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
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);
72 dma_resv_unlock(bo->base.resv);
80 err = dma_fence_wait(bo->moving, true);
81 if (unlikely(err != 0)) {
82 ret = (err != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
88 dma_fence_put(bo->moving);
95 static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object *bo,
96 unsigned long page_offset)
98 struct ttm_bo_device *bdev = bo->bdev;
100 if (bdev->driver->io_mem_pfn)
101 return bdev->driver->io_mem_pfn(bo, page_offset);
103 return ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT)
108 * ttm_bo_vm_reserve - Reserve a buffer object in a retryable vm callback
109 * @bo: The buffer object
110 * @vmf: The fault structure handed to the callback
112 * vm callbacks like fault() and *_mkwrite() allow for the mm_sem to be dropped
113 * during long waits, and after the wait the callback will be restarted. This
114 * is to allow other threads using the same virtual memory space concurrent
115 * access to map(), unmap() completely unrelated buffer objects. TTM buffer
116 * object reservations sometimes wait for GPU and should therefore be
117 * considered long waits. This function reserves the buffer object interruptibly
118 * taking this into account. Starvation is avoided by the vm system not
119 * allowing too many repeated restarts.
120 * This function is intended to be used in customized fault() and _mkwrite()
124 * 0 on success and the bo was reserved.
125 * VM_FAULT_RETRY if blocking wait.
126 * VM_FAULT_NOPAGE if blocking wait and retrying was not allowed.
128 vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo,
129 struct vm_fault *vmf)
132 * Work around locking order reversal in fault / nopfn
133 * between mmap_sem and bo_reserve: Perform a trylock operation
134 * for reserve, and if it fails, retry the fault after waiting
135 * for the buffer to become unreserved.
137 if (unlikely(!dma_resv_trylock(bo->base.resv))) {
138 if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
139 if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
141 up_read(&vmf->vma->vm_mm->mmap_sem);
142 if (!dma_resv_lock_interruptible(bo->base.resv,
144 dma_resv_unlock(bo->base.resv);
148 return VM_FAULT_RETRY;
151 if (dma_resv_lock_interruptible(bo->base.resv, NULL))
152 return VM_FAULT_NOPAGE;
157 EXPORT_SYMBOL(ttm_bo_vm_reserve);
159 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
161 * ttm_bo_vm_insert_huge - Insert a pfn for PUD or PMD faults
163 * @bo: The buffer object
164 * @page_offset: Page offset from bo start
165 * @fault_page_size: The size of the fault in pages.
166 * @pgprot: The page protections.
167 * Does additional checking whether it's possible to insert a PUD or PMD
168 * pfn and performs the insertion.
170 * Return: VM_FAULT_NOPAGE on successful insertion, VM_FAULT_FALLBACK if
171 * a huge fault was not possible, or on insertion error.
173 static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
174 struct ttm_buffer_object *bo,
176 pgoff_t fault_page_size,
183 struct ttm_tt *ttm = bo->ttm;
184 bool write = vmf->flags & FAULT_FLAG_WRITE;
186 /* Fault should not cross bo boundary. */
187 page_offset &= ~(fault_page_size - 1);
188 if (page_offset + fault_page_size > bo->num_pages)
191 if (bo->mem.bus.is_iomem)
192 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
194 pfn = page_to_pfn(ttm->pages[page_offset]);
196 /* pfn must be fault_page_size aligned. */
197 if ((pfn & (fault_page_size - 1)) != 0)
200 /* Check that memory is contiguous. */
201 if (!bo->mem.bus.is_iomem) {
202 for (i = 1; i < fault_page_size; ++i) {
203 if (page_to_pfn(ttm->pages[page_offset + i]) != pfn + i)
206 } else if (bo->bdev->driver->io_mem_pfn) {
207 for (i = 1; i < fault_page_size; ++i) {
208 if (ttm_bo_io_mem_pfn(bo, page_offset + i) != pfn + i)
213 pfnt = __pfn_to_pfn_t(pfn, PFN_DEV);
214 if (fault_page_size == (HPAGE_PMD_SIZE >> PAGE_SHIFT))
215 ret = vmf_insert_pfn_pmd_prot(vmf, pfnt, pgprot, write);
216 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
217 else if (fault_page_size == (HPAGE_PUD_SIZE >> PAGE_SHIFT))
218 ret = vmf_insert_pfn_pud_prot(vmf, pfnt, pgprot, write);
221 WARN_ON_ONCE(ret = VM_FAULT_FALLBACK);
223 if (ret != VM_FAULT_NOPAGE)
226 return VM_FAULT_NOPAGE;
228 count_vm_event(THP_FAULT_FALLBACK);
229 return VM_FAULT_FALLBACK;
232 static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
233 struct ttm_buffer_object *bo,
235 pgoff_t fault_page_size,
238 return VM_FAULT_FALLBACK;
243 * ttm_bo_vm_fault_reserved - TTM fault helper
244 * @vmf: The struct vm_fault given as argument to the fault callback
245 * @prot: The page protection to be used for this memory area.
246 * @num_prefault: Maximum number of prefault pages. The caller may want to
247 * specify this based on madvice settings and the size of the GPU object
248 * backed by the memory.
249 * @fault_page_size: The size of the fault in pages.
251 * This function inserts one or more page table entries pointing to the
252 * memory backing the buffer object, and then returns a return code
253 * instructing the caller to retry the page access.
256 * VM_FAULT_NOPAGE on success or pending signal
257 * VM_FAULT_SIGBUS on unspecified error
258 * VM_FAULT_OOM on out-of-memory
259 * VM_FAULT_RETRY if retryable wait
261 vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
263 pgoff_t num_prefault,
264 pgoff_t fault_page_size)
266 struct vm_area_struct *vma = vmf->vma;
267 struct ttm_buffer_object *bo = vma->vm_private_data;
268 struct ttm_bo_device *bdev = bo->bdev;
269 unsigned long page_offset;
270 unsigned long page_last;
272 struct ttm_tt *ttm = NULL;
276 vm_fault_t ret = VM_FAULT_NOPAGE;
277 unsigned long address = vmf->address;
278 struct ttm_mem_type_manager *man =
279 &bdev->man[bo->mem.mem_type];
282 * Refuse to fault imported pages. This should be handled
283 * (if at all) by redirecting mmap to the exporter.
285 if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG))
286 return VM_FAULT_SIGBUS;
288 if (bdev->driver->fault_reserve_notify) {
289 struct dma_fence *moving = dma_fence_get(bo->moving);
291 err = bdev->driver->fault_reserve_notify(bo);
297 return VM_FAULT_NOPAGE;
299 return VM_FAULT_SIGBUS;
302 if (bo->moving != moving) {
303 spin_lock(&ttm_bo_glob.lru_lock);
304 ttm_bo_move_to_lru_tail(bo, NULL);
305 spin_unlock(&ttm_bo_glob.lru_lock);
307 dma_fence_put(moving);
311 * Wait for buffer data in transit, due to a pipelined
314 ret = ttm_bo_vm_fault_idle(bo, vmf);
315 if (unlikely(ret != 0))
318 err = ttm_mem_io_lock(man, true);
319 if (unlikely(err != 0))
320 return VM_FAULT_NOPAGE;
321 err = ttm_mem_io_reserve_vm(bo);
322 if (unlikely(err != 0)) {
323 ret = VM_FAULT_SIGBUS;
327 page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
328 vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
329 page_last = vma_pages(vma) + vma->vm_pgoff -
330 drm_vma_node_start(&bo->base.vma_node);
332 if (unlikely(page_offset >= bo->num_pages)) {
333 ret = VM_FAULT_SIGBUS;
337 prot = ttm_io_prot(bo->mem.placement, prot);
338 if (!bo->mem.bus.is_iomem) {
339 struct ttm_operation_ctx ctx = {
340 .interruptible = false,
341 .no_wait_gpu = false,
342 .flags = TTM_OPT_FLAG_FORCE_ALLOC
347 if (ttm_tt_populate(bo->ttm, &ctx)) {
352 /* Iomem should not be marked encrypted */
353 prot = pgprot_decrypted(prot);
356 /* We don't prefault on huge faults. Yet. */
357 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && fault_page_size != 1) {
358 ret = ttm_bo_vm_insert_huge(vmf, bo, page_offset,
359 fault_page_size, prot);
364 * Speculatively prefault a number of pages. Only error on
367 for (i = 0; i < num_prefault; ++i) {
368 if (bo->mem.bus.is_iomem) {
369 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
371 page = ttm->pages[page_offset];
372 if (unlikely(!page && i == 0)) {
375 } else if (unlikely(!page)) {
378 page->index = drm_vma_node_start(&bo->base.vma_node) +
380 pfn = page_to_pfn(page);
384 * Note that the value of @prot at this point may differ from
385 * the value of @vma->vm_page_prot in the caching- and
386 * encryption bits. This is because the exact location of the
387 * data may not be known at mmap() time and may also change
388 * at arbitrary times while the data is mmap'ed.
389 * See vmf_insert_mixed_prot() for a discussion.
391 if (vma->vm_flags & VM_MIXEDMAP)
392 ret = vmf_insert_mixed_prot(vma, address,
393 __pfn_to_pfn_t(pfn, PFN_DEV),
396 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
398 /* Never error on prefaulted PTEs */
399 if (unlikely((ret & VM_FAULT_ERROR))) {
406 address += PAGE_SIZE;
407 if (unlikely(++page_offset >= page_last))
410 ret = VM_FAULT_NOPAGE;
412 ttm_mem_io_unlock(man);
415 EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
417 vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
419 struct vm_area_struct *vma = vmf->vma;
421 struct ttm_buffer_object *bo = vma->vm_private_data;
424 ret = ttm_bo_vm_reserve(bo, vmf);
428 prot = vma->vm_page_prot;
429 ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT, 1);
430 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
433 dma_resv_unlock(bo->base.resv);
437 EXPORT_SYMBOL(ttm_bo_vm_fault);
439 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
441 * ttm_pgprot_is_wrprotecting - Is a page protection value write-protecting?
442 * @prot: The page protection value
444 * Return: true if @prot is write-protecting. false otherwise.
446 static bool ttm_pgprot_is_wrprotecting(pgprot_t prot)
449 * This is meant to say "pgprot_wrprotect(prot) == prot" in a generic
450 * way. Unfortunately there is no generic pgprot_wrprotect.
452 return pte_val(pte_wrprotect(__pte(pgprot_val(prot)))) ==
456 static vm_fault_t ttm_bo_vm_huge_fault(struct vm_fault *vmf,
457 enum page_entry_size pe_size)
459 struct vm_area_struct *vma = vmf->vma;
461 struct ttm_buffer_object *bo = vma->vm_private_data;
463 pgoff_t fault_page_size = 0;
464 bool write = vmf->flags & FAULT_FLAG_WRITE;
468 fault_page_size = HPAGE_PMD_SIZE >> PAGE_SHIFT;
470 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
472 fault_page_size = HPAGE_PUD_SIZE >> PAGE_SHIFT;
477 return VM_FAULT_FALLBACK;
480 /* Fallback on write dirty-tracking or COW */
481 if (write && ttm_pgprot_is_wrprotecting(vma->vm_page_prot))
482 return VM_FAULT_FALLBACK;
484 ret = ttm_bo_vm_reserve(bo, vmf);
488 prot = vm_get_page_prot(vma->vm_flags);
489 ret = ttm_bo_vm_fault_reserved(vmf, prot, 1, fault_page_size);
490 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
493 dma_resv_unlock(bo->base.resv);
499 void ttm_bo_vm_open(struct vm_area_struct *vma)
501 struct ttm_buffer_object *bo = vma->vm_private_data;
503 WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
507 EXPORT_SYMBOL(ttm_bo_vm_open);
509 void ttm_bo_vm_close(struct vm_area_struct *vma)
511 struct ttm_buffer_object *bo = vma->vm_private_data;
514 vma->vm_private_data = NULL;
516 EXPORT_SYMBOL(ttm_bo_vm_close);
518 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
519 unsigned long offset,
520 uint8_t *buf, int len, int write)
522 unsigned long page = offset >> PAGE_SHIFT;
523 unsigned long bytes_left = len;
526 /* Copy a page at a time, that way no extra virtual address
529 offset -= page << PAGE_SHIFT;
531 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
532 struct ttm_bo_kmap_obj map;
536 ret = ttm_bo_kmap(bo, page, 1, &map);
540 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
541 WARN_ON_ONCE(is_iomem);
543 memcpy(ptr, buf, bytes);
545 memcpy(buf, ptr, bytes);
552 } while (bytes_left);
557 int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
558 void *buf, int len, int write)
560 unsigned long offset = (addr) - vma->vm_start;
561 struct ttm_buffer_object *bo = vma->vm_private_data;
564 if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->num_pages)
567 ret = ttm_bo_reserve(bo, true, false, NULL);
571 switch (bo->mem.mem_type) {
573 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
574 ret = ttm_tt_swapin(bo->ttm);
575 if (unlikely(ret != 0))
580 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
583 if (bo->bdev->driver->access_memory)
584 ret = bo->bdev->driver->access_memory(
585 bo, offset, buf, len, write);
590 ttm_bo_unreserve(bo);
594 EXPORT_SYMBOL(ttm_bo_vm_access);
596 static const struct vm_operations_struct ttm_bo_vm_ops = {
597 .fault = ttm_bo_vm_fault,
598 .open = ttm_bo_vm_open,
599 .close = ttm_bo_vm_close,
600 .access = ttm_bo_vm_access,
601 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
602 .huge_fault = ttm_bo_vm_huge_fault,
606 static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
607 unsigned long offset,
610 struct drm_vma_offset_node *node;
611 struct ttm_buffer_object *bo = NULL;
613 drm_vma_offset_lock_lookup(bdev->vma_manager);
615 node = drm_vma_offset_lookup_locked(bdev->vma_manager, offset, pages);
617 bo = container_of(node, struct ttm_buffer_object,
619 bo = ttm_bo_get_unless_zero(bo);
622 drm_vma_offset_unlock_lookup(bdev->vma_manager);
625 pr_err("Could not find buffer object to map\n");
630 static void ttm_bo_mmap_vma_setup(struct ttm_buffer_object *bo, struct vm_area_struct *vma)
632 vma->vm_ops = &ttm_bo_vm_ops;
635 * Note: We're transferring the bo reference to
636 * vma->vm_private_data here.
639 vma->vm_private_data = bo;
642 * We'd like to use VM_PFNMAP on shared mappings, where
643 * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
644 * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
645 * bad for performance. Until that has been sorted out, use
646 * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
648 vma->vm_flags |= VM_MIXEDMAP;
649 vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
652 int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
653 struct ttm_bo_device *bdev)
655 struct ttm_bo_driver *driver;
656 struct ttm_buffer_object *bo;
659 if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET_START))
662 bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
666 driver = bo->bdev->driver;
667 if (unlikely(!driver->verify_access)) {
671 ret = driver->verify_access(bo, filp);
672 if (unlikely(ret != 0))
675 ttm_bo_mmap_vma_setup(bo, vma);
681 EXPORT_SYMBOL(ttm_bo_mmap);
683 int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
686 ttm_bo_mmap_vma_setup(bo, vma);
689 EXPORT_SYMBOL(ttm_bo_mmap_obj);