xref: /openbsd-src/sys/uvm/uvm_device.c (revision d13be5d47e4149db2549a9828e244d59dbc43f15)
1 /*	$OpenBSD: uvm_device.c,v 1.40 2011/07/03 18:34:14 oga Exp $	*/
2 /*	$NetBSD: uvm_device.c,v 1.30 2000/11/25 06:27:59 chs Exp $	*/
3 
4 /*
5  *
6  * Copyright (c) 1997 Charles D. Cranor and Washington University.
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. All advertising materials mentioning features or use of this software
18  *    must display the following acknowledgement:
19  *      This product includes software developed by Charles D. Cranor and
20  *      Washington University.
21  * 4. The name of the author may not be used to endorse or promote products
22  *    derived from this software without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
25  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
26  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
27  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
28  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
29  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
30  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
31  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
32  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
33  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
34  *
35  * from: Id: uvm_device.c,v 1.1.2.9 1998/02/06 05:11:47 chs Exp
36  */
37 
38 /*
39  * uvm_device.c: the device pager.
40  */
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/conf.h>
45 #include <sys/proc.h>
46 #include <sys/malloc.h>
47 #include <sys/mutex.h>
48 #include <sys/vnode.h>
49 
50 #include <uvm/uvm.h>
51 #include <uvm/uvm_device.h>
52 
53 /*
54  * private global data structure
55  *
56  * we keep a list of active device objects in the system.
57  */
58 
59 LIST_HEAD(, uvm_device) udv_list = LIST_HEAD_INITIALIZER(udv_list);
60 struct mutex udv_lock = MUTEX_INITIALIZER(IPL_NONE);
61 
62 /*
63  * functions
64  */
65 
66 static void             udv_reference(struct uvm_object *);
67 static void             udv_detach(struct uvm_object *);
68 static int		udv_fault(struct uvm_faultinfo *, vaddr_t,
69 				       vm_page_t *, int, int, vm_fault_t,
70 				       vm_prot_t, int);
71 static boolean_t        udv_flush(struct uvm_object *, voff_t, voff_t,
72 				       int);
73 
74 /*
75  * master pager structure
76  */
77 
78 struct uvm_pagerops uvm_deviceops = {
79 	NULL,		/* inited statically */
80 	udv_reference,
81 	udv_detach,
82 	udv_fault,
83 	udv_flush,
84 };
85 
86 /*
87  * the ops.
88  */
89 
90 
91 /*
92  * udv_attach
93  *
94  * get a VM object that is associated with a device.   allocate a new
95  * one if needed.
96  *
97  * => caller must _not_ already be holding the lock on the uvm_object.
98  * => in fact, nothing should be locked so that we can sleep here.
99  *
100  * The last two arguments (off and size) are only used for access checking.
101  */
102 struct uvm_object *
103 udv_attach(void *arg, vm_prot_t accessprot, voff_t off, vsize_t size)
104 {
105 	dev_t device = *((dev_t *)arg);
106 	struct uvm_device *udv, *lcv;
107 	paddr_t (*mapfn)(dev_t, off_t, int);
108 
109 	/*
110 	 * before we do anything, ensure this device supports mmap
111 	 */
112 
113 	mapfn = cdevsw[major(device)].d_mmap;
114 	if (mapfn == NULL ||
115 	    mapfn == (paddr_t (*)(dev_t, off_t, int)) enodev ||
116 	    mapfn == (paddr_t (*)(dev_t, off_t, int)) nullop)
117 		return(NULL);
118 
119 	/*
120 	 * Negative offsets on the object are not allowed.
121 	 */
122 
123 	if (off < 0)
124 		return(NULL);
125 
126 	/*
127 	 * Check that the specified range of the device allows the
128 	 * desired protection.
129 	 *
130 	 * XXX assumes VM_PROT_* == PROT_*
131 	 * XXX clobbers off and size, but nothing else here needs them.
132 	 */
133 
134 	while (size != 0) {
135 		if ((*mapfn)(device, off, accessprot) == -1)
136 			return (NULL);
137 		off += PAGE_SIZE; size -= PAGE_SIZE;
138 	}
139 
140 	/*
141 	 * keep looping until we get it
142 	 */
143 
144 	for (;;) {
145 
146 		/*
147 		 * first, attempt to find it on the main list
148 		 */
149 
150 		mtx_enter(&udv_lock);
151 		LIST_FOREACH(lcv, &udv_list, u_list) {
152 			if (device == lcv->u_device)
153 				break;
154 		}
155 
156 		/*
157 		 * got it on main list.  put a hold on it and unlock udv_lock.
158 		 */
159 
160 		if (lcv) {
161 
162 			/*
163 			 * if someone else has a hold on it, sleep and start
164 			 * over again. Else, we need take HOLD flag so we
165 			 * don't have to re-order locking here.
166 			 */
167 
168 			if (lcv->u_flags & UVM_DEVICE_HOLD) {
169 				lcv->u_flags |= UVM_DEVICE_WANTED;
170 				msleep(lcv, &udv_lock, PVM | PNORELOCK,
171 				    "udv_attach", 0);
172 				continue;
173 			}
174 
175 			/* we are now holding it */
176 			lcv->u_flags |= UVM_DEVICE_HOLD;
177 			mtx_leave(&udv_lock);
178 
179 			/*
180 			 * bump reference count, unhold, return.
181 			 */
182 
183 			simple_lock(&lcv->u_obj.vmobjlock);
184 			lcv->u_obj.uo_refs++;
185 			simple_unlock(&lcv->u_obj.vmobjlock);
186 
187 			mtx_enter(&udv_lock);
188 			if (lcv->u_flags & UVM_DEVICE_WANTED)
189 				wakeup(lcv);
190 			lcv->u_flags &= ~(UVM_DEVICE_WANTED|UVM_DEVICE_HOLD);
191 			mtx_leave(&udv_lock);
192 			return(&lcv->u_obj);
193 		}
194 
195 		/*
196 		 * did not find it on main list.   need to malloc a new one.
197 		 */
198 
199 		mtx_leave(&udv_lock);
200 		/* NOTE: we could sleep in the following malloc() */
201 		udv = malloc(sizeof(*udv), M_TEMP, M_WAITOK);
202 		mtx_enter(&udv_lock);
203 
204 		/*
205 		 * now we have to double check to make sure no one added it
206 		 * to the list while we were sleeping...
207 		 */
208 
209 		LIST_FOREACH(lcv, &udv_list, u_list) {
210 			if (device == lcv->u_device)
211 				break;
212 		}
213 
214 		/*
215 		 * did we lose a race to someone else?
216 		 * free our memory and retry.
217 		 */
218 
219 		if (lcv) {
220 			mtx_leave(&udv_lock);
221 			free(udv, M_TEMP);
222 			continue;
223 		}
224 
225 		/*
226 		 * we have it!   init the data structures, add to list
227 		 * and return.
228 		 */
229 
230 		uvm_objinit(&udv->u_obj, &uvm_deviceops, 1);
231 		udv->u_flags = 0;
232 		udv->u_device = device;
233 		LIST_INSERT_HEAD(&udv_list, udv, u_list);
234 		mtx_leave(&udv_lock);
235 		return(&udv->u_obj);
236 	}
237 	/*NOTREACHED*/
238 }
239 
240 /*
241  * udv_reference
242  *
243  * add a reference to a VM object.   Note that the reference count must
244  * already be one (the passed in reference) so there is no chance of the
245  * udv being released or locked out here.
246  *
247  * => caller must call with object unlocked.
248  */
249 
250 static void
251 udv_reference(struct uvm_object *uobj)
252 {
253 
254 	simple_lock(&uobj->vmobjlock);
255 	uobj->uo_refs++;
256 	simple_unlock(&uobj->vmobjlock);
257 }
258 
259 /*
260  * udv_detach
261  *
262  * remove a reference to a VM object.
263  *
264  * => caller must call with object unlocked and map locked.
265  */
266 
267 static void
268 udv_detach(struct uvm_object *uobj)
269 {
270 	struct uvm_device *udv = (struct uvm_device *)uobj;
271 
272 	/*
273 	 * loop until done
274 	 */
275 again:
276 	simple_lock(&uobj->vmobjlock);
277 	if (uobj->uo_refs > 1) {
278 		uobj->uo_refs--;
279 		simple_unlock(&uobj->vmobjlock);
280 		return;
281 	}
282 	KASSERT(uobj->uo_npages == 0 && RB_EMPTY(&uobj->memt));
283 
284 	/*
285 	 * is it being held?   if so, wait until others are done.
286 	 */
287 
288 	mtx_enter(&udv_lock);
289 	if (udv->u_flags & UVM_DEVICE_HOLD) {
290 		udv->u_flags |= UVM_DEVICE_WANTED;
291 		/*
292 		 * lock interleaving. -- this is ok in this case since the
293 		 * locks are both IPL_NONE
294 		 */
295 		simple_unlock(&uobj->vmobjlock);
296 		msleep(udv, &udv_lock, PVM | PNORELOCK, "udv_detach", 0);
297 		goto again;
298 	}
299 
300 	/*
301 	 * got it!   nuke it now.
302 	 */
303 
304 	LIST_REMOVE(udv, u_list);
305 	if (udv->u_flags & UVM_DEVICE_WANTED)
306 		wakeup(udv);
307 	mtx_leave(&udv_lock);
308 	simple_unlock(&uobj->vmobjlock);
309 	free(udv, M_TEMP);
310 }
311 
312 
313 /*
314  * udv_flush
315  *
316  * flush pages out of a uvm object.   a no-op for devices.
317  */
318 
319 static boolean_t
320 udv_flush(struct uvm_object *uobj, voff_t start, voff_t stop, int flags)
321 {
322 
323 	return(TRUE);
324 }
325 
326 /*
327  * udv_fault: non-standard fault routine for device "pages"
328  *
329  * => rather than having a "get" function, we have a fault routine
330  *	since we don't return vm_pages we need full control over the
331  *	pmap_enter map in
332  * => all the usual fault data structured are locked by the caller
333  *	(i.e. maps(read), amap (if any), uobj)
334  * => on return, we unlock all fault data structures
335  * => flags: PGO_ALLPAGES: get all of the pages
336  *	     PGO_LOCKED: fault data structures are locked
337  *    XXX: currently PGO_LOCKED is always required ... consider removing
338  *	it as a flag
339  * => NOTE: vaddr is the VA of pps[0] in ufi->entry, _NOT_ pps[centeridx]
340  */
341 
342 static int
343 udv_fault(struct uvm_faultinfo *ufi, vaddr_t vaddr, vm_page_t *pps, int npages,
344     int centeridx, vm_fault_t fault_type, vm_prot_t access_type, int flags)
345 {
346 	struct vm_map_entry *entry = ufi->entry;
347 	struct uvm_object *uobj = entry->object.uvm_obj;
348 	struct uvm_device *udv = (struct uvm_device *)uobj;
349 	vaddr_t curr_va;
350 	off_t curr_offset;
351 	paddr_t paddr;
352 	int lcv, retval;
353 	dev_t device;
354 	paddr_t (*mapfn)(dev_t, off_t, int);
355 	vm_prot_t mapprot;
356 
357 	/*
358 	 * we do not allow device mappings to be mapped copy-on-write
359 	 * so we kill any attempt to do so here.
360 	 */
361 
362 	if (UVM_ET_ISCOPYONWRITE(entry)) {
363 		uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj, NULL);
364 		return(VM_PAGER_ERROR);
365 	}
366 
367 	/*
368 	 * get device map function.
369 	 */
370 
371 	device = udv->u_device;
372 	mapfn = cdevsw[major(device)].d_mmap;
373 
374 	/*
375 	 * now we must determine the offset in udv to use and the VA to
376 	 * use for pmap_enter.  note that we always use orig_map's pmap
377 	 * for pmap_enter (even if we have a submap).   since virtual
378 	 * addresses in a submap must match the main map, this is ok.
379 	 */
380 
381 	/* udv offset = (offset from start of entry) + entry's offset */
382 	curr_offset = entry->offset + (vaddr - entry->start);
383 	/* pmap va = vaddr (virtual address of pps[0]) */
384 	curr_va = vaddr;
385 
386 	/*
387 	 * loop over the page range entering in as needed
388 	 */
389 
390 	retval = VM_PAGER_OK;
391 	for (lcv = 0 ; lcv < npages ; lcv++, curr_offset += PAGE_SIZE,
392 	    curr_va += PAGE_SIZE) {
393 		if ((flags & PGO_ALLPAGES) == 0 && lcv != centeridx)
394 			continue;
395 
396 		if (pps[lcv] == PGO_DONTCARE)
397 			continue;
398 
399 		paddr = (*mapfn)(device, curr_offset, access_type);
400 		if (paddr == -1) {
401 			retval = VM_PAGER_ERROR;
402 			break;
403 		}
404 		mapprot = ufi->entry->protection;
405 		if (pmap_enter(ufi->orig_map->pmap, curr_va, paddr,
406 		    mapprot, PMAP_CANFAIL | mapprot) != 0) {
407 			/*
408 			 * pmap_enter() didn't have the resource to
409 			 * enter this mapping.  Unlock everything,
410 			 * wait for the pagedaemon to free up some
411 			 * pages, and then tell uvm_fault() to start
412 			 * the fault again.
413 			 *
414 			 * XXX Needs some rethinking for the PGO_ALLPAGES
415 			 * XXX case.
416 			 */
417 			uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap,
418 			    uobj, NULL);
419 
420 			/* sync what we have so far */
421 			pmap_update(ufi->orig_map->pmap);
422 			uvm_wait("udv_fault");
423 			return (VM_PAGER_REFAULT);
424 		}
425 	}
426 
427 	uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj, NULL);
428 	pmap_update(ufi->orig_map->pmap);
429 	return (retval);
430 }
431