xref: /netbsd-src/sys/uvm/uvm_bio.c (revision 82d56013d7b633d116a93943de88e08335357a7c)
1 /*	$NetBSD: uvm_bio.c,v 1.126 2021/04/01 06:26:26 simonb 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 /*
33  * uvm_bio.c: buffered i/o object mapping cache
34  */
35 
36 #include <sys/cdefs.h>
37 __KERNEL_RCSID(0, "$NetBSD: uvm_bio.c,v 1.126 2021/04/01 06:26:26 simonb Exp $");
38 
39 #include "opt_uvmhist.h"
40 #include "opt_ubc.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kmem.h>
45 #include <sys/kernel.h>
46 #include <sys/proc.h>
47 #include <sys/sysctl.h>
48 #include <sys/vnode.h>
49 #include <sys/bitops.h>		/* for ilog2() */
50 
51 #include <uvm/uvm.h>
52 #include <uvm/uvm_pdpolicy.h>
53 
54 #ifdef PMAP_DIRECT
55 #  define UBC_USE_PMAP_DIRECT
56 #endif
57 
58 /*
59  * local functions
60  */
61 
62 static int	ubc_fault(struct uvm_faultinfo *, vaddr_t, struct vm_page **,
63 			  int, int, vm_prot_t, int);
64 static struct ubc_map *ubc_find_mapping(struct uvm_object *, voff_t);
65 static int	ubchash_stats(struct hashstat_sysctl *hs, bool fill);
66 #ifdef UBC_USE_PMAP_DIRECT
67 static int __noinline ubc_uiomove_direct(struct uvm_object *, struct uio *, vsize_t,
68 			  int, int);
69 static void __noinline ubc_zerorange_direct(struct uvm_object *, off_t, size_t, int);
70 
71 /* XXX disabled by default until the kinks are worked out. */
72 bool ubc_direct = false;
73 #endif
74 
75 /*
76  * local data structues
77  */
78 
79 #define UBC_HASH(uobj, offset) 						\
80 	(((((u_long)(uobj)) >> 8) + (((u_long)(offset)) >> PAGE_SHIFT)) & \
81 				ubc_object.hashmask)
82 
83 #define UBC_QUEUE(offset)						\
84 	(&ubc_object.inactive[(((u_long)(offset)) >> ubc_winshift) &	\
85 			     (UBC_NQUEUES - 1)])
86 
87 #define UBC_UMAP_ADDR(u)						\
88 	(vaddr_t)(ubc_object.kva + (((u) - ubc_object.umap) << ubc_winshift))
89 
90 
91 #define UMAP_PAGES_LOCKED	0x0001
92 #define UMAP_MAPPING_CACHED	0x0002
93 
94 struct ubc_map {
95 	struct uvm_object *	uobj;		/* mapped object */
96 	voff_t			offset;		/* offset into uobj */
97 	voff_t			writeoff;	/* write offset */
98 	vsize_t			writelen;	/* write len */
99 	int			refcount;	/* refcount on mapping */
100 	int			flags;		/* extra state */
101 	int			advice;
102 
103 	LIST_ENTRY(ubc_map)	hash;		/* hash table */
104 	TAILQ_ENTRY(ubc_map)	inactive;	/* inactive queue */
105 	LIST_ENTRY(ubc_map)	list;		/* per-object list */
106 };
107 
108 TAILQ_HEAD(ubc_inactive_head, ubc_map);
109 static struct ubc_object {
110 	struct uvm_object uobj;		/* glue for uvm_map() */
111 	char *kva;			/* where ubc_object is mapped */
112 	struct ubc_map *umap;		/* array of ubc_map's */
113 
114 	LIST_HEAD(, ubc_map) *hash;	/* hashtable for cached ubc_map's */
115 	u_long hashmask;		/* mask for hashtable */
116 
117 	struct ubc_inactive_head *inactive;
118 					/* inactive queues for ubc_map's */
119 } ubc_object;
120 
121 const struct uvm_pagerops ubc_pager = {
122 	.pgo_fault = ubc_fault,
123 	/* ... rest are NULL */
124 };
125 
126 /* Use value at least as big as maximum page size supported by architecture */
127 #define UBC_MAX_WINSHIFT	\
128     ((1 << UBC_WINSHIFT) > MAX_PAGE_SIZE ? UBC_WINSHIFT : ilog2(MAX_PAGE_SIZE))
129 
130 int ubc_nwins = UBC_NWINS;
131 const int ubc_winshift = UBC_MAX_WINSHIFT;
132 const int ubc_winsize = 1 << UBC_MAX_WINSHIFT;
133 #if defined(PMAP_PREFER)
134 int ubc_nqueues;
135 #define UBC_NQUEUES ubc_nqueues
136 #else
137 #define UBC_NQUEUES 1
138 #endif
139 
140 #if defined(UBC_STATS)
141 
142 #define	UBC_EVCNT_DEFINE(name) \
143 struct evcnt ubc_evcnt_##name = \
144 EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "ubc", #name); \
145 EVCNT_ATTACH_STATIC(ubc_evcnt_##name);
146 #define	UBC_EVCNT_INCR(name) ubc_evcnt_##name.ev_count++
147 
148 #else /* defined(UBC_STATS) */
149 
150 #define	UBC_EVCNT_DEFINE(name)	/* nothing */
151 #define	UBC_EVCNT_INCR(name)	/* nothing */
152 
153 #endif /* defined(UBC_STATS) */
154 
155 UBC_EVCNT_DEFINE(wincachehit)
156 UBC_EVCNT_DEFINE(wincachemiss)
157 UBC_EVCNT_DEFINE(faultbusy)
158 
159 /*
160  * ubc_init
161  *
162  * init pager private data structures.
163  */
164 
165 void
166 ubc_init(void)
167 {
168 	/*
169 	 * Make sure ubc_winshift is sane.
170 	 */
171 	KASSERT(ubc_winshift >= PAGE_SHIFT);
172 
173 	/*
174 	 * init ubc_object.
175 	 * alloc and init ubc_map's.
176 	 * init inactive queues.
177 	 * alloc and init hashtable.
178 	 * map in ubc_object.
179 	 */
180 
181 	uvm_obj_init(&ubc_object.uobj, &ubc_pager, true, UVM_OBJ_KERN);
182 
183 	ubc_object.umap = kmem_zalloc(ubc_nwins * sizeof(struct ubc_map),
184 	    KM_SLEEP);
185 	if (ubc_object.umap == NULL)
186 		panic("ubc_init: failed to allocate ubc_map");
187 
188 	vaddr_t va = (vaddr_t)1L;
189 #ifdef PMAP_PREFER
190 	PMAP_PREFER(0, &va, 0, 0);	/* kernel is never topdown */
191 	ubc_nqueues = va >> ubc_winshift;
192 	if (ubc_nqueues == 0) {
193 		ubc_nqueues = 1;
194 	}
195 #endif
196 	ubc_object.inactive = kmem_alloc(UBC_NQUEUES *
197 	    sizeof(struct ubc_inactive_head), KM_SLEEP);
198 	for (int i = 0; i < UBC_NQUEUES; i++) {
199 		TAILQ_INIT(&ubc_object.inactive[i]);
200 	}
201 	for (int i = 0; i < ubc_nwins; i++) {
202 		struct ubc_map *umap;
203 		umap = &ubc_object.umap[i];
204 		TAILQ_INSERT_TAIL(&ubc_object.inactive[i & (UBC_NQUEUES - 1)],
205 				  umap, inactive);
206 	}
207 
208 	ubc_object.hash = hashinit(ubc_nwins, HASH_LIST, true,
209 	    &ubc_object.hashmask);
210 	for (int i = 0; i <= ubc_object.hashmask; i++) {
211 		LIST_INIT(&ubc_object.hash[i]);
212 	}
213 
214 	if (uvm_map(kernel_map, (vaddr_t *)&ubc_object.kva,
215 		    ubc_nwins << ubc_winshift, &ubc_object.uobj, 0, (vsize_t)va,
216 		    UVM_MAPFLAG(UVM_PROT_RW, UVM_PROT_RW, UVM_INH_NONE,
217 				UVM_ADV_RANDOM, UVM_FLAG_NOMERGE)) != 0) {
218 		panic("ubc_init: failed to map ubc_object");
219 	}
220 
221 	hashstat_register("ubchash", ubchash_stats);
222 }
223 
224 void
225 ubchist_init(void)
226 {
227 
228 	UVMHIST_INIT(ubchist, 300);
229 }
230 
231 /*
232  * ubc_fault_page: helper of ubc_fault to handle a single page.
233  *
234  * => Caller has UVM object locked.
235  * => Caller will perform pmap_update().
236  */
237 
238 static inline int
239 ubc_fault_page(const struct uvm_faultinfo *ufi, const struct ubc_map *umap,
240     struct vm_page *pg, vm_prot_t prot, vm_prot_t access_type, vaddr_t va)
241 {
242 	vm_prot_t mask;
243 	int error;
244 	bool rdonly;
245 
246 	KASSERT(rw_write_held(pg->uobject->vmobjlock));
247 
248 	KASSERT((pg->flags & PG_FAKE) == 0);
249 	if (pg->flags & PG_RELEASED) {
250 		uvm_pagefree(pg);
251 		return 0;
252 	}
253 	if (pg->loan_count != 0) {
254 
255 		/*
256 		 * Avoid unneeded loan break, if possible.
257 		 */
258 
259 		if ((access_type & VM_PROT_WRITE) == 0) {
260 			prot &= ~VM_PROT_WRITE;
261 		}
262 		if (prot & VM_PROT_WRITE) {
263 			struct vm_page *newpg;
264 
265 			newpg = uvm_loanbreak(pg);
266 			if (newpg == NULL) {
267 				uvm_page_unbusy(&pg, 1);
268 				return ENOMEM;
269 			}
270 			pg = newpg;
271 		}
272 	}
273 
274 	/*
275 	 * Note that a page whose backing store is partially allocated
276 	 * is marked as PG_RDONLY.
277 	 *
278 	 * it's a responsibility of ubc_alloc's caller to allocate backing
279 	 * blocks before writing to the window.
280 	 */
281 
282 	KASSERT((pg->flags & PG_RDONLY) == 0 ||
283 	    (access_type & VM_PROT_WRITE) == 0 ||
284 	    pg->offset < umap->writeoff ||
285 	    pg->offset + PAGE_SIZE > umap->writeoff + umap->writelen);
286 
287 	rdonly = uvm_pagereadonly_p(pg);
288 	mask = rdonly ? ~VM_PROT_WRITE : VM_PROT_ALL;
289 
290 	error = pmap_enter(ufi->orig_map->pmap, va, VM_PAGE_TO_PHYS(pg),
291 	    prot & mask, PMAP_CANFAIL | (access_type & mask));
292 
293 	uvm_pagelock(pg);
294 	uvm_pageactivate(pg);
295 	uvm_pagewakeup(pg);
296 	uvm_pageunlock(pg);
297 	pg->flags &= ~PG_BUSY;
298 	UVM_PAGE_OWN(pg, NULL);
299 
300 	return error;
301 }
302 
303 /*
304  * ubc_fault: fault routine for ubc mapping
305  */
306 
307 static int
308 ubc_fault(struct uvm_faultinfo *ufi, vaddr_t ign1, struct vm_page **ign2,
309     int ign3, int ign4, vm_prot_t access_type, int flags)
310 {
311 	struct uvm_object *uobj;
312 	struct ubc_map *umap;
313 	vaddr_t va, eva, ubc_offset, slot_offset;
314 	struct vm_page *pgs[howmany(ubc_winsize, MIN_PAGE_SIZE)];
315 	int i, error, npages;
316 	vm_prot_t prot;
317 
318 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
319 
320 	/*
321 	 * no need to try with PGO_LOCKED...
322 	 * we don't need to have the map locked since we know that
323 	 * no one will mess with it until our reference is released.
324 	 */
325 
326 	if (flags & PGO_LOCKED) {
327 		uvmfault_unlockall(ufi, NULL, &ubc_object.uobj);
328 		flags &= ~PGO_LOCKED;
329 	}
330 
331 	va = ufi->orig_rvaddr;
332 	ubc_offset = va - (vaddr_t)ubc_object.kva;
333 	umap = &ubc_object.umap[ubc_offset >> ubc_winshift];
334 	KASSERT(umap->refcount != 0);
335 	KASSERT((umap->flags & UMAP_PAGES_LOCKED) == 0);
336 	slot_offset = ubc_offset & (ubc_winsize - 1);
337 
338 	/*
339 	 * some platforms cannot write to individual bytes atomically, so
340 	 * software has to do read/modify/write of larger quantities instead.
341 	 * this means that the access_type for "write" operations
342 	 * can be VM_PROT_READ, which confuses us mightily.
343 	 *
344 	 * deal with this by resetting access_type based on the info
345 	 * that ubc_alloc() stores for us.
346 	 */
347 
348 	access_type = umap->writelen ? VM_PROT_WRITE : VM_PROT_READ;
349 	UVMHIST_LOG(ubchist, "va %#jx ubc_offset %#jx access_type %jd",
350 	    va, ubc_offset, access_type, 0);
351 
352 	if ((access_type & VM_PROT_WRITE) != 0) {
353 #ifndef PRIxOFF		/* XXX */
354 #define PRIxOFF "jx"	/* XXX */
355 #endif			/* XXX */
356 		KASSERTMSG((trunc_page(umap->writeoff) <= slot_offset),
357 		    "out of range write: slot=%#"PRIxVSIZE" off=%#"PRIxOFF,
358 		    slot_offset, (intmax_t)umap->writeoff);
359 		KASSERTMSG((slot_offset < umap->writeoff + umap->writelen),
360 		    "out of range write: slot=%#"PRIxVADDR
361 		        " off=%#"PRIxOFF" len=%#"PRIxVSIZE,
362 		    slot_offset, (intmax_t)umap->writeoff, umap->writelen);
363 	}
364 
365 	/* no umap locking needed since we have a ref on the umap */
366 	uobj = umap->uobj;
367 
368 	if ((access_type & VM_PROT_WRITE) == 0) {
369 		npages = (ubc_winsize - slot_offset) >> PAGE_SHIFT;
370 	} else {
371 		npages = (round_page(umap->offset + umap->writeoff +
372 		    umap->writelen) - (umap->offset + slot_offset))
373 		    >> PAGE_SHIFT;
374 		flags |= PGO_PASTEOF;
375 	}
376 
377 again:
378 	memset(pgs, 0, sizeof (pgs));
379 	rw_enter(uobj->vmobjlock, RW_WRITER);
380 
381 	UVMHIST_LOG(ubchist, "slot_offset %#jx writeoff %#jx writelen %#jx ",
382 	    slot_offset, umap->writeoff, umap->writelen, 0);
383 	UVMHIST_LOG(ubchist, "getpages uobj %#jx offset %#jx npages %jd",
384 	    (uintptr_t)uobj, umap->offset + slot_offset, npages, 0);
385 
386 	error = (*uobj->pgops->pgo_get)(uobj, umap->offset + slot_offset, pgs,
387 	    &npages, 0, access_type, umap->advice, flags | PGO_NOBLOCKALLOC |
388 	    PGO_NOTIMESTAMP);
389 	UVMHIST_LOG(ubchist, "getpages error %jd npages %jd", error, npages, 0,
390 	    0);
391 
392 	if (error == EAGAIN) {
393 		kpause("ubc_fault", false, hz >> 2, NULL);
394 		goto again;
395 	}
396 	if (error) {
397 		return error;
398 	}
399 
400 	/*
401 	 * For virtually-indexed, virtually-tagged caches we should avoid
402 	 * creating writable mappings when we do not absolutely need them,
403 	 * since the "compatible alias" trick does not work on such caches.
404 	 * Otherwise, we can always map the pages writable.
405 	 */
406 
407 #ifdef PMAP_CACHE_VIVT
408 	prot = VM_PROT_READ | access_type;
409 #else
410 	prot = VM_PROT_READ | VM_PROT_WRITE;
411 #endif
412 
413 	va = ufi->orig_rvaddr;
414 	eva = ufi->orig_rvaddr + (npages << PAGE_SHIFT);
415 
416 	UVMHIST_LOG(ubchist, "va %#jx eva %#jx", va, eva, 0, 0);
417 
418 	/*
419 	 * Note: normally all returned pages would have the same UVM object.
420 	 * However, layered file-systems and e.g. tmpfs, may return pages
421 	 * which belong to underlying UVM object.  In such case, lock is
422 	 * shared amongst the objects.
423 	 */
424 	rw_enter(uobj->vmobjlock, RW_WRITER);
425 	for (i = 0; va < eva; i++, va += PAGE_SIZE) {
426 		struct vm_page *pg;
427 
428 		UVMHIST_LOG(ubchist, "pgs[%jd] = %#jx", i, (uintptr_t)pgs[i],
429 		    0, 0);
430 		pg = pgs[i];
431 
432 		if (pg == NULL || pg == PGO_DONTCARE) {
433 			continue;
434 		}
435 		KASSERT(uobj->vmobjlock == pg->uobject->vmobjlock);
436 		error = ubc_fault_page(ufi, umap, pg, prot, access_type, va);
437 		if (error) {
438 			/*
439 			 * Flush (there might be pages entered), drop the lock,
440 			 * and perform uvm_wait().  Note: page will re-fault.
441 			 */
442 			pmap_update(ufi->orig_map->pmap);
443 			rw_exit(uobj->vmobjlock);
444 			uvm_wait("ubc_fault");
445 			rw_enter(uobj->vmobjlock, RW_WRITER);
446 		}
447 	}
448 	/* Must make VA visible before the unlock. */
449 	pmap_update(ufi->orig_map->pmap);
450 	rw_exit(uobj->vmobjlock);
451 
452 	return 0;
453 }
454 
455 /*
456  * local functions
457  */
458 
459 static struct ubc_map *
460 ubc_find_mapping(struct uvm_object *uobj, voff_t offset)
461 {
462 	struct ubc_map *umap;
463 
464 	LIST_FOREACH(umap, &ubc_object.hash[UBC_HASH(uobj, offset)], hash) {
465 		if (umap->uobj == uobj && umap->offset == offset) {
466 			return umap;
467 		}
468 	}
469 	return NULL;
470 }
471 
472 
473 /*
474  * ubc interface functions
475  */
476 
477 /*
478  * ubc_alloc:  allocate a file mapping window
479  */
480 
481 static void * __noinline
482 ubc_alloc(struct uvm_object *uobj, voff_t offset, vsize_t *lenp, int advice,
483     int flags, struct vm_page **pgs, int *npagesp)
484 {
485 	vaddr_t slot_offset, va;
486 	struct ubc_map *umap;
487 	voff_t umap_offset;
488 	int error;
489 	UVMHIST_FUNC(__func__);
490 	UVMHIST_CALLARGS(ubchist, "uobj %#jx offset %#jx len %#jx",
491 	    (uintptr_t)uobj, offset, *lenp, 0);
492 
493 	KASSERT(*lenp > 0);
494 	umap_offset = (offset & ~((voff_t)ubc_winsize - 1));
495 	slot_offset = (vaddr_t)(offset & ((voff_t)ubc_winsize - 1));
496 	*lenp = MIN(*lenp, ubc_winsize - slot_offset);
497 	KASSERT(*lenp > 0);
498 
499 	rw_enter(ubc_object.uobj.vmobjlock, RW_WRITER);
500 again:
501 	/*
502 	 * The UVM object is already referenced.
503 	 * Lock order: UBC object -> ubc_map::uobj.
504 	 */
505 	umap = ubc_find_mapping(uobj, umap_offset);
506 	if (umap == NULL) {
507 		struct uvm_object *oobj;
508 
509 		UBC_EVCNT_INCR(wincachemiss);
510 		umap = TAILQ_FIRST(UBC_QUEUE(offset));
511 		if (umap == NULL) {
512 			rw_exit(ubc_object.uobj.vmobjlock);
513 			kpause("ubc_alloc", false, hz >> 2, NULL);
514 			rw_enter(ubc_object.uobj.vmobjlock, RW_WRITER);
515 			goto again;
516 		}
517 
518 		va = UBC_UMAP_ADDR(umap);
519 		oobj = umap->uobj;
520 
521 		/*
522 		 * Remove from old hash (if any), add to new hash.
523 		 */
524 
525 		if (oobj != NULL) {
526 			/*
527 			 * Mapping must be removed before the list entry,
528 			 * since there is a race with ubc_purge().
529 			 */
530 			if (umap->flags & UMAP_MAPPING_CACHED) {
531 				umap->flags &= ~UMAP_MAPPING_CACHED;
532 				rw_enter(oobj->vmobjlock, RW_WRITER);
533 				pmap_remove(pmap_kernel(), va,
534 				    va + ubc_winsize);
535 				pmap_update(pmap_kernel());
536 				rw_exit(oobj->vmobjlock);
537 			}
538 			LIST_REMOVE(umap, hash);
539 			LIST_REMOVE(umap, list);
540 		} else {
541 			KASSERT((umap->flags & UMAP_MAPPING_CACHED) == 0);
542 		}
543 		umap->uobj = uobj;
544 		umap->offset = umap_offset;
545 		LIST_INSERT_HEAD(&ubc_object.hash[UBC_HASH(uobj, umap_offset)],
546 		    umap, hash);
547 		LIST_INSERT_HEAD(&uobj->uo_ubc, umap, list);
548 	} else {
549 		UBC_EVCNT_INCR(wincachehit);
550 		va = UBC_UMAP_ADDR(umap);
551 	}
552 
553 	if (umap->refcount == 0) {
554 		TAILQ_REMOVE(UBC_QUEUE(offset), umap, inactive);
555 	}
556 
557 	if (flags & UBC_WRITE) {
558 		KASSERTMSG(umap->writeoff == 0 && umap->writelen == 0,
559 		    "ubc_alloc: concurrent writes to uobj %p", uobj);
560 		umap->writeoff = slot_offset;
561 		umap->writelen = *lenp;
562 	}
563 
564 	umap->refcount++;
565 	umap->advice = advice;
566 	rw_exit(ubc_object.uobj.vmobjlock);
567 	UVMHIST_LOG(ubchist, "umap %#jx refs %jd va %#jx flags %#jx",
568 	    (uintptr_t)umap, umap->refcount, (uintptr_t)va, flags);
569 
570 	if (flags & UBC_FAULTBUSY) {
571 		int npages = (*lenp + (offset & (PAGE_SIZE - 1)) +
572 		    PAGE_SIZE - 1) >> PAGE_SHIFT;
573 		int gpflags =
574 		    PGO_SYNCIO|PGO_OVERWRITE|PGO_PASTEOF|PGO_NOBLOCKALLOC|
575 		    PGO_NOTIMESTAMP;
576 		int i;
577 		KDASSERT(flags & UBC_WRITE);
578 		KASSERT(npages <= *npagesp);
579 		KASSERT(umap->refcount == 1);
580 
581 		UBC_EVCNT_INCR(faultbusy);
582 again_faultbusy:
583 		rw_enter(uobj->vmobjlock, RW_WRITER);
584 		if (umap->flags & UMAP_MAPPING_CACHED) {
585 			umap->flags &= ~UMAP_MAPPING_CACHED;
586 			pmap_remove(pmap_kernel(), va, va + ubc_winsize);
587 		}
588 		memset(pgs, 0, *npagesp * sizeof(pgs[0]));
589 
590 		error = (*uobj->pgops->pgo_get)(uobj, trunc_page(offset), pgs,
591 		    &npages, 0, VM_PROT_READ | VM_PROT_WRITE, advice, gpflags);
592 		UVMHIST_LOG(ubchist, "faultbusy getpages %jd", error, 0, 0, 0);
593 		if (error) {
594 			/*
595 			 * Flush: the mapping above might have been removed.
596 			 */
597 			pmap_update(pmap_kernel());
598 			goto out;
599 		}
600 		for (i = 0; i < npages; i++) {
601 			struct vm_page *pg = pgs[i];
602 
603 			KASSERT(pg->uobject == uobj);
604 			if (pg->loan_count != 0) {
605 				rw_enter(uobj->vmobjlock, RW_WRITER);
606 				if (pg->loan_count != 0) {
607 					pg = uvm_loanbreak(pg);
608 				}
609 				if (pg == NULL) {
610 					pmap_kremove(va, ubc_winsize);
611 					pmap_update(pmap_kernel());
612 					uvm_page_unbusy(pgs, npages);
613 					rw_exit(uobj->vmobjlock);
614 					uvm_wait("ubc_alloc");
615 					goto again_faultbusy;
616 				}
617 				rw_exit(uobj->vmobjlock);
618 				pgs[i] = pg;
619 			}
620 			pmap_kenter_pa(
621 			    va + trunc_page(slot_offset) + (i << PAGE_SHIFT),
622 			    VM_PAGE_TO_PHYS(pg),
623 			    VM_PROT_READ | VM_PROT_WRITE, 0);
624 		}
625 		pmap_update(pmap_kernel());
626 		umap->flags |= UMAP_PAGES_LOCKED;
627 		*npagesp = npages;
628 	} else {
629 		KASSERT((umap->flags & UMAP_PAGES_LOCKED) == 0);
630 	}
631 
632 out:
633 	return (void *)(va + slot_offset);
634 }
635 
636 /*
637  * ubc_release:  free a file mapping window.
638  */
639 
640 static void __noinline
641 ubc_release(void *va, int flags, struct vm_page **pgs, int npages)
642 {
643 	struct ubc_map *umap;
644 	struct uvm_object *uobj;
645 	vaddr_t umapva;
646 	bool unmapped;
647 	UVMHIST_FUNC(__func__);
648 	UVMHIST_CALLARGS(ubchist, "va %#jx", (uintptr_t)va, 0, 0, 0);
649 
650 	umap = &ubc_object.umap[((char *)va - ubc_object.kva) >> ubc_winshift];
651 	umapva = UBC_UMAP_ADDR(umap);
652 	uobj = umap->uobj;
653 	KASSERT(uobj != NULL);
654 
655 	if (umap->flags & UMAP_PAGES_LOCKED) {
656 		const voff_t endoff = umap->writeoff + umap->writelen;
657 		const voff_t zerolen = round_page(endoff) - endoff;
658 
659 		KASSERT(npages == (round_page(endoff) -
660 		    trunc_page(umap->writeoff)) >> PAGE_SHIFT);
661 		KASSERT((umap->flags & UMAP_MAPPING_CACHED) == 0);
662 		if (zerolen) {
663 			memset((char *)umapva + endoff, 0, zerolen);
664 		}
665 		umap->flags &= ~UMAP_PAGES_LOCKED;
666 		rw_enter(uobj->vmobjlock, RW_WRITER);
667 		for (u_int i = 0; i < npages; i++) {
668 			struct vm_page *pg = pgs[i];
669 #ifdef DIAGNOSTIC
670 			paddr_t pa;
671 			bool rv;
672 			rv = pmap_extract(pmap_kernel(), umapva +
673 			    umap->writeoff + (i << PAGE_SHIFT), &pa);
674 			KASSERT(rv);
675 			KASSERT(PHYS_TO_VM_PAGE(pa) == pg);
676 #endif
677 			pg->flags &= ~PG_FAKE;
678 			KASSERTMSG(uvm_pagegetdirty(pg) ==
679 			    UVM_PAGE_STATUS_DIRTY,
680 			    "page %p not dirty", pg);
681 			KASSERT(pg->loan_count == 0);
682 			if (uvmpdpol_pageactivate_p(pg)) {
683 				uvm_pagelock(pg);
684 				uvm_pageactivate(pg);
685 				uvm_pageunlock(pg);
686 			}
687 		}
688 		pmap_kremove(umapva, ubc_winsize);
689 		pmap_update(pmap_kernel());
690 		uvm_page_unbusy(pgs, npages);
691 		rw_exit(uobj->vmobjlock);
692 		unmapped = true;
693 	} else {
694 		unmapped = false;
695 	}
696 
697 	rw_enter(ubc_object.uobj.vmobjlock, RW_WRITER);
698 	umap->writeoff = 0;
699 	umap->writelen = 0;
700 	umap->refcount--;
701 	if (umap->refcount == 0) {
702 		if (flags & UBC_UNMAP) {
703 			/*
704 			 * Invalidate any cached mappings if requested.
705 			 * This is typically used to avoid leaving
706 			 * incompatible cache aliases around indefinitely.
707 			 */
708 			rw_enter(uobj->vmobjlock, RW_WRITER);
709 			pmap_remove(pmap_kernel(), umapva,
710 				    umapva + ubc_winsize);
711 			pmap_update(pmap_kernel());
712 			rw_exit(uobj->vmobjlock);
713 
714 			umap->flags &= ~UMAP_MAPPING_CACHED;
715 			LIST_REMOVE(umap, hash);
716 			LIST_REMOVE(umap, list);
717 			umap->uobj = NULL;
718 			TAILQ_INSERT_HEAD(UBC_QUEUE(umap->offset), umap,
719 			    inactive);
720 		} else {
721 			if (!unmapped) {
722 				umap->flags |= UMAP_MAPPING_CACHED;
723 			}
724 			TAILQ_INSERT_TAIL(UBC_QUEUE(umap->offset), umap,
725 			    inactive);
726 		}
727 	}
728 	UVMHIST_LOG(ubchist, "umap %#jx refs %jd", (uintptr_t)umap,
729 	    umap->refcount, 0, 0);
730 	rw_exit(ubc_object.uobj.vmobjlock);
731 }
732 
733 /*
734  * ubc_uiomove: move data to/from an object.
735  */
736 
737 int
738 ubc_uiomove(struct uvm_object *uobj, struct uio *uio, vsize_t todo, int advice,
739     int flags)
740 {
741 	const bool overwrite = (flags & UBC_FAULTBUSY) != 0;
742 	struct vm_page *pgs[howmany(ubc_winsize, MIN_PAGE_SIZE)];
743 	voff_t off;
744 	int error, npages;
745 
746 	KASSERT(todo <= uio->uio_resid);
747 	KASSERT(((flags & UBC_WRITE) != 0 && uio->uio_rw == UIO_WRITE) ||
748 	    ((flags & UBC_READ) != 0 && uio->uio_rw == UIO_READ));
749 
750 #ifdef UBC_USE_PMAP_DIRECT
751 	/*
752 	 * during direct access pages need to be held busy to prevent them
753 	 * changing identity, and therefore if we read or write an object
754 	 * into a mapped view of same we could deadlock while faulting.
755 	 *
756 	 * avoid the problem by disallowing direct access if the object
757 	 * might be visible somewhere via mmap().
758 	 *
759 	 * XXX concurrent reads cause thundering herd issues with PG_BUSY.
760 	 * In the future enable by default for writes or if ncpu<=2, and
761 	 * make the toggle override that.
762 	 */
763 	if ((ubc_direct && (flags & UBC_ISMAPPED) == 0) ||
764 	    (flags & UBC_FAULTBUSY) != 0) {
765 		return ubc_uiomove_direct(uobj, uio, todo, advice, flags);
766 	}
767 #endif
768 
769 	off = uio->uio_offset;
770 	error = 0;
771 	while (todo > 0) {
772 		vsize_t bytelen = todo;
773 		void *win;
774 
775 		npages = __arraycount(pgs);
776 		win = ubc_alloc(uobj, off, &bytelen, advice, flags, pgs,
777 		    &npages);
778 		if (error == 0) {
779 			error = uiomove(win, bytelen, uio);
780 		}
781 		if (error != 0 && overwrite) {
782 			/*
783 			 * if we haven't initialized the pages yet,
784 			 * do it now.  it's safe to use memset here
785 			 * because we just mapped the pages above.
786 			 */
787 			memset(win, 0, bytelen);
788 		}
789 		ubc_release(win, flags, pgs, npages);
790 		off += bytelen;
791 		todo -= bytelen;
792 		if (error != 0 && (flags & UBC_PARTIALOK) != 0) {
793 			break;
794 		}
795 	}
796 
797 	return error;
798 }
799 
800 /*
801  * ubc_zerorange: set a range of bytes in an object to zero.
802  */
803 
804 void
805 ubc_zerorange(struct uvm_object *uobj, off_t off, size_t len, int flags)
806 {
807 	struct vm_page *pgs[howmany(ubc_winsize, MIN_PAGE_SIZE)];
808 	int npages;
809 
810 #ifdef UBC_USE_PMAP_DIRECT
811 	if (ubc_direct || (flags & UBC_FAULTBUSY) != 0) {
812 		ubc_zerorange_direct(uobj, off, len, flags);
813 		return;
814 	}
815 #endif
816 
817 	/*
818 	 * XXXUBC invent kzero() and use it
819 	 */
820 
821 	while (len) {
822 		void *win;
823 		vsize_t bytelen = len;
824 
825 		npages = __arraycount(pgs);
826 		win = ubc_alloc(uobj, off, &bytelen, UVM_ADV_NORMAL, UBC_WRITE,
827 		    pgs, &npages);
828 		memset(win, 0, bytelen);
829 		ubc_release(win, flags, pgs, npages);
830 
831 		off += bytelen;
832 		len -= bytelen;
833 	}
834 }
835 
836 #ifdef UBC_USE_PMAP_DIRECT
837 /* Copy data using direct map */
838 
839 /*
840  * ubc_alloc_direct:  allocate a file mapping window using direct map
841  */
842 static int __noinline
843 ubc_alloc_direct(struct uvm_object *uobj, voff_t offset, vsize_t *lenp,
844     int advice, int flags, struct vm_page **pgs, int *npages)
845 {
846 	voff_t pgoff;
847 	int error;
848 	int gpflags = flags | PGO_NOTIMESTAMP | PGO_SYNCIO;
849 	int access_type = VM_PROT_READ;
850 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
851 
852 	if (flags & UBC_WRITE) {
853 		if (flags & UBC_FAULTBUSY)
854 			gpflags |= PGO_OVERWRITE | PGO_NOBLOCKALLOC;
855 #if 0
856 		KASSERT(!UVM_OBJ_NEEDS_WRITEFAULT(uobj));
857 #endif
858 
859 		/*
860 		 * Tell genfs_getpages() we already have the journal lock,
861 		 * allow allocation past current EOF.
862 		 */
863 		gpflags |= PGO_JOURNALLOCKED | PGO_PASTEOF;
864 		access_type |= VM_PROT_WRITE;
865 	} else {
866 		/* Don't need the empty blocks allocated, PG_RDONLY is okay */
867 		gpflags |= PGO_NOBLOCKALLOC;
868 	}
869 
870 	pgoff = (offset & PAGE_MASK);
871 	*lenp = MIN(*lenp, ubc_winsize - pgoff);
872 
873 again:
874 	*npages = (*lenp + pgoff + PAGE_SIZE - 1) >> PAGE_SHIFT;
875 	KASSERT((*npages * PAGE_SIZE) <= ubc_winsize);
876 	KASSERT(*lenp + pgoff <= ubc_winsize);
877 	memset(pgs, 0, *npages * sizeof(pgs[0]));
878 
879 	rw_enter(uobj->vmobjlock, RW_WRITER);
880 	error = (*uobj->pgops->pgo_get)(uobj, trunc_page(offset), pgs,
881 	    npages, 0, access_type, advice, gpflags);
882 	UVMHIST_LOG(ubchist, "alloc_direct getpages %jd", error, 0, 0, 0);
883 	if (error) {
884 		if (error == EAGAIN) {
885 			kpause("ubc_alloc_directg", false, hz >> 2, NULL);
886 			goto again;
887 		}
888 		return error;
889 	}
890 
891 	rw_enter(uobj->vmobjlock, RW_WRITER);
892 	for (int i = 0; i < *npages; i++) {
893 		struct vm_page *pg = pgs[i];
894 
895 		KASSERT(pg != NULL);
896 		KASSERT(pg != PGO_DONTCARE);
897 		KASSERT((pg->flags & PG_FAKE) == 0 || (gpflags & PGO_OVERWRITE));
898 		KASSERT(pg->uobject->vmobjlock == uobj->vmobjlock);
899 
900 		/* Avoid breaking loan if possible, only do it on write */
901 		if ((flags & UBC_WRITE) && pg->loan_count != 0) {
902 			pg = uvm_loanbreak(pg);
903 			if (pg == NULL) {
904 				uvm_page_unbusy(pgs, *npages);
905 				rw_exit(uobj->vmobjlock);
906 				uvm_wait("ubc_alloc_directl");
907 				goto again;
908 			}
909 			pgs[i] = pg;
910 		}
911 
912 		/* Page must be writable by now */
913 		KASSERT((pg->flags & PG_RDONLY) == 0 || (flags & UBC_WRITE) == 0);
914 
915 		/*
916 		 * XXX For aobj pages.  No managed mapping - mark the page
917 		 * dirty.
918 		 */
919 		if ((flags & UBC_WRITE) != 0) {
920 			uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
921 		}
922 	}
923 	rw_exit(uobj->vmobjlock);
924 
925 	return 0;
926 }
927 
928 static void __noinline
929 ubc_direct_release(struct uvm_object *uobj,
930 	int flags, struct vm_page **pgs, int npages)
931 {
932 	rw_enter(uobj->vmobjlock, RW_WRITER);
933 	for (int i = 0; i < npages; i++) {
934 		struct vm_page *pg = pgs[i];
935 
936 		pg->flags &= ~PG_BUSY;
937 		UVM_PAGE_OWN(pg, NULL);
938 		if (pg->flags & PG_RELEASED) {
939 			pg->flags &= ~PG_RELEASED;
940 			uvm_pagefree(pg);
941 			continue;
942 		}
943 
944 		if (uvm_pagewanted_p(pg) || uvmpdpol_pageactivate_p(pg)) {
945 			uvm_pagelock(pg);
946 			uvm_pageactivate(pg);
947 			uvm_pagewakeup(pg);
948 			uvm_pageunlock(pg);
949 		}
950 
951 		/* Page was changed, no longer fake and neither clean. */
952 		if (flags & UBC_WRITE) {
953 			KASSERTMSG(uvm_pagegetdirty(pg) ==
954 			    UVM_PAGE_STATUS_DIRTY,
955 			    "page %p not dirty", pg);
956 			pg->flags &= ~PG_FAKE;
957 		}
958 	}
959 	rw_exit(uobj->vmobjlock);
960 }
961 
962 static int
963 ubc_uiomove_process(void *win, size_t len, void *arg)
964 {
965 	struct uio *uio = (struct uio *)arg;
966 
967 	return uiomove(win, len, uio);
968 }
969 
970 static int
971 ubc_zerorange_process(void *win, size_t len, void *arg)
972 {
973 	memset(win, 0, len);
974 	return 0;
975 }
976 
977 static int __noinline
978 ubc_uiomove_direct(struct uvm_object *uobj, struct uio *uio, vsize_t todo, int advice,
979     int flags)
980 {
981 	const bool overwrite = (flags & UBC_FAULTBUSY) != 0;
982 	voff_t off;
983 	int error, npages;
984 	struct vm_page *pgs[howmany(ubc_winsize, MIN_PAGE_SIZE)];
985 
986 	KASSERT(todo <= uio->uio_resid);
987 	KASSERT(((flags & UBC_WRITE) != 0 && uio->uio_rw == UIO_WRITE) ||
988 	    ((flags & UBC_READ) != 0 && uio->uio_rw == UIO_READ));
989 
990 	off = uio->uio_offset;
991 	error = 0;
992 	while (todo > 0) {
993 		vsize_t bytelen = todo;
994 
995 		error = ubc_alloc_direct(uobj, off, &bytelen, advice, flags,
996 		    pgs, &npages);
997 		if (error != 0) {
998 			/* can't do anything, failed to get the pages */
999 			break;
1000 		}
1001 
1002 		if (error == 0) {
1003 			error = uvm_direct_process(pgs, npages, off, bytelen,
1004 			    ubc_uiomove_process, uio);
1005 		}
1006 
1007 		if (overwrite) {
1008 			voff_t endoff;
1009 
1010 			/*
1011 			 * if we haven't initialized the pages yet due to an
1012 			 * error above, do it now.
1013 			 */
1014 			if (error != 0) {
1015 				(void) uvm_direct_process(pgs, npages, off,
1016 				    bytelen, ubc_zerorange_process, NULL);
1017 			}
1018 
1019 			off += bytelen;
1020 			todo -= bytelen;
1021 			endoff = off & (PAGE_SIZE - 1);
1022 
1023 			/*
1024 			 * zero out the remaining portion of the final page
1025 			 * (if any).
1026 			 */
1027 			if (todo == 0 && endoff != 0) {
1028 				vsize_t zlen = PAGE_SIZE - endoff;
1029 				(void) uvm_direct_process(pgs + npages - 1, 1,
1030 				    off, zlen, ubc_zerorange_process, NULL);
1031 			}
1032 		} else {
1033 			off += bytelen;
1034 			todo -= bytelen;
1035 		}
1036 
1037 		ubc_direct_release(uobj, flags, pgs, npages);
1038 
1039 		if (error != 0 && ISSET(flags, UBC_PARTIALOK)) {
1040 			break;
1041 		}
1042 	}
1043 
1044 	return error;
1045 }
1046 
1047 static void __noinline
1048 ubc_zerorange_direct(struct uvm_object *uobj, off_t off, size_t todo, int flags)
1049 {
1050 	int error, npages;
1051 	struct vm_page *pgs[howmany(ubc_winsize, MIN_PAGE_SIZE)];
1052 
1053 	flags |= UBC_WRITE;
1054 
1055 	error = 0;
1056 	while (todo > 0) {
1057 		vsize_t bytelen = todo;
1058 
1059 		error = ubc_alloc_direct(uobj, off, &bytelen, UVM_ADV_NORMAL,
1060 		    flags, pgs, &npages);
1061 		if (error != 0) {
1062 			/* can't do anything, failed to get the pages */
1063 			break;
1064 		}
1065 
1066 		error = uvm_direct_process(pgs, npages, off, bytelen,
1067 		    ubc_zerorange_process, NULL);
1068 
1069 		ubc_direct_release(uobj, flags, pgs, npages);
1070 
1071 		off += bytelen;
1072 		todo -= bytelen;
1073 	}
1074 }
1075 
1076 #endif /* UBC_USE_PMAP_DIRECT */
1077 
1078 /*
1079  * ubc_purge: disassociate ubc_map structures from an empty uvm_object.
1080  */
1081 
1082 void
1083 ubc_purge(struct uvm_object *uobj)
1084 {
1085 	struct ubc_map *umap;
1086 	vaddr_t va;
1087 
1088 	KASSERT(uobj->uo_npages == 0);
1089 
1090 	/*
1091 	 * Safe to check without lock held, as ubc_alloc() removes
1092 	 * the mapping and list entry in the correct order.
1093 	 */
1094 	if (__predict_true(LIST_EMPTY(&uobj->uo_ubc))) {
1095 		return;
1096 	}
1097 	rw_enter(ubc_object.uobj.vmobjlock, RW_WRITER);
1098 	while ((umap = LIST_FIRST(&uobj->uo_ubc)) != NULL) {
1099 		KASSERT(umap->refcount == 0);
1100 		for (va = 0; va < ubc_winsize; va += PAGE_SIZE) {
1101 			KASSERT(!pmap_extract(pmap_kernel(),
1102 			    va + UBC_UMAP_ADDR(umap), NULL));
1103 		}
1104 		LIST_REMOVE(umap, list);
1105 		LIST_REMOVE(umap, hash);
1106 		umap->flags &= ~UMAP_MAPPING_CACHED;
1107 		umap->uobj = NULL;
1108 	}
1109 	rw_exit(ubc_object.uobj.vmobjlock);
1110 }
1111 
1112 static int
1113 ubchash_stats(struct hashstat_sysctl *hs, bool fill)
1114 {
1115 	struct ubc_map *umap;
1116 	uint64_t chain;
1117 
1118 	strlcpy(hs->hash_name, "ubchash", sizeof(hs->hash_name));
1119 	strlcpy(hs->hash_desc, "ubc object hash", sizeof(hs->hash_desc));
1120 	if (!fill)
1121 		return 0;
1122 
1123 	hs->hash_size = ubc_object.hashmask + 1;
1124 
1125 	for (size_t i = 0; i < hs->hash_size; i++) {
1126 		chain = 0;
1127 		rw_enter(ubc_object.uobj.vmobjlock, RW_READER);
1128 		LIST_FOREACH(umap, &ubc_object.hash[i], hash) {
1129 			chain++;
1130 		}
1131 		rw_exit(ubc_object.uobj.vmobjlock);
1132 		if (chain > 0) {
1133 			hs->hash_used++;
1134 			hs->hash_items += chain;
1135 			if (chain > hs->hash_maxchain)
1136 				hs->hash_maxchain = chain;
1137 		}
1138 		preempt_point();
1139 	}
1140 
1141 	return 0;
1142 }
1143