xref: /netbsd-src/sys/uvm/uvm_vnode.c (revision 5dd36a3bc8bf2a9dec29ceb6349550414570c447)
1 /*	$NetBSD: uvm_vnode.c,v 1.107 2020/02/27 22:12:54 ad Exp $	*/
2 
3 /*
4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
5  * Copyright (c) 1991, 1993
6  *      The Regents of the University of California.
7  * Copyright (c) 1990 University of Utah.
8  *
9  * All rights reserved.
10  *
11  * This code is derived from software contributed to Berkeley by
12  * the Systems Programming Group of the University of Utah Computer
13  * Science Department.
14  *
15  * Redistribution and use in source and binary forms, with or without
16  * modification, are permitted provided that the following conditions
17  * are met:
18  * 1. Redistributions of source code must retain the above copyright
19  *    notice, this list of conditions and the following disclaimer.
20  * 2. Redistributions in binary form must reproduce the above copyright
21  *    notice, this list of conditions and the following disclaimer in the
22  *    documentation and/or other materials provided with the distribution.
23  * 3. Neither the name of the University nor the names of its contributors
24  *    may be used to endorse or promote products derived from this software
25  *    without specific prior written permission.
26  *
27  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
28  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
29  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
30  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
31  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
32  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
33  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
34  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
35  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
36  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37  * SUCH DAMAGE.
38  *
39  *      @(#)vnode_pager.c       8.8 (Berkeley) 2/13/94
40  * from: Id: uvm_vnode.c,v 1.1.2.26 1998/02/02 20:38:07 chuck Exp
41  */
42 
43 /*
44  * uvm_vnode.c: the vnode pager.
45  */
46 
47 #include <sys/cdefs.h>
48 __KERNEL_RCSID(0, "$NetBSD: uvm_vnode.c,v 1.107 2020/02/27 22:12:54 ad Exp $");
49 
50 #ifdef _KERNEL_OPT
51 #include "opt_uvmhist.h"
52 #endif
53 
54 #include <sys/param.h>
55 #include <sys/systm.h>
56 #include <sys/kernel.h>
57 #include <sys/vnode.h>
58 #include <sys/disklabel.h>
59 #include <sys/ioctl.h>
60 #include <sys/fcntl.h>
61 #include <sys/conf.h>
62 #include <sys/pool.h>
63 #include <sys/mount.h>
64 
65 #include <miscfs/specfs/specdev.h>
66 
67 #include <uvm/uvm.h>
68 #include <uvm/uvm_readahead.h>
69 #include <uvm/uvm_page_array.h>
70 
71 #ifdef UVMHIST
72 UVMHIST_DEFINE(ubchist);
73 #endif
74 
75 /*
76  * functions
77  */
78 
79 static void	uvn_alloc_ractx(struct uvm_object *);
80 static void	uvn_detach(struct uvm_object *);
81 static int	uvn_get(struct uvm_object *, voff_t, struct vm_page **, int *,
82 			int, vm_prot_t, int, int);
83 static int	uvn_put(struct uvm_object *, voff_t, voff_t, int);
84 static void	uvn_reference(struct uvm_object *);
85 
86 static int	uvn_findpage(struct uvm_object *, voff_t, struct vm_page **,
87 			     unsigned int, struct uvm_page_array *a,
88 			     unsigned int);
89 
90 /*
91  * master pager structure
92  */
93 
94 const struct uvm_pagerops uvm_vnodeops = {
95 	.pgo_reference = uvn_reference,
96 	.pgo_detach = uvn_detach,
97 	.pgo_get = uvn_get,
98 	.pgo_put = uvn_put,
99 };
100 
101 /*
102  * the ops!
103  */
104 
105 /*
106  * uvn_reference
107  *
108  * duplicate a reference to a VM object.  Note that the reference
109  * count must already be at least one (the passed in reference) so
110  * there is no chance of the uvn being killed or locked out here.
111  *
112  * => caller must call with object unlocked.
113  * => caller must be using the same accessprot as was used at attach time
114  */
115 
116 static void
117 uvn_reference(struct uvm_object *uobj)
118 {
119 	vref((struct vnode *)uobj);
120 }
121 
122 
123 /*
124  * uvn_detach
125  *
126  * remove a reference to a VM object.
127  *
128  * => caller must call with object unlocked and map locked.
129  */
130 
131 static void
132 uvn_detach(struct uvm_object *uobj)
133 {
134 	vrele((struct vnode *)uobj);
135 }
136 
137 /*
138  * uvn_put: flush page data to backing store.
139  *
140  * => object must be locked on entry!   VOP_PUTPAGES must unlock it.
141  * => flags: PGO_SYNCIO -- use sync. I/O
142  */
143 
144 static int
145 uvn_put(struct uvm_object *uobj, voff_t offlo, voff_t offhi, int flags)
146 {
147 	struct vnode *vp = (struct vnode *)uobj;
148 	int error;
149 
150 	KASSERT(rw_write_held(uobj->vmobjlock));
151 	error = VOP_PUTPAGES(vp, offlo, offhi, flags);
152 
153 	return error;
154 }
155 
156 
157 /*
158  * uvn_get: get pages (synchronously) from backing store
159  *
160  * => prefer map unlocked (not required)
161  * => object must be locked!  we will _unlock_ it before starting any I/O.
162  * => flags: PGO_ALLPAGES: get all of the pages
163  *           PGO_LOCKED: fault data structures are locked
164  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
165  * => NOTE: caller must check for released pages!!
166  */
167 
168 static int
169 uvn_get(struct uvm_object *uobj, voff_t offset,
170     struct vm_page **pps /* IN/OUT */,
171     int *npagesp /* IN (OUT if PGO_LOCKED)*/,
172     int centeridx, vm_prot_t access_type, int advice, int flags)
173 {
174 	struct vnode *vp = (struct vnode *)uobj;
175 	int error;
176 
177 	UVMHIST_FUNC("uvn_get"); UVMHIST_CALLED(ubchist);
178 
179 	UVMHIST_LOG(ubchist, "vp %#jx off 0x%jx", (uintptr_t)vp, (int)offset,
180 	    0, 0);
181 
182 	if (vp->v_type == VREG && (access_type & VM_PROT_WRITE) == 0
183 	    && (flags & PGO_LOCKED) == 0) {
184 		uvn_alloc_ractx(uobj);
185 		uvm_ra_request(vp->v_ractx, advice, uobj, offset,
186 		    *npagesp << PAGE_SHIFT);
187 	}
188 
189 	error = VOP_GETPAGES(vp, offset, pps, npagesp, centeridx,
190 			     access_type, advice, flags);
191 
192 	KASSERT(((flags & PGO_LOCKED) != 0 && rw_lock_held(uobj->vmobjlock)) ||
193 	    (flags & PGO_LOCKED) == 0);
194 	return error;
195 }
196 
197 
198 /*
199  * uvn_findpages:
200  * return the page for the uobj and offset requested, allocating if needed.
201  * => uobj must be locked.
202  * => returned pages will be BUSY.
203  */
204 
205 int
206 uvn_findpages(struct uvm_object *uobj, voff_t offset, unsigned int *npagesp,
207     struct vm_page **pgs, struct uvm_page_array *a, unsigned int flags)
208 {
209 	unsigned int count, found, npages;
210 	int i, rv;
211 	struct uvm_page_array a_store;
212 
213 	if (a == NULL) {
214 		a = &a_store;
215 		uvm_page_array_init(a);
216 	}
217 	count = found = 0;
218 	npages = *npagesp;
219 	if (flags & UFP_BACKWARD) {
220 		for (i = npages - 1; i >= 0; i--, offset -= PAGE_SIZE) {
221 			rv = uvn_findpage(uobj, offset, &pgs[i], flags, a,
222 			    i + 1);
223 			if (rv == 0) {
224 				if (flags & UFP_DIRTYONLY)
225 					break;
226 			} else
227 				found++;
228 			count++;
229 		}
230 	} else {
231 		for (i = 0; i < npages; i++, offset += PAGE_SIZE) {
232 			rv = uvn_findpage(uobj, offset, &pgs[i], flags, a,
233 			    npages - i);
234 			if (rv == 0) {
235 				if (flags & UFP_DIRTYONLY)
236 					break;
237 			} else
238 				found++;
239 			count++;
240 		}
241 	}
242 	if (a == &a_store) {
243 		uvm_page_array_fini(a);
244 	}
245 	*npagesp = count;
246 	return (found);
247 }
248 
249 /*
250  * uvn_findpage: find a single page
251  *
252  * if a suitable page was found, put it in *pgp and return 1.
253  * otherwise return 0.
254  */
255 
256 static int
257 uvn_findpage(struct uvm_object *uobj, voff_t offset, struct vm_page **pgp,
258     unsigned int flags, struct uvm_page_array *a, unsigned int nleft)
259 {
260 	struct vm_page *pg;
261 	const unsigned int fillflags =
262 	    ((flags & UFP_BACKWARD) ? UVM_PAGE_ARRAY_FILL_BACKWARD : 0) |
263 	    ((flags & UFP_DIRTYONLY) ?
264 	    (UVM_PAGE_ARRAY_FILL_DIRTY|UVM_PAGE_ARRAY_FILL_DENSE) : 0);
265 	UVMHIST_FUNC("uvn_findpage"); UVMHIST_CALLED(ubchist);
266 	UVMHIST_LOG(ubchist, "vp %#jx off 0x%jx", (uintptr_t)uobj, offset,
267 	    0, 0);
268 
269 	KASSERT(rw_write_held(uobj->vmobjlock));
270 
271 	if (*pgp != NULL) {
272 		UVMHIST_LOG(ubchist, "dontcare", 0,0,0,0);
273 		goto skip_offset;
274 	}
275 	for (;;) {
276 		/*
277 		 * look for an existing page.
278 		 *
279 		 * XXX fragile API
280 		 * note that the array can be the one supplied by the caller of
281 		 * uvn_findpages.  in that case, fillflags used by the caller
282 		 * might not match strictly with ours.
283 		 * in particular, the caller might have filled the array
284 		 * without DENSE but passed us UFP_DIRTYONLY (thus DENSE).
285 		 */
286 		pg = uvm_page_array_fill_and_peek(a, uobj, offset, nleft,
287 		    fillflags);
288 		if (pg != NULL && pg->offset != offset) {
289 			KASSERT(
290 			    ((fillflags & UVM_PAGE_ARRAY_FILL_BACKWARD) != 0)
291 			    == (pg->offset < offset));
292 			KASSERT(uvm_pagelookup(uobj, offset) == NULL
293 			    || ((fillflags & UVM_PAGE_ARRAY_FILL_DIRTY) != 0 &&
294 			    radix_tree_get_tag(&uobj->uo_pages,
295 			    offset >> PAGE_SHIFT, UVM_PAGE_DIRTY_TAG) == 0));
296 			pg = NULL;
297 			if ((fillflags & UVM_PAGE_ARRAY_FILL_DENSE) != 0) {
298 				UVMHIST_LOG(ubchist, "dense", 0,0,0,0);
299 				return 0;
300 			}
301 		}
302 
303 		/* nope?  allocate one now */
304 		if (pg == NULL) {
305 			if (flags & UFP_NOALLOC) {
306 				UVMHIST_LOG(ubchist, "noalloc", 0,0,0,0);
307 				return 0;
308 			}
309 			pg = uvm_pagealloc(uobj, offset, NULL,
310 			    UVM_FLAG_COLORMATCH);
311 			if (pg == NULL) {
312 				if (flags & UFP_NOWAIT) {
313 					UVMHIST_LOG(ubchist, "nowait",0,0,0,0);
314 					return 0;
315 				}
316 				rw_exit(uobj->vmobjlock);
317 				uvm_wait("uvn_fp1");
318 				uvm_page_array_clear(a);
319 				rw_enter(uobj->vmobjlock, RW_WRITER);
320 				continue;
321 			}
322 			UVMHIST_LOG(ubchist, "alloced %#jx (color %ju)",
323 			    (uintptr_t)pg, VM_PGCOLOR(pg), 0, 0);
324 			KASSERTMSG(uvm_pagegetdirty(pg) ==
325 			    UVM_PAGE_STATUS_CLEAN, "page %p not clean", pg);
326 			break;
327 		} else if (flags & UFP_NOCACHE) {
328 			UVMHIST_LOG(ubchist, "nocache",0,0,0,0);
329 			goto skip;
330 		}
331 
332 		/* page is there, see if we need to wait on it */
333 		if ((pg->flags & PG_BUSY) != 0) {
334 			if (flags & UFP_NOWAIT) {
335 				UVMHIST_LOG(ubchist, "nowait",0,0,0,0);
336 				goto skip;
337 			}
338 			pg->flags |= PG_WANTED;
339 			UVMHIST_LOG(ubchist, "wait %#jx (color %ju)",
340 			    (uintptr_t)pg, VM_PGCOLOR(pg), 0, 0);
341 			UVM_UNLOCK_AND_WAIT_RW(pg, uobj->vmobjlock, 0,
342 					       "uvn_fp2", 0);
343 			uvm_page_array_clear(a);
344 			rw_enter(uobj->vmobjlock, RW_WRITER);
345 			continue;
346 		}
347 
348 		/* skip PG_RDONLY pages if requested */
349 		if ((flags & UFP_NORDONLY) && (pg->flags & PG_RDONLY)) {
350 			UVMHIST_LOG(ubchist, "nordonly",0,0,0,0);
351 			goto skip;
352 		}
353 
354 		/* stop on clean pages if requested */
355 		if (flags & UFP_DIRTYONLY) {
356 			const bool dirty = uvm_pagecheckdirty(pg, false);
357 			if (!dirty) {
358 				UVMHIST_LOG(ubchist, "dirtonly", 0,0,0,0);
359 				return 0;
360 			}
361 		}
362 
363 		/* mark the page BUSY and we're done. */
364 		pg->flags |= PG_BUSY;
365 		UVM_PAGE_OWN(pg, "uvn_findpage");
366 		UVMHIST_LOG(ubchist, "found %#jx (color %ju)",
367 		    (uintptr_t)pg, VM_PGCOLOR(pg), 0, 0);
368 		uvm_page_array_advance(a);
369 		break;
370 	}
371 	*pgp = pg;
372 	return 1;
373 
374  skip_offset:
375 	/*
376 	 * skip this offset
377 	 */
378 	pg = uvm_page_array_peek(a);
379 	if (pg != NULL) {
380 		if (pg->offset == offset) {
381 			uvm_page_array_advance(a);
382 		} else {
383 			KASSERT((fillflags & UVM_PAGE_ARRAY_FILL_DENSE) == 0);
384 		}
385 	}
386 	return 0;
387 
388  skip:
389 	/*
390 	 * skip this page
391 	 */
392 	KASSERT(pg != NULL);
393 	uvm_page_array_advance(a);
394 	return 0;
395 }
396 
397 /*
398  * uvm_vnp_setsize: grow or shrink a vnode uobj
399  *
400  * grow   => just update size value
401  * shrink => toss un-needed pages
402  *
403  * => we assume that the caller has a reference of some sort to the
404  *	vnode in question so that it will not be yanked out from under
405  *	us.
406  */
407 
408 void
409 uvm_vnp_setsize(struct vnode *vp, voff_t newsize)
410 {
411 	struct uvm_object *uobj = &vp->v_uobj;
412 	voff_t pgend = round_page(newsize);
413 	voff_t oldsize;
414 	UVMHIST_FUNC("uvm_vnp_setsize"); UVMHIST_CALLED(ubchist);
415 
416 	rw_enter(uobj->vmobjlock, RW_WRITER);
417 	UVMHIST_LOG(ubchist, "vp %#jx old 0x%jx new 0x%jx",
418 	    (uintptr_t)vp, vp->v_size, newsize, 0);
419 
420 	/*
421 	 * now check if the size has changed: if we shrink we had better
422 	 * toss some pages...
423 	 */
424 
425 	KASSERT(newsize != VSIZENOTSET && newsize >= 0);
426 	KASSERT(vp->v_size <= vp->v_writesize);
427 	KASSERT(vp->v_size == vp->v_writesize ||
428 	    newsize == vp->v_writesize || newsize <= vp->v_size);
429 
430 	oldsize = vp->v_writesize;
431 
432 	/*
433 	 * check whether size shrinks
434 	 * if old size hasn't been set, there are no pages to drop
435 	 * if there was an integer overflow in pgend, then this is no shrink
436 	 */
437 	if (oldsize > pgend && oldsize != VSIZENOTSET && pgend >= 0) {
438 		(void) uvn_put(uobj, pgend, 0, PGO_FREE | PGO_SYNCIO);
439 		rw_enter(uobj->vmobjlock, RW_WRITER);
440 	}
441 	mutex_enter(vp->v_interlock);
442 	vp->v_size = vp->v_writesize = newsize;
443 	mutex_exit(vp->v_interlock);
444 	rw_exit(uobj->vmobjlock);
445 }
446 
447 void
448 uvm_vnp_setwritesize(struct vnode *vp, voff_t newsize)
449 {
450 
451 	rw_enter(vp->v_uobj.vmobjlock, RW_WRITER);
452 	KASSERT(newsize != VSIZENOTSET && newsize >= 0);
453 	KASSERT(vp->v_size != VSIZENOTSET);
454 	KASSERT(vp->v_writesize != VSIZENOTSET);
455 	KASSERT(vp->v_size <= vp->v_writesize);
456 	KASSERT(vp->v_size <= newsize);
457 	mutex_enter(vp->v_interlock);
458 	vp->v_writesize = newsize;
459 	mutex_exit(vp->v_interlock);
460 	rw_exit(vp->v_uobj.vmobjlock);
461 }
462 
463 bool
464 uvn_text_p(struct uvm_object *uobj)
465 {
466 	struct vnode *vp = (struct vnode *)uobj;
467 
468 	/*
469 	 * v_interlock is not held here, but VI_EXECMAP is only ever changed
470 	 * with the vmobjlock held too.
471 	 */
472 	return (vp->v_iflag & VI_EXECMAP) != 0;
473 }
474 
475 bool
476 uvn_clean_p(struct uvm_object *uobj)
477 {
478 
479 	return radix_tree_empty_tagged_tree_p(&uobj->uo_pages,
480             UVM_PAGE_DIRTY_TAG);
481 }
482 
483 bool
484 uvn_needs_writefault_p(struct uvm_object *uobj)
485 {
486 	struct vnode *vp = (struct vnode *)uobj;
487 
488 	/*
489 	 * v_interlock is not held here, but VI_WRMAP and VI_WRMAPDIRTY are
490 	 * only ever changed with the vmobjlock held too, or when it's known
491 	 * the uvm_object contains no pages (VI_PAGES clear).
492 	 */
493 	return uvn_clean_p(uobj) ||
494 	    (vp->v_iflag & (VI_WRMAP|VI_WRMAPDIRTY)) == VI_WRMAP;
495 }
496 
497 static void
498 uvn_alloc_ractx(struct uvm_object *uobj)
499 {
500 	struct vnode *vp = (struct vnode *)uobj;
501 	struct uvm_ractx *ra = NULL;
502 
503 	KASSERT(rw_write_held(uobj->vmobjlock));
504 
505 	if (vp->v_type != VREG) {
506 		return;
507 	}
508 	if (vp->v_ractx != NULL) {
509 		return;
510 	}
511 	if (vp->v_ractx == NULL) {
512 		rw_exit(uobj->vmobjlock);
513 		ra = uvm_ra_allocctx();
514 		rw_enter(uobj->vmobjlock, RW_WRITER);
515 		if (ra != NULL && vp->v_ractx == NULL) {
516 			vp->v_ractx = ra;
517 			ra = NULL;
518 		}
519 	}
520 	if (ra != NULL) {
521 		uvm_ra_freectx(ra);
522 	}
523 }
524