xref: /netbsd/sys/uvm/uvm_device.c (revision bf9ec67e)
1 /*	$NetBSD: uvm_device.c,v 1.40 2002/02/28 21:00:23 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 1.40 2002/02/28 21:00:23 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 __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 	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 	mapfn = cdevsw[major(device)].d_mmap;
134 	if (mapfn == NULL ||
135 	    mapfn == (dev_type_mmap((*))) enodev ||
136 	    mapfn == (dev_type_mmap((*))) nullop)
137 		return(NULL);
138 
139 	/*
140 	 * Negative offsets on the object are not allowed.
141 	 */
142 
143 	if (off < 0)
144 		return(NULL);
145 
146 	/*
147 	 * Check that the specified range of the device allows the
148 	 * desired protection.
149 	 *
150 	 * XXX assumes VM_PROT_* == PROT_*
151 	 * XXX clobbers off and size, but nothing else here needs them.
152 	 */
153 
154 	while (size != 0) {
155 		if ((*mapfn)(device, off, accessprot) == -1)
156 			return (NULL);
157 		off += PAGE_SIZE; size -= PAGE_SIZE;
158 	}
159 
160 	/*
161 	 * keep looping until we get it
162 	 */
163 
164 	for (;;) {
165 
166 		/*
167 		 * first, attempt to find it on the main list
168 		 */
169 
170 		simple_lock(&udv_lock);
171 		LIST_FOREACH(lcv, &udv_list, u_list) {
172 			if (device == lcv->u_device)
173 				break;
174 		}
175 
176 		/*
177 		 * got it on main list.  put a hold on it and unlock udv_lock.
178 		 */
179 
180 		if (lcv) {
181 
182 			/*
183 			 * if someone else has a hold on it, sleep and start
184 			 * over again.
185 			 */
186 
187 			if (lcv->u_flags & UVM_DEVICE_HOLD) {
188 				lcv->u_flags |= UVM_DEVICE_WANTED;
189 				UVM_UNLOCK_AND_WAIT(lcv, &udv_lock, FALSE,
190 				    "udv_attach",0);
191 				continue;
192 			}
193 
194 			/* we are now holding it */
195 			lcv->u_flags |= UVM_DEVICE_HOLD;
196 			simple_unlock(&udv_lock);
197 
198 			/*
199 			 * bump reference count, unhold, return.
200 			 */
201 
202 			simple_lock(&lcv->u_obj.vmobjlock);
203 			lcv->u_obj.uo_refs++;
204 			simple_unlock(&lcv->u_obj.vmobjlock);
205 
206 			simple_lock(&udv_lock);
207 			if (lcv->u_flags & UVM_DEVICE_WANTED)
208 				wakeup(lcv);
209 			lcv->u_flags &= ~(UVM_DEVICE_WANTED|UVM_DEVICE_HOLD);
210 			simple_unlock(&udv_lock);
211 			return(&lcv->u_obj);
212 		}
213 
214 		/*
215 		 * did not find it on main list.   need to malloc a new one.
216 		 */
217 
218 		simple_unlock(&udv_lock);
219 		/* NOTE: we could sleep in the following malloc() */
220 		MALLOC(udv, struct uvm_device *, sizeof(*udv), M_TEMP,
221 		       M_WAITOK);
222 		simple_lock(&udv_lock);
223 
224 		/*
225 		 * now we have to double check to make sure no one added it
226 		 * to the list while we were sleeping...
227 		 */
228 
229 		LIST_FOREACH(lcv, &udv_list, u_list) {
230 			if (device == lcv->u_device)
231 				break;
232 		}
233 
234 		/*
235 		 * did we lose a race to someone else?
236 		 * free our memory and retry.
237 		 */
238 
239 		if (lcv) {
240 			simple_unlock(&udv_lock);
241 			FREE(udv, M_TEMP);
242 			continue;
243 		}
244 
245 		/*
246 		 * we have it!   init the data structures, add to list
247 		 * and return.
248 		 */
249 
250 		simple_lock_init(&udv->u_obj.vmobjlock);
251 		udv->u_obj.pgops = &uvm_deviceops;
252 		TAILQ_INIT(&udv->u_obj.memq);
253 		udv->u_obj.uo_npages = 0;
254 		udv->u_obj.uo_refs = 1;
255 		udv->u_flags = 0;
256 		udv->u_device = device;
257 		LIST_INSERT_HEAD(&udv_list, udv, u_list);
258 		simple_unlock(&udv_lock);
259 		return(&udv->u_obj);
260 	}
261 	/*NOTREACHED*/
262 }
263 
264 /*
265  * udv_reference
266  *
267  * add a reference to a VM object.   Note that the reference count must
268  * already be one (the passed in reference) so there is no chance of the
269  * udv being released or locked out here.
270  *
271  * => caller must call with object unlocked.
272  */
273 
274 static void
275 udv_reference(uobj)
276 	struct uvm_object *uobj;
277 {
278 	UVMHIST_FUNC("udv_reference"); UVMHIST_CALLED(maphist);
279 
280 	simple_lock(&uobj->vmobjlock);
281 	uobj->uo_refs++;
282 	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)",
283 		    uobj, uobj->uo_refs,0,0);
284 	simple_unlock(&uobj->vmobjlock);
285 }
286 
287 /*
288  * udv_detach
289  *
290  * remove a reference to a VM object.
291  *
292  * => caller must call with object unlocked and map locked.
293  */
294 
295 static void
296 udv_detach(uobj)
297 	struct uvm_object *uobj;
298 {
299 	struct uvm_device *udv = (struct uvm_device *)uobj;
300 	UVMHIST_FUNC("udv_detach"); UVMHIST_CALLED(maphist);
301 
302 	/*
303 	 * loop until done
304 	 */
305 again:
306 	simple_lock(&uobj->vmobjlock);
307 	if (uobj->uo_refs > 1) {
308 		uobj->uo_refs--;
309 		simple_unlock(&uobj->vmobjlock);
310 		UVMHIST_LOG(maphist," <- done, uobj=0x%x, ref=%d",
311 			  uobj,uobj->uo_refs,0,0);
312 		return;
313 	}
314 
315 	/*
316 	 * is it being held?   if so, wait until others are done.
317 	 */
318 
319 	simple_lock(&udv_lock);
320 	if (udv->u_flags & UVM_DEVICE_HOLD) {
321 		udv->u_flags |= UVM_DEVICE_WANTED;
322 		simple_unlock(&uobj->vmobjlock);
323 		UVM_UNLOCK_AND_WAIT(udv, &udv_lock, FALSE, "udv_detach",0);
324 		goto again;
325 	}
326 
327 	/*
328 	 * got it!   nuke it now.
329 	 */
330 
331 	LIST_REMOVE(udv, u_list);
332 	if (udv->u_flags & UVM_DEVICE_WANTED)
333 		wakeup(udv);
334 	simple_unlock(&udv_lock);
335 	simple_unlock(&uobj->vmobjlock);
336 	FREE(udv, M_TEMP);
337 	UVMHIST_LOG(maphist," <- done, freed uobj=0x%x", uobj,0,0,0);
338 }
339 
340 /*
341  * udv_fault: non-standard fault routine for device "pages"
342  *
343  * => rather than having a "get" function, we have a fault routine
344  *	since we don't return vm_pages we need full control over the
345  *	pmap_enter map in
346  * => all the usual fault data structured are locked by the caller
347  *	(i.e. maps(read), amap (if any), uobj)
348  * => on return, we unlock all fault data structures
349  * => flags: PGO_ALLPAGES: get all of the pages
350  *	     PGO_LOCKED: fault data structures are locked
351  *    XXX: currently PGO_LOCKED is always required ... consider removing
352  *	it as a flag
353  * => NOTE: vaddr is the VA of pps[0] in ufi->entry, _NOT_ pps[centeridx]
354  */
355 
356 static int
357 udv_fault(ufi, vaddr, pps, npages, centeridx, fault_type, access_type, flags)
358 	struct uvm_faultinfo *ufi;
359 	vaddr_t vaddr;
360 	struct vm_page **pps;
361 	int npages, centeridx, flags;
362 	vm_fault_t fault_type;
363 	vm_prot_t access_type;
364 {
365 	struct vm_map_entry *entry = ufi->entry;
366 	struct uvm_object *uobj = entry->object.uvm_obj;
367 	struct uvm_device *udv = (struct uvm_device *)uobj;
368 	vaddr_t curr_va;
369 	off_t curr_offset;
370 	paddr_t paddr, mdpgno;
371 	int lcv, retval;
372 	dev_t device;
373 	paddr_t (*mapfn) __P((dev_t, off_t, int));
374 	vm_prot_t mapprot;
375 	UVMHIST_FUNC("udv_fault"); UVMHIST_CALLED(maphist);
376 	UVMHIST_LOG(maphist,"  flags=%d", flags,0,0,0);
377 
378 	/*
379 	 * we do not allow device mappings to be mapped copy-on-write
380 	 * so we kill any attempt to do so here.
381 	 */
382 
383 	if (UVM_ET_ISCOPYONWRITE(entry)) {
384 		UVMHIST_LOG(maphist, "<- failed -- COW entry (etype=0x%x)",
385 		entry->etype, 0,0,0);
386 		uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj, NULL);
387 		return(EIO);
388 	}
389 
390 	/*
391 	 * get device map function.
392 	 */
393 
394 	device = udv->u_device;
395 	mapfn = cdevsw[major(device)].d_mmap;
396 
397 	/*
398 	 * now we must determine the offset in udv to use and the VA to
399 	 * use for pmap_enter.  note that we always use orig_map's pmap
400 	 * for pmap_enter (even if we have a submap).   since virtual
401 	 * addresses in a submap must match the main map, this is ok.
402 	 */
403 
404 	/* udv offset = (offset from start of entry) + entry's offset */
405 	curr_offset = entry->offset + (vaddr - entry->start);
406 	/* pmap va = vaddr (virtual address of pps[0]) */
407 	curr_va = vaddr;
408 
409 	/*
410 	 * loop over the page range entering in as needed
411 	 */
412 
413 	retval = 0;
414 	for (lcv = 0 ; lcv < npages ; lcv++, curr_offset += PAGE_SIZE,
415 	    curr_va += PAGE_SIZE) {
416 		if ((flags & PGO_ALLPAGES) == 0 && lcv != centeridx)
417 			continue;
418 
419 		if (pps[lcv] == PGO_DONTCARE)
420 			continue;
421 
422 		mdpgno = (*mapfn)(device, curr_offset, access_type);
423 		if (mdpgno == -1) {
424 			retval = EIO;
425 			break;
426 		}
427 		paddr = pmap_phys_address(mdpgno);
428 		mapprot = ufi->entry->protection;
429 		UVMHIST_LOG(maphist,
430 		    "  MAPPING: device: pm=0x%x, va=0x%x, pa=0x%lx, at=%d",
431 		    ufi->orig_map->pmap, curr_va, paddr, mapprot);
432 		if (pmap_enter(ufi->orig_map->pmap, curr_va, paddr,
433 		    mapprot, PMAP_CANFAIL | mapprot) != 0) {
434 			/*
435 			 * pmap_enter() didn't have the resource to
436 			 * enter this mapping.  Unlock everything,
437 			 * wait for the pagedaemon to free up some
438 			 * pages, and then tell uvm_fault() to start
439 			 * the fault again.
440 			 *
441 			 * XXX Needs some rethinking for the PGO_ALLPAGES
442 			 * XXX case.
443 			 */
444 			uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap,
445 			    uobj, NULL);
446 			pmap_update(ufi->orig_map->pmap);	/* sync what we have so far */
447 			uvm_wait("udv_fault");
448 			return (ERESTART);
449 		}
450 	}
451 
452 	uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj, NULL);
453 	pmap_update(ufi->orig_map->pmap);
454 	return (retval);
455 }
456