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_bo_driver.h>
35 #include <drm/ttm/ttm_placement.h>
36 #include <drm/drm_vma_manager.h>
37 #include <drm/drm_drv.h>
38 #include <drm/drm_managed.h>
40 #include <linux/pfn_t.h>
41 #include <linux/rbtree.h>
42 #include <linux/module.h>
43 #include <linux/uaccess.h>
44 #include <linux/mem_encrypt.h>
46 static vm_fault_t ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
52 if (likely(!bo->moving))
56 * Quick non-stalling check for idle.
58 if (dma_fence_is_signaled(bo->moving))
62 * If possible, avoid waiting for GPU with mmap_lock
63 * held. We only do this if the fault allows retry and this
64 * is the first attempt.
66 if (fault_flag_allow_retry_first(vmf->flags)) {
68 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
72 mmap_read_unlock(vmf->vma->vm_mm);
73 (void) dma_fence_wait(bo->moving, true);
74 dma_resv_unlock(bo->base.resv);
82 err = dma_fence_wait(bo->moving, true);
83 if (unlikely(err != 0)) {
84 ret = (err != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
90 dma_fence_put(bo->moving);
97 static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object *bo,
98 unsigned long page_offset)
100 struct ttm_device *bdev = bo->bdev;
102 if (bdev->funcs->io_mem_pfn)
103 return bdev->funcs->io_mem_pfn(bo, page_offset);
105 return (bo->resource->bus.offset >> PAGE_SHIFT) + page_offset;
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_lock 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_lock
144 if (fault_flag_allow_retry_first(vmf->flags)) {
145 if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
147 mmap_read_unlock(vmf->vma->vm_mm);
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;
162 * Refuse to fault imported pages. This should be handled
163 * (if at all) by redirecting mmap to the exporter.
165 if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
166 dma_resv_unlock(bo->base.resv);
167 return VM_FAULT_SIGBUS;
172 EXPORT_SYMBOL(ttm_bo_vm_reserve);
174 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
176 * ttm_bo_vm_insert_huge - Insert a pfn for PUD or PMD faults
178 * @bo: The buffer object
179 * @page_offset: Page offset from bo start
180 * @fault_page_size: The size of the fault in pages.
181 * @pgprot: The page protections.
182 * Does additional checking whether it's possible to insert a PUD or PMD
183 * pfn and performs the insertion.
185 * Return: VM_FAULT_NOPAGE on successful insertion, VM_FAULT_FALLBACK if
186 * a huge fault was not possible, or on insertion error.
188 static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
189 struct ttm_buffer_object *bo,
191 pgoff_t fault_page_size,
198 struct ttm_tt *ttm = bo->ttm;
199 bool write = vmf->flags & FAULT_FLAG_WRITE;
201 /* Fault should not cross bo boundary. */
202 page_offset &= ~(fault_page_size - 1);
203 if (page_offset + fault_page_size > bo->resource->num_pages)
206 if (bo->resource->bus.is_iomem)
207 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
209 pfn = page_to_pfn(ttm->pages[page_offset]);
211 /* pfn must be fault_page_size aligned. */
212 if ((pfn & (fault_page_size - 1)) != 0)
215 /* Check that memory is contiguous. */
216 if (!bo->resource->bus.is_iomem) {
217 for (i = 1; i < fault_page_size; ++i) {
218 if (page_to_pfn(ttm->pages[page_offset + i]) != pfn + i)
221 } else if (bo->bdev->funcs->io_mem_pfn) {
222 for (i = 1; i < fault_page_size; ++i) {
223 if (ttm_bo_io_mem_pfn(bo, page_offset + i) != pfn + i)
228 pfnt = __pfn_to_pfn_t(pfn, PFN_DEV);
229 if (fault_page_size == (HPAGE_PMD_SIZE >> PAGE_SHIFT))
230 ret = vmf_insert_pfn_pmd_prot(vmf, pfnt, pgprot, write);
231 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
232 else if (fault_page_size == (HPAGE_PUD_SIZE >> PAGE_SHIFT))
233 ret = vmf_insert_pfn_pud_prot(vmf, pfnt, pgprot, write);
236 WARN_ON_ONCE(ret = VM_FAULT_FALLBACK);
238 if (ret != VM_FAULT_NOPAGE)
241 return VM_FAULT_NOPAGE;
243 count_vm_event(THP_FAULT_FALLBACK);
244 return VM_FAULT_FALLBACK;
247 static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
248 struct ttm_buffer_object *bo,
250 pgoff_t fault_page_size,
253 return VM_FAULT_FALLBACK;
258 * ttm_bo_vm_fault_reserved - TTM fault helper
259 * @vmf: The struct vm_fault given as argument to the fault callback
260 * @prot: The page protection to be used for this memory area.
261 * @num_prefault: Maximum number of prefault pages. The caller may want to
262 * specify this based on madvice settings and the size of the GPU object
263 * backed by the memory.
264 * @fault_page_size: The size of the fault in pages.
266 * This function inserts one or more page table entries pointing to the
267 * memory backing the buffer object, and then returns a return code
268 * instructing the caller to retry the page access.
271 * VM_FAULT_NOPAGE on success or pending signal
272 * VM_FAULT_SIGBUS on unspecified error
273 * VM_FAULT_OOM on out-of-memory
274 * VM_FAULT_RETRY if retryable wait
276 vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
278 pgoff_t num_prefault,
279 pgoff_t fault_page_size)
281 struct vm_area_struct *vma = vmf->vma;
282 struct ttm_buffer_object *bo = vma->vm_private_data;
283 struct ttm_device *bdev = bo->bdev;
284 unsigned long page_offset;
285 unsigned long page_last;
287 struct ttm_tt *ttm = NULL;
291 vm_fault_t ret = VM_FAULT_NOPAGE;
292 unsigned long address = vmf->address;
295 * Wait for buffer data in transit, due to a pipelined
298 ret = ttm_bo_vm_fault_idle(bo, vmf);
299 if (unlikely(ret != 0))
302 err = ttm_mem_io_reserve(bdev, bo->resource);
303 if (unlikely(err != 0))
304 return VM_FAULT_SIGBUS;
306 page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
307 vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
308 page_last = vma_pages(vma) + vma->vm_pgoff -
309 drm_vma_node_start(&bo->base.vma_node);
311 if (unlikely(page_offset >= bo->resource->num_pages))
312 return VM_FAULT_SIGBUS;
314 prot = ttm_io_prot(bo, bo->resource, prot);
315 if (!bo->resource->bus.is_iomem) {
316 struct ttm_operation_ctx ctx = {
317 .interruptible = false,
318 .no_wait_gpu = false,
323 if (ttm_tt_populate(bdev, bo->ttm, &ctx))
326 /* Iomem should not be marked encrypted */
327 prot = pgprot_decrypted(prot);
330 /* We don't prefault on huge faults. Yet. */
331 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && fault_page_size != 1)
332 return ttm_bo_vm_insert_huge(vmf, bo, page_offset,
333 fault_page_size, prot);
336 * Speculatively prefault a number of pages. Only error on
339 for (i = 0; i < num_prefault; ++i) {
340 if (bo->resource->bus.is_iomem) {
341 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
343 page = ttm->pages[page_offset];
344 if (unlikely(!page && i == 0)) {
346 } else if (unlikely(!page)) {
349 page->index = drm_vma_node_start(&bo->base.vma_node) +
351 pfn = page_to_pfn(page);
355 * Note that the value of @prot at this point may differ from
356 * the value of @vma->vm_page_prot in the caching- and
357 * encryption bits. This is because the exact location of the
358 * data may not be known at mmap() time and may also change
359 * at arbitrary times while the data is mmap'ed.
360 * See vmf_insert_mixed_prot() for a discussion.
362 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
364 /* Never error on prefaulted PTEs */
365 if (unlikely((ret & VM_FAULT_ERROR))) {
367 return VM_FAULT_NOPAGE;
372 address += PAGE_SIZE;
373 if (unlikely(++page_offset >= page_last))
378 EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
380 static void ttm_bo_release_dummy_page(struct drm_device *dev, void *res)
382 struct page *dummy_page = (struct page *)res;
384 __free_page(dummy_page);
387 vm_fault_t ttm_bo_vm_dummy_page(struct vm_fault *vmf, pgprot_t prot)
389 struct vm_area_struct *vma = vmf->vma;
390 struct ttm_buffer_object *bo = vma->vm_private_data;
391 struct drm_device *ddev = bo->base.dev;
392 vm_fault_t ret = VM_FAULT_NOPAGE;
393 unsigned long address;
397 /* Allocate new dummy page to map all the VA range in this VMA to it*/
398 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
402 /* Set the page to be freed using drmm release action */
403 if (drmm_add_action_or_reset(ddev, ttm_bo_release_dummy_page, page))
406 pfn = page_to_pfn(page);
408 /* Prefault the entire VMA range right away to avoid further faults */
409 for (address = vma->vm_start; address < vma->vm_end;
410 address += PAGE_SIZE)
411 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
415 EXPORT_SYMBOL(ttm_bo_vm_dummy_page);
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;
422 struct drm_device *ddev = bo->base.dev;
426 ret = ttm_bo_vm_reserve(bo, vmf);
430 prot = vma->vm_page_prot;
431 if (drm_dev_enter(ddev, &idx)) {
432 ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT, 1);
435 ret = ttm_bo_vm_dummy_page(vmf, prot);
437 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
440 dma_resv_unlock(bo->base.resv);
444 EXPORT_SYMBOL(ttm_bo_vm_fault);
446 void ttm_bo_vm_open(struct vm_area_struct *vma)
448 struct ttm_buffer_object *bo = vma->vm_private_data;
450 WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
454 EXPORT_SYMBOL(ttm_bo_vm_open);
456 void ttm_bo_vm_close(struct vm_area_struct *vma)
458 struct ttm_buffer_object *bo = vma->vm_private_data;
461 vma->vm_private_data = NULL;
463 EXPORT_SYMBOL(ttm_bo_vm_close);
465 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
466 unsigned long offset,
467 uint8_t *buf, int len, int write)
469 unsigned long page = offset >> PAGE_SHIFT;
470 unsigned long bytes_left = len;
473 /* Copy a page at a time, that way no extra virtual address
476 offset -= page << PAGE_SHIFT;
478 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
479 struct ttm_bo_kmap_obj map;
483 ret = ttm_bo_kmap(bo, page, 1, &map);
487 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
488 WARN_ON_ONCE(is_iomem);
490 memcpy(ptr, buf, bytes);
492 memcpy(buf, ptr, bytes);
499 } while (bytes_left);
504 int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
505 void *buf, int len, int write)
507 struct ttm_buffer_object *bo = vma->vm_private_data;
508 unsigned long offset = (addr) - vma->vm_start +
509 ((vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node))
513 if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->resource->num_pages)
516 ret = ttm_bo_reserve(bo, true, false, NULL);
520 switch (bo->resource->mem_type) {
522 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
523 ret = ttm_tt_swapin(bo->ttm);
524 if (unlikely(ret != 0))
529 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
532 if (bo->bdev->funcs->access_memory)
533 ret = bo->bdev->funcs->access_memory(
534 bo, offset, buf, len, write);
539 ttm_bo_unreserve(bo);
543 EXPORT_SYMBOL(ttm_bo_vm_access);
545 static const struct vm_operations_struct ttm_bo_vm_ops = {
546 .fault = ttm_bo_vm_fault,
547 .open = ttm_bo_vm_open,
548 .close = ttm_bo_vm_close,
549 .access = ttm_bo_vm_access,
552 int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
554 /* Enforce no COW since would have really strange behavior with it. */
555 if (is_cow_mapping(vma->vm_flags))
561 * Drivers may want to override the vm_ops field. Otherwise we
562 * use TTM's default callbacks.
565 vma->vm_ops = &ttm_bo_vm_ops;
568 * Note: We're transferring the bo reference to
569 * vma->vm_private_data here.
572 vma->vm_private_data = bo;
574 vma->vm_flags |= VM_PFNMAP;
575 vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
578 EXPORT_SYMBOL(ttm_bo_mmap_obj);