xref: /netbsd-src/sys/uvm/uvm_device.c (revision 181254a7b1bdde6873432bffef2d2decc4b5c22f)
1 /*	$NetBSD: uvm_device.c,v 1.71 2020/07/09 05:57:15 skrll Exp $	*/
2 
3 /*
4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
21  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
22  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
23  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
25  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26  *
27  * from: Id: uvm_device.c,v 1.1.2.9 1998/02/06 05:11:47 chs Exp
28  */
29 
30 /*
31  * uvm_device.c: the device pager.
32  */
33 
34 #include <sys/cdefs.h>
35 __KERNEL_RCSID(0, "$NetBSD: uvm_device.c,v 1.71 2020/07/09 05:57:15 skrll Exp $");
36 
37 #include "opt_uvmhist.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/conf.h>
42 #include <sys/proc.h>
43 #include <sys/kmem.h>
44 
45 #include <uvm/uvm.h>
46 #include <uvm/uvm_device.h>
47 #include <uvm/uvm_pmap.h>
48 
49 /*
50  * private global data structure
51  *
52  * we keep a list of active device objects in the system.
53  */
54 
55 LIST_HEAD(udv_list_struct, uvm_device);
56 static struct udv_list_struct udv_list;
57 static kmutex_t udv_lock __cacheline_aligned;
58 
59 /*
60  * functions
61  */
62 
63 static void	udv_init(void);
64 static void	udv_reference(struct uvm_object *);
65 static void	udv_detach(struct uvm_object *);
66 static int	udv_fault(struct uvm_faultinfo *, vaddr_t,
67 			  struct vm_page **, int, int, vm_prot_t,
68 			  int);
69 
70 /*
71  * master pager structure
72  */
73 
74 const struct uvm_pagerops uvm_deviceops = {
75 	.pgo_init = udv_init,
76 	.pgo_reference = udv_reference,
77 	.pgo_detach = udv_detach,
78 	.pgo_fault = udv_fault,
79 };
80 
81 /*
82  * the ops!
83  */
84 
85 /*
86  * udv_init
87  *
88  * init pager private data structures.
89  */
90 
91 static void
92 udv_init(void)
93 {
94 	LIST_INIT(&udv_list);
95 	mutex_init(&udv_lock, MUTEX_DEFAULT, IPL_NONE);
96 }
97 
98 /*
99  * udv_attach
100  *
101  * get a VM object that is associated with a device.   allocate a new
102  * one if needed.
103  *
104  * => caller must _not_ already be holding the lock on the uvm_object.
105  * => in fact, nothing should be locked so that we can sleep here.
106  */
107 
108 struct uvm_object *
109 udv_attach(dev_t device, vm_prot_t accessprot,
110     voff_t off,		/* used only for access check */
111     vsize_t size	/* used only for access check */)
112 {
113 	struct uvm_device *udv, *lcv;
114 	const struct cdevsw *cdev;
115 	dev_type_mmap((*mapfn));
116 
117 	UVMHIST_FUNC(__func__);
118 	UVMHIST_CALLARGS(maphist, "(device=%#jx)", device,0,0,0);
119 
120 	/*
121 	 * before we do anything, ensure this device supports mmap
122 	 */
123 
124 	cdev = cdevsw_lookup(device);
125 	if (cdev == NULL) {
126 		return (NULL);
127 	}
128 	mapfn = cdev->d_mmap;
129 	if (mapfn == NULL || mapfn == nommap) {
130 		return(NULL);
131 	}
132 
133 	/*
134 	 * Negative offsets on the object are not allowed.
135 	 */
136 
137 	if ((cdev->d_flag & D_NEGOFFSAFE) == 0 &&
138 	    off != UVM_UNKNOWN_OFFSET && off < 0)
139 		return(NULL);
140 
141 	/*
142 	 * Check that the specified range of the device allows the
143 	 * desired protection.
144 	 *
145 	 * XXX assumes VM_PROT_* == PROT_*
146 	 * XXX clobbers off and size, but nothing else here needs them.
147 	 */
148 
149 	while (size != 0) {
150 		if (cdev_mmap(device, off, accessprot) == -1) {
151 			return (NULL);
152 		}
153 		off += PAGE_SIZE; size -= PAGE_SIZE;
154 	}
155 
156 	/*
157 	 * keep looping until we get it
158 	 */
159 
160 	for (;;) {
161 
162 		/*
163 		 * first, attempt to find it on the main list
164 		 */
165 
166 		mutex_enter(&udv_lock);
167 		LIST_FOREACH(lcv, &udv_list, u_list) {
168 			if (device == lcv->u_device)
169 				break;
170 		}
171 
172 		/*
173 		 * got it on main list.  put a hold on it and unlock udv_lock.
174 		 */
175 
176 		if (lcv) {
177 
178 			/*
179 			 * if someone else has a hold on it, sleep and start
180 			 * over again.
181 			 */
182 
183 			if (lcv->u_flags & UVM_DEVICE_HOLD) {
184 				lcv->u_flags |= UVM_DEVICE_WANTED;
185 				UVM_UNLOCK_AND_WAIT(lcv, &udv_lock, false,
186 				    "udv_attach",0);
187 				continue;
188 			}
189 
190 			/* we are now holding it */
191 			lcv->u_flags |= UVM_DEVICE_HOLD;
192 			mutex_exit(&udv_lock);
193 
194 			/*
195 			 * bump reference count, unhold, return.
196 			 */
197 
198 			rw_enter(lcv->u_obj.vmobjlock, RW_WRITER);
199 			lcv->u_obj.uo_refs++;
200 			rw_exit(lcv->u_obj.vmobjlock);
201 
202 			mutex_enter(&udv_lock);
203 			if (lcv->u_flags & UVM_DEVICE_WANTED)
204 				wakeup(lcv);
205 			lcv->u_flags &= ~(UVM_DEVICE_WANTED|UVM_DEVICE_HOLD);
206 			mutex_exit(&udv_lock);
207 			return(&lcv->u_obj);
208 		}
209 
210 		/*
211 		 * Did not find it on main list.  Need to allocate a new one.
212 		 */
213 
214 		mutex_exit(&udv_lock);
215 
216 		/* Note: both calls may allocate memory and sleep. */
217 		udv = kmem_alloc(sizeof(*udv), KM_SLEEP);
218 		uvm_obj_init(&udv->u_obj, &uvm_deviceops, true, 1);
219 
220 		mutex_enter(&udv_lock);
221 
222 		/*
223 		 * now we have to double check to make sure no one added it
224 		 * to the list while we were sleeping...
225 		 */
226 
227 		LIST_FOREACH(lcv, &udv_list, u_list) {
228 			if (device == lcv->u_device)
229 				break;
230 		}
231 
232 		/*
233 		 * did we lose a race to someone else?
234 		 * free our memory and retry.
235 		 */
236 
237 		if (lcv) {
238 			mutex_exit(&udv_lock);
239 			uvm_obj_destroy(&udv->u_obj, true);
240 			kmem_free(udv, sizeof(*udv));
241 			continue;
242 		}
243 
244 		/*
245 		 * we have it!   init the data structures, add to list
246 		 * and return.
247 		 */
248 
249 		udv->u_flags = 0;
250 		udv->u_device = device;
251 		LIST_INSERT_HEAD(&udv_list, udv, u_list);
252 		mutex_exit(&udv_lock);
253 		return(&udv->u_obj);
254 	}
255 	/*NOTREACHED*/
256 }
257 
258 /*
259  * udv_reference
260  *
261  * add a reference to a VM object.   Note that the reference count must
262  * already be one (the passed in reference) so there is no chance of the
263  * udv being released or locked out here.
264  *
265  * => caller must call with object unlocked.
266  */
267 
268 static void
269 udv_reference(struct uvm_object *uobj)
270 {
271 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
272 
273 	rw_enter(uobj->vmobjlock, RW_WRITER);
274 	uobj->uo_refs++;
275 	UVMHIST_LOG(maphist, "<- done (uobj=%#jx, ref = %jd)",
276 	    (uintptr_t)uobj, uobj->uo_refs,0,0);
277 	rw_exit(uobj->vmobjlock);
278 }
279 
280 /*
281  * udv_detach
282  *
283  * remove a reference to a VM object.
284  *
285  * => caller must call with object unlocked and map locked.
286  */
287 
288 static void
289 udv_detach(struct uvm_object *uobj)
290 {
291 	struct uvm_device *udv = (struct uvm_device *)uobj;
292 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
293 
294 	/*
295 	 * loop until done
296 	 */
297 again:
298 	rw_enter(uobj->vmobjlock, RW_WRITER);
299 	if (uobj->uo_refs > 1) {
300 		uobj->uo_refs--;
301 		rw_exit(uobj->vmobjlock);
302 		UVMHIST_LOG(maphist," <- done, uobj=%#jx, ref=%jd",
303 		    (uintptr_t)uobj,uobj->uo_refs,0,0);
304 		return;
305 	}
306 
307 	/*
308 	 * is it being held?   if so, wait until others are done.
309 	 */
310 
311 	mutex_enter(&udv_lock);
312 	if (udv->u_flags & UVM_DEVICE_HOLD) {
313 		udv->u_flags |= UVM_DEVICE_WANTED;
314 		rw_exit(uobj->vmobjlock);
315 		UVM_UNLOCK_AND_WAIT(udv, &udv_lock, false, "udv_detach",0);
316 		goto again;
317 	}
318 
319 	/*
320 	 * got it!   nuke it now.
321 	 */
322 
323 	LIST_REMOVE(udv, u_list);
324 	if (udv->u_flags & UVM_DEVICE_WANTED)
325 		wakeup(udv);
326 	mutex_exit(&udv_lock);
327 	rw_exit(uobj->vmobjlock);
328 
329 	uvm_obj_destroy(uobj, true);
330 	kmem_free(udv, sizeof(*udv));
331 	UVMHIST_LOG(maphist," <- done, freed uobj=%#jx", (uintptr_t)uobj,
332 	    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_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 	vaddr_t curr_va;
360 	off_t curr_offset;
361 	paddr_t paddr, mdpgno;
362 	u_int mmapflags;
363 	int lcv, retval;
364 	dev_t device;
365 	vm_prot_t mapprot;
366 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
367 	UVMHIST_LOG(maphist,"  flags=%jd", flags,0,0,0);
368 
369 	/*
370 	 * we do not allow device mappings to be mapped copy-on-write
371 	 * so we kill any attempt to do so here.
372 	 */
373 
374 	if (UVM_ET_ISCOPYONWRITE(entry)) {
375 		UVMHIST_LOG(maphist, "<- failed -- COW entry (etype=%#jx)",
376 		    entry->etype, 0,0,0);
377 		uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj);
378 		return(EIO);
379 	}
380 
381 	/*
382 	 * get device map function.
383 	 */
384 
385 	device = udv->u_device;
386 	if (cdevsw_lookup(device) == NULL) {
387 		/* XXX This should not happen */
388 		uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj);
389 		return (EIO);
390 	}
391 
392 	/*
393 	 * now we must determine the offset in udv to use and the VA to
394 	 * use for pmap_enter.  note that we always use orig_map's pmap
395 	 * for pmap_enter (even if we have a submap).   since virtual
396 	 * addresses in a submap must match the main map, this is ok.
397 	 */
398 
399 	/* udv offset = (offset from start of entry) + entry's offset */
400 	curr_offset = entry->offset + (vaddr - entry->start);
401 	/* pmap va = vaddr (virtual address of pps[0]) */
402 	curr_va = vaddr;
403 
404 	/*
405 	 * loop over the page range entering in as needed
406 	 */
407 
408 	retval = 0;
409 	for (lcv = 0 ; lcv < npages ; lcv++, curr_offset += PAGE_SIZE,
410 	    curr_va += PAGE_SIZE) {
411 		if ((flags & PGO_ALLPAGES) == 0 && lcv != centeridx)
412 			continue;
413 
414 		if (pps[lcv] == PGO_DONTCARE)
415 			continue;
416 
417 		mdpgno = cdev_mmap(device, curr_offset, access_type);
418 		if (mdpgno == -1) {
419 			retval = EIO;
420 			break;
421 		}
422 		paddr = pmap_phys_address(mdpgno);
423 		mmapflags = pmap_mmap_flags(mdpgno);
424 		mapprot = ufi->entry->protection;
425 		UVMHIST_LOG(maphist,
426 		    "  MAPPING: device: pm=%#jx, va=%#jx, pa=%#jx, at=%jd",
427 		    (uintptr_t)ufi->orig_map->pmap, curr_va, paddr, mapprot);
428 		if (pmap_enter(ufi->orig_map->pmap, curr_va, paddr, mapprot,
429 		    PMAP_CANFAIL | mapprot | mmapflags) != 0) {
430 			/*
431 			 * pmap_enter() didn't have the resource to
432 			 * enter this mapping.  Unlock everything,
433 			 * wait for the pagedaemon to free up some
434 			 * pages, and then tell uvm_fault() to start
435 			 * the fault again.
436 			 *
437 			 * XXX Needs some rethinking for the PGO_ALLPAGES
438 			 * XXX case.
439 			 */
440 			pmap_update(ufi->orig_map->pmap);	/* sync what we have so far */
441 			uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap,
442 			    uobj);
443 			return ENOMEM;
444 		}
445 	}
446 
447 	pmap_update(ufi->orig_map->pmap);
448 	uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj);
449 	return (retval);
450 }
451