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, struct ttm_mem_reg *new_mem)
54 struct ttm_tt *ttm = bo->ttm;
55 struct ttm_mem_reg *old_mem = &bo->mem;
58 if (old_mem->mem_type != TTM_PL_SYSTEM) {
60 ttm_bo_free_old_node(bo);
61 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM,
63 old_mem->mem_type = TTM_PL_SYSTEM;
66 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement);
67 if (unlikely(ret != 0))
70 if (new_mem->mem_type != TTM_PL_SYSTEM) {
71 ret = ttm_tt_bind(ttm, new_mem);
72 if (unlikely(ret != 0))
77 new_mem->mm_node = NULL;
81 EXPORT_SYMBOL(ttm_bo_move_ttm);
83 int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
86 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
87 unsigned long bus_offset;
88 unsigned long bus_size;
89 unsigned long bus_base;
94 ret = ttm_bo_pci_offset(bdev, mem, &bus_base, &bus_offset, &bus_size);
95 if (ret || bus_size == 0)
98 if (!(man->flags & TTM_MEMTYPE_FLAG_NEEDS_IOREMAP))
99 addr = (void *)(((u8 *) man->io_addr) + bus_offset);
101 if (mem->placement & TTM_PL_FLAG_WC)
102 addr = ioremap_wc(bus_base + bus_offset, bus_size);
104 addr = ioremap_nocache(bus_base + bus_offset, bus_size);
112 void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
115 struct ttm_mem_type_manager *man;
117 man = &bdev->man[mem->mem_type];
119 if (virtual && (man->flags & TTM_MEMTYPE_FLAG_NEEDS_IOREMAP))
123 static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
126 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
128 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
131 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
132 iowrite32(ioread32(srcP++), dstP++);
136 static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
140 struct page *d = ttm_tt_get_page(ttm, page);
146 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
149 dst = kmap_atomic_prot(d, KM_USER0, prot);
151 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
152 dst = vmap(&d, 1, 0, prot);
159 memcpy_fromio(dst, src, PAGE_SIZE);
162 kunmap_atomic(dst, KM_USER0);
164 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
173 static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
177 struct page *s = ttm_tt_get_page(ttm, page);
183 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
185 src = kmap_atomic_prot(s, KM_USER0, prot);
187 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
188 src = vmap(&s, 1, 0, prot);
195 memcpy_toio(dst, src, PAGE_SIZE);
198 kunmap_atomic(src, KM_USER0);
200 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
209 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
210 bool evict, bool no_wait, struct ttm_mem_reg *new_mem)
212 struct ttm_bo_device *bdev = bo->bdev;
213 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
214 struct ttm_tt *ttm = bo->ttm;
215 struct ttm_mem_reg *old_mem = &bo->mem;
216 struct ttm_mem_reg old_copy = *old_mem;
222 unsigned long add = 0;
225 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap);
228 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap);
232 if (old_iomap == NULL && new_iomap == NULL)
234 if (old_iomap == NULL && ttm == NULL)
240 if ((old_mem->mem_type == new_mem->mem_type) &&
241 (new_mem->mm_node->start <
242 old_mem->mm_node->start + old_mem->mm_node->size)) {
244 add = new_mem->num_pages - 1;
247 for (i = 0; i < new_mem->num_pages; ++i) {
248 page = i * dir + add;
249 if (old_iomap == NULL) {
250 pgprot_t prot = ttm_io_prot(old_mem->placement,
252 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
254 } else if (new_iomap == NULL) {
255 pgprot_t prot = ttm_io_prot(new_mem->placement,
257 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
260 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
266 ttm_bo_free_old_node(bo);
269 new_mem->mm_node = NULL;
271 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && (ttm != NULL)) {
278 ttm_mem_reg_iounmap(bdev, new_mem, new_iomap);
280 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap);
283 EXPORT_SYMBOL(ttm_bo_move_memcpy);
285 static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
291 * ttm_buffer_object_transfer
293 * @bo: A pointer to a struct ttm_buffer_object.
294 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
295 * holding the data of @bo with the old placement.
297 * This is a utility function that may be called after an accelerated move
298 * has been scheduled. A new buffer object is created as a placeholder for
299 * the old data while it's being copied. When that buffer object is idle,
300 * it can be destroyed, releasing the space of the old placement.
305 static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
306 struct ttm_buffer_object **new_obj)
308 struct ttm_buffer_object *fbo;
309 struct ttm_bo_device *bdev = bo->bdev;
310 struct ttm_bo_driver *driver = bdev->driver;
312 fbo = kzalloc(sizeof(*fbo), GFP_KERNEL);
319 * Fix up members that we shouldn't copy directly:
320 * TODO: Explicit member copy would probably be better here.
323 spin_lock_init(&fbo->lock);
324 init_waitqueue_head(&fbo->event_queue);
325 INIT_LIST_HEAD(&fbo->ddestroy);
326 INIT_LIST_HEAD(&fbo->lru);
327 INIT_LIST_HEAD(&fbo->swap);
330 fbo->sync_obj = driver->sync_obj_ref(bo->sync_obj);
331 if (fbo->mem.mm_node)
332 fbo->mem.mm_node->private = (void *)fbo;
333 kref_init(&fbo->list_kref);
334 kref_init(&fbo->kref);
335 fbo->destroy = &ttm_transfered_destroy;
341 pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
343 #if defined(__i386__) || defined(__x86_64__)
344 if (caching_flags & TTM_PL_FLAG_WC)
345 tmp = pgprot_writecombine(tmp);
346 else if (boot_cpu_data.x86 > 3)
347 tmp = pgprot_noncached(tmp);
349 #elif defined(__powerpc__)
350 if (!(caching_flags & TTM_PL_FLAG_CACHED)) {
351 pgprot_val(tmp) |= _PAGE_NO_CACHE;
352 if (caching_flags & TTM_PL_FLAG_UNCACHED)
353 pgprot_val(tmp) |= _PAGE_GUARDED;
356 #if defined(__ia64__)
357 if (caching_flags & TTM_PL_FLAG_WC)
358 tmp = pgprot_writecombine(tmp);
360 tmp = pgprot_noncached(tmp);
362 #if defined(__sparc__)
363 if (!(caching_flags & TTM_PL_FLAG_CACHED))
364 tmp = pgprot_noncached(tmp);
368 EXPORT_SYMBOL(ttm_io_prot);
370 static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
371 unsigned long bus_base,
372 unsigned long bus_offset,
373 unsigned long bus_size,
374 struct ttm_bo_kmap_obj *map)
376 struct ttm_bo_device *bdev = bo->bdev;
377 struct ttm_mem_reg *mem = &bo->mem;
378 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
380 if (!(man->flags & TTM_MEMTYPE_FLAG_NEEDS_IOREMAP)) {
381 map->bo_kmap_type = ttm_bo_map_premapped;
382 map->virtual = (void *)(((u8 *) man->io_addr) + bus_offset);
384 map->bo_kmap_type = ttm_bo_map_iomap;
385 if (mem->placement & TTM_PL_FLAG_WC)
386 map->virtual = ioremap_wc(bus_base + bus_offset,
389 map->virtual = ioremap_nocache(bus_base + bus_offset,
392 return (!map->virtual) ? -ENOMEM : 0;
395 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
396 unsigned long start_page,
397 unsigned long num_pages,
398 struct ttm_bo_kmap_obj *map)
400 struct ttm_mem_reg *mem = &bo->mem; pgprot_t prot;
401 struct ttm_tt *ttm = bo->ttm;
406 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
408 * We're mapping a single page, and the desired
409 * page protection is consistent with the bo.
412 map->bo_kmap_type = ttm_bo_map_kmap;
413 map->page = ttm_tt_get_page(ttm, start_page);
414 map->virtual = kmap(map->page);
417 * Populate the part we're mapping;
419 for (i = start_page; i < start_page + num_pages; ++i) {
420 d = ttm_tt_get_page(ttm, i);
426 * We need to use vmap to get the desired page protection
427 * or to make the buffer object look contiguous.
429 prot = (mem->placement & TTM_PL_FLAG_CACHED) ?
431 ttm_io_prot(mem->placement, PAGE_KERNEL);
432 map->bo_kmap_type = ttm_bo_map_vmap;
433 map->virtual = vmap(ttm->pages + start_page, num_pages,
436 return (!map->virtual) ? -ENOMEM : 0;
439 int ttm_bo_kmap(struct ttm_buffer_object *bo,
440 unsigned long start_page, unsigned long num_pages,
441 struct ttm_bo_kmap_obj *map)
444 unsigned long bus_base;
445 unsigned long bus_offset;
446 unsigned long bus_size;
448 BUG_ON(!list_empty(&bo->swap));
450 if (num_pages > bo->num_pages)
452 if (start_page > bo->num_pages)
455 if (num_pages > 1 && !DRM_SUSER(DRM_CURPROC))
458 ret = ttm_bo_pci_offset(bo->bdev, &bo->mem, &bus_base,
459 &bus_offset, &bus_size);
463 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
465 bus_offset += start_page << PAGE_SHIFT;
466 bus_size = num_pages << PAGE_SHIFT;
467 return ttm_bo_ioremap(bo, bus_base, bus_offset, bus_size, map);
470 EXPORT_SYMBOL(ttm_bo_kmap);
472 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
476 switch (map->bo_kmap_type) {
477 case ttm_bo_map_iomap:
478 iounmap(map->virtual);
480 case ttm_bo_map_vmap:
481 vunmap(map->virtual);
483 case ttm_bo_map_kmap:
486 case ttm_bo_map_premapped:
494 EXPORT_SYMBOL(ttm_bo_kunmap);
496 int ttm_bo_pfn_prot(struct ttm_buffer_object *bo,
497 unsigned long dst_offset,
498 unsigned long *pfn, pgprot_t *prot)
500 struct ttm_mem_reg *mem = &bo->mem;
501 struct ttm_bo_device *bdev = bo->bdev;
502 unsigned long bus_offset;
503 unsigned long bus_size;
504 unsigned long bus_base;
506 ret = ttm_bo_pci_offset(bdev, mem, &bus_base, &bus_offset,
511 *pfn = (bus_base + bus_offset + dst_offset) >> PAGE_SHIFT;
516 *pfn = page_to_pfn(ttm_tt_get_page(bo->ttm,
519 *prot = (mem->placement & TTM_PL_FLAG_CACHED) ?
520 PAGE_KERNEL : ttm_io_prot(mem->placement, PAGE_KERNEL);
525 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
528 bool evict, bool no_wait,
529 struct ttm_mem_reg *new_mem)
531 struct ttm_bo_device *bdev = bo->bdev;
532 struct ttm_bo_driver *driver = bdev->driver;
533 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
534 struct ttm_mem_reg *old_mem = &bo->mem;
536 struct ttm_buffer_object *ghost_obj;
537 void *tmp_obj = NULL;
539 spin_lock(&bo->lock);
541 tmp_obj = bo->sync_obj;
544 bo->sync_obj = driver->sync_obj_ref(sync_obj);
545 bo->sync_obj_arg = sync_obj_arg;
547 ret = ttm_bo_wait(bo, false, false, false);
548 spin_unlock(&bo->lock);
550 driver->sync_obj_unref(&tmp_obj);
554 ttm_bo_free_old_node(bo);
555 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
557 ttm_tt_unbind(bo->ttm);
558 ttm_tt_destroy(bo->ttm);
563 * This should help pipeline ordinary buffer moves.
565 * Hang old buffer memory on a new buffer object,
566 * and leave it to be released when the GPU
567 * operation has completed.
570 set_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
571 spin_unlock(&bo->lock);
573 driver->sync_obj_unref(&tmp_obj);
575 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
580 * If we're not moving to fixed memory, the TTM object
581 * needs to stay alive. Otherwhise hang it on the ghost
582 * bo to be unbound and destroyed.
585 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED))
586 ghost_obj->ttm = NULL;
590 ttm_bo_unreserve(ghost_obj);
591 ttm_bo_unref(&ghost_obj);
595 new_mem->mm_node = NULL;
599 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);