xref: /netbsd-src/sys/uvm/uvm_device.c (revision e5548b402ae4c44fb816de42c7bba9581ce23ef5)
1 /*	$NetBSD: uvm_device.c,v 1.46 2005/12/11 12:25:29 christos 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 1.46 2005/12/11 12:25:29 christos 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(void);
71 static void	udv_reference(struct uvm_object *);
72 static void	udv_detach(struct uvm_object *);
73 static int	udv_fault(struct uvm_faultinfo *, vaddr_t,
74 			  struct vm_page **, int, int, vm_fault_t, vm_prot_t,
75 			  int);
76 
77 /*
78  * master pager structure
79  */
80 
81 struct uvm_pagerops uvm_deviceops = {
82 	udv_init,
83 	udv_reference,
84 	udv_detach,
85 	udv_fault,
86 };
87 
88 /*
89  * the ops!
90  */
91 
92 /*
93  * udv_init
94  *
95  * init pager private data structures.
96  */
97 
98 static void
99 udv_init(void)
100 {
101 	LIST_INIT(&udv_list);
102 	simple_lock_init(&udv_lock);
103 }
104 
105 /*
106  * udv_attach
107  *
108  * get a VM object that is associated with a device.   allocate a new
109  * one if needed.
110  *
111  * => caller must _not_ already be holding the lock on the uvm_object.
112  * => in fact, nothing should be locked so that we can sleep here.
113  */
114 
115 struct uvm_object *
116 udv_attach(void *arg, vm_prot_t accessprot,
117     voff_t off,		/* used only for access check */
118     vsize_t size	/* used only for access check */)
119 {
120 	dev_t device = *((dev_t *)arg);
121 	struct uvm_device *udv, *lcv;
122 	const struct cdevsw *cdev;
123 	dev_type_mmap((*mapfn));
124 
125 	UVMHIST_FUNC("udv_attach"); UVMHIST_CALLED(maphist);
126 
127 	UVMHIST_LOG(maphist, "(device=0x%x)", device,0,0,0);
128 
129 	/*
130 	 * before we do anything, ensure this device supports mmap
131 	 */
132 
133 	cdev = cdevsw_lookup(device);
134 	if (cdev == NULL)
135 		return (NULL);
136 	mapfn = cdev->d_mmap;
137 	if (mapfn == NULL || mapfn == nommap || mapfn == nullmmap)
138 		return(NULL);
139 
140 	/*
141 	 * Negative offsets on the object are not allowed.
142 	 */
143 
144 	if (off < 0)
145 		return(NULL);
146 
147 	/*
148 	 * Check that the specified range of the device allows the
149 	 * desired protection.
150 	 *
151 	 * XXX assumes VM_PROT_* == PROT_*
152 	 * XXX clobbers off and size, but nothing else here needs them.
153 	 */
154 
155 	while (size != 0) {
156 		if ((*mapfn)(device, off, accessprot) == -1)
157 			return (NULL);
158 		off += PAGE_SIZE; size -= PAGE_SIZE;
159 	}
160 
161 	/*
162 	 * keep looping until we get it
163 	 */
164 
165 	for (;;) {
166 
167 		/*
168 		 * first, attempt to find it on the main list
169 		 */
170 
171 		simple_lock(&udv_lock);
172 		LIST_FOREACH(lcv, &udv_list, u_list) {
173 			if (device == lcv->u_device)
174 				break;
175 		}
176 
177 		/*
178 		 * got it on main list.  put a hold on it and unlock udv_lock.
179 		 */
180 
181 		if (lcv) {
182 
183 			/*
184 			 * if someone else has a hold on it, sleep and start
185 			 * over again.
186 			 */
187 
188 			if (lcv->u_flags & UVM_DEVICE_HOLD) {
189 				lcv->u_flags |= UVM_DEVICE_WANTED;
190 				UVM_UNLOCK_AND_WAIT(lcv, &udv_lock, FALSE,
191 				    "udv_attach",0);
192 				continue;
193 			}
194 
195 			/* we are now holding it */
196 			lcv->u_flags |= UVM_DEVICE_HOLD;
197 			simple_unlock(&udv_lock);
198 
199 			/*
200 			 * bump reference count, unhold, return.
201 			 */
202 
203 			simple_lock(&lcv->u_obj.vmobjlock);
204 			lcv->u_obj.uo_refs++;
205 			simple_unlock(&lcv->u_obj.vmobjlock);
206 
207 			simple_lock(&udv_lock);
208 			if (lcv->u_flags & UVM_DEVICE_WANTED)
209 				wakeup(lcv);
210 			lcv->u_flags &= ~(UVM_DEVICE_WANTED|UVM_DEVICE_HOLD);
211 			simple_unlock(&udv_lock);
212 			return(&lcv->u_obj);
213 		}
214 
215 		/*
216 		 * did not find it on main list.   need to malloc a new one.
217 		 */
218 
219 		simple_unlock(&udv_lock);
220 		/* NOTE: we could sleep in the following malloc() */
221 		MALLOC(udv, struct uvm_device *, sizeof(*udv), M_TEMP,
222 		       M_WAITOK);
223 		simple_lock(&udv_lock);
224 
225 		/*
226 		 * now we have to double check to make sure no one added it
227 		 * to the list while we were sleeping...
228 		 */
229 
230 		LIST_FOREACH(lcv, &udv_list, u_list) {
231 			if (device == lcv->u_device)
232 				break;
233 		}
234 
235 		/*
236 		 * did we lose a race to someone else?
237 		 * free our memory and retry.
238 		 */
239 
240 		if (lcv) {
241 			simple_unlock(&udv_lock);
242 			FREE(udv, M_TEMP);
243 			continue;
244 		}
245 
246 		/*
247 		 * we have it!   init the data structures, add to list
248 		 * and return.
249 		 */
250 
251 		UVM_OBJ_INIT(&udv->u_obj, &uvm_deviceops, 1);
252 		udv->u_flags = 0;
253 		udv->u_device = device;
254 		LIST_INSERT_HEAD(&udv_list, udv, u_list);
255 		simple_unlock(&udv_lock);
256 		return(&udv->u_obj);
257 	}
258 	/*NOTREACHED*/
259 }
260 
261 /*
262  * udv_reference
263  *
264  * add a reference to a VM object.   Note that the reference count must
265  * already be one (the passed in reference) so there is no chance of the
266  * udv being released or locked out here.
267  *
268  * => caller must call with object unlocked.
269  */
270 
271 static void
272 udv_reference(struct uvm_object *uobj)
273 {
274 	UVMHIST_FUNC("udv_reference"); UVMHIST_CALLED(maphist);
275 
276 	simple_lock(&uobj->vmobjlock);
277 	uobj->uo_refs++;
278 	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)",
279 		    uobj, uobj->uo_refs,0,0);
280 	simple_unlock(&uobj->vmobjlock);
281 }
282 
283 /*
284  * udv_detach
285  *
286  * remove a reference to a VM object.
287  *
288  * => caller must call with object unlocked and map locked.
289  */
290 
291 static void
292 udv_detach(struct uvm_object *uobj)
293 {
294 	struct uvm_device *udv = (struct uvm_device *)uobj;
295 	UVMHIST_FUNC("udv_detach"); UVMHIST_CALLED(maphist);
296 
297 	/*
298 	 * loop until done
299 	 */
300 again:
301 	simple_lock(&uobj->vmobjlock);
302 	if (uobj->uo_refs > 1) {
303 		uobj->uo_refs--;
304 		simple_unlock(&uobj->vmobjlock);
305 		UVMHIST_LOG(maphist," <- done, uobj=0x%x, ref=%d",
306 			  uobj,uobj->uo_refs,0,0);
307 		return;
308 	}
309 
310 	/*
311 	 * is it being held?   if so, wait until others are done.
312 	 */
313 
314 	simple_lock(&udv_lock);
315 	if (udv->u_flags & UVM_DEVICE_HOLD) {
316 		udv->u_flags |= UVM_DEVICE_WANTED;
317 		simple_unlock(&uobj->vmobjlock);
318 		UVM_UNLOCK_AND_WAIT(udv, &udv_lock, FALSE, "udv_detach",0);
319 		goto again;
320 	}
321 
322 	/*
323 	 * got it!   nuke it now.
324 	 */
325 
326 	LIST_REMOVE(udv, u_list);
327 	if (udv->u_flags & UVM_DEVICE_WANTED)
328 		wakeup(udv);
329 	simple_unlock(&udv_lock);
330 	simple_unlock(&uobj->vmobjlock);
331 	FREE(udv, M_TEMP);
332 	UVMHIST_LOG(maphist," <- done, freed uobj=0x%x", uobj,0,0,0);
333 }
334 
335 /*
336  * udv_fault: non-standard fault routine for device "pages"
337  *
338  * => rather than having a "get" function, we have a fault routine
339  *	since we don't return vm_pages we need full control over the
340  *	pmap_enter map in
341  * => all the usual fault data structured are locked by the caller
342  *	(i.e. maps(read), amap (if any), uobj)
343  * => on return, we unlock all fault data structures
344  * => flags: PGO_ALLPAGES: get all of the pages
345  *	     PGO_LOCKED: fault data structures are locked
346  *    XXX: currently PGO_LOCKED is always required ... consider removing
347  *	it as a flag
348  * => NOTE: vaddr is the VA of pps[0] in ufi->entry, _NOT_ pps[centeridx]
349  */
350 
351 static int
352 udv_fault(struct uvm_faultinfo *ufi, vaddr_t vaddr, struct vm_page **pps,
353     int npages, int centeridx, vm_fault_t fault_type, vm_prot_t access_type,
354     int flags)
355 {
356 	struct vm_map_entry *entry = ufi->entry;
357 	struct uvm_object *uobj = entry->object.uvm_obj;
358 	struct uvm_device *udv = (struct uvm_device *)uobj;
359 	const struct cdevsw *cdev;
360 	vaddr_t curr_va;
361 	off_t curr_offset;
362 	paddr_t paddr, mdpgno;
363 	int lcv, retval;
364 	dev_t device;
365 	paddr_t (*mapfn)(dev_t, off_t, int);
366 	vm_prot_t mapprot;
367 	UVMHIST_FUNC("udv_fault"); UVMHIST_CALLED(maphist);
368 	UVMHIST_LOG(maphist,"  flags=%d", flags,0,0,0);
369 
370 	/*
371 	 * we do not allow device mappings to be mapped copy-on-write
372 	 * so we kill any attempt to do so here.
373 	 */
374 
375 	if (UVM_ET_ISCOPYONWRITE(entry)) {
376 		UVMHIST_LOG(maphist, "<- failed -- COW entry (etype=0x%x)",
377 		entry->etype, 0,0,0);
378 		uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj, NULL);
379 		return(EIO);
380 	}
381 
382 	/*
383 	 * get device map function.
384 	 */
385 
386 	device = udv->u_device;
387 	cdev = cdevsw_lookup(device);
388 	if (cdev == NULL) {
389 		uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj, NULL);
390 		return (EIO);
391 	}
392 	mapfn = cdev->d_mmap;
393 
394 	/*
395 	 * now we must determine the offset in udv to use and the VA to
396 	 * use for pmap_enter.  note that we always use orig_map's pmap
397 	 * for pmap_enter (even if we have a submap).   since virtual
398 	 * addresses in a submap must match the main map, this is ok.
399 	 */
400 
401 	/* udv offset = (offset from start of entry) + entry's offset */
402 	curr_offset = entry->offset + (vaddr - entry->start);
403 	/* pmap va = vaddr (virtual address of pps[0]) */
404 	curr_va = vaddr;
405 
406 	/*
407 	 * loop over the page range entering in as needed
408 	 */
409 
410 	retval = 0;
411 	for (lcv = 0 ; lcv < npages ; lcv++, curr_offset += PAGE_SIZE,
412 	    curr_va += PAGE_SIZE) {
413 		if ((flags & PGO_ALLPAGES) == 0 && lcv != centeridx)
414 			continue;
415 
416 		if (pps[lcv] == PGO_DONTCARE)
417 			continue;
418 
419 		mdpgno = (*mapfn)(device, curr_offset, access_type);
420 		if (mdpgno == -1) {
421 			retval = EIO;
422 			break;
423 		}
424 		paddr = pmap_phys_address(mdpgno);
425 		mapprot = ufi->entry->protection;
426 		UVMHIST_LOG(maphist,
427 		    "  MAPPING: device: pm=0x%x, va=0x%x, pa=0x%lx, at=%d",
428 		    ufi->orig_map->pmap, curr_va, paddr, mapprot);
429 		if (pmap_enter(ufi->orig_map->pmap, curr_va, paddr,
430 		    mapprot, PMAP_CANFAIL | mapprot) != 0) {
431 			/*
432 			 * pmap_enter() didn't have the resource to
433 			 * enter this mapping.  Unlock everything,
434 			 * wait for the pagedaemon to free up some
435 			 * pages, and then tell uvm_fault() to start
436 			 * the fault again.
437 			 *
438 			 * XXX Needs some rethinking for the PGO_ALLPAGES
439 			 * XXX case.
440 			 */
441 			uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap,
442 			    uobj, NULL);
443 			pmap_update(ufi->orig_map->pmap);	/* sync what we have so far */
444 			uvm_wait("udv_fault");
445 			return (ERESTART);
446 		}
447 	}
448 
449 	uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj, NULL);
450 	pmap_update(ufi->orig_map->pmap);
451 	return (retval);
452 }
453