xref: /freebsd/sys/vm/device_pager.c (revision 1123986db5f0b0b3bfc68840a6d8e12f93d6ca3e)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1990 University of Utah.
5  * Copyright (c) 1991, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * the Systems Programming Group of the University of Utah Computer
10  * Science Department.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  */
36 
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/conf.h>
40 #include <sys/lock.h>
41 #include <sys/proc.h>
42 #include <sys/mutex.h>
43 #include <sys/mman.h>
44 #include <sys/rwlock.h>
45 #include <sys/sx.h>
46 #include <sys/user.h>
47 #include <sys/vmmeter.h>
48 
49 #include <vm/vm.h>
50 #include <vm/vm_param.h>
51 #include <vm/vm_object.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_pager.h>
54 #include <vm/vm_radix.h>
55 #include <vm/vm_phys.h>
56 #include <vm/vm_radix.h>
57 #include <vm/uma.h>
58 
59 static void dev_pager_init(void);
60 static vm_object_t dev_pager_alloc(void *, vm_ooffset_t, vm_prot_t,
61     vm_ooffset_t, struct ucred *);
62 static void dev_pager_dealloc(vm_object_t);
63 static int dev_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *);
64 static void dev_pager_putpages(vm_object_t, vm_page_t *, int, int, int *);
65 static boolean_t dev_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
66 static void dev_pager_free_page(vm_object_t object, vm_page_t m);
67 static int dev_pager_populate(vm_object_t object, vm_pindex_t pidx,
68     int fault_type, vm_prot_t, vm_pindex_t *first, vm_pindex_t *last);
69 
70 /* list of device pager objects */
71 static struct pagerlst dev_pager_object_list;
72 /* protect list manipulation */
73 static struct mtx dev_pager_mtx;
74 
75 const struct pagerops devicepagerops = {
76 	.pgo_kvme_type = KVME_TYPE_DEVICE,
77 	.pgo_init =	dev_pager_init,
78 	.pgo_alloc =	dev_pager_alloc,
79 	.pgo_dealloc =	dev_pager_dealloc,
80 	.pgo_getpages =	dev_pager_getpages,
81 	.pgo_putpages =	dev_pager_putpages,
82 	.pgo_haspage =	dev_pager_haspage,
83 };
84 
85 const struct pagerops mgtdevicepagerops = {
86 	.pgo_kvme_type = KVME_TYPE_MGTDEVICE,
87 	.pgo_alloc =	dev_pager_alloc,
88 	.pgo_dealloc =	dev_pager_dealloc,
89 	.pgo_getpages =	dev_pager_getpages,
90 	.pgo_putpages =	dev_pager_putpages,
91 	.pgo_haspage =	dev_pager_haspage,
92 	.pgo_populate =	dev_pager_populate,
93 };
94 
95 static int old_dev_pager_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
96     vm_ooffset_t foff, struct ucred *cred, u_short *color);
97 static void old_dev_pager_dtor(void *handle);
98 static int old_dev_pager_fault(vm_object_t object, vm_ooffset_t offset,
99     int prot, vm_page_t *mres);
100 static void old_dev_pager_path(void *handle, char *path, size_t len);
101 
102 static const struct cdev_pager_ops old_dev_pager_ops = {
103 	.cdev_pg_ctor =	old_dev_pager_ctor,
104 	.cdev_pg_dtor =	old_dev_pager_dtor,
105 	.cdev_pg_fault = old_dev_pager_fault,
106 	.cdev_pg_path = old_dev_pager_path
107 };
108 
109 static void
dev_pager_init(void)110 dev_pager_init(void)
111 {
112 
113 	TAILQ_INIT(&dev_pager_object_list);
114 	mtx_init(&dev_pager_mtx, "dev_pager list", NULL, MTX_DEF);
115 }
116 
117 vm_object_t
cdev_pager_lookup(void * handle)118 cdev_pager_lookup(void *handle)
119 {
120 	vm_object_t object;
121 
122 again:
123 	mtx_lock(&dev_pager_mtx);
124 	object = vm_pager_object_lookup(&dev_pager_object_list, handle);
125 	if (object != NULL && object->un_pager.devp.handle == NULL) {
126 		msleep(&object->un_pager.devp.handle, &dev_pager_mtx,
127 		    PVM | PDROP, "cdplkp", 0);
128 		vm_object_deallocate(object);
129 		goto again;
130 	}
131 	mtx_unlock(&dev_pager_mtx);
132 	return (object);
133 }
134 
135 vm_object_t
cdev_pager_allocate(void * handle,enum obj_type tp,const struct cdev_pager_ops * ops,vm_ooffset_t size,vm_prot_t prot,vm_ooffset_t foff,struct ucred * cred)136 cdev_pager_allocate(void *handle, enum obj_type tp,
137     const struct cdev_pager_ops *ops, vm_ooffset_t size, vm_prot_t prot,
138     vm_ooffset_t foff, struct ucred *cred)
139 {
140 	vm_object_t object;
141 	vm_pindex_t pindex;
142 
143 	KASSERT(handle != NULL, ("device pager with NULL handle"));
144 
145 	if (tp != OBJT_DEVICE && tp != OBJT_MGTDEVICE)
146 		return (NULL);
147 	KASSERT(tp == OBJT_MGTDEVICE || ops->cdev_pg_populate == NULL,
148 	    ("populate on unmanaged device pager"));
149 
150 	/*
151 	 * Offset should be page aligned.
152 	 */
153 	if (foff & PAGE_MASK)
154 		return (NULL);
155 
156 	/*
157 	 * Treat the mmap(2) file offset as an unsigned value for a
158 	 * device mapping.  This, in effect, allows a user to pass all
159 	 * possible off_t values as the mapping cookie to the driver.  At
160 	 * this point, we know that both foff and size are a multiple
161 	 * of the page size.  Do a check to avoid wrap.
162 	 */
163 	size = round_page(size);
164 	pindex = OFF_TO_IDX(foff) + OFF_TO_IDX(size);
165 	if (pindex > OBJ_MAX_SIZE || pindex < OFF_TO_IDX(foff) ||
166 	    pindex < OFF_TO_IDX(size))
167 		return (NULL);
168 
169 again:
170 	mtx_lock(&dev_pager_mtx);
171 
172 	/*
173 	 * Look up pager, creating as necessary.
174 	 */
175 	object = vm_pager_object_lookup(&dev_pager_object_list, handle);
176 	if (object == NULL) {
177 		vm_object_t object1;
178 
179 		/*
180 		 * Allocate object and associate it with the pager.  Initialize
181 		 * the object's pg_color based upon the physical address of the
182 		 * device's memory.
183 		 */
184 		mtx_unlock(&dev_pager_mtx);
185 		object1 = vm_object_allocate(tp, pindex);
186 		mtx_lock(&dev_pager_mtx);
187 		object = vm_pager_object_lookup(&dev_pager_object_list, handle);
188 		if (object != NULL) {
189 			object1->type = OBJT_DEAD;
190 			vm_object_deallocate(object1);
191 			object1 = NULL;
192 			if (object->un_pager.devp.handle == NULL) {
193 				msleep(&object->un_pager.devp.handle,
194 				    &dev_pager_mtx, PVM | PDROP, "cdplkp", 0);
195 				vm_object_deallocate(object);
196 				goto again;
197 			}
198 
199 			/*
200 			 * We raced with other thread while allocating object.
201 			 */
202 			if (pindex > object->size)
203 				object->size = pindex;
204 			KASSERT(object->type == tp,
205 			    ("Inconsistent device pager type %p %d",
206 			    object, tp));
207 			KASSERT(object->un_pager.devp.ops == ops,
208 			    ("Inconsistent devops %p %p", object, ops));
209 		} else {
210 			u_short color;
211 
212 			object = object1;
213 			object1 = NULL;
214 			object->handle = handle;
215 			object->un_pager.devp.ops = ops;
216 			TAILQ_INIT(&object->un_pager.devp.devp_pglist);
217 			TAILQ_INSERT_TAIL(&dev_pager_object_list, object,
218 			    pager_object_list);
219 			mtx_unlock(&dev_pager_mtx);
220 			if (ops->cdev_pg_populate != NULL)
221 				vm_object_set_flag(object, OBJ_POPULATE);
222 			if (ops->cdev_pg_ctor(handle, size, prot, foff,
223 			    cred, &color) != 0) {
224 				mtx_lock(&dev_pager_mtx);
225 				TAILQ_REMOVE(&dev_pager_object_list, object,
226 				    pager_object_list);
227 				wakeup(&object->un_pager.devp.handle);
228 				mtx_unlock(&dev_pager_mtx);
229 				object->type = OBJT_DEAD;
230 				vm_object_deallocate(object);
231 				object = NULL;
232 				mtx_lock(&dev_pager_mtx);
233 			} else {
234 				mtx_lock(&dev_pager_mtx);
235 				object->flags |= OBJ_COLORED;
236 				object->pg_color = color;
237 				object->un_pager.devp.handle = handle;
238 				wakeup(&object->un_pager.devp.handle);
239 			}
240 		}
241 		MPASS(object1 == NULL);
242 	} else {
243 		if (object->un_pager.devp.handle == NULL) {
244 			msleep(&object->un_pager.devp.handle,
245 			    &dev_pager_mtx, PVM | PDROP, "cdplkp", 0);
246 			vm_object_deallocate(object);
247 			goto again;
248 		}
249 		if (pindex > object->size)
250 			object->size = pindex;
251 		KASSERT(object->type == tp,
252 		    ("Inconsistent device pager type %p %d", object, tp));
253 	}
254 	mtx_unlock(&dev_pager_mtx);
255 	return (object);
256 }
257 
258 static vm_object_t
dev_pager_alloc(void * handle,vm_ooffset_t size,vm_prot_t prot,vm_ooffset_t foff,struct ucred * cred)259 dev_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
260     vm_ooffset_t foff, struct ucred *cred)
261 {
262 
263 	return (cdev_pager_allocate(handle, OBJT_DEVICE, &old_dev_pager_ops,
264 	    size, prot, foff, cred));
265 }
266 
267 void
cdev_pager_get_path(vm_object_t object,char * path,size_t sz)268 cdev_pager_get_path(vm_object_t object, char *path, size_t sz)
269 {
270 	if (object->un_pager.devp.ops->cdev_pg_path != NULL)
271 		object->un_pager.devp.ops->cdev_pg_path(
272 		    object->un_pager.devp.handle, path, sz);
273 }
274 
275 void
cdev_pager_free_page(vm_object_t object,vm_page_t m)276 cdev_pager_free_page(vm_object_t object, vm_page_t m)
277 {
278 
279 	if (object->type == OBJT_MGTDEVICE) {
280 		struct pctrie_iter pages;
281 
282 		vm_page_iter_init(&pages, object);
283 		vm_radix_iter_lookup(&pages, m->pindex);
284 		cdev_mgtdev_pager_free_page(&pages, m);
285 	} else if (object->type == OBJT_DEVICE)
286 		dev_pager_free_page(object, m);
287 	else
288 		KASSERT(false,
289 		    ("Invalid device type obj %p m %p", object, m));
290 }
291 
292 void
cdev_mgtdev_pager_free_page(struct pctrie_iter * pages,vm_page_t m)293 cdev_mgtdev_pager_free_page(struct pctrie_iter *pages, vm_page_t m)
294 {
295 	pmap_remove_all(m);
296 	vm_page_iter_remove(pages, m);
297 }
298 
299 void
cdev_mgtdev_pager_free_pages(vm_object_t object)300 cdev_mgtdev_pager_free_pages(vm_object_t object)
301 {
302 	struct pctrie_iter pages;
303 	vm_page_t m;
304 
305 	vm_page_iter_init(&pages, object);
306 	VM_OBJECT_WLOCK(object);
307 retry:
308 	KASSERT(pctrie_iter_is_reset(&pages),
309 	    ("%s: pctrie_iter not reset for retry", __func__));
310 	for (m = vm_radix_iter_lookup_ge(&pages, 0); m != NULL;
311 	    m = vm_radix_iter_step(&pages)) {
312 		if (!vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL)) {
313 			pctrie_iter_reset(&pages);
314 			goto retry;
315 		}
316 		cdev_mgtdev_pager_free_page(&pages, m);
317 	}
318 	VM_OBJECT_WUNLOCK(object);
319 }
320 
321 static void
dev_pager_free_page(vm_object_t object,vm_page_t m)322 dev_pager_free_page(vm_object_t object, vm_page_t m)
323 {
324 
325 	VM_OBJECT_ASSERT_WLOCKED(object);
326 	KASSERT((object->type == OBJT_DEVICE &&
327 	    (m->oflags & VPO_UNMANAGED) != 0),
328 	    ("Managed device or page obj %p m %p", object, m));
329 	TAILQ_REMOVE(&object->un_pager.devp.devp_pglist, m, plinks.q);
330 	vm_page_putfake(m);
331 }
332 
333 static void
dev_pager_dealloc(vm_object_t object)334 dev_pager_dealloc(vm_object_t object)
335 {
336 	vm_page_t m;
337 
338 	VM_OBJECT_WUNLOCK(object);
339 	object->un_pager.devp.ops->cdev_pg_dtor(object->un_pager.devp.handle);
340 
341 	mtx_lock(&dev_pager_mtx);
342 	TAILQ_REMOVE(&dev_pager_object_list, object, pager_object_list);
343 	mtx_unlock(&dev_pager_mtx);
344 	VM_OBJECT_WLOCK(object);
345 
346 	if (object->type == OBJT_DEVICE) {
347 		/*
348 		 * Free up our fake pages.
349 		 */
350 		while ((m = TAILQ_FIRST(&object->un_pager.devp.devp_pglist))
351 		    != NULL) {
352 			if (vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL) == 0)
353 				continue;
354 
355 			dev_pager_free_page(object, m);
356 		}
357 	}
358 	object->handle = NULL;
359 	object->type = OBJT_DEAD;
360 }
361 
362 static int
dev_pager_getpages(vm_object_t object,vm_page_t * ma,int count,int * rbehind,int * rahead)363 dev_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind,
364     int *rahead)
365 {
366 	int error;
367 
368 	/* Since our haspage reports zero after/before, the count is 1. */
369 	KASSERT(count == 1, ("%s: count %d", __func__, count));
370 	if (object->un_pager.devp.ops->cdev_pg_fault == NULL)
371 		return (VM_PAGER_FAIL);
372 	VM_OBJECT_WLOCK(object);
373 	error = object->un_pager.devp.ops->cdev_pg_fault(object,
374 	    IDX_TO_OFF(ma[0]->pindex), PROT_READ, &ma[0]);
375 
376 	VM_OBJECT_ASSERT_WLOCKED(object);
377 
378 	if (error == VM_PAGER_OK) {
379 		KASSERT((object->type == OBJT_DEVICE &&
380 		     (ma[0]->oflags & VPO_UNMANAGED) != 0) ||
381 		    (object->type == OBJT_MGTDEVICE &&
382 		     (ma[0]->oflags & VPO_UNMANAGED) == 0),
383 		    ("Wrong page type %p %p", ma[0], object));
384 		if (object->type == OBJT_DEVICE) {
385 			TAILQ_INSERT_TAIL(&object->un_pager.devp.devp_pglist,
386 			    ma[0], plinks.q);
387 		}
388 		if (rbehind)
389 			*rbehind = 0;
390 		if (rahead)
391 			*rahead = 0;
392 	}
393 	VM_OBJECT_WUNLOCK(object);
394 
395 	return (error);
396 }
397 
398 static int
dev_pager_populate(vm_object_t object,vm_pindex_t pidx,int fault_type,vm_prot_t max_prot,vm_pindex_t * first,vm_pindex_t * last)399 dev_pager_populate(vm_object_t object, vm_pindex_t pidx, int fault_type,
400     vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last)
401 {
402 
403 	VM_OBJECT_ASSERT_WLOCKED(object);
404 	if (object->un_pager.devp.ops->cdev_pg_populate == NULL)
405 		return (VM_PAGER_FAIL);
406 	return (object->un_pager.devp.ops->cdev_pg_populate(object, pidx,
407 	    fault_type, max_prot, first, last));
408 }
409 
410 static int
old_dev_pager_fault(vm_object_t object,vm_ooffset_t offset,int prot,vm_page_t * mres)411 old_dev_pager_fault(vm_object_t object, vm_ooffset_t offset, int prot,
412     vm_page_t *mres)
413 {
414 	vm_paddr_t paddr;
415 	vm_page_t m_paddr, page;
416 	struct cdev *dev;
417 	struct cdevsw *csw;
418 	struct file *fpop;
419 	struct thread *td;
420 	vm_memattr_t memattr, memattr1;
421 	int ref, ret;
422 
423 	memattr = object->memattr;
424 
425 	VM_OBJECT_WUNLOCK(object);
426 
427 	dev = object->handle;
428 	csw = dev_refthread(dev, &ref);
429 	if (csw == NULL) {
430 		VM_OBJECT_WLOCK(object);
431 		return (VM_PAGER_FAIL);
432 	}
433 	td = curthread;
434 	fpop = td->td_fpop;
435 	td->td_fpop = NULL;
436 	ret = csw->d_mmap(dev, offset, &paddr, prot, &memattr);
437 	td->td_fpop = fpop;
438 	dev_relthread(dev, ref);
439 	if (ret != 0) {
440 		printf(
441 	    "WARNING: dev_pager_getpage: map function returns error %d", ret);
442 		VM_OBJECT_WLOCK(object);
443 		return (VM_PAGER_FAIL);
444 	}
445 
446 	/* If "paddr" is a real page, perform a sanity check on "memattr". */
447 	if ((m_paddr = vm_phys_paddr_to_vm_page(paddr)) != NULL &&
448 	    (memattr1 = pmap_page_get_memattr(m_paddr)) != memattr) {
449 		/*
450 		 * For the /dev/mem d_mmap routine to return the
451 		 * correct memattr, pmap_page_get_memattr() needs to
452 		 * be called, which we do there.
453 		 */
454 		if ((csw->d_flags & D_MEM) == 0) {
455 			printf("WARNING: Device driver %s has set "
456 			    "\"memattr\" inconsistently (drv %u pmap %u).\n",
457 			    csw->d_name, memattr, memattr1);
458 		}
459 		memattr = memattr1;
460 	}
461 	if (((*mres)->flags & PG_FICTITIOUS) != 0) {
462 		/*
463 		 * If the passed in result page is a fake page, update it with
464 		 * the new physical address.
465 		 */
466 		page = *mres;
467 		VM_OBJECT_WLOCK(object);
468 		vm_page_updatefake(page, paddr, memattr);
469 	} else {
470 		/*
471 		 * Replace the passed in reqpage page with our own fake page and
472 		 * free up the all of the original pages.
473 		 */
474 		page = vm_page_getfake(paddr, memattr);
475 		VM_OBJECT_WLOCK(object);
476 		vm_page_replace(page, object, (*mres)->pindex, *mres);
477 		*mres = page;
478 	}
479 	vm_page_valid(page);
480 	return (VM_PAGER_OK);
481 }
482 
483 static void
dev_pager_putpages(vm_object_t object,vm_page_t * m,int count,int flags,int * rtvals)484 dev_pager_putpages(vm_object_t object, vm_page_t *m, int count, int flags,
485     int *rtvals)
486 {
487 
488 	panic("dev_pager_putpage called");
489 }
490 
491 static boolean_t
dev_pager_haspage(vm_object_t object,vm_pindex_t pindex,int * before,int * after)492 dev_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
493     int *after)
494 {
495 
496 	if (before != NULL)
497 		*before = 0;
498 	if (after != NULL)
499 		*after = 0;
500 	return (TRUE);
501 }
502 
503 static int
old_dev_pager_ctor(void * handle,vm_ooffset_t size,vm_prot_t prot,vm_ooffset_t foff,struct ucred * cred,u_short * color)504 old_dev_pager_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
505     vm_ooffset_t foff, struct ucred *cred, u_short *color)
506 {
507 	struct cdev *dev;
508 	struct cdevsw *csw;
509 	vm_memattr_t dummy;
510 	vm_ooffset_t off;
511 	vm_paddr_t paddr;
512 	unsigned int npages;
513 	int ref;
514 
515 	/*
516 	 * Make sure this device can be mapped.
517 	 */
518 	dev = handle;
519 	csw = dev_refthread(dev, &ref);
520 	if (csw == NULL)
521 		return (ENXIO);
522 
523 	/*
524 	 * Check that the specified range of the device allows the desired
525 	 * protection.
526 	 *
527 	 * XXX assumes VM_PROT_* == PROT_*
528 	 */
529 	npages = OFF_TO_IDX(size);
530 	paddr = 0; /* Make paddr initialized for the case of size == 0. */
531 	for (off = foff; npages--; off += PAGE_SIZE) {
532 		if (csw->d_mmap(dev, off, &paddr, (int)prot, &dummy) != 0) {
533 			dev_relthread(dev, ref);
534 			return (EINVAL);
535 		}
536 	}
537 
538 	dev_ref(dev);
539 	dev_relthread(dev, ref);
540 	*color = atop(paddr) - OFF_TO_IDX(off - PAGE_SIZE);
541 	return (0);
542 }
543 
544 static void
old_dev_pager_dtor(void * handle)545 old_dev_pager_dtor(void *handle)
546 {
547 
548 	dev_rel(handle);
549 }
550 
551 static void
old_dev_pager_path(void * handle,char * path,size_t len)552 old_dev_pager_path(void *handle, char *path, size_t len)
553 {
554 	struct cdev *cdev = handle;
555 
556 	if (cdev != NULL)
557 		dev_copyname(cdev, path, len);
558 }
559