1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2 /**************************************************************************
3 *
4 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
5 * All Rights Reserved.
6 *
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:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
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.
26 *
27 **************************************************************************/
28 /*
29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
30 */
31
32 #define pr_fmt(fmt) "[TTM] " fmt
33
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>
38 #include <linux/mm.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>
44
45 #define TTM_BO_VM_NUM_PREFAULT 16
46
ttm_bo_vm_fault_idle(struct ttm_buffer_object * bo,struct vm_fault * vmf)47 static vm_fault_t ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
48 struct vm_fault *vmf)
49 {
50 vm_fault_t ret = 0;
51 int err = 0;
52
53 if (likely(!bo->moving))
54 goto out_unlock;
55
56 /*
57 * Quick non-stalling check for idle.
58 */
59 if (dma_fence_is_signaled(bo->moving))
60 goto out_clear;
61
62 /*
63 * If possible, avoid waiting for GPU with mmap_sem
64 * held.
65 */
66 if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
67 ret = VM_FAULT_RETRY;
68 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
69 goto out_unlock;
70
71 ttm_bo_get(bo);
72 up_read(&vmf->vma->vm_mm->mmap_sem);
73 (void) dma_fence_wait(bo->moving, true);
74 dma_resv_unlock(bo->base.resv);
75 ttm_bo_put(bo);
76 goto out_unlock;
77 }
78
79 /*
80 * Ordinary wait.
81 */
82 err = dma_fence_wait(bo->moving, true);
83 if (unlikely(err != 0)) {
84 ret = (err != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
85 VM_FAULT_NOPAGE;
86 goto out_unlock;
87 }
88
89 out_clear:
90 dma_fence_put(bo->moving);
91 bo->moving = NULL;
92
93 out_unlock:
94 return ret;
95 }
96
ttm_bo_io_mem_pfn(struct ttm_buffer_object * bo,unsigned long page_offset)97 static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object *bo,
98 unsigned long page_offset)
99 {
100 struct ttm_bo_device *bdev = bo->bdev;
101
102 if (bdev->driver->io_mem_pfn)
103 return bdev->driver->io_mem_pfn(bo, page_offset);
104
105 return ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT)
106 + page_offset;
107 }
108
ttm_bo_vm_fault(struct vm_fault * vmf)109 static vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
110 {
111 struct vm_area_struct *vma = vmf->vma;
112 struct ttm_buffer_object *bo = (struct ttm_buffer_object *)
113 vma->vm_private_data;
114 struct ttm_bo_device *bdev = bo->bdev;
115 unsigned long page_offset;
116 unsigned long page_last;
117 unsigned long pfn;
118 struct ttm_tt *ttm = NULL;
119 struct page *page;
120 int err;
121 int i;
122 vm_fault_t ret = VM_FAULT_NOPAGE;
123 unsigned long address = vmf->address;
124 struct ttm_mem_type_manager *man =
125 &bdev->man[bo->mem.mem_type];
126 struct vm_area_struct cvma;
127
128 /*
129 * Work around locking order reversal in fault / nopfn
130 * between mmap_sem and bo_reserve: Perform a trylock operation
131 * for reserve, and if it fails, retry the fault after waiting
132 * for the buffer to become unreserved.
133 */
134 if (unlikely(!dma_resv_trylock(bo->base.resv))) {
135 if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
136 if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
137 ttm_bo_get(bo);
138 up_read(&vmf->vma->vm_mm->mmap_sem);
139 (void) ttm_bo_wait_unreserved(bo);
140 ttm_bo_put(bo);
141 }
142
143 return VM_FAULT_RETRY;
144 }
145
146 /*
147 * If we'd want to change locking order to
148 * mmap_sem -> bo::reserve, we'd use a blocking reserve here
149 * instead of retrying the fault...
150 */
151 return VM_FAULT_NOPAGE;
152 }
153
154 /*
155 * Refuse to fault imported pages. This should be handled
156 * (if at all) by redirecting mmap to the exporter.
157 */
158 if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
159 ret = VM_FAULT_SIGBUS;
160 goto out_unlock;
161 }
162
163 if (bdev->driver->fault_reserve_notify) {
164 struct dma_fence *moving = dma_fence_get(bo->moving);
165
166 err = bdev->driver->fault_reserve_notify(bo);
167 switch (err) {
168 case 0:
169 break;
170 case -EBUSY:
171 case -ERESTARTSYS:
172 ret = VM_FAULT_NOPAGE;
173 goto out_unlock;
174 default:
175 ret = VM_FAULT_SIGBUS;
176 goto out_unlock;
177 }
178
179 if (bo->moving != moving) {
180 spin_lock(&bdev->glob->lru_lock);
181 ttm_bo_move_to_lru_tail(bo, NULL);
182 spin_unlock(&bdev->glob->lru_lock);
183 }
184 dma_fence_put(moving);
185 }
186
187 /*
188 * Wait for buffer data in transit, due to a pipelined
189 * move.
190 */
191 ret = ttm_bo_vm_fault_idle(bo, vmf);
192 if (unlikely(ret != 0)) {
193 if (ret == VM_FAULT_RETRY &&
194 !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
195 /* The BO has already been unreserved. */
196 return ret;
197 }
198
199 goto out_unlock;
200 }
201
202 err = ttm_mem_io_lock(man, true);
203 if (unlikely(err != 0)) {
204 ret = VM_FAULT_NOPAGE;
205 goto out_unlock;
206 }
207 err = ttm_mem_io_reserve_vm(bo);
208 if (unlikely(err != 0)) {
209 ret = VM_FAULT_SIGBUS;
210 goto out_io_unlock;
211 }
212
213 page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
214 vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
215 page_last = vma_pages(vma) + vma->vm_pgoff -
216 drm_vma_node_start(&bo->base.vma_node);
217
218 if (unlikely(page_offset >= bo->num_pages)) {
219 ret = VM_FAULT_SIGBUS;
220 goto out_io_unlock;
221 }
222
223 /*
224 * Make a local vma copy to modify the page_prot member
225 * and vm_flags if necessary. The vma parameter is protected
226 * by mmap_sem in write mode.
227 */
228 cvma = *vma;
229 cvma.vm_page_prot = vm_get_page_prot(cvma.vm_flags);
230
231 if (bo->mem.bus.is_iomem) {
232 cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
233 cvma.vm_page_prot);
234 } else {
235 struct ttm_operation_ctx ctx = {
236 .interruptible = false,
237 .no_wait_gpu = false,
238 .flags = TTM_OPT_FLAG_FORCE_ALLOC
239
240 };
241
242 ttm = bo->ttm;
243 cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
244 cvma.vm_page_prot);
245
246 /* Allocate all page at once, most common usage */
247 if (ttm_tt_populate(ttm, &ctx)) {
248 ret = VM_FAULT_OOM;
249 goto out_io_unlock;
250 }
251 }
252
253 /*
254 * Speculatively prefault a number of pages. Only error on
255 * first page.
256 */
257 for (i = 0; i < TTM_BO_VM_NUM_PREFAULT; ++i) {
258 if (bo->mem.bus.is_iomem) {
259 /* Iomem should not be marked encrypted */
260 cvma.vm_page_prot = pgprot_decrypted(cvma.vm_page_prot);
261 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
262 } else {
263 page = ttm->pages[page_offset];
264 if (unlikely(!page && i == 0)) {
265 ret = VM_FAULT_OOM;
266 goto out_io_unlock;
267 } else if (unlikely(!page)) {
268 break;
269 }
270 page->index = drm_vma_node_start(&bo->base.vma_node) +
271 page_offset;
272 pfn = page_to_pfn(page);
273 }
274
275 if (vma->vm_flags & VM_MIXEDMAP)
276 ret = vmf_insert_mixed(&cvma, address,
277 __pfn_to_pfn_t(pfn, PFN_DEV));
278 else
279 ret = vmf_insert_pfn(&cvma, address, pfn);
280
281 /* Never error on prefaulted PTEs */
282 if (unlikely((ret & VM_FAULT_ERROR))) {
283 if (i == 0)
284 goto out_io_unlock;
285 else
286 break;
287 }
288
289 address += PAGE_SIZE;
290 if (unlikely(++page_offset >= page_last))
291 break;
292 }
293 ret = VM_FAULT_NOPAGE;
294 out_io_unlock:
295 ttm_mem_io_unlock(man);
296 out_unlock:
297 dma_resv_unlock(bo->base.resv);
298 return ret;
299 }
300
ttm_bo_vm_open(struct vm_area_struct * vma)301 static void ttm_bo_vm_open(struct vm_area_struct *vma)
302 {
303 struct ttm_buffer_object *bo =
304 (struct ttm_buffer_object *)vma->vm_private_data;
305
306 WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
307
308 ttm_bo_get(bo);
309 }
310
ttm_bo_vm_close(struct vm_area_struct * vma)311 static void ttm_bo_vm_close(struct vm_area_struct *vma)
312 {
313 struct ttm_buffer_object *bo = (struct ttm_buffer_object *)vma->vm_private_data;
314
315 ttm_bo_put(bo);
316 vma->vm_private_data = NULL;
317 }
318
ttm_bo_vm_access_kmap(struct ttm_buffer_object * bo,unsigned long offset,uint8_t * buf,int len,int write)319 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
320 unsigned long offset,
321 uint8_t *buf, int len, int write)
322 {
323 unsigned long page = offset >> PAGE_SHIFT;
324 unsigned long bytes_left = len;
325 int ret;
326
327 /* Copy a page at a time, that way no extra virtual address
328 * mapping is needed
329 */
330 offset -= page << PAGE_SHIFT;
331 do {
332 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
333 struct ttm_bo_kmap_obj map;
334 void *ptr;
335 bool is_iomem;
336
337 ret = ttm_bo_kmap(bo, page, 1, &map);
338 if (ret)
339 return ret;
340
341 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
342 WARN_ON_ONCE(is_iomem);
343 if (write)
344 memcpy(ptr, buf, bytes);
345 else
346 memcpy(buf, ptr, bytes);
347 ttm_bo_kunmap(&map);
348
349 page++;
350 buf += bytes;
351 bytes_left -= bytes;
352 offset = 0;
353 } while (bytes_left);
354
355 return len;
356 }
357
ttm_bo_vm_access(struct vm_area_struct * vma,unsigned long addr,void * buf,int len,int write)358 static int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
359 void *buf, int len, int write)
360 {
361 unsigned long offset = (addr) - vma->vm_start;
362 struct ttm_buffer_object *bo = vma->vm_private_data;
363 int ret;
364
365 if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->num_pages)
366 return -EIO;
367
368 ret = ttm_bo_reserve(bo, true, false, NULL);
369 if (ret)
370 return ret;
371
372 switch (bo->mem.mem_type) {
373 case TTM_PL_SYSTEM:
374 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
375 ret = ttm_tt_swapin(bo->ttm);
376 if (unlikely(ret != 0))
377 return ret;
378 }
379 /* fall through */
380 case TTM_PL_TT:
381 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
382 break;
383 default:
384 if (bo->bdev->driver->access_memory)
385 ret = bo->bdev->driver->access_memory(
386 bo, offset, buf, len, write);
387 else
388 ret = -EIO;
389 }
390
391 ttm_bo_unreserve(bo);
392
393 return ret;
394 }
395
396 static const struct vm_operations_struct ttm_bo_vm_ops = {
397 .fault = ttm_bo_vm_fault,
398 .open = ttm_bo_vm_open,
399 .close = ttm_bo_vm_close,
400 .access = ttm_bo_vm_access
401 };
402
ttm_bo_vm_lookup(struct ttm_bo_device * bdev,unsigned long offset,unsigned long pages)403 static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
404 unsigned long offset,
405 unsigned long pages)
406 {
407 struct drm_vma_offset_node *node;
408 struct ttm_buffer_object *bo = NULL;
409
410 drm_vma_offset_lock_lookup(&bdev->vma_manager);
411
412 node = drm_vma_offset_lookup_locked(&bdev->vma_manager, offset, pages);
413 if (likely(node)) {
414 bo = container_of(node, struct ttm_buffer_object,
415 base.vma_node);
416 bo = ttm_bo_get_unless_zero(bo);
417 }
418
419 drm_vma_offset_unlock_lookup(&bdev->vma_manager);
420
421 if (!bo)
422 pr_err("Could not find buffer object to map\n");
423
424 return bo;
425 }
426
ttm_bo_mmap(struct file * filp,struct vm_area_struct * vma,struct ttm_bo_device * bdev)427 int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
428 struct ttm_bo_device *bdev)
429 {
430 struct ttm_bo_driver *driver;
431 struct ttm_buffer_object *bo;
432 int ret;
433
434 if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET_START))
435 return -EINVAL;
436
437 bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
438 if (unlikely(!bo))
439 return -EINVAL;
440
441 driver = bo->bdev->driver;
442 if (unlikely(!driver->verify_access)) {
443 ret = -EPERM;
444 goto out_unref;
445 }
446 ret = driver->verify_access(bo, filp);
447 if (unlikely(ret != 0))
448 goto out_unref;
449
450 vma->vm_ops = &ttm_bo_vm_ops;
451
452 /*
453 * Note: We're transferring the bo reference to
454 * vma->vm_private_data here.
455 */
456
457 vma->vm_private_data = bo;
458
459 /*
460 * We'd like to use VM_PFNMAP on shared mappings, where
461 * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
462 * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
463 * bad for performance. Until that has been sorted out, use
464 * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
465 */
466 vma->vm_flags |= VM_MIXEDMAP;
467 vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
468 return 0;
469 out_unref:
470 ttm_bo_put(bo);
471 return ret;
472 }
473 EXPORT_SYMBOL(ttm_bo_mmap);
474
ttm_fbdev_mmap(struct vm_area_struct * vma,struct ttm_buffer_object * bo)475 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
476 {
477 if (vma->vm_pgoff != 0)
478 return -EACCES;
479
480 ttm_bo_get(bo);
481
482 vma->vm_ops = &ttm_bo_vm_ops;
483 vma->vm_private_data = bo;
484 vma->vm_flags |= VM_MIXEDMAP;
485 vma->vm_flags |= VM_IO | VM_DONTEXPAND;
486 return 0;
487 }
488 EXPORT_SYMBOL(ttm_fbdev_mmap);
489
