]>
Commit | Line | Data |
---|---|---|
cd5351f4 | 1 | /* |
8bb0daff | 2 | * drivers/gpu/drm/omapdrm/omap_gem.c |
cd5351f4 RC |
3 | * |
4 | * Copyright (C) 2011 Texas Instruments | |
5 | * Author: Rob Clark <rob.clark@linaro.org> | |
6 | * | |
7 | * This program is free software; you can redistribute it and/or modify it | |
8 | * under the terms of the GNU General Public License version 2 as published by | |
9 | * the Free Software Foundation. | |
10 | * | |
11 | * This program is distributed in the hope that it will be useful, but WITHOUT | |
12 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for | |
14 | * more details. | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License along with | |
17 | * this program. If not, see <http://www.gnu.org/licenses/>. | |
18 | */ | |
19 | ||
cd5351f4 | 20 | #include <linux/shmem_fs.h> |
2d278f54 LP |
21 | #include <linux/spinlock.h> |
22 | ||
0de23977 | 23 | #include <drm/drm_vma_manager.h> |
cd5351f4 RC |
24 | |
25 | #include "omap_drv.h" | |
f7f9f453 | 26 | #include "omap_dmm_tiler.h" |
cd5351f4 RC |
27 | |
28 | /* remove these once drm core helpers are merged */ | |
801d5bc6 | 29 | struct page **_drm_gem_get_pages(struct drm_gem_object *obj, gfp_t gfpmask); |
cd5351f4 RC |
30 | void _drm_gem_put_pages(struct drm_gem_object *obj, struct page **pages, |
31 | bool dirty, bool accessed); | |
f7f9f453 | 32 | int _drm_gem_create_mmap_offset_size(struct drm_gem_object *obj, size_t size); |
cd5351f4 RC |
33 | |
34 | /* | |
35 | * GEM buffer object implementation. | |
36 | */ | |
37 | ||
38 | #define to_omap_bo(x) container_of(x, struct omap_gem_object, base) | |
39 | ||
40 | /* note: we use upper 8 bits of flags for driver-internal flags: */ | |
41 | #define OMAP_BO_DMA 0x01000000 /* actually is physically contiguous */ | |
42 | #define OMAP_BO_EXT_SYNC 0x02000000 /* externally allocated sync object */ | |
43 | #define OMAP_BO_EXT_MEM 0x04000000 /* externally allocated memory */ | |
44 | ||
45 | ||
46 | struct omap_gem_object { | |
47 | struct drm_gem_object base; | |
48 | ||
f6b6036e RC |
49 | struct list_head mm_list; |
50 | ||
cd5351f4 RC |
51 | uint32_t flags; |
52 | ||
f7f9f453 RC |
53 | /** width/height for tiled formats (rounded up to slot boundaries) */ |
54 | uint16_t width, height; | |
55 | ||
a6a91827 RC |
56 | /** roll applied when mapping to DMM */ |
57 | uint32_t roll; | |
58 | ||
cd5351f4 RC |
59 | /** |
60 | * If buffer is allocated physically contiguous, the OMAP_BO_DMA flag | |
f7f9f453 RC |
61 | * is set and the paddr is valid. Also if the buffer is remapped in |
62 | * TILER and paddr_cnt > 0, then paddr is valid. But if you are using | |
63 | * the physical address and OMAP_BO_DMA is not set, then you should | |
64 | * be going thru omap_gem_{get,put}_paddr() to ensure the mapping is | |
65 | * not removed from under your feet. | |
cd5351f4 RC |
66 | * |
67 | * Note that OMAP_BO_SCANOUT is a hint from userspace that DMA capable | |
68 | * buffer is requested, but doesn't mean that it is. Use the | |
69 | * OMAP_BO_DMA flag to determine if the buffer has a DMA capable | |
70 | * physical address. | |
71 | */ | |
72 | dma_addr_t paddr; | |
73 | ||
f7f9f453 RC |
74 | /** |
75 | * # of users of paddr | |
76 | */ | |
77 | uint32_t paddr_cnt; | |
78 | ||
79 | /** | |
80 | * tiler block used when buffer is remapped in DMM/TILER. | |
81 | */ | |
82 | struct tiler_block *block; | |
83 | ||
cd5351f4 RC |
84 | /** |
85 | * Array of backing pages, if allocated. Note that pages are never | |
86 | * allocated for buffers originally allocated from contiguous memory | |
87 | */ | |
88 | struct page **pages; | |
89 | ||
f3bc9d24 RC |
90 | /** addresses corresponding to pages in above array */ |
91 | dma_addr_t *addrs; | |
92 | ||
cd5351f4 RC |
93 | /** |
94 | * Virtual address, if mapped. | |
95 | */ | |
96 | void *vaddr; | |
97 | ||
98 | /** | |
99 | * sync-object allocated on demand (if needed) | |
100 | * | |
101 | * Per-buffer sync-object for tracking pending and completed hw/dma | |
102 | * read and write operations. The layout in memory is dictated by | |
103 | * the SGX firmware, which uses this information to stall the command | |
104 | * stream if a surface is not ready yet. | |
105 | * | |
106 | * Note that when buffer is used by SGX, the sync-object needs to be | |
107 | * allocated from a special heap of sync-objects. This way many sync | |
108 | * objects can be packed in a page, and not waste GPU virtual address | |
109 | * space. Because of this we have to have a omap_gem_set_sync_object() | |
110 | * API to allow replacement of the syncobj after it has (potentially) | |
111 | * already been allocated. A bit ugly but I haven't thought of a | |
112 | * better alternative. | |
113 | */ | |
114 | struct { | |
115 | uint32_t write_pending; | |
116 | uint32_t write_complete; | |
117 | uint32_t read_pending; | |
118 | uint32_t read_complete; | |
119 | } *sync; | |
120 | }; | |
121 | ||
c5b1247b RC |
122 | static int get_pages(struct drm_gem_object *obj, struct page ***pages); |
123 | static uint64_t mmap_offset(struct drm_gem_object *obj); | |
124 | ||
f7f9f453 RC |
125 | /* To deal with userspace mmap'ings of 2d tiled buffers, which (a) are |
126 | * not necessarily pinned in TILER all the time, and (b) when they are | |
127 | * they are not necessarily page aligned, we reserve one or more small | |
128 | * regions in each of the 2d containers to use as a user-GART where we | |
129 | * can create a second page-aligned mapping of parts of the buffer | |
130 | * being accessed from userspace. | |
131 | * | |
132 | * Note that we could optimize slightly when we know that multiple | |
133 | * tiler containers are backed by the same PAT.. but I'll leave that | |
134 | * for later.. | |
135 | */ | |
136 | #define NUM_USERGART_ENTRIES 2 | |
137 | struct usergart_entry { | |
138 | struct tiler_block *block; /* the reserved tiler block */ | |
139 | dma_addr_t paddr; | |
140 | struct drm_gem_object *obj; /* the current pinned obj */ | |
141 | pgoff_t obj_pgoff; /* page offset of obj currently | |
142 | mapped in */ | |
143 | }; | |
144 | static struct { | |
145 | struct usergart_entry entry[NUM_USERGART_ENTRIES]; | |
146 | int height; /* height in rows */ | |
147 | int height_shift; /* ilog2(height in rows) */ | |
148 | int slot_shift; /* ilog2(width per slot) */ | |
149 | int stride_pfn; /* stride in pages */ | |
150 | int last; /* index of last used entry */ | |
151 | } *usergart; | |
152 | ||
153 | static void evict_entry(struct drm_gem_object *obj, | |
154 | enum tiler_fmt fmt, struct usergart_entry *entry) | |
155 | { | |
6796cb16 DH |
156 | struct omap_gem_object *omap_obj = to_omap_bo(obj); |
157 | int n = usergart[fmt].height; | |
158 | size_t size = PAGE_SIZE * n; | |
159 | loff_t off = mmap_offset(obj) + | |
160 | (entry->obj_pgoff << PAGE_SHIFT); | |
161 | const int m = 1 + ((omap_obj->width << fmt) / PAGE_SIZE); | |
162 | ||
163 | if (m > 1) { | |
164 | int i; | |
165 | /* if stride > than PAGE_SIZE then sparse mapping: */ | |
166 | for (i = n; i > 0; i--) { | |
167 | unmap_mapping_range(obj->dev->anon_inode->i_mapping, | |
168 | off, PAGE_SIZE, 1); | |
169 | off += PAGE_SIZE * m; | |
e559895a | 170 | } |
6796cb16 DH |
171 | } else { |
172 | unmap_mapping_range(obj->dev->anon_inode->i_mapping, | |
173 | off, size, 1); | |
f7f9f453 RC |
174 | } |
175 | ||
176 | entry->obj = NULL; | |
177 | } | |
178 | ||
179 | /* Evict a buffer from usergart, if it is mapped there */ | |
180 | static void evict(struct drm_gem_object *obj) | |
181 | { | |
182 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
183 | ||
184 | if (omap_obj->flags & OMAP_BO_TILED) { | |
185 | enum tiler_fmt fmt = gem2fmt(omap_obj->flags); | |
186 | int i; | |
187 | ||
188 | if (!usergart) | |
189 | return; | |
190 | ||
191 | for (i = 0; i < NUM_USERGART_ENTRIES; i++) { | |
192 | struct usergart_entry *entry = &usergart[fmt].entry[i]; | |
193 | if (entry->obj == obj) | |
194 | evict_entry(obj, fmt, entry); | |
195 | } | |
196 | } | |
197 | } | |
198 | ||
cd5351f4 RC |
199 | /* GEM objects can either be allocated from contiguous memory (in which |
200 | * case obj->filp==NULL), or w/ shmem backing (obj->filp!=NULL). But non | |
201 | * contiguous buffers can be remapped in TILER/DMM if they need to be | |
202 | * contiguous... but we don't do this all the time to reduce pressure | |
203 | * on TILER/DMM space when we know at allocation time that the buffer | |
204 | * will need to be scanned out. | |
205 | */ | |
206 | static inline bool is_shmem(struct drm_gem_object *obj) | |
207 | { | |
208 | return obj->filp != NULL; | |
209 | } | |
210 | ||
8b6b569e RC |
211 | /** |
212 | * shmem buffers that are mapped cached can simulate coherency via using | |
213 | * page faulting to keep track of dirty pages | |
214 | */ | |
215 | static inline bool is_cached_coherent(struct drm_gem_object *obj) | |
216 | { | |
217 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
218 | return is_shmem(obj) && | |
219 | ((omap_obj->flags & OMAP_BO_CACHE_MASK) == OMAP_BO_CACHED); | |
220 | } | |
221 | ||
cd5351f4 RC |
222 | static DEFINE_SPINLOCK(sync_lock); |
223 | ||
224 | /** ensure backing pages are allocated */ | |
225 | static int omap_gem_attach_pages(struct drm_gem_object *obj) | |
226 | { | |
8b6b569e | 227 | struct drm_device *dev = obj->dev; |
cd5351f4 RC |
228 | struct omap_gem_object *omap_obj = to_omap_bo(obj); |
229 | struct page **pages; | |
d4eb23a9 EG |
230 | int npages = obj->size >> PAGE_SHIFT; |
231 | int i, ret; | |
8b6b569e | 232 | dma_addr_t *addrs; |
cd5351f4 RC |
233 | |
234 | WARN_ON(omap_obj->pages); | |
235 | ||
0cdbe8ac | 236 | pages = drm_gem_get_pages(obj); |
cd5351f4 RC |
237 | if (IS_ERR(pages)) { |
238 | dev_err(obj->dev->dev, "could not get pages: %ld\n", PTR_ERR(pages)); | |
239 | return PTR_ERR(pages); | |
240 | } | |
241 | ||
f3bc9d24 RC |
242 | /* for non-cached buffers, ensure the new pages are clean because |
243 | * DSS, GPU, etc. are not cache coherent: | |
244 | */ | |
245 | if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) { | |
23d84ed9 | 246 | addrs = kmalloc(npages * sizeof(*addrs), GFP_KERNEL); |
d4eb23a9 EG |
247 | if (!addrs) { |
248 | ret = -ENOMEM; | |
249 | goto free_pages; | |
250 | } | |
251 | ||
f3bc9d24 | 252 | for (i = 0; i < npages; i++) { |
8b6b569e | 253 | addrs[i] = dma_map_page(dev->dev, pages[i], |
f3bc9d24 RC |
254 | 0, PAGE_SIZE, DMA_BIDIRECTIONAL); |
255 | } | |
8b6b569e | 256 | } else { |
23d84ed9 | 257 | addrs = kzalloc(npages * sizeof(*addrs), GFP_KERNEL); |
d4eb23a9 EG |
258 | if (!addrs) { |
259 | ret = -ENOMEM; | |
260 | goto free_pages; | |
261 | } | |
f3bc9d24 RC |
262 | } |
263 | ||
8b6b569e | 264 | omap_obj->addrs = addrs; |
cd5351f4 | 265 | omap_obj->pages = pages; |
8b6b569e | 266 | |
cd5351f4 | 267 | return 0; |
d4eb23a9 EG |
268 | |
269 | free_pages: | |
ddcd09d6 | 270 | drm_gem_put_pages(obj, pages, true, false); |
d4eb23a9 EG |
271 | |
272 | return ret; | |
cd5351f4 RC |
273 | } |
274 | ||
275 | /** release backing pages */ | |
276 | static void omap_gem_detach_pages(struct drm_gem_object *obj) | |
277 | { | |
278 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
f3bc9d24 RC |
279 | |
280 | /* for non-cached buffers, ensure the new pages are clean because | |
281 | * DSS, GPU, etc. are not cache coherent: | |
282 | */ | |
283 | if (omap_obj->flags & (OMAP_BO_WC|OMAP_BO_UNCACHED)) { | |
284 | int i, npages = obj->size >> PAGE_SHIFT; | |
285 | for (i = 0; i < npages; i++) { | |
286 | dma_unmap_page(obj->dev->dev, omap_obj->addrs[i], | |
287 | PAGE_SIZE, DMA_BIDIRECTIONAL); | |
288 | } | |
f3bc9d24 RC |
289 | } |
290 | ||
8b6b569e RC |
291 | kfree(omap_obj->addrs); |
292 | omap_obj->addrs = NULL; | |
293 | ||
ddcd09d6 | 294 | drm_gem_put_pages(obj, omap_obj->pages, true, false); |
cd5351f4 RC |
295 | omap_obj->pages = NULL; |
296 | } | |
297 | ||
6ad11bc3 RC |
298 | /* get buffer flags */ |
299 | uint32_t omap_gem_flags(struct drm_gem_object *obj) | |
300 | { | |
301 | return to_omap_bo(obj)->flags; | |
302 | } | |
303 | ||
cd5351f4 | 304 | /** get mmap offset */ |
c5b1247b | 305 | static uint64_t mmap_offset(struct drm_gem_object *obj) |
cd5351f4 | 306 | { |
f6b6036e | 307 | struct drm_device *dev = obj->dev; |
0de23977 DH |
308 | int ret; |
309 | size_t size; | |
f6b6036e RC |
310 | |
311 | WARN_ON(!mutex_is_locked(&dev->struct_mutex)); | |
312 | ||
0de23977 DH |
313 | /* Make it mmapable */ |
314 | size = omap_gem_mmap_size(obj); | |
ddcd09d6 | 315 | ret = drm_gem_create_mmap_offset_size(obj, size); |
0de23977 DH |
316 | if (ret) { |
317 | dev_err(dev->dev, "could not allocate mmap offset\n"); | |
318 | return 0; | |
cd5351f4 RC |
319 | } |
320 | ||
0de23977 | 321 | return drm_vma_node_offset_addr(&obj->vma_node); |
cd5351f4 RC |
322 | } |
323 | ||
c5b1247b RC |
324 | uint64_t omap_gem_mmap_offset(struct drm_gem_object *obj) |
325 | { | |
326 | uint64_t offset; | |
327 | mutex_lock(&obj->dev->struct_mutex); | |
328 | offset = mmap_offset(obj); | |
329 | mutex_unlock(&obj->dev->struct_mutex); | |
330 | return offset; | |
331 | } | |
332 | ||
f7f9f453 RC |
333 | /** get mmap size */ |
334 | size_t omap_gem_mmap_size(struct drm_gem_object *obj) | |
335 | { | |
336 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
337 | size_t size = obj->size; | |
338 | ||
339 | if (omap_obj->flags & OMAP_BO_TILED) { | |
340 | /* for tiled buffers, the virtual size has stride rounded up | |
341 | * to 4kb.. (to hide the fact that row n+1 might start 16kb or | |
342 | * 32kb later!). But we don't back the entire buffer with | |
343 | * pages, only the valid picture part.. so need to adjust for | |
344 | * this in the size used to mmap and generate mmap offset | |
345 | */ | |
346 | size = tiler_vsize(gem2fmt(omap_obj->flags), | |
347 | omap_obj->width, omap_obj->height); | |
348 | } | |
349 | ||
350 | return size; | |
351 | } | |
352 | ||
3c810c61 RC |
353 | /* get tiled size, returns -EINVAL if not tiled buffer */ |
354 | int omap_gem_tiled_size(struct drm_gem_object *obj, uint16_t *w, uint16_t *h) | |
355 | { | |
356 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
357 | if (omap_obj->flags & OMAP_BO_TILED) { | |
358 | *w = omap_obj->width; | |
359 | *h = omap_obj->height; | |
360 | return 0; | |
361 | } | |
362 | return -EINVAL; | |
363 | } | |
f7f9f453 RC |
364 | |
365 | /* Normal handling for the case of faulting in non-tiled buffers */ | |
366 | static int fault_1d(struct drm_gem_object *obj, | |
367 | struct vm_area_struct *vma, struct vm_fault *vmf) | |
368 | { | |
369 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
370 | unsigned long pfn; | |
371 | pgoff_t pgoff; | |
372 | ||
373 | /* We don't use vmf->pgoff since that has the fake offset: */ | |
374 | pgoff = ((unsigned long)vmf->virtual_address - | |
375 | vma->vm_start) >> PAGE_SHIFT; | |
376 | ||
377 | if (omap_obj->pages) { | |
8b6b569e | 378 | omap_gem_cpu_sync(obj, pgoff); |
f7f9f453 RC |
379 | pfn = page_to_pfn(omap_obj->pages[pgoff]); |
380 | } else { | |
381 | BUG_ON(!(omap_obj->flags & OMAP_BO_DMA)); | |
382 | pfn = (omap_obj->paddr >> PAGE_SHIFT) + pgoff; | |
383 | } | |
384 | ||
385 | VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address, | |
386 | pfn, pfn << PAGE_SHIFT); | |
387 | ||
388 | return vm_insert_mixed(vma, (unsigned long)vmf->virtual_address, pfn); | |
389 | } | |
390 | ||
391 | /* Special handling for the case of faulting in 2d tiled buffers */ | |
392 | static int fault_2d(struct drm_gem_object *obj, | |
393 | struct vm_area_struct *vma, struct vm_fault *vmf) | |
394 | { | |
395 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
396 | struct usergart_entry *entry; | |
397 | enum tiler_fmt fmt = gem2fmt(omap_obj->flags); | |
398 | struct page *pages[64]; /* XXX is this too much to have on stack? */ | |
399 | unsigned long pfn; | |
400 | pgoff_t pgoff, base_pgoff; | |
401 | void __user *vaddr; | |
402 | int i, ret, slots; | |
403 | ||
e559895a RC |
404 | /* |
405 | * Note the height of the slot is also equal to the number of pages | |
406 | * that need to be mapped in to fill 4kb wide CPU page. If the slot | |
407 | * height is 64, then 64 pages fill a 4kb wide by 64 row region. | |
408 | */ | |
409 | const int n = usergart[fmt].height; | |
410 | const int n_shift = usergart[fmt].height_shift; | |
411 | ||
412 | /* | |
413 | * If buffer width in bytes > PAGE_SIZE then the virtual stride is | |
414 | * rounded up to next multiple of PAGE_SIZE.. this need to be taken | |
415 | * into account in some of the math, so figure out virtual stride | |
416 | * in pages | |
f7f9f453 | 417 | */ |
e559895a | 418 | const int m = 1 + ((omap_obj->width << fmt) / PAGE_SIZE); |
f7f9f453 RC |
419 | |
420 | /* We don't use vmf->pgoff since that has the fake offset: */ | |
421 | pgoff = ((unsigned long)vmf->virtual_address - | |
422 | vma->vm_start) >> PAGE_SHIFT; | |
423 | ||
e559895a RC |
424 | /* |
425 | * Actual address we start mapping at is rounded down to previous slot | |
f7f9f453 RC |
426 | * boundary in the y direction: |
427 | */ | |
e559895a | 428 | base_pgoff = round_down(pgoff, m << n_shift); |
f7f9f453 | 429 | |
e559895a | 430 | /* figure out buffer width in slots */ |
f7f9f453 RC |
431 | slots = omap_obj->width >> usergart[fmt].slot_shift; |
432 | ||
e559895a RC |
433 | vaddr = vmf->virtual_address - ((pgoff - base_pgoff) << PAGE_SHIFT); |
434 | ||
435 | entry = &usergart[fmt].entry[usergart[fmt].last]; | |
436 | ||
f7f9f453 RC |
437 | /* evict previous buffer using this usergart entry, if any: */ |
438 | if (entry->obj) | |
439 | evict_entry(entry->obj, fmt, entry); | |
440 | ||
441 | entry->obj = obj; | |
442 | entry->obj_pgoff = base_pgoff; | |
443 | ||
e559895a RC |
444 | /* now convert base_pgoff to phys offset from virt offset: */ |
445 | base_pgoff = (base_pgoff >> n_shift) * slots; | |
446 | ||
447 | /* for wider-than 4k.. figure out which part of the slot-row we want: */ | |
448 | if (m > 1) { | |
449 | int off = pgoff % m; | |
450 | entry->obj_pgoff += off; | |
451 | base_pgoff /= m; | |
452 | slots = min(slots - (off << n_shift), n); | |
453 | base_pgoff += off << n_shift; | |
454 | vaddr += off << PAGE_SHIFT; | |
455 | } | |
456 | ||
457 | /* | |
458 | * Map in pages. Beyond the valid pixel part of the buffer, we set | |
459 | * pages[i] to NULL to get a dummy page mapped in.. if someone | |
460 | * reads/writes it they will get random/undefined content, but at | |
461 | * least it won't be corrupting whatever other random page used to | |
462 | * be mapped in, or other undefined behavior. | |
f7f9f453 RC |
463 | */ |
464 | memcpy(pages, &omap_obj->pages[base_pgoff], | |
465 | sizeof(struct page *) * slots); | |
466 | memset(pages + slots, 0, | |
e559895a | 467 | sizeof(struct page *) * (n - slots)); |
f7f9f453 | 468 | |
a6a91827 | 469 | ret = tiler_pin(entry->block, pages, ARRAY_SIZE(pages), 0, true); |
f7f9f453 RC |
470 | if (ret) { |
471 | dev_err(obj->dev->dev, "failed to pin: %d\n", ret); | |
472 | return ret; | |
473 | } | |
474 | ||
f7f9f453 RC |
475 | pfn = entry->paddr >> PAGE_SHIFT; |
476 | ||
477 | VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address, | |
478 | pfn, pfn << PAGE_SHIFT); | |
479 | ||
e559895a | 480 | for (i = n; i > 0; i--) { |
f7f9f453 RC |
481 | vm_insert_mixed(vma, (unsigned long)vaddr, pfn); |
482 | pfn += usergart[fmt].stride_pfn; | |
e559895a | 483 | vaddr += PAGE_SIZE * m; |
f7f9f453 RC |
484 | } |
485 | ||
486 | /* simple round-robin: */ | |
487 | usergart[fmt].last = (usergart[fmt].last + 1) % NUM_USERGART_ENTRIES; | |
488 | ||
489 | return 0; | |
490 | } | |
491 | ||
cd5351f4 RC |
492 | /** |
493 | * omap_gem_fault - pagefault handler for GEM objects | |
494 | * @vma: the VMA of the GEM object | |
495 | * @vmf: fault detail | |
496 | * | |
497 | * Invoked when a fault occurs on an mmap of a GEM managed area. GEM | |
498 | * does most of the work for us including the actual map/unmap calls | |
499 | * but we need to do the actual page work. | |
500 | * | |
501 | * The VMA was set up by GEM. In doing so it also ensured that the | |
502 | * vma->vm_private_data points to the GEM object that is backing this | |
503 | * mapping. | |
504 | */ | |
505 | int omap_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf) | |
506 | { | |
507 | struct drm_gem_object *obj = vma->vm_private_data; | |
508 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
509 | struct drm_device *dev = obj->dev; | |
510 | struct page **pages; | |
cd5351f4 RC |
511 | int ret; |
512 | ||
513 | /* Make sure we don't parallel update on a fault, nor move or remove | |
514 | * something from beneath our feet | |
515 | */ | |
516 | mutex_lock(&dev->struct_mutex); | |
517 | ||
518 | /* if a shmem backed object, make sure we have pages attached now */ | |
519 | ret = get_pages(obj, &pages); | |
ae053039 | 520 | if (ret) |
cd5351f4 | 521 | goto fail; |
cd5351f4 RC |
522 | |
523 | /* where should we do corresponding put_pages().. we are mapping | |
524 | * the original page, rather than thru a GART, so we can't rely | |
525 | * on eviction to trigger this. But munmap() or all mappings should | |
526 | * probably trigger put_pages()? | |
527 | */ | |
528 | ||
f7f9f453 RC |
529 | if (omap_obj->flags & OMAP_BO_TILED) |
530 | ret = fault_2d(obj, vma, vmf); | |
531 | else | |
532 | ret = fault_1d(obj, vma, vmf); | |
cd5351f4 | 533 | |
cd5351f4 RC |
534 | |
535 | fail: | |
536 | mutex_unlock(&dev->struct_mutex); | |
537 | switch (ret) { | |
538 | case 0: | |
539 | case -ERESTARTSYS: | |
540 | case -EINTR: | |
541 | return VM_FAULT_NOPAGE; | |
542 | case -ENOMEM: | |
543 | return VM_FAULT_OOM; | |
544 | default: | |
545 | return VM_FAULT_SIGBUS; | |
546 | } | |
547 | } | |
548 | ||
549 | /** We override mainly to fix up some of the vm mapping flags.. */ | |
550 | int omap_gem_mmap(struct file *filp, struct vm_area_struct *vma) | |
551 | { | |
cd5351f4 RC |
552 | int ret; |
553 | ||
554 | ret = drm_gem_mmap(filp, vma); | |
555 | if (ret) { | |
556 | DBG("mmap failed: %d", ret); | |
557 | return ret; | |
558 | } | |
559 | ||
8b6b569e RC |
560 | return omap_gem_mmap_obj(vma->vm_private_data, vma); |
561 | } | |
562 | ||
563 | int omap_gem_mmap_obj(struct drm_gem_object *obj, | |
564 | struct vm_area_struct *vma) | |
565 | { | |
566 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
cd5351f4 RC |
567 | |
568 | vma->vm_flags &= ~VM_PFNMAP; | |
569 | vma->vm_flags |= VM_MIXEDMAP; | |
570 | ||
571 | if (omap_obj->flags & OMAP_BO_WC) { | |
572 | vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags)); | |
573 | } else if (omap_obj->flags & OMAP_BO_UNCACHED) { | |
574 | vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags)); | |
575 | } else { | |
8b6b569e RC |
576 | /* |
577 | * We do have some private objects, at least for scanout buffers | |
578 | * on hardware without DMM/TILER. But these are allocated write- | |
579 | * combine | |
580 | */ | |
581 | if (WARN_ON(!obj->filp)) | |
582 | return -EINVAL; | |
583 | ||
584 | /* | |
585 | * Shunt off cached objs to shmem file so they have their own | |
586 | * address_space (so unmap_mapping_range does what we want, | |
587 | * in particular in the case of mmap'd dmabufs) | |
588 | */ | |
589 | fput(vma->vm_file); | |
8b6b569e | 590 | vma->vm_pgoff = 0; |
cb0942b8 | 591 | vma->vm_file = get_file(obj->filp); |
8b6b569e | 592 | |
cd5351f4 RC |
593 | vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); |
594 | } | |
595 | ||
8b6b569e | 596 | return 0; |
cd5351f4 RC |
597 | } |
598 | ||
8b6b569e | 599 | |
cd5351f4 RC |
600 | /** |
601 | * omap_gem_dumb_create - create a dumb buffer | |
602 | * @drm_file: our client file | |
603 | * @dev: our device | |
604 | * @args: the requested arguments copied from userspace | |
605 | * | |
606 | * Allocate a buffer suitable for use for a frame buffer of the | |
607 | * form described by user space. Give userspace a handle by which | |
608 | * to reference it. | |
609 | */ | |
610 | int omap_gem_dumb_create(struct drm_file *file, struct drm_device *dev, | |
611 | struct drm_mode_create_dumb *args) | |
612 | { | |
613 | union omap_gem_size gsize; | |
614 | ||
bdb2b933 | 615 | args->pitch = align_pitch(0, args->width, args->bpp); |
cd5351f4 RC |
616 | args->size = PAGE_ALIGN(args->pitch * args->height); |
617 | ||
618 | gsize = (union omap_gem_size){ | |
619 | .bytes = args->size, | |
620 | }; | |
621 | ||
622 | return omap_gem_new_handle(dev, file, gsize, | |
623 | OMAP_BO_SCANOUT | OMAP_BO_WC, &args->handle); | |
624 | } | |
625 | ||
cd5351f4 RC |
626 | /** |
627 | * omap_gem_dumb_map - buffer mapping for dumb interface | |
628 | * @file: our drm client file | |
629 | * @dev: drm device | |
630 | * @handle: GEM handle to the object (from dumb_create) | |
631 | * | |
632 | * Do the necessary setup to allow the mapping of the frame buffer | |
633 | * into user memory. We don't have to do much here at the moment. | |
634 | */ | |
635 | int omap_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev, | |
636 | uint32_t handle, uint64_t *offset) | |
637 | { | |
638 | struct drm_gem_object *obj; | |
639 | int ret = 0; | |
640 | ||
cd5351f4 RC |
641 | /* GEM does all our handle to object mapping */ |
642 | obj = drm_gem_object_lookup(dev, file, handle); | |
643 | if (obj == NULL) { | |
644 | ret = -ENOENT; | |
645 | goto fail; | |
646 | } | |
647 | ||
648 | *offset = omap_gem_mmap_offset(obj); | |
649 | ||
650 | drm_gem_object_unreference_unlocked(obj); | |
651 | ||
652 | fail: | |
cd5351f4 RC |
653 | return ret; |
654 | } | |
655 | ||
a6a91827 RC |
656 | /* Set scrolling position. This allows us to implement fast scrolling |
657 | * for console. | |
9b55b95a RC |
658 | * |
659 | * Call only from non-atomic contexts. | |
a6a91827 RC |
660 | */ |
661 | int omap_gem_roll(struct drm_gem_object *obj, uint32_t roll) | |
662 | { | |
663 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
664 | uint32_t npages = obj->size >> PAGE_SHIFT; | |
665 | int ret = 0; | |
666 | ||
667 | if (roll > npages) { | |
668 | dev_err(obj->dev->dev, "invalid roll: %d\n", roll); | |
669 | return -EINVAL; | |
670 | } | |
671 | ||
a6a91827 RC |
672 | omap_obj->roll = roll; |
673 | ||
af69592a RC |
674 | mutex_lock(&obj->dev->struct_mutex); |
675 | ||
a6a91827 RC |
676 | /* if we aren't mapped yet, we don't need to do anything */ |
677 | if (omap_obj->block) { | |
678 | struct page **pages; | |
679 | ret = get_pages(obj, &pages); | |
680 | if (ret) | |
681 | goto fail; | |
682 | ret = tiler_pin(omap_obj->block, pages, npages, roll, true); | |
683 | if (ret) | |
684 | dev_err(obj->dev->dev, "could not repin: %d\n", ret); | |
685 | } | |
686 | ||
687 | fail: | |
688 | mutex_unlock(&obj->dev->struct_mutex); | |
689 | ||
690 | return ret; | |
691 | } | |
692 | ||
8b6b569e RC |
693 | /* Sync the buffer for CPU access.. note pages should already be |
694 | * attached, ie. omap_gem_get_pages() | |
695 | */ | |
696 | void omap_gem_cpu_sync(struct drm_gem_object *obj, int pgoff) | |
697 | { | |
698 | struct drm_device *dev = obj->dev; | |
699 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
700 | ||
701 | if (is_cached_coherent(obj) && omap_obj->addrs[pgoff]) { | |
702 | dma_unmap_page(dev->dev, omap_obj->addrs[pgoff], | |
703 | PAGE_SIZE, DMA_BIDIRECTIONAL); | |
704 | omap_obj->addrs[pgoff] = 0; | |
705 | } | |
706 | } | |
707 | ||
708 | /* sync the buffer for DMA access */ | |
709 | void omap_gem_dma_sync(struct drm_gem_object *obj, | |
710 | enum dma_data_direction dir) | |
711 | { | |
712 | struct drm_device *dev = obj->dev; | |
713 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
714 | ||
715 | if (is_cached_coherent(obj)) { | |
716 | int i, npages = obj->size >> PAGE_SHIFT; | |
717 | struct page **pages = omap_obj->pages; | |
718 | bool dirty = false; | |
719 | ||
720 | for (i = 0; i < npages; i++) { | |
721 | if (!omap_obj->addrs[i]) { | |
722 | omap_obj->addrs[i] = dma_map_page(dev->dev, pages[i], 0, | |
723 | PAGE_SIZE, DMA_BIDIRECTIONAL); | |
724 | dirty = true; | |
725 | } | |
726 | } | |
727 | ||
728 | if (dirty) { | |
729 | unmap_mapping_range(obj->filp->f_mapping, 0, | |
730 | omap_gem_mmap_size(obj), 1); | |
731 | } | |
732 | } | |
733 | } | |
734 | ||
cd5351f4 RC |
735 | /* Get physical address for DMA.. if 'remap' is true, and the buffer is not |
736 | * already contiguous, remap it to pin in physically contiguous memory.. (ie. | |
737 | * map in TILER) | |
738 | */ | |
739 | int omap_gem_get_paddr(struct drm_gem_object *obj, | |
740 | dma_addr_t *paddr, bool remap) | |
741 | { | |
a6a91827 | 742 | struct omap_drm_private *priv = obj->dev->dev_private; |
cd5351f4 RC |
743 | struct omap_gem_object *omap_obj = to_omap_bo(obj); |
744 | int ret = 0; | |
745 | ||
f7f9f453 RC |
746 | mutex_lock(&obj->dev->struct_mutex); |
747 | ||
a6a91827 | 748 | if (remap && is_shmem(obj) && priv->has_dmm) { |
f7f9f453 RC |
749 | if (omap_obj->paddr_cnt == 0) { |
750 | struct page **pages; | |
a6a91827 | 751 | uint32_t npages = obj->size >> PAGE_SHIFT; |
f7f9f453 RC |
752 | enum tiler_fmt fmt = gem2fmt(omap_obj->flags); |
753 | struct tiler_block *block; | |
a6a91827 | 754 | |
f7f9f453 RC |
755 | BUG_ON(omap_obj->block); |
756 | ||
757 | ret = get_pages(obj, &pages); | |
758 | if (ret) | |
759 | goto fail; | |
760 | ||
f7f9f453 RC |
761 | if (omap_obj->flags & OMAP_BO_TILED) { |
762 | block = tiler_reserve_2d(fmt, | |
763 | omap_obj->width, | |
764 | omap_obj->height, 0); | |
765 | } else { | |
766 | block = tiler_reserve_1d(obj->size); | |
767 | } | |
768 | ||
769 | if (IS_ERR(block)) { | |
770 | ret = PTR_ERR(block); | |
771 | dev_err(obj->dev->dev, | |
772 | "could not remap: %d (%d)\n", ret, fmt); | |
773 | goto fail; | |
774 | } | |
775 | ||
776 | /* TODO: enable async refill.. */ | |
a6a91827 RC |
777 | ret = tiler_pin(block, pages, npages, |
778 | omap_obj->roll, true); | |
f7f9f453 RC |
779 | if (ret) { |
780 | tiler_release(block); | |
781 | dev_err(obj->dev->dev, | |
782 | "could not pin: %d\n", ret); | |
783 | goto fail; | |
784 | } | |
785 | ||
786 | omap_obj->paddr = tiler_ssptr(block); | |
787 | omap_obj->block = block; | |
788 | ||
2d31ca3a | 789 | DBG("got paddr: %pad", &omap_obj->paddr); |
f7f9f453 RC |
790 | } |
791 | ||
792 | omap_obj->paddr_cnt++; | |
793 | ||
794 | *paddr = omap_obj->paddr; | |
795 | } else if (omap_obj->flags & OMAP_BO_DMA) { | |
796 | *paddr = omap_obj->paddr; | |
797 | } else { | |
798 | ret = -EINVAL; | |
8b6b569e | 799 | goto fail; |
cd5351f4 RC |
800 | } |
801 | ||
f7f9f453 RC |
802 | fail: |
803 | mutex_unlock(&obj->dev->struct_mutex); | |
cd5351f4 RC |
804 | |
805 | return ret; | |
806 | } | |
807 | ||
808 | /* Release physical address, when DMA is no longer being performed.. this | |
809 | * could potentially unpin and unmap buffers from TILER | |
810 | */ | |
393a949f | 811 | void omap_gem_put_paddr(struct drm_gem_object *obj) |
cd5351f4 | 812 | { |
f7f9f453 | 813 | struct omap_gem_object *omap_obj = to_omap_bo(obj); |
393a949f | 814 | int ret; |
f7f9f453 RC |
815 | |
816 | mutex_lock(&obj->dev->struct_mutex); | |
817 | if (omap_obj->paddr_cnt > 0) { | |
818 | omap_obj->paddr_cnt--; | |
819 | if (omap_obj->paddr_cnt == 0) { | |
820 | ret = tiler_unpin(omap_obj->block); | |
821 | if (ret) { | |
822 | dev_err(obj->dev->dev, | |
823 | "could not unpin pages: %d\n", ret); | |
f7f9f453 RC |
824 | } |
825 | ret = tiler_release(omap_obj->block); | |
826 | if (ret) { | |
827 | dev_err(obj->dev->dev, | |
828 | "could not release unmap: %d\n", ret); | |
829 | } | |
3f4d17c4 | 830 | omap_obj->paddr = 0; |
f7f9f453 RC |
831 | omap_obj->block = NULL; |
832 | } | |
833 | } | |
393a949f | 834 | |
f7f9f453 | 835 | mutex_unlock(&obj->dev->struct_mutex); |
cd5351f4 RC |
836 | } |
837 | ||
3c810c61 RC |
838 | /* Get rotated scanout address (only valid if already pinned), at the |
839 | * specified orientation and x,y offset from top-left corner of buffer | |
840 | * (only valid for tiled 2d buffers) | |
841 | */ | |
842 | int omap_gem_rotated_paddr(struct drm_gem_object *obj, uint32_t orient, | |
843 | int x, int y, dma_addr_t *paddr) | |
844 | { | |
845 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
846 | int ret = -EINVAL; | |
847 | ||
848 | mutex_lock(&obj->dev->struct_mutex); | |
849 | if ((omap_obj->paddr_cnt > 0) && omap_obj->block && | |
850 | (omap_obj->flags & OMAP_BO_TILED)) { | |
851 | *paddr = tiler_tsptr(omap_obj->block, orient, x, y); | |
852 | ret = 0; | |
853 | } | |
854 | mutex_unlock(&obj->dev->struct_mutex); | |
855 | return ret; | |
856 | } | |
857 | ||
858 | /* Get tiler stride for the buffer (only valid for 2d tiled buffers) */ | |
859 | int omap_gem_tiled_stride(struct drm_gem_object *obj, uint32_t orient) | |
860 | { | |
861 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
862 | int ret = -EINVAL; | |
863 | if (omap_obj->flags & OMAP_BO_TILED) | |
864 | ret = tiler_stride(gem2fmt(omap_obj->flags), orient); | |
865 | return ret; | |
866 | } | |
867 | ||
cd5351f4 RC |
868 | /* acquire pages when needed (for example, for DMA where physically |
869 | * contiguous buffer is not required | |
870 | */ | |
871 | static int get_pages(struct drm_gem_object *obj, struct page ***pages) | |
872 | { | |
873 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
874 | int ret = 0; | |
875 | ||
876 | if (is_shmem(obj) && !omap_obj->pages) { | |
877 | ret = omap_gem_attach_pages(obj); | |
878 | if (ret) { | |
879 | dev_err(obj->dev->dev, "could not attach pages\n"); | |
880 | return ret; | |
881 | } | |
882 | } | |
883 | ||
884 | /* TODO: even phys-contig.. we should have a list of pages? */ | |
885 | *pages = omap_obj->pages; | |
886 | ||
887 | return 0; | |
888 | } | |
889 | ||
6ad11bc3 RC |
890 | /* if !remap, and we don't have pages backing, then fail, rather than |
891 | * increasing the pin count (which we don't really do yet anyways, | |
892 | * because we don't support swapping pages back out). And 'remap' | |
893 | * might not be quite the right name, but I wanted to keep it working | |
894 | * similarly to omap_gem_get_paddr(). Note though that mutex is not | |
895 | * aquired if !remap (because this can be called in atomic ctxt), | |
896 | * but probably omap_gem_get_paddr() should be changed to work in the | |
897 | * same way. If !remap, a matching omap_gem_put_pages() call is not | |
898 | * required (and should not be made). | |
899 | */ | |
900 | int omap_gem_get_pages(struct drm_gem_object *obj, struct page ***pages, | |
901 | bool remap) | |
cd5351f4 RC |
902 | { |
903 | int ret; | |
6ad11bc3 RC |
904 | if (!remap) { |
905 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
906 | if (!omap_obj->pages) | |
907 | return -ENOMEM; | |
908 | *pages = omap_obj->pages; | |
909 | return 0; | |
910 | } | |
cd5351f4 RC |
911 | mutex_lock(&obj->dev->struct_mutex); |
912 | ret = get_pages(obj, pages); | |
913 | mutex_unlock(&obj->dev->struct_mutex); | |
914 | return ret; | |
915 | } | |
916 | ||
917 | /* release pages when DMA no longer being performed */ | |
918 | int omap_gem_put_pages(struct drm_gem_object *obj) | |
919 | { | |
920 | /* do something here if we dynamically attach/detach pages.. at | |
921 | * least they would no longer need to be pinned if everyone has | |
922 | * released the pages.. | |
923 | */ | |
924 | return 0; | |
925 | } | |
926 | ||
f7f9f453 RC |
927 | /* Get kernel virtual address for CPU access.. this more or less only |
928 | * exists for omap_fbdev. This should be called with struct_mutex | |
929 | * held. | |
cd5351f4 RC |
930 | */ |
931 | void *omap_gem_vaddr(struct drm_gem_object *obj) | |
932 | { | |
933 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
696e3ca3 | 934 | WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex)); |
f7f9f453 RC |
935 | if (!omap_obj->vaddr) { |
936 | struct page **pages; | |
937 | int ret = get_pages(obj, &pages); | |
938 | if (ret) | |
939 | return ERR_PTR(ret); | |
940 | omap_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT, | |
941 | VM_MAP, pgprot_writecombine(PAGE_KERNEL)); | |
942 | } | |
cd5351f4 RC |
943 | return omap_obj->vaddr; |
944 | } | |
945 | ||
e78edba1 AG |
946 | #ifdef CONFIG_PM |
947 | /* re-pin objects in DMM in resume path: */ | |
948 | int omap_gem_resume(struct device *dev) | |
949 | { | |
950 | struct drm_device *drm_dev = dev_get_drvdata(dev); | |
951 | struct omap_drm_private *priv = drm_dev->dev_private; | |
952 | struct omap_gem_object *omap_obj; | |
953 | int ret = 0; | |
954 | ||
955 | list_for_each_entry(omap_obj, &priv->obj_list, mm_list) { | |
956 | if (omap_obj->block) { | |
957 | struct drm_gem_object *obj = &omap_obj->base; | |
958 | uint32_t npages = obj->size >> PAGE_SHIFT; | |
959 | WARN_ON(!omap_obj->pages); /* this can't happen */ | |
960 | ret = tiler_pin(omap_obj->block, | |
961 | omap_obj->pages, npages, | |
962 | omap_obj->roll, true); | |
963 | if (ret) { | |
964 | dev_err(dev, "could not repin: %d\n", ret); | |
965 | return ret; | |
966 | } | |
967 | } | |
968 | } | |
969 | ||
970 | return 0; | |
971 | } | |
972 | #endif | |
973 | ||
f6b6036e RC |
974 | #ifdef CONFIG_DEBUG_FS |
975 | void omap_gem_describe(struct drm_gem_object *obj, struct seq_file *m) | |
976 | { | |
f6b6036e | 977 | struct omap_gem_object *omap_obj = to_omap_bo(obj); |
0de23977 | 978 | uint64_t off; |
f6b6036e | 979 | |
0de23977 | 980 | off = drm_vma_node_start(&obj->vma_node); |
f6b6036e | 981 | |
2d31ca3a | 982 | seq_printf(m, "%08x: %2d (%2d) %08llx %pad (%2d) %p %4d", |
f6b6036e | 983 | omap_obj->flags, obj->name, obj->refcount.refcount.counter, |
2d31ca3a | 984 | off, &omap_obj->paddr, omap_obj->paddr_cnt, |
f6b6036e RC |
985 | omap_obj->vaddr, omap_obj->roll); |
986 | ||
987 | if (omap_obj->flags & OMAP_BO_TILED) { | |
988 | seq_printf(m, " %dx%d", omap_obj->width, omap_obj->height); | |
989 | if (omap_obj->block) { | |
990 | struct tcm_area *area = &omap_obj->block->area; | |
991 | seq_printf(m, " (%dx%d, %dx%d)", | |
992 | area->p0.x, area->p0.y, | |
993 | area->p1.x, area->p1.y); | |
994 | } | |
995 | } else { | |
996 | seq_printf(m, " %d", obj->size); | |
997 | } | |
998 | ||
999 | seq_printf(m, "\n"); | |
1000 | } | |
1001 | ||
1002 | void omap_gem_describe_objects(struct list_head *list, struct seq_file *m) | |
1003 | { | |
1004 | struct omap_gem_object *omap_obj; | |
1005 | int count = 0; | |
1006 | size_t size = 0; | |
1007 | ||
1008 | list_for_each_entry(omap_obj, list, mm_list) { | |
1009 | struct drm_gem_object *obj = &omap_obj->base; | |
1010 | seq_printf(m, " "); | |
1011 | omap_gem_describe(obj, m); | |
1012 | count++; | |
1013 | size += obj->size; | |
1014 | } | |
1015 | ||
1016 | seq_printf(m, "Total %d objects, %zu bytes\n", count, size); | |
1017 | } | |
1018 | #endif | |
1019 | ||
cd5351f4 RC |
1020 | /* Buffer Synchronization: |
1021 | */ | |
1022 | ||
1023 | struct omap_gem_sync_waiter { | |
1024 | struct list_head list; | |
1025 | struct omap_gem_object *omap_obj; | |
1026 | enum omap_gem_op op; | |
1027 | uint32_t read_target, write_target; | |
1028 | /* notify called w/ sync_lock held */ | |
1029 | void (*notify)(void *arg); | |
1030 | void *arg; | |
1031 | }; | |
1032 | ||
1033 | /* list of omap_gem_sync_waiter.. the notify fxn gets called back when | |
1034 | * the read and/or write target count is achieved which can call a user | |
1035 | * callback (ex. to kick 3d and/or 2d), wakeup blocked task (prep for | |
1036 | * cpu access), etc. | |
1037 | */ | |
1038 | static LIST_HEAD(waiters); | |
1039 | ||
1040 | static inline bool is_waiting(struct omap_gem_sync_waiter *waiter) | |
1041 | { | |
1042 | struct omap_gem_object *omap_obj = waiter->omap_obj; | |
1043 | if ((waiter->op & OMAP_GEM_READ) && | |
f2cff0f3 | 1044 | (omap_obj->sync->write_complete < waiter->write_target)) |
cd5351f4 RC |
1045 | return true; |
1046 | if ((waiter->op & OMAP_GEM_WRITE) && | |
f2cff0f3 | 1047 | (omap_obj->sync->read_complete < waiter->read_target)) |
cd5351f4 RC |
1048 | return true; |
1049 | return false; | |
1050 | } | |
1051 | ||
1052 | /* macro for sync debug.. */ | |
1053 | #define SYNCDBG 0 | |
1054 | #define SYNC(fmt, ...) do { if (SYNCDBG) \ | |
1055 | printk(KERN_ERR "%s:%d: "fmt"\n", \ | |
1056 | __func__, __LINE__, ##__VA_ARGS__); \ | |
1057 | } while (0) | |
1058 | ||
1059 | ||
1060 | static void sync_op_update(void) | |
1061 | { | |
1062 | struct omap_gem_sync_waiter *waiter, *n; | |
1063 | list_for_each_entry_safe(waiter, n, &waiters, list) { | |
1064 | if (!is_waiting(waiter)) { | |
1065 | list_del(&waiter->list); | |
1066 | SYNC("notify: %p", waiter); | |
1067 | waiter->notify(waiter->arg); | |
1068 | kfree(waiter); | |
1069 | } | |
1070 | } | |
1071 | } | |
1072 | ||
1073 | static inline int sync_op(struct drm_gem_object *obj, | |
1074 | enum omap_gem_op op, bool start) | |
1075 | { | |
1076 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
1077 | int ret = 0; | |
1078 | ||
1079 | spin_lock(&sync_lock); | |
1080 | ||
1081 | if (!omap_obj->sync) { | |
1082 | omap_obj->sync = kzalloc(sizeof(*omap_obj->sync), GFP_ATOMIC); | |
1083 | if (!omap_obj->sync) { | |
1084 | ret = -ENOMEM; | |
1085 | goto unlock; | |
1086 | } | |
1087 | } | |
1088 | ||
1089 | if (start) { | |
1090 | if (op & OMAP_GEM_READ) | |
1091 | omap_obj->sync->read_pending++; | |
1092 | if (op & OMAP_GEM_WRITE) | |
1093 | omap_obj->sync->write_pending++; | |
1094 | } else { | |
1095 | if (op & OMAP_GEM_READ) | |
1096 | omap_obj->sync->read_complete++; | |
1097 | if (op & OMAP_GEM_WRITE) | |
1098 | omap_obj->sync->write_complete++; | |
1099 | sync_op_update(); | |
1100 | } | |
1101 | ||
1102 | unlock: | |
1103 | spin_unlock(&sync_lock); | |
1104 | ||
1105 | return ret; | |
1106 | } | |
1107 | ||
1108 | /* it is a bit lame to handle updates in this sort of polling way, but | |
1109 | * in case of PVR, the GPU can directly update read/write complete | |
1110 | * values, and not really tell us which ones it updated.. this also | |
1111 | * means that sync_lock is not quite sufficient. So we'll need to | |
1112 | * do something a bit better when it comes time to add support for | |
1113 | * separate 2d hw.. | |
1114 | */ | |
1115 | void omap_gem_op_update(void) | |
1116 | { | |
1117 | spin_lock(&sync_lock); | |
1118 | sync_op_update(); | |
1119 | spin_unlock(&sync_lock); | |
1120 | } | |
1121 | ||
1122 | /* mark the start of read and/or write operation */ | |
1123 | int omap_gem_op_start(struct drm_gem_object *obj, enum omap_gem_op op) | |
1124 | { | |
1125 | return sync_op(obj, op, true); | |
1126 | } | |
1127 | ||
1128 | int omap_gem_op_finish(struct drm_gem_object *obj, enum omap_gem_op op) | |
1129 | { | |
1130 | return sync_op(obj, op, false); | |
1131 | } | |
1132 | ||
1133 | static DECLARE_WAIT_QUEUE_HEAD(sync_event); | |
1134 | ||
1135 | static void sync_notify(void *arg) | |
1136 | { | |
1137 | struct task_struct **waiter_task = arg; | |
1138 | *waiter_task = NULL; | |
1139 | wake_up_all(&sync_event); | |
1140 | } | |
1141 | ||
1142 | int omap_gem_op_sync(struct drm_gem_object *obj, enum omap_gem_op op) | |
1143 | { | |
1144 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
1145 | int ret = 0; | |
1146 | if (omap_obj->sync) { | |
1147 | struct task_struct *waiter_task = current; | |
1148 | struct omap_gem_sync_waiter *waiter = | |
1149 | kzalloc(sizeof(*waiter), GFP_KERNEL); | |
1150 | ||
ae053039 | 1151 | if (!waiter) |
cd5351f4 | 1152 | return -ENOMEM; |
cd5351f4 RC |
1153 | |
1154 | waiter->omap_obj = omap_obj; | |
1155 | waiter->op = op; | |
1156 | waiter->read_target = omap_obj->sync->read_pending; | |
1157 | waiter->write_target = omap_obj->sync->write_pending; | |
1158 | waiter->notify = sync_notify; | |
1159 | waiter->arg = &waiter_task; | |
1160 | ||
1161 | spin_lock(&sync_lock); | |
1162 | if (is_waiting(waiter)) { | |
1163 | SYNC("waited: %p", waiter); | |
1164 | list_add_tail(&waiter->list, &waiters); | |
1165 | spin_unlock(&sync_lock); | |
1166 | ret = wait_event_interruptible(sync_event, | |
1167 | (waiter_task == NULL)); | |
1168 | spin_lock(&sync_lock); | |
1169 | if (waiter_task) { | |
1170 | SYNC("interrupted: %p", waiter); | |
1171 | /* we were interrupted */ | |
1172 | list_del(&waiter->list); | |
1173 | waiter_task = NULL; | |
1174 | } else { | |
1175 | /* freed in sync_op_update() */ | |
1176 | waiter = NULL; | |
1177 | } | |
1178 | } | |
1179 | spin_unlock(&sync_lock); | |
d2c87e2d | 1180 | kfree(waiter); |
cd5351f4 RC |
1181 | } |
1182 | return ret; | |
1183 | } | |
1184 | ||
1185 | /* call fxn(arg), either synchronously or asynchronously if the op | |
1186 | * is currently blocked.. fxn() can be called from any context | |
1187 | * | |
1188 | * (TODO for now fxn is called back from whichever context calls | |
1189 | * omap_gem_op_update().. but this could be better defined later | |
1190 | * if needed) | |
1191 | * | |
1192 | * TODO more code in common w/ _sync().. | |
1193 | */ | |
1194 | int omap_gem_op_async(struct drm_gem_object *obj, enum omap_gem_op op, | |
1195 | void (*fxn)(void *arg), void *arg) | |
1196 | { | |
1197 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
1198 | if (omap_obj->sync) { | |
1199 | struct omap_gem_sync_waiter *waiter = | |
1200 | kzalloc(sizeof(*waiter), GFP_ATOMIC); | |
1201 | ||
ae053039 | 1202 | if (!waiter) |
cd5351f4 | 1203 | return -ENOMEM; |
cd5351f4 RC |
1204 | |
1205 | waiter->omap_obj = omap_obj; | |
1206 | waiter->op = op; | |
1207 | waiter->read_target = omap_obj->sync->read_pending; | |
1208 | waiter->write_target = omap_obj->sync->write_pending; | |
1209 | waiter->notify = fxn; | |
1210 | waiter->arg = arg; | |
1211 | ||
1212 | spin_lock(&sync_lock); | |
1213 | if (is_waiting(waiter)) { | |
1214 | SYNC("waited: %p", waiter); | |
1215 | list_add_tail(&waiter->list, &waiters); | |
1216 | spin_unlock(&sync_lock); | |
1217 | return 0; | |
1218 | } | |
1219 | ||
1220 | spin_unlock(&sync_lock); | |
15ec2ca9 SP |
1221 | |
1222 | kfree(waiter); | |
cd5351f4 RC |
1223 | } |
1224 | ||
1225 | /* no waiting.. */ | |
1226 | fxn(arg); | |
1227 | ||
1228 | return 0; | |
1229 | } | |
1230 | ||
1231 | /* special API so PVR can update the buffer to use a sync-object allocated | |
1232 | * from it's sync-obj heap. Only used for a newly allocated (from PVR's | |
1233 | * perspective) sync-object, so we overwrite the new syncobj w/ values | |
1234 | * from the already allocated syncobj (if there is one) | |
1235 | */ | |
1236 | int omap_gem_set_sync_object(struct drm_gem_object *obj, void *syncobj) | |
1237 | { | |
1238 | struct omap_gem_object *omap_obj = to_omap_bo(obj); | |
1239 | int ret = 0; | |
1240 | ||
1241 | spin_lock(&sync_lock); | |
1242 | ||
1243 | if ((omap_obj->flags & OMAP_BO_EXT_SYNC) && !syncobj) { | |
1244 | /* clearing a previously set syncobj */ | |
e6200964 PH |
1245 | syncobj = kmemdup(omap_obj->sync, sizeof(*omap_obj->sync), |
1246 | GFP_ATOMIC); | |
cd5351f4 RC |
1247 | if (!syncobj) { |
1248 | ret = -ENOMEM; | |
1249 | goto unlock; | |
1250 | } | |
cd5351f4 RC |
1251 | omap_obj->flags &= ~OMAP_BO_EXT_SYNC; |
1252 | omap_obj->sync = syncobj; | |
1253 | } else if (syncobj && !(omap_obj->flags & OMAP_BO_EXT_SYNC)) { | |
1254 | /* replacing an existing syncobj */ | |
1255 | if (omap_obj->sync) { | |
1256 | memcpy(syncobj, omap_obj->sync, sizeof(*omap_obj->sync)); | |
1257 | kfree(omap_obj->sync); | |
1258 | } | |
1259 | omap_obj->flags |= OMAP_BO_EXT_SYNC; | |
1260 | omap_obj->sync = syncobj; | |
1261 | } | |
1262 | ||
1263 | unlock: | |
1264 | spin_unlock(&sync_lock); | |
1265 | return ret; | |
1266 | } | |
1267 | ||
cd5351f4 RC |
1268 | /* don't call directly.. called from GEM core when it is time to actually |
1269 | * free the object.. | |
1270 | */ | |
1271 | void omap_gem_free_object(struct drm_gem_object *obj) | |
1272 | { | |
1273 | struct drm_device *dev = obj->dev; | |
76c4055f | 1274 | struct omap_drm_private *priv = dev->dev_private; |
cd5351f4 RC |
1275 | struct omap_gem_object *omap_obj = to_omap_bo(obj); |
1276 | ||
f7f9f453 RC |
1277 | evict(obj); |
1278 | ||
f6b6036e RC |
1279 | WARN_ON(!mutex_is_locked(&dev->struct_mutex)); |
1280 | ||
76c4055f | 1281 | spin_lock(&priv->list_lock); |
f6b6036e | 1282 | list_del(&omap_obj->mm_list); |
76c4055f | 1283 | spin_unlock(&priv->list_lock); |
f6b6036e | 1284 | |
0de23977 | 1285 | drm_gem_free_mmap_offset(obj); |
cd5351f4 | 1286 | |
9a0774e0 RC |
1287 | /* this means the object is still pinned.. which really should |
1288 | * not happen. I think.. | |
1289 | */ | |
1290 | WARN_ON(omap_obj->paddr_cnt > 0); | |
1291 | ||
cd5351f4 RC |
1292 | /* don't free externally allocated backing memory */ |
1293 | if (!(omap_obj->flags & OMAP_BO_EXT_MEM)) { | |
ae053039 | 1294 | if (omap_obj->pages) |
cd5351f4 | 1295 | omap_gem_detach_pages(obj); |
ae053039 | 1296 | |
cd5351f4 RC |
1297 | if (!is_shmem(obj)) { |
1298 | dma_free_writecombine(dev->dev, obj->size, | |
1299 | omap_obj->vaddr, omap_obj->paddr); | |
f7f9f453 RC |
1300 | } else if (omap_obj->vaddr) { |
1301 | vunmap(omap_obj->vaddr); | |
cd5351f4 RC |
1302 | } |
1303 | } | |
1304 | ||
1305 | /* don't free externally allocated syncobj */ | |
ae053039 | 1306 | if (!(omap_obj->flags & OMAP_BO_EXT_SYNC)) |
cd5351f4 | 1307 | kfree(omap_obj->sync); |
cd5351f4 RC |
1308 | |
1309 | drm_gem_object_release(obj); | |
1310 | ||
1311 | kfree(obj); | |
1312 | } | |
1313 | ||
1314 | /* convenience method to construct a GEM buffer object, and userspace handle */ | |
1315 | int omap_gem_new_handle(struct drm_device *dev, struct drm_file *file, | |
1316 | union omap_gem_size gsize, uint32_t flags, uint32_t *handle) | |
1317 | { | |
1318 | struct drm_gem_object *obj; | |
1319 | int ret; | |
1320 | ||
1321 | obj = omap_gem_new(dev, gsize, flags); | |
1322 | if (!obj) | |
1323 | return -ENOMEM; | |
1324 | ||
1325 | ret = drm_gem_handle_create(file, obj, handle); | |
1326 | if (ret) { | |
1327 | drm_gem_object_release(obj); | |
1328 | kfree(obj); /* TODO isn't there a dtor to call? just copying i915 */ | |
1329 | return ret; | |
1330 | } | |
1331 | ||
1332 | /* drop reference from allocate - handle holds it now */ | |
1333 | drm_gem_object_unreference_unlocked(obj); | |
1334 | ||
1335 | return 0; | |
1336 | } | |
1337 | ||
1338 | /* GEM buffer object constructor */ | |
1339 | struct drm_gem_object *omap_gem_new(struct drm_device *dev, | |
1340 | union omap_gem_size gsize, uint32_t flags) | |
1341 | { | |
a6a91827 | 1342 | struct omap_drm_private *priv = dev->dev_private; |
cd5351f4 RC |
1343 | struct omap_gem_object *omap_obj; |
1344 | struct drm_gem_object *obj = NULL; | |
ab5a60c3 | 1345 | struct address_space *mapping; |
cd5351f4 RC |
1346 | size_t size; |
1347 | int ret; | |
1348 | ||
1349 | if (flags & OMAP_BO_TILED) { | |
f7f9f453 RC |
1350 | if (!usergart) { |
1351 | dev_err(dev->dev, "Tiled buffers require DMM\n"); | |
1352 | goto fail; | |
1353 | } | |
1354 | ||
1355 | /* tiled buffers are always shmem paged backed.. when they are | |
1356 | * scanned out, they are remapped into DMM/TILER | |
1357 | */ | |
1358 | flags &= ~OMAP_BO_SCANOUT; | |
1359 | ||
1360 | /* currently don't allow cached buffers.. there is some caching | |
1361 | * stuff that needs to be handled better | |
1362 | */ | |
7cb0d6c1 TV |
1363 | flags &= ~(OMAP_BO_CACHED|OMAP_BO_WC|OMAP_BO_UNCACHED); |
1364 | flags |= tiler_get_cpu_cache_flags(); | |
cd5351f4 | 1365 | |
f7f9f453 RC |
1366 | /* align dimensions to slot boundaries... */ |
1367 | tiler_align(gem2fmt(flags), | |
1368 | &gsize.tiled.width, &gsize.tiled.height); | |
1369 | ||
1370 | /* ...and calculate size based on aligned dimensions */ | |
1371 | size = tiler_size(gem2fmt(flags), | |
1372 | gsize.tiled.width, gsize.tiled.height); | |
1373 | } else { | |
1374 | size = PAGE_ALIGN(gsize.bytes); | |
1375 | } | |
cd5351f4 RC |
1376 | |
1377 | omap_obj = kzalloc(sizeof(*omap_obj), GFP_KERNEL); | |
78110bb8 | 1378 | if (!omap_obj) |
a903e3b6 | 1379 | return NULL; |
f6b6036e | 1380 | |
cd5351f4 RC |
1381 | obj = &omap_obj->base; |
1382 | ||
a6a91827 RC |
1383 | if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) { |
1384 | /* attempt to allocate contiguous memory if we don't | |
1385 | * have DMM for remappign discontiguous buffers | |
1386 | */ | |
cd5351f4 RC |
1387 | omap_obj->vaddr = dma_alloc_writecombine(dev->dev, size, |
1388 | &omap_obj->paddr, GFP_KERNEL); | |
a903e3b6 TV |
1389 | if (!omap_obj->vaddr) { |
1390 | kfree(omap_obj); | |
ae053039 | 1391 | |
a903e3b6 TV |
1392 | return NULL; |
1393 | } | |
1394 | ||
1395 | flags |= OMAP_BO_DMA; | |
cd5351f4 RC |
1396 | } |
1397 | ||
a903e3b6 TV |
1398 | spin_lock(&priv->list_lock); |
1399 | list_add(&omap_obj->mm_list, &priv->obj_list); | |
1400 | spin_unlock(&priv->list_lock); | |
1401 | ||
cd5351f4 RC |
1402 | omap_obj->flags = flags; |
1403 | ||
f7f9f453 RC |
1404 | if (flags & OMAP_BO_TILED) { |
1405 | omap_obj->width = gsize.tiled.width; | |
1406 | omap_obj->height = gsize.tiled.height; | |
1407 | } | |
1408 | ||
ab5a60c3 | 1409 | if (flags & (OMAP_BO_DMA|OMAP_BO_EXT_MEM)) { |
89c8233f | 1410 | drm_gem_private_object_init(dev, obj, size); |
ab5a60c3 | 1411 | } else { |
cd5351f4 | 1412 | ret = drm_gem_object_init(dev, obj, size); |
ab5a60c3 DH |
1413 | if (ret) |
1414 | goto fail; | |
cd5351f4 | 1415 | |
ab5a60c3 DH |
1416 | mapping = file_inode(obj->filp)->i_mapping; |
1417 | mapping_set_gfp_mask(mapping, GFP_USER | __GFP_DMA32); | |
1418 | } | |
cd5351f4 RC |
1419 | |
1420 | return obj; | |
1421 | ||
1422 | fail: | |
ae053039 | 1423 | if (obj) |
cd5351f4 | 1424 | omap_gem_free_object(obj); |
ae053039 | 1425 | |
cd5351f4 RC |
1426 | return NULL; |
1427 | } | |
f7f9f453 RC |
1428 | |
1429 | /* init/cleanup.. if DMM is used, we need to set some stuff up.. */ | |
1430 | void omap_gem_init(struct drm_device *dev) | |
1431 | { | |
a6a91827 | 1432 | struct omap_drm_private *priv = dev->dev_private; |
f7f9f453 RC |
1433 | const enum tiler_fmt fmts[] = { |
1434 | TILFMT_8BIT, TILFMT_16BIT, TILFMT_32BIT | |
1435 | }; | |
5c137797 | 1436 | int i, j; |
f7f9f453 | 1437 | |
e5e4e9b7 | 1438 | if (!dmm_is_available()) { |
f7f9f453 | 1439 | /* DMM only supported on OMAP4 and later, so this isn't fatal */ |
5c137797 | 1440 | dev_warn(dev->dev, "DMM not available, disable DMM support\n"); |
f7f9f453 RC |
1441 | return; |
1442 | } | |
1443 | ||
78110bb8 JP |
1444 | usergart = kcalloc(3, sizeof(*usergart), GFP_KERNEL); |
1445 | if (!usergart) | |
b369839b | 1446 | return; |
f7f9f453 RC |
1447 | |
1448 | /* reserve 4k aligned/wide regions for userspace mappings: */ | |
1449 | for (i = 0; i < ARRAY_SIZE(fmts); i++) { | |
1450 | uint16_t h = 1, w = PAGE_SIZE >> i; | |
1451 | tiler_align(fmts[i], &w, &h); | |
1452 | /* note: since each region is 1 4kb page wide, and minimum | |
1453 | * number of rows, the height ends up being the same as the | |
1454 | * # of pages in the region | |
1455 | */ | |
1456 | usergart[i].height = h; | |
1457 | usergart[i].height_shift = ilog2(h); | |
3c810c61 | 1458 | usergart[i].stride_pfn = tiler_stride(fmts[i], 0) >> PAGE_SHIFT; |
f7f9f453 RC |
1459 | usergart[i].slot_shift = ilog2((PAGE_SIZE / h) >> i); |
1460 | for (j = 0; j < NUM_USERGART_ENTRIES; j++) { | |
1461 | struct usergart_entry *entry = &usergart[i].entry[j]; | |
1462 | struct tiler_block *block = | |
1463 | tiler_reserve_2d(fmts[i], w, h, | |
1464 | PAGE_SIZE); | |
1465 | if (IS_ERR(block)) { | |
1466 | dev_err(dev->dev, | |
1467 | "reserve failed: %d, %d, %ld\n", | |
1468 | i, j, PTR_ERR(block)); | |
1469 | return; | |
1470 | } | |
1471 | entry->paddr = tiler_ssptr(block); | |
1472 | entry->block = block; | |
1473 | ||
2d31ca3a RK |
1474 | DBG("%d:%d: %dx%d: paddr=%pad stride=%d", i, j, w, h, |
1475 | &entry->paddr, | |
f7f9f453 RC |
1476 | usergart[i].stride_pfn << PAGE_SHIFT); |
1477 | } | |
1478 | } | |
a6a91827 RC |
1479 | |
1480 | priv->has_dmm = true; | |
f7f9f453 RC |
1481 | } |
1482 | ||
1483 | void omap_gem_deinit(struct drm_device *dev) | |
1484 | { | |
1485 | /* I believe we can rely on there being no more outstanding GEM | |
1486 | * objects which could depend on usergart/dmm at this point. | |
1487 | */ | |
f7f9f453 RC |
1488 | kfree(usergart); |
1489 | } |