xref: /netbsd-src/sys/uvm/uvm_bio.c (revision 3b01aba77a7a698587faaae455bbfe740923c1f5)
1 /*	$NetBSD: uvm_bio.c,v 1.16 2001/07/18 16:44:39 thorpej Exp $	*/
2 
3 /*
4  * Copyright (c) 1998 Chuck Silvers.
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  * 3. The name of the author may not be used to endorse or promote products
16  *    derived from this software without specific prior written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
23  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
24  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
25  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
26  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  */
31 
32 #include "opt_uvmhist.h"
33 
34 /*
35  * uvm_bio.c: buffered i/o vnode mapping cache
36  */
37 
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/malloc.h>
42 #include <sys/kernel.h>
43 #include <sys/vnode.h>
44 
45 #include <uvm/uvm.h>
46 #include <uvm/uvm_page.h>
47 
48 /*
49  * global data structures
50  */
51 
52 /*
53  * local functions
54  */
55 
56 static int	ubc_fault __P((struct uvm_faultinfo *, vaddr_t,
57 		    struct vm_page **, int, int, vm_fault_t, vm_prot_t, int));
58 static struct ubc_map *ubc_find_mapping __P((struct uvm_object *, voff_t));
59 
60 /*
61  * local data structues
62  */
63 
64 #define UBC_HASH(uobj, offset) (((((u_long)(uobj)) >> 8) + \
65 				 (((u_long)(offset)) >> PAGE_SHIFT)) & \
66 				ubc_object.hashmask)
67 
68 #define UBC_QUEUE(offset) (&ubc_object.inactive[((offset) >> ubc_winshift) & \
69 					       (UBC_NQUEUES - 1)])
70 
71 struct ubc_map
72 {
73 	struct uvm_object *	uobj;		/* mapped object */
74 	voff_t			offset;		/* offset into uobj */
75 	int			refcount;	/* refcount on mapping */
76 	voff_t			writeoff;	/* overwrite offset */
77 	vsize_t			writelen;	/* overwrite len */
78 
79 	LIST_ENTRY(ubc_map)	hash;		/* hash table */
80 	TAILQ_ENTRY(ubc_map)	inactive;	/* inactive queue */
81 };
82 
83 static struct ubc_object
84 {
85 	struct uvm_object uobj;		/* glue for uvm_map() */
86 	char *kva;			/* where ubc_object is mapped */
87 	struct ubc_map *umap;		/* array of ubc_map's */
88 
89 	LIST_HEAD(, ubc_map) *hash;	/* hashtable for cached ubc_map's */
90 	u_long hashmask;		/* mask for hashtable */
91 
92 	TAILQ_HEAD(ubc_inactive_head, ubc_map) *inactive;
93 					/* inactive queues for ubc_map's */
94 
95 } ubc_object;
96 
97 struct uvm_pagerops ubc_pager =
98 {
99 	NULL,		/* init */
100 	NULL,		/* reference */
101 	NULL,		/* detach */
102 	ubc_fault,	/* fault */
103 	/* ... rest are NULL */
104 };
105 
106 int ubc_nwins = UBC_NWINS;
107 int ubc_winshift = UBC_WINSHIFT;
108 int ubc_winsize;
109 #ifdef PMAP_PREFER
110 int ubc_nqueues;
111 boolean_t ubc_release_unmap = FALSE;
112 #define UBC_NQUEUES ubc_nqueues
113 #define UBC_RELEASE_UNMAP ubc_release_unmap
114 #else
115 #define UBC_NQUEUES 1
116 #define UBC_RELEASE_UNMAP FALSE
117 #endif
118 
119 /*
120  * ubc_init
121  *
122  * init pager private data structures.
123  */
124 
125 void
126 ubc_init(void)
127 {
128 	struct ubc_map *umap;
129 	vaddr_t va;
130 	int i;
131 
132 	/*
133 	 * Make sure ubc_winshift is sane.
134 	 */
135 	if (ubc_winshift < PAGE_SHIFT)
136 		ubc_winshift = PAGE_SHIFT;
137 
138 	/*
139 	 * init ubc_object.
140 	 * alloc and init ubc_map's.
141 	 * init inactive queues.
142 	 * alloc and init hashtable.
143 	 * map in ubc_object.
144 	 */
145 
146 	simple_lock_init(&ubc_object.uobj.vmobjlock);
147 	ubc_object.uobj.pgops = &ubc_pager;
148 	TAILQ_INIT(&ubc_object.uobj.memq);
149 	ubc_object.uobj.uo_npages = 0;
150 	ubc_object.uobj.uo_refs = UVM_OBJ_KERN;
151 
152 	ubc_object.umap = malloc(ubc_nwins * sizeof(struct ubc_map),
153 				 M_TEMP, M_NOWAIT);
154 	if (ubc_object.umap == NULL)
155 		panic("ubc_init: failed to allocate ubc_map");
156 	memset(ubc_object.umap, 0, ubc_nwins * sizeof(struct ubc_map));
157 
158 	va = (vaddr_t)1L;
159 #ifdef PMAP_PREFER
160 	PMAP_PREFER(0, &va);
161 	ubc_nqueues = va >> ubc_winshift;
162 	if (ubc_nqueues == 0) {
163 		ubc_nqueues = 1;
164 	}
165 	if (ubc_nqueues != 1) {
166 		ubc_release_unmap = TRUE;
167 	}
168 #endif
169 	ubc_winsize = 1 << ubc_winshift;
170 	ubc_object.inactive = malloc(UBC_NQUEUES *
171 				     sizeof(struct ubc_inactive_head),
172 				     M_TEMP, M_NOWAIT);
173 	if (ubc_object.inactive == NULL)
174 		panic("ubc_init: failed to allocate inactive queue heads");
175 	for (i = 0; i < UBC_NQUEUES; i++) {
176 		TAILQ_INIT(&ubc_object.inactive[i]);
177 	}
178 	for (i = 0; i < ubc_nwins; i++) {
179 		umap = &ubc_object.umap[i];
180 		TAILQ_INSERT_TAIL(&ubc_object.inactive[i & (UBC_NQUEUES - 1)],
181 				  umap, inactive);
182 	}
183 
184 	ubc_object.hash = hashinit(ubc_nwins, HASH_LIST, M_TEMP, M_NOWAIT,
185 				   &ubc_object.hashmask);
186 	for (i = 0; i <= ubc_object.hashmask; i++) {
187 		LIST_INIT(&ubc_object.hash[i]);
188 	}
189 
190 	if (uvm_map(kernel_map, (vaddr_t *)&ubc_object.kva,
191 		    ubc_nwins << ubc_winshift, &ubc_object.uobj, 0, (vsize_t)va,
192 		    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
193 				UVM_ADV_RANDOM, UVM_FLAG_NOMERGE)) != 0) {
194 		panic("ubc_init: failed to map ubc_object\n");
195 	}
196 	UVMHIST_INIT(ubchist, 300);
197 }
198 
199 
200 /*
201  * ubc_fault: fault routine for ubc mapping
202  */
203 int
204 ubc_fault(ufi, ign1, ign2, ign3, ign4, fault_type, access_type, flags)
205 	struct uvm_faultinfo *ufi;
206 	vaddr_t ign1;
207 	struct vm_page **ign2;
208 	int ign3, ign4;
209 	vm_fault_t fault_type;
210 	vm_prot_t access_type;
211 	int flags;
212 {
213 	struct uvm_object *uobj;
214 	struct vnode *vp;
215 	struct ubc_map *umap;
216 	vaddr_t va, eva, ubc_offset, slot_offset;
217 	int i, error, rv, npages;
218 	struct vm_page *pgs[(1 << ubc_winshift) >> PAGE_SHIFT], *pg;
219 	UVMHIST_FUNC("ubc_fault");  UVMHIST_CALLED(ubchist);
220 
221 	/*
222 	 * no need to try with PGO_LOCKED...
223 	 * we don't need to have the map locked since we know that
224 	 * no one will mess with it until our reference is released.
225 	 */
226 	if (flags & PGO_LOCKED) {
227 #if 0
228 		return EBUSY;
229 #else
230 		uvmfault_unlockall(ufi, NULL, &ubc_object.uobj, NULL);
231 		flags &= ~PGO_LOCKED;
232 #endif
233 	}
234 
235 	va = ufi->orig_rvaddr;
236 	ubc_offset = va - (vaddr_t)ubc_object.kva;
237 
238 	UVMHIST_LOG(ubchist, "va 0x%lx ubc_offset 0x%lx at %d",
239 		    va, ubc_offset, access_type,0);
240 
241 	umap = &ubc_object.umap[ubc_offset >> ubc_winshift];
242 	KASSERT(umap->refcount != 0);
243 	slot_offset = trunc_page(ubc_offset & (ubc_winsize - 1));
244 
245 	/* no umap locking needed since we have a ref on the umap */
246 	uobj = umap->uobj;
247 	vp = (struct vnode *)uobj;
248 	KASSERT(uobj != NULL);
249 
250 	npages = (ubc_winsize - slot_offset) >> PAGE_SHIFT;
251 
252 	/*
253 	 * XXXUBC
254 	 * if npages is more than 1 we have to be sure that
255 	 * we set PGO_OVERWRITE correctly.
256 	 */
257 	if (access_type == VM_PROT_WRITE) {
258 		npages = 1;
259 	}
260 
261 again:
262 	memset(pgs, 0, sizeof (pgs));
263 	simple_lock(&uobj->vmobjlock);
264 
265 	UVMHIST_LOG(ubchist, "slot_offset 0x%x writeoff 0x%x writelen 0x%x "
266 		    "u_size 0x%x", slot_offset, umap->writeoff, umap->writelen,
267 		    vp->v_uvm.u_size);
268 
269 	if (access_type & VM_PROT_WRITE &&
270 	    slot_offset >= umap->writeoff &&
271 	    (slot_offset + PAGE_SIZE <= umap->writeoff + umap->writelen ||
272 	     slot_offset + PAGE_SIZE >= vp->v_uvm.u_size - umap->offset)) {
273 		UVMHIST_LOG(ubchist, "setting PGO_OVERWRITE", 0,0,0,0);
274 		flags |= PGO_OVERWRITE;
275 	}
276 	else { UVMHIST_LOG(ubchist, "NOT setting PGO_OVERWRITE", 0,0,0,0); }
277 	/* XXX be sure to zero any part of the page past EOF */
278 
279 	/*
280 	 * XXX
281 	 * ideally we'd like to pre-fault all of the pages we're overwriting.
282 	 * so for PGO_OVERWRITE, we should call VOP_GETPAGES() with all of the
283 	 * pages in [writeoff, writeoff+writesize] instead of just the one.
284 	 */
285 
286 	UVMHIST_LOG(ubchist, "getpages vp %p offset 0x%x npages %d",
287 		    uobj, umap->offset + slot_offset, npages, 0);
288 
289 	error = VOP_GETPAGES(vp, umap->offset + slot_offset, pgs, &npages, 0,
290 	    access_type, 0, flags);
291 	UVMHIST_LOG(ubchist, "getpages error %d npages %d", error, npages,0,0);
292 
293 	if (error == EAGAIN) {
294 		tsleep(&lbolt, PVM, "ubc_fault", 0);
295 		goto again;
296 	}
297 	if (error) {
298 		return error;
299 	}
300 	if (npages == 0) {
301 		return 0;
302 	}
303 
304 	va = ufi->orig_rvaddr;
305 	eva = ufi->orig_rvaddr + (npages << PAGE_SHIFT);
306 
307 	UVMHIST_LOG(ubchist, "va 0x%lx eva 0x%lx", va, eva, 0,0);
308 	simple_lock(&uobj->vmobjlock);
309 	for (i = 0; va < eva; i++, va += PAGE_SIZE) {
310 		UVMHIST_LOG(ubchist, "pgs[%d] = %p", i, pgs[i],0,0);
311 		pg = pgs[i];
312 
313 		if (pg == NULL || pg == PGO_DONTCARE) {
314 			continue;
315 		}
316 		if (pg->flags & PG_WANTED) {
317 			wakeup(pg);
318 		}
319 		KASSERT((pg->flags & PG_FAKE) == 0);
320 		if (pg->flags & PG_RELEASED) {
321 			rv = uobj->pgops->pgo_releasepg(pg, NULL);
322 			KASSERT(rv);
323 			continue;
324 		}
325 		KASSERT(access_type == VM_PROT_READ ||
326 			(pg->flags & PG_RDONLY) == 0);
327 
328 		uvm_lock_pageq();
329 		uvm_pageactivate(pg);
330 		uvm_unlock_pageq();
331 
332 		pmap_enter(ufi->orig_map->pmap, va, VM_PAGE_TO_PHYS(pg),
333 			   VM_PROT_READ | VM_PROT_WRITE, access_type);
334 
335 		pg->flags &= ~(PG_BUSY);
336 		UVM_PAGE_OWN(pg, NULL);
337 	}
338 	simple_unlock(&uobj->vmobjlock);
339 	pmap_update();
340 	return 0;
341 }
342 
343 /*
344  * local functions
345  */
346 
347 struct ubc_map *
348 ubc_find_mapping(uobj, offset)
349 	struct uvm_object *uobj;
350 	voff_t offset;
351 {
352 	struct ubc_map *umap;
353 
354 	LIST_FOREACH(umap, &ubc_object.hash[UBC_HASH(uobj, offset)], hash) {
355 		if (umap->uobj == uobj && umap->offset == offset) {
356 			return umap;
357 		}
358 	}
359 	return NULL;
360 }
361 
362 
363 /*
364  * ubc interface functions
365  */
366 
367 /*
368  * ubc_alloc:  allocate a buffer mapping
369  */
370 void *
371 ubc_alloc(uobj, offset, lenp, flags)
372 	struct uvm_object *uobj;
373 	voff_t offset;
374 	vsize_t *lenp;
375 	int flags;
376 {
377 	int s;
378 	vaddr_t slot_offset, va;
379 	struct ubc_map *umap;
380 	voff_t umap_offset;
381 	UVMHIST_FUNC("ubc_alloc"); UVMHIST_CALLED(ubchist);
382 
383 	UVMHIST_LOG(ubchist, "uobj %p offset 0x%lx len 0x%lx filesize 0x%x",
384 		    uobj, offset, *lenp, ((struct uvm_vnode *)uobj)->u_size);
385 
386 	umap_offset = (offset & ~((voff_t)ubc_winsize - 1));
387 	slot_offset = (vaddr_t)(offset & ((voff_t)ubc_winsize - 1));
388 	*lenp = min(*lenp, ubc_winsize - slot_offset);
389 
390 	/*
391 	 * the vnode is always locked here, so we don't need to add a ref.
392 	 */
393 
394 	s = splbio();
395 
396 again:
397 	simple_lock(&ubc_object.uobj.vmobjlock);
398 	umap = ubc_find_mapping(uobj, umap_offset);
399 	if (umap == NULL) {
400 		umap = TAILQ_FIRST(UBC_QUEUE(offset));
401 		if (umap == NULL) {
402 			simple_unlock(&ubc_object.uobj.vmobjlock);
403 			tsleep(&lbolt, PVM, "ubc_alloc", 0);
404 			goto again;
405 		}
406 
407 		/*
408 		 * remove from old hash (if any),
409 		 * add to new hash.
410 		 */
411 
412 		if (umap->uobj != NULL) {
413 			LIST_REMOVE(umap, hash);
414 		}
415 
416 		umap->uobj = uobj;
417 		umap->offset = umap_offset;
418 
419 		LIST_INSERT_HEAD(&ubc_object.hash[UBC_HASH(uobj, umap_offset)],
420 				 umap, hash);
421 
422 		va = (vaddr_t)(ubc_object.kva +
423 			       ((umap - ubc_object.umap) << ubc_winshift));
424 		pmap_remove(pmap_kernel(), va, va + ubc_winsize);
425 		pmap_update();
426 	}
427 
428 	if (umap->refcount == 0) {
429 		TAILQ_REMOVE(UBC_QUEUE(offset), umap, inactive);
430 	}
431 
432 #ifdef DIAGNOSTIC
433 	if ((flags & UBC_WRITE) &&
434 	    (umap->writeoff || umap->writelen)) {
435 		panic("ubc_fault: concurrent writes vp %p", uobj);
436 	}
437 #endif
438 	if (flags & UBC_WRITE) {
439 		umap->writeoff = slot_offset;
440 		umap->writelen = *lenp;
441 	}
442 
443 	umap->refcount++;
444 	simple_unlock(&ubc_object.uobj.vmobjlock);
445 	splx(s);
446 	UVMHIST_LOG(ubchist, "umap %p refs %d va %p",
447 		    umap, umap->refcount,
448 		    ubc_object.kva + ((umap - ubc_object.umap) << ubc_winshift),
449 		    0);
450 
451 	return ubc_object.kva +
452 		((umap - ubc_object.umap) << ubc_winshift) + slot_offset;
453 }
454 
455 
456 void
457 ubc_release(va, wlen)
458 	void *va;
459 	vsize_t wlen;
460 {
461 	struct ubc_map *umap;
462 	struct uvm_object *uobj;
463 	int s;
464 	UVMHIST_FUNC("ubc_release"); UVMHIST_CALLED(ubchist);
465 
466 	UVMHIST_LOG(ubchist, "va %p", va,0,0,0);
467 
468 	s = splbio();
469 	simple_lock(&ubc_object.uobj.vmobjlock);
470 
471 	umap = &ubc_object.umap[((char *)va - ubc_object.kva) >> ubc_winshift];
472 	uobj = umap->uobj;
473 	KASSERT(uobj != NULL);
474 
475 	umap->writeoff = 0;
476 	umap->writelen = 0;
477 	umap->refcount--;
478 	if (umap->refcount == 0) {
479 		if (UBC_RELEASE_UNMAP &&
480 		    (((struct vnode *)uobj)->v_flag & VTEXT)) {
481 			vaddr_t va;
482 
483 			/*
484 			 * if this file is the executable image of
485 			 * some process, that process will likely have
486 			 * the file mapped at an alignment other than
487 			 * what PMAP_PREFER() would like.  we'd like
488 			 * to have process text be able to use the
489 			 * cache even if someone is also reading the
490 			 * file, so invalidate mappings of such files
491 			 * as soon as possible.
492 			 */
493 
494 			va = (vaddr_t)(ubc_object.kva +
495 			    ((umap - ubc_object.umap) << ubc_winshift));
496 			pmap_remove(pmap_kernel(), va, va + ubc_winsize);
497 			pmap_update();
498 			LIST_REMOVE(umap, hash);
499 			umap->uobj = NULL;
500 			TAILQ_INSERT_HEAD(UBC_QUEUE(umap->offset), umap,
501 			    inactive);
502 		} else {
503 			TAILQ_INSERT_TAIL(UBC_QUEUE(umap->offset), umap,
504 			    inactive);
505 		}
506 	}
507 	UVMHIST_LOG(ubchist, "umap %p refs %d", umap, umap->refcount,0,0);
508 	simple_unlock(&ubc_object.uobj.vmobjlock);
509 	splx(s);
510 }
511 
512 
513 /*
514  * removing a range of mappings from the ubc mapping cache.
515  */
516 
517 void
518 ubc_flush(uobj, start, end)
519 	struct uvm_object *uobj;
520 	voff_t start, end;
521 {
522 	struct ubc_map *umap;
523 	vaddr_t va;
524 	int s;
525 	UVMHIST_FUNC("ubc_flush");  UVMHIST_CALLED(ubchist);
526 
527 	UVMHIST_LOG(ubchist, "uobj %p start 0x%lx end 0x%lx",
528 		    uobj, start, end,0);
529 
530 	s = splbio();
531 	simple_lock(&ubc_object.uobj.vmobjlock);
532 	for (umap = ubc_object.umap;
533 	     umap < &ubc_object.umap[ubc_nwins];
534 	     umap++) {
535 
536 		if (umap->uobj != uobj ||
537 		    umap->offset < start ||
538 		    (umap->offset >= end && end != 0) ||
539 		    umap->refcount > 0) {
540 			continue;
541 		}
542 
543 		/*
544 		 * remove from hash,
545 		 * move to head of inactive queue.
546 		 */
547 
548 		va = (vaddr_t)(ubc_object.kva +
549 			       ((umap - ubc_object.umap) << ubc_winshift));
550 		pmap_remove(pmap_kernel(), va, va + ubc_winsize);
551 		pmap_update();
552 
553 		LIST_REMOVE(umap, hash);
554 		umap->uobj = NULL;
555 		TAILQ_REMOVE(UBC_QUEUE(umap->offset), umap, inactive);
556 		TAILQ_INSERT_HEAD(UBC_QUEUE(umap->offset), umap, inactive);
557 	}
558 	simple_unlock(&ubc_object.uobj.vmobjlock);
559 	splx(s);
560 }
561