1 /*
2  * Copyright 2008 Advanced Micro Devices, Inc.
3  * Copyright 2008 Red Hat Inc.
4  * Copyright 2009 Jerome Glisse.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the "Software"),
8  * to deal in the Software without restriction, including without limitation
9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10  * and/or sell copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  *
24  * Authors: Dave Airlie
25  *          Alex Deucher
26  *          Jerome Glisse
27  */
28 #include <linux/seq_file.h>
29 #include <linux/slab.h>
30 #include "drmP.h"
31 #include "radeon_drm.h"
32 #include "radeon_reg.h"
33 #include "radeon.h"
34 #include "atom.h"
35 
36 int radeon_debugfs_ib_init(struct radeon_device *rdev);
37 
38 void radeon_ib_bogus_cleanup(struct radeon_device *rdev)
39 {
40 	struct radeon_ib *ib, *n;
41 
42 	list_for_each_entry_safe(ib, n, &rdev->ib_pool.bogus_ib, list) {
43 		list_del(&ib->list);
44 		vfree(ib->ptr);
45 		kfree(ib);
46 	}
47 }
48 
49 void radeon_ib_bogus_add(struct radeon_device *rdev, struct radeon_ib *ib)
50 {
51 	struct radeon_ib *bib;
52 
53 	bib = kmalloc(sizeof(*bib), GFP_KERNEL);
54 	if (bib == NULL)
55 		return;
56 	bib->ptr = vmalloc(ib->length_dw * 4);
57 	if (bib->ptr == NULL) {
58 		kfree(bib);
59 		return;
60 	}
61 	memcpy(bib->ptr, ib->ptr, ib->length_dw * 4);
62 	bib->length_dw = ib->length_dw;
63 	mutex_lock(&rdev->ib_pool.mutex);
64 	list_add_tail(&bib->list, &rdev->ib_pool.bogus_ib);
65 	mutex_unlock(&rdev->ib_pool.mutex);
66 }
67 
68 /*
69  * IB.
70  */
71 int radeon_ib_get(struct radeon_device *rdev, struct radeon_ib **ib)
72 {
73 	struct radeon_fence *fence;
74 	struct radeon_ib *nib;
75 	int r = 0, i, c;
76 
77 	*ib = NULL;
78 	r = radeon_fence_create(rdev, &fence);
79 	if (r) {
80 		dev_err(rdev->dev, "failed to create fence for new IB\n");
81 		return r;
82 	}
83 	mutex_lock(&rdev->ib_pool.mutex);
84 	for (i = rdev->ib_pool.head_id, c = 0, nib = NULL; c < RADEON_IB_POOL_SIZE; c++, i++) {
85 		i &= (RADEON_IB_POOL_SIZE - 1);
86 		if (rdev->ib_pool.ibs[i].free) {
87 			nib = &rdev->ib_pool.ibs[i];
88 			break;
89 		}
90 	}
91 	if (nib == NULL) {
92 		/* This should never happen, it means we allocated all
93 		 * IB and haven't scheduled one yet, return EBUSY to
94 		 * userspace hoping that on ioctl recall we get better
95 		 * luck
96 		 */
97 		dev_err(rdev->dev, "no free indirect buffer !\n");
98 		mutex_unlock(&rdev->ib_pool.mutex);
99 		radeon_fence_unref(&fence);
100 		return -EBUSY;
101 	}
102 	rdev->ib_pool.head_id = (nib->idx + 1) & (RADEON_IB_POOL_SIZE - 1);
103 	nib->free = false;
104 	if (nib->fence) {
105 		mutex_unlock(&rdev->ib_pool.mutex);
106 		r = radeon_fence_wait(nib->fence, false);
107 		if (r) {
108 			dev_err(rdev->dev, "error waiting fence of IB(%u:0x%016lX:%u)\n",
109 				nib->idx, (unsigned long)nib->gpu_addr, nib->length_dw);
110 			mutex_lock(&rdev->ib_pool.mutex);
111 			nib->free = true;
112 			mutex_unlock(&rdev->ib_pool.mutex);
113 			radeon_fence_unref(&fence);
114 			return r;
115 		}
116 		mutex_lock(&rdev->ib_pool.mutex);
117 	}
118 	radeon_fence_unref(&nib->fence);
119 	nib->fence = fence;
120 	nib->length_dw = 0;
121 	mutex_unlock(&rdev->ib_pool.mutex);
122 	*ib = nib;
123 	return 0;
124 }
125 
126 void radeon_ib_free(struct radeon_device *rdev, struct radeon_ib **ib)
127 {
128 	struct radeon_ib *tmp = *ib;
129 
130 	*ib = NULL;
131 	if (tmp == NULL) {
132 		return;
133 	}
134 	if (!tmp->fence->emited)
135 		radeon_fence_unref(&tmp->fence);
136 	mutex_lock(&rdev->ib_pool.mutex);
137 	tmp->free = true;
138 	mutex_unlock(&rdev->ib_pool.mutex);
139 }
140 
141 int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib)
142 {
143 	int r = 0;
144 
145 	if (!ib->length_dw || !rdev->cp.ready) {
146 		/* TODO: Nothings in the ib we should report. */
147 		DRM_ERROR("radeon: couldn't schedule IB(%u).\n", ib->idx);
148 		return -EINVAL;
149 	}
150 
151 	/* 64 dwords should be enough for fence too */
152 	r = radeon_ring_lock(rdev, 64);
153 	if (r) {
154 		DRM_ERROR("radeon: scheduling IB failed (%d).\n", r);
155 		return r;
156 	}
157 	radeon_ring_ib_execute(rdev, ib);
158 	radeon_fence_emit(rdev, ib->fence);
159 	mutex_lock(&rdev->ib_pool.mutex);
160 	/* once scheduled IB is considered free and protected by the fence */
161 	ib->free = true;
162 	mutex_unlock(&rdev->ib_pool.mutex);
163 	radeon_ring_unlock_commit(rdev);
164 	return 0;
165 }
166 
167 int radeon_ib_pool_init(struct radeon_device *rdev)
168 {
169 	void *ptr;
170 	uint64_t gpu_addr;
171 	int i;
172 	int r = 0;
173 
174 	if (rdev->ib_pool.robj)
175 		return 0;
176 	INIT_LIST_HEAD(&rdev->ib_pool.bogus_ib);
177 	/* Allocate 1M object buffer */
178 	r = radeon_bo_create(rdev, RADEON_IB_POOL_SIZE*64*1024,
179 			     PAGE_SIZE, true, RADEON_GEM_DOMAIN_GTT,
180 			     &rdev->ib_pool.robj);
181 	if (r) {
182 		DRM_ERROR("radeon: failed to ib pool (%d).\n", r);
183 		return r;
184 	}
185 	r = radeon_bo_reserve(rdev->ib_pool.robj, false);
186 	if (unlikely(r != 0))
187 		return r;
188 	r = radeon_bo_pin(rdev->ib_pool.robj, RADEON_GEM_DOMAIN_GTT, &gpu_addr);
189 	if (r) {
190 		radeon_bo_unreserve(rdev->ib_pool.robj);
191 		DRM_ERROR("radeon: failed to pin ib pool (%d).\n", r);
192 		return r;
193 	}
194 	r = radeon_bo_kmap(rdev->ib_pool.robj, &ptr);
195 	radeon_bo_unreserve(rdev->ib_pool.robj);
196 	if (r) {
197 		DRM_ERROR("radeon: failed to map ib poll (%d).\n", r);
198 		return r;
199 	}
200 	for (i = 0; i < RADEON_IB_POOL_SIZE; i++) {
201 		unsigned offset;
202 
203 		offset = i * 64 * 1024;
204 		rdev->ib_pool.ibs[i].gpu_addr = gpu_addr + offset;
205 		rdev->ib_pool.ibs[i].ptr = ptr + offset;
206 		rdev->ib_pool.ibs[i].idx = i;
207 		rdev->ib_pool.ibs[i].length_dw = 0;
208 		rdev->ib_pool.ibs[i].free = true;
209 	}
210 	rdev->ib_pool.head_id = 0;
211 	rdev->ib_pool.ready = true;
212 	DRM_INFO("radeon: ib pool ready.\n");
213 	if (radeon_debugfs_ib_init(rdev)) {
214 		DRM_ERROR("Failed to register debugfs file for IB !\n");
215 	}
216 	return r;
217 }
218 
219 void radeon_ib_pool_fini(struct radeon_device *rdev)
220 {
221 	int r;
222 	struct radeon_bo *robj;
223 
224 	if (!rdev->ib_pool.ready) {
225 		return;
226 	}
227 	mutex_lock(&rdev->ib_pool.mutex);
228 	radeon_ib_bogus_cleanup(rdev);
229 	robj = rdev->ib_pool.robj;
230 	rdev->ib_pool.robj = NULL;
231 	mutex_unlock(&rdev->ib_pool.mutex);
232 
233 	if (robj) {
234 		r = radeon_bo_reserve(robj, false);
235 		if (likely(r == 0)) {
236 			radeon_bo_kunmap(robj);
237 			radeon_bo_unpin(robj);
238 			radeon_bo_unreserve(robj);
239 		}
240 		radeon_bo_unref(&robj);
241 	}
242 }
243 
244 
245 /*
246  * Ring.
247  */
248 void radeon_ring_free_size(struct radeon_device *rdev)
249 {
250 	if (rdev->wb.enabled)
251 		rdev->cp.rptr = rdev->wb.wb[RADEON_WB_CP_RPTR_OFFSET/4];
252 	else {
253 		if (rdev->family >= CHIP_R600)
254 			rdev->cp.rptr = RREG32(R600_CP_RB_RPTR);
255 		else
256 			rdev->cp.rptr = RREG32(RADEON_CP_RB_RPTR);
257 	}
258 	/* This works because ring_size is a power of 2 */
259 	rdev->cp.ring_free_dw = (rdev->cp.rptr + (rdev->cp.ring_size / 4));
260 	rdev->cp.ring_free_dw -= rdev->cp.wptr;
261 	rdev->cp.ring_free_dw &= rdev->cp.ptr_mask;
262 	if (!rdev->cp.ring_free_dw) {
263 		rdev->cp.ring_free_dw = rdev->cp.ring_size / 4;
264 	}
265 }
266 
267 int radeon_ring_alloc(struct radeon_device *rdev, unsigned ndw)
268 {
269 	int r;
270 
271 	/* Align requested size with padding so unlock_commit can
272 	 * pad safely */
273 	ndw = (ndw + rdev->cp.align_mask) & ~rdev->cp.align_mask;
274 	while (ndw > (rdev->cp.ring_free_dw - 1)) {
275 		radeon_ring_free_size(rdev);
276 		if (ndw < rdev->cp.ring_free_dw) {
277 			break;
278 		}
279 		r = radeon_fence_wait_next(rdev);
280 		if (r)
281 			return r;
282 	}
283 	rdev->cp.count_dw = ndw;
284 	rdev->cp.wptr_old = rdev->cp.wptr;
285 	return 0;
286 }
287 
288 int radeon_ring_lock(struct radeon_device *rdev, unsigned ndw)
289 {
290 	int r;
291 
292 	mutex_lock(&rdev->cp.mutex);
293 	r = radeon_ring_alloc(rdev, ndw);
294 	if (r) {
295 		mutex_unlock(&rdev->cp.mutex);
296 		return r;
297 	}
298 	return 0;
299 }
300 
301 void radeon_ring_commit(struct radeon_device *rdev)
302 {
303 	unsigned count_dw_pad;
304 	unsigned i;
305 
306 	/* We pad to match fetch size */
307 	count_dw_pad = (rdev->cp.align_mask + 1) -
308 		       (rdev->cp.wptr & rdev->cp.align_mask);
309 	for (i = 0; i < count_dw_pad; i++) {
310 		radeon_ring_write(rdev, 2 << 30);
311 	}
312 	DRM_MEMORYBARRIER();
313 	radeon_cp_commit(rdev);
314 }
315 
316 void radeon_ring_unlock_commit(struct radeon_device *rdev)
317 {
318 	radeon_ring_commit(rdev);
319 	mutex_unlock(&rdev->cp.mutex);
320 }
321 
322 void radeon_ring_unlock_undo(struct radeon_device *rdev)
323 {
324 	rdev->cp.wptr = rdev->cp.wptr_old;
325 	mutex_unlock(&rdev->cp.mutex);
326 }
327 
328 int radeon_ring_init(struct radeon_device *rdev, unsigned ring_size)
329 {
330 	int r;
331 
332 	rdev->cp.ring_size = ring_size;
333 	/* Allocate ring buffer */
334 	if (rdev->cp.ring_obj == NULL) {
335 		r = radeon_bo_create(rdev, rdev->cp.ring_size, PAGE_SIZE, true,
336 					RADEON_GEM_DOMAIN_GTT,
337 					&rdev->cp.ring_obj);
338 		if (r) {
339 			dev_err(rdev->dev, "(%d) ring create failed\n", r);
340 			return r;
341 		}
342 		r = radeon_bo_reserve(rdev->cp.ring_obj, false);
343 		if (unlikely(r != 0))
344 			return r;
345 		r = radeon_bo_pin(rdev->cp.ring_obj, RADEON_GEM_DOMAIN_GTT,
346 					&rdev->cp.gpu_addr);
347 		if (r) {
348 			radeon_bo_unreserve(rdev->cp.ring_obj);
349 			dev_err(rdev->dev, "(%d) ring pin failed\n", r);
350 			return r;
351 		}
352 		r = radeon_bo_kmap(rdev->cp.ring_obj,
353 				       (void **)&rdev->cp.ring);
354 		radeon_bo_unreserve(rdev->cp.ring_obj);
355 		if (r) {
356 			dev_err(rdev->dev, "(%d) ring map failed\n", r);
357 			return r;
358 		}
359 	}
360 	rdev->cp.ptr_mask = (rdev->cp.ring_size / 4) - 1;
361 	rdev->cp.ring_free_dw = rdev->cp.ring_size / 4;
362 	return 0;
363 }
364 
365 void radeon_ring_fini(struct radeon_device *rdev)
366 {
367 	int r;
368 	struct radeon_bo *ring_obj;
369 
370 	mutex_lock(&rdev->cp.mutex);
371 	ring_obj = rdev->cp.ring_obj;
372 	rdev->cp.ring = NULL;
373 	rdev->cp.ring_obj = NULL;
374 	mutex_unlock(&rdev->cp.mutex);
375 
376 	if (ring_obj) {
377 		r = radeon_bo_reserve(ring_obj, false);
378 		if (likely(r == 0)) {
379 			radeon_bo_kunmap(ring_obj);
380 			radeon_bo_unpin(ring_obj);
381 			radeon_bo_unreserve(ring_obj);
382 		}
383 		radeon_bo_unref(&ring_obj);
384 	}
385 }
386 
387 
388 /*
389  * Debugfs info
390  */
391 #if defined(CONFIG_DEBUG_FS)
392 static int radeon_debugfs_ib_info(struct seq_file *m, void *data)
393 {
394 	struct drm_info_node *node = (struct drm_info_node *) m->private;
395 	struct radeon_ib *ib = node->info_ent->data;
396 	unsigned i;
397 
398 	if (ib == NULL) {
399 		return 0;
400 	}
401 	seq_printf(m, "IB %04u\n", ib->idx);
402 	seq_printf(m, "IB fence %p\n", ib->fence);
403 	seq_printf(m, "IB size %05u dwords\n", ib->length_dw);
404 	for (i = 0; i < ib->length_dw; i++) {
405 		seq_printf(m, "[%05u]=0x%08X\n", i, ib->ptr[i]);
406 	}
407 	return 0;
408 }
409 
410 static int radeon_debugfs_ib_bogus_info(struct seq_file *m, void *data)
411 {
412 	struct drm_info_node *node = (struct drm_info_node *) m->private;
413 	struct radeon_device *rdev = node->info_ent->data;
414 	struct radeon_ib *ib;
415 	unsigned i;
416 
417 	mutex_lock(&rdev->ib_pool.mutex);
418 	if (list_empty(&rdev->ib_pool.bogus_ib)) {
419 		mutex_unlock(&rdev->ib_pool.mutex);
420 		seq_printf(m, "no bogus IB recorded\n");
421 		return 0;
422 	}
423 	ib = list_first_entry(&rdev->ib_pool.bogus_ib, struct radeon_ib, list);
424 	list_del_init(&ib->list);
425 	mutex_unlock(&rdev->ib_pool.mutex);
426 	seq_printf(m, "IB size %05u dwords\n", ib->length_dw);
427 	for (i = 0; i < ib->length_dw; i++) {
428 		seq_printf(m, "[%05u]=0x%08X\n", i, ib->ptr[i]);
429 	}
430 	vfree(ib->ptr);
431 	kfree(ib);
432 	return 0;
433 }
434 
435 static struct drm_info_list radeon_debugfs_ib_list[RADEON_IB_POOL_SIZE];
436 static char radeon_debugfs_ib_names[RADEON_IB_POOL_SIZE][32];
437 
438 static struct drm_info_list radeon_debugfs_ib_bogus_info_list[] = {
439 	{"radeon_ib_bogus", radeon_debugfs_ib_bogus_info, 0, NULL},
440 };
441 #endif
442 
443 int radeon_debugfs_ib_init(struct radeon_device *rdev)
444 {
445 #if defined(CONFIG_DEBUG_FS)
446 	unsigned i;
447 	int r;
448 
449 	radeon_debugfs_ib_bogus_info_list[0].data = rdev;
450 	r = radeon_debugfs_add_files(rdev, radeon_debugfs_ib_bogus_info_list, 1);
451 	if (r)
452 		return r;
453 	for (i = 0; i < RADEON_IB_POOL_SIZE; i++) {
454 		sprintf(radeon_debugfs_ib_names[i], "radeon_ib_%04u", i);
455 		radeon_debugfs_ib_list[i].name = radeon_debugfs_ib_names[i];
456 		radeon_debugfs_ib_list[i].show = &radeon_debugfs_ib_info;
457 		radeon_debugfs_ib_list[i].driver_features = 0;
458 		radeon_debugfs_ib_list[i].data = &rdev->ib_pool.ibs[i];
459 	}
460 	return radeon_debugfs_add_files(rdev, radeon_debugfs_ib_list,
461 					RADEON_IB_POOL_SIZE);
462 #else
463 	return 0;
464 #endif
465 }
466