1 /**************************************************************************
3 * Copyright (c) 2006-2009 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 #define pr_fmt(fmt) "[TTM] " fmt
33 #include <drm/ttm/ttm_module.h>
34 #include <drm/ttm/ttm_bo_driver.h>
35 #include <drm/ttm/ttm_placement.h>
36 #include <drm/drm_vma_manager.h>
38 #include <linux/pfn_t.h>
39 #include <linux/rbtree.h>
40 #include <linux/module.h>
41 #include <linux/uaccess.h>
42 #include <linux/mem_encrypt.h>
44 #define TTM_BO_VM_NUM_PREFAULT 16
46 static int ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
51 if (likely(!bo->moving))
55 * Quick non-stalling check for idle.
57 if (dma_fence_is_signaled(bo->moving))
61 * If possible, avoid waiting for GPU with mmap_sem
64 if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
66 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
70 up_read(&vmf->vma->vm_mm->mmap_sem);
71 (void) dma_fence_wait(bo->moving, true);
80 ret = dma_fence_wait(bo->moving, true);
81 if (unlikely(ret != 0)) {
82 ret = (ret != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
88 dma_fence_put(bo->moving);
95 static int ttm_bo_vm_fault(struct vm_fault *vmf)
97 struct vm_area_struct *vma = vmf->vma;
98 struct ttm_buffer_object *bo = (struct ttm_buffer_object *)
100 struct ttm_bo_device *bdev = bo->bdev;
101 unsigned long page_offset;
102 unsigned long page_last;
104 struct ttm_tt *ttm = NULL;
108 unsigned long address = vmf->address;
109 int retval = VM_FAULT_NOPAGE;
110 struct ttm_mem_type_manager *man =
111 &bdev->man[bo->mem.mem_type];
112 struct vm_area_struct cvma;
115 * Work around locking order reversal in fault / nopfn
116 * between mmap_sem and bo_reserve: Perform a trylock operation
117 * for reserve, and if it fails, retry the fault after waiting
118 * for the buffer to become unreserved.
120 ret = ttm_bo_reserve(bo, true, true, NULL);
121 if (unlikely(ret != 0)) {
123 return VM_FAULT_NOPAGE;
125 if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
126 if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
127 ttm_bo_reference(bo);
128 up_read(&vmf->vma->vm_mm->mmap_sem);
129 (void) ttm_bo_wait_unreserved(bo);
133 return VM_FAULT_RETRY;
137 * If we'd want to change locking order to
138 * mmap_sem -> bo::reserve, we'd use a blocking reserve here
139 * instead of retrying the fault...
141 return VM_FAULT_NOPAGE;
145 * Refuse to fault imported pages. This should be handled
146 * (if at all) by redirecting mmap to the exporter.
148 if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
149 retval = VM_FAULT_SIGBUS;
153 if (bdev->driver->fault_reserve_notify) {
154 ret = bdev->driver->fault_reserve_notify(bo);
160 retval = VM_FAULT_NOPAGE;
163 retval = VM_FAULT_SIGBUS;
169 * Wait for buffer data in transit, due to a pipelined
172 ret = ttm_bo_vm_fault_idle(bo, vmf);
173 if (unlikely(ret != 0)) {
176 if (retval == VM_FAULT_RETRY &&
177 !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
178 /* The BO has already been unreserved. */
185 ret = ttm_mem_io_lock(man, true);
186 if (unlikely(ret != 0)) {
187 retval = VM_FAULT_NOPAGE;
190 ret = ttm_mem_io_reserve_vm(bo);
191 if (unlikely(ret != 0)) {
192 retval = VM_FAULT_SIGBUS;
196 page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
197 vma->vm_pgoff - drm_vma_node_start(&bo->vma_node);
198 page_last = vma_pages(vma) + vma->vm_pgoff -
199 drm_vma_node_start(&bo->vma_node);
201 if (unlikely(page_offset >= bo->num_pages)) {
202 retval = VM_FAULT_SIGBUS;
207 * Make a local vma copy to modify the page_prot member
208 * and vm_flags if necessary. The vma parameter is protected
209 * by mmap_sem in write mode.
212 cvma.vm_page_prot = vm_get_page_prot(cvma.vm_flags);
214 if (bo->mem.bus.is_iomem) {
215 cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
219 cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
222 /* Allocate all page at once, most common usage */
223 if (ttm->bdev->driver->ttm_tt_populate(ttm)) {
224 retval = VM_FAULT_OOM;
230 * Speculatively prefault a number of pages. Only error on
233 for (i = 0; i < TTM_BO_VM_NUM_PREFAULT; ++i) {
234 if (bo->mem.bus.is_iomem) {
235 /* Iomem should not be marked encrypted */
236 cvma.vm_page_prot = pgprot_decrypted(cvma.vm_page_prot);
237 pfn = bdev->driver->io_mem_pfn(bo, page_offset);
239 page = ttm->pages[page_offset];
240 if (unlikely(!page && i == 0)) {
241 retval = VM_FAULT_OOM;
243 } else if (unlikely(!page)) {
246 page->mapping = vma->vm_file->f_mapping;
247 page->index = drm_vma_node_start(&bo->vma_node) +
249 pfn = page_to_pfn(page);
252 if (vma->vm_flags & VM_MIXEDMAP)
253 ret = vm_insert_mixed(&cvma, address,
254 __pfn_to_pfn_t(pfn, PFN_DEV));
256 ret = vm_insert_pfn(&cvma, address, pfn);
259 * Somebody beat us to this PTE or prefaulting to
260 * an already populated PTE, or prefaulting error.
263 if (unlikely((ret == -EBUSY) || (ret != 0 && i > 0)))
265 else if (unlikely(ret != 0)) {
267 (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
271 address += PAGE_SIZE;
272 if (unlikely(++page_offset >= page_last))
276 ttm_mem_io_unlock(man);
278 ttm_bo_unreserve(bo);
282 static void ttm_bo_vm_open(struct vm_area_struct *vma)
284 struct ttm_buffer_object *bo =
285 (struct ttm_buffer_object *)vma->vm_private_data;
287 WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
289 (void)ttm_bo_reference(bo);
292 static void ttm_bo_vm_close(struct vm_area_struct *vma)
294 struct ttm_buffer_object *bo = (struct ttm_buffer_object *)vma->vm_private_data;
297 vma->vm_private_data = NULL;
300 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
301 unsigned long offset,
302 uint8_t *buf, int len, int write)
304 unsigned long page = offset >> PAGE_SHIFT;
305 unsigned long bytes_left = len;
308 /* Copy a page at a time, that way no extra virtual address
311 offset -= page << PAGE_SHIFT;
313 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
314 struct ttm_bo_kmap_obj map;
318 ret = ttm_bo_kmap(bo, page, 1, &map);
322 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
323 WARN_ON_ONCE(is_iomem);
325 memcpy(ptr, buf, bytes);
327 memcpy(buf, ptr, bytes);
334 } while (bytes_left);
339 static int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
340 void *buf, int len, int write)
342 unsigned long offset = (addr) - vma->vm_start;
343 struct ttm_buffer_object *bo = vma->vm_private_data;
346 if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->num_pages)
349 ret = ttm_bo_reserve(bo, true, false, NULL);
353 switch (bo->mem.mem_type) {
355 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
356 ret = ttm_tt_swapin(bo->ttm);
357 if (unlikely(ret != 0))
362 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
365 if (bo->bdev->driver->access_memory)
366 ret = bo->bdev->driver->access_memory(
367 bo, offset, buf, len, write);
372 ttm_bo_unreserve(bo);
377 static const struct vm_operations_struct ttm_bo_vm_ops = {
378 .fault = ttm_bo_vm_fault,
379 .open = ttm_bo_vm_open,
380 .close = ttm_bo_vm_close,
381 .access = ttm_bo_vm_access
384 static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
385 unsigned long offset,
388 struct drm_vma_offset_node *node;
389 struct ttm_buffer_object *bo = NULL;
391 drm_vma_offset_lock_lookup(&bdev->vma_manager);
393 node = drm_vma_offset_lookup_locked(&bdev->vma_manager, offset, pages);
395 bo = container_of(node, struct ttm_buffer_object, vma_node);
396 if (!kref_get_unless_zero(&bo->kref))
400 drm_vma_offset_unlock_lookup(&bdev->vma_manager);
403 pr_err("Could not find buffer object to map\n");
408 unsigned long ttm_bo_default_io_mem_pfn(struct ttm_buffer_object *bo,
409 unsigned long page_offset)
411 return ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT)
414 EXPORT_SYMBOL(ttm_bo_default_io_mem_pfn);
416 int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
417 struct ttm_bo_device *bdev)
419 struct ttm_bo_driver *driver;
420 struct ttm_buffer_object *bo;
423 bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
427 driver = bo->bdev->driver;
428 if (unlikely(!driver->verify_access)) {
432 ret = driver->verify_access(bo, filp);
433 if (unlikely(ret != 0))
436 vma->vm_ops = &ttm_bo_vm_ops;
439 * Note: We're transferring the bo reference to
440 * vma->vm_private_data here.
443 vma->vm_private_data = bo;
446 * We'd like to use VM_PFNMAP on shared mappings, where
447 * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
448 * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
449 * bad for performance. Until that has been sorted out, use
450 * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
452 vma->vm_flags |= VM_MIXEDMAP;
453 vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
459 EXPORT_SYMBOL(ttm_bo_mmap);
461 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
463 if (vma->vm_pgoff != 0)
466 vma->vm_ops = &ttm_bo_vm_ops;
467 vma->vm_private_data = ttm_bo_reference(bo);
468 vma->vm_flags |= VM_MIXEDMAP;
469 vma->vm_flags |= VM_IO | VM_DONTEXPAND;
472 EXPORT_SYMBOL(ttm_fbdev_mmap);