xref: /titanic_41/usr/src/uts/common/fs/ufs/ufs_directio.c (revision f875b4ebb1dd9fdbeb043557cab38ab3bf7f6e01)
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 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /* Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T */
27 /* All Rights Reserved */
28 
29 /*
30  * Portions of this source code were derived from Berkeley 4.3 BSD
31  * under license from the Regents of the University of California.
32  */
33 
34 #pragma ident	"%Z%%M%	%I%	%E% SMI"
35 
36 #include <sys/types.h>
37 #include <sys/t_lock.h>
38 #include <sys/param.h>
39 #include <sys/time.h>
40 #include <sys/systm.h>
41 #include <sys/sysmacros.h>
42 #include <sys/resource.h>
43 #include <sys/signal.h>
44 #include <sys/cred.h>
45 #include <sys/user.h>
46 #include <sys/buf.h>
47 #include <sys/vfs.h>
48 #include <sys/vnode.h>
49 #include <sys/proc.h>
50 #include <sys/disp.h>
51 #include <sys/file.h>
52 #include <sys/fcntl.h>
53 #include <sys/flock.h>
54 #include <sys/kmem.h>
55 #include <sys/uio.h>
56 #include <sys/dnlc.h>
57 #include <sys/conf.h>
58 #include <sys/mman.h>
59 #include <sys/pathname.h>
60 #include <sys/debug.h>
61 #include <sys/vmsystm.h>
62 #include <sys/cmn_err.h>
63 #include <sys/filio.h>
64 #include <sys/atomic.h>
65 
66 #include <sys/fssnap_if.h>
67 #include <sys/fs/ufs_fs.h>
68 #include <sys/fs/ufs_lockfs.h>
69 #include <sys/fs/ufs_filio.h>
70 #include <sys/fs/ufs_inode.h>
71 #include <sys/fs/ufs_fsdir.h>
72 #include <sys/fs/ufs_quota.h>
73 #include <sys/fs/ufs_trans.h>
74 #include <sys/fs/ufs_panic.h>
75 #include <sys/dirent.h>		/* must be AFTER <sys/fs/fsdir.h>! */
76 #include <sys/errno.h>
77 
78 #include <sys/filio.h>		/* _FIOIO */
79 
80 #include <vm/hat.h>
81 #include <vm/page.h>
82 #include <vm/pvn.h>
83 #include <vm/as.h>
84 #include <vm/seg.h>
85 #include <vm/seg_map.h>
86 #include <vm/seg_vn.h>
87 #include <vm/seg_kmem.h>
88 #include <vm/rm.h>
89 #include <sys/swap.h>
90 #include <sys/epm.h>
91 
92 #include <fs/fs_subr.h>
93 
94 static void	*ufs_directio_zero_buf;
95 static int	ufs_directio_zero_len	= 8192;
96 
97 int	ufs_directio_enabled = 1;	/* feature is enabled */
98 
99 /*
100  * for kstats reader
101  */
102 struct ufs_directio_kstats {
103 	kstat_named_t	logical_reads;
104 	kstat_named_t	phys_reads;
105 	kstat_named_t	hole_reads;
106 	kstat_named_t	nread;
107 	kstat_named_t	logical_writes;
108 	kstat_named_t	phys_writes;
109 	kstat_named_t	nwritten;
110 	kstat_named_t	nflushes;
111 } ufs_directio_kstats = {
112 	{ "logical_reads",	KSTAT_DATA_UINT64 },
113 	{ "phys_reads",		KSTAT_DATA_UINT64 },
114 	{ "hole_reads",		KSTAT_DATA_UINT64 },
115 	{ "nread",		KSTAT_DATA_UINT64 },
116 	{ "logical_writes",	KSTAT_DATA_UINT64 },
117 	{ "phys_writes",	KSTAT_DATA_UINT64 },
118 	{ "nwritten",		KSTAT_DATA_UINT64 },
119 	{ "nflushes",		KSTAT_DATA_UINT64 },
120 };
121 
122 kstat_t	*ufs_directio_kstatsp;
123 
124 /*
125  * use kmem_cache_create for direct-physio buffers. This has shown
126  * a better cache distribution compared to buffers on the
127  * stack. It also avoids semaphore construction/deconstruction
128  * per request
129  */
130 struct directio_buf {
131 	struct directio_buf	*next;
132 	char		*addr;
133 	size_t		nbytes;
134 	struct buf	buf;
135 };
136 static struct kmem_cache *directio_buf_cache;
137 
138 
139 /* ARGSUSED */
140 static int
141 directio_buf_constructor(void *dbp, void *cdrarg, int kmflags)
142 {
143 	bioinit((struct buf *)&((struct directio_buf *)dbp)->buf);
144 	return (0);
145 }
146 
147 /* ARGSUSED */
148 static void
149 directio_buf_destructor(void *dbp, void *cdrarg)
150 {
151 	biofini((struct buf *)&((struct directio_buf *)dbp)->buf);
152 }
153 
154 void
155 directio_bufs_init(void)
156 {
157 	directio_buf_cache = kmem_cache_create("directio_buf_cache",
158 	    sizeof (struct directio_buf), 0,
159 	    directio_buf_constructor, directio_buf_destructor,
160 	    NULL, NULL, NULL, 0);
161 }
162 
163 void
164 ufs_directio_init(void)
165 {
166 	/*
167 	 * kstats
168 	 */
169 	ufs_directio_kstatsp = kstat_create("ufs", 0,
170 	    "directio", "ufs", KSTAT_TYPE_NAMED,
171 	    sizeof (ufs_directio_kstats) / sizeof (kstat_named_t),
172 	    KSTAT_FLAG_VIRTUAL | KSTAT_FLAG_WRITABLE);
173 	if (ufs_directio_kstatsp) {
174 		ufs_directio_kstatsp->ks_data = (void *)&ufs_directio_kstats;
175 		kstat_install(ufs_directio_kstatsp);
176 	}
177 	/*
178 	 * kzero is broken so we have to use a private buf of zeroes
179 	 */
180 	ufs_directio_zero_buf = kmem_zalloc(ufs_directio_zero_len, KM_SLEEP);
181 	directio_bufs_init();
182 }
183 
184 /*
185  * Wait for the first direct IO operation to finish
186  */
187 static int
188 directio_wait_one(struct directio_buf *dbp, long *bytes_iop)
189 {
190 	buf_t	*bp;
191 	int	error;
192 
193 	/*
194 	 * Wait for IO to finish
195 	 */
196 	bp = &dbp->buf;
197 	error = biowait(bp);
198 
199 	/*
200 	 * bytes_io will be used to figure out a resid
201 	 * for the caller. The resid is approximated by reporting
202 	 * the bytes following the first failed IO as the residual.
203 	 *
204 	 * I am cautious about using b_resid because I
205 	 * am not sure how well the disk drivers maintain it.
206 	 */
207 	if (error)
208 		if (bp->b_resid)
209 			*bytes_iop = bp->b_bcount - bp->b_resid;
210 		else
211 			*bytes_iop = 0;
212 	else
213 		*bytes_iop += bp->b_bcount;
214 	/*
215 	 * Release direct IO resources
216 	 */
217 	bp->b_flags &= ~(B_BUSY|B_WANTED|B_PHYS|B_SHADOW);
218 	kmem_cache_free(directio_buf_cache, dbp);
219 	return (error);
220 }
221 
222 /*
223  * Wait for all of the direct IO operations to finish
224  */
225 
226 uint32_t	ufs_directio_drop_kpri = 0;	/* enable kpri hack */
227 
228 static int
229 directio_wait(struct directio_buf *tail, long *bytes_iop)
230 {
231 	int	error = 0, newerror;
232 	struct directio_buf	*dbp;
233 	uint_t	kpri_req_save;
234 
235 	/*
236 	 * The linked list of directio buf structures is maintained
237 	 * in reverse order (tail->last request->penultimate request->...)
238 	 */
239 	/*
240 	 * This is the k_pri_req hack. Large numbers of threads
241 	 * sleeping with kernel priority will cause scheduler thrashing
242 	 * on an MP machine. This can be seen running Oracle using
243 	 * directio to ufs files. Sleep at normal priority here to
244 	 * more closely mimic physio to a device partition. This
245 	 * workaround is disabled by default as a niced thread could
246 	 * be starved from running while holding i_rwlock and i_contents.
247 	 */
248 	if (ufs_directio_drop_kpri) {
249 		kpri_req_save = curthread->t_kpri_req;
250 		curthread->t_kpri_req = 0;
251 	}
252 	while ((dbp = tail) != NULL) {
253 		tail = dbp->next;
254 		newerror = directio_wait_one(dbp, bytes_iop);
255 		if (error == 0)
256 			error = newerror;
257 	}
258 	if (ufs_directio_drop_kpri)
259 		curthread->t_kpri_req = kpri_req_save;
260 	return (error);
261 }
262 /*
263  * Initiate direct IO request
264  */
265 static void
266 directio_start(struct ufsvfs *ufsvfsp, dev_t dev, size_t nbytes,
267 	offset_t offset, char *addr, enum seg_rw rw, struct proc *procp,
268 	struct directio_buf **tailp, page_t **pplist)
269 {
270 	buf_t *bp;
271 	struct directio_buf *dbp;
272 
273 	/*
274 	 * Allocate a directio buf header
275 	 *   Note - list is maintained in reverse order.
276 	 *   directio_wait_one() depends on this fact when
277 	 *   adjusting the ``bytes_io'' param. bytes_io
278 	 *   is used to compute a residual in the case of error.
279 	 */
280 	dbp = kmem_cache_alloc(directio_buf_cache, KM_SLEEP);
281 	dbp->next = *tailp;
282 	*tailp = dbp;
283 
284 	/*
285 	 * Initialize buf header
286 	 */
287 	dbp->addr = addr;
288 	dbp->nbytes = nbytes;
289 	bp = &dbp->buf;
290 	bp->b_edev = dev;
291 	bp->b_lblkno = btodt(offset);
292 	bp->b_bcount = nbytes;
293 	bp->b_un.b_addr = addr;
294 	bp->b_proc = procp;
295 
296 	/*
297 	 * Note that S_WRITE implies B_READ and vice versa: a read(2)
298 	 * will B_READ data from the filesystem and S_WRITE it into
299 	 * the user's buffer; a write(2) will S_READ data from the
300 	 * user's buffer and B_WRITE it to the filesystem.
301 	 */
302 	if (rw == S_WRITE) {
303 		bp->b_flags = B_BUSY | B_PHYS | B_READ;
304 		ufs_directio_kstats.phys_reads.value.ui64++;
305 		ufs_directio_kstats.nread.value.ui64 += nbytes;
306 	} else {
307 		bp->b_flags = B_BUSY | B_PHYS | B_WRITE;
308 		ufs_directio_kstats.phys_writes.value.ui64++;
309 		ufs_directio_kstats.nwritten.value.ui64 += nbytes;
310 	}
311 	bp->b_shadow = pplist;
312 	if (pplist != NULL)
313 		bp->b_flags |= B_SHADOW;
314 
315 	/*
316 	 * Issue I/O request.
317 	 */
318 	ufsvfsp->vfs_iotstamp = lbolt;
319 	if (ufsvfsp->vfs_snapshot)
320 		fssnap_strategy(&ufsvfsp->vfs_snapshot, bp);
321 	else
322 		(void) bdev_strategy(bp);
323 
324 	if (rw == S_WRITE)
325 		lwp_stat_update(LWP_STAT_OUBLK, 1);
326 	else
327 		lwp_stat_update(LWP_STAT_INBLK, 1);
328 
329 }
330 
331 uint32_t	ufs_shared_writes;	/* writes done w/ lock shared */
332 uint32_t	ufs_cur_writes;		/* # concurrent writes */
333 uint32_t	ufs_maxcur_writes;	/* high water concurrent writes */
334 uint32_t	ufs_posix_hits;		/* writes done /w lock excl. */
335 
336 /*
337  * Force POSIX syncronous data integrity on all writes for testing.
338  */
339 uint32_t	ufs_force_posix_sdi = 0;
340 
341 /*
342  * Direct Write
343  */
344 
345 int
346 ufs_directio_write(struct inode *ip, uio_t *arg_uio, int ioflag, int rewrite,
347 	cred_t *cr, int *statusp)
348 {
349 	long		resid, bytes_written;
350 	u_offset_t	size, uoff;
351 	uio_t		*uio = arg_uio;
352 	rlim64_t	limit = uio->uio_llimit;
353 	int		on, n, error, newerror, len, has_holes;
354 	daddr_t		bn;
355 	size_t		nbytes;
356 	struct fs	*fs;
357 	vnode_t		*vp;
358 	iovec_t		*iov;
359 	struct ufsvfs	*ufsvfsp = ip->i_ufsvfs;
360 	struct proc	*procp;
361 	struct as	*as;
362 	struct directio_buf	*tail;
363 	int		exclusive, ncur, bmap_peek;
364 	uio_t		copy_uio;
365 	iovec_t		copy_iov;
366 	char		*copy_base;
367 	long		copy_resid;
368 
369 	/*
370 	 * assume that directio isn't possible (normal case)
371 	 */
372 	*statusp = DIRECTIO_FAILURE;
373 
374 	/*
375 	 * Don't go direct
376 	 */
377 	if (ufs_directio_enabled == 0)
378 		return (0);
379 
380 	/*
381 	 * mapped file; nevermind
382 	 */
383 	if (ip->i_mapcnt)
384 		return (0);
385 
386 	/*
387 	 * CAN WE DO DIRECT IO?
388 	 */
389 	uoff = uio->uio_loffset;
390 	resid = uio->uio_resid;
391 
392 	/*
393 	 * beyond limit
394 	 */
395 	if (uoff + resid > limit)
396 		return (0);
397 
398 	/*
399 	 * must be sector aligned
400 	 */
401 	if ((uoff & (u_offset_t)(DEV_BSIZE - 1)) || (resid & (DEV_BSIZE - 1)))
402 		return (0);
403 
404 	/*
405 	 * SHOULD WE DO DIRECT IO?
406 	 */
407 	size = ip->i_size;
408 	has_holes = -1;
409 
410 	/*
411 	 * only on regular files; no metadata
412 	 */
413 	if (((ip->i_mode & IFMT) != IFREG) || ip->i_ufsvfs->vfs_qinod == ip)
414 		return (0);
415 
416 	/*
417 	 * Synchronous, allocating writes run very slow in Direct-Mode
418 	 * 	XXX - can be fixed with bmap_write changes for large writes!!!
419 	 *	XXX - can be fixed for updates to "almost-full" files
420 	 *	XXX - WARNING - system hangs if bmap_write() has to
421 	 * 			allocate lots of pages since pageout
422 	 * 			suspends on locked inode
423 	 */
424 	if (!rewrite && (ip->i_flag & ISYNC)) {
425 		if ((uoff + resid) > size)
426 			return (0);
427 		has_holes = bmap_has_holes(ip);
428 		if (has_holes)
429 			return (0);
430 	}
431 
432 	/*
433 	 * Each iovec must be short aligned and sector aligned.  If
434 	 * one is not, then kmem_alloc a new buffer and copy all of
435 	 * the smaller buffers into the new buffer.  This new
436 	 * buffer will be short aligned and sector aligned.
437 	 */
438 	iov = uio->uio_iov;
439 	nbytes = uio->uio_iovcnt;
440 	while (nbytes--) {
441 		if (((uint_t)iov->iov_len & (DEV_BSIZE - 1)) != 0 ||
442 		    (intptr_t)(iov->iov_base) & 1) {
443 			copy_resid = uio->uio_resid;
444 			copy_base = kmem_alloc(copy_resid, KM_NOSLEEP);
445 			if (copy_base == NULL)
446 				return (0);
447 			copy_iov.iov_base = copy_base;
448 			copy_iov.iov_len = copy_resid;
449 			copy_uio.uio_iov = &copy_iov;
450 			copy_uio.uio_iovcnt = 1;
451 			copy_uio.uio_segflg = UIO_SYSSPACE;
452 			copy_uio.uio_extflg = UIO_COPY_DEFAULT;
453 			copy_uio.uio_loffset = uio->uio_loffset;
454 			copy_uio.uio_resid = uio->uio_resid;
455 			copy_uio.uio_llimit = uio->uio_llimit;
456 			error = uiomove(copy_base, copy_resid, UIO_WRITE, uio);
457 			if (error) {
458 				kmem_free(copy_base, copy_resid);
459 				return (0);
460 			}
461 			uio = &copy_uio;
462 			break;
463 		}
464 		iov++;
465 	}
466 
467 	/*
468 	 * From here on down, all error exits must go to errout and
469 	 * not simply return a 0.
470 	 */
471 
472 	/*
473 	 * DIRECTIO
474 	 */
475 
476 	fs = ip->i_fs;
477 
478 	/*
479 	 * POSIX check. If attempting a concurrent re-write, make sure
480 	 * that this will be a single request to the driver to meet
481 	 * POSIX synchronous data integrity requirements.
482 	 */
483 	bmap_peek = 0;
484 	if (rewrite && ((ioflag & FDSYNC) || ufs_force_posix_sdi)) {
485 		int upgrade = 0;
486 
487 		/* check easy conditions first */
488 		if (uio->uio_iovcnt != 1 || resid > ufsvfsp->vfs_ioclustsz) {
489 			upgrade = 1;
490 		} else {
491 			/* now look for contiguous allocation */
492 			len = (ssize_t)blkroundup(fs, resid);
493 			error = bmap_read(ip, uoff, &bn, &len);
494 			if (error || bn == UFS_HOLE || len == 0)
495 				goto errout;
496 			/* save a call to bmap_read later */
497 			bmap_peek = 1;
498 			if (len < resid)
499 				upgrade = 1;
500 		}
501 		if (upgrade) {
502 			rw_exit(&ip->i_contents);
503 			rw_enter(&ip->i_contents, RW_WRITER);
504 			ufs_posix_hits++;
505 		}
506 	}
507 
508 
509 	/*
510 	 * allocate space
511 	 */
512 
513 	/*
514 	 * If attempting a re-write, there is no allocation to do.
515 	 * bmap_write would trip an ASSERT if i_contents is held shared.
516 	 */
517 	if (rewrite)
518 		goto skip_alloc;
519 
520 	do {
521 		on = (int)blkoff(fs, uoff);
522 		n = (int)MIN(fs->fs_bsize - on, resid);
523 		if ((uoff + n) > ip->i_size) {
524 			error = bmap_write(ip, uoff, (int)(on + n),
525 			    (int)(uoff & (offset_t)MAXBOFFSET) == 0,
526 			    NULL, cr);
527 			/* Caller is responsible for updating i_seq if needed */
528 			if (error)
529 				break;
530 			ip->i_size = uoff + n;
531 			ip->i_flag |= IATTCHG;
532 		} else if (n == MAXBSIZE) {
533 			error = bmap_write(ip, uoff, (int)(on + n),
534 			    BI_ALLOC_ONLY, NULL, cr);
535 			/* Caller is responsible for updating i_seq if needed */
536 		} else {
537 			if (has_holes < 0)
538 				has_holes = bmap_has_holes(ip);
539 			if (has_holes) {
540 				uint_t	blk_size;
541 				u_offset_t offset;
542 
543 				offset = uoff & (offset_t)fs->fs_bmask;
544 				blk_size = (int)blksize(fs, ip,
545 				    (daddr_t)lblkno(fs, offset));
546 				error = bmap_write(ip, uoff, blk_size,
547 				    BI_NORMAL, NULL, cr);
548 				/*
549 				 * Caller is responsible for updating
550 				 * i_seq if needed
551 				 */
552 			} else
553 				error = 0;
554 		}
555 		if (error)
556 			break;
557 		uoff += n;
558 		resid -= n;
559 		/*
560 		 * if file has grown larger than 2GB, set flag
561 		 * in superblock if not already set
562 		 */
563 		if ((ip->i_size > MAXOFF32_T) &&
564 		    !(fs->fs_flags & FSLARGEFILES)) {
565 			ASSERT(ufsvfsp->vfs_lfflags & UFS_LARGEFILES);
566 			mutex_enter(&ufsvfsp->vfs_lock);
567 			fs->fs_flags |= FSLARGEFILES;
568 			ufs_sbwrite(ufsvfsp);
569 			mutex_exit(&ufsvfsp->vfs_lock);
570 		}
571 	} while (resid);
572 
573 	if (error) {
574 		/*
575 		 * restore original state
576 		 */
577 		if (resid) {
578 			if (size == ip->i_size)
579 				goto errout;
580 			(void) ufs_itrunc(ip, size, 0, cr);
581 		}
582 		/*
583 		 * try non-directio path
584 		 */
585 		goto errout;
586 	}
587 skip_alloc:
588 
589 	/*
590 	 * get rid of cached pages
591 	 */
592 	vp = ITOV(ip);
593 	exclusive = rw_write_held(&ip->i_contents);
594 	if (vn_has_cached_data(vp)) {
595 		if (!exclusive) {
596 			/*
597 			 * Still holding i_rwlock, so no allocations
598 			 * can happen after dropping contents.
599 			 */
600 			rw_exit(&ip->i_contents);
601 			rw_enter(&ip->i_contents, RW_WRITER);
602 		}
603 		(void) VOP_PUTPAGE(vp, (offset_t)0, (size_t)0, B_INVAL, cr);
604 		if (vn_has_cached_data(vp))
605 			goto errout;
606 		if (!exclusive)
607 			rw_downgrade(&ip->i_contents);
608 		ufs_directio_kstats.nflushes.value.ui64++;
609 	}
610 
611 	/*
612 	 * Direct Writes
613 	 */
614 
615 	if (!exclusive) {
616 		ufs_shared_writes++;
617 		ncur = atomic_add_32_nv(&ufs_cur_writes, 1);
618 		if (ncur > ufs_maxcur_writes)
619 			ufs_maxcur_writes = ncur;
620 	}
621 
622 	/*
623 	 * proc and as are for VM operations in directio_start()
624 	 */
625 	if (uio->uio_segflg == UIO_USERSPACE) {
626 		procp = ttoproc(curthread);
627 		as = procp->p_as;
628 	} else {
629 		procp = NULL;
630 		as = &kas;
631 	}
632 	*statusp = DIRECTIO_SUCCESS;
633 	error = 0;
634 	newerror = 0;
635 	resid = uio->uio_resid;
636 	bytes_written = 0;
637 	ufs_directio_kstats.logical_writes.value.ui64++;
638 	while (error == 0 && newerror == 0 && resid && uio->uio_iovcnt) {
639 		size_t pglck_len, pglck_size;
640 		caddr_t pglck_base;
641 		page_t **pplist, **spplist;
642 
643 		tail = NULL;
644 
645 		/*
646 		 * Adjust number of bytes
647 		 */
648 		iov = uio->uio_iov;
649 		pglck_len = (size_t)MIN(iov->iov_len, resid);
650 		pglck_base = iov->iov_base;
651 		if (pglck_len == 0) {
652 			uio->uio_iov++;
653 			uio->uio_iovcnt--;
654 			continue;
655 		}
656 
657 		/*
658 		 * Try to Lock down the largest chunck of pages possible.
659 		 */
660 		pglck_len = (size_t)MIN(pglck_len,  ufsvfsp->vfs_ioclustsz);
661 		error = as_pagelock(as, &pplist, pglck_base, pglck_len, S_READ);
662 
663 		if (error)
664 			break;
665 
666 		pglck_size = pglck_len;
667 		while (pglck_len) {
668 
669 			nbytes = pglck_len;
670 			uoff = uio->uio_loffset;
671 
672 			if (!bmap_peek) {
673 
674 				/*
675 				 * Re-adjust number of bytes to contiguous
676 				 * range. May have already called bmap_read
677 				 * in the case of a concurrent rewrite.
678 				 */
679 				len = (ssize_t)blkroundup(fs, nbytes);
680 				error = bmap_read(ip, uoff, &bn, &len);
681 				if (error)
682 					break;
683 				if (bn == UFS_HOLE || len == 0)
684 					break;
685 			}
686 			nbytes = (size_t)MIN(nbytes, len);
687 			bmap_peek = 0;
688 
689 			/*
690 			 * Get the pagelist pointer for this offset to be
691 			 * passed to directio_start.
692 			 */
693 
694 			if (pplist != NULL)
695 				spplist = pplist +
696 				    btop((uintptr_t)iov->iov_base -
697 				    ((uintptr_t)pglck_base & PAGEMASK));
698 			else
699 				spplist = NULL;
700 
701 			/*
702 			 * Kick off the direct write requests
703 			 */
704 			directio_start(ufsvfsp, ip->i_dev, nbytes, ldbtob(bn),
705 			    iov->iov_base, S_READ, procp, &tail, spplist);
706 
707 			/*
708 			 * Adjust pointers and counters
709 			 */
710 			iov->iov_len -= nbytes;
711 			iov->iov_base += nbytes;
712 			uio->uio_loffset += nbytes;
713 			resid -= nbytes;
714 			pglck_len -= nbytes;
715 		}
716 
717 		/*
718 		 * Wait for outstanding requests
719 		 */
720 		newerror = directio_wait(tail, &bytes_written);
721 
722 		/*
723 		 * Release VM resources
724 		 */
725 		as_pageunlock(as, pplist, pglck_base, pglck_size, S_READ);
726 
727 	}
728 
729 	if (!exclusive) {
730 		atomic_add_32(&ufs_cur_writes, -1);
731 		/*
732 		 * If this write was done shared, readers may
733 		 * have pulled in unmodified pages. Get rid of
734 		 * these potentially stale pages.
735 		 */
736 		if (vn_has_cached_data(vp)) {
737 			rw_exit(&ip->i_contents);
738 			rw_enter(&ip->i_contents, RW_WRITER);
739 			(void) VOP_PUTPAGE(vp, (offset_t)0, (size_t)0,
740 			    B_INVAL, cr);
741 			ufs_directio_kstats.nflushes.value.ui64++;
742 			rw_downgrade(&ip->i_contents);
743 		}
744 	}
745 
746 	/*
747 	 * If error, adjust resid to begin at the first
748 	 * un-writable byte.
749 	 */
750 	if (error == 0)
751 		error = newerror;
752 	if (error)
753 		resid = uio->uio_resid - bytes_written;
754 	arg_uio->uio_resid = resid;
755 
756 	if (!rewrite) {
757 		ip->i_flag |= IUPD | ICHG;
758 		/* Caller will update i_seq */
759 		TRANS_INODE(ip->i_ufsvfs, ip);
760 	}
761 	/*
762 	 * If there is a residual; adjust the EOF if necessary
763 	 */
764 	if (resid) {
765 		if (size != ip->i_size) {
766 			if (uio->uio_loffset > size)
767 				size = uio->uio_loffset;
768 			(void) ufs_itrunc(ip, size, 0, cr);
769 		}
770 	}
771 
772 	if (uio == &copy_uio)
773 		kmem_free(copy_base, copy_resid);
774 
775 	return (error);
776 
777 errout:
778 	if (uio == &copy_uio)
779 		kmem_free(copy_base, copy_resid);
780 
781 	return (0);
782 }
783 /*
784  * Direct read of a hole
785  */
786 static int
787 directio_hole(struct uio *uio, size_t nbytes)
788 {
789 	int		error = 0, nzero;
790 	uio_t		phys_uio;
791 	iovec_t		phys_iov;
792 
793 	ufs_directio_kstats.hole_reads.value.ui64++;
794 	ufs_directio_kstats.nread.value.ui64 += nbytes;
795 
796 	phys_iov.iov_base = uio->uio_iov->iov_base;
797 	phys_iov.iov_len = nbytes;
798 
799 	phys_uio.uio_iov = &phys_iov;
800 	phys_uio.uio_iovcnt = 1;
801 	phys_uio.uio_resid = phys_iov.iov_len;
802 	phys_uio.uio_segflg = uio->uio_segflg;
803 	phys_uio.uio_extflg = uio->uio_extflg;
804 	while (error == 0 && phys_uio.uio_resid) {
805 		nzero = (int)MIN(phys_iov.iov_len, ufs_directio_zero_len);
806 		error = uiomove(ufs_directio_zero_buf, nzero, UIO_READ,
807 		    &phys_uio);
808 	}
809 	return (error);
810 }
811 
812 /*
813  * Direct Read
814  */
815 int
816 ufs_directio_read(struct inode *ip, uio_t *uio, cred_t *cr, int *statusp)
817 {
818 	ssize_t		resid, bytes_read;
819 	u_offset_t	size, uoff;
820 	int		error, newerror, len;
821 	size_t		nbytes;
822 	struct fs	*fs;
823 	vnode_t		*vp;
824 	daddr_t		bn;
825 	iovec_t		*iov;
826 	struct ufsvfs	*ufsvfsp = ip->i_ufsvfs;
827 	struct proc	*procp;
828 	struct as	*as;
829 	struct directio_buf	*tail;
830 
831 	/*
832 	 * assume that directio isn't possible (normal case)
833 	 */
834 	*statusp = DIRECTIO_FAILURE;
835 
836 	/*
837 	 * Don't go direct
838 	 */
839 	if (ufs_directio_enabled == 0)
840 		return (0);
841 
842 	/*
843 	 * mapped file; nevermind
844 	 */
845 	if (ip->i_mapcnt)
846 		return (0);
847 
848 	/*
849 	 * CAN WE DO DIRECT IO?
850 	 */
851 	/*
852 	 * must be sector aligned
853 	 */
854 	uoff = uio->uio_loffset;
855 	resid = uio->uio_resid;
856 	if ((uoff & (u_offset_t)(DEV_BSIZE - 1)) || (resid & (DEV_BSIZE - 1)))
857 		return (0);
858 	/*
859 	 * must be short aligned and sector aligned
860 	 */
861 	iov = uio->uio_iov;
862 	nbytes = uio->uio_iovcnt;
863 	while (nbytes--) {
864 		if (((size_t)iov->iov_len & (DEV_BSIZE - 1)) != 0)
865 			return (0);
866 		if ((intptr_t)(iov++->iov_base) & 1)
867 			return (0);
868 	}
869 
870 	/*
871 	 * DIRECTIO
872 	 */
873 	fs = ip->i_fs;
874 
875 	/*
876 	 * don't read past EOF
877 	 */
878 	size = ip->i_size;
879 
880 	/*
881 	 * The file offset is past EOF so bail out here; we don't want
882 	 * to update uio_resid and make it look like we read something.
883 	 * We say that direct I/O was a success to avoid having rdip()
884 	 * go through the same "read past EOF logic".
885 	 */
886 	if (uoff >= size) {
887 		*statusp = DIRECTIO_SUCCESS;
888 		return (0);
889 	}
890 
891 	/*
892 	 * The read would extend past EOF so make it smaller.
893 	 */
894 	if ((uoff + resid) > size) {
895 		resid = size - uoff;
896 		/*
897 		 * recheck sector alignment
898 		 */
899 		if (resid & (DEV_BSIZE - 1))
900 			return (0);
901 	}
902 
903 	/*
904 	 * At this point, we know there is some real work to do.
905 	 */
906 	ASSERT(resid);
907 
908 	/*
909 	 * get rid of cached pages
910 	 */
911 	vp = ITOV(ip);
912 	if (vn_has_cached_data(vp)) {
913 		rw_exit(&ip->i_contents);
914 		rw_enter(&ip->i_contents, RW_WRITER);
915 		(void) VOP_PUTPAGE(vp, (offset_t)0, (size_t)0, B_INVAL, cr);
916 		if (vn_has_cached_data(vp))
917 			return (0);
918 		rw_downgrade(&ip->i_contents);
919 		ufs_directio_kstats.nflushes.value.ui64++;
920 	}
921 	/*
922 	 * Direct Reads
923 	 */
924 
925 	/*
926 	 * proc and as are for VM operations in directio_start()
927 	 */
928 	if (uio->uio_segflg == UIO_USERSPACE) {
929 		procp = ttoproc(curthread);
930 		as = procp->p_as;
931 	} else {
932 		procp = NULL;
933 		as = &kas;
934 	}
935 
936 	*statusp = DIRECTIO_SUCCESS;
937 	error = 0;
938 	newerror = 0;
939 	bytes_read = 0;
940 	ufs_directio_kstats.logical_reads.value.ui64++;
941 	while (error == 0 && newerror == 0 && resid && uio->uio_iovcnt) {
942 		size_t pglck_len, pglck_size;
943 		caddr_t pglck_base;
944 		page_t **pplist, **spplist;
945 
946 		tail = NULL;
947 
948 		/*
949 		 * Adjust number of bytes
950 		 */
951 		iov = uio->uio_iov;
952 		pglck_len = (size_t)MIN(iov->iov_len, resid);
953 		pglck_base = iov->iov_base;
954 		if (pglck_len == 0) {
955 			uio->uio_iov++;
956 			uio->uio_iovcnt--;
957 			continue;
958 		}
959 
960 		/*
961 		 * Try to Lock down the largest chunck of pages possible.
962 		 */
963 		pglck_len = (size_t)MIN(pglck_len,  ufsvfsp->vfs_ioclustsz);
964 		error = as_pagelock(as, &pplist, pglck_base,
965 		    pglck_len, S_WRITE);
966 
967 		if (error)
968 			break;
969 
970 		pglck_size = pglck_len;
971 		while (pglck_len) {
972 
973 			nbytes = pglck_len;
974 			uoff = uio->uio_loffset;
975 
976 			/*
977 			 * Re-adjust number of bytes to contiguous range
978 			 */
979 			len = (ssize_t)blkroundup(fs, nbytes);
980 			error = bmap_read(ip, uoff, &bn, &len);
981 			if (error)
982 				break;
983 
984 			if (bn == UFS_HOLE) {
985 				nbytes = (size_t)MIN(fs->fs_bsize -
986 				    (long)blkoff(fs, uoff), nbytes);
987 				error = directio_hole(uio, nbytes);
988 				/*
989 				 * Hole reads are not added to the list
990 				 * processed by directio_wait() below so
991 				 * account for bytes read here.
992 				 */
993 				if (!error)
994 					bytes_read += nbytes;
995 			} else {
996 				nbytes = (size_t)MIN(nbytes, len);
997 
998 				/*
999 				 * Get the pagelist pointer for this offset
1000 				 * to be passed to directio_start.
1001 				 */
1002 				if (pplist != NULL)
1003 					spplist = pplist +
1004 					    btop((uintptr_t)iov->iov_base -
1005 					    ((uintptr_t)pglck_base & PAGEMASK));
1006 				else
1007 					spplist = NULL;
1008 
1009 				/*
1010 				 * Kick off the direct read requests
1011 				 */
1012 				directio_start(ufsvfsp, ip->i_dev, nbytes,
1013 				    ldbtob(bn), iov->iov_base,
1014 				    S_WRITE, procp, &tail, spplist);
1015 			}
1016 
1017 			if (error)
1018 				break;
1019 
1020 			/*
1021 			 * Adjust pointers and counters
1022 			 */
1023 			iov->iov_len -= nbytes;
1024 			iov->iov_base += nbytes;
1025 			uio->uio_loffset += nbytes;
1026 			resid -= nbytes;
1027 			pglck_len -= nbytes;
1028 		}
1029 
1030 		/*
1031 		 * Wait for outstanding requests
1032 		 */
1033 		newerror = directio_wait(tail, &bytes_read);
1034 		/*
1035 		 * Release VM resources
1036 		 */
1037 		as_pageunlock(as, pplist, pglck_base, pglck_size, S_WRITE);
1038 
1039 	}
1040 
1041 	/*
1042 	 * If error, adjust resid to begin at the first
1043 	 * un-read byte.
1044 	 */
1045 	if (error == 0)
1046 		error = newerror;
1047 	uio->uio_resid -= bytes_read;
1048 	return (error);
1049 }
1050