]>
Commit | Line | Data |
---|---|---|
d38ceaf9 AD |
1 | /* |
2 | * Copyright 2008 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 "Software"), | |
7 | * to deal in the Software without restriction, including without limitation | |
8 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, | |
9 | * and/or sell copies of the Software, and to permit persons to whom the | |
10 | * Software is furnished to do so, subject to the following conditions: | |
11 | * | |
12 | * The above copyright notice and this permission notice (including the next | |
13 | * paragraph) shall be included in all copies or substantial portions of the | |
14 | * 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 | * PRECISION INSIGHT AND/OR ITS SUPPLIERS 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 OTHER | |
22 | * DEALINGS IN THE SOFTWARE. | |
23 | * | |
24 | * Authors: | |
25 | * Jerome Glisse <glisse@freedesktop.org> | |
26 | */ | |
27 | #include <linux/list_sort.h> | |
28 | #include <drm/drmP.h> | |
29 | #include <drm/amdgpu_drm.h> | |
30 | #include "amdgpu.h" | |
31 | #include "amdgpu_trace.h" | |
32 | ||
33 | #define AMDGPU_CS_MAX_PRIORITY 32u | |
34 | #define AMDGPU_CS_NUM_BUCKETS (AMDGPU_CS_MAX_PRIORITY + 1) | |
35 | ||
36 | /* This is based on the bucket sort with O(n) time complexity. | |
37 | * An item with priority "i" is added to bucket[i]. The lists are then | |
38 | * concatenated in descending order. | |
39 | */ | |
40 | struct amdgpu_cs_buckets { | |
41 | struct list_head bucket[AMDGPU_CS_NUM_BUCKETS]; | |
42 | }; | |
43 | ||
44 | static void amdgpu_cs_buckets_init(struct amdgpu_cs_buckets *b) | |
45 | { | |
46 | unsigned i; | |
47 | ||
48 | for (i = 0; i < AMDGPU_CS_NUM_BUCKETS; i++) | |
49 | INIT_LIST_HEAD(&b->bucket[i]); | |
50 | } | |
51 | ||
52 | static void amdgpu_cs_buckets_add(struct amdgpu_cs_buckets *b, | |
53 | struct list_head *item, unsigned priority) | |
54 | { | |
55 | /* Since buffers which appear sooner in the relocation list are | |
56 | * likely to be used more often than buffers which appear later | |
57 | * in the list, the sort mustn't change the ordering of buffers | |
58 | * with the same priority, i.e. it must be stable. | |
59 | */ | |
60 | list_add_tail(item, &b->bucket[min(priority, AMDGPU_CS_MAX_PRIORITY)]); | |
61 | } | |
62 | ||
63 | static void amdgpu_cs_buckets_get_list(struct amdgpu_cs_buckets *b, | |
64 | struct list_head *out_list) | |
65 | { | |
66 | unsigned i; | |
67 | ||
68 | /* Connect the sorted buckets in the output list. */ | |
69 | for (i = 0; i < AMDGPU_CS_NUM_BUCKETS; i++) { | |
70 | list_splice(&b->bucket[i], out_list); | |
71 | } | |
72 | } | |
73 | ||
74 | int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type, | |
75 | u32 ip_instance, u32 ring, | |
76 | struct amdgpu_ring **out_ring) | |
77 | { | |
78 | /* Right now all IPs have only one instance - multiple rings. */ | |
79 | if (ip_instance != 0) { | |
80 | DRM_ERROR("invalid ip instance: %d\n", ip_instance); | |
81 | return -EINVAL; | |
82 | } | |
83 | ||
84 | switch (ip_type) { | |
85 | default: | |
86 | DRM_ERROR("unknown ip type: %d\n", ip_type); | |
87 | return -EINVAL; | |
88 | case AMDGPU_HW_IP_GFX: | |
89 | if (ring < adev->gfx.num_gfx_rings) { | |
90 | *out_ring = &adev->gfx.gfx_ring[ring]; | |
91 | } else { | |
92 | DRM_ERROR("only %d gfx rings are supported now\n", | |
93 | adev->gfx.num_gfx_rings); | |
94 | return -EINVAL; | |
95 | } | |
96 | break; | |
97 | case AMDGPU_HW_IP_COMPUTE: | |
98 | if (ring < adev->gfx.num_compute_rings) { | |
99 | *out_ring = &adev->gfx.compute_ring[ring]; | |
100 | } else { | |
101 | DRM_ERROR("only %d compute rings are supported now\n", | |
102 | adev->gfx.num_compute_rings); | |
103 | return -EINVAL; | |
104 | } | |
105 | break; | |
106 | case AMDGPU_HW_IP_DMA: | |
c113ea1c AD |
107 | if (ring < adev->sdma.num_instances) { |
108 | *out_ring = &adev->sdma.instance[ring].ring; | |
d38ceaf9 | 109 | } else { |
c113ea1c AD |
110 | DRM_ERROR("only %d SDMA rings are supported\n", |
111 | adev->sdma.num_instances); | |
d38ceaf9 AD |
112 | return -EINVAL; |
113 | } | |
114 | break; | |
115 | case AMDGPU_HW_IP_UVD: | |
116 | *out_ring = &adev->uvd.ring; | |
117 | break; | |
118 | case AMDGPU_HW_IP_VCE: | |
119 | if (ring < 2){ | |
120 | *out_ring = &adev->vce.ring[ring]; | |
121 | } else { | |
122 | DRM_ERROR("only two VCE rings are supported\n"); | |
123 | return -EINVAL; | |
124 | } | |
125 | break; | |
126 | } | |
127 | return 0; | |
128 | } | |
129 | ||
049fc527 CZ |
130 | struct amdgpu_cs_parser *amdgpu_cs_parser_create(struct amdgpu_device *adev, |
131 | struct drm_file *filp, | |
132 | struct amdgpu_ctx *ctx, | |
133 | struct amdgpu_ib *ibs, | |
134 | uint32_t num_ibs) | |
135 | { | |
136 | struct amdgpu_cs_parser *parser; | |
137 | int i; | |
138 | ||
139 | parser = kzalloc(sizeof(struct amdgpu_cs_parser), GFP_KERNEL); | |
140 | if (!parser) | |
141 | return NULL; | |
142 | ||
143 | parser->adev = adev; | |
144 | parser->filp = filp; | |
145 | parser->ctx = ctx; | |
146 | parser->ibs = ibs; | |
147 | parser->num_ibs = num_ibs; | |
049fc527 CZ |
148 | for (i = 0; i < num_ibs; i++) |
149 | ibs[i].ctx = ctx; | |
150 | ||
151 | return parser; | |
152 | } | |
153 | ||
d38ceaf9 AD |
154 | int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data) |
155 | { | |
156 | union drm_amdgpu_cs *cs = data; | |
157 | uint64_t *chunk_array_user; | |
1d263474 | 158 | uint64_t *chunk_array; |
d38ceaf9 | 159 | struct amdgpu_fpriv *fpriv = p->filp->driver_priv; |
54313503 DC |
160 | unsigned size; |
161 | int i; | |
1d263474 | 162 | int ret; |
d38ceaf9 | 163 | |
1d263474 DC |
164 | if (cs->in.num_chunks == 0) |
165 | return 0; | |
166 | ||
167 | chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL); | |
168 | if (!chunk_array) | |
169 | return -ENOMEM; | |
d38ceaf9 | 170 | |
3cb485f3 CK |
171 | p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id); |
172 | if (!p->ctx) { | |
1d263474 DC |
173 | ret = -EINVAL; |
174 | goto free_chunk; | |
3cb485f3 | 175 | } |
1d263474 | 176 | |
a3348bb8 | 177 | p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle); |
d38ceaf9 AD |
178 | |
179 | /* get chunks */ | |
180 | INIT_LIST_HEAD(&p->validated); | |
028423b0 | 181 | chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks); |
d38ceaf9 AD |
182 | if (copy_from_user(chunk_array, chunk_array_user, |
183 | sizeof(uint64_t)*cs->in.num_chunks)) { | |
1d263474 DC |
184 | ret = -EFAULT; |
185 | goto put_bo_list; | |
d38ceaf9 AD |
186 | } |
187 | ||
188 | p->nchunks = cs->in.num_chunks; | |
e60b344f | 189 | p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk), |
d38ceaf9 | 190 | GFP_KERNEL); |
1d263474 DC |
191 | if (!p->chunks) { |
192 | ret = -ENOMEM; | |
193 | goto put_bo_list; | |
d38ceaf9 AD |
194 | } |
195 | ||
196 | for (i = 0; i < p->nchunks; i++) { | |
197 | struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL; | |
198 | struct drm_amdgpu_cs_chunk user_chunk; | |
199 | uint32_t __user *cdata; | |
200 | ||
028423b0 | 201 | chunk_ptr = (void __user *)(unsigned long)chunk_array[i]; |
d38ceaf9 AD |
202 | if (copy_from_user(&user_chunk, chunk_ptr, |
203 | sizeof(struct drm_amdgpu_cs_chunk))) { | |
1d263474 DC |
204 | ret = -EFAULT; |
205 | i--; | |
206 | goto free_partial_kdata; | |
d38ceaf9 AD |
207 | } |
208 | p->chunks[i].chunk_id = user_chunk.chunk_id; | |
209 | p->chunks[i].length_dw = user_chunk.length_dw; | |
d38ceaf9 AD |
210 | |
211 | size = p->chunks[i].length_dw; | |
028423b0 | 212 | cdata = (void __user *)(unsigned long)user_chunk.chunk_data; |
d38ceaf9 AD |
213 | p->chunks[i].user_ptr = cdata; |
214 | ||
215 | p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t)); | |
216 | if (p->chunks[i].kdata == NULL) { | |
1d263474 DC |
217 | ret = -ENOMEM; |
218 | i--; | |
219 | goto free_partial_kdata; | |
d38ceaf9 AD |
220 | } |
221 | size *= sizeof(uint32_t); | |
222 | if (copy_from_user(p->chunks[i].kdata, cdata, size)) { | |
1d263474 DC |
223 | ret = -EFAULT; |
224 | goto free_partial_kdata; | |
d38ceaf9 AD |
225 | } |
226 | ||
9a5e8fb1 CK |
227 | switch (p->chunks[i].chunk_id) { |
228 | case AMDGPU_CHUNK_ID_IB: | |
229 | p->num_ibs++; | |
230 | break; | |
231 | ||
232 | case AMDGPU_CHUNK_ID_FENCE: | |
d38ceaf9 AD |
233 | size = sizeof(struct drm_amdgpu_cs_chunk_fence); |
234 | if (p->chunks[i].length_dw * sizeof(uint32_t) >= size) { | |
235 | uint32_t handle; | |
236 | struct drm_gem_object *gobj; | |
237 | struct drm_amdgpu_cs_chunk_fence *fence_data; | |
238 | ||
239 | fence_data = (void *)p->chunks[i].kdata; | |
240 | handle = fence_data->handle; | |
241 | gobj = drm_gem_object_lookup(p->adev->ddev, | |
242 | p->filp, handle); | |
243 | if (gobj == NULL) { | |
1d263474 DC |
244 | ret = -EINVAL; |
245 | goto free_partial_kdata; | |
d38ceaf9 AD |
246 | } |
247 | ||
248 | p->uf.bo = gem_to_amdgpu_bo(gobj); | |
249 | p->uf.offset = fence_data->offset; | |
250 | } else { | |
1d263474 DC |
251 | ret = -EINVAL; |
252 | goto free_partial_kdata; | |
d38ceaf9 | 253 | } |
9a5e8fb1 CK |
254 | break; |
255 | ||
2b48d323 CK |
256 | case AMDGPU_CHUNK_ID_DEPENDENCIES: |
257 | break; | |
258 | ||
9a5e8fb1 | 259 | default: |
1d263474 DC |
260 | ret = -EINVAL; |
261 | goto free_partial_kdata; | |
d38ceaf9 AD |
262 | } |
263 | } | |
264 | ||
e60b344f | 265 | |
b203dd95 | 266 | p->ibs = kcalloc(p->num_ibs, sizeof(struct amdgpu_ib), GFP_KERNEL); |
1d263474 DC |
267 | if (!p->ibs) { |
268 | ret = -ENOMEM; | |
269 | goto free_all_kdata; | |
270 | } | |
d38ceaf9 | 271 | |
d38ceaf9 | 272 | kfree(chunk_array); |
1d263474 DC |
273 | return 0; |
274 | ||
275 | free_all_kdata: | |
276 | i = p->nchunks - 1; | |
277 | free_partial_kdata: | |
278 | for (; i >= 0; i--) | |
279 | drm_free_large(p->chunks[i].kdata); | |
280 | kfree(p->chunks); | |
281 | put_bo_list: | |
282 | if (p->bo_list) | |
283 | amdgpu_bo_list_put(p->bo_list); | |
284 | amdgpu_ctx_put(p->ctx); | |
285 | free_chunk: | |
286 | kfree(chunk_array); | |
287 | ||
288 | return ret; | |
d38ceaf9 AD |
289 | } |
290 | ||
291 | /* Returns how many bytes TTM can move per IB. | |
292 | */ | |
293 | static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev) | |
294 | { | |
295 | u64 real_vram_size = adev->mc.real_vram_size; | |
296 | u64 vram_usage = atomic64_read(&adev->vram_usage); | |
297 | ||
298 | /* This function is based on the current VRAM usage. | |
299 | * | |
300 | * - If all of VRAM is free, allow relocating the number of bytes that | |
301 | * is equal to 1/4 of the size of VRAM for this IB. | |
302 | ||
303 | * - If more than one half of VRAM is occupied, only allow relocating | |
304 | * 1 MB of data for this IB. | |
305 | * | |
306 | * - From 0 to one half of used VRAM, the threshold decreases | |
307 | * linearly. | |
308 | * __________________ | |
309 | * 1/4 of -|\ | | |
310 | * VRAM | \ | | |
311 | * | \ | | |
312 | * | \ | | |
313 | * | \ | | |
314 | * | \ | | |
315 | * | \ | | |
316 | * | \________|1 MB | |
317 | * |----------------| | |
318 | * VRAM 0 % 100 % | |
319 | * used used | |
320 | * | |
321 | * Note: It's a threshold, not a limit. The threshold must be crossed | |
322 | * for buffer relocations to stop, so any buffer of an arbitrary size | |
323 | * can be moved as long as the threshold isn't crossed before | |
324 | * the relocation takes place. We don't want to disable buffer | |
325 | * relocations completely. | |
326 | * | |
327 | * The idea is that buffers should be placed in VRAM at creation time | |
328 | * and TTM should only do a minimum number of relocations during | |
329 | * command submission. In practice, you need to submit at least | |
330 | * a dozen IBs to move all buffers to VRAM if they are in GTT. | |
331 | * | |
332 | * Also, things can get pretty crazy under memory pressure and actual | |
333 | * VRAM usage can change a lot, so playing safe even at 50% does | |
334 | * consistently increase performance. | |
335 | */ | |
336 | ||
337 | u64 half_vram = real_vram_size >> 1; | |
338 | u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage; | |
339 | u64 bytes_moved_threshold = half_free_vram >> 1; | |
340 | return max(bytes_moved_threshold, 1024*1024ull); | |
341 | } | |
342 | ||
a5b75058 CK |
343 | int amdgpu_cs_list_validate(struct amdgpu_device *adev, |
344 | struct amdgpu_vm *vm, | |
345 | struct list_head *validated) | |
d38ceaf9 | 346 | { |
d38ceaf9 | 347 | struct amdgpu_bo_list_entry *lobj; |
d38ceaf9 AD |
348 | struct amdgpu_bo *bo; |
349 | u64 bytes_moved = 0, initial_bytes_moved; | |
350 | u64 bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(adev); | |
351 | int r; | |
352 | ||
a5b75058 | 353 | list_for_each_entry(lobj, validated, tv.head) { |
d38ceaf9 AD |
354 | bo = lobj->robj; |
355 | if (!bo->pin_count) { | |
356 | u32 domain = lobj->prefered_domains; | |
357 | u32 current_domain = | |
358 | amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type); | |
359 | ||
360 | /* Check if this buffer will be moved and don't move it | |
361 | * if we have moved too many buffers for this IB already. | |
362 | * | |
363 | * Note that this allows moving at least one buffer of | |
364 | * any size, because it doesn't take the current "bo" | |
365 | * into account. We don't want to disallow buffer moves | |
366 | * completely. | |
367 | */ | |
270e869d | 368 | if ((lobj->allowed_domains & current_domain) != 0 && |
d38ceaf9 AD |
369 | (domain & current_domain) == 0 && /* will be moved */ |
370 | bytes_moved > bytes_moved_threshold) { | |
371 | /* don't move it */ | |
372 | domain = current_domain; | |
373 | } | |
374 | ||
375 | retry: | |
376 | amdgpu_ttm_placement_from_domain(bo, domain); | |
377 | initial_bytes_moved = atomic64_read(&adev->num_bytes_moved); | |
378 | r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false); | |
379 | bytes_moved += atomic64_read(&adev->num_bytes_moved) - | |
380 | initial_bytes_moved; | |
381 | ||
382 | if (unlikely(r)) { | |
383 | if (r != -ERESTARTSYS && domain != lobj->allowed_domains) { | |
384 | domain = lobj->allowed_domains; | |
385 | goto retry; | |
386 | } | |
d38ceaf9 AD |
387 | return r; |
388 | } | |
389 | } | |
390 | lobj->bo_va = amdgpu_vm_bo_find(vm, bo); | |
391 | } | |
392 | return 0; | |
393 | } | |
394 | ||
395 | static int amdgpu_cs_parser_relocs(struct amdgpu_cs_parser *p) | |
396 | { | |
397 | struct amdgpu_fpriv *fpriv = p->filp->driver_priv; | |
398 | struct amdgpu_cs_buckets buckets; | |
a5b75058 | 399 | struct list_head duplicates; |
840d5144 | 400 | bool need_mmap_lock = false; |
d38ceaf9 AD |
401 | int i, r; |
402 | ||
840d5144 | 403 | if (p->bo_list) { |
404 | need_mmap_lock = p->bo_list->has_userptr; | |
405 | amdgpu_cs_buckets_init(&buckets); | |
406 | for (i = 0; i < p->bo_list->num_entries; i++) | |
407 | amdgpu_cs_buckets_add(&buckets, &p->bo_list->array[i].tv.head, | |
408 | p->bo_list->array[i].priority); | |
d38ceaf9 | 409 | |
840d5144 | 410 | amdgpu_cs_buckets_get_list(&buckets, &p->validated); |
411 | } | |
d38ceaf9 | 412 | |
d38ceaf9 AD |
413 | p->vm_bos = amdgpu_vm_get_bos(p->adev, &fpriv->vm, |
414 | &p->validated); | |
415 | ||
d38ceaf9 AD |
416 | if (need_mmap_lock) |
417 | down_read(¤t->mm->mmap_sem); | |
418 | ||
a5b75058 CK |
419 | INIT_LIST_HEAD(&duplicates); |
420 | r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true, &duplicates); | |
421 | if (unlikely(r != 0)) | |
422 | goto error_reserve; | |
423 | ||
424 | r = amdgpu_cs_list_validate(p->adev, &fpriv->vm, &p->validated); | |
425 | if (r) | |
426 | goto error_validate; | |
427 | ||
428 | r = amdgpu_cs_list_validate(p->adev, &fpriv->vm, &duplicates); | |
429 | ||
430 | error_validate: | |
431 | if (r) | |
432 | ttm_eu_backoff_reservation(&p->ticket, &p->validated); | |
d38ceaf9 | 433 | |
a5b75058 | 434 | error_reserve: |
d38ceaf9 AD |
435 | if (need_mmap_lock) |
436 | up_read(¤t->mm->mmap_sem); | |
437 | ||
438 | return r; | |
439 | } | |
440 | ||
441 | static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p) | |
442 | { | |
443 | struct amdgpu_bo_list_entry *e; | |
444 | int r; | |
445 | ||
446 | list_for_each_entry(e, &p->validated, tv.head) { | |
447 | struct reservation_object *resv = e->robj->tbo.resv; | |
448 | r = amdgpu_sync_resv(p->adev, &p->ibs[0].sync, resv, p->filp); | |
449 | ||
450 | if (r) | |
451 | return r; | |
452 | } | |
453 | return 0; | |
454 | } | |
455 | ||
456 | static int cmp_size_smaller_first(void *priv, struct list_head *a, | |
457 | struct list_head *b) | |
458 | { | |
459 | struct amdgpu_bo_list_entry *la = list_entry(a, struct amdgpu_bo_list_entry, tv.head); | |
460 | struct amdgpu_bo_list_entry *lb = list_entry(b, struct amdgpu_bo_list_entry, tv.head); | |
461 | ||
462 | /* Sort A before B if A is smaller. */ | |
463 | return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages; | |
464 | } | |
465 | ||
049fc527 CZ |
466 | static void amdgpu_cs_parser_fini_early(struct amdgpu_cs_parser *parser, int error, bool backoff) |
467 | { | |
d38ceaf9 AD |
468 | if (!error) { |
469 | /* Sort the buffer list from the smallest to largest buffer, | |
470 | * which affects the order of buffers in the LRU list. | |
471 | * This assures that the smallest buffers are added first | |
472 | * to the LRU list, so they are likely to be later evicted | |
473 | * first, instead of large buffers whose eviction is more | |
474 | * expensive. | |
475 | * | |
476 | * This slightly lowers the number of bytes moved by TTM | |
477 | * per frame under memory pressure. | |
478 | */ | |
479 | list_sort(NULL, &parser->validated, cmp_size_smaller_first); | |
480 | ||
481 | ttm_eu_fence_buffer_objects(&parser->ticket, | |
482 | &parser->validated, | |
483 | &parser->ibs[parser->num_ibs-1].fence->base); | |
484 | } else if (backoff) { | |
485 | ttm_eu_backoff_reservation(&parser->ticket, | |
486 | &parser->validated); | |
487 | } | |
049fc527 | 488 | } |
d38ceaf9 | 489 | |
049fc527 CZ |
490 | static void amdgpu_cs_parser_fini_late(struct amdgpu_cs_parser *parser) |
491 | { | |
492 | unsigned i; | |
3cb485f3 CK |
493 | if (parser->ctx) |
494 | amdgpu_ctx_put(parser->ctx); | |
a3348bb8 CZ |
495 | if (parser->bo_list) |
496 | amdgpu_bo_list_put(parser->bo_list); | |
497 | ||
d38ceaf9 AD |
498 | drm_free_large(parser->vm_bos); |
499 | for (i = 0; i < parser->nchunks; i++) | |
500 | drm_free_large(parser->chunks[i].kdata); | |
501 | kfree(parser->chunks); | |
e4a58a28 CK |
502 | if (parser->ibs) |
503 | for (i = 0; i < parser->num_ibs; i++) | |
504 | amdgpu_ib_free(parser->adev, &parser->ibs[i]); | |
505 | kfree(parser->ibs); | |
506 | if (parser->uf.bo) | |
507 | drm_gem_object_unreference_unlocked(&parser->uf.bo->gem_base); | |
bb977d37 | 508 | kfree(parser); |
d38ceaf9 AD |
509 | } |
510 | ||
351dba73 CK |
511 | /** |
512 | * cs_parser_fini() - clean parser states | |
513 | * @parser: parser structure holding parsing context. | |
514 | * @error: error number | |
515 | * | |
516 | * If error is set than unvalidate buffer, otherwise just free memory | |
517 | * used by parsing context. | |
518 | **/ | |
519 | static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff) | |
520 | { | |
521 | amdgpu_cs_parser_fini_early(parser, error, backoff); | |
522 | amdgpu_cs_parser_fini_late(parser); | |
523 | } | |
524 | ||
d38ceaf9 AD |
525 | static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p, |
526 | struct amdgpu_vm *vm) | |
527 | { | |
528 | struct amdgpu_device *adev = p->adev; | |
529 | struct amdgpu_bo_va *bo_va; | |
530 | struct amdgpu_bo *bo; | |
531 | int i, r; | |
532 | ||
533 | r = amdgpu_vm_update_page_directory(adev, vm); | |
534 | if (r) | |
535 | return r; | |
536 | ||
05906dec BN |
537 | r = amdgpu_sync_fence(adev, &p->ibs[0].sync, vm->page_directory_fence); |
538 | if (r) | |
539 | return r; | |
540 | ||
d38ceaf9 AD |
541 | r = amdgpu_vm_clear_freed(adev, vm); |
542 | if (r) | |
543 | return r; | |
544 | ||
545 | if (p->bo_list) { | |
546 | for (i = 0; i < p->bo_list->num_entries; i++) { | |
91e1a520 CK |
547 | struct fence *f; |
548 | ||
d38ceaf9 AD |
549 | /* ignore duplicates */ |
550 | bo = p->bo_list->array[i].robj; | |
551 | if (!bo) | |
552 | continue; | |
553 | ||
554 | bo_va = p->bo_list->array[i].bo_va; | |
555 | if (bo_va == NULL) | |
556 | continue; | |
557 | ||
558 | r = amdgpu_vm_bo_update(adev, bo_va, &bo->tbo.mem); | |
559 | if (r) | |
560 | return r; | |
561 | ||
bb1e38a4 | 562 | f = bo_va->last_pt_update; |
91e1a520 CK |
563 | r = amdgpu_sync_fence(adev, &p->ibs[0].sync, f); |
564 | if (r) | |
565 | return r; | |
d38ceaf9 | 566 | } |
b495bd3a CK |
567 | |
568 | } | |
569 | ||
570 | r = amdgpu_vm_clear_invalids(adev, vm, &p->ibs[0].sync); | |
571 | ||
572 | if (amdgpu_vm_debug && p->bo_list) { | |
573 | /* Invalidate all BOs to test for userspace bugs */ | |
574 | for (i = 0; i < p->bo_list->num_entries; i++) { | |
575 | /* ignore duplicates */ | |
576 | bo = p->bo_list->array[i].robj; | |
577 | if (!bo) | |
578 | continue; | |
579 | ||
580 | amdgpu_vm_bo_invalidate(adev, bo); | |
581 | } | |
d38ceaf9 AD |
582 | } |
583 | ||
b495bd3a | 584 | return r; |
d38ceaf9 AD |
585 | } |
586 | ||
587 | static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev, | |
588 | struct amdgpu_cs_parser *parser) | |
589 | { | |
590 | struct amdgpu_fpriv *fpriv = parser->filp->driver_priv; | |
591 | struct amdgpu_vm *vm = &fpriv->vm; | |
592 | struct amdgpu_ring *ring; | |
593 | int i, r; | |
594 | ||
595 | if (parser->num_ibs == 0) | |
596 | return 0; | |
597 | ||
598 | /* Only for UVD/VCE VM emulation */ | |
599 | for (i = 0; i < parser->num_ibs; i++) { | |
600 | ring = parser->ibs[i].ring; | |
601 | if (ring->funcs->parse_cs) { | |
602 | r = amdgpu_ring_parse_cs(ring, parser, i); | |
603 | if (r) | |
604 | return r; | |
605 | } | |
606 | } | |
607 | ||
d38ceaf9 AD |
608 | r = amdgpu_bo_vm_update_pte(parser, vm); |
609 | if (r) { | |
610 | goto out; | |
611 | } | |
612 | amdgpu_cs_sync_rings(parser); | |
049fc527 CZ |
613 | if (!amdgpu_enable_scheduler) |
614 | r = amdgpu_ib_schedule(adev, parser->num_ibs, parser->ibs, | |
615 | parser->filp); | |
d38ceaf9 AD |
616 | |
617 | out: | |
d38ceaf9 AD |
618 | return r; |
619 | } | |
620 | ||
621 | static int amdgpu_cs_handle_lockup(struct amdgpu_device *adev, int r) | |
622 | { | |
623 | if (r == -EDEADLK) { | |
624 | r = amdgpu_gpu_reset(adev); | |
625 | if (!r) | |
626 | r = -EAGAIN; | |
627 | } | |
628 | return r; | |
629 | } | |
630 | ||
631 | static int amdgpu_cs_ib_fill(struct amdgpu_device *adev, | |
632 | struct amdgpu_cs_parser *parser) | |
633 | { | |
634 | struct amdgpu_fpriv *fpriv = parser->filp->driver_priv; | |
635 | struct amdgpu_vm *vm = &fpriv->vm; | |
636 | int i, j; | |
637 | int r; | |
638 | ||
639 | for (i = 0, j = 0; i < parser->nchunks && j < parser->num_ibs; i++) { | |
640 | struct amdgpu_cs_chunk *chunk; | |
641 | struct amdgpu_ib *ib; | |
642 | struct drm_amdgpu_cs_chunk_ib *chunk_ib; | |
d38ceaf9 | 643 | struct amdgpu_ring *ring; |
d38ceaf9 AD |
644 | |
645 | chunk = &parser->chunks[i]; | |
646 | ib = &parser->ibs[j]; | |
647 | chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata; | |
648 | ||
649 | if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB) | |
650 | continue; | |
651 | ||
d38ceaf9 AD |
652 | r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type, |
653 | chunk_ib->ip_instance, chunk_ib->ring, | |
654 | &ring); | |
3ccec53c | 655 | if (r) |
d38ceaf9 | 656 | return r; |
d38ceaf9 AD |
657 | |
658 | if (ring->funcs->parse_cs) { | |
4802ce11 | 659 | struct amdgpu_bo_va_mapping *m; |
3ccec53c | 660 | struct amdgpu_bo *aobj = NULL; |
4802ce11 CK |
661 | uint64_t offset; |
662 | uint8_t *kptr; | |
3ccec53c | 663 | |
4802ce11 CK |
664 | m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start, |
665 | &aobj); | |
3ccec53c MO |
666 | if (!aobj) { |
667 | DRM_ERROR("IB va_start is invalid\n"); | |
668 | return -EINVAL; | |
d38ceaf9 AD |
669 | } |
670 | ||
4802ce11 CK |
671 | if ((chunk_ib->va_start + chunk_ib->ib_bytes) > |
672 | (m->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) { | |
673 | DRM_ERROR("IB va_start+ib_bytes is invalid\n"); | |
674 | return -EINVAL; | |
675 | } | |
676 | ||
3ccec53c | 677 | /* the IB should be reserved at this point */ |
4802ce11 | 678 | r = amdgpu_bo_kmap(aobj, (void **)&kptr); |
d38ceaf9 | 679 | if (r) { |
d38ceaf9 AD |
680 | return r; |
681 | } | |
682 | ||
4802ce11 CK |
683 | offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE; |
684 | kptr += chunk_ib->va_start - offset; | |
685 | ||
d38ceaf9 AD |
686 | r = amdgpu_ib_get(ring, NULL, chunk_ib->ib_bytes, ib); |
687 | if (r) { | |
688 | DRM_ERROR("Failed to get ib !\n"); | |
d38ceaf9 AD |
689 | return r; |
690 | } | |
691 | ||
692 | memcpy(ib->ptr, kptr, chunk_ib->ib_bytes); | |
693 | amdgpu_bo_kunmap(aobj); | |
d38ceaf9 AD |
694 | } else { |
695 | r = amdgpu_ib_get(ring, vm, 0, ib); | |
696 | if (r) { | |
697 | DRM_ERROR("Failed to get ib !\n"); | |
d38ceaf9 AD |
698 | return r; |
699 | } | |
700 | ||
701 | ib->gpu_addr = chunk_ib->va_start; | |
702 | } | |
d38ceaf9 | 703 | |
3ccec53c | 704 | ib->length_dw = chunk_ib->ib_bytes / 4; |
de807f81 | 705 | ib->flags = chunk_ib->flags; |
3cb485f3 | 706 | ib->ctx = parser->ctx; |
d38ceaf9 AD |
707 | j++; |
708 | } | |
709 | ||
710 | if (!parser->num_ibs) | |
711 | return 0; | |
712 | ||
713 | /* add GDS resources to first IB */ | |
714 | if (parser->bo_list) { | |
715 | struct amdgpu_bo *gds = parser->bo_list->gds_obj; | |
716 | struct amdgpu_bo *gws = parser->bo_list->gws_obj; | |
717 | struct amdgpu_bo *oa = parser->bo_list->oa_obj; | |
718 | struct amdgpu_ib *ib = &parser->ibs[0]; | |
719 | ||
720 | if (gds) { | |
721 | ib->gds_base = amdgpu_bo_gpu_offset(gds); | |
722 | ib->gds_size = amdgpu_bo_size(gds); | |
723 | } | |
724 | if (gws) { | |
725 | ib->gws_base = amdgpu_bo_gpu_offset(gws); | |
726 | ib->gws_size = amdgpu_bo_size(gws); | |
727 | } | |
728 | if (oa) { | |
729 | ib->oa_base = amdgpu_bo_gpu_offset(oa); | |
730 | ib->oa_size = amdgpu_bo_size(oa); | |
731 | } | |
732 | } | |
d38ceaf9 AD |
733 | /* wrap the last IB with user fence */ |
734 | if (parser->uf.bo) { | |
735 | struct amdgpu_ib *ib = &parser->ibs[parser->num_ibs - 1]; | |
736 | ||
737 | /* UVD & VCE fw doesn't support user fences */ | |
738 | if (ib->ring->type == AMDGPU_RING_TYPE_UVD || | |
739 | ib->ring->type == AMDGPU_RING_TYPE_VCE) | |
740 | return -EINVAL; | |
741 | ||
742 | ib->user = &parser->uf; | |
743 | } | |
744 | ||
745 | return 0; | |
746 | } | |
747 | ||
2b48d323 CK |
748 | static int amdgpu_cs_dependencies(struct amdgpu_device *adev, |
749 | struct amdgpu_cs_parser *p) | |
750 | { | |
76a1ea61 | 751 | struct amdgpu_fpriv *fpriv = p->filp->driver_priv; |
2b48d323 CK |
752 | struct amdgpu_ib *ib; |
753 | int i, j, r; | |
754 | ||
755 | if (!p->num_ibs) | |
756 | return 0; | |
757 | ||
758 | /* Add dependencies to first IB */ | |
759 | ib = &p->ibs[0]; | |
760 | for (i = 0; i < p->nchunks; ++i) { | |
761 | struct drm_amdgpu_cs_chunk_dep *deps; | |
762 | struct amdgpu_cs_chunk *chunk; | |
763 | unsigned num_deps; | |
764 | ||
765 | chunk = &p->chunks[i]; | |
766 | ||
767 | if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES) | |
768 | continue; | |
769 | ||
770 | deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata; | |
771 | num_deps = chunk->length_dw * 4 / | |
772 | sizeof(struct drm_amdgpu_cs_chunk_dep); | |
773 | ||
774 | for (j = 0; j < num_deps; ++j) { | |
2b48d323 | 775 | struct amdgpu_ring *ring; |
76a1ea61 | 776 | struct amdgpu_ctx *ctx; |
21c16bf6 | 777 | struct fence *fence; |
2b48d323 CK |
778 | |
779 | r = amdgpu_cs_get_ring(adev, deps[j].ip_type, | |
780 | deps[j].ip_instance, | |
781 | deps[j].ring, &ring); | |
782 | if (r) | |
783 | return r; | |
784 | ||
76a1ea61 CK |
785 | ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id); |
786 | if (ctx == NULL) | |
787 | return -EINVAL; | |
788 | ||
21c16bf6 CK |
789 | fence = amdgpu_ctx_get_fence(ctx, ring, |
790 | deps[j].handle); | |
791 | if (IS_ERR(fence)) { | |
792 | r = PTR_ERR(fence); | |
76a1ea61 | 793 | amdgpu_ctx_put(ctx); |
2b48d323 | 794 | return r; |
91e1a520 | 795 | |
21c16bf6 CK |
796 | } else if (fence) { |
797 | r = amdgpu_sync_fence(adev, &ib->sync, fence); | |
798 | fence_put(fence); | |
799 | amdgpu_ctx_put(ctx); | |
800 | if (r) | |
801 | return r; | |
802 | } | |
2b48d323 CK |
803 | } |
804 | } | |
805 | ||
806 | return 0; | |
807 | } | |
808 | ||
4c7eb91c | 809 | static int amdgpu_cs_free_job(struct amdgpu_job *job) |
bb977d37 CZ |
810 | { |
811 | int i; | |
4c7eb91c JZ |
812 | if (job->ibs) |
813 | for (i = 0; i < job->num_ibs; i++) | |
814 | amdgpu_ib_free(job->adev, &job->ibs[i]); | |
815 | kfree(job->ibs); | |
816 | if (job->uf.bo) | |
817 | drm_gem_object_unreference_unlocked(&job->uf.bo->gem_base); | |
bb977d37 CZ |
818 | return 0; |
819 | } | |
820 | ||
049fc527 CZ |
821 | int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) |
822 | { | |
823 | struct amdgpu_device *adev = dev->dev_private; | |
824 | union drm_amdgpu_cs *cs = data; | |
f48b2659 CZ |
825 | struct amdgpu_fpriv *fpriv = filp->driver_priv; |
826 | struct amdgpu_vm *vm = &fpriv->vm; | |
049fc527 | 827 | struct amdgpu_cs_parser *parser; |
26a6980c CK |
828 | bool reserved_buffers = false; |
829 | int i, r; | |
049fc527 | 830 | |
0c418f10 | 831 | if (!adev->accel_working) |
049fc527 | 832 | return -EBUSY; |
2b48d323 | 833 | |
049fc527 CZ |
834 | parser = amdgpu_cs_parser_create(adev, filp, NULL, NULL, 0); |
835 | if (!parser) | |
836 | return -ENOMEM; | |
837 | r = amdgpu_cs_parser_init(parser, data); | |
d38ceaf9 | 838 | if (r) { |
049fc527 | 839 | DRM_ERROR("Failed to initialize parser !\n"); |
0c418f10 | 840 | amdgpu_cs_parser_fini(parser, r, false); |
d38ceaf9 AD |
841 | r = amdgpu_cs_handle_lockup(adev, r); |
842 | return r; | |
843 | } | |
f48b2659 | 844 | mutex_lock(&vm->mutex); |
26a6980c CK |
845 | r = amdgpu_cs_parser_relocs(parser); |
846 | if (r == -ENOMEM) | |
847 | DRM_ERROR("Not enough memory for command submission!\n"); | |
848 | else if (r && r != -ERESTARTSYS) | |
849 | DRM_ERROR("Failed to process the buffer list %d!\n", r); | |
850 | else if (!r) { | |
851 | reserved_buffers = true; | |
852 | r = amdgpu_cs_ib_fill(adev, parser); | |
853 | } | |
854 | ||
855 | if (!r) { | |
856 | r = amdgpu_cs_dependencies(adev, parser); | |
857 | if (r) | |
858 | DRM_ERROR("Failed in the dependencies handling %d!\n", r); | |
859 | } | |
860 | ||
861 | if (r) | |
862 | goto out; | |
863 | ||
864 | for (i = 0; i < parser->num_ibs; i++) | |
865 | trace_amdgpu_cs(parser, i); | |
866 | ||
867 | r = amdgpu_cs_ib_vm_chunk(adev, parser); | |
4fe63117 CZ |
868 | if (r) |
869 | goto out; | |
870 | ||
049fc527 | 871 | if (amdgpu_enable_scheduler && parser->num_ibs) { |
bb977d37 | 872 | struct amdgpu_job *job; |
3c4adead | 873 | struct amdgpu_ring * ring = parser->ibs->ring; |
bb977d37 | 874 | job = kzalloc(sizeof(struct amdgpu_job), GFP_KERNEL); |
4cfdcd9c DC |
875 | if (!job) { |
876 | r = -ENOMEM; | |
877 | goto out; | |
878 | } | |
4f839a24 | 879 | job->base.sched = &ring->sched; |
bb977d37 CZ |
880 | job->base.s_entity = &parser->ctx->rings[ring->idx].entity; |
881 | job->adev = parser->adev; | |
882 | job->ibs = parser->ibs; | |
883 | job->num_ibs = parser->num_ibs; | |
84f76ea6 | 884 | job->base.owner = parser->filp; |
bb977d37 CZ |
885 | mutex_init(&job->job_lock); |
886 | if (job->ibs[job->num_ibs - 1].user) { | |
e4a58a28 | 887 | job->uf = parser->uf; |
bb977d37 | 888 | job->ibs[job->num_ibs - 1].user = &job->uf; |
e4a58a28 | 889 | parser->uf.bo = NULL; |
bb977d37 CZ |
890 | } |
891 | ||
e4a58a28 CK |
892 | parser->ibs = NULL; |
893 | parser->num_ibs = 0; | |
894 | ||
bb977d37 CZ |
895 | job->free_job = amdgpu_cs_free_job; |
896 | mutex_lock(&job->job_lock); | |
a6db8a33 | 897 | r = amd_sched_entity_push_job(&job->base); |
f556cb0c | 898 | if (r) { |
bb977d37 CZ |
899 | mutex_unlock(&job->job_lock); |
900 | amdgpu_cs_free_job(job); | |
901 | kfree(job); | |
f556cb0c CZ |
902 | goto out; |
903 | } | |
ce882e6d | 904 | cs->out.handle = |
3a185a33 | 905 | amdgpu_ctx_add_fence(parser->ctx, ring, |
ce882e6d | 906 | &job->base.s_fence->base); |
e4a58a28 | 907 | job->ibs[job->num_ibs - 1].sequence = cs->out.handle; |
eb98d1c5 | 908 | |
c3b95d4f CZ |
909 | list_sort(NULL, &parser->validated, cmp_size_smaller_first); |
910 | ttm_eu_fence_buffer_objects(&parser->ticket, | |
911 | &parser->validated, | |
bb977d37 | 912 | &job->base.s_fence->base); |
c3b95d4f | 913 | |
bb977d37 CZ |
914 | mutex_unlock(&job->job_lock); |
915 | amdgpu_cs_parser_fini_late(parser); | |
f48b2659 | 916 | mutex_unlock(&vm->mutex); |
049fc527 | 917 | return 0; |
d38ceaf9 AD |
918 | } |
919 | ||
049fc527 | 920 | cs->out.handle = parser->ibs[parser->num_ibs - 1].sequence; |
d38ceaf9 | 921 | out: |
26a6980c | 922 | amdgpu_cs_parser_fini(parser, r, reserved_buffers); |
f48b2659 | 923 | mutex_unlock(&vm->mutex); |
d38ceaf9 AD |
924 | r = amdgpu_cs_handle_lockup(adev, r); |
925 | return r; | |
926 | } | |
927 | ||
928 | /** | |
929 | * amdgpu_cs_wait_ioctl - wait for a command submission to finish | |
930 | * | |
931 | * @dev: drm device | |
932 | * @data: data from userspace | |
933 | * @filp: file private | |
934 | * | |
935 | * Wait for the command submission identified by handle to finish. | |
936 | */ | |
937 | int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data, | |
938 | struct drm_file *filp) | |
939 | { | |
940 | union drm_amdgpu_wait_cs *wait = data; | |
941 | struct amdgpu_device *adev = dev->dev_private; | |
d38ceaf9 | 942 | unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout); |
03507c4f | 943 | struct amdgpu_ring *ring = NULL; |
66b3cf2a | 944 | struct amdgpu_ctx *ctx; |
21c16bf6 | 945 | struct fence *fence; |
d38ceaf9 AD |
946 | long r; |
947 | ||
21c16bf6 CK |
948 | r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance, |
949 | wait->in.ring, &ring); | |
950 | if (r) | |
951 | return r; | |
952 | ||
66b3cf2a JZ |
953 | ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id); |
954 | if (ctx == NULL) | |
955 | return -EINVAL; | |
d38ceaf9 | 956 | |
4b559c90 CZ |
957 | fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle); |
958 | if (IS_ERR(fence)) | |
959 | r = PTR_ERR(fence); | |
960 | else if (fence) { | |
961 | r = fence_wait_timeout(fence, true, timeout); | |
962 | fence_put(fence); | |
963 | } else | |
964 | r = 1; | |
049fc527 | 965 | |
66b3cf2a | 966 | amdgpu_ctx_put(ctx); |
d38ceaf9 AD |
967 | if (r < 0) |
968 | return r; | |
969 | ||
970 | memset(wait, 0, sizeof(*wait)); | |
971 | wait->out.status = (r == 0); | |
972 | ||
973 | return 0; | |
974 | } | |
975 | ||
976 | /** | |
977 | * amdgpu_cs_find_bo_va - find bo_va for VM address | |
978 | * | |
979 | * @parser: command submission parser context | |
980 | * @addr: VM address | |
981 | * @bo: resulting BO of the mapping found | |
982 | * | |
983 | * Search the buffer objects in the command submission context for a certain | |
984 | * virtual memory address. Returns allocation structure when found, NULL | |
985 | * otherwise. | |
986 | */ | |
987 | struct amdgpu_bo_va_mapping * | |
988 | amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser, | |
989 | uint64_t addr, struct amdgpu_bo **bo) | |
990 | { | |
991 | struct amdgpu_bo_list_entry *reloc; | |
992 | struct amdgpu_bo_va_mapping *mapping; | |
993 | ||
994 | addr /= AMDGPU_GPU_PAGE_SIZE; | |
995 | ||
996 | list_for_each_entry(reloc, &parser->validated, tv.head) { | |
997 | if (!reloc->bo_va) | |
998 | continue; | |
999 | ||
7fc11959 CK |
1000 | list_for_each_entry(mapping, &reloc->bo_va->valids, list) { |
1001 | if (mapping->it.start > addr || | |
1002 | addr > mapping->it.last) | |
1003 | continue; | |
1004 | ||
1005 | *bo = reloc->bo_va->bo; | |
1006 | return mapping; | |
1007 | } | |
1008 | ||
1009 | list_for_each_entry(mapping, &reloc->bo_va->invalids, list) { | |
d38ceaf9 AD |
1010 | if (mapping->it.start > addr || |
1011 | addr > mapping->it.last) | |
1012 | continue; | |
1013 | ||
1014 | *bo = reloc->bo_va->bo; | |
1015 | return mapping; | |
1016 | } | |
1017 | } | |
1018 | ||
1019 | return NULL; | |
1020 | } |