xref: /netbsd-src/sys/uvm/uvm_device.c (revision 220b5c059a84c51ea44107ea8951a57ffaecdc8c)
1 /*	$NetBSD: uvm_device.c,v 1.39 2001/11/10 07:36:59 lukem Exp $	*/
2 
3 /*
4  *
5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *      This product includes software developed by Charles D. Cranor and
19  *      Washington University.
20  * 4. The name of the author may not be used to endorse or promote products
21  *    derived from this software without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
24  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
25  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
26  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
27  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
28  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
32  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33  *
34  * from: Id: uvm_device.c,v 1.1.2.9 1998/02/06 05:11:47 chs Exp
35  */
36 
37 /*
38  * uvm_device.c: the device pager.
39  */
40 
41 #include <sys/cdefs.h>
42 __KERNEL_RCSID(0, "$NetBSD: uvm_device.c,v 1.39 2001/11/10 07:36:59 lukem Exp $");
43 
44 #include "opt_uvmhist.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/conf.h>
49 #include <sys/proc.h>
50 #include <sys/malloc.h>
51 #include <sys/vnode.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 struct simplelock 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     struct vm_page **, int, int, vm_fault_t, vm_prot_t, int));
75 
76 /*
77  * master pager structure
78  */
79 
80 struct uvm_pagerops uvm_deviceops = {
81 	udv_init,
82 	udv_reference,
83 	udv_detach,
84 	udv_fault,
85 };
86 
87 /*
88  * the ops!
89  */
90 
91 /*
92  * udv_init
93  *
94  * init pager private data structures.
95  */
96 
97 static void
98 udv_init(void)
99 {
100 	LIST_INIT(&udv_list);
101 	simple_lock_init(&udv_lock);
102 }
103 
104 /*
105  * udv_attach
106  *
107  * get a VM object that is associated with a device.   allocate a new
108  * one if needed.
109  *
110  * => caller must _not_ already be holding the lock on the uvm_object.
111  * => in fact, nothing should be locked so that we can sleep here.
112  */
113 
114 struct uvm_object *
115 udv_attach(arg, accessprot, off, size)
116 	void *arg;
117 	vm_prot_t accessprot;
118 	voff_t off;			/* used only for access check */
119 	vsize_t size;			/* used only for access check */
120 {
121 	dev_t device = *((dev_t *)arg);
122 	struct uvm_device *udv, *lcv;
123 	paddr_t (*mapfn) __P((dev_t, off_t, int));
124 	UVMHIST_FUNC("udv_attach"); UVMHIST_CALLED(maphist);
125 
126 	UVMHIST_LOG(maphist, "(device=0x%x)", device,0,0,0);
127 
128 	/*
129 	 * before we do anything, ensure this device supports mmap
130 	 */
131 
132 	mapfn = cdevsw[major(device)].d_mmap;
133 	if (mapfn == NULL ||
134 	    mapfn == (paddr_t (*) __P((dev_t, off_t, int))) enodev ||
135 	    mapfn == (paddr_t (*) __P((dev_t, off_t, int))) nullop)
136 		return(NULL);
137 
138 	/*
139 	 * Negative offsets on the object are not allowed.
140 	 */
141 
142 	if (off < 0)
143 		return(NULL);
144 
145 	/*
146 	 * Check that the specified range of the device allows the
147 	 * desired protection.
148 	 *
149 	 * XXX assumes VM_PROT_* == PROT_*
150 	 * XXX clobbers off and size, but nothing else here needs them.
151 	 */
152 
153 	while (size != 0) {
154 		if ((*mapfn)(device, off, accessprot) == -1)
155 			return (NULL);
156 		off += PAGE_SIZE; size -= PAGE_SIZE;
157 	}
158 
159 	/*
160 	 * keep looping until we get it
161 	 */
162 
163 	for (;;) {
164 
165 		/*
166 		 * first, attempt to find it on the main list
167 		 */
168 
169 		simple_lock(&udv_lock);
170 		LIST_FOREACH(lcv, &udv_list, u_list) {
171 			if (device == lcv->u_device)
172 				break;
173 		}
174 
175 		/*
176 		 * got it on main list.  put a hold on it and unlock udv_lock.
177 		 */
178 
179 		if (lcv) {
180 
181 			/*
182 			 * if someone else has a hold on it, sleep and start
183 			 * over again.
184 			 */
185 
186 			if (lcv->u_flags & UVM_DEVICE_HOLD) {
187 				lcv->u_flags |= UVM_DEVICE_WANTED;
188 				UVM_UNLOCK_AND_WAIT(lcv, &udv_lock, FALSE,
189 				    "udv_attach",0);
190 				continue;
191 			}
192 
193 			/* we are now holding it */
194 			lcv->u_flags |= UVM_DEVICE_HOLD;
195 			simple_unlock(&udv_lock);
196 
197 			/*
198 			 * bump reference count, unhold, return.
199 			 */
200 
201 			simple_lock(&lcv->u_obj.vmobjlock);
202 			lcv->u_obj.uo_refs++;
203 			simple_unlock(&lcv->u_obj.vmobjlock);
204 
205 			simple_lock(&udv_lock);
206 			if (lcv->u_flags & UVM_DEVICE_WANTED)
207 				wakeup(lcv);
208 			lcv->u_flags &= ~(UVM_DEVICE_WANTED|UVM_DEVICE_HOLD);
209 			simple_unlock(&udv_lock);
210 			return(&lcv->u_obj);
211 		}
212 
213 		/*
214 		 * did not find it on main list.   need to malloc a new one.
215 		 */
216 
217 		simple_unlock(&udv_lock);
218 		/* NOTE: we could sleep in the following malloc() */
219 		MALLOC(udv, struct uvm_device *, sizeof(*udv), M_TEMP,
220 		       M_WAITOK);
221 		simple_lock(&udv_lock);
222 
223 		/*
224 		 * now we have to double check to make sure no one added it
225 		 * to the list while we were sleeping...
226 		 */
227 
228 		LIST_FOREACH(lcv, &udv_list, u_list) {
229 			if (device == lcv->u_device)
230 				break;
231 		}
232 
233 		/*
234 		 * did we lose a race to someone else?
235 		 * free our memory and retry.
236 		 */
237 
238 		if (lcv) {
239 			simple_unlock(&udv_lock);
240 			FREE(udv, M_TEMP);
241 			continue;
242 		}
243 
244 		/*
245 		 * we have it!   init the data structures, add to list
246 		 * and return.
247 		 */
248 
249 		simple_lock_init(&udv->u_obj.vmobjlock);
250 		udv->u_obj.pgops = &uvm_deviceops;
251 		TAILQ_INIT(&udv->u_obj.memq);
252 		udv->u_obj.uo_npages = 0;
253 		udv->u_obj.uo_refs = 1;
254 		udv->u_flags = 0;
255 		udv->u_device = device;
256 		LIST_INSERT_HEAD(&udv_list, udv, u_list);
257 		simple_unlock(&udv_lock);
258 		return(&udv->u_obj);
259 	}
260 	/*NOTREACHED*/
261 }
262 
263 /*
264  * udv_reference
265  *
266  * add a reference to a VM object.   Note that the reference count must
267  * already be one (the passed in reference) so there is no chance of the
268  * udv being released or locked out here.
269  *
270  * => caller must call with object unlocked.
271  */
272 
273 static void
274 udv_reference(uobj)
275 	struct uvm_object *uobj;
276 {
277 	UVMHIST_FUNC("udv_reference"); UVMHIST_CALLED(maphist);
278 
279 	simple_lock(&uobj->vmobjlock);
280 	uobj->uo_refs++;
281 	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)",
282 		    uobj, uobj->uo_refs,0,0);
283 	simple_unlock(&uobj->vmobjlock);
284 }
285 
286 /*
287  * udv_detach
288  *
289  * remove a reference to a VM object.
290  *
291  * => caller must call with object unlocked and map locked.
292  */
293 
294 static void
295 udv_detach(uobj)
296 	struct uvm_object *uobj;
297 {
298 	struct uvm_device *udv = (struct uvm_device *)uobj;
299 	UVMHIST_FUNC("udv_detach"); UVMHIST_CALLED(maphist);
300 
301 	/*
302 	 * loop until done
303 	 */
304 again:
305 	simple_lock(&uobj->vmobjlock);
306 	if (uobj->uo_refs > 1) {
307 		uobj->uo_refs--;
308 		simple_unlock(&uobj->vmobjlock);
309 		UVMHIST_LOG(maphist," <- done, uobj=0x%x, ref=%d",
310 			  uobj,uobj->uo_refs,0,0);
311 		return;
312 	}
313 
314 	/*
315 	 * is it being held?   if so, wait until others are done.
316 	 */
317 
318 	simple_lock(&udv_lock);
319 	if (udv->u_flags & UVM_DEVICE_HOLD) {
320 		udv->u_flags |= UVM_DEVICE_WANTED;
321 		simple_unlock(&uobj->vmobjlock);
322 		UVM_UNLOCK_AND_WAIT(udv, &udv_lock, FALSE, "udv_detach",0);
323 		goto again;
324 	}
325 
326 	/*
327 	 * got it!   nuke it now.
328 	 */
329 
330 	LIST_REMOVE(udv, u_list);
331 	if (udv->u_flags & UVM_DEVICE_WANTED)
332 		wakeup(udv);
333 	simple_unlock(&udv_lock);
334 	simple_unlock(&uobj->vmobjlock);
335 	FREE(udv, M_TEMP);
336 	UVMHIST_LOG(maphist," <- done, freed uobj=0x%x", uobj,0,0,0);
337 }
338 
339 /*
340  * udv_fault: non-standard fault routine for device "pages"
341  *
342  * => rather than having a "get" function, we have a fault routine
343  *	since we don't return vm_pages we need full control over the
344  *	pmap_enter map in
345  * => all the usual fault data structured are locked by the caller
346  *	(i.e. maps(read), amap (if any), uobj)
347  * => on return, we unlock all fault data structures
348  * => flags: PGO_ALLPAGES: get all of the pages
349  *	     PGO_LOCKED: fault data structures are locked
350  *    XXX: currently PGO_LOCKED is always required ... consider removing
351  *	it as a flag
352  * => NOTE: vaddr is the VA of pps[0] in ufi->entry, _NOT_ pps[centeridx]
353  */
354 
355 static int
356 udv_fault(ufi, vaddr, pps, npages, centeridx, fault_type, access_type, flags)
357 	struct uvm_faultinfo *ufi;
358 	vaddr_t vaddr;
359 	struct vm_page **pps;
360 	int npages, centeridx, flags;
361 	vm_fault_t fault_type;
362 	vm_prot_t access_type;
363 {
364 	struct vm_map_entry *entry = ufi->entry;
365 	struct uvm_object *uobj = entry->object.uvm_obj;
366 	struct uvm_device *udv = (struct uvm_device *)uobj;
367 	vaddr_t curr_va;
368 	off_t curr_offset;
369 	paddr_t paddr, mdpgno;
370 	int lcv, retval;
371 	dev_t device;
372 	paddr_t (*mapfn) __P((dev_t, off_t, int));
373 	vm_prot_t mapprot;
374 	UVMHIST_FUNC("udv_fault"); UVMHIST_CALLED(maphist);
375 	UVMHIST_LOG(maphist,"  flags=%d", flags,0,0,0);
376 
377 	/*
378 	 * we do not allow device mappings to be mapped copy-on-write
379 	 * so we kill any attempt to do so here.
380 	 */
381 
382 	if (UVM_ET_ISCOPYONWRITE(entry)) {
383 		UVMHIST_LOG(maphist, "<- failed -- COW entry (etype=0x%x)",
384 		entry->etype, 0,0,0);
385 		uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj, NULL);
386 		return(EIO);
387 	}
388 
389 	/*
390 	 * get device map function.
391 	 */
392 
393 	device = udv->u_device;
394 	mapfn = cdevsw[major(device)].d_mmap;
395 
396 	/*
397 	 * now we must determine the offset in udv to use and the VA to
398 	 * use for pmap_enter.  note that we always use orig_map's pmap
399 	 * for pmap_enter (even if we have a submap).   since virtual
400 	 * addresses in a submap must match the main map, this is ok.
401 	 */
402 
403 	/* udv offset = (offset from start of entry) + entry's offset */
404 	curr_offset = entry->offset + (vaddr - entry->start);
405 	/* pmap va = vaddr (virtual address of pps[0]) */
406 	curr_va = vaddr;
407 
408 	/*
409 	 * loop over the page range entering in as needed
410 	 */
411 
412 	retval = 0;
413 	for (lcv = 0 ; lcv < npages ; lcv++, curr_offset += PAGE_SIZE,
414 	    curr_va += PAGE_SIZE) {
415 		if ((flags & PGO_ALLPAGES) == 0 && lcv != centeridx)
416 			continue;
417 
418 		if (pps[lcv] == PGO_DONTCARE)
419 			continue;
420 
421 		mdpgno = (*mapfn)(device, curr_offset, access_type);
422 		if (mdpgno == -1) {
423 			retval = EIO;
424 			break;
425 		}
426 		paddr = pmap_phys_address(mdpgno);
427 		mapprot = ufi->entry->protection;
428 		UVMHIST_LOG(maphist,
429 		    "  MAPPING: device: pm=0x%x, va=0x%x, pa=0x%lx, at=%d",
430 		    ufi->orig_map->pmap, curr_va, paddr, mapprot);
431 		if (pmap_enter(ufi->orig_map->pmap, curr_va, paddr,
432 		    mapprot, PMAP_CANFAIL | mapprot) != 0) {
433 			/*
434 			 * pmap_enter() didn't have the resource to
435 			 * enter this mapping.  Unlock everything,
436 			 * wait for the pagedaemon to free up some
437 			 * pages, and then tell uvm_fault() to start
438 			 * the fault again.
439 			 *
440 			 * XXX Needs some rethinking for the PGO_ALLPAGES
441 			 * XXX case.
442 			 */
443 			uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap,
444 			    uobj, NULL);
445 			pmap_update(ufi->orig_map->pmap);	/* sync what we have so far */
446 			uvm_wait("udv_fault");
447 			return (ERESTART);
448 		}
449 	}
450 
451 	uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj, NULL);
452 	pmap_update(ufi->orig_map->pmap);
453 	return (retval);
454 }
455