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_drv.h"
19 #include "msm_gpu.h"
20 #include "msm_gem.h"
21 
22 /*
23  * Cmdstream submission:
24  */
25 
26 /* make sure these don't conflict w/ MSM_SUBMIT_BO_x */
27 #define BO_VALID    0x8000   /* is current addr in cmdstream correct/valid? */
28 #define BO_LOCKED   0x4000
29 #define BO_PINNED   0x2000
30 
31 static struct msm_gem_submit *submit_create(struct drm_device *dev,
32 		struct msm_gpu *gpu, int nr_bos, int nr_cmds)
33 {
34 	struct msm_gem_submit *submit;
35 	int sz = sizeof(*submit) + (nr_bos * sizeof(submit->bos[0])) +
36 			(nr_cmds * sizeof(*submit->cmd));
37 
38 	submit = kmalloc(sz, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
39 	if (!submit)
40 		return NULL;
41 
42 	submit->dev = dev;
43 	submit->gpu = gpu;
44 	submit->fence = NULL;
45 	submit->pid = get_pid(task_pid(current));
46 	submit->cmd = (void *)&submit->bos[nr_bos];
47 
48 	/* initially, until copy_from_user() and bo lookup succeeds: */
49 	submit->nr_bos = 0;
50 	submit->nr_cmds = 0;
51 
52 	INIT_LIST_HEAD(&submit->node);
53 	INIT_LIST_HEAD(&submit->bo_list);
54 	ww_acquire_init(&submit->ticket, &reservation_ww_class);
55 
56 	return submit;
57 }
58 
59 void msm_gem_submit_free(struct msm_gem_submit *submit)
60 {
61 	fence_put(submit->fence);
62 	list_del(&submit->node);
63 	put_pid(submit->pid);
64 	kfree(submit);
65 }
66 
67 static int submit_lookup_objects(struct msm_gem_submit *submit,
68 		struct drm_msm_gem_submit *args, struct drm_file *file)
69 {
70 	unsigned i;
71 	int ret = 0;
72 
73 	spin_lock(&file->table_lock);
74 
75 	for (i = 0; i < args->nr_bos; i++) {
76 		struct drm_msm_gem_submit_bo submit_bo;
77 		struct drm_gem_object *obj;
78 		struct msm_gem_object *msm_obj;
79 		void __user *userptr =
80 			u64_to_user_ptr(args->bos + (i * sizeof(submit_bo)));
81 
82 		/* make sure we don't have garbage flags, in case we hit
83 		 * error path before flags is initialized:
84 		 */
85 		submit->bos[i].flags = 0;
86 
87 		ret = copy_from_user(&submit_bo, userptr, sizeof(submit_bo));
88 		if (ret) {
89 			ret = -EFAULT;
90 			goto out_unlock;
91 		}
92 
93 		if (submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) {
94 			DRM_ERROR("invalid flags: %x\n", submit_bo.flags);
95 			ret = -EINVAL;
96 			goto out_unlock;
97 		}
98 
99 		submit->bos[i].flags = submit_bo.flags;
100 		/* in validate_objects() we figure out if this is true: */
101 		submit->bos[i].iova  = submit_bo.presumed;
102 
103 		/* normally use drm_gem_object_lookup(), but for bulk lookup
104 		 * all under single table_lock just hit object_idr directly:
105 		 */
106 		obj = idr_find(&file->object_idr, submit_bo.handle);
107 		if (!obj) {
108 			DRM_ERROR("invalid handle %u at index %u\n", submit_bo.handle, i);
109 			ret = -EINVAL;
110 			goto out_unlock;
111 		}
112 
113 		msm_obj = to_msm_bo(obj);
114 
115 		if (!list_empty(&msm_obj->submit_entry)) {
116 			DRM_ERROR("handle %u at index %u already on submit list\n",
117 					submit_bo.handle, i);
118 			ret = -EINVAL;
119 			goto out_unlock;
120 		}
121 
122 		drm_gem_object_reference(obj);
123 
124 		submit->bos[i].obj = msm_obj;
125 
126 		list_add_tail(&msm_obj->submit_entry, &submit->bo_list);
127 	}
128 
129 out_unlock:
130 	submit->nr_bos = i;
131 	spin_unlock(&file->table_lock);
132 
133 	return ret;
134 }
135 
136 static void submit_unlock_unpin_bo(struct msm_gem_submit *submit, int i)
137 {
138 	struct msm_gem_object *msm_obj = submit->bos[i].obj;
139 
140 	if (submit->bos[i].flags & BO_PINNED)
141 		msm_gem_put_iova(&msm_obj->base, submit->gpu->id);
142 
143 	if (submit->bos[i].flags & BO_LOCKED)
144 		ww_mutex_unlock(&msm_obj->resv->lock);
145 
146 	if (!(submit->bos[i].flags & BO_VALID))
147 		submit->bos[i].iova = 0;
148 
149 	submit->bos[i].flags &= ~(BO_LOCKED | BO_PINNED);
150 }
151 
152 /* This is where we make sure all the bo's are reserved and pin'd: */
153 static int submit_lock_objects(struct msm_gem_submit *submit)
154 {
155 	int contended, slow_locked = -1, i, ret = 0;
156 
157 retry:
158 	for (i = 0; i < submit->nr_bos; i++) {
159 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
160 
161 		if (slow_locked == i)
162 			slow_locked = -1;
163 
164 		contended = i;
165 
166 		if (!(submit->bos[i].flags & BO_LOCKED)) {
167 			ret = ww_mutex_lock_interruptible(&msm_obj->resv->lock,
168 					&submit->ticket);
169 			if (ret)
170 				goto fail;
171 			submit->bos[i].flags |= BO_LOCKED;
172 		}
173 	}
174 
175 	ww_acquire_done(&submit->ticket);
176 
177 	return 0;
178 
179 fail:
180 	for (; i >= 0; i--)
181 		submit_unlock_unpin_bo(submit, i);
182 
183 	if (slow_locked > 0)
184 		submit_unlock_unpin_bo(submit, slow_locked);
185 
186 	if (ret == -EDEADLK) {
187 		struct msm_gem_object *msm_obj = submit->bos[contended].obj;
188 		/* we lost out in a seqno race, lock and retry.. */
189 		ret = ww_mutex_lock_slow_interruptible(&msm_obj->resv->lock,
190 				&submit->ticket);
191 		if (!ret) {
192 			submit->bos[contended].flags |= BO_LOCKED;
193 			slow_locked = contended;
194 			goto retry;
195 		}
196 	}
197 
198 	return ret;
199 }
200 
201 static int submit_fence_sync(struct msm_gem_submit *submit)
202 {
203 	int i, ret = 0;
204 
205 	for (i = 0; i < submit->nr_bos; i++) {
206 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
207 		bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE;
208 
209 		ret = msm_gem_sync_object(&msm_obj->base, submit->gpu->fctx, write);
210 		if (ret)
211 			break;
212 	}
213 
214 	return ret;
215 }
216 
217 static int submit_pin_objects(struct msm_gem_submit *submit)
218 {
219 	int i, ret = 0;
220 
221 	submit->valid = true;
222 
223 	for (i = 0; i < submit->nr_bos; i++) {
224 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
225 		uint32_t iova;
226 
227 		/* if locking succeeded, pin bo: */
228 		ret = msm_gem_get_iova_locked(&msm_obj->base,
229 				submit->gpu->id, &iova);
230 
231 		if (ret)
232 			break;
233 
234 		submit->bos[i].flags |= BO_PINNED;
235 
236 		if (iova == submit->bos[i].iova) {
237 			submit->bos[i].flags |= BO_VALID;
238 		} else {
239 			submit->bos[i].iova = iova;
240 			/* iova changed, so address in cmdstream is not valid: */
241 			submit->bos[i].flags &= ~BO_VALID;
242 			submit->valid = false;
243 		}
244 	}
245 
246 	return ret;
247 }
248 
249 static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
250 		struct msm_gem_object **obj, uint32_t *iova, bool *valid)
251 {
252 	if (idx >= submit->nr_bos) {
253 		DRM_ERROR("invalid buffer index: %u (out of %u)\n",
254 				idx, submit->nr_bos);
255 		return -EINVAL;
256 	}
257 
258 	if (obj)
259 		*obj = submit->bos[idx].obj;
260 	if (iova)
261 		*iova = submit->bos[idx].iova;
262 	if (valid)
263 		*valid = !!(submit->bos[idx].flags & BO_VALID);
264 
265 	return 0;
266 }
267 
268 /* process the reloc's and patch up the cmdstream as needed: */
269 static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj,
270 		uint32_t offset, uint32_t nr_relocs, uint64_t relocs)
271 {
272 	uint32_t i, last_offset = 0;
273 	uint32_t *ptr;
274 	int ret;
275 
276 	if (offset % 4) {
277 		DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset);
278 		return -EINVAL;
279 	}
280 
281 	/* For now, just map the entire thing.  Eventually we probably
282 	 * to do it page-by-page, w/ kmap() if not vmap()d..
283 	 */
284 	ptr = msm_gem_get_vaddr_locked(&obj->base);
285 
286 	if (IS_ERR(ptr)) {
287 		ret = PTR_ERR(ptr);
288 		DBG("failed to map: %d", ret);
289 		return ret;
290 	}
291 
292 	for (i = 0; i < nr_relocs; i++) {
293 		struct drm_msm_gem_submit_reloc submit_reloc;
294 		void __user *userptr =
295 			u64_to_user_ptr(relocs + (i * sizeof(submit_reloc)));
296 		uint32_t iova, off;
297 		bool valid;
298 
299 		ret = copy_from_user(&submit_reloc, userptr, sizeof(submit_reloc));
300 		if (ret)
301 			return -EFAULT;
302 
303 		if (submit_reloc.submit_offset % 4) {
304 			DRM_ERROR("non-aligned reloc offset: %u\n",
305 					submit_reloc.submit_offset);
306 			return -EINVAL;
307 		}
308 
309 		/* offset in dwords: */
310 		off = submit_reloc.submit_offset / 4;
311 
312 		if ((off >= (obj->base.size / 4)) ||
313 				(off < last_offset)) {
314 			DRM_ERROR("invalid offset %u at reloc %u\n", off, i);
315 			return -EINVAL;
316 		}
317 
318 		ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid);
319 		if (ret)
320 			return ret;
321 
322 		if (valid)
323 			continue;
324 
325 		iova += submit_reloc.reloc_offset;
326 
327 		if (submit_reloc.shift < 0)
328 			iova >>= -submit_reloc.shift;
329 		else
330 			iova <<= submit_reloc.shift;
331 
332 		ptr[off] = iova | submit_reloc.or;
333 
334 		last_offset = off;
335 	}
336 
337 	msm_gem_put_vaddr_locked(&obj->base);
338 
339 	return 0;
340 }
341 
342 static void submit_cleanup(struct msm_gem_submit *submit)
343 {
344 	unsigned i;
345 
346 	for (i = 0; i < submit->nr_bos; i++) {
347 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
348 		submit_unlock_unpin_bo(submit, i);
349 		list_del_init(&msm_obj->submit_entry);
350 		drm_gem_object_unreference(&msm_obj->base);
351 	}
352 
353 	ww_acquire_fini(&submit->ticket);
354 }
355 
356 int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
357 		struct drm_file *file)
358 {
359 	struct msm_drm_private *priv = dev->dev_private;
360 	struct drm_msm_gem_submit *args = data;
361 	struct msm_file_private *ctx = file->driver_priv;
362 	struct msm_gem_submit *submit;
363 	struct msm_gpu *gpu = priv->gpu;
364 	unsigned i;
365 	int ret;
366 
367 	if (!gpu)
368 		return -ENXIO;
369 
370 	/* for now, we just have 3d pipe.. eventually this would need to
371 	 * be more clever to dispatch to appropriate gpu module:
372 	 */
373 	if (args->pipe != MSM_PIPE_3D0)
374 		return -EINVAL;
375 
376 	ret = mutex_lock_interruptible(&dev->struct_mutex);
377 	if (ret)
378 		return ret;
379 
380 	submit = submit_create(dev, gpu, args->nr_bos, args->nr_cmds);
381 	if (!submit) {
382 		ret = -ENOMEM;
383 		goto out_unlock;
384 	}
385 
386 	ret = submit_lookup_objects(submit, args, file);
387 	if (ret)
388 		goto out;
389 
390 	ret = submit_lock_objects(submit);
391 	if (ret)
392 		goto out;
393 
394 	ret = submit_fence_sync(submit);
395 	if (ret)
396 		goto out;
397 
398 	ret = submit_pin_objects(submit);
399 	if (ret)
400 		goto out;
401 
402 	for (i = 0; i < args->nr_cmds; i++) {
403 		struct drm_msm_gem_submit_cmd submit_cmd;
404 		void __user *userptr =
405 			u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
406 		struct msm_gem_object *msm_obj;
407 		uint32_t iova;
408 
409 		ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
410 		if (ret) {
411 			ret = -EFAULT;
412 			goto out;
413 		}
414 
415 		/* validate input from userspace: */
416 		switch (submit_cmd.type) {
417 		case MSM_SUBMIT_CMD_BUF:
418 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
419 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
420 			break;
421 		default:
422 			DRM_ERROR("invalid type: %08x\n", submit_cmd.type);
423 			ret = -EINVAL;
424 			goto out;
425 		}
426 
427 		ret = submit_bo(submit, submit_cmd.submit_idx,
428 				&msm_obj, &iova, NULL);
429 		if (ret)
430 			goto out;
431 
432 		if (submit_cmd.size % 4) {
433 			DRM_ERROR("non-aligned cmdstream buffer size: %u\n",
434 					submit_cmd.size);
435 			ret = -EINVAL;
436 			goto out;
437 		}
438 
439 		if ((submit_cmd.size + submit_cmd.submit_offset) >=
440 				msm_obj->base.size) {
441 			DRM_ERROR("invalid cmdstream size: %u\n", submit_cmd.size);
442 			ret = -EINVAL;
443 			goto out;
444 		}
445 
446 		submit->cmd[i].type = submit_cmd.type;
447 		submit->cmd[i].size = submit_cmd.size / 4;
448 		submit->cmd[i].iova = iova + submit_cmd.submit_offset;
449 		submit->cmd[i].idx  = submit_cmd.submit_idx;
450 
451 		if (submit->valid)
452 			continue;
453 
454 		ret = submit_reloc(submit, msm_obj, submit_cmd.submit_offset,
455 				submit_cmd.nr_relocs, submit_cmd.relocs);
456 		if (ret)
457 			goto out;
458 	}
459 
460 	submit->nr_cmds = i;
461 
462 	ret = msm_gpu_submit(gpu, submit, ctx);
463 
464 	args->fence = submit->fence->seqno;
465 
466 out:
467 	submit_cleanup(submit);
468 	if (ret)
469 		msm_gem_submit_free(submit);
470 out_unlock:
471 	mutex_unlock(&dev->struct_mutex);
472 	return ret;
473 }
474