xref: /openbsd-src/sys/uvm/uvm_device.c (revision f1dd7b858388b4a23f4f67a4957ec5ff656ebbe8)
1 /*	$OpenBSD: uvm_device.c,v 1.61 2021/03/20 10:24:21 mpi Exp $	*/
2 /*	$NetBSD: uvm_device.c,v 1.30 2000/11/25 06:27:59 chs Exp $	*/
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  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  *
28  * from: Id: uvm_device.c,v 1.1.2.9 1998/02/06 05:11:47 chs Exp
29  */
30 
31 /*
32  * uvm_device.c: the device pager.
33  */
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/conf.h>
38 #include <sys/malloc.h>
39 #include <sys/mutex.h>
40 
41 #include <uvm/uvm.h>
42 #include <uvm/uvm_device.h>
43 
44 #if defined(__amd64__) || defined(__arm64__) || \
45     defined(__i386__) || defined(__loongson__) || \
46     defined(__macppc__) || defined(__powerpc64__) || \
47     defined(__sparc64__)
48 #include "drm.h"
49 #endif
50 
51 /*
52  * private global data structure
53  *
54  * we keep a list of active device objects in the system.
55  */
56 
57 LIST_HEAD(, uvm_device) udv_list = LIST_HEAD_INITIALIZER(udv_list);
58 struct mutex udv_lock = MUTEX_INITIALIZER(IPL_NONE);
59 
60 /*
61  * functions
62  */
63 static void             udv_reference(struct uvm_object *);
64 static void             udv_detach(struct uvm_object *);
65 static int		udv_fault(struct uvm_faultinfo *, vaddr_t,
66 				       vm_page_t *, int, int, vm_fault_t,
67 				       vm_prot_t, int);
68 static boolean_t        udv_flush(struct uvm_object *, voff_t, voff_t,
69 				       int);
70 
71 /*
72  * master pager structure
73  */
74 const struct uvm_pagerops uvm_deviceops = {
75 	.pgo_reference = udv_reference,
76 	.pgo_detach = udv_detach,
77 	.pgo_fault = udv_fault,
78 	.pgo_flush = udv_flush,
79 };
80 
81 /*
82  * the ops!
83  */
84 
85 
86 /*
87  * udv_attach
88  *
89  * get a VM object that is associated with a device.   allocate a new
90  * one if needed.
91  *
92  * => nothing should be locked so that we can sleep here.
93  *
94  * The last two arguments (off and size) are only used for access checking.
95  */
96 struct uvm_object *
97 udv_attach(dev_t device, vm_prot_t accessprot, voff_t off, vsize_t size)
98 {
99 	struct uvm_device *udv, *lcv;
100 	paddr_t (*mapfn)(dev_t, off_t, int);
101 #if NDRM > 0
102 	struct uvm_object *obj;
103 #endif
104 
105 	/*
106 	 * before we do anything, ensure this device supports mmap
107 	 */
108 	mapfn = cdevsw[major(device)].d_mmap;
109 	if (mapfn == NULL ||
110 	    mapfn == (paddr_t (*)(dev_t, off_t, int)) enodev ||
111 	    mapfn == (paddr_t (*)(dev_t, off_t, int)) nullop)
112 		return(NULL);
113 
114 	/*
115 	 * Negative offsets on the object are not allowed.
116 	 */
117 	if (off < 0)
118 		return(NULL);
119 
120 #if NDRM > 0
121 	obj = udv_attach_drm(device, accessprot, off, size);
122 	if (obj)
123 		return(obj);
124 #endif
125 
126 	/*
127 	 * Check that the specified range of the device allows the
128 	 * desired protection.
129 	 *
130 	 * XXX clobbers off and size, but nothing else here needs them.
131 	 */
132 	while (size != 0) {
133 		if ((*mapfn)(device, off, accessprot) == -1)
134 			return (NULL);
135 		off += PAGE_SIZE; size -= PAGE_SIZE;
136 	}
137 
138 	/*
139 	 * keep looping until we get it
140 	 */
141 	for (;;) {
142 		/*
143 		 * first, attempt to find it on the main list
144 		 */
145 		mtx_enter(&udv_lock);
146 		LIST_FOREACH(lcv, &udv_list, u_list) {
147 			if (device == lcv->u_device)
148 				break;
149 		}
150 
151 		/*
152 		 * got it on main list.  put a hold on it and unlock udv_lock.
153 		 */
154 		if (lcv) {
155 			/*
156 			 * if someone else has a hold on it, sleep and start
157 			 * over again. Else, we need take HOLD flag so we
158 			 * don't have to re-order locking here.
159 			 */
160 			if (lcv->u_flags & UVM_DEVICE_HOLD) {
161 				lcv->u_flags |= UVM_DEVICE_WANTED;
162 				msleep_nsec(lcv, &udv_lock, PVM | PNORELOCK,
163 				    "udv_attach", INFSLP);
164 				continue;
165 			}
166 
167 			/* we are now holding it */
168 			lcv->u_flags |= UVM_DEVICE_HOLD;
169 			mtx_leave(&udv_lock);
170 
171 			/*
172 			 * bump reference count, unhold, return.
173 			 */
174 			lcv->u_obj.uo_refs++;
175 
176 			mtx_enter(&udv_lock);
177 			if (lcv->u_flags & UVM_DEVICE_WANTED)
178 				wakeup(lcv);
179 			lcv->u_flags &= ~(UVM_DEVICE_WANTED|UVM_DEVICE_HOLD);
180 			mtx_leave(&udv_lock);
181 			return(&lcv->u_obj);
182 		}
183 
184 		/*
185 		 * Did not find it on main list.  Need to allocate a new one.
186 		 */
187 		mtx_leave(&udv_lock);
188 		/* NOTE: we could sleep in the following malloc() */
189 		udv = malloc(sizeof(*udv), M_TEMP, M_WAITOK);
190 		mtx_enter(&udv_lock);
191 
192 		/*
193 		 * now we have to double check to make sure no one added it
194 		 * to the list while we were sleeping...
195 		 */
196 		LIST_FOREACH(lcv, &udv_list, u_list) {
197 			if (device == lcv->u_device)
198 				break;
199 		}
200 
201 		/*
202 		 * did we lose a race to someone else?
203 		 * free our memory and retry.
204 		 */
205 		if (lcv) {
206 			mtx_leave(&udv_lock);
207 			free(udv, M_TEMP, sizeof(*udv));
208 			continue;
209 		}
210 
211 		/*
212 		 * we have it!   init the data structures, add to list
213 		 * and return.
214 		 */
215 		uvm_objinit(&udv->u_obj, &uvm_deviceops, 1);
216 		udv->u_flags = 0;
217 		udv->u_device = device;
218 		LIST_INSERT_HEAD(&udv_list, udv, u_list);
219 		mtx_leave(&udv_lock);
220 		return(&udv->u_obj);
221 	}
222 	/*NOTREACHED*/
223 }
224 
225 /*
226  * udv_reference
227  *
228  * add a reference to a VM object.   Note that the reference count must
229  * already be one (the passed in reference) so there is no chance of the
230  * udv being released or locked out here.
231  */
232 static void
233 udv_reference(struct uvm_object *uobj)
234 {
235 	KERNEL_ASSERT_LOCKED();
236 	uobj->uo_refs++;
237 }
238 
239 /*
240  * udv_detach
241  *
242  * remove a reference to a VM object.
243  */
244 static void
245 udv_detach(struct uvm_object *uobj)
246 {
247 	struct uvm_device *udv = (struct uvm_device *)uobj;
248 
249 	KERNEL_ASSERT_LOCKED();
250 
251 	/*
252 	 * loop until done
253 	 */
254 again:
255 	if (uobj->uo_refs > 1) {
256 		uobj->uo_refs--;
257 		return;
258 	}
259 	KASSERT(uobj->uo_npages == 0 && RBT_EMPTY(uvm_objtree, &uobj->memt));
260 
261 	/*
262 	 * is it being held?   if so, wait until others are done.
263 	 */
264 	mtx_enter(&udv_lock);
265 	if (udv->u_flags & UVM_DEVICE_HOLD) {
266 		udv->u_flags |= UVM_DEVICE_WANTED;
267 		/*
268 		 * lock interleaving. -- this is ok in this case since the
269 		 * locks are both IPL_NONE
270 		 */
271 		msleep_nsec(udv, &udv_lock, PVM | PNORELOCK, "udv_detach",
272 		    INFSLP);
273 		goto again;
274 	}
275 
276 	/*
277 	 * got it!   nuke it now.
278 	 */
279 	LIST_REMOVE(udv, u_list);
280 	if (udv->u_flags & UVM_DEVICE_WANTED)
281 		wakeup(udv);
282 	mtx_leave(&udv_lock);
283 	free(udv, M_TEMP, sizeof(*udv));
284 }
285 
286 
287 /*
288  * udv_flush
289  *
290  * flush pages out of a uvm object.   a no-op for devices.
291  */
292 static boolean_t
293 udv_flush(struct uvm_object *uobj, voff_t start, voff_t stop, int flags)
294 {
295 
296 	return(TRUE);
297 }
298 
299 /*
300  * udv_fault: non-standard fault routine for device "pages"
301  *
302  * => rather than having a "get" function, we have a fault routine
303  *	since we don't return vm_pages we need full control over the
304  *	pmap_enter map in
305  * => on return, we unlock all fault data structures
306  * => flags: PGO_ALLPAGES: get all of the pages
307  *	     PGO_LOCKED: fault data structures are locked
308  *    XXX: currently PGO_LOCKED is always required ... consider removing
309  *	it as a flag
310  * => NOTE: vaddr is the VA of pps[0] in ufi->entry, _NOT_ pps[centeridx]
311  */
312 static int
313 udv_fault(struct uvm_faultinfo *ufi, vaddr_t vaddr, vm_page_t *pps, int npages,
314     int centeridx, vm_fault_t fault_type, vm_prot_t access_type, int flags)
315 {
316 	struct vm_map_entry *entry = ufi->entry;
317 	struct uvm_object *uobj = entry->object.uvm_obj;
318 	struct uvm_device *udv = (struct uvm_device *)uobj;
319 	vaddr_t curr_va;
320 	off_t curr_offset;
321 	paddr_t paddr;
322 	int lcv, retval;
323 	dev_t device;
324 	paddr_t (*mapfn)(dev_t, off_t, int);
325 	vm_prot_t mapprot;
326 
327 	KERNEL_ASSERT_LOCKED();
328 
329 	/*
330 	 * we do not allow device mappings to be mapped copy-on-write
331 	 * so we kill any attempt to do so here.
332 	 */
333 	if (UVM_ET_ISCOPYONWRITE(entry)) {
334 		uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj);
335 		return(VM_PAGER_ERROR);
336 	}
337 
338 	/*
339 	 * get device map function.
340 	 */
341 	device = udv->u_device;
342 	mapfn = cdevsw[major(device)].d_mmap;
343 
344 	/*
345 	 * now we must determine the offset in udv to use and the VA to
346 	 * use for pmap_enter.  note that we always use orig_map's pmap
347 	 * for pmap_enter (even if we have a submap).   since virtual
348 	 * addresses in a submap must match the main map, this is ok.
349 	 */
350 	/* udv offset = (offset from start of entry) + entry's offset */
351 	curr_offset = entry->offset + (vaddr - entry->start);
352 	/* pmap va = vaddr (virtual address of pps[0]) */
353 	curr_va = vaddr;
354 
355 	/*
356 	 * loop over the page range entering in as needed
357 	 */
358 	retval = VM_PAGER_OK;
359 	for (lcv = 0 ; lcv < npages ; lcv++, curr_offset += PAGE_SIZE,
360 	    curr_va += PAGE_SIZE) {
361 		if ((flags & PGO_ALLPAGES) == 0 && lcv != centeridx)
362 			continue;
363 
364 		if (pps[lcv] == PGO_DONTCARE)
365 			continue;
366 
367 		paddr = (*mapfn)(device, curr_offset, access_type);
368 		if (paddr == -1) {
369 			retval = VM_PAGER_ERROR;
370 			break;
371 		}
372 		mapprot = ufi->entry->protection;
373 		if (pmap_enter(ufi->orig_map->pmap, curr_va, paddr,
374 		    mapprot, PMAP_CANFAIL | mapprot) != 0) {
375 			/*
376 			 * pmap_enter() didn't have the resource to
377 			 * enter this mapping.  Unlock everything,
378 			 * wait for the pagedaemon to free up some
379 			 * pages, and then tell uvm_fault() to start
380 			 * the fault again.
381 			 *
382 			 * XXX Needs some rethinking for the PGO_ALLPAGES
383 			 * XXX case.
384 			 */
385 			uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap,
386 			    uobj);
387 
388 			/* sync what we have so far */
389 			pmap_update(ufi->orig_map->pmap);
390 			uvm_wait("udv_fault");
391 			return (VM_PAGER_REFAULT);
392 		}
393 	}
394 
395 	uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj);
396 	pmap_update(ufi->orig_map->pmap);
397 	return (retval);
398 }
399