xref: /freebsd/sys/dev/drm2/ttm/ttm_bo_vm.c (revision c14aafed63f79923e6a3cc30f0069ece46d36fd7)
1 /**************************************************************************
2  *
3  * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
4  * All Rights Reserved.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the
8  * "Software"), to deal in the Software without restriction, including
9  * without limitation the rights to use, copy, modify, merge, publish,
10  * distribute, sub license, and/or sell copies of the Software, and to
11  * permit persons to whom the Software is furnished to do so, subject to
12  * the following conditions:
13  *
14  * The above copyright notice and this permission notice (including the
15  * next paragraph) shall be included in all copies or substantial portions
16  * of the Software.
17  *
18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24  * USE OR OTHER DEALINGS IN THE SOFTWARE.
25  *
26  **************************************************************************/
27 /*
28  * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
29  */
30 /*
31  * Copyright (c) 2013 The FreeBSD Foundation
32  * All rights reserved.
33  *
34  * Portions of this software were developed by Konstantin Belousov
35  * <kib@FreeBSD.org> under sponsorship from the FreeBSD Foundation.
36  */
37 
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
40 
41 #include "opt_vm.h"
42 
43 #include <dev/drm2/drmP.h>
44 #include <dev/drm2/ttm/ttm_module.h>
45 #include <dev/drm2/ttm/ttm_bo_driver.h>
46 #include <dev/drm2/ttm/ttm_placement.h>
47 
48 #include <vm/vm.h>
49 #include <vm/vm_page.h>
50 #include <vm/vm_pageout.h>
51 
52 #define TTM_BO_VM_NUM_PREFAULT 16
53 
54 RB_GENERATE(ttm_bo_device_buffer_objects, ttm_buffer_object, vm_rb,
55     ttm_bo_cmp_rb_tree_items);
56 
57 int
58 ttm_bo_cmp_rb_tree_items(struct ttm_buffer_object *a,
59     struct ttm_buffer_object *b)
60 {
61 
62 	if (a->vm_node->start < b->vm_node->start) {
63 		return (-1);
64 	} else if (a->vm_node->start > b->vm_node->start) {
65 		return (1);
66 	} else {
67 		return (0);
68 	}
69 }
70 
71 static struct ttm_buffer_object *ttm_bo_vm_lookup_rb(struct ttm_bo_device *bdev,
72 						     unsigned long page_start,
73 						     unsigned long num_pages)
74 {
75 	unsigned long cur_offset;
76 	struct ttm_buffer_object *bo;
77 	struct ttm_buffer_object *best_bo = NULL;
78 
79 	bo = RB_ROOT(&bdev->addr_space_rb);
80 	while (bo != NULL) {
81 		cur_offset = bo->vm_node->start;
82 		if (page_start >= cur_offset) {
83 			best_bo = bo;
84 			if (page_start == cur_offset)
85 				break;
86 			bo = RB_RIGHT(bo, vm_rb);
87 		} else
88 			bo = RB_LEFT(bo, vm_rb);
89 	}
90 
91 	if (unlikely(best_bo == NULL))
92 		return NULL;
93 
94 	if (unlikely((best_bo->vm_node->start + best_bo->num_pages) <
95 		     (page_start + num_pages)))
96 		return NULL;
97 
98 	return best_bo;
99 }
100 
101 static int
102 ttm_bo_vm_fault(vm_object_t vm_obj, vm_ooffset_t offset,
103     int prot, vm_page_t *mres)
104 {
105 
106 	struct ttm_buffer_object *bo = vm_obj->handle;
107 	struct ttm_bo_device *bdev = bo->bdev;
108 	struct ttm_tt *ttm = NULL;
109 	vm_page_t m, m1, oldm;
110 	int ret;
111 	int retval = VM_PAGER_OK;
112 	struct ttm_mem_type_manager *man =
113 		&bdev->man[bo->mem.mem_type];
114 
115 	vm_object_pip_add(vm_obj, 1);
116 	oldm = *mres;
117 	if (oldm != NULL) {
118 		vm_page_lock(oldm);
119 		vm_page_remove(oldm);
120 		vm_page_unlock(oldm);
121 		*mres = NULL;
122 	} else
123 		oldm = NULL;
124 retry:
125 	VM_OBJECT_WUNLOCK(vm_obj);
126 	m = NULL;
127 
128 reserve:
129 	ret = ttm_bo_reserve(bo, false, false, false, 0);
130 	if (unlikely(ret != 0)) {
131 		if (ret == -EBUSY) {
132 			kern_yield(0);
133 			goto reserve;
134 		}
135 	}
136 
137 	if (bdev->driver->fault_reserve_notify) {
138 		ret = bdev->driver->fault_reserve_notify(bo);
139 		switch (ret) {
140 		case 0:
141 			break;
142 		case -EBUSY:
143 		case -ERESTARTSYS:
144 		case -EINTR:
145 			kern_yield(0);
146 			goto reserve;
147 		default:
148 			retval = VM_PAGER_ERROR;
149 			goto out_unlock;
150 		}
151 	}
152 
153 	/*
154 	 * Wait for buffer data in transit, due to a pipelined
155 	 * move.
156 	 */
157 
158 	mtx_lock(&bdev->fence_lock);
159 	if (test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)) {
160 		/*
161 		 * Here, the behavior differs between Linux and FreeBSD.
162 		 *
163 		 * On Linux, the wait is interruptible (3rd argument to
164 		 * ttm_bo_wait). There must be some mechanism to resume
165 		 * page fault handling, once the signal is processed.
166 		 *
167 		 * On FreeBSD, the wait is uninteruptible. This is not a
168 		 * problem as we can't end up with an unkillable process
169 		 * here, because the wait will eventually time out.
170 		 *
171 		 * An example of this situation is the Xorg process
172 		 * which uses SIGALRM internally. The signal could
173 		 * interrupt the wait, causing the page fault to fail
174 		 * and the process to receive SIGSEGV.
175 		 */
176 		ret = ttm_bo_wait(bo, false, false, false);
177 		mtx_unlock(&bdev->fence_lock);
178 		if (unlikely(ret != 0)) {
179 			retval = VM_PAGER_ERROR;
180 			goto out_unlock;
181 		}
182 	} else
183 		mtx_unlock(&bdev->fence_lock);
184 
185 	ret = ttm_mem_io_lock(man, true);
186 	if (unlikely(ret != 0)) {
187 		retval = VM_PAGER_ERROR;
188 		goto out_unlock;
189 	}
190 	ret = ttm_mem_io_reserve_vm(bo);
191 	if (unlikely(ret != 0)) {
192 		retval = VM_PAGER_ERROR;
193 		goto out_io_unlock;
194 	}
195 
196 	/*
197 	 * Strictly, we're not allowed to modify vma->vm_page_prot here,
198 	 * since the mmap_sem is only held in read mode. However, we
199 	 * modify only the caching bits of vma->vm_page_prot and
200 	 * consider those bits protected by
201 	 * the bo->mutex, as we should be the only writers.
202 	 * There shouldn't really be any readers of these bits except
203 	 * within vm_insert_mixed()? fork?
204 	 *
205 	 * TODO: Add a list of vmas to the bo, and change the
206 	 * vma->vm_page_prot when the object changes caching policy, with
207 	 * the correct locks held.
208 	 */
209 	if (!bo->mem.bus.is_iomem) {
210 		/* Allocate all page at once, most common usage */
211 		ttm = bo->ttm;
212 		if (ttm->bdev->driver->ttm_tt_populate(ttm)) {
213 			retval = VM_PAGER_ERROR;
214 			goto out_io_unlock;
215 		}
216 	}
217 
218 	if (bo->mem.bus.is_iomem) {
219 		m = PHYS_TO_VM_PAGE(bo->mem.bus.base + bo->mem.bus.offset +
220 		    offset);
221 		KASSERT((m->flags & PG_FICTITIOUS) != 0,
222 		    ("physical address %#jx not fictitious",
223 		    (uintmax_t)(bo->mem.bus.base + bo->mem.bus.offset
224 		    + offset)));
225 		pmap_page_set_memattr(m, ttm_io_prot(bo->mem.placement));
226 	} else {
227 		ttm = bo->ttm;
228 		m = ttm->pages[OFF_TO_IDX(offset)];
229 		if (unlikely(!m)) {
230 			retval = VM_PAGER_ERROR;
231 			goto out_io_unlock;
232 		}
233 		pmap_page_set_memattr(m,
234 		    (bo->mem.placement & TTM_PL_FLAG_CACHED) ?
235 		    VM_MEMATTR_WRITE_BACK : ttm_io_prot(bo->mem.placement));
236 	}
237 
238 	VM_OBJECT_WLOCK(vm_obj);
239 	if (vm_page_busied(m)) {
240 		vm_page_lock(m);
241 		VM_OBJECT_WUNLOCK(vm_obj);
242 		vm_page_busy_sleep(m, "ttmpbs");
243 		VM_OBJECT_WLOCK(vm_obj);
244 		ttm_mem_io_unlock(man);
245 		ttm_bo_unreserve(bo);
246 		goto retry;
247 	}
248 	m1 = vm_page_lookup(vm_obj, OFF_TO_IDX(offset));
249 	if (m1 == NULL) {
250 		if (vm_page_insert(m, vm_obj, OFF_TO_IDX(offset))) {
251 			VM_OBJECT_WUNLOCK(vm_obj);
252 			VM_WAIT;
253 			VM_OBJECT_WLOCK(vm_obj);
254 			ttm_mem_io_unlock(man);
255 			ttm_bo_unreserve(bo);
256 			goto retry;
257 		}
258 	} else {
259 		KASSERT(m == m1,
260 		    ("inconsistent insert bo %p m %p m1 %p offset %jx",
261 		    bo, m, m1, (uintmax_t)offset));
262 	}
263 	m->valid = VM_PAGE_BITS_ALL;
264 	*mres = m;
265 	vm_page_xbusy(m);
266 
267 	if (oldm != NULL) {
268 		vm_page_lock(oldm);
269 		vm_page_free(oldm);
270 		vm_page_unlock(oldm);
271 	}
272 
273 out_io_unlock1:
274 	ttm_mem_io_unlock(man);
275 out_unlock1:
276 	ttm_bo_unreserve(bo);
277 	vm_object_pip_wakeup(vm_obj);
278 	return (retval);
279 
280 out_io_unlock:
281 	VM_OBJECT_WLOCK(vm_obj);
282 	goto out_io_unlock1;
283 
284 out_unlock:
285 	VM_OBJECT_WLOCK(vm_obj);
286 	goto out_unlock1;
287 }
288 
289 static int
290 ttm_bo_vm_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
291     vm_ooffset_t foff, struct ucred *cred, u_short *color)
292 {
293 
294 	/*
295 	 * On Linux, a reference to the buffer object is acquired here.
296 	 * The reason is that this function is not called when the
297 	 * mmap() is initialized, but only when a process forks for
298 	 * instance. Therefore on Linux, the reference on the bo is
299 	 * acquired either in ttm_bo_mmap() or ttm_bo_vm_open(). It's
300 	 * then released in ttm_bo_vm_close().
301 	 *
302 	 * Here, this function is called during mmap() intialization.
303 	 * Thus, the reference acquired in ttm_bo_mmap_single() is
304 	 * sufficient.
305 	 */
306 
307 	*color = 0;
308 	return (0);
309 }
310 
311 static void
312 ttm_bo_vm_dtor(void *handle)
313 {
314 	struct ttm_buffer_object *bo = handle;
315 
316 	ttm_bo_unref(&bo);
317 }
318 
319 static struct cdev_pager_ops ttm_pager_ops = {
320 	.cdev_pg_fault = ttm_bo_vm_fault,
321 	.cdev_pg_ctor = ttm_bo_vm_ctor,
322 	.cdev_pg_dtor = ttm_bo_vm_dtor
323 };
324 
325 int
326 ttm_bo_mmap_single(struct ttm_bo_device *bdev, vm_ooffset_t *offset, vm_size_t size,
327     struct vm_object **obj_res, int nprot)
328 {
329 	struct ttm_bo_driver *driver;
330 	struct ttm_buffer_object *bo;
331 	struct vm_object *vm_obj;
332 	int ret;
333 
334 	rw_wlock(&bdev->vm_lock);
335 	bo = ttm_bo_vm_lookup_rb(bdev, OFF_TO_IDX(*offset), OFF_TO_IDX(size));
336 	if (likely(bo != NULL))
337 		refcount_acquire(&bo->kref);
338 	rw_wunlock(&bdev->vm_lock);
339 
340 	if (unlikely(bo == NULL)) {
341 		printf("[TTM] Could not find buffer object to map\n");
342 		return (-EINVAL);
343 	}
344 
345 	driver = bo->bdev->driver;
346 	if (unlikely(!driver->verify_access)) {
347 		ret = -EPERM;
348 		goto out_unref;
349 	}
350 	ret = driver->verify_access(bo);
351 	if (unlikely(ret != 0))
352 		goto out_unref;
353 
354 	vm_obj = cdev_pager_allocate(bo, OBJT_MGTDEVICE, &ttm_pager_ops,
355 	    size, nprot, 0, curthread->td_ucred);
356 	if (vm_obj == NULL) {
357 		ret = -EINVAL;
358 		goto out_unref;
359 	}
360 	/*
361 	 * Note: We're transferring the bo reference to vm_obj->handle here.
362 	 */
363 	*offset = 0;
364 	*obj_res = vm_obj;
365 	return 0;
366 out_unref:
367 	ttm_bo_unref(&bo);
368 	return ret;
369 }
370 
371 void
372 ttm_bo_release_mmap(struct ttm_buffer_object *bo)
373 {
374 	vm_object_t vm_obj;
375 	vm_page_t m;
376 	int i;
377 
378 	vm_obj = cdev_pager_lookup(bo);
379 	if (vm_obj == NULL)
380 		return;
381 
382 	VM_OBJECT_WLOCK(vm_obj);
383 retry:
384 	for (i = 0; i < bo->num_pages; i++) {
385 		m = vm_page_lookup(vm_obj, i);
386 		if (m == NULL)
387 			continue;
388 		if (vm_page_sleep_if_busy(m, "ttm_unm"))
389 			goto retry;
390 		cdev_pager_free_page(vm_obj, m);
391 	}
392 	VM_OBJECT_WUNLOCK(vm_obj);
393 
394 	vm_object_deallocate(vm_obj);
395 }
396 
397 #if 0
398 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
399 {
400 	if (vma->vm_pgoff != 0)
401 		return -EACCES;
402 
403 	vma->vm_ops = &ttm_bo_vm_ops;
404 	vma->vm_private_data = ttm_bo_reference(bo);
405 	vma->vm_flags |= VM_IO | VM_MIXEDMAP | VM_DONTEXPAND;
406 	return 0;
407 }
408 
409 ssize_t ttm_bo_io(struct ttm_bo_device *bdev, struct file *filp,
410 		  const char __user *wbuf, char __user *rbuf, size_t count,
411 		  loff_t *f_pos, bool write)
412 {
413 	struct ttm_buffer_object *bo;
414 	struct ttm_bo_driver *driver;
415 	struct ttm_bo_kmap_obj map;
416 	unsigned long dev_offset = (*f_pos >> PAGE_SHIFT);
417 	unsigned long kmap_offset;
418 	unsigned long kmap_end;
419 	unsigned long kmap_num;
420 	size_t io_size;
421 	unsigned int page_offset;
422 	char *virtual;
423 	int ret;
424 	bool no_wait = false;
425 	bool dummy;
426 
427 	read_lock(&bdev->vm_lock);
428 	bo = ttm_bo_vm_lookup_rb(bdev, dev_offset, 1);
429 	if (likely(bo != NULL))
430 		ttm_bo_reference(bo);
431 	read_unlock(&bdev->vm_lock);
432 
433 	if (unlikely(bo == NULL))
434 		return -EFAULT;
435 
436 	driver = bo->bdev->driver;
437 	if (unlikely(!driver->verify_access)) {
438 		ret = -EPERM;
439 		goto out_unref;
440 	}
441 
442 	ret = driver->verify_access(bo, filp);
443 	if (unlikely(ret != 0))
444 		goto out_unref;
445 
446 	kmap_offset = dev_offset - bo->vm_node->start;
447 	if (unlikely(kmap_offset >= bo->num_pages)) {
448 		ret = -EFBIG;
449 		goto out_unref;
450 	}
451 
452 	page_offset = *f_pos & ~PAGE_MASK;
453 	io_size = bo->num_pages - kmap_offset;
454 	io_size = (io_size << PAGE_SHIFT) - page_offset;
455 	if (count < io_size)
456 		io_size = count;
457 
458 	kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT;
459 	kmap_num = kmap_end - kmap_offset + 1;
460 
461 	ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
462 
463 	switch (ret) {
464 	case 0:
465 		break;
466 	case -EBUSY:
467 		ret = -EAGAIN;
468 		goto out_unref;
469 	default:
470 		goto out_unref;
471 	}
472 
473 	ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map);
474 	if (unlikely(ret != 0)) {
475 		ttm_bo_unreserve(bo);
476 		goto out_unref;
477 	}
478 
479 	virtual = ttm_kmap_obj_virtual(&map, &dummy);
480 	virtual += page_offset;
481 
482 	if (write)
483 		ret = copy_from_user(virtual, wbuf, io_size);
484 	else
485 		ret = copy_to_user(rbuf, virtual, io_size);
486 
487 	ttm_bo_kunmap(&map);
488 	ttm_bo_unreserve(bo);
489 	ttm_bo_unref(&bo);
490 
491 	if (unlikely(ret != 0))
492 		return -EFBIG;
493 
494 	*f_pos += io_size;
495 
496 	return io_size;
497 out_unref:
498 	ttm_bo_unref(&bo);
499 	return ret;
500 }
501 
502 ssize_t ttm_bo_fbdev_io(struct ttm_buffer_object *bo, const char __user *wbuf,
503 			char __user *rbuf, size_t count, loff_t *f_pos,
504 			bool write)
505 {
506 	struct ttm_bo_kmap_obj map;
507 	unsigned long kmap_offset;
508 	unsigned long kmap_end;
509 	unsigned long kmap_num;
510 	size_t io_size;
511 	unsigned int page_offset;
512 	char *virtual;
513 	int ret;
514 	bool no_wait = false;
515 	bool dummy;
516 
517 	kmap_offset = (*f_pos >> PAGE_SHIFT);
518 	if (unlikely(kmap_offset >= bo->num_pages))
519 		return -EFBIG;
520 
521 	page_offset = *f_pos & ~PAGE_MASK;
522 	io_size = bo->num_pages - kmap_offset;
523 	io_size = (io_size << PAGE_SHIFT) - page_offset;
524 	if (count < io_size)
525 		io_size = count;
526 
527 	kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT;
528 	kmap_num = kmap_end - kmap_offset + 1;
529 
530 	ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
531 
532 	switch (ret) {
533 	case 0:
534 		break;
535 	case -EBUSY:
536 		return -EAGAIN;
537 	default:
538 		return ret;
539 	}
540 
541 	ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map);
542 	if (unlikely(ret != 0)) {
543 		ttm_bo_unreserve(bo);
544 		return ret;
545 	}
546 
547 	virtual = ttm_kmap_obj_virtual(&map, &dummy);
548 	virtual += page_offset;
549 
550 	if (write)
551 		ret = copy_from_user(virtual, wbuf, io_size);
552 	else
553 		ret = copy_to_user(rbuf, virtual, io_size);
554 
555 	ttm_bo_kunmap(&map);
556 	ttm_bo_unreserve(bo);
557 	ttm_bo_unref(&bo);
558 
559 	if (unlikely(ret != 0))
560 		return ret;
561 
562 	*f_pos += io_size;
563 
564 	return io_size;
565 }
566 #endif
567