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 <drm/ttm/ttm_bo_driver.h>
32 #include <drm/ttm/ttm_placement.h>
34 #include <linux/highmem.h>
35 #include <linux/wait.h>
36 #include <linux/slab.h>
37 #include <linux/vmalloc.h>
38 #include <linux/module.h>
40 void ttm_bo_free_old_node(struct ttm_buffer_object *bo)
42 ttm_bo_mem_put(bo, &bo->mem);
45 int ttm_bo_move_ttm(struct ttm_buffer_object *bo,
47 bool no_wait_gpu, struct ttm_mem_reg *new_mem)
49 struct ttm_tt *ttm = bo->ttm;
50 struct ttm_mem_reg *old_mem = &bo->mem;
53 if (old_mem->mem_type != TTM_PL_SYSTEM) {
55 ttm_bo_free_old_node(bo);
56 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM,
58 old_mem->mem_type = TTM_PL_SYSTEM;
61 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement);
62 if (unlikely(ret != 0))
65 if (new_mem->mem_type != TTM_PL_SYSTEM) {
66 ret = ttm_tt_bind(ttm, new_mem);
67 if (unlikely(ret != 0))
72 new_mem->mm_node = NULL;
76 EXPORT_SYMBOL(ttm_bo_move_ttm);
78 int ttm_mem_io_lock(struct ttm_mem_type_manager *man, bool interruptible)
80 if (likely(man->io_reserve_fastpath))
84 return mutex_lock_interruptible(&man->io_reserve_mutex);
86 mutex_lock(&man->io_reserve_mutex);
90 void ttm_mem_io_unlock(struct ttm_mem_type_manager *man)
92 if (likely(man->io_reserve_fastpath))
95 mutex_unlock(&man->io_reserve_mutex);
98 static int ttm_mem_io_evict(struct ttm_mem_type_manager *man)
100 struct ttm_buffer_object *bo;
102 if (!man->use_io_reserve_lru || list_empty(&man->io_reserve_lru))
105 bo = list_first_entry(&man->io_reserve_lru,
106 struct ttm_buffer_object,
108 list_del_init(&bo->io_reserve_lru);
109 ttm_bo_unmap_virtual_locked(bo);
114 static int ttm_mem_io_reserve(struct ttm_bo_device *bdev,
115 struct ttm_mem_reg *mem)
117 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
120 if (!bdev->driver->io_mem_reserve)
122 if (likely(man->io_reserve_fastpath))
123 return bdev->driver->io_mem_reserve(bdev, mem);
125 if (bdev->driver->io_mem_reserve &&
126 mem->bus.io_reserved_count++ == 0) {
128 ret = bdev->driver->io_mem_reserve(bdev, mem);
129 if (ret == -EAGAIN) {
130 ret = ttm_mem_io_evict(man);
138 static void ttm_mem_io_free(struct ttm_bo_device *bdev,
139 struct ttm_mem_reg *mem)
141 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
143 if (likely(man->io_reserve_fastpath))
146 if (bdev->driver->io_mem_reserve &&
147 --mem->bus.io_reserved_count == 0 &&
148 bdev->driver->io_mem_free)
149 bdev->driver->io_mem_free(bdev, mem);
153 int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo)
155 struct ttm_mem_reg *mem = &bo->mem;
158 if (!mem->bus.io_reserved_vm) {
159 struct ttm_mem_type_manager *man =
160 &bo->bdev->man[mem->mem_type];
162 ret = ttm_mem_io_reserve(bo->bdev, mem);
163 if (unlikely(ret != 0))
165 mem->bus.io_reserved_vm = true;
166 if (man->use_io_reserve_lru)
167 list_add_tail(&bo->io_reserve_lru,
168 &man->io_reserve_lru);
173 void ttm_mem_io_free_vm(struct ttm_buffer_object *bo)
175 struct ttm_mem_reg *mem = &bo->mem;
177 if (mem->bus.io_reserved_vm) {
178 mem->bus.io_reserved_vm = false;
179 list_del_init(&bo->io_reserve_lru);
180 ttm_mem_io_free(bo->bdev, mem);
184 int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
187 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
192 (void) ttm_mem_io_lock(man, false);
193 ret = ttm_mem_io_reserve(bdev, mem);
194 ttm_mem_io_unlock(man);
195 if (ret || !mem->bus.is_iomem)
199 addr = mem->bus.addr;
201 if (mem->placement & TTM_PL_FLAG_WC)
202 addr = ioremap_wc(mem->bus.base + mem->bus.offset, mem->bus.size);
204 addr = ioremap_nocache(mem->bus.base + mem->bus.offset, mem->bus.size);
206 (void) ttm_mem_io_lock(man, false);
207 ttm_mem_io_free(bdev, mem);
208 ttm_mem_io_unlock(man);
216 void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
219 struct ttm_mem_type_manager *man;
221 man = &bdev->man[mem->mem_type];
223 if (virtual && mem->bus.addr == NULL)
225 (void) ttm_mem_io_lock(man, false);
226 ttm_mem_io_free(bdev, mem);
227 ttm_mem_io_unlock(man);
230 static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
233 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
235 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
238 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
239 iowrite32(ioread32(srcP++), dstP++);
243 static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
247 struct page *d = ttm->pages[page];
253 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
256 dst = kmap_atomic_prot(d, prot);
258 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
259 dst = vmap(&d, 1, 0, prot);
266 memcpy_fromio(dst, src, PAGE_SIZE);
271 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
280 static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
284 struct page *s = ttm->pages[page];
290 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
292 src = kmap_atomic_prot(s, prot);
294 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
295 src = vmap(&s, 1, 0, prot);
302 memcpy_toio(dst, src, PAGE_SIZE);
307 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
316 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
317 bool evict, bool no_wait_gpu,
318 struct ttm_mem_reg *new_mem)
320 struct ttm_bo_device *bdev = bo->bdev;
321 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
322 struct ttm_tt *ttm = bo->ttm;
323 struct ttm_mem_reg *old_mem = &bo->mem;
324 struct ttm_mem_reg old_copy = *old_mem;
330 unsigned long add = 0;
333 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap);
336 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap);
340 if (old_iomap == NULL && new_iomap == NULL)
342 if (old_iomap == NULL && ttm == NULL)
345 if (ttm->state == tt_unpopulated) {
346 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
348 /* if we fail here don't nuke the mm node
349 * as the bo still owns it */
350 old_copy.mm_node = NULL;
358 if ((old_mem->mem_type == new_mem->mem_type) &&
359 (new_mem->start < old_mem->start + old_mem->size)) {
361 add = new_mem->num_pages - 1;
364 for (i = 0; i < new_mem->num_pages; ++i) {
365 page = i * dir + add;
366 if (old_iomap == NULL) {
367 pgprot_t prot = ttm_io_prot(old_mem->placement,
369 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
371 } else if (new_iomap == NULL) {
372 pgprot_t prot = ttm_io_prot(new_mem->placement,
374 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
377 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
379 /* failing here, means keep old copy as-is */
380 old_copy.mm_node = NULL;
388 new_mem->mm_node = NULL;
390 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && (ttm != NULL)) {
397 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap);
399 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap);
400 ttm_bo_mem_put(bo, &old_copy);
403 EXPORT_SYMBOL(ttm_bo_move_memcpy);
405 static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
411 * ttm_buffer_object_transfer
413 * @bo: A pointer to a struct ttm_buffer_object.
414 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
415 * holding the data of @bo with the old placement.
417 * This is a utility function that may be called after an accelerated move
418 * has been scheduled. A new buffer object is created as a placeholder for
419 * the old data while it's being copied. When that buffer object is idle,
420 * it can be destroyed, releasing the space of the old placement.
425 static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
426 struct ttm_buffer_object **new_obj)
428 struct ttm_buffer_object *fbo;
429 struct ttm_bo_device *bdev = bo->bdev;
430 struct ttm_bo_driver *driver = bdev->driver;
432 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
439 * Fix up members that we shouldn't copy directly:
440 * TODO: Explicit member copy would probably be better here.
443 init_waitqueue_head(&fbo->event_queue);
444 INIT_LIST_HEAD(&fbo->ddestroy);
445 INIT_LIST_HEAD(&fbo->lru);
446 INIT_LIST_HEAD(&fbo->swap);
447 INIT_LIST_HEAD(&fbo->io_reserve_lru);
449 atomic_set(&fbo->cpu_writers, 0);
451 spin_lock(&bdev->fence_lock);
453 fbo->sync_obj = driver->sync_obj_ref(bo->sync_obj);
455 fbo->sync_obj = NULL;
456 spin_unlock(&bdev->fence_lock);
457 kref_init(&fbo->list_kref);
458 kref_init(&fbo->kref);
459 fbo->destroy = &ttm_transfered_destroy;
466 pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
468 #if defined(__i386__) || defined(__x86_64__)
469 if (caching_flags & TTM_PL_FLAG_WC)
470 tmp = pgprot_writecombine(tmp);
471 else if (boot_cpu_data.x86 > 3)
472 tmp = pgprot_noncached(tmp);
474 #elif defined(__powerpc__)
475 if (!(caching_flags & TTM_PL_FLAG_CACHED)) {
476 pgprot_val(tmp) |= _PAGE_NO_CACHE;
477 if (caching_flags & TTM_PL_FLAG_UNCACHED)
478 pgprot_val(tmp) |= _PAGE_GUARDED;
481 #if defined(__ia64__)
482 if (caching_flags & TTM_PL_FLAG_WC)
483 tmp = pgprot_writecombine(tmp);
485 tmp = pgprot_noncached(tmp);
487 #if defined(__sparc__) || defined(__mips__)
488 if (!(caching_flags & TTM_PL_FLAG_CACHED))
489 tmp = pgprot_noncached(tmp);
493 EXPORT_SYMBOL(ttm_io_prot);
495 static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
496 unsigned long offset,
498 struct ttm_bo_kmap_obj *map)
500 struct ttm_mem_reg *mem = &bo->mem;
502 if (bo->mem.bus.addr) {
503 map->bo_kmap_type = ttm_bo_map_premapped;
504 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
506 map->bo_kmap_type = ttm_bo_map_iomap;
507 if (mem->placement & TTM_PL_FLAG_WC)
508 map->virtual = ioremap_wc(bo->mem.bus.base + bo->mem.bus.offset + offset,
511 map->virtual = ioremap_nocache(bo->mem.bus.base + bo->mem.bus.offset + offset,
514 return (!map->virtual) ? -ENOMEM : 0;
517 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
518 unsigned long start_page,
519 unsigned long num_pages,
520 struct ttm_bo_kmap_obj *map)
522 struct ttm_mem_reg *mem = &bo->mem; pgprot_t prot;
523 struct ttm_tt *ttm = bo->ttm;
528 if (ttm->state == tt_unpopulated) {
529 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
534 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
536 * We're mapping a single page, and the desired
537 * page protection is consistent with the bo.
540 map->bo_kmap_type = ttm_bo_map_kmap;
541 map->page = ttm->pages[start_page];
542 map->virtual = kmap(map->page);
545 * We need to use vmap to get the desired page protection
546 * or to make the buffer object look contiguous.
548 prot = (mem->placement & TTM_PL_FLAG_CACHED) ?
550 ttm_io_prot(mem->placement, PAGE_KERNEL);
551 map->bo_kmap_type = ttm_bo_map_vmap;
552 map->virtual = vmap(ttm->pages + start_page, num_pages,
555 return (!map->virtual) ? -ENOMEM : 0;
558 int ttm_bo_kmap(struct ttm_buffer_object *bo,
559 unsigned long start_page, unsigned long num_pages,
560 struct ttm_bo_kmap_obj *map)
562 struct ttm_mem_type_manager *man =
563 &bo->bdev->man[bo->mem.mem_type];
564 unsigned long offset, size;
567 BUG_ON(!list_empty(&bo->swap));
570 if (num_pages > bo->num_pages)
572 if (start_page > bo->num_pages)
575 if (num_pages > 1 && !DRM_SUSER(DRM_CURPROC))
578 (void) ttm_mem_io_lock(man, false);
579 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem);
580 ttm_mem_io_unlock(man);
583 if (!bo->mem.bus.is_iomem) {
584 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
586 offset = start_page << PAGE_SHIFT;
587 size = num_pages << PAGE_SHIFT;
588 return ttm_bo_ioremap(bo, offset, size, map);
591 EXPORT_SYMBOL(ttm_bo_kmap);
593 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
595 struct ttm_buffer_object *bo = map->bo;
596 struct ttm_mem_type_manager *man =
597 &bo->bdev->man[bo->mem.mem_type];
601 switch (map->bo_kmap_type) {
602 case ttm_bo_map_iomap:
603 iounmap(map->virtual);
605 case ttm_bo_map_vmap:
606 vunmap(map->virtual);
608 case ttm_bo_map_kmap:
611 case ttm_bo_map_premapped:
616 (void) ttm_mem_io_lock(man, false);
617 ttm_mem_io_free(map->bo->bdev, &map->bo->mem);
618 ttm_mem_io_unlock(man);
622 EXPORT_SYMBOL(ttm_bo_kunmap);
624 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
628 struct ttm_mem_reg *new_mem)
630 struct ttm_bo_device *bdev = bo->bdev;
631 struct ttm_bo_driver *driver = bdev->driver;
632 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
633 struct ttm_mem_reg *old_mem = &bo->mem;
635 struct ttm_buffer_object *ghost_obj;
636 void *tmp_obj = NULL;
638 spin_lock(&bdev->fence_lock);
640 tmp_obj = bo->sync_obj;
643 bo->sync_obj = driver->sync_obj_ref(sync_obj);
645 ret = ttm_bo_wait(bo, false, false, false);
646 spin_unlock(&bdev->fence_lock);
648 driver->sync_obj_unref(&tmp_obj);
652 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
654 ttm_tt_unbind(bo->ttm);
655 ttm_tt_destroy(bo->ttm);
658 ttm_bo_free_old_node(bo);
661 * This should help pipeline ordinary buffer moves.
663 * Hang old buffer memory on a new buffer object,
664 * and leave it to be released when the GPU
665 * operation has completed.
668 set_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
669 spin_unlock(&bdev->fence_lock);
671 driver->sync_obj_unref(&tmp_obj);
673 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
678 * If we're not moving to fixed memory, the TTM object
679 * needs to stay alive. Otherwhise hang it on the ghost
680 * bo to be unbound and destroyed.
683 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED))
684 ghost_obj->ttm = NULL;
688 ttm_bo_unreserve(ghost_obj);
689 ttm_bo_unref(&ghost_obj);
693 new_mem->mm_node = NULL;
697 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);