xref: /freebsd/sys/dev/drm2/ttm/ttm_bo_vm.c (revision 7944a87dce9355bea07b7cbd005cf5107fac61d2)
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/param.h>
39 #include <sys/pctrie.h>
40 #include "opt_vm.h"
41 
42 #include <dev/drm2/drmP.h>
43 #include <dev/drm2/ttm/ttm_module.h>
44 #include <dev/drm2/ttm/ttm_bo_driver.h>
45 #include <dev/drm2/ttm/ttm_placement.h>
46 
47 #include <vm/vm.h>
48 #include <vm/vm_page.h>
49 #include <vm/vm_pageout.h>
50 #include <vm/vm_radix.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
ttm_bo_cmp_rb_tree_items(struct ttm_buffer_object * a,struct ttm_buffer_object * b)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 
ttm_bo_vm_lookup_rb(struct ttm_bo_device * bdev,unsigned long page_start,unsigned long num_pages)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
ttm_bo_vm_fault(vm_object_t vm_obj,vm_ooffset_t offset,int prot,vm_page_t * mres)102 ttm_bo_vm_fault(vm_object_t vm_obj, vm_ooffset_t offset,
103     int prot, vm_page_t *mres)
104 {
105 	struct pctrie_iter pages;
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;
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 	if (*mres != NULL) {
117 		(void)vm_page_remove(*mres);
118 	}
119 	vm_page_iter_init(&pages, vm_obj);
120 retry:
121 	VM_OBJECT_WUNLOCK(vm_obj);
122 	m = NULL;
123 
124 reserve:
125 	ret = ttm_bo_reserve(bo, false, false, false, 0);
126 	if (unlikely(ret != 0)) {
127 		if (ret == -EBUSY) {
128 			kern_yield(PRI_USER);
129 			goto reserve;
130 		}
131 	}
132 
133 	if (bdev->driver->fault_reserve_notify) {
134 		ret = bdev->driver->fault_reserve_notify(bo);
135 		switch (ret) {
136 		case 0:
137 			break;
138 		case -EBUSY:
139 		case -ERESTARTSYS:
140 		case -EINTR:
141 			kern_yield(PRI_USER);
142 			goto reserve;
143 		default:
144 			retval = VM_PAGER_ERROR;
145 			goto out_unlock;
146 		}
147 	}
148 
149 	/*
150 	 * Wait for buffer data in transit, due to a pipelined
151 	 * move.
152 	 */
153 
154 	mtx_lock(&bdev->fence_lock);
155 	if (test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)) {
156 		/*
157 		 * Here, the behavior differs between Linux and FreeBSD.
158 		 *
159 		 * On Linux, the wait is interruptible (3rd argument to
160 		 * ttm_bo_wait). There must be some mechanism to resume
161 		 * page fault handling, once the signal is processed.
162 		 *
163 		 * On FreeBSD, the wait is uninteruptible. This is not a
164 		 * problem as we can't end up with an unkillable process
165 		 * here, because the wait will eventually time out.
166 		 *
167 		 * An example of this situation is the Xorg process
168 		 * which uses SIGALRM internally. The signal could
169 		 * interrupt the wait, causing the page fault to fail
170 		 * and the process to receive SIGSEGV.
171 		 */
172 		ret = ttm_bo_wait(bo, false, false, false);
173 		mtx_unlock(&bdev->fence_lock);
174 		if (unlikely(ret != 0)) {
175 			retval = VM_PAGER_ERROR;
176 			goto out_unlock;
177 		}
178 	} else
179 		mtx_unlock(&bdev->fence_lock);
180 
181 	ret = ttm_mem_io_lock(man, true);
182 	if (unlikely(ret != 0)) {
183 		retval = VM_PAGER_ERROR;
184 		goto out_unlock;
185 	}
186 	ret = ttm_mem_io_reserve_vm(bo);
187 	if (unlikely(ret != 0)) {
188 		retval = VM_PAGER_ERROR;
189 		goto out_io_unlock;
190 	}
191 
192 	/*
193 	 * Strictly, we're not allowed to modify vma->vm_page_prot here,
194 	 * since the mmap_sem is only held in read mode. However, we
195 	 * modify only the caching bits of vma->vm_page_prot and
196 	 * consider those bits protected by
197 	 * the bo->mutex, as we should be the only writers.
198 	 * There shouldn't really be any readers of these bits except
199 	 * within vm_insert_mixed()? fork?
200 	 *
201 	 * TODO: Add a list of vmas to the bo, and change the
202 	 * vma->vm_page_prot when the object changes caching policy, with
203 	 * the correct locks held.
204 	 */
205 	if (!bo->mem.bus.is_iomem) {
206 		/* Allocate all page at once, most common usage */
207 		ttm = bo->ttm;
208 		if (ttm->bdev->driver->ttm_tt_populate(ttm)) {
209 			retval = VM_PAGER_ERROR;
210 			goto out_io_unlock;
211 		}
212 	}
213 
214 	if (bo->mem.bus.is_iomem) {
215 		m = PHYS_TO_VM_PAGE(bo->mem.bus.base + bo->mem.bus.offset +
216 		    offset);
217 		KASSERT((m->flags & PG_FICTITIOUS) != 0,
218 		    ("physical address %#jx not fictitious",
219 		    (uintmax_t)(bo->mem.bus.base + bo->mem.bus.offset
220 		    + offset)));
221 		pmap_page_set_memattr(m, ttm_io_prot(bo->mem.placement));
222 	} else {
223 		ttm = bo->ttm;
224 		m = ttm->pages[OFF_TO_IDX(offset)];
225 		if (unlikely(!m)) {
226 			retval = VM_PAGER_ERROR;
227 			goto out_io_unlock;
228 		}
229 		pmap_page_set_memattr(m,
230 		    (bo->mem.placement & TTM_PL_FLAG_CACHED) ?
231 		    VM_MEMATTR_WRITE_BACK : ttm_io_prot(bo->mem.placement));
232 	}
233 
234 	VM_OBJECT_WLOCK(vm_obj);
235 	if (vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL) == 0) {
236 		ttm_mem_io_unlock(man);
237 		ttm_bo_unreserve(bo);
238 		goto retry;
239 	}
240 	pctrie_iter_reset(&pages);
241 	m1 = vm_radix_iter_lookup(&pages, OFF_TO_IDX(offset));
242 	/* XXX This looks like it should just be vm_page_replace? */
243 	if (m1 == NULL) {
244 		if (vm_page_iter_insert(
245 		    m, vm_obj, OFF_TO_IDX(offset), &pages) != 0) {
246 			vm_page_xunbusy(m);
247 			VM_OBJECT_WUNLOCK(vm_obj);
248 			vm_wait(vm_obj);
249 			VM_OBJECT_WLOCK(vm_obj);
250 			ttm_mem_io_unlock(man);
251 			ttm_bo_unreserve(bo);
252 			goto retry;
253 		}
254 	} else {
255 		KASSERT(m == m1,
256 		    ("inconsistent insert bo %p m %p m1 %p offset %jx",
257 		    bo, m, m1, (uintmax_t)offset));
258 	}
259 	vm_page_valid(m);
260 	if (*mres != NULL) {
261 		KASSERT(*mres != m, ("losing %p %p", *mres, m));
262 		vm_page_xunbusy(*mres);
263 		vm_page_free(*mres);
264 	}
265 	*mres = m;
266 
267 out_io_unlock1:
268 	ttm_mem_io_unlock(man);
269 out_unlock1:
270 	ttm_bo_unreserve(bo);
271 	vm_object_pip_wakeup(vm_obj);
272 	return (retval);
273 
274 out_io_unlock:
275 	VM_OBJECT_WLOCK(vm_obj);
276 	goto out_io_unlock1;
277 
278 out_unlock:
279 	VM_OBJECT_WLOCK(vm_obj);
280 	goto out_unlock1;
281 }
282 
283 static int
ttm_bo_vm_ctor(void * handle,vm_ooffset_t size,vm_prot_t prot,vm_ooffset_t foff,struct ucred * cred,u_short * color)284 ttm_bo_vm_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
285     vm_ooffset_t foff, struct ucred *cred, u_short *color)
286 {
287 
288 	/*
289 	 * On Linux, a reference to the buffer object is acquired here.
290 	 * The reason is that this function is not called when the
291 	 * mmap() is initialized, but only when a process forks for
292 	 * instance. Therefore on Linux, the reference on the bo is
293 	 * acquired either in ttm_bo_mmap() or ttm_bo_vm_open(). It's
294 	 * then released in ttm_bo_vm_close().
295 	 *
296 	 * Here, this function is called during mmap() initialization.
297 	 * Thus, the reference acquired in ttm_bo_mmap_single() is
298 	 * sufficient.
299 	 */
300 
301 	*color = 0;
302 	return (0);
303 }
304 
305 static void
ttm_bo_vm_dtor(void * handle)306 ttm_bo_vm_dtor(void *handle)
307 {
308 	struct ttm_buffer_object *bo = handle;
309 
310 	ttm_bo_unref(&bo);
311 }
312 
313 static struct cdev_pager_ops ttm_pager_ops = {
314 	.cdev_pg_fault = ttm_bo_vm_fault,
315 	.cdev_pg_ctor = ttm_bo_vm_ctor,
316 	.cdev_pg_dtor = ttm_bo_vm_dtor
317 };
318 
319 int
ttm_bo_mmap_single(struct ttm_bo_device * bdev,vm_ooffset_t * offset,vm_size_t size,struct vm_object ** obj_res,int nprot)320 ttm_bo_mmap_single(struct ttm_bo_device *bdev, vm_ooffset_t *offset, vm_size_t size,
321     struct vm_object **obj_res, int nprot)
322 {
323 	struct ttm_bo_driver *driver;
324 	struct ttm_buffer_object *bo;
325 	struct vm_object *vm_obj;
326 	int ret;
327 
328 	rw_wlock(&bdev->vm_lock);
329 	bo = ttm_bo_vm_lookup_rb(bdev, OFF_TO_IDX(*offset), OFF_TO_IDX(size));
330 	if (likely(bo != NULL))
331 		refcount_acquire(&bo->kref);
332 	rw_wunlock(&bdev->vm_lock);
333 
334 	if (unlikely(bo == NULL)) {
335 		printf("[TTM] Could not find buffer object to map\n");
336 		return (-EINVAL);
337 	}
338 
339 	driver = bo->bdev->driver;
340 	if (unlikely(!driver->verify_access)) {
341 		ret = -EPERM;
342 		goto out_unref;
343 	}
344 	ret = driver->verify_access(bo);
345 	if (unlikely(ret != 0))
346 		goto out_unref;
347 
348 	vm_obj = cdev_pager_allocate(bo, OBJT_MGTDEVICE, &ttm_pager_ops,
349 	    size, nprot, 0, curthread->td_ucred);
350 	if (vm_obj == NULL) {
351 		ret = -EINVAL;
352 		goto out_unref;
353 	}
354 	/*
355 	 * Note: We're transferring the bo reference to vm_obj->handle here.
356 	 */
357 	*offset = 0;
358 	*obj_res = vm_obj;
359 	return 0;
360 out_unref:
361 	ttm_bo_unref(&bo);
362 	return ret;
363 }
364 
365 void
ttm_bo_release_mmap(struct ttm_buffer_object * bo)366 ttm_bo_release_mmap(struct ttm_buffer_object *bo)
367 {
368 	vm_object_t vm_obj;
369 
370 	vm_obj = cdev_pager_lookup(bo);
371 	if (vm_obj != NULL) {
372 		cdev_mgtdev_pager_free_pages(vm_obj);
373 		vm_object_deallocate(vm_obj);
374 	}
375 }
376 
377 #if 0
378 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
379 {
380 	if (vma->vm_pgoff != 0)
381 		return -EACCES;
382 
383 	vma->vm_ops = &ttm_bo_vm_ops;
384 	vma->vm_private_data = ttm_bo_reference(bo);
385 	vma->vm_flags |= VM_IO | VM_MIXEDMAP | VM_DONTEXPAND;
386 	return 0;
387 }
388 
389 ssize_t ttm_bo_io(struct ttm_bo_device *bdev, struct file *filp,
390 		  const char __user *wbuf, char __user *rbuf, size_t count,
391 		  loff_t *f_pos, bool write)
392 {
393 	struct ttm_buffer_object *bo;
394 	struct ttm_bo_driver *driver;
395 	struct ttm_bo_kmap_obj map;
396 	unsigned long dev_offset = (*f_pos >> PAGE_SHIFT);
397 	unsigned long kmap_offset;
398 	unsigned long kmap_end;
399 	unsigned long kmap_num;
400 	size_t io_size;
401 	unsigned int page_offset;
402 	char *virtual;
403 	int ret;
404 	bool no_wait = false;
405 	bool dummy;
406 
407 	read_lock(&bdev->vm_lock);
408 	bo = ttm_bo_vm_lookup_rb(bdev, dev_offset, 1);
409 	if (likely(bo != NULL))
410 		ttm_bo_reference(bo);
411 	read_unlock(&bdev->vm_lock);
412 
413 	if (unlikely(bo == NULL))
414 		return -EFAULT;
415 
416 	driver = bo->bdev->driver;
417 	if (unlikely(!driver->verify_access)) {
418 		ret = -EPERM;
419 		goto out_unref;
420 	}
421 
422 	ret = driver->verify_access(bo, filp);
423 	if (unlikely(ret != 0))
424 		goto out_unref;
425 
426 	kmap_offset = dev_offset - bo->vm_node->start;
427 	if (unlikely(kmap_offset >= bo->num_pages)) {
428 		ret = -EFBIG;
429 		goto out_unref;
430 	}
431 
432 	page_offset = *f_pos & ~PAGE_MASK;
433 	io_size = bo->num_pages - kmap_offset;
434 	io_size = (io_size << PAGE_SHIFT) - page_offset;
435 	if (count < io_size)
436 		io_size = count;
437 
438 	kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT;
439 	kmap_num = kmap_end - kmap_offset + 1;
440 
441 	ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
442 
443 	switch (ret) {
444 	case 0:
445 		break;
446 	case -EBUSY:
447 		ret = -EAGAIN;
448 		goto out_unref;
449 	default:
450 		goto out_unref;
451 	}
452 
453 	ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map);
454 	if (unlikely(ret != 0)) {
455 		ttm_bo_unreserve(bo);
456 		goto out_unref;
457 	}
458 
459 	virtual = ttm_kmap_obj_virtual(&map, &dummy);
460 	virtual += page_offset;
461 
462 	if (write)
463 		ret = copy_from_user(virtual, wbuf, io_size);
464 	else
465 		ret = copy_to_user(rbuf, virtual, io_size);
466 
467 	ttm_bo_kunmap(&map);
468 	ttm_bo_unreserve(bo);
469 	ttm_bo_unref(&bo);
470 
471 	if (unlikely(ret != 0))
472 		return -EFBIG;
473 
474 	*f_pos += io_size;
475 
476 	return io_size;
477 out_unref:
478 	ttm_bo_unref(&bo);
479 	return ret;
480 }
481 
482 ssize_t ttm_bo_fbdev_io(struct ttm_buffer_object *bo, const char __user *wbuf,
483 			char __user *rbuf, size_t count, loff_t *f_pos,
484 			bool write)
485 {
486 	struct ttm_bo_kmap_obj map;
487 	unsigned long kmap_offset;
488 	unsigned long kmap_end;
489 	unsigned long kmap_num;
490 	size_t io_size;
491 	unsigned int page_offset;
492 	char *virtual;
493 	int ret;
494 	bool no_wait = false;
495 	bool dummy;
496 
497 	kmap_offset = (*f_pos >> PAGE_SHIFT);
498 	if (unlikely(kmap_offset >= bo->num_pages))
499 		return -EFBIG;
500 
501 	page_offset = *f_pos & ~PAGE_MASK;
502 	io_size = bo->num_pages - kmap_offset;
503 	io_size = (io_size << PAGE_SHIFT) - page_offset;
504 	if (count < io_size)
505 		io_size = count;
506 
507 	kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT;
508 	kmap_num = kmap_end - kmap_offset + 1;
509 
510 	ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
511 
512 	switch (ret) {
513 	case 0:
514 		break;
515 	case -EBUSY:
516 		return -EAGAIN;
517 	default:
518 		return ret;
519 	}
520 
521 	ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map);
522 	if (unlikely(ret != 0)) {
523 		ttm_bo_unreserve(bo);
524 		return ret;
525 	}
526 
527 	virtual = ttm_kmap_obj_virtual(&map, &dummy);
528 	virtual += page_offset;
529 
530 	if (write)
531 		ret = copy_from_user(virtual, wbuf, io_size);
532 	else
533 		ret = copy_to_user(rbuf, virtual, io_size);
534 
535 	ttm_bo_kunmap(&map);
536 	ttm_bo_unreserve(bo);
537 	ttm_bo_unref(&bo);
538 
539 	if (unlikely(ret != 0))
540 		return ret;
541 
542 	*f_pos += io_size;
543 
544 	return io_size;
545 }
546 #endif
547