xref: /freebsd-13-stable/sys/contrib/openzfs/module/zfs/zfs_vnops.c (revision e6c1e181ba7f666e02b073be104eb3e241097d83)
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 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
25  * Copyright (c) 2015 by Chunwei Chen. All rights reserved.
26  * Copyright 2017 Nexenta Systems, Inc.
27  */
28 
29 /* Portions Copyright 2007 Jeremy Teo */
30 /* Portions Copyright 2010 Robert Milkowski */
31 
32 #include <sys/types.h>
33 #include <sys/param.h>
34 #include <sys/time.h>
35 #include <sys/sysmacros.h>
36 #include <sys/vfs.h>
37 #include <sys/uio_impl.h>
38 #include <sys/file.h>
39 #include <sys/stat.h>
40 #include <sys/kmem.h>
41 #include <sys/cmn_err.h>
42 #include <sys/errno.h>
43 #include <sys/zfs_dir.h>
44 #include <sys/zfs_acl.h>
45 #include <sys/zfs_ioctl.h>
46 #include <sys/fs/zfs.h>
47 #include <sys/dmu.h>
48 #include <sys/dmu_objset.h>
49 #include <sys/spa.h>
50 #include <sys/txg.h>
51 #include <sys/dbuf.h>
52 #include <sys/policy.h>
53 #include <sys/zfs_vnops.h>
54 #include <sys/zfs_quota.h>
55 #include <sys/zfs_vfsops.h>
56 #include <sys/zfs_znode.h>
57 
58 
59 static ulong_t zfs_fsync_sync_cnt = 4;
60 
61 int
zfs_fsync(znode_t * zp,int syncflag,cred_t * cr)62 zfs_fsync(znode_t *zp, int syncflag, cred_t *cr)
63 {
64 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
65 
66 	(void) tsd_set(zfs_fsyncer_key, (void *)zfs_fsync_sync_cnt);
67 
68 	if (zfsvfs->z_os->os_sync != ZFS_SYNC_DISABLED) {
69 		ZFS_ENTER(zfsvfs);
70 		ZFS_VERIFY_ZP(zp);
71 		atomic_inc_32(&zp->z_sync_writes_cnt);
72 		zil_commit(zfsvfs->z_log, zp->z_id);
73 		atomic_dec_32(&zp->z_sync_writes_cnt);
74 		ZFS_EXIT(zfsvfs);
75 	}
76 	tsd_set(zfs_fsyncer_key, NULL);
77 
78 	return (0);
79 }
80 
81 
82 #if defined(SEEK_HOLE) && defined(SEEK_DATA)
83 /*
84  * Lseek support for finding holes (cmd == SEEK_HOLE) and
85  * data (cmd == SEEK_DATA). "off" is an in/out parameter.
86  */
87 static int
zfs_holey_common(znode_t * zp,ulong_t cmd,loff_t * off)88 zfs_holey_common(znode_t *zp, ulong_t cmd, loff_t *off)
89 {
90 	zfs_locked_range_t *lr;
91 	uint64_t noff = (uint64_t)*off; /* new offset */
92 	uint64_t file_sz;
93 	int error;
94 	boolean_t hole;
95 
96 	file_sz = zp->z_size;
97 	if (noff >= file_sz)  {
98 		return (SET_ERROR(ENXIO));
99 	}
100 
101 	if (cmd == F_SEEK_HOLE)
102 		hole = B_TRUE;
103 	else
104 		hole = B_FALSE;
105 
106 	/* Flush any mmap()'d data to disk */
107 	if (zn_has_cached_data(zp, 0, file_sz - 1))
108 		zn_flush_cached_data(zp, B_FALSE);
109 
110 	lr = zfs_rangelock_enter(&zp->z_rangelock, 0, UINT64_MAX, RL_READER);
111 	error = dmu_offset_next(ZTOZSB(zp)->z_os, zp->z_id, hole, &noff);
112 	zfs_rangelock_exit(lr);
113 
114 	if (error == ESRCH)
115 		return (SET_ERROR(ENXIO));
116 
117 	/* File was dirty, so fall back to using generic logic */
118 	if (error == EBUSY) {
119 		if (hole)
120 			*off = file_sz;
121 
122 		return (0);
123 	}
124 
125 	/*
126 	 * We could find a hole that begins after the logical end-of-file,
127 	 * because dmu_offset_next() only works on whole blocks.  If the
128 	 * EOF falls mid-block, then indicate that the "virtual hole"
129 	 * at the end of the file begins at the logical EOF, rather than
130 	 * at the end of the last block.
131 	 */
132 	if (noff > file_sz) {
133 		ASSERT(hole);
134 		noff = file_sz;
135 	}
136 
137 	if (noff < *off)
138 		return (error);
139 	*off = noff;
140 	return (error);
141 }
142 
143 int
zfs_holey(znode_t * zp,ulong_t cmd,loff_t * off)144 zfs_holey(znode_t *zp, ulong_t cmd, loff_t *off)
145 {
146 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
147 	int error;
148 
149 	ZFS_ENTER(zfsvfs);
150 	ZFS_VERIFY_ZP(zp);
151 
152 	error = zfs_holey_common(zp, cmd, off);
153 
154 	ZFS_EXIT(zfsvfs);
155 	return (error);
156 }
157 #endif /* SEEK_HOLE && SEEK_DATA */
158 
159 /*ARGSUSED*/
160 int
zfs_access(znode_t * zp,int mode,int flag,cred_t * cr)161 zfs_access(znode_t *zp, int mode, int flag, cred_t *cr)
162 {
163 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
164 	int error;
165 
166 	ZFS_ENTER(zfsvfs);
167 	ZFS_VERIFY_ZP(zp);
168 
169 	if (flag & V_ACE_MASK)
170 		error = zfs_zaccess(zp, mode, flag, B_FALSE, cr);
171 	else
172 		error = zfs_zaccess_rwx(zp, mode, flag, cr);
173 
174 	ZFS_EXIT(zfsvfs);
175 	return (error);
176 }
177 
178 static unsigned long zfs_vnops_read_chunk_size = 1024 * 1024; /* Tunable */
179 
180 /*
181  * Read bytes from specified file into supplied buffer.
182  *
183  *	IN:	zp	- inode of file to be read from.
184  *		uio	- structure supplying read location, range info,
185  *			  and return buffer.
186  *		ioflag	- O_SYNC flags; used to provide FRSYNC semantics.
187  *			  O_DIRECT flag; used to bypass page cache.
188  *		cr	- credentials of caller.
189  *
190  *	OUT:	uio	- updated offset and range, buffer filled.
191  *
192  *	RETURN:	0 on success, error code on failure.
193  *
194  * Side Effects:
195  *	inode - atime updated if byte count > 0
196  */
197 /* ARGSUSED */
198 int
zfs_read(struct znode * zp,zfs_uio_t * uio,int ioflag,cred_t * cr)199 zfs_read(struct znode *zp, zfs_uio_t *uio, int ioflag, cred_t *cr)
200 {
201 	int error = 0;
202 	boolean_t frsync = B_FALSE;
203 
204 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
205 	ZFS_ENTER(zfsvfs);
206 	ZFS_VERIFY_ZP(zp);
207 
208 	if (zp->z_pflags & ZFS_AV_QUARANTINED) {
209 		ZFS_EXIT(zfsvfs);
210 		return (SET_ERROR(EACCES));
211 	}
212 
213 	/* We don't copy out anything useful for directories. */
214 	if (Z_ISDIR(ZTOTYPE(zp))) {
215 		ZFS_EXIT(zfsvfs);
216 		return (SET_ERROR(EISDIR));
217 	}
218 
219 	/*
220 	 * Validate file offset
221 	 */
222 	if (zfs_uio_offset(uio) < (offset_t)0) {
223 		ZFS_EXIT(zfsvfs);
224 		return (SET_ERROR(EINVAL));
225 	}
226 
227 	/*
228 	 * Fasttrack empty reads
229 	 */
230 	if (zfs_uio_resid(uio) == 0) {
231 		ZFS_EXIT(zfsvfs);
232 		return (0);
233 	}
234 
235 #ifdef FRSYNC
236 	/*
237 	 * If we're in FRSYNC mode, sync out this znode before reading it.
238 	 * Only do this for non-snapshots.
239 	 *
240 	 * Some platforms do not support FRSYNC and instead map it
241 	 * to O_SYNC, which results in unnecessary calls to zil_commit. We
242 	 * only honor FRSYNC requests on platforms which support it.
243 	 */
244 	frsync = !!(ioflag & FRSYNC);
245 #endif
246 	if (zfsvfs->z_log &&
247 	    (frsync || zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS))
248 		zil_commit(zfsvfs->z_log, zp->z_id);
249 
250 	/*
251 	 * Lock the range against changes.
252 	 */
253 	zfs_locked_range_t *lr = zfs_rangelock_enter(&zp->z_rangelock,
254 	    zfs_uio_offset(uio), zfs_uio_resid(uio), RL_READER);
255 
256 	/*
257 	 * If we are reading past end-of-file we can skip
258 	 * to the end; but we might still need to set atime.
259 	 */
260 	if (zfs_uio_offset(uio) >= zp->z_size) {
261 		error = 0;
262 		goto out;
263 	}
264 
265 	ASSERT(zfs_uio_offset(uio) < zp->z_size);
266 #if defined(__linux__)
267 	ssize_t start_offset = zfs_uio_offset(uio);
268 #endif
269 	ssize_t n = MIN(zfs_uio_resid(uio), zp->z_size - zfs_uio_offset(uio));
270 	ssize_t start_resid = n;
271 
272 	while (n > 0) {
273 		ssize_t nbytes = MIN(n, zfs_vnops_read_chunk_size -
274 		    P2PHASE(zfs_uio_offset(uio), zfs_vnops_read_chunk_size));
275 #ifdef UIO_NOCOPY
276 		if (zfs_uio_segflg(uio) == UIO_NOCOPY)
277 			error = mappedread_sf(zp, nbytes, uio);
278 		else
279 #endif
280 		if (zn_has_cached_data(zp, zfs_uio_offset(uio),
281 		    zfs_uio_offset(uio) + nbytes - 1) && !(ioflag & O_DIRECT)) {
282 			error = mappedread(zp, nbytes, uio);
283 		} else {
284 			error = dmu_read_uio_dbuf(sa_get_db(zp->z_sa_hdl),
285 			    uio, nbytes);
286 		}
287 
288 		if (error) {
289 			/* convert checksum errors into IO errors */
290 			if (error == ECKSUM)
291 				error = SET_ERROR(EIO);
292 
293 #if defined(__linux__)
294 			/*
295 			 * if we actually read some bytes, bubbling EFAULT
296 			 * up to become EAGAIN isn't what we want here...
297 			 *
298 			 * ...on Linux, at least. On FBSD, doing this breaks.
299 			 */
300 			if (error == EFAULT &&
301 			    (zfs_uio_offset(uio) - start_offset) != 0)
302 				error = 0;
303 #endif
304 			break;
305 		}
306 
307 		n -= nbytes;
308 	}
309 
310 	int64_t nread = start_resid - n;
311 	dataset_kstats_update_read_kstats(&zfsvfs->z_kstat, nread);
312 	task_io_account_read(nread);
313 out:
314 	zfs_rangelock_exit(lr);
315 
316 	ZFS_ACCESSTIME_STAMP(zfsvfs, zp);
317 	ZFS_EXIT(zfsvfs);
318 	return (error);
319 }
320 
321 static void
zfs_clear_setid_bits_if_necessary(zfsvfs_t * zfsvfs,znode_t * zp,cred_t * cr,uint64_t * clear_setid_bits_txgp,dmu_tx_t * tx)322 zfs_clear_setid_bits_if_necessary(zfsvfs_t *zfsvfs, znode_t *zp, cred_t *cr,
323     uint64_t *clear_setid_bits_txgp, dmu_tx_t *tx)
324 {
325 	zilog_t *zilog = zfsvfs->z_log;
326 	const uint64_t uid = KUID_TO_SUID(ZTOUID(zp));
327 
328 	ASSERT(clear_setid_bits_txgp != NULL);
329 	ASSERT(tx != NULL);
330 
331 	/*
332 	 * Clear Set-UID/Set-GID bits on successful write if not
333 	 * privileged and at least one of the execute bits is set.
334 	 *
335 	 * It would be nice to do this after all writes have
336 	 * been done, but that would still expose the ISUID/ISGID
337 	 * to another app after the partial write is committed.
338 	 *
339 	 * Note: we don't call zfs_fuid_map_id() here because
340 	 * user 0 is not an ephemeral uid.
341 	 */
342 	mutex_enter(&zp->z_acl_lock);
343 	if ((zp->z_mode & (S_IXUSR | (S_IXUSR >> 3) | (S_IXUSR >> 6))) != 0 &&
344 	    (zp->z_mode & (S_ISUID | S_ISGID)) != 0 &&
345 	    secpolicy_vnode_setid_retain(zp, cr,
346 	    ((zp->z_mode & S_ISUID) != 0 && uid == 0)) != 0) {
347 		uint64_t newmode;
348 
349 		zp->z_mode &= ~(S_ISUID | S_ISGID);
350 		newmode = zp->z_mode;
351 		(void) sa_update(zp->z_sa_hdl, SA_ZPL_MODE(zfsvfs),
352 		    (void *)&newmode, sizeof (uint64_t), tx);
353 
354 		mutex_exit(&zp->z_acl_lock);
355 
356 		/*
357 		 * Make sure SUID/SGID bits will be removed when we replay the
358 		 * log. If the setid bits are keep coming back, don't log more
359 		 * than one TX_SETATTR per transaction group.
360 		 */
361 		if (*clear_setid_bits_txgp != dmu_tx_get_txg(tx)) {
362 			vattr_t va;
363 
364 			bzero(&va, sizeof (va));
365 			va.va_mask = AT_MODE;
366 			va.va_nodeid = zp->z_id;
367 			va.va_mode = newmode;
368 			zfs_log_setattr(zilog, tx, TX_SETATTR, zp, &va, AT_MODE,
369 			    NULL);
370 			*clear_setid_bits_txgp = dmu_tx_get_txg(tx);
371 		}
372 	} else {
373 		mutex_exit(&zp->z_acl_lock);
374 	}
375 }
376 
377 /*
378  * Write the bytes to a file.
379  *
380  *	IN:	zp	- znode of file to be written to.
381  *		uio	- structure supplying write location, range info,
382  *			  and data buffer.
383  *		ioflag	- O_APPEND flag set if in append mode.
384  *			  O_DIRECT flag; used to bypass page cache.
385  *		cr	- credentials of caller.
386  *
387  *	OUT:	uio	- updated offset and range.
388  *
389  *	RETURN:	0 if success
390  *		error code if failure
391  *
392  * Timestamps:
393  *	ip - ctime|mtime updated if byte count > 0
394  */
395 
396 /* ARGSUSED */
397 int
zfs_write(znode_t * zp,zfs_uio_t * uio,int ioflag,cred_t * cr)398 zfs_write(znode_t *zp, zfs_uio_t *uio, int ioflag, cred_t *cr)
399 {
400 	int error = 0, error1;
401 	ssize_t start_resid = zfs_uio_resid(uio);
402 	uint64_t clear_setid_bits_txg = 0;
403 
404 	/*
405 	 * Fasttrack empty write
406 	 */
407 	ssize_t n = start_resid;
408 	if (n == 0)
409 		return (0);
410 
411 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
412 	ZFS_ENTER(zfsvfs);
413 	ZFS_VERIFY_ZP(zp);
414 
415 	sa_bulk_attr_t bulk[4];
416 	int count = 0;
417 	uint64_t mtime[2], ctime[2];
418 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, &mtime, 16);
419 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, 16);
420 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_SIZE(zfsvfs), NULL,
421 	    &zp->z_size, 8);
422 	SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_FLAGS(zfsvfs), NULL,
423 	    &zp->z_pflags, 8);
424 
425 	/*
426 	 * Callers might not be able to detect properly that we are read-only,
427 	 * so check it explicitly here.
428 	 */
429 	if (zfs_is_readonly(zfsvfs)) {
430 		ZFS_EXIT(zfsvfs);
431 		return (SET_ERROR(EROFS));
432 	}
433 
434 	/*
435 	 * If immutable or not appending then return EPERM.
436 	 * Intentionally allow ZFS_READONLY through here.
437 	 * See zfs_zaccess_common()
438 	 */
439 	if ((zp->z_pflags & ZFS_IMMUTABLE) ||
440 	    ((zp->z_pflags & ZFS_APPENDONLY) && !(ioflag & O_APPEND) &&
441 	    (zfs_uio_offset(uio) < zp->z_size))) {
442 		ZFS_EXIT(zfsvfs);
443 		return (SET_ERROR(EPERM));
444 	}
445 
446 	/*
447 	 * Validate file offset
448 	 */
449 	offset_t woff = ioflag & O_APPEND ? zp->z_size : zfs_uio_offset(uio);
450 	if (woff < 0) {
451 		ZFS_EXIT(zfsvfs);
452 		return (SET_ERROR(EINVAL));
453 	}
454 
455 	const uint64_t max_blksz = zfsvfs->z_max_blksz;
456 
457 	/*
458 	 * Pre-fault the pages to ensure slow (eg NFS) pages
459 	 * don't hold up txg.
460 	 * Skip this if uio contains loaned arc_buf.
461 	 */
462 	if (zfs_uio_prefaultpages(MIN(n, max_blksz), uio)) {
463 		ZFS_EXIT(zfsvfs);
464 		return (SET_ERROR(EFAULT));
465 	}
466 
467 	/*
468 	 * If in append mode, set the io offset pointer to eof.
469 	 */
470 	zfs_locked_range_t *lr;
471 	if (ioflag & O_APPEND) {
472 		/*
473 		 * Obtain an appending range lock to guarantee file append
474 		 * semantics.  We reset the write offset once we have the lock.
475 		 */
476 		lr = zfs_rangelock_enter(&zp->z_rangelock, 0, n, RL_APPEND);
477 		woff = lr->lr_offset;
478 		if (lr->lr_length == UINT64_MAX) {
479 			/*
480 			 * We overlocked the file because this write will cause
481 			 * the file block size to increase.
482 			 * Note that zp_size cannot change with this lock held.
483 			 */
484 			woff = zp->z_size;
485 		}
486 		zfs_uio_setoffset(uio, woff);
487 	} else {
488 		/*
489 		 * Note that if the file block size will change as a result of
490 		 * this write, then this range lock will lock the entire file
491 		 * so that we can re-write the block safely.
492 		 */
493 		lr = zfs_rangelock_enter(&zp->z_rangelock, woff, n, RL_WRITER);
494 	}
495 
496 	if (zn_rlimit_fsize(zp, uio)) {
497 		zfs_rangelock_exit(lr);
498 		ZFS_EXIT(zfsvfs);
499 		return (SET_ERROR(EFBIG));
500 	}
501 
502 	const rlim64_t limit = MAXOFFSET_T;
503 
504 	if (woff >= limit) {
505 		zfs_rangelock_exit(lr);
506 		ZFS_EXIT(zfsvfs);
507 		return (SET_ERROR(EFBIG));
508 	}
509 
510 	if (n > limit - woff)
511 		n = limit - woff;
512 
513 	uint64_t end_size = MAX(zp->z_size, woff + n);
514 	zilog_t *zilog = zfsvfs->z_log;
515 
516 	const uint64_t uid = KUID_TO_SUID(ZTOUID(zp));
517 	const uint64_t gid = KGID_TO_SGID(ZTOGID(zp));
518 	const uint64_t projid = zp->z_projid;
519 
520 	/*
521 	 * Write the file in reasonable size chunks.  Each chunk is written
522 	 * in a separate transaction; this keeps the intent log records small
523 	 * and allows us to do more fine-grained space accounting.
524 	 */
525 	while (n > 0) {
526 		woff = zfs_uio_offset(uio);
527 
528 		if (zfs_id_overblockquota(zfsvfs, DMU_USERUSED_OBJECT, uid) ||
529 		    zfs_id_overblockquota(zfsvfs, DMU_GROUPUSED_OBJECT, gid) ||
530 		    (projid != ZFS_DEFAULT_PROJID &&
531 		    zfs_id_overblockquota(zfsvfs, DMU_PROJECTUSED_OBJECT,
532 		    projid))) {
533 			error = SET_ERROR(EDQUOT);
534 			break;
535 		}
536 
537 		arc_buf_t *abuf = NULL;
538 		if (n >= max_blksz && woff >= zp->z_size &&
539 		    P2PHASE(woff, max_blksz) == 0 &&
540 		    zp->z_blksz == max_blksz) {
541 			/*
542 			 * This write covers a full block.  "Borrow" a buffer
543 			 * from the dmu so that we can fill it before we enter
544 			 * a transaction.  This avoids the possibility of
545 			 * holding up the transaction if the data copy hangs
546 			 * up on a pagefault (e.g., from an NFS server mapping).
547 			 */
548 			size_t cbytes;
549 
550 			abuf = dmu_request_arcbuf(sa_get_db(zp->z_sa_hdl),
551 			    max_blksz);
552 			ASSERT(abuf != NULL);
553 			ASSERT(arc_buf_size(abuf) == max_blksz);
554 			if ((error = zfs_uiocopy(abuf->b_data, max_blksz,
555 			    UIO_WRITE, uio, &cbytes))) {
556 				dmu_return_arcbuf(abuf);
557 				break;
558 			}
559 			ASSERT3S(cbytes, ==, max_blksz);
560 		}
561 
562 		/*
563 		 * Start a transaction.
564 		 */
565 		dmu_tx_t *tx = dmu_tx_create(zfsvfs->z_os);
566 		dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE);
567 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)sa_get_db(zp->z_sa_hdl);
568 		DB_DNODE_ENTER(db);
569 		dmu_tx_hold_write_by_dnode(tx, DB_DNODE(db), woff,
570 		    MIN(n, max_blksz));
571 		DB_DNODE_EXIT(db);
572 		zfs_sa_upgrade_txholds(tx, zp);
573 		error = dmu_tx_assign(tx, TXG_WAIT);
574 		if (error) {
575 			dmu_tx_abort(tx);
576 			if (abuf != NULL)
577 				dmu_return_arcbuf(abuf);
578 			break;
579 		}
580 
581 		/*
582 		 * NB: We must call zfs_clear_setid_bits_if_necessary before
583 		 * committing the transaction!
584 		 */
585 
586 		/*
587 		 * If rangelock_enter() over-locked we grow the blocksize
588 		 * and then reduce the lock range.  This will only happen
589 		 * on the first iteration since rangelock_reduce() will
590 		 * shrink down lr_length to the appropriate size.
591 		 */
592 		if (lr->lr_length == UINT64_MAX) {
593 			uint64_t new_blksz;
594 
595 			if (zp->z_blksz > max_blksz) {
596 				/*
597 				 * File's blocksize is already larger than the
598 				 * "recordsize" property.  Only let it grow to
599 				 * the next power of 2.
600 				 */
601 				ASSERT(!ISP2(zp->z_blksz));
602 				new_blksz = MIN(end_size,
603 				    1 << highbit64(zp->z_blksz));
604 			} else {
605 				new_blksz = MIN(end_size, max_blksz);
606 			}
607 			zfs_grow_blocksize(zp, new_blksz, tx);
608 			zfs_rangelock_reduce(lr, woff, n);
609 		}
610 
611 		/*
612 		 * XXX - should we really limit each write to z_max_blksz?
613 		 * Perhaps we should use SPA_MAXBLOCKSIZE chunks?
614 		 */
615 		const ssize_t nbytes =
616 		    MIN(n, max_blksz - P2PHASE(woff, max_blksz));
617 
618 		ssize_t tx_bytes;
619 		if (abuf == NULL) {
620 			tx_bytes = zfs_uio_resid(uio);
621 			zfs_uio_fault_disable(uio, B_TRUE);
622 			error = dmu_write_uio_dbuf(sa_get_db(zp->z_sa_hdl),
623 			    uio, nbytes, tx);
624 			zfs_uio_fault_disable(uio, B_FALSE);
625 #ifdef __linux__
626 			if (error == EFAULT) {
627 				zfs_clear_setid_bits_if_necessary(zfsvfs, zp,
628 				    cr, &clear_setid_bits_txg, tx);
629 				dmu_tx_commit(tx);
630 				/*
631 				 * Account for partial writes before
632 				 * continuing the loop.
633 				 * Update needs to occur before the next
634 				 * zfs_uio_prefaultpages, or prefaultpages may
635 				 * error, and we may break the loop early.
636 				 */
637 				if (tx_bytes != zfs_uio_resid(uio))
638 					n -= tx_bytes - zfs_uio_resid(uio);
639 				if (zfs_uio_prefaultpages(MIN(n, max_blksz),
640 				    uio)) {
641 					break;
642 				}
643 				continue;
644 			}
645 #endif
646 			/*
647 			 * On FreeBSD, EFAULT should be propagated back to the
648 			 * VFS, which will handle faulting and will retry.
649 			 */
650 			if (error != 0 && error != EFAULT) {
651 				zfs_clear_setid_bits_if_necessary(zfsvfs, zp,
652 				    cr, &clear_setid_bits_txg, tx);
653 				dmu_tx_commit(tx);
654 				break;
655 			}
656 			tx_bytes -= zfs_uio_resid(uio);
657 		} else {
658 			/* Implied by abuf != NULL: */
659 			ASSERT3S(n, >=, max_blksz);
660 			ASSERT0(P2PHASE(woff, max_blksz));
661 			/*
662 			 * We can simplify nbytes to MIN(n, max_blksz) since
663 			 * P2PHASE(woff, max_blksz) is 0, and knowing
664 			 * n >= max_blksz lets us simplify further:
665 			 */
666 			ASSERT3S(nbytes, ==, max_blksz);
667 			/*
668 			 * Thus, we're writing a full block at a block-aligned
669 			 * offset and extending the file past EOF.
670 			 *
671 			 * dmu_assign_arcbuf_by_dbuf() will directly assign the
672 			 * arc buffer to a dbuf.
673 			 */
674 			error = dmu_assign_arcbuf_by_dbuf(
675 			    sa_get_db(zp->z_sa_hdl), woff, abuf, tx);
676 			if (error != 0) {
677 				/*
678 				 * XXX This might not be necessary if
679 				 * dmu_assign_arcbuf_by_dbuf is guaranteed
680 				 * to be atomic.
681 				 */
682 				zfs_clear_setid_bits_if_necessary(zfsvfs, zp,
683 				    cr, &clear_setid_bits_txg, tx);
684 				dmu_return_arcbuf(abuf);
685 				dmu_tx_commit(tx);
686 				break;
687 			}
688 			ASSERT3S(nbytes, <=, zfs_uio_resid(uio));
689 			zfs_uioskip(uio, nbytes);
690 			tx_bytes = nbytes;
691 		}
692 		if (tx_bytes &&
693 		    zn_has_cached_data(zp, woff, woff + tx_bytes - 1) &&
694 		    !(ioflag & O_DIRECT)) {
695 			update_pages(zp, woff, tx_bytes, zfsvfs->z_os);
696 		}
697 
698 		/*
699 		 * If we made no progress, we're done.  If we made even
700 		 * partial progress, update the znode and ZIL accordingly.
701 		 */
702 		if (tx_bytes == 0) {
703 			(void) sa_update(zp->z_sa_hdl, SA_ZPL_SIZE(zfsvfs),
704 			    (void *)&zp->z_size, sizeof (uint64_t), tx);
705 			dmu_tx_commit(tx);
706 			ASSERT(error != 0);
707 			break;
708 		}
709 
710 		zfs_clear_setid_bits_if_necessary(zfsvfs, zp, cr,
711 		    &clear_setid_bits_txg, tx);
712 
713 		zfs_tstamp_update_setup(zp, CONTENT_MODIFIED, mtime, ctime);
714 
715 		/*
716 		 * Update the file size (zp_size) if it has changed;
717 		 * account for possible concurrent updates.
718 		 */
719 		while ((end_size = zp->z_size) < zfs_uio_offset(uio)) {
720 			(void) atomic_cas_64(&zp->z_size, end_size,
721 			    zfs_uio_offset(uio));
722 			ASSERT(error == 0 || error == EFAULT);
723 		}
724 		/*
725 		 * If we are replaying and eof is non zero then force
726 		 * the file size to the specified eof. Note, there's no
727 		 * concurrency during replay.
728 		 */
729 		if (zfsvfs->z_replay && zfsvfs->z_replay_eof != 0)
730 			zp->z_size = zfsvfs->z_replay_eof;
731 
732 		error1 = sa_bulk_update(zp->z_sa_hdl, bulk, count, tx);
733 		if (error1 != 0)
734 			/* Avoid clobbering EFAULT. */
735 			error = error1;
736 
737 		/*
738 		 * NB: During replay, the TX_SETATTR record logged by
739 		 * zfs_clear_setid_bits_if_necessary must precede any of
740 		 * the TX_WRITE records logged here.
741 		 */
742 		zfs_log_write(zilog, tx, TX_WRITE, zp, woff, tx_bytes, ioflag,
743 		    NULL, NULL);
744 
745 		dmu_tx_commit(tx);
746 
747 		if (error != 0)
748 			break;
749 		ASSERT3S(tx_bytes, ==, nbytes);
750 		n -= nbytes;
751 
752 		if (n > 0) {
753 			if (zfs_uio_prefaultpages(MIN(n, max_blksz), uio)) {
754 				error = SET_ERROR(EFAULT);
755 				break;
756 			}
757 		}
758 	}
759 
760 	zfs_znode_update_vfs(zp);
761 	zfs_rangelock_exit(lr);
762 
763 	/*
764 	 * If we're in replay mode, or we made no progress, or the
765 	 * uio data is inaccessible return an error.  Otherwise, it's
766 	 * at least a partial write, so it's successful.
767 	 */
768 	if (zfsvfs->z_replay || zfs_uio_resid(uio) == start_resid ||
769 	    error == EFAULT) {
770 		ZFS_EXIT(zfsvfs);
771 		return (error);
772 	}
773 
774 	if (ioflag & (O_SYNC | O_DSYNC) ||
775 	    zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS)
776 		zil_commit(zilog, zp->z_id);
777 
778 	const int64_t nwritten = start_resid - zfs_uio_resid(uio);
779 	dataset_kstats_update_write_kstats(&zfsvfs->z_kstat, nwritten);
780 	task_io_account_write(nwritten);
781 
782 	ZFS_EXIT(zfsvfs);
783 	return (0);
784 }
785 
786 /*ARGSUSED*/
787 int
zfs_getsecattr(znode_t * zp,vsecattr_t * vsecp,int flag,cred_t * cr)788 zfs_getsecattr(znode_t *zp, vsecattr_t *vsecp, int flag, cred_t *cr)
789 {
790 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
791 	int error;
792 	boolean_t skipaclchk = (flag & ATTR_NOACLCHECK) ? B_TRUE : B_FALSE;
793 
794 	ZFS_ENTER(zfsvfs);
795 	ZFS_VERIFY_ZP(zp);
796 	error = zfs_getacl(zp, vsecp, skipaclchk, cr);
797 	ZFS_EXIT(zfsvfs);
798 
799 	return (error);
800 }
801 
802 /*ARGSUSED*/
803 int
zfs_setsecattr(znode_t * zp,vsecattr_t * vsecp,int flag,cred_t * cr)804 zfs_setsecattr(znode_t *zp, vsecattr_t *vsecp, int flag, cred_t *cr)
805 {
806 	zfsvfs_t *zfsvfs = ZTOZSB(zp);
807 	int error;
808 	boolean_t skipaclchk = (flag & ATTR_NOACLCHECK) ? B_TRUE : B_FALSE;
809 	zilog_t	*zilog = zfsvfs->z_log;
810 
811 	ZFS_ENTER(zfsvfs);
812 	ZFS_VERIFY_ZP(zp);
813 
814 	error = zfs_setacl(zp, vsecp, skipaclchk, cr);
815 
816 	if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS)
817 		zil_commit(zilog, 0);
818 
819 	ZFS_EXIT(zfsvfs);
820 	return (error);
821 }
822 
823 #ifdef ZFS_DEBUG
824 static int zil_fault_io = 0;
825 #endif
826 
827 static void zfs_get_done(zgd_t *zgd, int error);
828 
829 /*
830  * Get data to generate a TX_WRITE intent log record.
831  */
832 int
zfs_get_data(void * arg,uint64_t gen,lr_write_t * lr,char * buf,struct lwb * lwb,zio_t * zio)833 zfs_get_data(void *arg, uint64_t gen, lr_write_t *lr, char *buf,
834     struct lwb *lwb, zio_t *zio)
835 {
836 	zfsvfs_t *zfsvfs = arg;
837 	objset_t *os = zfsvfs->z_os;
838 	znode_t *zp;
839 	uint64_t object = lr->lr_foid;
840 	uint64_t offset = lr->lr_offset;
841 	uint64_t size = lr->lr_length;
842 	dmu_buf_t *db;
843 	zgd_t *zgd;
844 	int error = 0;
845 	uint64_t zp_gen;
846 
847 	ASSERT3P(lwb, !=, NULL);
848 	ASSERT3P(zio, !=, NULL);
849 	ASSERT3U(size, !=, 0);
850 
851 	/*
852 	 * Nothing to do if the file has been removed
853 	 */
854 	if (zfs_zget(zfsvfs, object, &zp) != 0)
855 		return (SET_ERROR(ENOENT));
856 	if (zp->z_unlinked) {
857 		/*
858 		 * Release the vnode asynchronously as we currently have the
859 		 * txg stopped from syncing.
860 		 */
861 		zfs_zrele_async(zp);
862 		return (SET_ERROR(ENOENT));
863 	}
864 	/* check if generation number matches */
865 	if (sa_lookup(zp->z_sa_hdl, SA_ZPL_GEN(zfsvfs), &zp_gen,
866 	    sizeof (zp_gen)) != 0) {
867 		zfs_zrele_async(zp);
868 		return (SET_ERROR(EIO));
869 	}
870 	if (zp_gen != gen) {
871 		zfs_zrele_async(zp);
872 		return (SET_ERROR(ENOENT));
873 	}
874 
875 	zgd = (zgd_t *)kmem_zalloc(sizeof (zgd_t), KM_SLEEP);
876 	zgd->zgd_lwb = lwb;
877 	zgd->zgd_private = zp;
878 
879 	/*
880 	 * Write records come in two flavors: immediate and indirect.
881 	 * For small writes it's cheaper to store the data with the
882 	 * log record (immediate); for large writes it's cheaper to
883 	 * sync the data and get a pointer to it (indirect) so that
884 	 * we don't have to write the data twice.
885 	 */
886 	if (buf != NULL) { /* immediate write */
887 		zgd->zgd_lr = zfs_rangelock_enter(&zp->z_rangelock,
888 		    offset, size, RL_READER);
889 		/* test for truncation needs to be done while range locked */
890 		if (offset >= zp->z_size) {
891 			error = SET_ERROR(ENOENT);
892 		} else {
893 			error = dmu_read(os, object, offset, size, buf,
894 			    DMU_READ_NO_PREFETCH);
895 		}
896 		ASSERT(error == 0 || error == ENOENT);
897 	} else { /* indirect write */
898 		/*
899 		 * Have to lock the whole block to ensure when it's
900 		 * written out and its checksum is being calculated
901 		 * that no one can change the data. We need to re-check
902 		 * blocksize after we get the lock in case it's changed!
903 		 */
904 		for (;;) {
905 			uint64_t blkoff;
906 			size = zp->z_blksz;
907 			blkoff = ISP2(size) ? P2PHASE(offset, size) : offset;
908 			offset -= blkoff;
909 			zgd->zgd_lr = zfs_rangelock_enter(&zp->z_rangelock,
910 			    offset, size, RL_READER);
911 			if (zp->z_blksz == size)
912 				break;
913 			offset += blkoff;
914 			zfs_rangelock_exit(zgd->zgd_lr);
915 		}
916 		/* test for truncation needs to be done while range locked */
917 		if (lr->lr_offset >= zp->z_size)
918 			error = SET_ERROR(ENOENT);
919 #ifdef ZFS_DEBUG
920 		if (zil_fault_io) {
921 			error = SET_ERROR(EIO);
922 			zil_fault_io = 0;
923 		}
924 #endif
925 		if (error == 0)
926 			error = dmu_buf_hold(os, object, offset, zgd, &db,
927 			    DMU_READ_NO_PREFETCH);
928 
929 		if (error == 0) {
930 			blkptr_t *bp = &lr->lr_blkptr;
931 
932 			zgd->zgd_db = db;
933 			zgd->zgd_bp = bp;
934 
935 			ASSERT(db->db_offset == offset);
936 			ASSERT(db->db_size == size);
937 
938 			error = dmu_sync(zio, lr->lr_common.lrc_txg,
939 			    zfs_get_done, zgd);
940 			ASSERT(error || lr->lr_length <= size);
941 
942 			/*
943 			 * On success, we need to wait for the write I/O
944 			 * initiated by dmu_sync() to complete before we can
945 			 * release this dbuf.  We will finish everything up
946 			 * in the zfs_get_done() callback.
947 			 */
948 			if (error == 0)
949 				return (0);
950 
951 			if (error == EALREADY) {
952 				lr->lr_common.lrc_txtype = TX_WRITE2;
953 				/*
954 				 * TX_WRITE2 relies on the data previously
955 				 * written by the TX_WRITE that caused
956 				 * EALREADY.  We zero out the BP because
957 				 * it is the old, currently-on-disk BP.
958 				 */
959 				zgd->zgd_bp = NULL;
960 				BP_ZERO(bp);
961 				error = 0;
962 			}
963 		}
964 	}
965 
966 	zfs_get_done(zgd, error);
967 
968 	return (error);
969 }
970 
971 
972 /* ARGSUSED */
973 static void
zfs_get_done(zgd_t * zgd,int error)974 zfs_get_done(zgd_t *zgd, int error)
975 {
976 	znode_t *zp = zgd->zgd_private;
977 
978 	if (zgd->zgd_db)
979 		dmu_buf_rele(zgd->zgd_db, zgd);
980 
981 	zfs_rangelock_exit(zgd->zgd_lr);
982 
983 	/*
984 	 * Release the vnode asynchronously as we currently have the
985 	 * txg stopped from syncing.
986 	 */
987 	zfs_zrele_async(zp);
988 
989 	kmem_free(zgd, sizeof (zgd_t));
990 }
991 
992 EXPORT_SYMBOL(zfs_access);
993 EXPORT_SYMBOL(zfs_fsync);
994 EXPORT_SYMBOL(zfs_holey);
995 EXPORT_SYMBOL(zfs_read);
996 EXPORT_SYMBOL(zfs_write);
997 EXPORT_SYMBOL(zfs_getsecattr);
998 EXPORT_SYMBOL(zfs_setsecattr);
999 
1000 ZFS_MODULE_PARAM(zfs_vnops, zfs_vnops_, read_chunk_size, ULONG, ZMOD_RW,
1001 	"Bytes to read per chunk");
1002