1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2 /**************************************************************************
4 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25 * USE OR OTHER DEALINGS IN THE SOFTWARE.
27 **************************************************************************/
29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
32 #define pr_fmt(fmt) "[TTM] " fmt
34 #include <drm/ttm/ttm_module.h>
35 #include <drm/ttm/ttm_bo_driver.h>
36 #include <drm/ttm/ttm_placement.h>
37 #include <drm/drm_vma_manager.h>
39 #include <linux/pfn_t.h>
40 #include <linux/rbtree.h>
41 #include <linux/module.h>
42 #include <linux/uaccess.h>
43 #include <linux/mem_encrypt.h>
45 static vm_fault_t 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
62 * held. We only do this if the fault allows retry and this
63 * is the first attempt.
65 if (fault_flag_allow_retry_first(vmf->flags)) {
67 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
71 up_read(&vmf->vma->vm_mm->mmap_sem);
72 (void) dma_fence_wait(bo->moving, true);
73 dma_resv_unlock(bo->base.resv);
81 err = dma_fence_wait(bo->moving, true);
82 if (unlikely(err != 0)) {
83 ret = (err != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
89 dma_fence_put(bo->moving);
96 static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object *bo,
97 unsigned long page_offset)
99 struct ttm_bo_device *bdev = bo->bdev;
101 if (bdev->driver->io_mem_pfn)
102 return bdev->driver->io_mem_pfn(bo, page_offset);
104 return ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT)
109 * ttm_bo_vm_reserve - Reserve a buffer object in a retryable vm callback
110 * @bo: The buffer object
111 * @vmf: The fault structure handed to the callback
113 * vm callbacks like fault() and *_mkwrite() allow for the mm_sem to be dropped
114 * during long waits, and after the wait the callback will be restarted. This
115 * is to allow other threads using the same virtual memory space concurrent
116 * access to map(), unmap() completely unrelated buffer objects. TTM buffer
117 * object reservations sometimes wait for GPU and should therefore be
118 * considered long waits. This function reserves the buffer object interruptibly
119 * taking this into account. Starvation is avoided by the vm system not
120 * allowing too many repeated restarts.
121 * This function is intended to be used in customized fault() and _mkwrite()
125 * 0 on success and the bo was reserved.
126 * VM_FAULT_RETRY if blocking wait.
127 * VM_FAULT_NOPAGE if blocking wait and retrying was not allowed.
129 vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo,
130 struct vm_fault *vmf)
133 * Work around locking order reversal in fault / nopfn
134 * between mmap_sem and bo_reserve: Perform a trylock operation
135 * for reserve, and if it fails, retry the fault after waiting
136 * for the buffer to become unreserved.
138 if (unlikely(!dma_resv_trylock(bo->base.resv))) {
140 * If the fault allows retry and this is the first
141 * fault attempt, we try to release the mmap_sem
144 if (fault_flag_allow_retry_first(vmf->flags)) {
145 if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
147 up_read(&vmf->vma->vm_mm->mmap_sem);
148 if (!dma_resv_lock_interruptible(bo->base.resv,
150 dma_resv_unlock(bo->base.resv);
154 return VM_FAULT_RETRY;
157 if (dma_resv_lock_interruptible(bo->base.resv, NULL))
158 return VM_FAULT_NOPAGE;
163 EXPORT_SYMBOL(ttm_bo_vm_reserve);
165 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
167 * ttm_bo_vm_insert_huge - Insert a pfn for PUD or PMD faults
169 * @bo: The buffer object
170 * @page_offset: Page offset from bo start
171 * @fault_page_size: The size of the fault in pages.
172 * @pgprot: The page protections.
173 * Does additional checking whether it's possible to insert a PUD or PMD
174 * pfn and performs the insertion.
176 * Return: VM_FAULT_NOPAGE on successful insertion, VM_FAULT_FALLBACK if
177 * a huge fault was not possible, or on insertion error.
179 static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
180 struct ttm_buffer_object *bo,
182 pgoff_t fault_page_size,
189 struct ttm_tt *ttm = bo->ttm;
190 bool write = vmf->flags & FAULT_FLAG_WRITE;
192 /* Fault should not cross bo boundary. */
193 page_offset &= ~(fault_page_size - 1);
194 if (page_offset + fault_page_size > bo->num_pages)
197 if (bo->mem.bus.is_iomem)
198 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
200 pfn = page_to_pfn(ttm->pages[page_offset]);
202 /* pfn must be fault_page_size aligned. */
203 if ((pfn & (fault_page_size - 1)) != 0)
206 /* Check that memory is contiguous. */
207 if (!bo->mem.bus.is_iomem) {
208 for (i = 1; i < fault_page_size; ++i) {
209 if (page_to_pfn(ttm->pages[page_offset + i]) != pfn + i)
212 } else if (bo->bdev->driver->io_mem_pfn) {
213 for (i = 1; i < fault_page_size; ++i) {
214 if (ttm_bo_io_mem_pfn(bo, page_offset + i) != pfn + i)
219 pfnt = __pfn_to_pfn_t(pfn, PFN_DEV);
220 if (fault_page_size == (HPAGE_PMD_SIZE >> PAGE_SHIFT))
221 ret = vmf_insert_pfn_pmd_prot(vmf, pfnt, pgprot, write);
222 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
223 else if (fault_page_size == (HPAGE_PUD_SIZE >> PAGE_SHIFT))
224 ret = vmf_insert_pfn_pud_prot(vmf, pfnt, pgprot, write);
227 WARN_ON_ONCE(ret = VM_FAULT_FALLBACK);
229 if (ret != VM_FAULT_NOPAGE)
232 return VM_FAULT_NOPAGE;
234 count_vm_event(THP_FAULT_FALLBACK);
235 return VM_FAULT_FALLBACK;
238 static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
239 struct ttm_buffer_object *bo,
241 pgoff_t fault_page_size,
244 return VM_FAULT_FALLBACK;
249 * ttm_bo_vm_fault_reserved - TTM fault helper
250 * @vmf: The struct vm_fault given as argument to the fault callback
251 * @prot: The page protection to be used for this memory area.
252 * @num_prefault: Maximum number of prefault pages. The caller may want to
253 * specify this based on madvice settings and the size of the GPU object
254 * backed by the memory.
255 * @fault_page_size: The size of the fault in pages.
257 * This function inserts one or more page table entries pointing to the
258 * memory backing the buffer object, and then returns a return code
259 * instructing the caller to retry the page access.
262 * VM_FAULT_NOPAGE on success or pending signal
263 * VM_FAULT_SIGBUS on unspecified error
264 * VM_FAULT_OOM on out-of-memory
265 * VM_FAULT_RETRY if retryable wait
267 vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
269 pgoff_t num_prefault,
270 pgoff_t fault_page_size)
272 struct vm_area_struct *vma = vmf->vma;
273 struct ttm_buffer_object *bo = vma->vm_private_data;
274 struct ttm_bo_device *bdev = bo->bdev;
275 unsigned long page_offset;
276 unsigned long page_last;
278 struct ttm_tt *ttm = NULL;
282 vm_fault_t ret = VM_FAULT_NOPAGE;
283 unsigned long address = vmf->address;
284 struct ttm_mem_type_manager *man =
285 &bdev->man[bo->mem.mem_type];
288 * Refuse to fault imported pages. This should be handled
289 * (if at all) by redirecting mmap to the exporter.
291 if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG))
292 return VM_FAULT_SIGBUS;
294 if (bdev->driver->fault_reserve_notify) {
295 struct dma_fence *moving = dma_fence_get(bo->moving);
297 err = bdev->driver->fault_reserve_notify(bo);
303 return VM_FAULT_NOPAGE;
305 return VM_FAULT_SIGBUS;
308 if (bo->moving != moving) {
309 spin_lock(&ttm_bo_glob.lru_lock);
310 ttm_bo_move_to_lru_tail(bo, NULL);
311 spin_unlock(&ttm_bo_glob.lru_lock);
313 dma_fence_put(moving);
317 * Wait for buffer data in transit, due to a pipelined
320 ret = ttm_bo_vm_fault_idle(bo, vmf);
321 if (unlikely(ret != 0))
324 err = ttm_mem_io_lock(man, true);
325 if (unlikely(err != 0))
326 return VM_FAULT_NOPAGE;
327 err = ttm_mem_io_reserve_vm(bo);
328 if (unlikely(err != 0)) {
329 ret = VM_FAULT_SIGBUS;
333 page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
334 vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
335 page_last = vma_pages(vma) + vma->vm_pgoff -
336 drm_vma_node_start(&bo->base.vma_node);
338 if (unlikely(page_offset >= bo->num_pages)) {
339 ret = VM_FAULT_SIGBUS;
343 prot = ttm_io_prot(bo->mem.placement, prot);
344 if (!bo->mem.bus.is_iomem) {
345 struct ttm_operation_ctx ctx = {
346 .interruptible = false,
347 .no_wait_gpu = false,
348 .flags = TTM_OPT_FLAG_FORCE_ALLOC
353 if (ttm_tt_populate(bo->ttm, &ctx)) {
358 /* Iomem should not be marked encrypted */
359 prot = pgprot_decrypted(prot);
362 /* We don't prefault on huge faults. Yet. */
363 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && fault_page_size != 1) {
364 ret = ttm_bo_vm_insert_huge(vmf, bo, page_offset,
365 fault_page_size, prot);
370 * Speculatively prefault a number of pages. Only error on
373 for (i = 0; i < num_prefault; ++i) {
374 if (bo->mem.bus.is_iomem) {
375 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
377 page = ttm->pages[page_offset];
378 if (unlikely(!page && i == 0)) {
381 } else if (unlikely(!page)) {
384 page->index = drm_vma_node_start(&bo->base.vma_node) +
386 pfn = page_to_pfn(page);
390 * Note that the value of @prot at this point may differ from
391 * the value of @vma->vm_page_prot in the caching- and
392 * encryption bits. This is because the exact location of the
393 * data may not be known at mmap() time and may also change
394 * at arbitrary times while the data is mmap'ed.
395 * See vmf_insert_mixed_prot() for a discussion.
397 if (vma->vm_flags & VM_MIXEDMAP)
398 ret = vmf_insert_mixed_prot(vma, address,
399 __pfn_to_pfn_t(pfn, PFN_DEV),
402 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
404 /* Never error on prefaulted PTEs */
405 if (unlikely((ret & VM_FAULT_ERROR))) {
412 address += PAGE_SIZE;
413 if (unlikely(++page_offset >= page_last))
416 ret = VM_FAULT_NOPAGE;
418 ttm_mem_io_unlock(man);
421 EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
423 vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
425 struct vm_area_struct *vma = vmf->vma;
427 struct ttm_buffer_object *bo = vma->vm_private_data;
430 ret = ttm_bo_vm_reserve(bo, vmf);
434 prot = vma->vm_page_prot;
435 ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT, 1);
436 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
439 dma_resv_unlock(bo->base.resv);
443 EXPORT_SYMBOL(ttm_bo_vm_fault);
445 void ttm_bo_vm_open(struct vm_area_struct *vma)
447 struct ttm_buffer_object *bo = vma->vm_private_data;
449 WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
453 EXPORT_SYMBOL(ttm_bo_vm_open);
455 void ttm_bo_vm_close(struct vm_area_struct *vma)
457 struct ttm_buffer_object *bo = vma->vm_private_data;
460 vma->vm_private_data = NULL;
462 EXPORT_SYMBOL(ttm_bo_vm_close);
464 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
465 unsigned long offset,
466 uint8_t *buf, int len, int write)
468 unsigned long page = offset >> PAGE_SHIFT;
469 unsigned long bytes_left = len;
472 /* Copy a page at a time, that way no extra virtual address
475 offset -= page << PAGE_SHIFT;
477 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
478 struct ttm_bo_kmap_obj map;
482 ret = ttm_bo_kmap(bo, page, 1, &map);
486 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
487 WARN_ON_ONCE(is_iomem);
489 memcpy(ptr, buf, bytes);
491 memcpy(buf, ptr, bytes);
498 } while (bytes_left);
503 int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
504 void *buf, int len, int write)
506 unsigned long offset = (addr) - vma->vm_start;
507 struct ttm_buffer_object *bo = vma->vm_private_data;
510 if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->num_pages)
513 ret = ttm_bo_reserve(bo, true, false, NULL);
517 switch (bo->mem.mem_type) {
519 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
520 ret = ttm_tt_swapin(bo->ttm);
521 if (unlikely(ret != 0))
526 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
529 if (bo->bdev->driver->access_memory)
530 ret = bo->bdev->driver->access_memory(
531 bo, offset, buf, len, write);
536 ttm_bo_unreserve(bo);
540 EXPORT_SYMBOL(ttm_bo_vm_access);
542 static const struct vm_operations_struct ttm_bo_vm_ops = {
543 .fault = ttm_bo_vm_fault,
544 .open = ttm_bo_vm_open,
545 .close = ttm_bo_vm_close,
546 .access = ttm_bo_vm_access,
549 static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
550 unsigned long offset,
553 struct drm_vma_offset_node *node;
554 struct ttm_buffer_object *bo = NULL;
556 drm_vma_offset_lock_lookup(bdev->vma_manager);
558 node = drm_vma_offset_lookup_locked(bdev->vma_manager, offset, pages);
560 bo = container_of(node, struct ttm_buffer_object,
562 bo = ttm_bo_get_unless_zero(bo);
565 drm_vma_offset_unlock_lookup(bdev->vma_manager);
568 pr_err("Could not find buffer object to map\n");
573 static void ttm_bo_mmap_vma_setup(struct ttm_buffer_object *bo, struct vm_area_struct *vma)
575 vma->vm_ops = &ttm_bo_vm_ops;
578 * Note: We're transferring the bo reference to
579 * vma->vm_private_data here.
582 vma->vm_private_data = bo;
585 * We'd like to use VM_PFNMAP on shared mappings, where
586 * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
587 * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
588 * bad for performance. Until that has been sorted out, use
589 * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
591 vma->vm_flags |= VM_MIXEDMAP;
592 vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
595 int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
596 struct ttm_bo_device *bdev)
598 struct ttm_bo_driver *driver;
599 struct ttm_buffer_object *bo;
602 if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET_START))
605 bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
609 driver = bo->bdev->driver;
610 if (unlikely(!driver->verify_access)) {
614 ret = driver->verify_access(bo, filp);
615 if (unlikely(ret != 0))
618 ttm_bo_mmap_vma_setup(bo, vma);
624 EXPORT_SYMBOL(ttm_bo_mmap);
626 int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
629 ttm_bo_mmap_vma_setup(bo, vma);
632 EXPORT_SYMBOL(ttm_bo_mmap_obj);