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_lock
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 mmap_read_unlock(vmf->vma->vm_mm);
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.offset >> PAGE_SHIFT) + page_offset;
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_lock 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))) {
139 * If the fault allows retry and this is the first
140 * fault attempt, we try to release the mmap_lock
143 if (fault_flag_allow_retry_first(vmf->flags)) {
144 if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
146 mmap_read_unlock(vmf->vma->vm_mm);
147 if (!dma_resv_lock_interruptible(bo->base.resv,
149 dma_resv_unlock(bo->base.resv);
153 return VM_FAULT_RETRY;
156 if (dma_resv_lock_interruptible(bo->base.resv, NULL))
157 return VM_FAULT_NOPAGE;
161 * Refuse to fault imported pages. This should be handled
162 * (if at all) by redirecting mmap to the exporter.
164 if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
165 dma_resv_unlock(bo->base.resv);
166 return VM_FAULT_SIGBUS;
171 EXPORT_SYMBOL(ttm_bo_vm_reserve);
173 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
175 * ttm_bo_vm_insert_huge - Insert a pfn for PUD or PMD faults
177 * @bo: The buffer object
178 * @page_offset: Page offset from bo start
179 * @fault_page_size: The size of the fault in pages.
180 * @pgprot: The page protections.
181 * Does additional checking whether it's possible to insert a PUD or PMD
182 * pfn and performs the insertion.
184 * Return: VM_FAULT_NOPAGE on successful insertion, VM_FAULT_FALLBACK if
185 * a huge fault was not possible, or on insertion error.
187 static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
188 struct ttm_buffer_object *bo,
190 pgoff_t fault_page_size,
197 struct ttm_tt *ttm = bo->ttm;
198 bool write = vmf->flags & FAULT_FLAG_WRITE;
200 /* Fault should not cross bo boundary. */
201 page_offset &= ~(fault_page_size - 1);
202 if (page_offset + fault_page_size > bo->num_pages)
205 if (bo->mem.bus.is_iomem)
206 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
208 pfn = page_to_pfn(ttm->pages[page_offset]);
210 /* pfn must be fault_page_size aligned. */
211 if ((pfn & (fault_page_size - 1)) != 0)
214 /* Check that memory is contiguous. */
215 if (!bo->mem.bus.is_iomem) {
216 for (i = 1; i < fault_page_size; ++i) {
217 if (page_to_pfn(ttm->pages[page_offset + i]) != pfn + i)
220 } else if (bo->bdev->driver->io_mem_pfn) {
221 for (i = 1; i < fault_page_size; ++i) {
222 if (ttm_bo_io_mem_pfn(bo, page_offset + i) != pfn + i)
227 pfnt = __pfn_to_pfn_t(pfn, PFN_DEV);
228 if (fault_page_size == (HPAGE_PMD_SIZE >> PAGE_SHIFT))
229 ret = vmf_insert_pfn_pmd_prot(vmf, pfnt, pgprot, write);
230 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
231 else if (fault_page_size == (HPAGE_PUD_SIZE >> PAGE_SHIFT))
232 ret = vmf_insert_pfn_pud_prot(vmf, pfnt, pgprot, write);
235 WARN_ON_ONCE(ret = VM_FAULT_FALLBACK);
237 if (ret != VM_FAULT_NOPAGE)
240 return VM_FAULT_NOPAGE;
242 count_vm_event(THP_FAULT_FALLBACK);
243 return VM_FAULT_FALLBACK;
246 static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
247 struct ttm_buffer_object *bo,
249 pgoff_t fault_page_size,
252 return VM_FAULT_FALLBACK;
257 * ttm_bo_vm_fault_reserved - TTM fault helper
258 * @vmf: The struct vm_fault given as argument to the fault callback
259 * @prot: The page protection to be used for this memory area.
260 * @num_prefault: Maximum number of prefault pages. The caller may want to
261 * specify this based on madvice settings and the size of the GPU object
262 * backed by the memory.
263 * @fault_page_size: The size of the fault in pages.
265 * This function inserts one or more page table entries pointing to the
266 * memory backing the buffer object, and then returns a return code
267 * instructing the caller to retry the page access.
270 * VM_FAULT_NOPAGE on success or pending signal
271 * VM_FAULT_SIGBUS on unspecified error
272 * VM_FAULT_OOM on out-of-memory
273 * VM_FAULT_RETRY if retryable wait
275 vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
277 pgoff_t num_prefault,
278 pgoff_t fault_page_size)
280 struct vm_area_struct *vma = vmf->vma;
281 struct ttm_buffer_object *bo = vma->vm_private_data;
282 struct ttm_bo_device *bdev = bo->bdev;
283 unsigned long page_offset;
284 unsigned long page_last;
286 struct ttm_tt *ttm = NULL;
290 vm_fault_t ret = VM_FAULT_NOPAGE;
291 unsigned long address = vmf->address;
294 * Wait for buffer data in transit, due to a pipelined
297 ret = ttm_bo_vm_fault_idle(bo, vmf);
298 if (unlikely(ret != 0))
301 err = ttm_mem_io_reserve(bdev, &bo->mem);
302 if (unlikely(err != 0))
303 return VM_FAULT_SIGBUS;
305 page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
306 vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
307 page_last = vma_pages(vma) + vma->vm_pgoff -
308 drm_vma_node_start(&bo->base.vma_node);
310 if (unlikely(page_offset >= bo->num_pages))
311 return VM_FAULT_SIGBUS;
313 prot = ttm_io_prot(bo, &bo->mem, prot);
314 if (!bo->mem.bus.is_iomem) {
315 struct ttm_operation_ctx ctx = {
316 .interruptible = false,
317 .no_wait_gpu = false,
322 if (ttm_tt_populate(bdev, bo->ttm, &ctx))
325 /* Iomem should not be marked encrypted */
326 prot = pgprot_decrypted(prot);
329 /* We don't prefault on huge faults. Yet. */
330 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && fault_page_size != 1)
331 return ttm_bo_vm_insert_huge(vmf, bo, page_offset,
332 fault_page_size, prot);
335 * Speculatively prefault a number of pages. Only error on
338 for (i = 0; i < num_prefault; ++i) {
339 if (bo->mem.bus.is_iomem) {
340 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
342 page = ttm->pages[page_offset];
343 if (unlikely(!page && i == 0)) {
345 } else if (unlikely(!page)) {
348 page->index = drm_vma_node_start(&bo->base.vma_node) +
350 pfn = page_to_pfn(page);
354 * Note that the value of @prot at this point may differ from
355 * the value of @vma->vm_page_prot in the caching- and
356 * encryption bits. This is because the exact location of the
357 * data may not be known at mmap() time and may also change
358 * at arbitrary times while the data is mmap'ed.
359 * See vmf_insert_mixed_prot() for a discussion.
361 if (vma->vm_flags & VM_MIXEDMAP)
362 ret = vmf_insert_mixed_prot(vma, address,
363 __pfn_to_pfn_t(pfn, PFN_DEV),
366 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
368 /* Never error on prefaulted PTEs */
369 if (unlikely((ret & VM_FAULT_ERROR))) {
371 return VM_FAULT_NOPAGE;
376 address += PAGE_SIZE;
377 if (unlikely(++page_offset >= page_last))
382 EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
384 vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
386 struct vm_area_struct *vma = vmf->vma;
388 struct ttm_buffer_object *bo = vma->vm_private_data;
391 ret = ttm_bo_vm_reserve(bo, vmf);
395 prot = vma->vm_page_prot;
396 ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT, 1);
397 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
400 dma_resv_unlock(bo->base.resv);
404 EXPORT_SYMBOL(ttm_bo_vm_fault);
406 void ttm_bo_vm_open(struct vm_area_struct *vma)
408 struct ttm_buffer_object *bo = vma->vm_private_data;
410 WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
414 EXPORT_SYMBOL(ttm_bo_vm_open);
416 void ttm_bo_vm_close(struct vm_area_struct *vma)
418 struct ttm_buffer_object *bo = vma->vm_private_data;
421 vma->vm_private_data = NULL;
423 EXPORT_SYMBOL(ttm_bo_vm_close);
425 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
426 unsigned long offset,
427 uint8_t *buf, int len, int write)
429 unsigned long page = offset >> PAGE_SHIFT;
430 unsigned long bytes_left = len;
433 /* Copy a page at a time, that way no extra virtual address
436 offset -= page << PAGE_SHIFT;
438 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
439 struct ttm_bo_kmap_obj map;
443 ret = ttm_bo_kmap(bo, page, 1, &map);
447 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
448 WARN_ON_ONCE(is_iomem);
450 memcpy(ptr, buf, bytes);
452 memcpy(buf, ptr, bytes);
459 } while (bytes_left);
464 int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
465 void *buf, int len, int write)
467 struct ttm_buffer_object *bo = vma->vm_private_data;
468 unsigned long offset = (addr) - vma->vm_start +
469 ((vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node))
473 if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->num_pages)
476 ret = ttm_bo_reserve(bo, true, false, NULL);
480 switch (bo->mem.mem_type) {
482 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
483 ret = ttm_tt_swapin(bo->ttm);
484 if (unlikely(ret != 0))
489 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
492 if (bo->bdev->driver->access_memory)
493 ret = bo->bdev->driver->access_memory(
494 bo, offset, buf, len, write);
499 ttm_bo_unreserve(bo);
503 EXPORT_SYMBOL(ttm_bo_vm_access);
505 static const struct vm_operations_struct ttm_bo_vm_ops = {
506 .fault = ttm_bo_vm_fault,
507 .open = ttm_bo_vm_open,
508 .close = ttm_bo_vm_close,
509 .access = ttm_bo_vm_access,
512 static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
513 unsigned long offset,
516 struct drm_vma_offset_node *node;
517 struct ttm_buffer_object *bo = NULL;
519 drm_vma_offset_lock_lookup(bdev->vma_manager);
521 node = drm_vma_offset_lookup_locked(bdev->vma_manager, offset, pages);
523 bo = container_of(node, struct ttm_buffer_object,
525 bo = ttm_bo_get_unless_zero(bo);
528 drm_vma_offset_unlock_lookup(bdev->vma_manager);
531 pr_err("Could not find buffer object to map\n");
536 static void ttm_bo_mmap_vma_setup(struct ttm_buffer_object *bo, struct vm_area_struct *vma)
538 vma->vm_ops = &ttm_bo_vm_ops;
541 * Note: We're transferring the bo reference to
542 * vma->vm_private_data here.
545 vma->vm_private_data = bo;
548 * We'd like to use VM_PFNMAP on shared mappings, where
549 * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
550 * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
551 * bad for performance. Until that has been sorted out, use
552 * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
554 vma->vm_flags |= VM_MIXEDMAP;
555 vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
558 int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
559 struct ttm_bo_device *bdev)
561 struct ttm_bo_driver *driver;
562 struct ttm_buffer_object *bo;
565 if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET_START))
568 bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
572 driver = bo->bdev->driver;
573 if (unlikely(!driver->verify_access)) {
577 ret = driver->verify_access(bo, filp);
578 if (unlikely(ret != 0))
581 ttm_bo_mmap_vma_setup(bo, vma);
587 EXPORT_SYMBOL(ttm_bo_mmap);
589 int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
592 ttm_bo_mmap_vma_setup(bo, vma);
595 EXPORT_SYMBOL(ttm_bo_mmap_obj);