}
129 		return false;
130 	}
131 	n = NULL;
132 	list_for_each(i, &rdev->fence_drv.emited) {
133 		fence = list_entry(i, struct radeon_fence, list);
134 		if (fence->seq == seq) {
135 			n = i;
136 			break;
137 		}
138 	}
139 	/* all fence previous to this one are considered as signaled */
140 	if (n) {
141 		i = n;
142 		do {
143 			n = i->prev;
144 			list_move_tail(i, &rdev->fence_drv.signaled);
145 			fence = list_entry(i, struct radeon_fence, list);
146 			fence->signaled = true;
147 			i = n;
148 		} while (i != &rdev->fence_drv.emited);
149 		wake = true;
150 	}
151 	return wake;
152 }
153 
154 static void radeon_fence_destroy(struct kref *kref)
155 {
156 	unsigned long irq_flags;
157         struct radeon_fence *fence;
158 
159 	fence = container_of(kref, struct radeon_fence, kref);
160 	write_lock_irqsave(&fence->rdev->fence_drv.lock, irq_flags);
161 	list_del(&fence->list);
162 	fence->emited = false;
163 	write_unlock_irqrestore(&fence->rdev->fence_drv.lock, irq_flags);
164 	kfree(fence);
165 }
166 
167 int radeon_fence_create(struct radeon_device *rdev, struct radeon_fence **fence)
168 {
169 	unsigned long irq_flags;
170 
171 	*fence = kmalloc(sizeof(struct radeon_fence), GFP_KERNEL);
172 	if ((*fence) == NULL) {
173 		return -ENOMEM;
174 	}
175 	kref_init(&((*fence)->kref));
176 	(*fence)->rdev = rdev;
177 	(*fence)->emited = false;
178 	(*fence)->signaled = false;
179 	(*fence)->seq = 0;
180 	INIT_LIST_HEAD(&(*fence)->list);
181 
182 	write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
183 	list_add_tail(&(*fence)->list, &rdev->fence_drv.created);
184 	write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
185 	return 0;
186 }
187 
188 
189 bool radeon_fence_signaled(struct radeon_fence *fence)
190 {
191 	unsigned long irq_flags;
192 	bool signaled = false;
193 
194 	if (!fence)
195 		return true;
196 
197 	if (fence->rdev->gpu_lockup)
198 		return true;
199 
200 	write_lock_irqsave(&fence->rdev->fence_drv.lock, irq_flags);
201 	signaled = fence->signaled;
202 	/* if we are shuting down report all fence as signaled */
203 	if (fence->rdev->shutdown) {
204 		signaled = true;
205 	}
206 	if (!fence->emited) {
207 		WARN(1, "Querying an unemited fence : %p !\n", fence);
208 		signaled = true;
209 	}
210 	if (!signaled) {
211 		radeon_fence_poll_locked(fence->rdev);
212 		signaled = fence->signaled;
213 	}
214 	write_unlock_irqrestore(&fence->rdev->fence_drv.lock, irq_flags);
215 	return signaled;
216 }
217 
218 int radeon_fence_wait(struct radeon_fence *fence, bool intr)
219 {
220 	struct radeon_device *rdev;
221 	unsigned long irq_flags, timeout;
222 	u32 seq;
223 	int r;
224 
225 	if (fence == NULL) {
226 		WARN(1, "Querying an invalid fence : %p !\n", fence);
227 		return 0;
228 	}
229 	rdev = fence->rdev;
230 	if (radeon_fence_signaled(fence)) {
231 		return 0;
232 	}
233 	timeout = rdev->fence_drv.last_timeout;
234 retry:
235 	/* save current sequence used to check for GPU lockup */
236 	seq = rdev->fence_drv.last_seq;
237 	trace_radeon_fence_wait_begin(rdev->ddev, seq);
238 	if (intr) {
239 		radeon_irq_kms_sw_irq_get(rdev);
240 		r = wait_event_interruptible_timeout(rdev->fence_drv.queue,
241 				radeon_fence_signaled(fence), timeout);
242 		radeon_irq_kms_sw_irq_put(rdev);
243 		if (unlikely(r < 0)) {
244 			return r;
245 		}
246 	} else {
247 		radeon_irq_kms_sw_irq_get(rdev);
248 		r = wait_event_timeout(rdev->fence_drv.queue,
249 			 radeon_fence_signaled(fence), timeout);
250 		radeon_irq_kms_sw_irq_put(rdev);
251 	}
252 	trace_radeon_fence_wait_end(rdev->ddev, seq);
253 	if (unlikely(!radeon_fence_signaled(fence))) {
254 		/* we were interrupted for some reason and fence isn't
255 		 * isn't signaled yet, resume wait
256 		 */
257 		if (r) {
258 			timeout = r;
259 			goto retry;
260 		}
261 		/* don't protect read access to rdev->fence_drv.last_seq
262 		 * if we experiencing a lockup the value doesn't change
263 		 */
264 		if (seq == rdev->fence_drv.last_seq && radeon_gpu_is_lockup(rdev)) {
265 			/* good news we believe it's a lockup */
266 			WARN(1, "GPU lockup (waiting for 0x%08X last fence id 0x%08X)\n",
267 			     fence->seq, seq);
268 			/* FIXME: what should we do ? marking everyone
269 			 * as signaled for now
270 			 */
271 			rdev->gpu_lockup = true;
272 			r = radeon_gpu_reset(rdev);
273 			if (r)
274 				return r;
275 			radeon_fence_write(rdev, fence->seq);
276 			rdev->gpu_lockup = false;
277 		}
278 		timeout = RADEON_FENCE_JIFFIES_TIMEOUT;
279 		write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
280 		rdev->fence_drv.last_timeout = RADEON_FENCE_JIFFIES_TIMEOUT;
281 		rdev->fence_drv.last_jiffies = jiffies;
282 		write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
283 		goto retry;
284 	}
285 	return 0;
286 }
287 
288 int radeon_fence_wait_next(struct radeon_device *rdev)
289 {
290 	unsigned long irq_flags;
291 	struct radeon_fence *fence;
292 	int r;
293 
294 	if (rdev->gpu_lockup) {
295 		return 0;
296 	}
297 	write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
298 	if (list_empty(&rdev->fence_drv.emited)) {
299 		write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
300 		return 0;
301 	}
302 	fence = list_entry(rdev->fence_drv.emited.next,
303 			   struct radeon_fence, list);
304 	radeon_fence_ref(fence);
305 	write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
306 	r = radeon_fence_wait(fence, false);
307 	radeon_fence_unref(&fence);
308 	return r;
309 }
310 
311 int radeon_fence_wait_last(struct radeon_device *rdev)
312 {
313 	unsigned long irq_flags;
314 	struct radeon_fence *fence;
315 	int r;
316 
317 	if (rdev->gpu_lockup) {
318 		return 0;
319 	}
320 	write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
321 	if (list_empty(&rdev->fence_drv.emited)) {
322 		write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
323 		return 0;
324 	}
325 	fence = list_entry(rdev->fence_drv.emited.prev,
326 			   struct radeon_fence, list);
327 	radeon_fence_ref(fence);
328 	write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
329 	r = radeon_fence_wait(fence, false);
330 	radeon_fence_unref(&fence);
331 	return r;
332 }
333 
334 struct radeon_fence *radeon_fence_ref(struct radeon_fence *fence)
335 {
336 	kref_get(&fence->kref);
337 	return fence;
338 }
339 
340 void radeon_fence_unref(struct radeon_fence **fence)
341 {
342 	struct radeon_fence *tmp = *fence;
343 
344 	*fence = NULL;
345 	if (tmp) {
346 		kref_put(&tmp->kref, radeon_fence_destroy);
347 	}
348 }
349 
350 void radeon_fence_process(struct radeon_device *rdev)
351 {
352 	unsigned long irq_flags;
353 	bool wake;
354 
355 	write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
356 	wake = radeon_fence_poll_locked(rdev);
357 	write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
358 	if (wake) {
359 		wake_up_all(&rdev->fence_drv.queue);
360 	}
361 }
362 
363 int radeon_fence_driver_init(struct radeon_device *rdev)
364 {
365 	unsigned long irq_flags;
366 	int r;
367 
368 	write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
369 	r = radeon_scratch_get(rdev, &rdev->fence_drv.scratch_reg);
370 	if (r) {
371 		dev_err(rdev->dev, "fence failed to get scratch register\n");
372 		write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
373 		return r;
374 	}
375 	radeon_fence_write(rdev, 0);
376 	atomic_set(&rdev->fence_drv.seq, 0);
377 	INIT_LIST_HEAD(&rdev->fence_drv.created);
378 	INIT_LIST_HEAD(&rdev->fence_drv.emited);
379 	INIT_LIST_HEAD(&rdev->fence_drv.signaled);
380 	init_waitqueue_head(&rdev->fence_drv.queue);
381 	rdev->fence_drv.initialized = true;
382 	write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
383 	if (radeon_debugfs_fence_init(rdev)) {
384 		dev_err(rdev->dev, "fence debugfs file creation failed\n");
385 	}
386 	return 0;
387 }
388 
389 void radeon_fence_driver_fini(struct radeon_device *rdev)
390 {
391 	unsigned long irq_flags;
392 
393 	if (!rdev->fence_drv.initialized)
394 		return;
395 	wake_up_all(&rdev->fence_drv.queue);
396 	write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
397 	radeon_scratch_free(rdev, rdev->fence_drv.scratch_reg);
398 	write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
399 	rdev->fence_drv.initialized = false;
400 }
401 
402 
403 /*
404  * Fence debugfs
405  */
406 #if defined(CONFIG_DEBUG_FS)
407 static int radeon_debugfs_fence_info(struct seq_file *m, void *data)
408 {
409 	struct drm_info_node *node = (struct drm_info_node *)m->private;
410 	struct drm_device *dev = node->minor->dev;
411 	struct radeon_device *rdev = dev->dev_private;
412 	struct radeon_fence *fence;
413 
414 	seq_printf(m, "Last signaled fence 0x%08X\n",
415 		   radeon_fence_read(rdev));
416 	if (!list_empty(&rdev->fence_drv.emited)) {
417 		   fence = list_entry(rdev->fence_drv.emited.prev,
418 				      struct radeon_fence, list);
419 		   seq_printf(m, "Last emited fence %p with 0x%08X\n",
420 			      fence,  fence->seq);
421 	}
422 	return 0;
423 }
424 
425 static struct drm_info_list radeon_debugfs_fence_list[] = {
426 	{"radeon_fence_info", &radeon_debugfs_fence_info, 0, NULL},
427 };
428 #endif
429 
430 int radeon_debugfs_fence_init(struct radeon_device *rdev)
431 {
432 #if defined(CONFIG_DEBUG_FS)
433 	return radeon_debugfs_add_files(rdev, radeon_debugfs_fence_list, 1);
434 #else
435 	return 0;
436 #endif
437 }
438