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 <linux/sync_file.h>
19
20 #include "msm_drv.h"
21 #include "msm_gpu.h"
22 #include "msm_gem.h"
23
24 /*
25 * Cmdstream submission:
26 */
27
28 /* make sure these don't conflict w/ MSM_SUBMIT_BO_x */
29 #define BO_VALID 0x8000 /* is current addr in cmdstream correct/valid? */
30 #define BO_LOCKED 0x4000
31 #define BO_PINNED 0x2000
32
submit_create(struct drm_device * dev,struct msm_gpu * gpu,uint32_t nr_bos,uint32_t nr_cmds)33 static struct msm_gem_submit *submit_create(struct drm_device *dev,
34 struct msm_gpu *gpu, uint32_t nr_bos, uint32_t nr_cmds)
35 {
36 struct msm_gem_submit *submit;
37 uint64_t sz = sizeof(*submit) + ((u64)nr_bos * sizeof(submit->bos[0])) +
38 ((u64)nr_cmds * sizeof(submit->cmd[0]));
39
40 if (sz > SIZE_MAX)
41 return NULL;
42
43 submit = kmalloc(sz, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
44 if (!submit)
45 return NULL;
46
47 submit->dev = dev;
48 submit->gpu = gpu;
49 submit->fence = NULL;
50 submit->pid = get_pid(task_pid(current));
51 submit->cmd = (void *)&submit->bos[nr_bos];
52
53 /* initially, until copy_from_user() and bo lookup succeeds: */
54 submit->nr_bos = 0;
55 submit->nr_cmds = 0;
56
57 INIT_LIST_HEAD(&submit->node);
58 INIT_LIST_HEAD(&submit->bo_list);
59 ww_acquire_init(&submit->ticket, &reservation_ww_class);
60
61 return submit;
62 }
63
msm_gem_submit_free(struct msm_gem_submit * submit)64 void msm_gem_submit_free(struct msm_gem_submit *submit)
65 {
66 fence_put(submit->fence);
67 list_del(&submit->node);
68 put_pid(submit->pid);
69 kfree(submit);
70 }
71
72 static inline unsigned long __must_check
copy_from_user_inatomic(void * to,const void __user * from,unsigned long n)73 copy_from_user_inatomic(void *to, const void __user *from, unsigned long n)
74 {
75 if (access_ok(VERIFY_READ, from, n))
76 return __copy_from_user_inatomic(to, from, n);
77 return -EFAULT;
78 }
79
submit_lookup_objects(struct msm_gem_submit * submit,struct drm_msm_gem_submit * args,struct drm_file * file)80 static int submit_lookup_objects(struct msm_gem_submit *submit,
81 struct drm_msm_gem_submit *args, struct drm_file *file)
82 {
83 unsigned i;
84 int ret = 0;
85
86 spin_lock(&file->table_lock);
87 pagefault_disable();
88
89 for (i = 0; i < args->nr_bos; i++) {
90 struct drm_msm_gem_submit_bo submit_bo;
91 struct drm_gem_object *obj;
92 struct msm_gem_object *msm_obj;
93 void __user *userptr =
94 u64_to_user_ptr(args->bos + (i * sizeof(submit_bo)));
95
96 /* make sure we don't have garbage flags, in case we hit
97 * error path before flags is initialized:
98 */
99 submit->bos[i].flags = 0;
100
101 ret = copy_from_user_inatomic(&submit_bo, userptr, sizeof(submit_bo));
102 if (unlikely(ret)) {
103 pagefault_enable();
104 spin_unlock(&file->table_lock);
105 ret = copy_from_user(&submit_bo, userptr, sizeof(submit_bo));
106 if (ret)
107 goto out;
108 spin_lock(&file->table_lock);
109 pagefault_disable();
110 }
111
112 if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) ||
113 !(submit_bo.flags & MSM_SUBMIT_BO_FLAGS)) {
114 DRM_ERROR("invalid flags: %x\n", submit_bo.flags);
115 ret = -EINVAL;
116 goto out_unlock;
117 }
118
119 submit->bos[i].flags = submit_bo.flags;
120 /* in validate_objects() we figure out if this is true: */
121 submit->bos[i].iova = submit_bo.presumed;
122
123 /* normally use drm_gem_object_lookup(), but for bulk lookup
124 * all under single table_lock just hit object_idr directly:
125 */
126 obj = idr_find(&file->object_idr, submit_bo.handle);
127 if (!obj) {
128 DRM_ERROR("invalid handle %u at index %u\n", submit_bo.handle, i);
129 ret = -EINVAL;
130 goto out_unlock;
131 }
132
133 msm_obj = to_msm_bo(obj);
134
135 if (!list_empty(&msm_obj->submit_entry)) {
136 DRM_ERROR("handle %u at index %u already on submit list\n",
137 submit_bo.handle, i);
138 ret = -EINVAL;
139 goto out_unlock;
140 }
141
142 drm_gem_object_reference(obj);
143
144 submit->bos[i].obj = msm_obj;
145
146 list_add_tail(&msm_obj->submit_entry, &submit->bo_list);
147 }
148
149 out_unlock:
150 pagefault_enable();
151 spin_unlock(&file->table_lock);
152
153 out:
154 submit->nr_bos = i;
155
156 return ret;
157 }
158
submit_unlock_unpin_bo(struct msm_gem_submit * submit,int i)159 static void submit_unlock_unpin_bo(struct msm_gem_submit *submit, int i)
160 {
161 struct msm_gem_object *msm_obj = submit->bos[i].obj;
162
163 if (submit->bos[i].flags & BO_PINNED)
164 msm_gem_put_iova(&msm_obj->base, submit->gpu->id);
165
166 if (submit->bos[i].flags & BO_LOCKED)
167 ww_mutex_unlock(&msm_obj->resv->lock);
168
169 if (!(submit->bos[i].flags & BO_VALID))
170 submit->bos[i].iova = 0;
171
172 submit->bos[i].flags &= ~(BO_LOCKED | BO_PINNED);
173 }
174
175 /* This is where we make sure all the bo's are reserved and pin'd: */
submit_lock_objects(struct msm_gem_submit * submit)176 static int submit_lock_objects(struct msm_gem_submit *submit)
177 {
178 int contended, slow_locked = -1, i, ret = 0;
179
180 retry:
181 for (i = 0; i < submit->nr_bos; i++) {
182 struct msm_gem_object *msm_obj = submit->bos[i].obj;
183
184 if (slow_locked == i)
185 slow_locked = -1;
186
187 contended = i;
188
189 if (!(submit->bos[i].flags & BO_LOCKED)) {
190 ret = ww_mutex_lock_interruptible(&msm_obj->resv->lock,
191 &submit->ticket);
192 if (ret)
193 goto fail;
194 submit->bos[i].flags |= BO_LOCKED;
195 }
196 }
197
198 ww_acquire_done(&submit->ticket);
199
200 return 0;
201
202 fail:
203 for (; i >= 0; i--)
204 submit_unlock_unpin_bo(submit, i);
205
206 if (slow_locked > 0)
207 submit_unlock_unpin_bo(submit, slow_locked);
208
209 if (ret == -EDEADLK) {
210 struct msm_gem_object *msm_obj = submit->bos[contended].obj;
211 /* we lost out in a seqno race, lock and retry.. */
212 ret = ww_mutex_lock_slow_interruptible(&msm_obj->resv->lock,
213 &submit->ticket);
214 if (!ret) {
215 submit->bos[contended].flags |= BO_LOCKED;
216 slow_locked = contended;
217 goto retry;
218 }
219 }
220
221 return ret;
222 }
223
submit_fence_sync(struct msm_gem_submit * submit)224 static int submit_fence_sync(struct msm_gem_submit *submit)
225 {
226 int i, ret = 0;
227
228 for (i = 0; i < submit->nr_bos; i++) {
229 struct msm_gem_object *msm_obj = submit->bos[i].obj;
230 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE;
231
232 ret = msm_gem_sync_object(&msm_obj->base, submit->gpu->fctx, write);
233 if (ret)
234 break;
235 }
236
237 return ret;
238 }
239
submit_pin_objects(struct msm_gem_submit * submit)240 static int submit_pin_objects(struct msm_gem_submit *submit)
241 {
242 int i, ret = 0;
243
244 submit->valid = true;
245
246 for (i = 0; i < submit->nr_bos; i++) {
247 struct msm_gem_object *msm_obj = submit->bos[i].obj;
248 uint32_t iova;
249
250 /* if locking succeeded, pin bo: */
251 ret = msm_gem_get_iova_locked(&msm_obj->base,
252 submit->gpu->id, &iova);
253
254 if (ret)
255 break;
256
257 submit->bos[i].flags |= BO_PINNED;
258
259 if (iova == submit->bos[i].iova) {
260 submit->bos[i].flags |= BO_VALID;
261 } else {
262 submit->bos[i].iova = iova;
263 /* iova changed, so address in cmdstream is not valid: */
264 submit->bos[i].flags &= ~BO_VALID;
265 submit->valid = false;
266 }
267 }
268
269 return ret;
270 }
271
submit_bo(struct msm_gem_submit * submit,uint32_t idx,struct msm_gem_object ** obj,uint32_t * iova,bool * valid)272 static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
273 struct msm_gem_object **obj, uint32_t *iova, bool *valid)
274 {
275 if (idx >= submit->nr_bos) {
276 DRM_ERROR("invalid buffer index: %u (out of %u)\n",
277 idx, submit->nr_bos);
278 return -EINVAL;
279 }
280
281 if (obj)
282 *obj = submit->bos[idx].obj;
283 if (iova)
284 *iova = submit->bos[idx].iova;
285 if (valid)
286 *valid = !!(submit->bos[idx].flags & BO_VALID);
287
288 return 0;
289 }
290
291 /* process the reloc's and patch up the cmdstream as needed: */
submit_reloc(struct msm_gem_submit * submit,struct msm_gem_object * obj,uint32_t offset,uint32_t nr_relocs,uint64_t relocs)292 static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj,
293 uint32_t offset, uint32_t nr_relocs, uint64_t relocs)
294 {
295 uint32_t i, last_offset = 0;
296 uint32_t *ptr;
297 int ret = 0;
298
299 if (offset % 4) {
300 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset);
301 return -EINVAL;
302 }
303
304 /* For now, just map the entire thing. Eventually we probably
305 * to do it page-by-page, w/ kmap() if not vmap()d..
306 */
307 ptr = msm_gem_get_vaddr_locked(&obj->base);
308
309 if (IS_ERR(ptr)) {
310 ret = PTR_ERR(ptr);
311 DBG("failed to map: %d", ret);
312 return ret;
313 }
314
315 for (i = 0; i < nr_relocs; i++) {
316 struct drm_msm_gem_submit_reloc submit_reloc;
317 void __user *userptr =
318 u64_to_user_ptr(relocs + (i * sizeof(submit_reloc)));
319 uint32_t iova, off;
320 bool valid;
321
322 ret = copy_from_user(&submit_reloc, userptr, sizeof(submit_reloc));
323 if (ret)
324 goto out;
325
326 if (submit_reloc.submit_offset % 4) {
327 DRM_ERROR("non-aligned reloc offset: %u\n",
328 submit_reloc.submit_offset);
329 ret = -EINVAL;
330 goto out;
331 }
332
333 /* offset in dwords: */
334 off = submit_reloc.submit_offset / 4;
335
336 if ((off >= (obj->base.size / 4)) ||
337 (off < last_offset)) {
338 DRM_ERROR("invalid offset %u at reloc %u\n", off, i);
339 ret = -EINVAL;
340 goto out;
341 }
342
343 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid);
344 if (ret)
345 goto out;
346
347 if (valid)
348 continue;
349
350 iova += submit_reloc.reloc_offset;
351
352 if (submit_reloc.shift < 0)
353 iova >>= -submit_reloc.shift;
354 else
355 iova <<= submit_reloc.shift;
356
357 ptr[off] = iova | submit_reloc.or;
358
359 last_offset = off;
360 }
361
362 out:
363 msm_gem_put_vaddr_locked(&obj->base);
364
365 return ret;
366 }
367
submit_cleanup(struct msm_gem_submit * submit)368 static void submit_cleanup(struct msm_gem_submit *submit)
369 {
370 unsigned i;
371
372 for (i = 0; i < submit->nr_bos; i++) {
373 struct msm_gem_object *msm_obj = submit->bos[i].obj;
374 submit_unlock_unpin_bo(submit, i);
375 list_del_init(&msm_obj->submit_entry);
376 drm_gem_object_unreference(&msm_obj->base);
377 }
378
379 ww_acquire_fini(&submit->ticket);
380 }
381
msm_ioctl_gem_submit(struct drm_device * dev,void * data,struct drm_file * file)382 int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
383 struct drm_file *file)
384 {
385 struct msm_drm_private *priv = dev->dev_private;
386 struct drm_msm_gem_submit *args = data;
387 struct msm_file_private *ctx = file->driver_priv;
388 struct msm_gem_submit *submit;
389 struct msm_gpu *gpu = priv->gpu;
390 struct fence *in_fence = NULL;
391 struct sync_file *sync_file = NULL;
392 int out_fence_fd = -1;
393 unsigned i;
394 int ret;
395
396 if (!gpu)
397 return -ENXIO;
398
399 /* for now, we just have 3d pipe.. eventually this would need to
400 * be more clever to dispatch to appropriate gpu module:
401 */
402 if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0)
403 return -EINVAL;
404
405 if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS)
406 return -EINVAL;
407
408 ret = mutex_lock_interruptible(&dev->struct_mutex);
409 if (ret)
410 return ret;
411
412 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
413 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
414 if (out_fence_fd < 0) {
415 ret = out_fence_fd;
416 goto out_unlock;
417 }
418 }
419 priv->struct_mutex_task = current;
420
421 submit = submit_create(dev, gpu, args->nr_bos, args->nr_cmds);
422 if (!submit) {
423 ret = -ENOMEM;
424 goto out_unlock;
425 }
426
427 ret = submit_lookup_objects(submit, args, file);
428 if (ret)
429 goto out;
430
431 ret = submit_lock_objects(submit);
432 if (ret)
433 goto out;
434
435 if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
436 in_fence = sync_file_get_fence(args->fence_fd);
437
438 if (!in_fence) {
439 ret = -EINVAL;
440 goto out;
441 }
442
443 /* TODO if we get an array-fence due to userspace merging multiple
444 * fences, we need a way to determine if all the backing fences
445 * are from our own context..
446 */
447
448 if (in_fence->context != gpu->fctx->context) {
449 ret = fence_wait(in_fence, true);
450 if (ret)
451 goto out;
452 }
453
454 }
455
456 if (!(args->fence & MSM_SUBMIT_NO_IMPLICIT)) {
457 ret = submit_fence_sync(submit);
458 if (ret)
459 goto out;
460 }
461
462 ret = submit_pin_objects(submit);
463 if (ret)
464 goto out;
465
466 for (i = 0; i < args->nr_cmds; i++) {
467 struct drm_msm_gem_submit_cmd submit_cmd;
468 void __user *userptr =
469 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
470 struct msm_gem_object *msm_obj;
471 uint32_t iova;
472
473 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
474 if (ret) {
475 ret = -EFAULT;
476 goto out;
477 }
478
479 /* validate input from userspace: */
480 switch (submit_cmd.type) {
481 case MSM_SUBMIT_CMD_BUF:
482 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
483 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
484 break;
485 default:
486 DRM_ERROR("invalid type: %08x\n", submit_cmd.type);
487 ret = -EINVAL;
488 goto out;
489 }
490
491 ret = submit_bo(submit, submit_cmd.submit_idx,
492 &msm_obj, &iova, NULL);
493 if (ret)
494 goto out;
495
496 if (submit_cmd.size % 4) {
497 DRM_ERROR("non-aligned cmdstream buffer size: %u\n",
498 submit_cmd.size);
499 ret = -EINVAL;
500 goto out;
501 }
502
503 if ((submit_cmd.size + submit_cmd.submit_offset) >=
504 msm_obj->base.size) {
505 DRM_ERROR("invalid cmdstream size: %u\n", submit_cmd.size);
506 ret = -EINVAL;
507 goto out;
508 }
509
510 submit->cmd[i].type = submit_cmd.type;
511 submit->cmd[i].size = submit_cmd.size / 4;
512 submit->cmd[i].iova = iova + submit_cmd.submit_offset;
513 submit->cmd[i].idx = submit_cmd.submit_idx;
514
515 if (submit->valid)
516 continue;
517
518 ret = submit_reloc(submit, msm_obj, submit_cmd.submit_offset,
519 submit_cmd.nr_relocs, submit_cmd.relocs);
520 if (ret)
521 goto out;
522 }
523
524 submit->nr_cmds = i;
525
526 submit->fence = msm_fence_alloc(gpu->fctx);
527 if (IS_ERR(submit->fence)) {
528 ret = PTR_ERR(submit->fence);
529 submit->fence = NULL;
530 goto out;
531 }
532
533 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
534 sync_file = sync_file_create(submit->fence);
535 if (!sync_file) {
536 ret = -ENOMEM;
537 goto out;
538 }
539 }
540
541 msm_gpu_submit(gpu, submit, ctx);
542
543 args->fence = submit->fence->seqno;
544
545 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
546 fd_install(out_fence_fd, sync_file->file);
547 args->fence_fd = out_fence_fd;
548 }
549
550 out:
551 if (in_fence)
552 fence_put(in_fence);
553 submit_cleanup(submit);
554 if (ret)
555 msm_gem_submit_free(submit);
556 out_unlock:
557 if (ret && (out_fence_fd >= 0))
558 put_unused_fd(out_fence_fd);
559 priv->struct_mutex_task = NULL;
560 mutex_unlock(&dev->struct_mutex);
561 return ret;
562 }
563