xref: /freebsd/sys/dev/drm2/ttm/ttm_bo_vm.c (revision 38e3125d6df98919983ed800471673ff0e5c2337)
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 #include "opt_vm.h"
40 
41 #include <dev/drm2/drmP.h>
42 #include <dev/drm2/ttm/ttm_module.h>
43 #include <dev/drm2/ttm/ttm_bo_driver.h>
44 #include <dev/drm2/ttm/ttm_placement.h>
45 
46 #include <vm/vm.h>
47 #include <vm/vm_page.h>
48 #include <vm/vm_pageout.h>
49 
50 #define TTM_BO_VM_NUM_PREFAULT 16
51 
52 RB_GENERATE(ttm_bo_device_buffer_objects, ttm_buffer_object, vm_rb,
53     ttm_bo_cmp_rb_tree_items);
54 
55 int
ttm_bo_cmp_rb_tree_items(struct ttm_buffer_object * a,struct ttm_buffer_object * b)56 ttm_bo_cmp_rb_tree_items(struct ttm_buffer_object *a,
57     struct ttm_buffer_object *b)
58 {
59 
60 	if (a->vm_node->start < b->vm_node->start) {
61 		return (-1);
62 	} else if (a->vm_node->start > b->vm_node->start) {
63 		return (1);
64 	} else {
65 		return (0);
66 	}
67 }
68 
ttm_bo_vm_lookup_rb(struct ttm_bo_device * bdev,unsigned long page_start,unsigned long num_pages)69 static struct ttm_buffer_object *ttm_bo_vm_lookup_rb(struct ttm_bo_device *bdev,
70 						     unsigned long page_start,
71 						     unsigned long num_pages)
72 {
73 	unsigned long cur_offset;
74 	struct ttm_buffer_object *bo;
75 	struct ttm_buffer_object *best_bo = NULL;
76 
77 	bo = RB_ROOT(&bdev->addr_space_rb);
78 	while (bo != NULL) {
79 		cur_offset = bo->vm_node->start;
80 		if (page_start >= cur_offset) {
81 			best_bo = bo;
82 			if (page_start == cur_offset)
83 				break;
84 			bo = RB_RIGHT(bo, vm_rb);
85 		} else
86 			bo = RB_LEFT(bo, vm_rb);
87 	}
88 
89 	if (unlikely(best_bo == NULL))
90 		return NULL;
91 
92 	if (unlikely((best_bo->vm_node->start + best_bo->num_pages) <
93 		     (page_start + num_pages)))
94 		return NULL;
95 
96 	return best_bo;
97 }
98 
99 static int
ttm_bo_vm_fault(vm_object_t vm_obj,vm_ooffset_t offset,int prot,vm_page_t * mres)100 ttm_bo_vm_fault(vm_object_t vm_obj, vm_ooffset_t offset,
101     int prot, vm_page_t *mres)
102 {
103 
104 	struct ttm_buffer_object *bo = vm_obj->handle;
105 	struct ttm_bo_device *bdev = bo->bdev;
106 	struct ttm_tt *ttm = NULL;
107 	vm_page_t m, m1;
108 	int ret;
109 	int retval = VM_PAGER_OK;
110 	struct ttm_mem_type_manager *man =
111 		&bdev->man[bo->mem.mem_type];
112 
113 	vm_object_pip_add(vm_obj, 1);
114 	if (*mres != NULL) {
115 		(void)vm_page_remove(*mres);
116 	}
117 retry:
118 	VM_OBJECT_WUNLOCK(vm_obj);
119 	m = NULL;
120 
121 reserve:
122 	ret = ttm_bo_reserve(bo, false, false, false, 0);
123 	if (unlikely(ret != 0)) {
124 		if (ret == -EBUSY) {
125 			kern_yield(PRI_USER);
126 			goto reserve;
127 		}
128 	}
129 
130 	if (bdev->driver->fault_reserve_notify) {
131 		ret = bdev->driver->fault_reserve_notify(bo);
132 		switch (ret) {
133 		case 0:
134 			break;
135 		case -EBUSY:
136 		case -ERESTARTSYS:
137 		case -EINTR:
138 			kern_yield(PRI_USER);
139 			goto reserve;
140 		default:
141 			retval = VM_PAGER_ERROR;
142 			goto out_unlock;
143 		}
144 	}
145 
146 	/*
147 	 * Wait for buffer data in transit, due to a pipelined
148 	 * move.
149 	 */
150 
151 	mtx_lock(&bdev->fence_lock);
152 	if (test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)) {
153 		/*
154 		 * Here, the behavior differs between Linux and FreeBSD.
155 		 *
156 		 * On Linux, the wait is interruptible (3rd argument to
157 		 * ttm_bo_wait). There must be some mechanism to resume
158 		 * page fault handling, once the signal is processed.
159 		 *
160 		 * On FreeBSD, the wait is uninteruptible. This is not a
161 		 * problem as we can't end up with an unkillable process
162 		 * here, because the wait will eventually time out.
163 		 *
164 		 * An example of this situation is the Xorg process
165 		 * which uses SIGALRM internally. The signal could
166 		 * interrupt the wait, causing the page fault to fail
167 		 * and the process to receive SIGSEGV.
168 		 */
169 		ret = ttm_bo_wait(bo, false, false, false);
170 		mtx_unlock(&bdev->fence_lock);
171 		if (unlikely(ret != 0)) {
172 			retval = VM_PAGER_ERROR;
173 			goto out_unlock;
174 		}
175 	} else
176 		mtx_unlock(&bdev->fence_lock);
177 
178 	ret = ttm_mem_io_lock(man, true);
179 	if (unlikely(ret != 0)) {
180 		retval = VM_PAGER_ERROR;
181 		goto out_unlock;
182 	}
183 	ret = ttm_mem_io_reserve_vm(bo);
184 	if (unlikely(ret != 0)) {
185 		retval = VM_PAGER_ERROR;
186 		goto out_io_unlock;
187 	}
188 
189 	/*
190 	 * Strictly, we're not allowed to modify vma->vm_page_prot here,
191 	 * since the mmap_sem is only held in read mode. However, we
192 	 * modify only the caching bits of vma->vm_page_prot and
193 	 * consider those bits protected by
194 	 * the bo->mutex, as we should be the only writers.
195 	 * There shouldn't really be any readers of these bits except
196 	 * within vm_insert_mixed()? fork?
197 	 *
198 	 * TODO: Add a list of vmas to the bo, and change the
199 	 * vma->vm_page_prot when the object changes caching policy, with
200 	 * the correct locks held.
201 	 */
202 	if (!bo->mem.bus.is_iomem) {
203 		/* Allocate all page at once, most common usage */
204 		ttm = bo->ttm;
205 		if (ttm->bdev->driver->ttm_tt_populate(ttm)) {
206 			retval = VM_PAGER_ERROR;
207 			goto out_io_unlock;
208 		}
209 	}
210 
211 	if (bo->mem.bus.is_iomem) {
212 		m = PHYS_TO_VM_PAGE(bo->mem.bus.base + bo->mem.bus.offset +
213 		    offset);
214 		KASSERT((m->flags & PG_FICTITIOUS) != 0,
215 		    ("physical address %#jx not fictitious",
216 		    (uintmax_t)(bo->mem.bus.base + bo->mem.bus.offset
217 		    + offset)));
218 		pmap_page_set_memattr(m, ttm_io_prot(bo->mem.placement));
219 	} else {
220 		ttm = bo->ttm;
221 		m = ttm->pages[OFF_TO_IDX(offset)];
222 		if (unlikely(!m)) {
223 			retval = VM_PAGER_ERROR;
224 			goto out_io_unlock;
225 		}
226 		pmap_page_set_memattr(m,
227 		    (bo->mem.placement & TTM_PL_FLAG_CACHED) ?
228 		    VM_MEMATTR_WRITE_BACK : ttm_io_prot(bo->mem.placement));
229 	}
230 
231 	VM_OBJECT_WLOCK(vm_obj);
232 	if (vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL) == 0) {
233 		ttm_mem_io_unlock(man);
234 		ttm_bo_unreserve(bo);
235 		goto retry;
236 	}
237 	m1 = vm_page_lookup(vm_obj, OFF_TO_IDX(offset));
238 	/* XXX This looks like it should just be vm_page_replace? */
239 	if (m1 == NULL) {
240 		if (vm_page_insert(m, vm_obj, OFF_TO_IDX(offset))) {
241 			vm_page_xunbusy(m);
242 			VM_OBJECT_WUNLOCK(vm_obj);
243 			vm_wait(vm_obj);
244 			VM_OBJECT_WLOCK(vm_obj);
245 			ttm_mem_io_unlock(man);
246 			ttm_bo_unreserve(bo);
247 			goto retry;
248 		}
249 	} else {
250 		KASSERT(m == m1,
251 		    ("inconsistent insert bo %p m %p m1 %p offset %jx",
252 		    bo, m, m1, (uintmax_t)offset));
253 	}
254 	vm_page_valid(m);
255 	if (*mres != NULL) {
256 		KASSERT(*mres != m, ("losing %p %p", *mres, m));
257 		vm_page_xunbusy(*mres);
258 		vm_page_free(*mres);
259 	}
260 	*mres = m;
261 
262 out_io_unlock1:
263 	ttm_mem_io_unlock(man);
264 out_unlock1:
265 	ttm_bo_unreserve(bo);
266 	vm_object_pip_wakeup(vm_obj);
267 	return (retval);
268 
269 out_io_unlock:
270 	VM_OBJECT_WLOCK(vm_obj);
271 	goto out_io_unlock1;
272 
273 out_unlock:
274 	VM_OBJECT_WLOCK(vm_obj);
275 	goto out_unlock1;
276 }
277 
278 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)279 ttm_bo_vm_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
280     vm_ooffset_t foff, struct ucred *cred, u_short *color)
281 {
282 
283 	/*
284 	 * On Linux, a reference to the buffer object is acquired here.
285 	 * The reason is that this function is not called when the
286 	 * mmap() is initialized, but only when a process forks for
287 	 * instance. Therefore on Linux, the reference on the bo is
288 	 * acquired either in ttm_bo_mmap() or ttm_bo_vm_open(). It's
289 	 * then released in ttm_bo_vm_close().
290 	 *
291 	 * Here, this function is called during mmap() initialization.
292 	 * Thus, the reference acquired in ttm_bo_mmap_single() is
293 	 * sufficient.
294 	 */
295 
296 	*color = 0;
297 	return (0);
298 }
299 
300 static void
ttm_bo_vm_dtor(void * handle)301 ttm_bo_vm_dtor(void *handle)
302 {
303 	struct ttm_buffer_object *bo = handle;
304 
305 	ttm_bo_unref(&bo);
306 }
307 
308 static struct cdev_pager_ops ttm_pager_ops = {
309 	.cdev_pg_fault = ttm_bo_vm_fault,
310 	.cdev_pg_ctor = ttm_bo_vm_ctor,
311 	.cdev_pg_dtor = ttm_bo_vm_dtor
312 };
313 
314 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)315 ttm_bo_mmap_single(struct ttm_bo_device *bdev, vm_ooffset_t *offset, vm_size_t size,
316     struct vm_object **obj_res, int nprot)
317 {
318 	struct ttm_bo_driver *driver;
319 	struct ttm_buffer_object *bo;
320 	struct vm_object *vm_obj;
321 	int ret;
322 
323 	rw_wlock(&bdev->vm_lock);
324 	bo = ttm_bo_vm_lookup_rb(bdev, OFF_TO_IDX(*offset), OFF_TO_IDX(size));
325 	if (likely(bo != NULL))
326 		refcount_acquire(&bo->kref);
327 	rw_wunlock(&bdev->vm_lock);
328 
329 	if (unlikely(bo == NULL)) {
330 		printf("[TTM] Could not find buffer object to map\n");
331 		return (-EINVAL);
332 	}
333 
334 	driver = bo->bdev->driver;
335 	if (unlikely(!driver->verify_access)) {
336 		ret = -EPERM;
337 		goto out_unref;
338 	}
339 	ret = driver->verify_access(bo);
340 	if (unlikely(ret != 0))
341 		goto out_unref;
342 
343 	vm_obj = cdev_pager_allocate(bo, OBJT_MGTDEVICE, &ttm_pager_ops,
344 	    size, nprot, 0, curthread->td_ucred);
345 	if (vm_obj == NULL) {
346 		ret = -EINVAL;
347 		goto out_unref;
348 	}
349 	/*
350 	 * Note: We're transferring the bo reference to vm_obj->handle here.
351 	 */
352 	*offset = 0;
353 	*obj_res = vm_obj;
354 	return 0;
355 out_unref:
356 	ttm_bo_unref(&bo);
357 	return ret;
358 }
359 
360 void
ttm_bo_release_mmap(struct ttm_buffer_object * bo)361 ttm_bo_release_mmap(struct ttm_buffer_object *bo)
362 {
363 	vm_object_t vm_obj;
364 
365 	vm_obj = cdev_pager_lookup(bo);
366 	if (vm_obj != NULL) {
367 		cdev_mgtdev_pager_free_pages(vm_obj);
368 		vm_object_deallocate(vm_obj);
369 	}
370 }
371 
372 #if 0
373 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
374 {
375 	if (vma->vm_pgoff != 0)
376 		return -EACCES;
377 
378 	vma->vm_ops = &ttm_bo_vm_ops;
379 	vma->vm_private_data = ttm_bo_reference(bo);
380 	vma->vm_flags |= VM_IO | VM_MIXEDMAP | VM_DONTEXPAND;
381 	return 0;
382 }
383 
384 ssize_t ttm_bo_io(struct ttm_bo_device *bdev, struct file *filp,
385 		  const char __user *wbuf, char __user *rbuf, size_t count,
386 		  loff_t *f_pos, bool write)
387 {
388 	struct ttm_buffer_object *bo;
389 	struct ttm_bo_driver *driver;
390 	struct ttm_bo_kmap_obj map;
391 	unsigned long dev_offset = (*f_pos >> PAGE_SHIFT);
392 	unsigned long kmap_offset;
393 	unsigned long kmap_end;
394 	unsigned long kmap_num;
395 	size_t io_size;
396 	unsigned int page_offset;
397 	char *virtual;
398 	int ret;
399 	bool no_wait = false;
400 	bool dummy;
401 
402 	read_lock(&bdev->vm_lock);
403 	bo = ttm_bo_vm_lookup_rb(bdev, dev_offset, 1);
404 	if (likely(bo != NULL))
405 		ttm_bo_reference(bo);
406 	read_unlock(&bdev->vm_lock);
407 
408 	if (unlikely(bo == NULL))
409 		return -EFAULT;
410 
411 	driver = bo->bdev->driver;
412 	if (unlikely(!driver->verify_access)) {
413 		ret = -EPERM;
414 		goto out_unref;
415 	}
416 
417 	ret = driver->verify_access(bo, filp);
418 	if (unlikely(ret != 0))
419 		goto out_unref;
420 
421 	kmap_offset = dev_offset - bo->vm_node->start;
422 	if (unlikely(kmap_offset >= bo->num_pages)) {
423 		ret = -EFBIG;
424 		goto out_unref;
425 	}
426 
427 	page_offset = *f_pos & ~PAGE_MASK;
428 	io_size = bo->num_pages - kmap_offset;
429 	io_size = (io_size << PAGE_SHIFT) - page_offset;
430 	if (count < io_size)
431 		io_size = count;
432 
433 	kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT;
434 	kmap_num = kmap_end - kmap_offset + 1;
435 
436 	ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
437 
438 	switch (ret) {
439 	case 0:
440 		break;
441 	case -EBUSY:
442 		ret = -EAGAIN;
443 		goto out_unref;
444 	default:
445 		goto out_unref;
446 	}
447 
448 	ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map);
449 	if (unlikely(ret != 0)) {
450 		ttm_bo_unreserve(bo);
451 		goto out_unref;
452 	}
453 
454 	virtual = ttm_kmap_obj_virtual(&map, &dummy);
455 	virtual += page_offset;
456 
457 	if (write)
458 		ret = copy_from_user(virtual, wbuf, io_size);
459 	else
460 		ret = copy_to_user(rbuf, virtual, io_size);
461 
462 	ttm_bo_kunmap(&map);
463 	ttm_bo_unreserve(bo);
464 	ttm_bo_unref(&bo);
465 
466 	if (unlikely(ret != 0))
467 		return -EFBIG;
468 
469 	*f_pos += io_size;
470 
471 	return io_size;
472 out_unref:
473 	ttm_bo_unref(&bo);
474 	return ret;
475 }
476 
477 ssize_t ttm_bo_fbdev_io(struct ttm_buffer_object *bo, const char __user *wbuf,
478 			char __user *rbuf, size_t count, loff_t *f_pos,
479 			bool write)
480 {
481 	struct ttm_bo_kmap_obj map;
482 	unsigned long kmap_offset;
483 	unsigned long kmap_end;
484 	unsigned long kmap_num;
485 	size_t io_size;
486 	unsigned int page_offset;
487 	char *virtual;
488 	int ret;
489 	bool no_wait = false;
490 	bool dummy;
491 
492 	kmap_offset = (*f_pos >> PAGE_SHIFT);
493 	if (unlikely(kmap_offset >= bo->num_pages))
494 		return -EFBIG;
495 
496 	page_offset = *f_pos & ~PAGE_MASK;
497 	io_size = bo->num_pages - kmap_offset;
498 	io_size = (io_size << PAGE_SHIFT) - page_offset;
499 	if (count < io_size)
500 		io_size = count;
501 
502 	kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT;
503 	kmap_num = kmap_end - kmap_offset + 1;
504 
505 	ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
506 
507 	switch (ret) {
508 	case 0:
509 		break;
510 	case -EBUSY:
511 		return -EAGAIN;
512 	default:
513 		return ret;
514 	}
515 
516 	ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map);
517 	if (unlikely(ret != 0)) {
518 		ttm_bo_unreserve(bo);
519 		return ret;
520 	}
521 
522 	virtual = ttm_kmap_obj_virtual(&map, &dummy);
523 	virtual += page_offset;
524 
525 	if (write)
526 		ret = copy_from_user(virtual, wbuf, io_size);
527 	else
528 		ret = copy_to_user(rbuf, virtual, io_size);
529 
530 	ttm_bo_kunmap(&map);
531 	ttm_bo_unreserve(bo);
532 	ttm_bo_unref(&bo);
533 
534 	if (unlikely(ret != 0))
535 		return ret;
536 
537 	*f_pos += io_size;
538 
539 	return io_size;
540 }
541 #endif
542