xref: /illumos-gate/usr/src/uts/common/os/vm_subr.c (revision e3ae4b35c024af1196582063ecee3ab79367227d)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 1986, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
26 /*	  All Rights Reserved  	*/
27 
28 /*
29  * University Copyright- Copyright (c) 1982, 1986, 1988
30  * The Regents of the University of California
31  * All Rights Reserved
32  *
33  * University Acknowledgment- Portions of this document are derived from
34  * software developed by the University of California, Berkeley, and its
35  * contributors.
36  */
37 
38 #include <sys/types.h>
39 #include <sys/t_lock.h>
40 #include <sys/param.h>
41 #include <sys/errno.h>
42 #include <sys/debug.h>
43 #include <sys/cmn_err.h>
44 #include <sys/kmem.h>
45 #include <sys/sysmacros.h>
46 #include <sys/inline.h>
47 #include <sys/buf.h>
48 #include <sys/uio.h>
49 #include <sys/user.h>
50 #include <sys/proc.h>
51 #include <sys/systm.h>
52 #include <sys/vmsystm.h>
53 #include <sys/cpuvar.h>
54 #include <sys/mman.h>
55 #include <sys/cred.h>
56 #include <sys/vnode.h>
57 #include <sys/file.h>
58 #include <sys/vm.h>
59 
60 #include <sys/swap.h>
61 #include <sys/vtrace.h>
62 #include <sys/fs/snode.h>
63 #include <sys/copyops.h>
64 #include <sys/conf.h>
65 #include <sys/sdt.h>
66 
67 #include <vm/anon.h>
68 #include <vm/hat.h>
69 #include <vm/as.h>
70 #include <vm/seg.h>
71 #include <vm/page.h>
72 #include <vm/seg_vn.h>
73 #include <vm/seg_kmem.h>
74 
75 extern int maxphys;
76 
77 void
78 minphys(struct buf *bp)
79 {
80 	if (bp->b_bcount > maxphys)
81 		bp->b_bcount = maxphys;
82 }
83 
84 /*
85  * use kmem_cache_create for physio buffers. This has shown
86  * a better cache distribution compared to buffers on the
87  * stack. It also avoids semaphore construction/deconstruction
88  * per request
89  */
90 
91 static struct kmem_cache *physio_buf_cache;
92 
93 /* ARGSUSED */
94 static int
95 physio_buf_constructor(void *buf, void *cdrarg, int kmflags)
96 {
97 	bioinit((struct buf *)buf);
98 	return (0);
99 }
100 
101 /* ARGSUSED */
102 static void
103 physio_buf_destructor(void *buf, void *cdrarg)
104 {
105 	biofini((struct buf *)buf);
106 }
107 
108 void
109 physio_bufs_init(void)
110 {
111 	physio_buf_cache = kmem_cache_create("physio_buf_cache",
112 	    sizeof (struct buf), 0, physio_buf_constructor,
113 	    physio_buf_destructor, NULL, NULL, NULL, 0);
114 }
115 
116 
117 
118 /*
119  * initiate raw I/O request
120  *
121  * allocate buf header if necessary
122  * adjust max size of each I/O request
123  * lock down user pages and verify access protections
124  * call driver's strategy routine to submit request
125  * wait for I/O completion
126  * unlock user pages and free allocated buf header
127  */
128 
129 int
130 default_physio(int (*strat)(struct buf *), struct buf *bp, dev_t dev,
131 	int rw, void (*mincnt)(struct buf *), struct uio *uio)
132 {
133 	struct iovec *iov;
134 	struct proc *procp;
135 	struct as *asp;
136 	ssize_t c;
137 	char *a;
138 	int error = 0;
139 	page_t **pplist;
140 	int allocbuf = 0;
141 
142 	TRACE_1(TR_FAC_PHYSIO, TR_PHYSIO_START, "physio_start: bp %p", bp);
143 
144 	if (rw == B_READ) {
145 		CPU_STATS_ADD_K(sys, phread, 1);
146 	} else {
147 		CPU_STATS_ADD_K(sys, phwrite, 1);
148 	}
149 
150 	TRACE_1(TR_FAC_PHYSIO, TR_PHYSIO_GETBUF_START,
151 	    "getbuf_start: bp %p", bp);
152 
153 	if (bp == NULL) {
154 		bp = kmem_cache_alloc(physio_buf_cache, KM_SLEEP);
155 		bp->b_iodone = NULL;
156 		bp->b_resid = 0;
157 		allocbuf = 1;
158 	}
159 	TRACE_1(TR_FAC_PHYSIO, TR_PHYSIO_GETBUF_END, "getbuf_end: bp %p", bp);
160 
161 	if (uio->uio_segflg == UIO_USERSPACE) {
162 		procp = ttoproc(curthread);
163 		asp = procp->p_as;
164 	} else {
165 		procp = NULL;
166 		asp = &kas;
167 	}
168 	ASSERT(SEMA_HELD(&bp->b_sem));
169 
170 	/*
171 	 * We need to prepare this buffer for the io:::start probe, including
172 	 * NULL'ing out the file, clearing the offset, and filling in the
173 	 * b_dip field.
174 	 */
175 	bp->b_file = NULL;
176 	bp->b_offset = -1;
177 
178 	if (dev != NODEV) {
179 		(void) devopsp[getmajor(dev)]->devo_getinfo(NULL,
180 		    DDI_INFO_DEVT2DEVINFO, (void *)dev, (void **)&bp->b_dip);
181 	} else {
182 		bp->b_dip = NULL;
183 	}
184 
185 	while (uio->uio_iovcnt > 0) {
186 		iov = uio->uio_iov;
187 
188 		bp->b_error = 0;
189 		bp->b_proc = procp;
190 
191 		while (iov->iov_len > 0) {
192 			if (uio->uio_resid == 0)
193 				break;
194 			if (uio->uio_loffset < 0) {
195 				error = EINVAL;
196 				break;
197 			}
198 #ifdef	_ILP32
199 			/*
200 			 * For 32-bit kernels, check against SPEC_MAXOFFSET_T
201 			 * which represents the maximum size that can be
202 			 * supported by the IO subsystem.
203 			 * XXX this code assumes a D_64BIT driver.
204 			 */
205 			if (uio->uio_loffset > SPEC_MAXOFFSET_T) {
206 				error = EINVAL;
207 				break;
208 			}
209 #endif	/* _ILP32 */
210 			bp->b_flags = B_BUSY | B_PHYS | rw;
211 			bp->b_edev = dev;
212 			bp->b_lblkno = btodt(uio->uio_loffset);
213 
214 			/*
215 			 * Don't count on b_addr remaining untouched by the
216 			 * code below (it may be reset because someone does
217 			 * a bp_mapin on the buffer) -- reset from the iov
218 			 * each time through, updating the iov's base address
219 			 * instead.
220 			 */
221 			a = bp->b_un.b_addr = iov->iov_base;
222 			bp->b_bcount = MIN(iov->iov_len, uio->uio_resid);
223 			(*mincnt)(bp);
224 			c = bp->b_bcount;
225 
226 			TRACE_1(TR_FAC_PHYSIO, TR_PHYSIO_LOCK_START,
227 			    "as_pagelock_start: bp %p", bp);
228 
229 			error = as_pagelock(asp, &pplist, a,
230 			    c, rw == B_READ? S_WRITE : S_READ);
231 
232 			TRACE_0(TR_FAC_PHYSIO, TR_PHYSIO_LOCK_END,
233 			    "as_pagelock_end:");
234 
235 			if (error != 0) {
236 				bp->b_flags |= B_ERROR;
237 				bp->b_error = error;
238 				bp->b_flags &= ~(B_BUSY|B_WANTED|B_PHYS);
239 				break;
240 			}
241 			bp->b_shadow = pplist;
242 			if (pplist != NULL) {
243 				bp->b_flags |= B_SHADOW;
244 			}
245 
246 			DTRACE_IO1(start, struct buf *, bp);
247 			bp->b_flags |= B_STARTED;
248 
249 			(void) (*strat)(bp);
250 			error = biowait(bp);
251 
252 			/*
253 			 * unlock the pages
254 			 */
255 			TRACE_1(TR_FAC_PHYSIO, TR_PHYSIO_UNLOCK_START,
256 			    "as_pageunlock_start: bp %p", bp);
257 
258 			as_pageunlock(asp, pplist, a, c,
259 			    rw == B_READ? S_WRITE : S_READ);
260 
261 			TRACE_0(TR_FAC_PHYSIO, TR_PHYSIO_UNLOCK_END,
262 			    "as_pageunlock_end:");
263 
264 			c -= bp->b_resid;
265 			iov->iov_base += c;
266 			iov->iov_len -= c;
267 			uio->uio_resid -= c;
268 			uio->uio_loffset += c;
269 			/* bp->b_resid - temp kludge for tape drives */
270 			if (bp->b_resid || error)
271 				break;
272 		}
273 		bp->b_flags &= ~(B_BUSY|B_WANTED|B_PHYS|B_SHADOW);
274 		/* bp->b_resid - temp kludge for tape drives */
275 		if (bp->b_resid || error)
276 			break;
277 		uio->uio_iov++;
278 		uio->uio_iovcnt--;
279 	}
280 
281 	if (allocbuf) {
282 		kmem_cache_free(physio_buf_cache, bp);
283 	}
284 
285 	TRACE_1(TR_FAC_PHYSIO, TR_PHYSIO_END, "physio_end: bp %p", bp);
286 
287 	return (error);
288 }
289 
290 /*
291  * Returns 0 on success, or an error on failure.
292  *
293  * This function is no longer a part of the DDI/DKI.
294  * However, for compatibility, its interface should not
295  * be changed and it should not be removed from the kernel.
296  */
297 int
298 useracc(void *addr, size_t count, int access)
299 {
300 	uint_t prot;
301 
302 	prot = PROT_USER | ((access == B_READ) ? PROT_READ : PROT_WRITE);
303 	return (as_checkprot(ttoproc(curthread)->p_as, addr, count, prot));
304 }
305 
306 #define	MAX_MAPIN_PAGES	8
307 
308 /*
309  * This function temporarily "borrows" user pages for kernel use. If
310  * "cow" is on, it also sets up copy-on-write protection (only feasible
311  * on MAP_PRIVATE segment) on the user mappings, to protect the borrowed
312  * pages from any changes by the user. The caller is responsible for
313  * unlocking and tearing down cow settings when it's done with the pages.
314  * For an example, see kcfree().
315  *
316  * Pages behind [uaddr..uaddr+*lenp] under address space "as" are locked
317  * (shared), and mapped into kernel address range [kaddr..kaddr+*lenp] if
318  * kaddr != -1. On entering this function, cached_ppp contains a list
319  * of pages that are mapped into [kaddr..kaddr+*lenp] already (from a
320  * previous call). Thus if same pages remain behind [uaddr..uaddr+*lenp],
321  * the kernel map won't need to be reloaded again.
322  *
323  * For cow == 1, if the pages are anonymous pages, it also bumps the anon
324  * reference count, and change the user-mapping to read-only. This
325  * scheme should work on all types of segment drivers. But to be safe,
326  * we check against segvn here.
327  *
328  * Since this function is used to emulate copyin() semantic, it checks
329  * to make sure the user-mappings allow "user-read".
330  *
331  * On exit "lenp" contains the number of bytes successfully locked and
332  * mapped in. For the unsuccessful ones, the caller can fall back to
333  * copyin().
334  *
335  * Error return:
336  * ENOTSUP - operation like this is not supported either on this segment
337  * type, or on this platform type.
338  */
339 int
340 cow_mapin(struct as *as, caddr_t uaddr, caddr_t kaddr, struct page **cached_ppp,
341     struct anon **app, size_t *lenp, int cow)
342 {
343 	struct		hat *hat;
344 	struct seg	*seg;
345 	caddr_t		base;
346 	page_t		*pp, *ppp[MAX_MAPIN_PAGES];
347 	long		i;
348 	int		flags;
349 	size_t		size, total = *lenp;
350 	char		first = 1;
351 	faultcode_t	res;
352 
353 	*lenp = 0;
354 	if (cow) {
355 		AS_LOCK_ENTER(as, RW_WRITER);
356 		seg = as_findseg(as, uaddr, 0);
357 		if ((seg == NULL) || ((base = seg->s_base) > uaddr) ||
358 		    (uaddr + total) > base + seg->s_size) {
359 			AS_LOCK_EXIT(as);
360 			return (EINVAL);
361 		}
362 		/*
363 		 * The COW scheme should work for all segment types.
364 		 * But to be safe, we check against segvn.
365 		 */
366 		if (seg->s_ops != &segvn_ops) {
367 			AS_LOCK_EXIT(as);
368 			return (ENOTSUP);
369 		} else if ((SEGOP_GETTYPE(seg, uaddr) & MAP_PRIVATE) == 0) {
370 			AS_LOCK_EXIT(as);
371 			return (ENOTSUP);
372 		}
373 	}
374 	hat = as->a_hat;
375 	size = total;
376 tryagain:
377 	/*
378 	 * If (cow), hat_softlock will also change the usr protection to RO.
379 	 * This is the first step toward setting up cow. Before we
380 	 * bump up an_refcnt, we can't allow any cow-fault on this
381 	 * address. Otherwise segvn_fault will change the protection back
382 	 * to RW upon seeing an_refcnt == 1.
383 	 * The solution is to hold the writer lock on "as".
384 	 */
385 	res = hat_softlock(hat, uaddr, &size, &ppp[0], cow ? HAT_COW : 0);
386 	size = total - size;
387 	*lenp += size;
388 	size = size >> PAGESHIFT;
389 	i = 0;
390 	while (i < size) {
391 		pp = ppp[i];
392 		if (cow) {
393 			kmutex_t *ahm;
394 			/*
395 			 * Another solution is to hold SE_EXCL on pp, and
396 			 * disable PROT_WRITE. This also works for MAP_SHARED
397 			 * segment. The disadvantage is that it locks the
398 			 * page from being used by anybody else.
399 			 */
400 			ahm = AH_MUTEX(pp->p_vnode, pp->p_offset);
401 			mutex_enter(ahm);
402 			*app = swap_anon(pp->p_vnode, pp->p_offset);
403 			/*
404 			 * Since we are holding the as lock, this avoids a
405 			 * potential race with anon_decref. (segvn_unmap and
406 			 * segvn_free needs the as writer lock to do anon_free.)
407 			 */
408 			if (*app != NULL) {
409 #if 0
410 				if ((*app)->an_refcnt == 0)
411 				/*
412 				 * Consider the following senario (unlikey
413 				 * though):
414 				 * 1. an_refcnt == 2
415 				 * 2. we solftlock the page.
416 				 * 3. cow ocurrs on this addr. So a new ap,
417 				 * page and mapping is established on addr.
418 				 * 4. an_refcnt drops to 1 (segvn_faultpage
419 				 * -> anon_decref(oldap))
420 				 * 5. the last ref to ap also drops (from
421 				 * another as). It ends up blocked inside
422 				 * anon_decref trying to get page's excl lock.
423 				 * 6. Later kcfree unlocks the page, call
424 				 * anon_decref -> oops, ap is gone already.
425 				 *
426 				 * Holding as writer lock solves all problems.
427 				 */
428 					*app = NULL;
429 				else
430 #endif
431 					(*app)->an_refcnt++;
432 			}
433 			mutex_exit(ahm);
434 		} else {
435 			*app = NULL;
436 		}
437 		if (kaddr != (caddr_t)-1) {
438 			if (pp != *cached_ppp) {
439 				if (*cached_ppp == NULL)
440 					flags = HAT_LOAD_LOCK | HAT_NOSYNC |
441 					    HAT_LOAD_NOCONSIST;
442 				else
443 					flags = HAT_LOAD_REMAP |
444 					    HAT_LOAD_NOCONSIST;
445 				/*
446 				 * In order to cache the kernel mapping after
447 				 * the user page is unlocked, we call
448 				 * hat_devload instead of hat_memload so
449 				 * that the kernel mapping we set up here is
450 				 * "invisible" to the rest of the world. This
451 				 * is not very pretty. But as long as the
452 				 * caller bears the responsibility of keeping
453 				 * cache consistency, we should be ok -
454 				 * HAT_NOCONSIST will get us a uncached
455 				 * mapping on VAC. hat_softlock will flush
456 				 * a VAC_WRITEBACK cache. Therefore the kaddr
457 				 * doesn't have to be of the same vcolor as
458 				 * uaddr.
459 				 * The alternative is - change hat_devload
460 				 * to get a cached mapping. Allocate a kaddr
461 				 * with the same vcolor as uaddr. Then
462 				 * hat_softlock won't need to flush the VAC.
463 				 */
464 				hat_devload(kas.a_hat, kaddr, PAGESIZE,
465 				    page_pptonum(pp), PROT_READ, flags);
466 				*cached_ppp = pp;
467 			}
468 			kaddr += PAGESIZE;
469 		}
470 		cached_ppp++;
471 		app++;
472 		++i;
473 	}
474 	if (cow) {
475 		AS_LOCK_EXIT(as);
476 	}
477 	if (first && res == FC_NOMAP) {
478 		/*
479 		 * If the address is not mapped yet, we call as_fault to
480 		 * fault the pages in. We could've fallen back to copy and
481 		 * let it fault in the pages. But for a mapped file, we
482 		 * normally reference each page only once. For zero-copy to
483 		 * be of any use, we'd better fall in the page now and try
484 		 * again.
485 		 */
486 		first = 0;
487 		size = size << PAGESHIFT;
488 		uaddr += size;
489 		total -= size;
490 		size = total;
491 		res = as_fault(as->a_hat, as, uaddr, size, F_INVAL, S_READ);
492 		if (cow)
493 			AS_LOCK_ENTER(as, RW_WRITER);
494 		goto tryagain;
495 	}
496 	switch (res) {
497 	case FC_NOSUPPORT:
498 		return (ENOTSUP);
499 	case FC_PROT:	/* Pretend we don't know about it. This will be */
500 			/* caught by the caller when uiomove fails. */
501 	case FC_NOMAP:
502 	case FC_OBJERR:
503 	default:
504 		return (0);
505 	}
506 }
507