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);
160 * ttm_bo_vm_fault_reserved - TTM fault helper
161 * @vmf: The struct vm_fault given as argument to the fault callback
162 * @prot: The page protection to be used for this memory area.
163 * @num_prefault: Maximum number of prefault pages. The caller may want to
164 * specify this based on madvice settings and the size of the GPU object
165 * backed by the memory.
167 * This function inserts one or more page table entries pointing to the
168 * memory backing the buffer object, and then returns a return code
169 * instructing the caller to retry the page access.
172 * VM_FAULT_NOPAGE on success or pending signal
173 * VM_FAULT_SIGBUS on unspecified error
174 * VM_FAULT_OOM on out-of-memory
175 * VM_FAULT_RETRY if retryable wait
177 vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
179 pgoff_t num_prefault)
181 struct vm_area_struct *vma = vmf->vma;
182 struct ttm_buffer_object *bo = vma->vm_private_data;
183 struct ttm_bo_device *bdev = bo->bdev;
184 unsigned long page_offset;
185 unsigned long page_last;
187 struct ttm_tt *ttm = NULL;
191 vm_fault_t ret = VM_FAULT_NOPAGE;
192 unsigned long address = vmf->address;
193 struct ttm_mem_type_manager *man =
194 &bdev->man[bo->mem.mem_type];
197 * Refuse to fault imported pages. This should be handled
198 * (if at all) by redirecting mmap to the exporter.
200 if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG))
201 return VM_FAULT_SIGBUS;
203 if (bdev->driver->fault_reserve_notify) {
204 struct dma_fence *moving = dma_fence_get(bo->moving);
206 err = bdev->driver->fault_reserve_notify(bo);
212 return VM_FAULT_NOPAGE;
214 return VM_FAULT_SIGBUS;
217 if (bo->moving != moving) {
218 spin_lock(&ttm_bo_glob.lru_lock);
219 ttm_bo_move_to_lru_tail(bo, NULL);
220 spin_unlock(&ttm_bo_glob.lru_lock);
222 dma_fence_put(moving);
226 * Wait for buffer data in transit, due to a pipelined
229 ret = ttm_bo_vm_fault_idle(bo, vmf);
230 if (unlikely(ret != 0))
233 err = ttm_mem_io_lock(man, true);
234 if (unlikely(err != 0))
235 return VM_FAULT_NOPAGE;
236 err = ttm_mem_io_reserve_vm(bo);
237 if (unlikely(err != 0)) {
238 ret = VM_FAULT_SIGBUS;
242 page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
243 vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
244 page_last = vma_pages(vma) + vma->vm_pgoff -
245 drm_vma_node_start(&bo->base.vma_node);
247 if (unlikely(page_offset >= bo->num_pages)) {
248 ret = VM_FAULT_SIGBUS;
252 prot = ttm_io_prot(bo->mem.placement, prot);
253 if (!bo->mem.bus.is_iomem) {
254 struct ttm_operation_ctx ctx = {
255 .interruptible = false,
256 .no_wait_gpu = false,
257 .flags = TTM_OPT_FLAG_FORCE_ALLOC
262 if (ttm_tt_populate(bo->ttm, &ctx)) {
267 /* Iomem should not be marked encrypted */
268 prot = pgprot_decrypted(prot);
272 * Speculatively prefault a number of pages. Only error on
275 for (i = 0; i < num_prefault; ++i) {
276 if (bo->mem.bus.is_iomem) {
277 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
279 page = ttm->pages[page_offset];
280 if (unlikely(!page && i == 0)) {
283 } else if (unlikely(!page)) {
286 page->index = drm_vma_node_start(&bo->base.vma_node) +
288 pfn = page_to_pfn(page);
292 * Note that the value of @prot at this point may differ from
293 * the value of @vma->vm_page_prot in the caching- and
294 * encryption bits. This is because the exact location of the
295 * data may not be known at mmap() time and may also change
296 * at arbitrary times while the data is mmap'ed.
297 * See vmf_insert_mixed_prot() for a discussion.
299 if (vma->vm_flags & VM_MIXEDMAP)
300 ret = vmf_insert_mixed_prot(vma, address,
301 __pfn_to_pfn_t(pfn, PFN_DEV),
304 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
306 /* Never error on prefaulted PTEs */
307 if (unlikely((ret & VM_FAULT_ERROR))) {
314 address += PAGE_SIZE;
315 if (unlikely(++page_offset >= page_last))
318 ret = VM_FAULT_NOPAGE;
320 ttm_mem_io_unlock(man);
323 EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
325 vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
327 struct vm_area_struct *vma = vmf->vma;
329 struct ttm_buffer_object *bo = vma->vm_private_data;
332 ret = ttm_bo_vm_reserve(bo, vmf);
336 prot = vma->vm_page_prot;
337 ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT);
338 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
341 dma_resv_unlock(bo->base.resv);
345 EXPORT_SYMBOL(ttm_bo_vm_fault);
347 void ttm_bo_vm_open(struct vm_area_struct *vma)
349 struct ttm_buffer_object *bo = vma->vm_private_data;
351 WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
355 EXPORT_SYMBOL(ttm_bo_vm_open);
357 void ttm_bo_vm_close(struct vm_area_struct *vma)
359 struct ttm_buffer_object *bo = vma->vm_private_data;
362 vma->vm_private_data = NULL;
364 EXPORT_SYMBOL(ttm_bo_vm_close);
366 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
367 unsigned long offset,
368 uint8_t *buf, int len, int write)
370 unsigned long page = offset >> PAGE_SHIFT;
371 unsigned long bytes_left = len;
374 /* Copy a page at a time, that way no extra virtual address
377 offset -= page << PAGE_SHIFT;
379 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
380 struct ttm_bo_kmap_obj map;
384 ret = ttm_bo_kmap(bo, page, 1, &map);
388 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
389 WARN_ON_ONCE(is_iomem);
391 memcpy(ptr, buf, bytes);
393 memcpy(buf, ptr, bytes);
400 } while (bytes_left);
405 int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
406 void *buf, int len, int write)
408 unsigned long offset = (addr) - vma->vm_start;
409 struct ttm_buffer_object *bo = vma->vm_private_data;
412 if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->num_pages)
415 ret = ttm_bo_reserve(bo, true, false, NULL);
419 switch (bo->mem.mem_type) {
421 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
422 ret = ttm_tt_swapin(bo->ttm);
423 if (unlikely(ret != 0))
428 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
431 if (bo->bdev->driver->access_memory)
432 ret = bo->bdev->driver->access_memory(
433 bo, offset, buf, len, write);
438 ttm_bo_unreserve(bo);
442 EXPORT_SYMBOL(ttm_bo_vm_access);
444 static const struct vm_operations_struct ttm_bo_vm_ops = {
445 .fault = ttm_bo_vm_fault,
446 .open = ttm_bo_vm_open,
447 .close = ttm_bo_vm_close,
448 .access = ttm_bo_vm_access
451 static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
452 unsigned long offset,
455 struct drm_vma_offset_node *node;
456 struct ttm_buffer_object *bo = NULL;
458 drm_vma_offset_lock_lookup(bdev->vma_manager);
460 node = drm_vma_offset_lookup_locked(bdev->vma_manager, offset, pages);
462 bo = container_of(node, struct ttm_buffer_object,
464 bo = ttm_bo_get_unless_zero(bo);
467 drm_vma_offset_unlock_lookup(bdev->vma_manager);
470 pr_err("Could not find buffer object to map\n");
475 static void ttm_bo_mmap_vma_setup(struct ttm_buffer_object *bo, struct vm_area_struct *vma)
477 vma->vm_ops = &ttm_bo_vm_ops;
480 * Note: We're transferring the bo reference to
481 * vma->vm_private_data here.
484 vma->vm_private_data = bo;
487 * We'd like to use VM_PFNMAP on shared mappings, where
488 * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
489 * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
490 * bad for performance. Until that has been sorted out, use
491 * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
493 vma->vm_flags |= VM_MIXEDMAP;
494 vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
497 int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
498 struct ttm_bo_device *bdev)
500 struct ttm_bo_driver *driver;
501 struct ttm_buffer_object *bo;
504 if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET_START))
507 bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
511 driver = bo->bdev->driver;
512 if (unlikely(!driver->verify_access)) {
516 ret = driver->verify_access(bo, filp);
517 if (unlikely(ret != 0))
520 ttm_bo_mmap_vma_setup(bo, vma);
526 EXPORT_SYMBOL(ttm_bo_mmap);
528 int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
531 ttm_bo_mmap_vma_setup(bo, vma);
534 EXPORT_SYMBOL(ttm_bo_mmap_obj);