diff options
-rw-r--r-- | drivers/gpu/drm/ttm/Makefile | 2 | ||||
-rw-r--r-- | drivers/gpu/drm/ttm/ttm_memory.c | 7 | ||||
-rw-r--r-- | drivers/gpu/drm/ttm/ttm_page_alloc.c | 711 | ||||
-rw-r--r-- | drivers/gpu/drm/ttm/ttm_tt.c | 44 | ||||
-rw-r--r-- | include/drm/ttm/ttm_page_alloc.h | 70 |
5 files changed, 809 insertions, 25 deletions
diff --git a/drivers/gpu/drm/ttm/Makefile b/drivers/gpu/drm/ttm/Makefile index 1e138f5bae09..4256e2006476 100644 --- a/drivers/gpu/drm/ttm/Makefile +++ b/drivers/gpu/drm/ttm/Makefile | |||
@@ -4,6 +4,6 @@ | |||
4 | ccflags-y := -Iinclude/drm | 4 | ccflags-y := -Iinclude/drm |
5 | ttm-y := ttm_agp_backend.o ttm_memory.o ttm_tt.o ttm_bo.o \ | 5 | ttm-y := ttm_agp_backend.o ttm_memory.o ttm_tt.o ttm_bo.o \ |
6 | ttm_bo_util.o ttm_bo_vm.o ttm_module.o ttm_global.o \ | 6 | ttm_bo_util.o ttm_bo_vm.o ttm_module.o ttm_global.o \ |
7 | ttm_object.o ttm_lock.o ttm_execbuf_util.o | 7 | ttm_object.o ttm_lock.o ttm_execbuf_util.o ttm_page_alloc.o |
8 | 8 | ||
9 | obj-$(CONFIG_DRM_TTM) += ttm.o | 9 | obj-$(CONFIG_DRM_TTM) += ttm.o |
diff --git a/drivers/gpu/drm/ttm/ttm_memory.c b/drivers/gpu/drm/ttm/ttm_memory.c index c40e5f48e9a1..daff8a87977e 100644 --- a/drivers/gpu/drm/ttm/ttm_memory.c +++ b/drivers/gpu/drm/ttm/ttm_memory.c | |||
@@ -27,6 +27,7 @@ | |||
27 | 27 | ||
28 | #include "ttm/ttm_memory.h" | 28 | #include "ttm/ttm_memory.h" |
29 | #include "ttm/ttm_module.h" | 29 | #include "ttm/ttm_module.h" |
30 | #include "ttm/ttm_page_alloc.h" | ||
30 | #include <linux/spinlock.h> | 31 | #include <linux/spinlock.h> |
31 | #include <linux/sched.h> | 32 | #include <linux/sched.h> |
32 | #include <linux/wait.h> | 33 | #include <linux/wait.h> |
@@ -392,6 +393,7 @@ int ttm_mem_global_init(struct ttm_mem_global *glob) | |||
392 | "Zone %7s: Available graphics memory: %llu kiB.\n", | 393 | "Zone %7s: Available graphics memory: %llu kiB.\n", |
393 | zone->name, (unsigned long long) zone->max_mem >> 10); | 394 | zone->name, (unsigned long long) zone->max_mem >> 10); |
394 | } | 395 | } |
396 | ttm_page_alloc_init(glob->zone_kernel->max_mem/(2*PAGE_SIZE)); | ||
395 | return 0; | 397 | return 0; |
396 | out_no_zone: | 398 | out_no_zone: |
397 | ttm_mem_global_release(glob); | 399 | ttm_mem_global_release(glob); |
@@ -404,6 +406,9 @@ void ttm_mem_global_release(struct ttm_mem_global *glob) | |||
404 | unsigned int i; | 406 | unsigned int i; |
405 | struct ttm_mem_zone *zone; | 407 | struct ttm_mem_zone *zone; |
406 | 408 | ||
409 | /* let the page allocator first stop the shrink work. */ | ||
410 | ttm_page_alloc_fini(); | ||
411 | |||
407 | flush_workqueue(glob->swap_queue); | 412 | flush_workqueue(glob->swap_queue); |
408 | destroy_workqueue(glob->swap_queue); | 413 | destroy_workqueue(glob->swap_queue); |
409 | glob->swap_queue = NULL; | 414 | glob->swap_queue = NULL; |
@@ -411,7 +416,7 @@ void ttm_mem_global_release(struct ttm_mem_global *glob) | |||
411 | zone = glob->zones[i]; | 416 | zone = glob->zones[i]; |
412 | kobject_del(&zone->kobj); | 417 | kobject_del(&zone->kobj); |
413 | kobject_put(&zone->kobj); | 418 | kobject_put(&zone->kobj); |
414 | } | 419 | } |
415 | kobject_del(&glob->kobj); | 420 | kobject_del(&glob->kobj); |
416 | kobject_put(&glob->kobj); | 421 | kobject_put(&glob->kobj); |
417 | } | 422 | } |
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc.c b/drivers/gpu/drm/ttm/ttm_page_alloc.c new file mode 100644 index 000000000000..f46e40be0797 --- /dev/null +++ b/drivers/gpu/drm/ttm/ttm_page_alloc.c | |||
@@ -0,0 +1,711 @@ | |||
1 | /* | ||
2 | * Copyright (c) Red Hat Inc. | ||
3 | |||
4 | * Permission is hereby granted, free of charge, to any person obtaining a | ||
5 | * copy of this software and associated documentation files (the "Software"), | ||
6 | * to deal in the Software without restriction, including without limitation | ||
7 | * the rights to use, copy, modify, merge, publish, distribute, sub license, | ||
8 | * and/or sell copies of the Software, and to permit persons to whom the | ||
9 | * Software is furnished to do so, subject to the following conditions: | ||
10 | * | ||
11 | * The above copyright notice and this permission notice (including the | ||
12 | * next paragraph) shall be included in all copies or substantial portions | ||
13 | * of the Software. | ||
14 | * | ||
15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | ||
16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | ||
17 | * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL | ||
18 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER | ||
19 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING | ||
20 | * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER | ||
21 | * DEALINGS IN THE SOFTWARE. | ||
22 | * | ||
23 | * Authors: Dave Airlie <airlied@redhat.com> | ||
24 | * Jerome Glisse <jglisse@redhat.com> | ||
25 | * Pauli Nieminen <suokkos@gmail.com> | ||
26 | */ | ||
27 | |||
28 | /* simple list based uncached page pool | ||
29 | * - Pool collects resently freed pages for reuse | ||
30 | * - Use page->lru to keep a free list | ||
31 | * - doesn't track currently in use pages | ||
32 | */ | ||
33 | #include <linux/list.h> | ||
34 | #include <linux/spinlock.h> | ||
35 | #include <linux/highmem.h> | ||
36 | #include <linux/mm_types.h> | ||
37 | #include <linux/mm.h> | ||
38 | |||
39 | #include <asm/atomic.h> | ||
40 | #include <asm/agp.h> | ||
41 | |||
42 | #include "ttm/ttm_bo_driver.h" | ||
43 | #include "ttm/ttm_page_alloc.h" | ||
44 | |||
45 | |||
46 | #define NUM_PAGES_TO_ALLOC (PAGE_SIZE/sizeof(struct page *)) | ||
47 | #define SMALL_ALLOCATION 16 | ||
48 | #define FREE_ALL_PAGES (~0U) | ||
49 | /* times are in msecs */ | ||
50 | #define PAGE_FREE_INTERVAL 1000 | ||
51 | |||
52 | /** | ||
53 | * struct ttm_page_pool - Pool to reuse recently allocated uc/wc pages. | ||
54 | * | ||
55 | * @lock: Protects the shared pool from concurrnet access. Must be used with | ||
56 | * irqsave/irqrestore variants because pool allocator maybe called from | ||
57 | * delayed work. | ||
58 | * @fill_lock: Prevent concurrent calls to fill. | ||
59 | * @list: Pool of free uc/wc pages for fast reuse. | ||
60 | * @gfp_flags: Flags to pass for alloc_page. | ||
61 | * @npages: Number of pages in pool. | ||
62 | */ | ||
63 | struct ttm_page_pool { | ||
64 | spinlock_t lock; | ||
65 | bool fill_lock; | ||
66 | struct list_head list; | ||
67 | int gfp_flags; | ||
68 | unsigned npages; | ||
69 | }; | ||
70 | |||
71 | struct ttm_pool_opts { | ||
72 | unsigned alloc_size; | ||
73 | unsigned max_size; | ||
74 | unsigned small; | ||
75 | }; | ||
76 | |||
77 | #define NUM_POOLS 4 | ||
78 | |||
79 | /** | ||
80 | * struct ttm_pool_manager - Holds memory pools for fst allocation | ||
81 | * | ||
82 | * Manager is read only object for pool code so it doesn't need locking. | ||
83 | * | ||
84 | * @free_interval: minimum number of jiffies between freeing pages from pool. | ||
85 | * @page_alloc_inited: reference counting for pool allocation. | ||
86 | * @work: Work that is used to shrink the pool. Work is only run when there is | ||
87 | * some pages to free. | ||
88 | * @small_allocation: Limit in number of pages what is small allocation. | ||
89 | * | ||
90 | * @pools: All pool objects in use. | ||
91 | **/ | ||
92 | struct ttm_pool_manager { | ||
93 | struct shrinker mm_shrink; | ||
94 | atomic_t page_alloc_inited; | ||
95 | struct ttm_pool_opts options; | ||
96 | |||
97 | union { | ||
98 | struct ttm_page_pool pools[NUM_POOLS]; | ||
99 | struct { | ||
100 | struct ttm_page_pool wc_pool; | ||
101 | struct ttm_page_pool uc_pool; | ||
102 | struct ttm_page_pool wc_pool_dma32; | ||
103 | struct ttm_page_pool uc_pool_dma32; | ||
104 | } ; | ||
105 | }; | ||
106 | }; | ||
107 | |||
108 | static struct ttm_pool_manager _manager = { | ||
109 | .page_alloc_inited = ATOMIC_INIT(0) | ||
110 | }; | ||
111 | |||
112 | #ifdef CONFIG_X86 | ||
113 | /* TODO: add this to x86 like _uc, this version here is inefficient */ | ||
114 | static int set_pages_array_wc(struct page **pages, int addrinarray) | ||
115 | { | ||
116 | int i; | ||
117 | |||
118 | for (i = 0; i < addrinarray; i++) | ||
119 | set_memory_wc((unsigned long)page_address(pages[i]), 1); | ||
120 | return 0; | ||
121 | } | ||
122 | #else | ||
123 | static int set_pages_array_wb(struct page **pages, int addrinarray) | ||
124 | { | ||
125 | #ifdef TTM_HAS_AGP | ||
126 | int i; | ||
127 | |||
128 | for (i = 0; i < addrinarray; i++) | ||
129 | unmap_page_from_agp(pages[i]); | ||
130 | #endif | ||
131 | return 0; | ||
132 | } | ||
133 | |||
134 | static int set_pages_array_wc(struct page **pages, int addrinarray) | ||
135 | { | ||
136 | #ifdef TTM_HAS_AGP | ||
137 | int i; | ||
138 | |||
139 | for (i = 0; i < addrinarray; i++) | ||
140 | map_page_into_agp(pages[i]); | ||
141 | #endif | ||
142 | return 0; | ||
143 | } | ||
144 | |||
145 | static int set_pages_array_uc(struct page **pages, int addrinarray) | ||
146 | { | ||
147 | #ifdef TTM_HAS_AGP | ||
148 | int i; | ||
149 | |||
150 | for (i = 0; i < addrinarray; i++) | ||
151 | map_page_into_agp(pages[i]); | ||
152 | #endif | ||
153 | return 0; | ||
154 | } | ||
155 | #endif | ||
156 | |||
157 | /** | ||
158 | * Select the right pool or requested caching state and ttm flags. */ | ||
159 | static struct ttm_page_pool *ttm_get_pool(int flags, | ||
160 | enum ttm_caching_state cstate) | ||
161 | { | ||
162 | int pool_index; | ||
163 | |||
164 | if (cstate == tt_cached) | ||
165 | return NULL; | ||
166 | |||
167 | if (cstate == tt_wc) | ||
168 | pool_index = 0x0; | ||
169 | else | ||
170 | pool_index = 0x1; | ||
171 | |||
172 | if (flags & TTM_PAGE_FLAG_DMA32) | ||
173 | pool_index |= 0x2; | ||
174 | |||
175 | return &_manager.pools[pool_index]; | ||
176 | } | ||
177 | |||
178 | /* set memory back to wb and free the pages. */ | ||
179 | static void ttm_pages_put(struct page *pages[], unsigned npages) | ||
180 | { | ||
181 | unsigned i; | ||
182 | if (set_pages_array_wb(pages, npages)) | ||
183 | printk(KERN_ERR "[ttm] Failed to set %d pages to wb!\n", | ||
184 | npages); | ||
185 | for (i = 0; i < npages; ++i) | ||
186 | __free_page(pages[i]); | ||
187 | } | ||
188 | |||
189 | static void ttm_pool_update_free_locked(struct ttm_page_pool *pool, | ||
190 | unsigned freed_pages) | ||
191 | { | ||
192 | pool->npages -= freed_pages; | ||
193 | } | ||
194 | |||
195 | /** | ||
196 | * Free pages from pool. | ||
197 | * | ||
198 | * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC | ||
199 | * number of pages in one go. | ||
200 | * | ||
201 | * @pool: to free the pages from | ||
202 | * @free_all: If set to true will free all pages in pool | ||
203 | **/ | ||
204 | static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free) | ||
205 | { | ||
206 | unsigned long irq_flags; | ||
207 | struct page *p; | ||
208 | struct page **pages_to_free; | ||
209 | unsigned freed_pages = 0, | ||
210 | npages_to_free = nr_free; | ||
211 | |||
212 | if (NUM_PAGES_TO_ALLOC < nr_free) | ||
213 | npages_to_free = NUM_PAGES_TO_ALLOC; | ||
214 | |||
215 | pages_to_free = kmalloc(npages_to_free * sizeof(struct page *), | ||
216 | GFP_KERNEL); | ||
217 | if (!pages_to_free) { | ||
218 | printk(KERN_ERR "Failed to allocate memory for pool free operation.\n"); | ||
219 | return 0; | ||
220 | } | ||
221 | |||
222 | restart: | ||
223 | spin_lock_irqsave(&pool->lock, irq_flags); | ||
224 | |||
225 | list_for_each_entry_reverse(p, &pool->list, lru) { | ||
226 | if (freed_pages >= npages_to_free) | ||
227 | break; | ||
228 | |||
229 | pages_to_free[freed_pages++] = p; | ||
230 | /* We can only remove NUM_PAGES_TO_ALLOC at a time. */ | ||
231 | if (freed_pages >= NUM_PAGES_TO_ALLOC) { | ||
232 | /* remove range of pages from the pool */ | ||
233 | __list_del(p->lru.prev, &pool->list); | ||
234 | |||
235 | ttm_pool_update_free_locked(pool, freed_pages); | ||
236 | /** | ||
237 | * Because changing page caching is costly | ||
238 | * we unlock the pool to prevent stalling. | ||
239 | */ | ||
240 | spin_unlock_irqrestore(&pool->lock, irq_flags); | ||
241 | |||
242 | ttm_pages_put(pages_to_free, freed_pages); | ||
243 | if (likely(nr_free != FREE_ALL_PAGES)) | ||
244 | nr_free -= freed_pages; | ||
245 | |||
246 | if (NUM_PAGES_TO_ALLOC >= nr_free) | ||
247 | npages_to_free = nr_free; | ||
248 | else | ||
249 | npages_to_free = NUM_PAGES_TO_ALLOC; | ||
250 | |||
251 | freed_pages = 0; | ||
252 | |||
253 | /* free all so restart the processing */ | ||
254 | if (nr_free) | ||
255 | goto restart; | ||
256 | |||
257 | /* Not allowed to fall tough or break because | ||
258 | * following context is inside spinlock while we are | ||
259 | * outside here. | ||
260 | */ | ||
261 | goto out; | ||
262 | |||
263 | } | ||
264 | } | ||
265 | |||
266 | |||
267 | /* remove range of pages from the pool */ | ||
268 | if (freed_pages) { | ||
269 | __list_del(&p->lru, &pool->list); | ||
270 | |||
271 | ttm_pool_update_free_locked(pool, freed_pages); | ||
272 | nr_free -= freed_pages; | ||
273 | } | ||
274 | |||
275 | spin_unlock_irqrestore(&pool->lock, irq_flags); | ||
276 | |||
277 | if (freed_pages) | ||
278 | ttm_pages_put(pages_to_free, freed_pages); | ||
279 | out: | ||
280 | kfree(pages_to_free); | ||
281 | return nr_free; | ||
282 | } | ||
283 | |||
284 | /* Get good estimation how many pages are free in pools */ | ||
285 | static int ttm_pool_get_num_unused_pages(void) | ||
286 | { | ||
287 | unsigned i; | ||
288 | int total = 0; | ||
289 | for (i = 0; i < NUM_POOLS; ++i) | ||
290 | total += _manager.pools[i].npages; | ||
291 | |||
292 | return total; | ||
293 | } | ||
294 | |||
295 | /** | ||
296 | * Calback for mm to request pool to reduce number of page held. | ||
297 | */ | ||
298 | static int ttm_pool_mm_shrink(int shrink_pages, gfp_t gfp_mask) | ||
299 | { | ||
300 | static atomic_t start_pool = ATOMIC_INIT(0); | ||
301 | unsigned i; | ||
302 | unsigned pool_offset = atomic_add_return(1, &start_pool); | ||
303 | struct ttm_page_pool *pool; | ||
304 | |||
305 | pool_offset = pool_offset % NUM_POOLS; | ||
306 | /* select start pool in round robin fashion */ | ||
307 | for (i = 0; i < NUM_POOLS; ++i) { | ||
308 | unsigned nr_free = shrink_pages; | ||
309 | if (shrink_pages == 0) | ||
310 | break; | ||
311 | pool = &_manager.pools[(i + pool_offset)%NUM_POOLS]; | ||
312 | shrink_pages = ttm_page_pool_free(pool, nr_free); | ||
313 | } | ||
314 | /* return estimated number of unused pages in pool */ | ||
315 | return ttm_pool_get_num_unused_pages(); | ||
316 | } | ||
317 | |||
318 | static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager) | ||
319 | { | ||
320 | manager->mm_shrink.shrink = &ttm_pool_mm_shrink; | ||
321 | manager->mm_shrink.seeks = 1; | ||
322 | register_shrinker(&manager->mm_shrink); | ||
323 | } | ||
324 | |||
325 | static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager) | ||
326 | { | ||
327 | unregister_shrinker(&manager->mm_shrink); | ||
328 | } | ||
329 | |||
330 | static int ttm_set_pages_caching(struct page **pages, | ||
331 | enum ttm_caching_state cstate, unsigned cpages) | ||
332 | { | ||
333 | int r = 0; | ||
334 | /* Set page caching */ | ||
335 | switch (cstate) { | ||
336 | case tt_uncached: | ||
337 | r = set_pages_array_uc(pages, cpages); | ||
338 | if (r) | ||
339 | printk(KERN_ERR "[ttm] Failed to set %d pages to uc!\n", | ||
340 | cpages); | ||
341 | break; | ||
342 | case tt_wc: | ||
343 | r = set_pages_array_wc(pages, cpages); | ||
344 | if (r) | ||
345 | printk(KERN_ERR "[ttm] Failed to set %d pages to wc!\n", | ||
346 | cpages); | ||
347 | break; | ||
348 | default: | ||
349 | break; | ||
350 | } | ||
351 | return r; | ||
352 | } | ||
353 | |||
354 | /** | ||
355 | * Free pages the pages that failed to change the caching state. If there is | ||
356 | * any pages that have changed their caching state already put them to the | ||
357 | * pool. | ||
358 | */ | ||
359 | static void ttm_handle_caching_state_failure(struct list_head *pages, | ||
360 | int ttm_flags, enum ttm_caching_state cstate, | ||
361 | struct page **failed_pages, unsigned cpages) | ||
362 | { | ||
363 | unsigned i; | ||
364 | /* Failed pages has to be reed */ | ||
365 | for (i = 0; i < cpages; ++i) { | ||
366 | list_del(&failed_pages[i]->lru); | ||
367 | __free_page(failed_pages[i]); | ||
368 | } | ||
369 | } | ||
370 | |||
371 | /** | ||
372 | * Allocate new pages with correct caching. | ||
373 | * | ||
374 | * This function is reentrant if caller updates count depending on number of | ||
375 | * pages returned in pages array. | ||
376 | */ | ||
377 | static int ttm_alloc_new_pages(struct list_head *pages, int gfp_flags, | ||
378 | int ttm_flags, enum ttm_caching_state cstate, unsigned count) | ||
379 | { | ||
380 | struct page **caching_array; | ||
381 | struct page *p; | ||
382 | int r = 0; | ||
383 | unsigned i, cpages; | ||
384 | unsigned max_cpages = min(count, | ||
385 | (unsigned)(PAGE_SIZE/sizeof(struct page *))); | ||
386 | |||
387 | /* allocate array for page caching change */ | ||
388 | caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL); | ||
389 | |||
390 | if (!caching_array) { | ||
391 | printk(KERN_ERR "[ttm] unable to allocate table for new pages."); | ||
392 | return -ENOMEM; | ||
393 | } | ||
394 | |||
395 | for (i = 0, cpages = 0; i < count; ++i) { | ||
396 | p = alloc_page(gfp_flags); | ||
397 | |||
398 | if (!p) { | ||
399 | printk(KERN_ERR "[ttm] unable to get page %u\n", i); | ||
400 | |||
401 | /* store already allocated pages in the pool after | ||
402 | * setting the caching state */ | ||
403 | if (cpages) { | ||
404 | r = ttm_set_pages_caching(caching_array, cstate, cpages); | ||
405 | if (r) | ||
406 | ttm_handle_caching_state_failure(pages, | ||
407 | ttm_flags, cstate, | ||
408 | caching_array, cpages); | ||
409 | } | ||
410 | r = -ENOMEM; | ||
411 | goto out; | ||
412 | } | ||
413 | |||
414 | #ifdef CONFIG_HIGHMEM | ||
415 | /* gfp flags of highmem page should never be dma32 so we | ||
416 | * we should be fine in such case | ||
417 | */ | ||
418 | if (!PageHighMem(p)) | ||
419 | #endif | ||
420 | { | ||
421 | caching_array[cpages++] = p; | ||
422 | if (cpages == max_cpages) { | ||
423 | |||
424 | r = ttm_set_pages_caching(caching_array, | ||
425 | cstate, cpages); | ||
426 | if (r) { | ||
427 | ttm_handle_caching_state_failure(pages, | ||
428 | ttm_flags, cstate, | ||
429 | caching_array, cpages); | ||
430 | goto out; | ||
431 | } | ||
432 | cpages = 0; | ||
433 | } | ||
434 | } | ||
435 | |||
436 | list_add(&p->lru, pages); | ||
437 | } | ||
438 | |||
439 | if (cpages) { | ||
440 | r = ttm_set_pages_caching(caching_array, cstate, cpages); | ||
441 | if (r) | ||
442 | ttm_handle_caching_state_failure(pages, | ||
443 | ttm_flags, cstate, | ||
444 | caching_array, cpages); | ||
445 | } | ||
446 | out: | ||
447 | kfree(caching_array); | ||
448 | |||
449 | return r; | ||
450 | } | ||
451 | |||
452 | /** | ||
453 | * Fill the given pool if there isn't enough pages and requested number of | ||
454 | * pages is small. | ||
455 | */ | ||
456 | static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool, | ||
457 | int ttm_flags, enum ttm_caching_state cstate, unsigned count, | ||
458 | unsigned long *irq_flags) | ||
459 | { | ||
460 | struct page *p; | ||
461 | int r; | ||
462 | unsigned cpages = 0; | ||
463 | /** | ||
464 | * Only allow one pool fill operation at a time. | ||
465 | * If pool doesn't have enough pages for the allocation new pages are | ||
466 | * allocated from outside of pool. | ||
467 | */ | ||
468 | if (pool->fill_lock) | ||
469 | return; | ||
470 | |||
471 | pool->fill_lock = true; | ||
472 | |||
473 | /* If allocation request is small and there is not enough | ||
474 | * pages in pool we fill the pool first */ | ||
475 | if (count < _manager.options.small | ||
476 | && count > pool->npages) { | ||
477 | struct list_head new_pages; | ||
478 | unsigned alloc_size = _manager.options.alloc_size; | ||
479 | |||
480 | /** | ||
481 | * Can't change page caching if in irqsave context. We have to | ||
482 | * drop the pool->lock. | ||
483 | */ | ||
484 | spin_unlock_irqrestore(&pool->lock, *irq_flags); | ||
485 | |||
486 | INIT_LIST_HEAD(&new_pages); | ||
487 | r = ttm_alloc_new_pages(&new_pages, pool->gfp_flags, ttm_flags, | ||
488 | cstate, alloc_size); | ||
489 | spin_lock_irqsave(&pool->lock, *irq_flags); | ||
490 | |||
491 | if (!r) { | ||
492 | list_splice(&new_pages, &pool->list); | ||
493 | pool->npages += alloc_size; | ||
494 | } else { | ||
495 | printk(KERN_ERR "[ttm] Failed to fill pool (%p).", pool); | ||
496 | /* If we have any pages left put them to the pool. */ | ||
497 | list_for_each_entry(p, &pool->list, lru) { | ||
498 | ++cpages; | ||
499 | } | ||
500 | list_splice(&new_pages, &pool->list); | ||
501 | pool->npages += cpages; | ||
502 | } | ||
503 | |||
504 | } | ||
505 | pool->fill_lock = false; | ||
506 | } | ||
507 | |||
508 | /** | ||
509 | * Cut count nubmer of pages from the pool and put them to return list | ||
510 | * | ||
511 | * @return count of pages still to allocate to fill the request. | ||
512 | */ | ||
513 | static unsigned ttm_page_pool_get_pages(struct ttm_page_pool *pool, | ||
514 | struct list_head *pages, int ttm_flags, | ||
515 | enum ttm_caching_state cstate, unsigned count) | ||
516 | { | ||
517 | unsigned long irq_flags; | ||
518 | struct list_head *p; | ||
519 | unsigned i; | ||
520 | |||
521 | spin_lock_irqsave(&pool->lock, irq_flags); | ||
522 | ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count, &irq_flags); | ||
523 | |||
524 | if (count >= pool->npages) { | ||
525 | /* take all pages from the pool */ | ||
526 | list_splice_init(&pool->list, pages); | ||
527 | count -= pool->npages; | ||
528 | pool->npages = 0; | ||
529 | goto out; | ||
530 | } | ||
531 | /* find the last pages to include for requested number of pages. Split | ||
532 | * pool to begin and halves to reduce search space. */ | ||
533 | if (count <= pool->npages/2) { | ||
534 | i = 0; | ||
535 | list_for_each(p, &pool->list) { | ||
536 | if (++i == count) | ||
537 | break; | ||
538 | } | ||
539 | } else { | ||
540 | i = pool->npages + 1; | ||
541 | list_for_each_prev(p, &pool->list) { | ||
542 | if (--i == count) | ||
543 | break; | ||
544 | } | ||
545 | } | ||
546 | /* Cut count number of pages from pool */ | ||
547 | list_cut_position(pages, &pool->list, p); | ||
548 | pool->npages -= count; | ||
549 | count = 0; | ||
550 | out: | ||
551 | spin_unlock_irqrestore(&pool->lock, irq_flags); | ||
552 | return count; | ||
553 | } | ||
554 | |||
555 | /* | ||
556 | * On success pages list will hold count number of correctly | ||
557 | * cached pages. | ||
558 | */ | ||
559 | int ttm_get_pages(struct list_head *pages, int flags, | ||
560 | enum ttm_caching_state cstate, unsigned count) | ||
561 | { | ||
562 | struct ttm_page_pool *pool = ttm_get_pool(flags, cstate); | ||
563 | struct page *p = NULL; | ||
564 | int gfp_flags = 0; | ||
565 | int r; | ||
566 | |||
567 | /* set zero flag for page allocation if required */ | ||
568 | if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) | ||
569 | gfp_flags |= __GFP_ZERO; | ||
570 | |||
571 | /* No pool for cached pages */ | ||
572 | if (pool == NULL) { | ||
573 | if (flags & TTM_PAGE_FLAG_DMA32) | ||
574 | gfp_flags |= GFP_DMA32; | ||
575 | else | ||
576 | gfp_flags |= __GFP_HIGHMEM; | ||
577 | |||
578 | for (r = 0; r < count; ++r) { | ||
579 | p = alloc_page(gfp_flags); | ||
580 | if (!p) { | ||
581 | |||
582 | printk(KERN_ERR "[ttm] unable to allocate page."); | ||
583 | return -ENOMEM; | ||
584 | } | ||
585 | |||
586 | list_add(&p->lru, pages); | ||
587 | } | ||
588 | return 0; | ||
589 | } | ||
590 | |||
591 | |||
592 | /* combine zero flag to pool flags */ | ||
593 | gfp_flags |= pool->gfp_flags; | ||
594 | |||
595 | /* First we take pages from the pool */ | ||
596 | count = ttm_page_pool_get_pages(pool, pages, flags, cstate, count); | ||
597 | |||
598 | /* clear the pages coming from the pool if requested */ | ||
599 | if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) { | ||
600 | list_for_each_entry(p, pages, lru) { | ||
601 | clear_page(page_address(p)); | ||
602 | } | ||
603 | } | ||
604 | |||
605 | /* If pool didn't have enough pages allocate new one. */ | ||
606 | if (count > 0) { | ||
607 | /* ttm_alloc_new_pages doesn't reference pool so we can run | ||
608 | * multiple requests in parallel. | ||
609 | **/ | ||
610 | r = ttm_alloc_new_pages(pages, gfp_flags, flags, cstate, count); | ||
611 | if (r) { | ||
612 | /* If there is any pages in the list put them back to | ||
613 | * the pool. */ | ||
614 | printk(KERN_ERR "[ttm] Failed to allocate extra pages " | ||
615 | "for large request."); | ||
616 | ttm_put_pages(pages, 0, flags, cstate); | ||
617 | return r; | ||
618 | } | ||
619 | } | ||
620 | |||
621 | |||
622 | return 0; | ||
623 | } | ||
624 | |||
625 | /* Put all pages in pages list to correct pool to wait for reuse */ | ||
626 | void ttm_put_pages(struct list_head *pages, unsigned page_count, int flags, | ||
627 | enum ttm_caching_state cstate) | ||
628 | { | ||
629 | unsigned long irq_flags; | ||
630 | struct ttm_page_pool *pool = ttm_get_pool(flags, cstate); | ||
631 | struct page *p, *tmp; | ||
632 | |||
633 | if (pool == NULL) { | ||
634 | /* No pool for this memory type so free the pages */ | ||
635 | |||
636 | list_for_each_entry_safe(p, tmp, pages, lru) { | ||
637 | __free_page(p); | ||
638 | } | ||
639 | /* Make the pages list empty */ | ||
640 | INIT_LIST_HEAD(pages); | ||
641 | return; | ||
642 | } | ||
643 | if (page_count == 0) { | ||
644 | list_for_each_entry_safe(p, tmp, pages, lru) { | ||
645 | ++page_count; | ||
646 | } | ||
647 | } | ||
648 | |||
649 | spin_lock_irqsave(&pool->lock, irq_flags); | ||
650 | list_splice_init(pages, &pool->list); | ||
651 | pool->npages += page_count; | ||
652 | /* Check that we don't go over the pool limit */ | ||
653 | page_count = 0; | ||
654 | if (pool->npages > _manager.options.max_size) { | ||
655 | page_count = pool->npages - _manager.options.max_size; | ||
656 | /* free at least NUM_PAGES_TO_ALLOC number of pages | ||
657 | * to reduce calls to set_memory_wb */ | ||
658 | if (page_count < NUM_PAGES_TO_ALLOC) | ||
659 | page_count = NUM_PAGES_TO_ALLOC; | ||
660 | } | ||
661 | spin_unlock_irqrestore(&pool->lock, irq_flags); | ||
662 | if (page_count) | ||
663 | ttm_page_pool_free(pool, page_count); | ||
664 | } | ||
665 | |||
666 | static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, int flags) | ||
667 | { | ||
668 | spin_lock_init(&pool->lock); | ||
669 | pool->fill_lock = false; | ||
670 | INIT_LIST_HEAD(&pool->list); | ||
671 | pool->npages = 0; | ||
672 | pool->gfp_flags = flags; | ||
673 | } | ||
674 | |||
675 | int ttm_page_alloc_init(unsigned max_pages) | ||
676 | { | ||
677 | if (atomic_add_return(1, &_manager.page_alloc_inited) > 1) | ||
678 | return 0; | ||
679 | |||
680 | printk(KERN_INFO "[ttm] Initializing pool allocator.\n"); | ||
681 | |||
682 | ttm_page_pool_init_locked(&_manager.wc_pool, GFP_HIGHUSER); | ||
683 | |||
684 | ttm_page_pool_init_locked(&_manager.uc_pool, GFP_HIGHUSER); | ||
685 | |||
686 | ttm_page_pool_init_locked(&_manager.wc_pool_dma32, GFP_USER | GFP_DMA32); | ||
687 | |||
688 | ttm_page_pool_init_locked(&_manager.uc_pool_dma32, GFP_USER | GFP_DMA32); | ||
689 | |||
690 | _manager.options.max_size = max_pages; | ||
691 | _manager.options.small = SMALL_ALLOCATION; | ||
692 | _manager.options.alloc_size = NUM_PAGES_TO_ALLOC; | ||
693 | |||
694 | ttm_pool_mm_shrink_init(&_manager); | ||
695 | |||
696 | return 0; | ||
697 | } | ||
698 | |||
699 | void ttm_page_alloc_fini() | ||
700 | { | ||
701 | int i; | ||
702 | |||
703 | if (atomic_sub_return(1, &_manager.page_alloc_inited) > 0) | ||
704 | return; | ||
705 | |||
706 | printk(KERN_INFO "[ttm] Finilizing pool allocator.\n"); | ||
707 | ttm_pool_mm_shrink_fini(&_manager); | ||
708 | |||
709 | for (i = 0; i < NUM_POOLS; ++i) | ||
710 | ttm_page_pool_free(&_manager.pools[i], FREE_ALL_PAGES); | ||
711 | } | ||
diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c index bab6cd8d8a1e..a3269ef831c0 100644 --- a/drivers/gpu/drm/ttm/ttm_tt.c +++ b/drivers/gpu/drm/ttm/ttm_tt.c | |||
@@ -38,6 +38,7 @@ | |||
38 | #include "ttm/ttm_module.h" | 38 | #include "ttm/ttm_module.h" |
39 | #include "ttm/ttm_bo_driver.h" | 39 | #include "ttm/ttm_bo_driver.h" |
40 | #include "ttm/ttm_placement.h" | 40 | #include "ttm/ttm_placement.h" |
41 | #include "ttm/ttm_page_alloc.h" | ||
41 | 42 | ||
42 | static int ttm_tt_swapin(struct ttm_tt *ttm); | 43 | static int ttm_tt_swapin(struct ttm_tt *ttm); |
43 | 44 | ||
@@ -55,21 +56,6 @@ static void ttm_tt_free_page_directory(struct ttm_tt *ttm) | |||
55 | ttm->pages = NULL; | 56 | ttm->pages = NULL; |
56 | } | 57 | } |
57 | 58 | ||
58 | static struct page *ttm_tt_alloc_page(unsigned page_flags) | ||
59 | { | ||
60 | gfp_t gfp_flags = GFP_USER; | ||
61 | |||
62 | if (page_flags & TTM_PAGE_FLAG_ZERO_ALLOC) | ||
63 | gfp_flags |= __GFP_ZERO; | ||
64 | |||
65 | if (page_flags & TTM_PAGE_FLAG_DMA32) | ||
66 | gfp_flags |= __GFP_DMA32; | ||
67 | else | ||
68 | gfp_flags |= __GFP_HIGHMEM; | ||
69 | |||
70 | return alloc_page(gfp_flags); | ||
71 | } | ||
72 | |||
73 | static void ttm_tt_free_user_pages(struct ttm_tt *ttm) | 59 | static void ttm_tt_free_user_pages(struct ttm_tt *ttm) |
74 | { | 60 | { |
75 | int write; | 61 | int write; |
@@ -110,15 +96,21 @@ static void ttm_tt_free_user_pages(struct ttm_tt *ttm) | |||
110 | static struct page *__ttm_tt_get_page(struct ttm_tt *ttm, int index) | 96 | static struct page *__ttm_tt_get_page(struct ttm_tt *ttm, int index) |
111 | { | 97 | { |
112 | struct page *p; | 98 | struct page *p; |
99 | struct list_head h; | ||
113 | struct ttm_mem_global *mem_glob = ttm->glob->mem_glob; | 100 | struct ttm_mem_global *mem_glob = ttm->glob->mem_glob; |
114 | int ret; | 101 | int ret; |
115 | 102 | ||
116 | while (NULL == (p = ttm->pages[index])) { | 103 | while (NULL == (p = ttm->pages[index])) { |
117 | p = ttm_tt_alloc_page(ttm->page_flags); | ||
118 | 104 | ||
119 | if (!p) | 105 | INIT_LIST_HEAD(&h); |
106 | |||
107 | ret = ttm_get_pages(&h, ttm->page_flags, ttm->caching_state, 1); | ||
108 | |||
109 | if (ret != 0) | ||
120 | return NULL; | 110 | return NULL; |
121 | 111 | ||
112 | p = list_first_entry(&h, struct page, lru); | ||
113 | |||
122 | ret = ttm_mem_global_alloc_page(mem_glob, p, false, false); | 114 | ret = ttm_mem_global_alloc_page(mem_glob, p, false, false); |
123 | if (unlikely(ret != 0)) | 115 | if (unlikely(ret != 0)) |
124 | goto out_err; | 116 | goto out_err; |
@@ -227,10 +219,10 @@ static int ttm_tt_set_caching(struct ttm_tt *ttm, | |||
227 | if (ttm->caching_state == c_state) | 219 | if (ttm->caching_state == c_state) |
228 | return 0; | 220 | return 0; |
229 | 221 | ||
230 | if (c_state != tt_cached) { | 222 | if (ttm->state == tt_unpopulated) { |
231 | ret = ttm_tt_populate(ttm); | 223 | /* Change caching but don't populate */ |
232 | if (unlikely(ret != 0)) | 224 | ttm->caching_state = c_state; |
233 | return ret; | 225 | return 0; |
234 | } | 226 | } |
235 | 227 | ||
236 | if (ttm->caching_state == tt_cached) | 228 | if (ttm->caching_state == tt_cached) |
@@ -281,13 +273,17 @@ EXPORT_SYMBOL(ttm_tt_set_placement_caching); | |||
281 | static void ttm_tt_free_alloced_pages(struct ttm_tt *ttm) | 273 | static void ttm_tt_free_alloced_pages(struct ttm_tt *ttm) |
282 | { | 274 | { |
283 | int i; | 275 | int i; |
276 | unsigned count = 0; | ||
277 | struct list_head h; | ||
284 | struct page *cur_page; | 278 | struct page *cur_page; |
285 | struct ttm_backend *be = ttm->be; | 279 | struct ttm_backend *be = ttm->be; |
286 | 280 | ||
281 | INIT_LIST_HEAD(&h); | ||
282 | |||
287 | if (be) | 283 | if (be) |
288 | be->func->clear(be); | 284 | be->func->clear(be); |
289 | (void)ttm_tt_set_caching(ttm, tt_cached); | ||
290 | for (i = 0; i < ttm->num_pages; ++i) { | 285 | for (i = 0; i < ttm->num_pages; ++i) { |
286 | |||
291 | cur_page = ttm->pages[i]; | 287 | cur_page = ttm->pages[i]; |
292 | ttm->pages[i] = NULL; | 288 | ttm->pages[i] = NULL; |
293 | if (cur_page) { | 289 | if (cur_page) { |
@@ -297,9 +293,11 @@ static void ttm_tt_free_alloced_pages(struct ttm_tt *ttm) | |||
297 | "Leaking pages.\n"); | 293 | "Leaking pages.\n"); |
298 | ttm_mem_global_free_page(ttm->glob->mem_glob, | 294 | ttm_mem_global_free_page(ttm->glob->mem_glob, |
299 | cur_page); | 295 | cur_page); |
300 | __free_page(cur_page); | 296 | list_add(&cur_page->lru, &h); |
297 | count++; | ||
301 | } | 298 | } |
302 | } | 299 | } |
300 | ttm_put_pages(&h, count, ttm->page_flags, ttm->caching_state); | ||
303 | ttm->state = tt_unpopulated; | 301 | ttm->state = tt_unpopulated; |
304 | ttm->first_himem_page = ttm->num_pages; | 302 | ttm->first_himem_page = ttm->num_pages; |
305 | ttm->last_lomem_page = -1; | 303 | ttm->last_lomem_page = -1; |
diff --git a/include/drm/ttm/ttm_page_alloc.h b/include/drm/ttm/ttm_page_alloc.h new file mode 100644 index 000000000000..043d817b8164 --- /dev/null +++ b/include/drm/ttm/ttm_page_alloc.h | |||
@@ -0,0 +1,70 @@ | |||
1 | /* | ||
2 | * Copyright (c) Red Hat Inc. | ||
3 | |||
4 | * Permission is hereby granted, free of charge, to any person obtaining a | ||
5 | * copy of this software and associated documentation files (the "Software"), | ||
6 | * to deal in the Software without restriction, including without limitation | ||
7 | * the rights to use, copy, modify, merge, publish, distribute, sub license, | ||
8 | * and/or sell copies of the Software, and to permit persons to whom the | ||
9 | * Software is furnished to do so, subject to the following conditions: | ||
10 | * | ||
11 | * The above copyright notice and this permission notice (including the | ||
12 | * next paragraph) shall be included in all copies or substantial portions | ||
13 | * of the Software. | ||
14 | * | ||
15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | ||
16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | ||
17 | * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL | ||
18 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER | ||
19 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING | ||
20 | * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER | ||
21 | * DEALINGS IN THE SOFTWARE. | ||
22 | * | ||
23 | * Authors: Dave Airlie <airlied@redhat.com> | ||
24 | * Jerome Glisse <jglisse@redhat.com> | ||
25 | */ | ||
26 | #ifndef TTM_PAGE_ALLOC | ||
27 | #define TTM_PAGE_ALLOC | ||
28 | |||
29 | #include "ttm_bo_driver.h" | ||
30 | #include "ttm_memory.h" | ||
31 | |||
32 | /** | ||
33 | * Get count number of pages from pool to pages list. | ||
34 | * | ||
35 | * @pages: heado of empty linked list where pages are filled. | ||
36 | * @flags: ttm flags for page allocation. | ||
37 | * @cstate: ttm caching state for the page. | ||
38 | * @count: number of pages to allocate. | ||
39 | */ | ||
40 | int ttm_get_pages(struct list_head *pages, | ||
41 | int flags, | ||
42 | enum ttm_caching_state cstate, | ||
43 | unsigned count); | ||
44 | /** | ||
45 | * Put linked list of pages to pool. | ||
46 | * | ||
47 | * @pages: list of pages to free. | ||
48 | * @page_count: number of pages in the list. Zero can be passed for unknown | ||
49 | * count. | ||
50 | * @flags: ttm flags for page allocation. | ||
51 | * @cstate: ttm caching state. | ||
52 | */ | ||
53 | void ttm_put_pages(struct list_head *pages, | ||
54 | unsigned page_count, | ||
55 | int flags, | ||
56 | enum ttm_caching_state cstate); | ||
57 | /** | ||
58 | * Initialize pool allocator. | ||
59 | * | ||
60 | * Pool allocator is internaly reference counted so it can be initialized | ||
61 | * multiple times but ttm_page_alloc_fini has to be called same number of | ||
62 | * times. | ||
63 | */ | ||
64 | int ttm_page_alloc_init(unsigned max_pages); | ||
65 | /** | ||
66 | * Free pool allocator. | ||
67 | */ | ||
68 | void ttm_page_alloc_fini(void); | ||
69 | |||
70 | #endif | ||