diff options
Diffstat (limited to 'drivers/gpu/nvgpu/common/mm/vidmem.c')
-rw-r--r-- | drivers/gpu/nvgpu/common/mm/vidmem.c | 554 |
1 files changed, 554 insertions, 0 deletions
diff --git a/drivers/gpu/nvgpu/common/mm/vidmem.c b/drivers/gpu/nvgpu/common/mm/vidmem.c new file mode 100644 index 00000000..3526fce5 --- /dev/null +++ b/drivers/gpu/nvgpu/common/mm/vidmem.c | |||
@@ -0,0 +1,554 @@ | |||
1 | /* | ||
2 | * Copyright (c) 2017, NVIDIA CORPORATION. All rights reserved. | ||
3 | * | ||
4 | * Permission is hereby granted, free of charge, to any person obtaining a | ||
5 | * copy of this software and associated documentation files (the "Software"), | ||
6 | * to deal in the Software without restriction, including without limitation | ||
7 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, | ||
8 | * and/or sell copies of the Software, and to permit persons to whom the | ||
9 | * Software is furnished to do so, subject to the following conditions: | ||
10 | * | ||
11 | * The above copyright notice and this permission notice shall be included in | ||
12 | * all copies or substantial portions of the Software. | ||
13 | * | ||
14 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | ||
15 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | ||
16 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | ||
17 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER | ||
18 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING | ||
19 | * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER | ||
20 | * DEALINGS IN THE SOFTWARE. | ||
21 | */ | ||
22 | |||
23 | #include <linux/scatterlist.h> | ||
24 | |||
25 | #include <nvgpu/timers.h> | ||
26 | #include <nvgpu/dma.h> | ||
27 | #include <nvgpu/vidmem.h> | ||
28 | #include <nvgpu/page_allocator.h> | ||
29 | #include <nvgpu/enabled.h> | ||
30 | |||
31 | #include "gk20a/gk20a.h" | ||
32 | #include "gk20a/mm_gk20a.h" | ||
33 | |||
34 | /* | ||
35 | * This is expected to be called from the shutdown path (or the error path in | ||
36 | * the vidmem init code). As such we do not expect new vidmem frees to be | ||
37 | * enqueued. | ||
38 | */ | ||
39 | void nvgpu_vidmem_destroy(struct gk20a *g) | ||
40 | { | ||
41 | struct nvgpu_timeout timeout; | ||
42 | |||
43 | nvgpu_timeout_init(g, &timeout, 100, NVGPU_TIMER_RETRY_TIMER); | ||
44 | |||
45 | /* | ||
46 | * Ensure that the thread runs one last time to flush anything in the | ||
47 | * queue. | ||
48 | */ | ||
49 | nvgpu_cond_signal_interruptible(&g->mm.vidmem.clearing_thread_cond); | ||
50 | |||
51 | /* | ||
52 | * Wait for at most 1 second before just continuing on. It doesn't make | ||
53 | * sense to hang the system over some potential memory leaks. | ||
54 | */ | ||
55 | do { | ||
56 | bool empty; | ||
57 | |||
58 | nvgpu_mutex_acquire(&g->mm.vidmem.clear_list_mutex); | ||
59 | empty = nvgpu_list_empty(&g->mm.vidmem.clear_list_head); | ||
60 | nvgpu_mutex_release(&g->mm.vidmem.clear_list_mutex); | ||
61 | |||
62 | if (empty) | ||
63 | break; | ||
64 | |||
65 | nvgpu_msleep(10); | ||
66 | } while (!nvgpu_timeout_expired(&timeout)); | ||
67 | |||
68 | /* | ||
69 | * Kill the vidmem clearing thread now. This will wake the thread up | ||
70 | * automatically and cause the wait_interruptible condition trigger. | ||
71 | */ | ||
72 | nvgpu_thread_stop(&g->mm.vidmem.clearing_thread); | ||
73 | |||
74 | if (nvgpu_alloc_initialized(&g->mm.vidmem.allocator)) | ||
75 | nvgpu_alloc_destroy(&g->mm.vidmem.allocator); | ||
76 | } | ||
77 | |||
78 | static int __nvgpu_vidmem_do_clear_all(struct gk20a *g) | ||
79 | { | ||
80 | struct mm_gk20a *mm = &g->mm; | ||
81 | struct gk20a_fence *gk20a_fence_out = NULL; | ||
82 | u64 region2_base = 0; | ||
83 | int err = 0; | ||
84 | |||
85 | if (mm->vidmem.ce_ctx_id == (u32)~0) | ||
86 | return -EINVAL; | ||
87 | |||
88 | vidmem_dbg(g, "Clearing all VIDMEM:"); | ||
89 | |||
90 | err = gk20a_ce_execute_ops(g, | ||
91 | mm->vidmem.ce_ctx_id, | ||
92 | 0, | ||
93 | mm->vidmem.base, | ||
94 | mm->vidmem.bootstrap_base - mm->vidmem.base, | ||
95 | 0x00000000, | ||
96 | NVGPU_CE_DST_LOCATION_LOCAL_FB, | ||
97 | NVGPU_CE_MEMSET, | ||
98 | NULL, | ||
99 | 0, | ||
100 | NULL); | ||
101 | if (err) { | ||
102 | nvgpu_err(g, | ||
103 | "Failed to clear vidmem region 1 : %d", err); | ||
104 | return err; | ||
105 | } | ||
106 | |||
107 | region2_base = mm->vidmem.bootstrap_base + mm->vidmem.bootstrap_size; | ||
108 | |||
109 | err = gk20a_ce_execute_ops(g, | ||
110 | mm->vidmem.ce_ctx_id, | ||
111 | 0, | ||
112 | region2_base, | ||
113 | mm->vidmem.size - region2_base, | ||
114 | 0x00000000, | ||
115 | NVGPU_CE_DST_LOCATION_LOCAL_FB, | ||
116 | NVGPU_CE_MEMSET, | ||
117 | NULL, | ||
118 | 0, | ||
119 | &gk20a_fence_out); | ||
120 | if (err) { | ||
121 | nvgpu_err(g, | ||
122 | "Failed to clear vidmem region 2 : %d", err); | ||
123 | return err; | ||
124 | } | ||
125 | |||
126 | if (gk20a_fence_out) { | ||
127 | struct nvgpu_timeout timeout; | ||
128 | |||
129 | nvgpu_timeout_init(g, &timeout, | ||
130 | gk20a_get_gr_idle_timeout(g), | ||
131 | NVGPU_TIMER_CPU_TIMER); | ||
132 | |||
133 | do { | ||
134 | err = gk20a_fence_wait(g, gk20a_fence_out, | ||
135 | gk20a_get_gr_idle_timeout(g)); | ||
136 | } while (err == -ERESTARTSYS && | ||
137 | !nvgpu_timeout_expired(&timeout)); | ||
138 | |||
139 | gk20a_fence_put(gk20a_fence_out); | ||
140 | if (err) { | ||
141 | nvgpu_err(g, | ||
142 | "fence wait failed for CE execute ops"); | ||
143 | return err; | ||
144 | } | ||
145 | } | ||
146 | |||
147 | mm->vidmem.cleared = true; | ||
148 | |||
149 | vidmem_dbg(g, "Done!"); | ||
150 | |||
151 | return 0; | ||
152 | } | ||
153 | |||
154 | void nvgpu_vidmem_thread_pause_sync(struct mm_gk20a *mm) | ||
155 | { | ||
156 | /* | ||
157 | * On the first increment of the pause_count (0 -> 1) take the pause | ||
158 | * lock and prevent the vidmem clearing thread from processing work | ||
159 | * items. | ||
160 | * | ||
161 | * Otherwise the increment is all that's needed - it's essentially a | ||
162 | * ref-count for the number of pause() calls. | ||
163 | * | ||
164 | * The sync component is implemented by waiting for the lock to be | ||
165 | * released by the clearing thread in case the thread is currently | ||
166 | * processing work items. | ||
167 | */ | ||
168 | if (nvgpu_atomic_inc_return(&mm->vidmem.pause_count) == 1) | ||
169 | nvgpu_mutex_acquire(&mm->vidmem.clearing_thread_lock); | ||
170 | |||
171 | vidmem_dbg(mm->g, "Clearing thread paused; new count=%d", | ||
172 | nvgpu_atomic_read(&mm->vidmem.pause_count)); | ||
173 | } | ||
174 | |||
175 | void nvgpu_vidmem_thread_unpause(struct mm_gk20a *mm) | ||
176 | { | ||
177 | vidmem_dbg(mm->g, "Unpausing clearing thread; current count=%d", | ||
178 | nvgpu_atomic_read(&mm->vidmem.pause_count)); | ||
179 | |||
180 | /* | ||
181 | * And on the last decrement (1 -> 0) release the pause lock and let | ||
182 | * the vidmem clearing thread continue. | ||
183 | */ | ||
184 | if (nvgpu_atomic_dec_return(&mm->vidmem.pause_count) == 0) { | ||
185 | nvgpu_mutex_release(&mm->vidmem.clearing_thread_lock); | ||
186 | vidmem_dbg(mm->g, " > Clearing thread really unpaused!"); | ||
187 | } | ||
188 | } | ||
189 | |||
190 | int nvgpu_vidmem_clear_list_enqueue(struct gk20a *g, struct nvgpu_mem *mem) | ||
191 | { | ||
192 | struct mm_gk20a *mm = &g->mm; | ||
193 | |||
194 | /* | ||
195 | * Crap. Can't enqueue new vidmem bufs! CE may be gone! | ||
196 | * | ||
197 | * However, an errant app can hold a vidmem dma_buf FD open past when | ||
198 | * the nvgpu driver has exited. Thus when the FD does get closed | ||
199 | * eventually the dma_buf release function will try to call the vidmem | ||
200 | * free function which will attempt to enqueue the vidmem into the | ||
201 | * vidmem clearing thread. | ||
202 | */ | ||
203 | if (nvgpu_is_enabled(g, NVGPU_DRIVER_IS_DYING)) | ||
204 | return -ENOSYS; | ||
205 | |||
206 | nvgpu_mutex_acquire(&mm->vidmem.clear_list_mutex); | ||
207 | nvgpu_list_add_tail(&mem->clear_list_entry, | ||
208 | &mm->vidmem.clear_list_head); | ||
209 | nvgpu_atomic64_add(mem->aligned_size, &mm->vidmem.bytes_pending); | ||
210 | nvgpu_mutex_release(&mm->vidmem.clear_list_mutex); | ||
211 | |||
212 | nvgpu_cond_signal_interruptible(&mm->vidmem.clearing_thread_cond); | ||
213 | |||
214 | return 0; | ||
215 | } | ||
216 | |||
217 | static struct nvgpu_mem *nvgpu_vidmem_clear_list_dequeue(struct mm_gk20a *mm) | ||
218 | { | ||
219 | struct nvgpu_mem *mem = NULL; | ||
220 | |||
221 | nvgpu_mutex_acquire(&mm->vidmem.clear_list_mutex); | ||
222 | if (!nvgpu_list_empty(&mm->vidmem.clear_list_head)) { | ||
223 | mem = nvgpu_list_first_entry(&mm->vidmem.clear_list_head, | ||
224 | nvgpu_mem, clear_list_entry); | ||
225 | nvgpu_list_del(&mem->clear_list_entry); | ||
226 | } | ||
227 | nvgpu_mutex_release(&mm->vidmem.clear_list_mutex); | ||
228 | |||
229 | return mem; | ||
230 | } | ||
231 | |||
232 | static void nvgpu_vidmem_clear_pending_allocs(struct mm_gk20a *mm) | ||
233 | { | ||
234 | struct gk20a *g = mm->g; | ||
235 | struct nvgpu_mem *mem; | ||
236 | |||
237 | vidmem_dbg(g, "Running VIDMEM clearing thread:"); | ||
238 | |||
239 | while ((mem = nvgpu_vidmem_clear_list_dequeue(mm)) != NULL) { | ||
240 | nvgpu_vidmem_clear(g, mem); | ||
241 | |||
242 | WARN_ON(nvgpu_atomic64_sub_return(mem->aligned_size, | ||
243 | &g->mm.vidmem.bytes_pending) < 0); | ||
244 | mem->size = 0; | ||
245 | mem->aperture = APERTURE_INVALID; | ||
246 | |||
247 | __nvgpu_mem_free_vidmem_alloc(g, mem); | ||
248 | nvgpu_kfree(g, mem); | ||
249 | } | ||
250 | |||
251 | vidmem_dbg(g, "Done!"); | ||
252 | } | ||
253 | |||
254 | static int nvgpu_vidmem_clear_pending_allocs_thr(void *mm_ptr) | ||
255 | { | ||
256 | struct mm_gk20a *mm = mm_ptr; | ||
257 | |||
258 | /* | ||
259 | * Simple thread who's sole job is to periodically clear userspace | ||
260 | * vidmem allocations that have been recently freed. | ||
261 | * | ||
262 | * Since it doesn't make sense to run unless there's pending work a | ||
263 | * condition field is used to wait for work. When the DMA API frees a | ||
264 | * userspace vidmem buf it enqueues it into the clear list and alerts us | ||
265 | * that we have some work to do. | ||
266 | */ | ||
267 | |||
268 | while (!nvgpu_thread_should_stop(&mm->vidmem.clearing_thread)) { | ||
269 | int ret; | ||
270 | |||
271 | /* | ||
272 | * Wait for work but also make sure we should not be paused. | ||
273 | */ | ||
274 | ret = NVGPU_COND_WAIT_INTERRUPTIBLE( | ||
275 | &mm->vidmem.clearing_thread_cond, | ||
276 | nvgpu_thread_should_stop( | ||
277 | &mm->vidmem.clearing_thread) || | ||
278 | !nvgpu_list_empty(&mm->vidmem.clear_list_head), | ||
279 | 0); | ||
280 | if (ret == -ERESTARTSYS) | ||
281 | continue; | ||
282 | |||
283 | /* | ||
284 | * Use this lock to implement a pause mechanism. By taking this | ||
285 | * lock some other code can prevent this thread from processing | ||
286 | * work items. | ||
287 | */ | ||
288 | if (!nvgpu_mutex_tryacquire(&mm->vidmem.clearing_thread_lock)) | ||
289 | continue; | ||
290 | |||
291 | nvgpu_vidmem_clear_pending_allocs(mm); | ||
292 | |||
293 | nvgpu_mutex_release(&mm->vidmem.clearing_thread_lock); | ||
294 | } | ||
295 | |||
296 | return 0; | ||
297 | } | ||
298 | |||
299 | int nvgpu_vidmem_init(struct mm_gk20a *mm) | ||
300 | { | ||
301 | struct gk20a *g = mm->g; | ||
302 | size_t size = g->ops.mm.get_vidmem_size ? | ||
303 | g->ops.mm.get_vidmem_size(g) : 0; | ||
304 | u64 bootstrap_base, bootstrap_size, base; | ||
305 | u64 default_page_size = SZ_64K; | ||
306 | int err; | ||
307 | |||
308 | static struct nvgpu_alloc_carveout wpr_co = | ||
309 | NVGPU_CARVEOUT("wpr-region", 0, SZ_16M); | ||
310 | |||
311 | if (!size) | ||
312 | return 0; | ||
313 | |||
314 | vidmem_dbg(g, "init begin"); | ||
315 | |||
316 | wpr_co.base = size - SZ_256M; | ||
317 | bootstrap_base = wpr_co.base; | ||
318 | bootstrap_size = SZ_16M; | ||
319 | base = default_page_size; | ||
320 | |||
321 | /* | ||
322 | * Bootstrap allocator for use before the CE is initialized (CE | ||
323 | * initialization requires vidmem but we want to use the CE to zero | ||
324 | * out vidmem before allocating it... | ||
325 | */ | ||
326 | err = nvgpu_page_allocator_init(g, &g->mm.vidmem.bootstrap_allocator, | ||
327 | "vidmem-bootstrap", | ||
328 | bootstrap_base, bootstrap_size, | ||
329 | SZ_4K, 0); | ||
330 | |||
331 | err = nvgpu_page_allocator_init(g, &g->mm.vidmem.allocator, | ||
332 | "vidmem", | ||
333 | base, size - base, | ||
334 | default_page_size, | ||
335 | GPU_ALLOC_4K_VIDMEM_PAGES); | ||
336 | if (err) { | ||
337 | nvgpu_err(g, "Failed to register vidmem for size %zu: %d", | ||
338 | size, err); | ||
339 | return err; | ||
340 | } | ||
341 | |||
342 | /* Reserve bootstrap region in vidmem allocator */ | ||
343 | nvgpu_alloc_reserve_carveout(&g->mm.vidmem.allocator, &wpr_co); | ||
344 | |||
345 | mm->vidmem.base = base; | ||
346 | mm->vidmem.size = size - base; | ||
347 | mm->vidmem.bootstrap_base = bootstrap_base; | ||
348 | mm->vidmem.bootstrap_size = bootstrap_size; | ||
349 | |||
350 | err = nvgpu_cond_init(&mm->vidmem.clearing_thread_cond); | ||
351 | if (err) | ||
352 | goto fail; | ||
353 | |||
354 | nvgpu_atomic64_set(&mm->vidmem.bytes_pending, 0); | ||
355 | nvgpu_init_list_node(&mm->vidmem.clear_list_head); | ||
356 | nvgpu_mutex_init(&mm->vidmem.clear_list_mutex); | ||
357 | nvgpu_mutex_init(&mm->vidmem.clearing_thread_lock); | ||
358 | nvgpu_mutex_init(&mm->vidmem.first_clear_mutex); | ||
359 | nvgpu_atomic_set(&mm->vidmem.pause_count, 0); | ||
360 | |||
361 | /* | ||
362 | * Start the thread off in the paused state. The thread doesn't have to | ||
363 | * be running for this to work. It will be woken up later on in | ||
364 | * finalize_poweron(). We won't necessarily have a CE context yet | ||
365 | * either, so hypothetically one could cause a race where we try to | ||
366 | * clear a vidmem struct before we have a CE context to do so. | ||
367 | */ | ||
368 | nvgpu_vidmem_thread_pause_sync(mm); | ||
369 | |||
370 | err = nvgpu_thread_create(&mm->vidmem.clearing_thread, mm, | ||
371 | nvgpu_vidmem_clear_pending_allocs_thr, | ||
372 | "vidmem-clear"); | ||
373 | if (err) | ||
374 | goto fail; | ||
375 | |||
376 | vidmem_dbg(g, "VIDMEM Total: %zu MB", size >> 20); | ||
377 | vidmem_dbg(g, "VIDMEM Ranges:"); | ||
378 | vidmem_dbg(g, " 0x%-10llx -> 0x%-10llx Primary", | ||
379 | mm->vidmem.base, mm->vidmem.base + mm->vidmem.size); | ||
380 | vidmem_dbg(g, " 0x%-10llx -> 0x%-10llx Bootstrap", | ||
381 | mm->vidmem.bootstrap_base, | ||
382 | mm->vidmem.bootstrap_base + mm->vidmem.bootstrap_size); | ||
383 | vidmem_dbg(g, "VIDMEM carveouts:"); | ||
384 | vidmem_dbg(g, " 0x%-10llx -> 0x%-10llx %s", | ||
385 | wpr_co.base, wpr_co.base + wpr_co.length, wpr_co.name); | ||
386 | |||
387 | return 0; | ||
388 | |||
389 | fail: | ||
390 | nvgpu_cond_destroy(&mm->vidmem.clearing_thread_cond); | ||
391 | nvgpu_vidmem_destroy(g); | ||
392 | return err; | ||
393 | } | ||
394 | |||
395 | int nvgpu_vidmem_get_space(struct gk20a *g, u64 *space) | ||
396 | { | ||
397 | struct nvgpu_allocator *allocator = &g->mm.vidmem.allocator; | ||
398 | |||
399 | gk20a_dbg_fn(""); | ||
400 | |||
401 | if (!nvgpu_alloc_initialized(allocator)) | ||
402 | return -ENOSYS; | ||
403 | |||
404 | nvgpu_mutex_acquire(&g->mm.vidmem.clear_list_mutex); | ||
405 | *space = nvgpu_alloc_space(allocator) + | ||
406 | nvgpu_atomic64_read(&g->mm.vidmem.bytes_pending); | ||
407 | nvgpu_mutex_release(&g->mm.vidmem.clear_list_mutex); | ||
408 | return 0; | ||
409 | } | ||
410 | |||
411 | int nvgpu_vidmem_clear(struct gk20a *g, struct nvgpu_mem *mem) | ||
412 | { | ||
413 | struct gk20a_fence *gk20a_fence_out = NULL; | ||
414 | struct gk20a_fence *gk20a_last_fence = NULL; | ||
415 | struct nvgpu_page_alloc *alloc = NULL; | ||
416 | void *sgl = NULL; | ||
417 | int err = 0; | ||
418 | |||
419 | if (g->mm.vidmem.ce_ctx_id == (u32)~0) | ||
420 | return -EINVAL; | ||
421 | |||
422 | alloc = mem->vidmem_alloc; | ||
423 | |||
424 | vidmem_dbg(g, "Clearing VIDMEM buf:"); | ||
425 | |||
426 | nvgpu_sgt_for_each_sgl(sgl, &alloc->sgt) { | ||
427 | if (gk20a_last_fence) | ||
428 | gk20a_fence_put(gk20a_last_fence); | ||
429 | |||
430 | err = gk20a_ce_execute_ops(g, | ||
431 | g->mm.vidmem.ce_ctx_id, | ||
432 | 0, | ||
433 | nvgpu_sgt_get_phys(&alloc->sgt, sgl), | ||
434 | nvgpu_sgt_get_length(&alloc->sgt, sgl), | ||
435 | 0x00000000, | ||
436 | NVGPU_CE_DST_LOCATION_LOCAL_FB, | ||
437 | NVGPU_CE_MEMSET, | ||
438 | NULL, | ||
439 | 0, | ||
440 | &gk20a_fence_out); | ||
441 | |||
442 | if (err) { | ||
443 | nvgpu_err(g, | ||
444 | "Failed gk20a_ce_execute_ops[%d]", err); | ||
445 | return err; | ||
446 | } | ||
447 | |||
448 | vidmem_dbg(g, " > [0x%llx +0x%llx]", | ||
449 | nvgpu_sgt_get_phys(&alloc->sgt, sgl), | ||
450 | nvgpu_sgt_get_length(&alloc->sgt, sgl)); | ||
451 | |||
452 | gk20a_last_fence = gk20a_fence_out; | ||
453 | } | ||
454 | |||
455 | if (gk20a_last_fence) { | ||
456 | struct nvgpu_timeout timeout; | ||
457 | |||
458 | nvgpu_timeout_init(g, &timeout, | ||
459 | gk20a_get_gr_idle_timeout(g), | ||
460 | NVGPU_TIMER_CPU_TIMER); | ||
461 | |||
462 | do { | ||
463 | err = gk20a_fence_wait(g, gk20a_last_fence, | ||
464 | gk20a_get_gr_idle_timeout(g)); | ||
465 | } while (err == -ERESTARTSYS && | ||
466 | !nvgpu_timeout_expired(&timeout)); | ||
467 | |||
468 | gk20a_fence_put(gk20a_last_fence); | ||
469 | if (err) | ||
470 | nvgpu_err(g, | ||
471 | "fence wait failed for CE execute ops"); | ||
472 | } | ||
473 | |||
474 | vidmem_dbg(g, " Done"); | ||
475 | |||
476 | return err; | ||
477 | } | ||
478 | |||
479 | static int nvgpu_vidmem_clear_all(struct gk20a *g) | ||
480 | { | ||
481 | int err; | ||
482 | |||
483 | if (g->mm.vidmem.cleared) | ||
484 | return 0; | ||
485 | |||
486 | nvgpu_mutex_acquire(&g->mm.vidmem.first_clear_mutex); | ||
487 | if (!g->mm.vidmem.cleared) { | ||
488 | err = __nvgpu_vidmem_do_clear_all(g); | ||
489 | if (err) { | ||
490 | nvgpu_mutex_release(&g->mm.vidmem.first_clear_mutex); | ||
491 | nvgpu_err(g, "failed to clear whole vidmem"); | ||
492 | return err; | ||
493 | } | ||
494 | } | ||
495 | nvgpu_mutex_release(&g->mm.vidmem.first_clear_mutex); | ||
496 | |||
497 | return 0; | ||
498 | } | ||
499 | |||
500 | struct nvgpu_vidmem_buf *nvgpu_vidmem_user_alloc(struct gk20a *g, size_t bytes) | ||
501 | { | ||
502 | struct nvgpu_vidmem_buf *buf; | ||
503 | int err; | ||
504 | |||
505 | err = nvgpu_vidmem_clear_all(g); | ||
506 | if (err) | ||
507 | return NULL; | ||
508 | |||
509 | buf = nvgpu_kzalloc(g, sizeof(*buf)); | ||
510 | if (!buf) | ||
511 | return NULL; | ||
512 | |||
513 | buf->g = g; | ||
514 | buf->mem = nvgpu_kzalloc(g, sizeof(*buf->mem)); | ||
515 | if (!buf->mem) | ||
516 | goto fail; | ||
517 | |||
518 | err = nvgpu_dma_alloc_vid(g, bytes, buf->mem); | ||
519 | if (err) | ||
520 | goto fail; | ||
521 | |||
522 | /* | ||
523 | * Alerts the DMA API that when we free this vidmem buf we have to | ||
524 | * clear it to avoid leaking data to userspace. | ||
525 | */ | ||
526 | buf->mem->mem_flags |= NVGPU_MEM_FLAG_USER_MEM; | ||
527 | |||
528 | return buf; | ||
529 | |||
530 | fail: | ||
531 | /* buf will never be NULL here. */ | ||
532 | nvgpu_kfree(g, buf->mem); | ||
533 | nvgpu_kfree(g, buf); | ||
534 | return NULL; | ||
535 | } | ||
536 | |||
537 | void nvgpu_vidmem_buf_free(struct gk20a *g, struct nvgpu_vidmem_buf *buf) | ||
538 | { | ||
539 | /* | ||
540 | * In some error paths it's convenient to be able to "free" a NULL buf. | ||
541 | */ | ||
542 | if (!buf) | ||
543 | return; | ||
544 | |||
545 | nvgpu_dma_free(g, buf->mem); | ||
546 | |||
547 | /* | ||
548 | * We don't free buf->mem here. This is handled by nvgpu_dma_free()! | ||
549 | * Since these buffers are cleared in the background the nvgpu_mem | ||
550 | * struct must live on through that. We transfer ownership here to the | ||
551 | * DMA API and let the DMA API free the buffer. | ||
552 | */ | ||
553 | nvgpu_kfree(g, buf); | ||
554 | } | ||