1 /* $NetBSD: ttm_bo_vm.c,v 1.3 2021/12/18 23:45:44 riastradh Exp $ */
2
3 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
4 /**************************************************************************
5 *
6 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
7 * All Rights Reserved.
8 *
9 * Permission is hereby granted, free of charge, to any person obtaining a
10 * copy of this software and associated documentation files (the
11 * "Software"), to deal in the Software without restriction, including
12 * without limitation the rights to use, copy, modify, merge, publish,
13 * distribute, sub license, and/or sell copies of the Software, and to
14 * permit persons to whom the Software is furnished to do so, subject to
15 * the following conditions:
16 *
17 * The above copyright notice and this permission notice (including the
18 * next paragraph) shall be included in all copies or substantial portions
19 * of the Software.
20 *
21 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
22 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
23 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
24 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
25 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
26 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
27 * USE OR OTHER DEALINGS IN THE SOFTWARE.
28 *
29 **************************************************************************/
30 /*
31 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
32 */
33
34 #include <sys/cdefs.h>
35 __KERNEL_RCSID(0, "$NetBSD: ttm_bo_vm.c,v 1.3 2021/12/18 23:45:44 riastradh Exp $");
36
37 #define pr_fmt(fmt) "[TTM] " fmt
38
39 #include <drm/ttm/ttm_module.h>
40 #include <drm/ttm/ttm_bo_driver.h>
41 #include <drm/ttm/ttm_placement.h>
42 #include <drm/drm_vma_manager.h>
43 #include <linux/mm.h>
44 #include <linux/pfn_t.h>
45 #include <linux/rbtree.h>
46 #include <linux/module.h>
47 #include <linux/uaccess.h>
48 #include <linux/mem_encrypt.h>
49
ttm_bo_vm_fault_idle(struct ttm_buffer_object * bo,struct vm_fault * vmf)50 static vm_fault_t ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
51 struct vm_fault *vmf)
52 {
53 vm_fault_t ret = 0;
54 int err = 0;
55
56 if (likely(!bo->moving))
57 goto out_unlock;
58
59 /*
60 * Quick non-stalling check for idle.
61 */
62 if (dma_fence_is_signaled(bo->moving))
63 goto out_clear;
64
65 /*
66 * If possible, avoid waiting for GPU with mmap_sem
67 * held.
68 */
69 if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
70 ret = VM_FAULT_RETRY;
71 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
72 goto out_unlock;
73
74 ttm_bo_get(bo);
75 up_read(&vmf->vma->vm_mm->mmap_sem);
76 (void) dma_fence_wait(bo->moving, true);
77 dma_resv_unlock(bo->base.resv);
78 ttm_bo_put(bo);
79 goto out_unlock;
80 }
81
82 /*
83 * Ordinary wait.
84 */
85 err = dma_fence_wait(bo->moving, true);
86 if (unlikely(err != 0)) {
87 ret = (err != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
88 VM_FAULT_NOPAGE;
89 goto out_unlock;
90 }
91
92 out_clear:
93 dma_fence_put(bo->moving);
94 bo->moving = NULL;
95
96 out_unlock:
97 return ret;
98 }
99
ttm_bo_io_mem_pfn(struct ttm_buffer_object * bo,unsigned long page_offset)100 static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object *bo,
101 unsigned long page_offset)
102 {
103 struct ttm_bo_device *bdev = bo->bdev;
104
105 if (bdev->driver->io_mem_pfn)
106 return bdev->driver->io_mem_pfn(bo, page_offset);
107
108 return ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT)
109 + page_offset;
110 }
111
112 /**
113 * ttm_bo_vm_reserve - Reserve a buffer object in a retryable vm callback
114 * @bo: The buffer object
115 * @vmf: The fault structure handed to the callback
116 *
117 * vm callbacks like fault() and *_mkwrite() allow for the mm_sem to be dropped
118 * during long waits, and after the wait the callback will be restarted. This
119 * is to allow other threads using the same virtual memory space concurrent
120 * access to map(), unmap() completely unrelated buffer objects. TTM buffer
121 * object reservations sometimes wait for GPU and should therefore be
122 * considered long waits. This function reserves the buffer object interruptibly
123 * taking this into account. Starvation is avoided by the vm system not
124 * allowing too many repeated restarts.
125 * This function is intended to be used in customized fault() and _mkwrite()
126 * handlers.
127 *
128 * Return:
129 * 0 on success and the bo was reserved.
130 * VM_FAULT_RETRY if blocking wait.
131 * VM_FAULT_NOPAGE if blocking wait and retrying was not allowed.
132 */
ttm_bo_vm_reserve(struct ttm_buffer_object * bo,struct vm_fault * vmf)133 vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo,
134 struct vm_fault *vmf)
135 {
136 /*
137 * Work around locking order reversal in fault / nopfn
138 * between mmap_sem and bo_reserve: Perform a trylock operation
139 * for reserve, and if it fails, retry the fault after waiting
140 * for the buffer to become unreserved.
141 */
142 if (unlikely(!dma_resv_trylock(bo->base.resv))) {
143 if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
144 if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
145 ttm_bo_get(bo);
146 up_read(&vmf->vma->vm_mm->mmap_sem);
147 if (!dma_resv_lock_interruptible(bo->base.resv,
148 NULL))
149 dma_resv_unlock(bo->base.resv);
150 ttm_bo_put(bo);
151 }
152
153 return VM_FAULT_RETRY;
154 }
155
156 if (dma_resv_lock_interruptible(bo->base.resv, NULL))
157 return VM_FAULT_NOPAGE;
158 }
159
160 return 0;
161 }
162 EXPORT_SYMBOL(ttm_bo_vm_reserve);
163
164 /**
165 * ttm_bo_vm_fault_reserved - TTM fault helper
166 * @vmf: The struct vm_fault given as argument to the fault callback
167 * @prot: The page protection to be used for this memory area.
168 * @num_prefault: Maximum number of prefault pages. The caller may want to
169 * specify this based on madvice settings and the size of the GPU object
170 * backed by the memory.
171 *
172 * This function inserts one or more page table entries pointing to the
173 * memory backing the buffer object, and then returns a return code
174 * instructing the caller to retry the page access.
175 *
176 * Return:
177 * VM_FAULT_NOPAGE on success or pending signal
178 * VM_FAULT_SIGBUS on unspecified error
179 * VM_FAULT_OOM on out-of-memory
180 * VM_FAULT_RETRY if retryable wait
181 */
ttm_bo_vm_fault_reserved(struct vm_fault * vmf,pgprot_t prot,pgoff_t num_prefault)182 vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
183 pgprot_t prot,
184 pgoff_t num_prefault)
185 {
186 struct vm_area_struct *vma = vmf->vma;
187 struct ttm_buffer_object *bo = vma->vm_private_data;
188 struct ttm_bo_device *bdev = bo->bdev;
189 unsigned long page_offset;
190 unsigned long page_last;
191 unsigned long pfn;
192 struct ttm_tt *ttm = NULL;
193 struct page *page;
194 int err;
195 pgoff_t i;
196 vm_fault_t ret = VM_FAULT_NOPAGE;
197 unsigned long address = vmf->address;
198 struct ttm_mem_type_manager *man =
199 &bdev->man[bo->mem.mem_type];
200
201 /*
202 * Refuse to fault imported pages. This should be handled
203 * (if at all) by redirecting mmap to the exporter.
204 */
205 if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG))
206 return VM_FAULT_SIGBUS;
207
208 if (bdev->driver->fault_reserve_notify) {
209 struct dma_fence *moving = dma_fence_get(bo->moving);
210
211 err = bdev->driver->fault_reserve_notify(bo);
212 switch (err) {
213 case 0:
214 break;
215 case -EBUSY:
216 case -ERESTARTSYS:
217 return VM_FAULT_NOPAGE;
218 default:
219 return VM_FAULT_SIGBUS;
220 }
221
222 if (bo->moving != moving) {
223 spin_lock(&ttm_bo_glob.lru_lock);
224 ttm_bo_move_to_lru_tail(bo, NULL);
225 spin_unlock(&ttm_bo_glob.lru_lock);
226 }
227 dma_fence_put(moving);
228 }
229
230 /*
231 * Wait for buffer data in transit, due to a pipelined
232 * move.
233 */
234 ret = ttm_bo_vm_fault_idle(bo, vmf);
235 if (unlikely(ret != 0))
236 return ret;
237
238 err = ttm_mem_io_lock(man, true);
239 if (unlikely(err != 0))
240 return VM_FAULT_NOPAGE;
241 err = ttm_mem_io_reserve_vm(bo);
242 if (unlikely(err != 0)) {
243 ret = VM_FAULT_SIGBUS;
244 goto out_io_unlock;
245 }
246
247 page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
248 vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
249 page_last = vma_pages(vma) + vma->vm_pgoff -
250 drm_vma_node_start(&bo->base.vma_node);
251
252 if (unlikely(page_offset >= bo->num_pages)) {
253 ret = VM_FAULT_SIGBUS;
254 goto out_io_unlock;
255 }
256
257 prot = ttm_io_prot(bo->mem.placement, prot);
258 if (!bo->mem.bus.is_iomem) {
259 struct ttm_operation_ctx ctx = {
260 .interruptible = false,
261 .no_wait_gpu = false,
262 .flags = TTM_OPT_FLAG_FORCE_ALLOC
263
264 };
265
266 ttm = bo->ttm;
267 if (ttm_tt_populate(bo->ttm, &ctx)) {
268 ret = VM_FAULT_OOM;
269 goto out_io_unlock;
270 }
271 } else {
272 /* Iomem should not be marked encrypted */
273 prot = pgprot_decrypted(prot);
274 }
275
276 /*
277 * Speculatively prefault a number of pages. Only error on
278 * first page.
279 */
280 for (i = 0; i < num_prefault; ++i) {
281 if (bo->mem.bus.is_iomem) {
282 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
283 } else {
284 page = ttm->pages[page_offset];
285 if (unlikely(!page && i == 0)) {
286 ret = VM_FAULT_OOM;
287 goto out_io_unlock;
288 } else if (unlikely(!page)) {
289 break;
290 }
291 page->index = drm_vma_node_start(&bo->base.vma_node) +
292 page_offset;
293 pfn = page_to_pfn(page);
294 }
295
296 /*
297 * Note that the value of @prot at this point may differ from
298 * the value of @vma->vm_page_prot in the caching- and
299 * encryption bits. This is because the exact location of the
300 * data may not be known at mmap() time and may also change
301 * at arbitrary times while the data is mmap'ed.
302 * See vmf_insert_mixed_prot() for a discussion.
303 */
304 if (vma->vm_flags & VM_MIXEDMAP)
305 ret = vmf_insert_mixed_prot(vma, address,
306 __pfn_to_pfn_t(pfn, PFN_DEV),
307 prot);
308 else
309 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
310
311 /* Never error on prefaulted PTEs */
312 if (unlikely((ret & VM_FAULT_ERROR))) {
313 if (i == 0)
314 goto out_io_unlock;
315 else
316 break;
317 }
318
319 address += PAGE_SIZE;
320 if (unlikely(++page_offset >= page_last))
321 break;
322 }
323 ret = VM_FAULT_NOPAGE;
324 out_io_unlock:
325 ttm_mem_io_unlock(man);
326 return ret;
327 }
328 EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
329
ttm_bo_vm_fault(struct vm_fault * vmf)330 vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
331 {
332 struct vm_area_struct *vma = vmf->vma;
333 pgprot_t prot;
334 struct ttm_buffer_object *bo = vma->vm_private_data;
335 vm_fault_t ret;
336
337 ret = ttm_bo_vm_reserve(bo, vmf);
338 if (ret)
339 return ret;
340
341 prot = vma->vm_page_prot;
342 ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT);
343 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
344 return ret;
345
346 dma_resv_unlock(bo->base.resv);
347
348 return ret;
349 }
350 EXPORT_SYMBOL(ttm_bo_vm_fault);
351
ttm_bo_vm_open(struct vm_area_struct * vma)352 void ttm_bo_vm_open(struct vm_area_struct *vma)
353 {
354 struct ttm_buffer_object *bo = vma->vm_private_data;
355
356 WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
357
358 ttm_bo_get(bo);
359 }
360 EXPORT_SYMBOL(ttm_bo_vm_open);
361
ttm_bo_vm_close(struct vm_area_struct * vma)362 void ttm_bo_vm_close(struct vm_area_struct *vma)
363 {
364 struct ttm_buffer_object *bo = vma->vm_private_data;
365
366 ttm_bo_put(bo);
367 vma->vm_private_data = NULL;
368 }
369 EXPORT_SYMBOL(ttm_bo_vm_close);
370
ttm_bo_vm_access_kmap(struct ttm_buffer_object * bo,unsigned long offset,uint8_t * buf,int len,int write)371 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
372 unsigned long offset,
373 uint8_t *buf, int len, int write)
374 {
375 unsigned long page = offset >> PAGE_SHIFT;
376 unsigned long bytes_left = len;
377 int ret;
378
379 /* Copy a page at a time, that way no extra virtual address
380 * mapping is needed
381 */
382 offset -= page << PAGE_SHIFT;
383 do {
384 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
385 struct ttm_bo_kmap_obj map;
386 void *ptr;
387 bool is_iomem;
388
389 ret = ttm_bo_kmap(bo, page, 1, &map);
390 if (ret)
391 return ret;
392
393 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
394 WARN_ON_ONCE(is_iomem);
395 if (write)
396 memcpy(ptr, buf, bytes);
397 else
398 memcpy(buf, ptr, bytes);
399 ttm_bo_kunmap(&map);
400
401 page++;
402 buf += bytes;
403 bytes_left -= bytes;
404 offset = 0;
405 } while (bytes_left);
406
407 return len;
408 }
409
ttm_bo_vm_access(struct vm_area_struct * vma,unsigned long addr,void * buf,int len,int write)410 int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
411 void *buf, int len, int write)
412 {
413 unsigned long offset = (addr) - vma->vm_start;
414 struct ttm_buffer_object *bo = vma->vm_private_data;
415 int ret;
416
417 if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->num_pages)
418 return -EIO;
419
420 ret = ttm_bo_reserve(bo, true, false, NULL);
421 if (ret)
422 return ret;
423
424 switch (bo->mem.mem_type) {
425 case TTM_PL_SYSTEM:
426 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
427 ret = ttm_tt_swapin(bo->ttm);
428 if (unlikely(ret != 0))
429 return ret;
430 }
431 /* fall through */
432 case TTM_PL_TT:
433 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
434 break;
435 default:
436 if (bo->bdev->driver->access_memory)
437 ret = bo->bdev->driver->access_memory(
438 bo, offset, buf, len, write);
439 else
440 ret = -EIO;
441 }
442
443 ttm_bo_unreserve(bo);
444
445 return ret;
446 }
447 EXPORT_SYMBOL(ttm_bo_vm_access);
448
449 static const struct vm_operations_struct ttm_bo_vm_ops = {
450 .fault = ttm_bo_vm_fault,
451 .open = ttm_bo_vm_open,
452 .close = ttm_bo_vm_close,
453 .access = ttm_bo_vm_access
454 };
455
ttm_bo_vm_lookup(struct ttm_bo_device * bdev,unsigned long offset,unsigned long pages)456 static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
457 unsigned long offset,
458 unsigned long pages)
459 {
460 struct drm_vma_offset_node *node;
461 struct ttm_buffer_object *bo = NULL;
462
463 drm_vma_offset_lock_lookup(bdev->vma_manager);
464
465 node = drm_vma_offset_lookup_locked(bdev->vma_manager, offset, pages);
466 if (likely(node)) {
467 bo = container_of(node, struct ttm_buffer_object,
468 base.vma_node);
469 bo = ttm_bo_get_unless_zero(bo);
470 }
471
472 drm_vma_offset_unlock_lookup(bdev->vma_manager);
473
474 if (!bo)
475 pr_err("Could not find buffer object to map\n");
476
477 return bo;
478 }
479
ttm_bo_mmap_vma_setup(struct ttm_buffer_object * bo,struct vm_area_struct * vma)480 static void ttm_bo_mmap_vma_setup(struct ttm_buffer_object *bo, struct vm_area_struct *vma)
481 {
482 vma->vm_ops = &ttm_bo_vm_ops;
483
484 /*
485 * Note: We're transferring the bo reference to
486 * vma->vm_private_data here.
487 */
488
489 vma->vm_private_data = bo;
490
491 /*
492 * We'd like to use VM_PFNMAP on shared mappings, where
493 * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
494 * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
495 * bad for performance. Until that has been sorted out, use
496 * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
497 */
498 vma->vm_flags |= VM_MIXEDMAP;
499 vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
500 }
501
ttm_bo_mmap(struct file * filp,struct vm_area_struct * vma,struct ttm_bo_device * bdev)502 int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
503 struct ttm_bo_device *bdev)
504 {
505 struct ttm_bo_driver *driver;
506 struct ttm_buffer_object *bo;
507 int ret;
508
509 if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET_START))
510 return -EINVAL;
511
512 bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
513 if (unlikely(!bo))
514 return -EINVAL;
515
516 driver = bo->bdev->driver;
517 if (unlikely(!driver->verify_access)) {
518 ret = -EPERM;
519 goto out_unref;
520 }
521 ret = driver->verify_access(bo, filp);
522 if (unlikely(ret != 0))
523 goto out_unref;
524
525 ttm_bo_mmap_vma_setup(bo, vma);
526 return 0;
527 out_unref:
528 ttm_bo_put(bo);
529 return ret;
530 }
531 EXPORT_SYMBOL(ttm_bo_mmap);
532
ttm_bo_mmap_obj(struct vm_area_struct * vma,struct ttm_buffer_object * bo)533 int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
534 {
535 ttm_bo_get(bo);
536 ttm_bo_mmap_vma_setup(bo, vma);
537 return 0;
538 }
539 EXPORT_SYMBOL(ttm_bo_mmap_obj);
540