1 /**************************************************************************
3 * Copyright (c) 2007-2009 VMware, Inc., Palo Alto, CA., USA
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
28 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
31 #include "ttm/ttm_bo_driver.h"
32 #include "ttm/ttm_placement.h"
34 #include <linux/highmem.h>
35 #include <linux/wait.h>
36 #include <linux/vmalloc.h>
37 #include <linux/module.h>
39 void ttm_bo_free_old_node(struct ttm_buffer_object *bo)
41 struct ttm_mem_reg *old_mem = &bo->mem;
43 if (old_mem->mm_node) {
44 spin_lock(&bo->glob->lru_lock);
45 drm_mm_put_block(old_mem->mm_node);
46 spin_unlock(&bo->glob->lru_lock);
48 old_mem->mm_node = NULL;
51 int ttm_bo_move_ttm(struct ttm_buffer_object *bo,
52 bool evict, bool no_wait_reserve,
53 bool no_wait_gpu, struct ttm_mem_reg *new_mem)
55 struct ttm_tt *ttm = bo->ttm;
56 struct ttm_mem_reg *old_mem = &bo->mem;
59 if (old_mem->mem_type != TTM_PL_SYSTEM) {
61 ttm_bo_free_old_node(bo);
62 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM,
64 old_mem->mem_type = TTM_PL_SYSTEM;
67 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement);
68 if (unlikely(ret != 0))
71 if (new_mem->mem_type != TTM_PL_SYSTEM) {
72 ret = ttm_tt_bind(ttm, new_mem);
73 if (unlikely(ret != 0))
78 new_mem->mm_node = NULL;
82 EXPORT_SYMBOL(ttm_bo_move_ttm);
84 int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
87 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
88 unsigned long bus_offset;
89 unsigned long bus_size;
90 unsigned long bus_base;
95 ret = ttm_bo_pci_offset(bdev, mem, &bus_base, &bus_offset, &bus_size);
96 if (ret || bus_size == 0)
99 if (!(man->flags & TTM_MEMTYPE_FLAG_NEEDS_IOREMAP))
100 addr = (void *)(((u8 *) man->io_addr) + bus_offset);
102 if (mem->placement & TTM_PL_FLAG_WC)
103 addr = ioremap_wc(bus_base + bus_offset, bus_size);
105 addr = ioremap_nocache(bus_base + bus_offset, bus_size);
113 void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
116 struct ttm_mem_type_manager *man;
118 man = &bdev->man[mem->mem_type];
120 if (virtual && (man->flags & TTM_MEMTYPE_FLAG_NEEDS_IOREMAP))
124 static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
127 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
129 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
132 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
133 iowrite32(ioread32(srcP++), dstP++);
137 static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
141 struct page *d = ttm_tt_get_page(ttm, page);
147 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
150 dst = kmap_atomic_prot(d, KM_USER0, prot);
152 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
153 dst = vmap(&d, 1, 0, prot);
160 memcpy_fromio(dst, src, PAGE_SIZE);
163 kunmap_atomic(dst, KM_USER0);
165 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
174 static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
178 struct page *s = ttm_tt_get_page(ttm, page);
184 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
186 src = kmap_atomic_prot(s, KM_USER0, prot);
188 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
189 src = vmap(&s, 1, 0, prot);
196 memcpy_toio(dst, src, PAGE_SIZE);
199 kunmap_atomic(src, KM_USER0);
201 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
210 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
211 bool evict, bool no_wait_reserve, bool no_wait_gpu,
212 struct ttm_mem_reg *new_mem)
214 struct ttm_bo_device *bdev = bo->bdev;
215 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
216 struct ttm_tt *ttm = bo->ttm;
217 struct ttm_mem_reg *old_mem = &bo->mem;
218 struct ttm_mem_reg old_copy = *old_mem;
224 unsigned long add = 0;
227 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap);
230 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap);
234 if (old_iomap == NULL && new_iomap == NULL)
236 if (old_iomap == NULL && ttm == NULL)
242 if ((old_mem->mem_type == new_mem->mem_type) &&
243 (new_mem->mm_node->start <
244 old_mem->mm_node->start + old_mem->mm_node->size)) {
246 add = new_mem->num_pages - 1;
249 for (i = 0; i < new_mem->num_pages; ++i) {
250 page = i * dir + add;
251 if (old_iomap == NULL) {
252 pgprot_t prot = ttm_io_prot(old_mem->placement,
254 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
256 } else if (new_iomap == NULL) {
257 pgprot_t prot = ttm_io_prot(new_mem->placement,
259 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
262 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
268 ttm_bo_free_old_node(bo);
271 new_mem->mm_node = NULL;
273 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && (ttm != NULL)) {
280 ttm_mem_reg_iounmap(bdev, new_mem, new_iomap);
282 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap);
285 EXPORT_SYMBOL(ttm_bo_move_memcpy);
287 static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
293 * ttm_buffer_object_transfer
295 * @bo: A pointer to a struct ttm_buffer_object.
296 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
297 * holding the data of @bo with the old placement.
299 * This is a utility function that may be called after an accelerated move
300 * has been scheduled. A new buffer object is created as a placeholder for
301 * the old data while it's being copied. When that buffer object is idle,
302 * it can be destroyed, releasing the space of the old placement.
307 static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
308 struct ttm_buffer_object **new_obj)
310 struct ttm_buffer_object *fbo;
311 struct ttm_bo_device *bdev = bo->bdev;
312 struct ttm_bo_driver *driver = bdev->driver;
314 fbo = kzalloc(sizeof(*fbo), GFP_KERNEL);
321 * Fix up members that we shouldn't copy directly:
322 * TODO: Explicit member copy would probably be better here.
325 spin_lock_init(&fbo->lock);
326 init_waitqueue_head(&fbo->event_queue);
327 INIT_LIST_HEAD(&fbo->ddestroy);
328 INIT_LIST_HEAD(&fbo->lru);
329 INIT_LIST_HEAD(&fbo->swap);
332 fbo->sync_obj = driver->sync_obj_ref(bo->sync_obj);
333 if (fbo->mem.mm_node)
334 fbo->mem.mm_node->private = (void *)fbo;
335 kref_init(&fbo->list_kref);
336 kref_init(&fbo->kref);
337 fbo->destroy = &ttm_transfered_destroy;
343 pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
345 #if defined(__i386__) || defined(__x86_64__)
346 if (caching_flags & TTM_PL_FLAG_WC)
347 tmp = pgprot_writecombine(tmp);
348 else if (boot_cpu_data.x86 > 3)
349 tmp = pgprot_noncached(tmp);
351 #elif defined(__powerpc__)
352 if (!(caching_flags & TTM_PL_FLAG_CACHED)) {
353 pgprot_val(tmp) |= _PAGE_NO_CACHE;
354 if (caching_flags & TTM_PL_FLAG_UNCACHED)
355 pgprot_val(tmp) |= _PAGE_GUARDED;
358 #if defined(__ia64__)
359 if (caching_flags & TTM_PL_FLAG_WC)
360 tmp = pgprot_writecombine(tmp);
362 tmp = pgprot_noncached(tmp);
364 #if defined(__sparc__)
365 if (!(caching_flags & TTM_PL_FLAG_CACHED))
366 tmp = pgprot_noncached(tmp);
370 EXPORT_SYMBOL(ttm_io_prot);
372 static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
373 unsigned long bus_base,
374 unsigned long bus_offset,
375 unsigned long bus_size,
376 struct ttm_bo_kmap_obj *map)
378 struct ttm_bo_device *bdev = bo->bdev;
379 struct ttm_mem_reg *mem = &bo->mem;
380 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
382 if (!(man->flags & TTM_MEMTYPE_FLAG_NEEDS_IOREMAP)) {
383 map->bo_kmap_type = ttm_bo_map_premapped;
384 map->virtual = (void *)(((u8 *) man->io_addr) + bus_offset);
386 map->bo_kmap_type = ttm_bo_map_iomap;
387 if (mem->placement & TTM_PL_FLAG_WC)
388 map->virtual = ioremap_wc(bus_base + bus_offset,
391 map->virtual = ioremap_nocache(bus_base + bus_offset,
394 return (!map->virtual) ? -ENOMEM : 0;
397 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
398 unsigned long start_page,
399 unsigned long num_pages,
400 struct ttm_bo_kmap_obj *map)
402 struct ttm_mem_reg *mem = &bo->mem; pgprot_t prot;
403 struct ttm_tt *ttm = bo->ttm;
408 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
410 * We're mapping a single page, and the desired
411 * page protection is consistent with the bo.
414 map->bo_kmap_type = ttm_bo_map_kmap;
415 map->page = ttm_tt_get_page(ttm, start_page);
416 map->virtual = kmap(map->page);
419 * Populate the part we're mapping;
421 for (i = start_page; i < start_page + num_pages; ++i) {
422 d = ttm_tt_get_page(ttm, i);
428 * We need to use vmap to get the desired page protection
429 * or to make the buffer object look contiguous.
431 prot = (mem->placement & TTM_PL_FLAG_CACHED) ?
433 ttm_io_prot(mem->placement, PAGE_KERNEL);
434 map->bo_kmap_type = ttm_bo_map_vmap;
435 map->virtual = vmap(ttm->pages + start_page, num_pages,
438 return (!map->virtual) ? -ENOMEM : 0;
441 int ttm_bo_kmap(struct ttm_buffer_object *bo,
442 unsigned long start_page, unsigned long num_pages,
443 struct ttm_bo_kmap_obj *map)
446 unsigned long bus_base;
447 unsigned long bus_offset;
448 unsigned long bus_size;
450 BUG_ON(!list_empty(&bo->swap));
452 if (num_pages > bo->num_pages)
454 if (start_page > bo->num_pages)
457 if (num_pages > 1 && !DRM_SUSER(DRM_CURPROC))
460 ret = ttm_bo_pci_offset(bo->bdev, &bo->mem, &bus_base,
461 &bus_offset, &bus_size);
465 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
467 bus_offset += start_page << PAGE_SHIFT;
468 bus_size = num_pages << PAGE_SHIFT;
469 return ttm_bo_ioremap(bo, bus_base, bus_offset, bus_size, map);
472 EXPORT_SYMBOL(ttm_bo_kmap);
474 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
478 switch (map->bo_kmap_type) {
479 case ttm_bo_map_iomap:
480 iounmap(map->virtual);
482 case ttm_bo_map_vmap:
483 vunmap(map->virtual);
485 case ttm_bo_map_kmap:
488 case ttm_bo_map_premapped:
496 EXPORT_SYMBOL(ttm_bo_kunmap);
498 int ttm_bo_pfn_prot(struct ttm_buffer_object *bo,
499 unsigned long dst_offset,
500 unsigned long *pfn, pgprot_t *prot)
502 struct ttm_mem_reg *mem = &bo->mem;
503 struct ttm_bo_device *bdev = bo->bdev;
504 unsigned long bus_offset;
505 unsigned long bus_size;
506 unsigned long bus_base;
508 ret = ttm_bo_pci_offset(bdev, mem, &bus_base, &bus_offset,
513 *pfn = (bus_base + bus_offset + dst_offset) >> PAGE_SHIFT;
518 *pfn = page_to_pfn(ttm_tt_get_page(bo->ttm,
521 *prot = (mem->placement & TTM_PL_FLAG_CACHED) ?
522 PAGE_KERNEL : ttm_io_prot(mem->placement, PAGE_KERNEL);
527 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
530 bool evict, bool no_wait_reserve,
532 struct ttm_mem_reg *new_mem)
534 struct ttm_bo_device *bdev = bo->bdev;
535 struct ttm_bo_driver *driver = bdev->driver;
536 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
537 struct ttm_mem_reg *old_mem = &bo->mem;
539 struct ttm_buffer_object *ghost_obj;
540 void *tmp_obj = NULL;
542 spin_lock(&bo->lock);
544 tmp_obj = bo->sync_obj;
547 bo->sync_obj = driver->sync_obj_ref(sync_obj);
548 bo->sync_obj_arg = sync_obj_arg;
550 ret = ttm_bo_wait(bo, false, false, false);
551 spin_unlock(&bo->lock);
553 driver->sync_obj_unref(&tmp_obj);
557 ttm_bo_free_old_node(bo);
558 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
560 ttm_tt_unbind(bo->ttm);
561 ttm_tt_destroy(bo->ttm);
566 * This should help pipeline ordinary buffer moves.
568 * Hang old buffer memory on a new buffer object,
569 * and leave it to be released when the GPU
570 * operation has completed.
573 set_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
574 spin_unlock(&bo->lock);
576 driver->sync_obj_unref(&tmp_obj);
578 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
583 * If we're not moving to fixed memory, the TTM object
584 * needs to stay alive. Otherwhise hang it on the ghost
585 * bo to be unbound and destroyed.
588 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED))
589 ghost_obj->ttm = NULL;
593 ttm_bo_unreserve(ghost_obj);
594 ttm_bo_unref(&ghost_obj);
598 new_mem->mm_node = NULL;
602 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);