xref: /openbsd-src/sys/uvm/uvm_device.c (revision b2ea75c1b17e1a9a339660e7ed45cd24946b230e)
1 /*	$OpenBSD: uvm_device.c,v 1.11 2001/08/11 10:57:22 art Exp $	*/
2 /*	$NetBSD: uvm_device.c,v 1.22 2000/05/28 10:21:55 drochner 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/vnode.h>
48 
49 #include <vm/vm.h>
50 #include <vm/vm_page.h>
51 #include <vm/vm_kern.h>
52 
53 #include <uvm/uvm.h>
54 #include <uvm/uvm_device.h>
55 
56 /*
57  * private global data structure
58  *
59  * we keep a list of active device objects in the system.
60  */
61 
62 LIST_HEAD(udv_list_struct, uvm_device);
63 static struct udv_list_struct udv_list;
64 static simple_lock_data_t udv_lock;
65 
66 /*
67  * functions
68  */
69 
70 static void		udv_init __P((void));
71 static void             udv_reference __P((struct uvm_object *));
72 static void             udv_detach __P((struct uvm_object *));
73 static int		udv_fault __P((struct uvm_faultinfo *, vaddr_t,
74 				       vm_page_t *, int, int, vm_fault_t,
75 				       vm_prot_t, int));
76 static int		udv_asyncget __P((struct uvm_object *, voff_t,
77 				       int));
78 static boolean_t        udv_flush __P((struct uvm_object *, voff_t, voff_t,
79 				       int));
80 static int		udv_put __P((struct uvm_object *, vm_page_t *,
81 					int, boolean_t));
82 
83 /*
84  * master pager structure
85  */
86 
87 struct uvm_pagerops uvm_deviceops = {
88 	udv_init,
89 	udv_reference,
90 	udv_detach,
91 	udv_fault,
92 	udv_flush,
93 	NULL,		/* no get function since we have udv_fault */
94 	udv_asyncget,
95 	udv_put,
96 	NULL,		/* no cluster function */
97 	NULL,		/* no put cluster function */
98 	NULL,		/* no AIO-DONE function since no async i/o */
99 	NULL,		/* no releasepg function since no normal pages */
100 };
101 
102 /*
103  * the ops!
104  */
105 
106 /*
107  * udv_init
108  *
109  * init pager private data structures.
110  */
111 
112 void
113 udv_init()
114 {
115 
116 	LIST_INIT(&udv_list);
117 	simple_lock_init(&udv_lock);
118 }
119 
120 /*
121  * udv_attach
122  *
123  * get a VM object that is associated with a device.   allocate a new
124  * one if needed.
125  *
126  * => caller must _not_ already be holding the lock on the uvm_object.
127  * => in fact, nothing should be locked so that we can sleep here.
128  */
129 struct uvm_object *
130 udv_attach(arg, accessprot, off, size)
131 	void *arg;
132 	vm_prot_t accessprot;
133 	voff_t off;			/* used only for access check */
134 	vsize_t size;			/* used only for access check */
135 {
136 	dev_t device = *((dev_t *) arg);
137 	struct uvm_device *udv, *lcv;
138 	int (*mapfn) __P((dev_t, int, int));
139 	UVMHIST_FUNC("udv_attach"); UVMHIST_CALLED(maphist);
140 
141 	UVMHIST_LOG(maphist, "(device=0x%x)", device,0,0,0);
142 
143 	/*
144 	 * before we do anything, ensure this device supports mmap
145 	 */
146 
147 	mapfn = cdevsw[major(device)].d_mmap;
148 	if (mapfn == NULL ||
149 			mapfn == (int (*) __P((dev_t, int, int))) enodev ||
150 			mapfn == (int (*) __P((dev_t, int, int))) nullop)
151 		return(NULL);
152 
153 	/*
154 	 * As long as the device d_mmap interface gets an "int"
155 	 * offset, we have to watch out not to overflow its
156 	 * numeric range. (assuming it will be interpreted as
157 	 * "unsigned")
158 	 */
159 	if (((off + size - 1) & (u_int)-1) != off + size - 1)
160 		return (0);
161 
162 	/*
163 	 * Check that the specified range of the device allows the
164 	 * desired protection.
165 	 *
166 	 * XXX assumes VM_PROT_* == PROT_*
167 	 * XXX clobbers off and size, but nothing else here needs them.
168 	 */
169 
170 	while (size != 0) {
171 		if ((*mapfn)(device, off, accessprot) == -1)
172 			return (NULL);
173 		off += PAGE_SIZE; size -= PAGE_SIZE;
174 	}
175 
176 	/*
177 	 * keep looping until we get it
178 	 */
179 
180 	while (1) {
181 
182 		/*
183 		 * first, attempt to find it on the main list
184 		 */
185 
186 		simple_lock(&udv_lock);
187 		for (lcv = udv_list.lh_first ; lcv != NULL ; lcv = lcv->u_list.le_next) {
188 			if (device == lcv->u_device)
189 				break;
190 		}
191 
192 		/*
193 		 * got it on main list.  put a hold on it and unlock udv_lock.
194 		 */
195 
196 		if (lcv) {
197 
198 			/*
199 			 * if someone else has a hold on it, sleep and start
200 			 * over again.
201 			 */
202 
203 			if (lcv->u_flags & UVM_DEVICE_HOLD) {
204 				lcv->u_flags |= UVM_DEVICE_WANTED;
205 				UVM_UNLOCK_AND_WAIT(lcv, &udv_lock, FALSE,
206 				    "udv_attach",0);
207 				continue;
208 			}
209 
210 			/* we are now holding it */
211 			lcv->u_flags |= UVM_DEVICE_HOLD;
212 			simple_unlock(&udv_lock);
213 
214 			/*
215 			 * bump reference count, unhold, return.
216 			 */
217 
218 			simple_lock(&lcv->u_obj.vmobjlock);
219 			lcv->u_obj.uo_refs++;
220 			simple_unlock(&lcv->u_obj.vmobjlock);
221 
222 			simple_lock(&udv_lock);
223 			if (lcv->u_flags & UVM_DEVICE_WANTED)
224 				wakeup(lcv);
225 			lcv->u_flags &= ~(UVM_DEVICE_WANTED|UVM_DEVICE_HOLD);
226 			simple_unlock(&udv_lock);
227 			return(&lcv->u_obj);
228 		}
229 
230 		/*
231 		 * did not find it on main list.   need to malloc a new one.
232 		 */
233 
234 		simple_unlock(&udv_lock);
235 		/* NOTE: we could sleep in the following malloc() */
236 		MALLOC(udv, struct uvm_device *, sizeof(*udv), M_TEMP, M_WAITOK);
237 		simple_lock(&udv_lock);
238 
239 		/*
240 		 * now we have to double check to make sure no one added it
241 		 * to the list while we were sleeping...
242 		 */
243 
244 		for (lcv = udv_list.lh_first ; lcv != NULL ;
245 		    lcv = lcv->u_list.le_next) {
246 			if (device == lcv->u_device)
247 				break;
248 		}
249 
250 		/*
251 		 * did we lose a race to someone else?   free our memory and retry.
252 		 */
253 
254 		if (lcv) {
255 			simple_unlock(&udv_lock);
256 			FREE(udv, M_TEMP);
257 			continue;
258 		}
259 
260 		/*
261 		 * we have it!   init the data structures, add to list
262 		 * and return.
263 		 */
264 
265 		simple_lock_init(&udv->u_obj.vmobjlock);
266 		udv->u_obj.pgops = &uvm_deviceops;
267 		TAILQ_INIT(&udv->u_obj.memq);	/* not used, but be safe */
268 		udv->u_obj.uo_npages = 0;
269 		udv->u_obj.uo_refs = 1;
270 		udv->u_flags = 0;
271 		udv->u_device = device;
272 		LIST_INSERT_HEAD(&udv_list, udv, u_list);
273 		simple_unlock(&udv_lock);
274 
275 		return(&udv->u_obj);
276 
277 	}  /* while(1) loop */
278 
279 	/*NOTREACHED*/
280 }
281 
282 /*
283  * udv_reference
284  *
285  * add a reference to a VM object.   Note that the reference count must
286  * already be one (the passed in reference) so there is no chance of the
287  * udv being released or locked out here.
288  *
289  * => caller must call with object unlocked.
290  */
291 
292 static void
293 udv_reference(uobj)
294 	struct uvm_object *uobj;
295 {
296 	UVMHIST_FUNC("udv_reference"); UVMHIST_CALLED(maphist);
297 
298 	simple_lock(&uobj->vmobjlock);
299 	uobj->uo_refs++;
300 	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)",
301 	uobj, uobj->uo_refs,0,0);
302 	simple_unlock(&uobj->vmobjlock);
303 }
304 
305 /*
306  * udv_detach
307  *
308  * remove a reference to a VM object.
309  *
310  * => caller must call with object unlocked and map locked.
311  */
312 
313 static void
314 udv_detach(uobj)
315 	struct uvm_object *uobj;
316 {
317 	struct uvm_device *udv = (struct uvm_device *) uobj;
318 	UVMHIST_FUNC("udv_detach"); UVMHIST_CALLED(maphist);
319 
320 	/*
321 	 * loop until done
322 	 */
323 
324 	while (1) {
325 		simple_lock(&uobj->vmobjlock);
326 
327 		if (uobj->uo_refs > 1) {
328 			uobj->uo_refs--;			/* drop ref! */
329 			simple_unlock(&uobj->vmobjlock);
330 			UVMHIST_LOG(maphist," <- done, uobj=0x%x, ref=%d",
331 				  uobj,uobj->uo_refs,0,0);
332 			return;
333 		}
334 
335 #ifdef DIAGNOSTIC
336 		if (uobj->uo_npages || uobj->memq.tqh_first)
337 			panic("udv_detach: pages in a device object?");
338 #endif
339 
340 		/*
341 		 * now lock udv_lock
342 		 */
343 		simple_lock(&udv_lock);
344 
345 		/*
346 		 * is it being held?   if so, wait until others are done.
347 		 */
348 		if (udv->u_flags & UVM_DEVICE_HOLD) {
349 
350 			/*
351 			 * want it
352 			 */
353 			udv->u_flags |= UVM_DEVICE_WANTED;
354 			simple_unlock(&uobj->vmobjlock);
355 			UVM_UNLOCK_AND_WAIT(udv, &udv_lock, FALSE, "udv_detach",0);
356 			continue;
357 		}
358 
359 		/*
360 		 * got it!   nuke it now.
361 		 */
362 
363 		LIST_REMOVE(udv, u_list);
364 		if (udv->u_flags & UVM_DEVICE_WANTED)
365 			wakeup(udv);
366 		FREE(udv, M_TEMP);
367 		break;	/* DONE! */
368 
369 	}	/* while (1) loop */
370 
371 	UVMHIST_LOG(maphist," <- done, freed uobj=0x%x", uobj,0,0,0);
372 	return;
373 }
374 
375 
376 /*
377  * udv_flush
378  *
379  * flush pages out of a uvm object.   a no-op for devices.
380  */
381 
382 static boolean_t udv_flush(uobj, start, stop, flags)
383 	struct uvm_object *uobj;
384 	voff_t start, stop;
385 	int flags;
386 {
387 
388 	return(TRUE);
389 }
390 
391 /*
392  * udv_fault: non-standard fault routine for device "pages"
393  *
394  * => rather than having a "get" function, we have a fault routine
395  *	since we don't return vm_pages we need full control over the
396  *	pmap_enter map in
397  * => all the usual fault data structured are locked by the caller
398  *	(i.e. maps(read), amap (if any), uobj)
399  * => on return, we unlock all fault data structures
400  * => flags: PGO_ALLPAGES: get all of the pages
401  *	     PGO_LOCKED: fault data structures are locked
402  *    XXX: currently PGO_LOCKED is always required ... consider removing
403  *	it as a flag
404  * => NOTE: vaddr is the VA of pps[0] in ufi->entry, _NOT_ pps[centeridx]
405  */
406 
407 static int
408 udv_fault(ufi, vaddr, pps, npages, centeridx, fault_type, access_type, flags)
409 	struct uvm_faultinfo *ufi;
410 	vaddr_t vaddr;
411 	vm_page_t *pps;
412 	int npages, centeridx, flags;
413 	vm_fault_t fault_type;
414 	vm_prot_t access_type;
415 {
416 	struct vm_map_entry *entry = ufi->entry;
417 	struct uvm_object *uobj = entry->object.uvm_obj;
418 	struct uvm_device *udv = (struct uvm_device *)uobj;
419 	vaddr_t curr_va;
420 	int curr_offset;
421 	paddr_t paddr;
422 	int lcv, retval, mdpgno;
423 	dev_t device;
424 	int (*mapfn) __P((dev_t, int, int));
425 	vm_prot_t mapprot;
426 	UVMHIST_FUNC("udv_fault"); UVMHIST_CALLED(maphist);
427 	UVMHIST_LOG(maphist,"  flags=%d", flags,0,0,0);
428 
429 	/*
430 	 * XXX: !PGO_LOCKED calls are currently not allowed (or used)
431 	 */
432 
433 	if ((flags & PGO_LOCKED) == 0)
434 		panic("udv_fault: !PGO_LOCKED fault");
435 
436 	/*
437 	 * we do not allow device mappings to be mapped copy-on-write
438 	 * so we kill any attempt to do so here.
439 	 */
440 
441 	if (UVM_ET_ISCOPYONWRITE(entry)) {
442 		UVMHIST_LOG(maphist, "<- failed -- COW entry (etype=0x%x)",
443 		entry->etype, 0,0,0);
444 		uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj, NULL);
445 		return(VM_PAGER_ERROR);
446 	}
447 
448 	/*
449 	 * get device map function.
450 	 */
451 	device = udv->u_device;
452 	mapfn = cdevsw[major(device)].d_mmap;
453 
454 	/*
455 	 * now we must determine the offset in udv to use and the VA to
456 	 * use for pmap_enter.  note that we always use orig_map's pmap
457 	 * for pmap_enter (even if we have a submap).   since virtual
458 	 * addresses in a submap must match the main map, this is ok.
459 	 */
460 	/* udv offset = (offset from start of entry) + entry's offset */
461 	curr_offset = (int)((vaddr - entry->start) + entry->offset);
462 	/* pmap va = vaddr (virtual address of pps[0]) */
463 	curr_va = vaddr;
464 
465 	/*
466 	 * loop over the page range entering in as needed
467 	 */
468 
469 	retval = VM_PAGER_OK;
470 	for (lcv = 0 ; lcv < npages ; lcv++, curr_offset += PAGE_SIZE,
471 	    curr_va += PAGE_SIZE) {
472 		if ((flags & PGO_ALLPAGES) == 0 && lcv != centeridx)
473 			continue;
474 
475 		if (pps[lcv] == PGO_DONTCARE)
476 			continue;
477 
478 		mdpgno = (*mapfn)(device, curr_offset, access_type);
479 		if (mdpgno == -1) {
480 			retval = VM_PAGER_ERROR;
481 			break;
482 		}
483 		paddr = pmap_phys_address(mdpgno);
484 		mapprot = ufi->entry->protection;
485 		UVMHIST_LOG(maphist,
486 		    "  MAPPING: device: pm=0x%x, va=0x%x, pa=0x%x, at=%d",
487 		    ufi->orig_map->pmap, curr_va, (int)paddr, mapprot);
488 		if (pmap_enter(ufi->orig_map->pmap, curr_va, paddr,
489 		    mapprot, PMAP_CANFAIL | mapprot) != KERN_SUCCESS) {
490 			/*
491 			 * pmap_enter() didn't have the resource to
492 			 * enter this mapping.  Unlock everything,
493 			 * wait for the pagedaemon to free up some
494 			 * pages, and then tell uvm_fault() to start
495 			 * the fault again.
496 			 *
497 			 * XXX Needs some rethinking for the PGO_ALLPAGES
498 			 * XXX case.
499 			 */
500 			uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap,
501 			    uobj, NULL);
502 			uvm_wait("udv_fault");
503 			return (VM_PAGER_REFAULT);
504 		}
505 	}
506 
507 	uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj, NULL);
508 	return (retval);
509 }
510 
511 /*
512  * udv_asyncget: start async I/O to bring pages into ram
513  *
514  * => caller must lock object(???XXX: see if this is best)
515  * => a no-op for devices
516  */
517 
518 static int
519 udv_asyncget(uobj, offset, npages)
520 	struct uvm_object *uobj;
521 	voff_t offset;
522 	int npages;
523 {
524 
525 	return(KERN_SUCCESS);
526 }
527 
528 /*
529  * udv_put: flush page data to backing store.
530  *
531  * => this function should never be called (since we never have any
532  *	page structures to "put")
533  */
534 
535 static int
536 udv_put(uobj, pps, npages, flags)
537 	struct uvm_object *uobj;
538 	struct vm_page **pps;
539 	int npages, flags;
540 {
541 
542 	panic("udv_put: trying to page out to a device!");
543 }
544