xref: /xnu-11215/bsd/vm/vnode_pager.c (revision 33de042d)
1 /*
2  * Copyright (c) 2000-2020 Apple Inc. All rights reserved.
3  *
4  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5  *
6  * This file contains Original Code and/or Modifications of Original Code
7  * as defined in and that are subject to the Apple Public Source License
8  * Version 2.0 (the 'License'). You may not use this file except in
9  * compliance with the License. The rights granted to you under the License
10  * may not be used to create, or enable the creation or redistribution of,
11  * unlawful or unlicensed copies of an Apple operating system, or to
12  * circumvent, violate, or enable the circumvention or violation of, any
13  * terms of an Apple operating system software license agreement.
14  *
15  * Please obtain a copy of the License at
16  * http://www.opensource.apple.com/apsl/ and read it before using this file.
17  *
18  * The Original Code and all software distributed under the License are
19  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23  * Please see the License for the specific language governing rights and
24  * limitations under the License.
25  *
26  * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27  */
28 /*
29  * Mach Operating System
30  * Copyright (c) 1987 Carnegie-Mellon University
31  * All rights reserved.  The CMU software License Agreement specifies
32  * the terms and conditions for use and redistribution.
33  */
34 /*
35  *	File:	vnode_pager.c
36  *
37  *	"Swap" pager that pages to/from vnodes.  Also
38  *	handles demand paging from files.
39  *
40  */
41 
42 #include <mach/boolean.h>
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/user.h>
46 #include <sys/proc.h>
47 #include <sys/kauth.h>
48 #include <sys/buf.h>
49 #include <sys/uio.h>
50 #include <sys/vnode_internal.h>
51 #include <sys/namei.h>
52 #include <sys/mount_internal.h> /* needs internal due to fhandle_t */
53 #include <sys/ubc_internal.h>
54 #include <sys/lock.h>
55 #include <sys/disk.h>           /* For DKIOC calls */
56 
57 #include <mach/mach_types.h>
58 #include <mach/memory_object_types.h>
59 #include <mach/vm_map.h>
60 #include <mach/mach_vm.h>
61 #include <mach/upl.h>
62 #include <mach/sdt.h>
63 
64 #include <vm/vm_map.h>
65 #include <vm/vm_kern.h>
66 #include <kern/zalloc.h>
67 #include <libkern/libkern.h>
68 
69 #include <vm/vnode_pager.h>
70 #include <vm/vm_pageout.h>
71 #include <vm/vm_ubc.h>
72 
73 #include <kern/assert.h>
74 #include <sys/kdebug.h>
75 #include <nfs/nfs.h>
76 
77 #include <vm/vm_protos_internal.h>
78 
79 #include <sys/kdebug.h>
80 #include <sys/kdebug_triage.h>
81 #include <vfs/vfs_disk_conditioner.h>
82 
83 void
vnode_pager_throttle(void)84 vnode_pager_throttle(void)
85 {
86 	if (current_uthread()->uu_lowpri_window) {
87 		throttle_lowpri_io(1);
88 	}
89 }
90 
91 boolean_t
vnode_pager_isSSD(vnode_t vp)92 vnode_pager_isSSD(vnode_t vp)
93 {
94 	return disk_conditioner_mount_is_ssd(vp->v_mount);
95 }
96 
97 #if FBDP_DEBUG_OBJECT_NO_PAGER
98 bool
vnode_pager_forced_unmount(vnode_t vp)99 vnode_pager_forced_unmount(vnode_t vp)
100 {
101 	mount_t mnt;
102 	mnt = vnode_mount(vp);
103 	if (!mnt) {
104 		return false;
105 	}
106 	return vfs_isforce(mnt);
107 }
108 #endif /* FBDP_DEBUG_OBJECT_NO_PAGER */
109 
110 #if CONFIG_IOSCHED
111 void
vnode_pager_issue_reprioritize_io(struct vnode * devvp,uint64_t blkno,uint32_t len,int priority)112 vnode_pager_issue_reprioritize_io(struct vnode *devvp, uint64_t blkno, uint32_t len, int priority)
113 {
114 	u_int32_t blocksize = 0;
115 	dk_extent_t extent;
116 	dk_set_tier_t set_tier;
117 	int error = 0;
118 
119 	error = VNOP_IOCTL(devvp, DKIOCGETBLOCKSIZE, (caddr_t)&blocksize, 0, vfs_context_kernel());
120 	if (error) {
121 		return;
122 	}
123 
124 	memset(&extent, 0, sizeof(dk_extent_t));
125 	memset(&set_tier, 0, sizeof(dk_set_tier_t));
126 
127 	extent.offset = blkno * (u_int64_t) blocksize;
128 	extent.length = len;
129 
130 	set_tier.extents = &extent;
131 	set_tier.extentsCount = 1;
132 	set_tier.tier = (uint8_t)priority;
133 
134 	error = VNOP_IOCTL(devvp, DKIOCSETTIER, (caddr_t)&set_tier, 0, vfs_context_kernel());
135 	return;
136 }
137 #endif
138 
139 void
vnode_pager_was_dirtied(struct vnode * vp,vm_object_offset_t s_offset,vm_object_offset_t e_offset)140 vnode_pager_was_dirtied(
141 	struct vnode            *vp,
142 	vm_object_offset_t      s_offset,
143 	vm_object_offset_t      e_offset)
144 {
145 	cluster_update_state(vp, s_offset, e_offset, TRUE);
146 }
147 
148 uint32_t
vnode_pager_isinuse(struct vnode * vp)149 vnode_pager_isinuse(struct vnode *vp)
150 {
151 	if (vp->v_usecount > vp->v_kusecount) {
152 		return 1;
153 	}
154 	return 0;
155 }
156 
157 uint32_t
vnode_pager_return_throttle_io_limit(struct vnode * vp,uint32_t * limit)158 vnode_pager_return_throttle_io_limit(struct vnode *vp, uint32_t *limit)
159 {
160 	return cluster_throttle_io_limit(vp, limit);
161 }
162 
163 vm_object_offset_t
vnode_pager_get_filesize(struct vnode * vp)164 vnode_pager_get_filesize(struct vnode *vp)
165 {
166 	return (vm_object_offset_t) ubc_getsize(vp);
167 }
168 
169 extern int safe_getpath(struct vnode *dvp, char *leafname, char *path, int _len, int *truncated_path);
170 
171 kern_return_t
vnode_pager_get_name(struct vnode * vp,char * pathname,vm_size_t pathname_len,char * filename,vm_size_t filename_len,boolean_t * truncated_path_p)172 vnode_pager_get_name(
173 	struct vnode    *vp,
174 	char            *pathname,
175 	vm_size_t       pathname_len,
176 	char            *filename,
177 	vm_size_t       filename_len,
178 	boolean_t       *truncated_path_p)
179 {
180 	*truncated_path_p = FALSE;
181 	if (pathname != NULL) {
182 		/* get the path name */
183 		safe_getpath(vp, NULL,
184 		    pathname, (int) pathname_len,
185 		    truncated_path_p);
186 	}
187 	if ((pathname == NULL || *truncated_path_p) &&
188 	    filename != NULL) {
189 		/* get the file name */
190 		const char *name;
191 
192 		name = vnode_getname_printable(vp);
193 		strlcpy(filename, name, (size_t) filename_len);
194 		vnode_putname_printable(name);
195 	}
196 	return KERN_SUCCESS;
197 }
198 
199 kern_return_t
vnode_pager_get_mtime(struct vnode * vp,struct timespec * current_mtime,struct timespec * cs_mtime)200 vnode_pager_get_mtime(
201 	struct vnode    *vp,
202 	struct timespec *current_mtime,
203 	struct timespec *cs_mtime)
204 {
205 	vnode_mtime(vp, current_mtime, vfs_context_current());
206 	if (cs_mtime != NULL) {
207 		ubc_get_cs_mtime(vp, cs_mtime);
208 	}
209 	return KERN_SUCCESS;
210 }
211 
212 kern_return_t
vnode_pager_get_cs_blobs(struct vnode * vp,void ** blobs)213 vnode_pager_get_cs_blobs(
214 	struct vnode    *vp,
215 	void            **blobs)
216 {
217 	*blobs = ubc_get_cs_blobs(vp);
218 	return KERN_SUCCESS;
219 }
220 
221 /*
222  * vnode_trim:
223  * Used to call the DKIOCUNMAP ioctl on the underlying disk device for the specified vnode.
224  * Trims the region at offset bytes into the file, for length bytes.
225  *
226  * Care must be taken to ensure that the vnode is sufficiently reference counted at the time this
227  * function is called; no iocounts or usecounts are taken on the vnode.
228  * This function is non-idempotent in error cases;  We cannot un-discard the blocks if only some of them
229  * are successfully discarded.
230  */
231 u_int32_t
vnode_trim(struct vnode * vp,off_t offset,size_t length)232 vnode_trim(
233 	struct vnode *vp,
234 	off_t offset,
235 	size_t length)
236 {
237 	daddr64_t io_blockno;    /* Block number corresponding to the start of the extent */
238 	size_t io_bytecount;    /* Number of bytes in current extent for the specified range */
239 	size_t trimmed = 0;
240 	off_t current_offset = offset;
241 	size_t remaining_length = length;
242 	int error = 0;
243 	u_int32_t blocksize = 0;
244 	struct vnode *devvp;
245 	dk_extent_t extent;
246 	dk_unmap_t unmap;
247 
248 
249 	/* Get the underlying device vnode */
250 	devvp = vp->v_mount->mnt_devvp;
251 
252 	/* Figure out the underlying device block size */
253 	error  = VNOP_IOCTL(devvp, DKIOCGETBLOCKSIZE, (caddr_t)&blocksize, 0, vfs_context_kernel());
254 	if (error) {
255 		goto trim_exit;
256 	}
257 
258 	/*
259 	 * We may not get the entire range from offset -> offset+length in a single
260 	 * extent from the blockmap call.  Keep looping/going until we are sure we've hit
261 	 * the whole range or if we encounter an error.
262 	 */
263 	while (trimmed < length) {
264 		/*
265 		 * VNOP_BLOCKMAP will tell us the logical to physical block number mapping for the
266 		 * specified offset.  It returns blocks in contiguous chunks, so if the logical range is
267 		 * broken into multiple extents, it must be called multiple times, increasing the offset
268 		 * in each call to ensure that the entire range is covered.
269 		 */
270 		error = VNOP_BLOCKMAP(vp, current_offset, remaining_length,
271 		    &io_blockno, &io_bytecount, NULL, VNODE_READ | VNODE_BLOCKMAP_NO_TRACK, NULL);
272 
273 		if (error) {
274 			goto trim_exit;
275 		}
276 		/*
277 		 * We have a contiguous run.  Prepare & issue the ioctl for the device.
278 		 * the DKIOCUNMAP ioctl takes offset in bytes from the start of the device.
279 		 */
280 		memset(&extent, 0, sizeof(dk_extent_t));
281 		memset(&unmap, 0, sizeof(dk_unmap_t));
282 		extent.offset = (uint64_t) io_blockno * (u_int64_t) blocksize;
283 		extent.length = io_bytecount;
284 		unmap.extents = &extent;
285 		unmap.extentsCount = 1;
286 		error = VNOP_IOCTL(devvp, DKIOCUNMAP, (caddr_t)&unmap, 0, vfs_context_kernel());
287 
288 		if (error) {
289 			goto trim_exit;
290 		}
291 		remaining_length = remaining_length - io_bytecount;
292 		trimmed = trimmed + io_bytecount;
293 		current_offset = current_offset + io_bytecount;
294 	}
295 trim_exit:
296 
297 	return error;
298 }
299 
300 pager_return_t
vnode_pageout(struct vnode * vp,upl_t upl,upl_offset_t upl_offset,vm_object_offset_t f_offset,upl_size_t size,int flags,int * errorp)301 vnode_pageout(struct vnode *vp,
302     upl_t                   upl,
303     upl_offset_t            upl_offset,
304     vm_object_offset_t      f_offset,
305     upl_size_t              size,
306     int                     flags,
307     int                     *errorp)
308 {
309 	int             result = PAGER_SUCCESS;
310 	int             error = 0;
311 	int             error_ret = 0;
312 	daddr64_t blkno;
313 	int isize;
314 	int pg_index;
315 	int base_index;
316 	upl_offset_t offset;
317 	upl_page_info_t *pl;
318 	vfs_context_t ctx = vfs_context_current();      /* pager context */
319 
320 	isize = (int)size;
321 
322 	/*
323 	 * This call is non-blocking and does not ever fail but it can
324 	 * only be made when there is other explicit synchronization
325 	 * with reclaiming of the vnode which, in this path, is provided
326 	 * by the paging in progress counter.
327 	 *
328 	 * In addition, this may also be entered via explicit ubc_msync
329 	 * calls or vm_swapfile_io where the existing iocount provides
330 	 * the necessary synchronization. Ideally we would not take an
331 	 * additional iocount here in the cases where an explcit iocount
332 	 * has already been taken but this call doesn't cause a deadlock
333 	 * as other forms of vnode_get* might if this thread has already
334 	 * taken an iocount.
335 	 */
336 	error = vnode_getalways_from_pager(vp);
337 	if (error != 0) {
338 		/* This can't happen */
339 		panic("vnode_getalways returned %d for vp %p", error, vp);
340 	}
341 
342 	if (isize <= 0) {
343 		result    = PAGER_ERROR;
344 		error_ret = EINVAL;
345 		goto out;
346 	}
347 
348 	if (UBCINFOEXISTS(vp) == 0) {
349 		result    = PAGER_ERROR;
350 		error_ret = EINVAL;
351 
352 		if (upl && !(flags & UPL_NOCOMMIT)) {
353 			ubc_upl_abort_range(upl, upl_offset, size, UPL_ABORT_FREE_ON_EMPTY);
354 		}
355 		goto out;
356 	}
357 	if (!(flags & UPL_VNODE_PAGER)) {
358 		/*
359 		 * This is a pageout from the default pager,
360 		 * just go ahead and call vnop_pageout since
361 		 * it has already sorted out the dirty ranges
362 		 */
363 		KDBG_RELEASE(
364 			VMDBG_CODE(DBG_VM_VNODE_PAGEOUT) | DBG_FUNC_START,
365 			size, 1);
366 
367 		if ((error_ret = VNOP_PAGEOUT(vp, upl, upl_offset, (off_t)f_offset,
368 		    (size_t)size, flags, ctx))) {
369 			result = PAGER_ERROR;
370 		}
371 
372 		KDBG_RELEASE(
373 			VMDBG_CODE(DBG_VM_VNODE_PAGEOUT) | DBG_FUNC_END,
374 			size, 1);
375 
376 		goto out;
377 	}
378 	if (upl == NULL) {
379 		int                     request_flags;
380 
381 		if (vp->v_mount->mnt_vtable->vfc_vfsflags & VFC_VFSVNOP_PAGEOUTV2) {
382 			/*
383 			 * filesystem has requested the new form of VNOP_PAGEOUT for file
384 			 * backed objects... we will not grab the UPL befofe calling VNOP_PAGEOUT...
385 			 * it is the fileystem's responsibility to grab the range we're denoting
386 			 * via 'f_offset' and 'size' into a UPL... this allows the filesystem to first
387 			 * take any locks it needs, before effectively locking the pages into a UPL...
388 			 */
389 			KDBG_RELEASE(VMDBG_CODE(DBG_VM_VNODE_PAGEOUT) | DBG_FUNC_START,
390 			    size, (int)f_offset);
391 
392 			if ((error_ret = VNOP_PAGEOUT(vp, NULL, upl_offset, (off_t)f_offset,
393 			    size, flags, ctx))) {
394 				result = PAGER_ERROR;
395 			}
396 			KDBG_RELEASE(VMDBG_CODE(DBG_VM_VNODE_PAGEOUT) | DBG_FUNC_END,
397 			    size);
398 
399 			goto out;
400 		}
401 		if (flags & UPL_MSYNC) {
402 			request_flags = UPL_UBC_MSYNC | UPL_RET_ONLY_DIRTY;
403 		} else {
404 			request_flags = UPL_UBC_PAGEOUT | UPL_RET_ONLY_DIRTY;
405 		}
406 
407 		if (ubc_create_upl_kernel(vp, f_offset, size, &upl, &pl, request_flags, VM_KERN_MEMORY_FILE) != KERN_SUCCESS) {
408 			result    = PAGER_ERROR;
409 			error_ret = EINVAL;
410 			goto out;
411 		}
412 		upl_offset = 0;
413 	} else {
414 		pl = ubc_upl_pageinfo(upl);
415 	}
416 
417 	/*
418 	 * Ignore any non-present pages at the end of the
419 	 * UPL so that we aren't looking at a upl that
420 	 * may already have been freed by the preceeding
421 	 * aborts/completions.
422 	 */
423 	base_index = upl_offset / PAGE_SIZE;
424 
425 	for (pg_index = (upl_offset + isize) / PAGE_SIZE; pg_index > base_index;) {
426 		if (upl_page_present(pl, --pg_index)) {
427 			break;
428 		}
429 		if (pg_index == base_index) {
430 			/*
431 			 * no pages were returned, so release
432 			 * our hold on the upl and leave
433 			 */
434 			if (!(flags & UPL_NOCOMMIT)) {
435 				ubc_upl_abort_range(upl, upl_offset, isize, UPL_ABORT_FREE_ON_EMPTY);
436 			}
437 
438 			goto out;
439 		}
440 	}
441 	isize = ((pg_index + 1) - base_index) * PAGE_SIZE;
442 
443 	/*
444 	 * we come here for pageouts to 'real' files and
445 	 * for msyncs...  the upl may not contain any
446 	 * dirty pages.. it's our responsibility to sort
447 	 * through it and find the 'runs' of dirty pages
448 	 * to call VNOP_PAGEOUT on...
449 	 */
450 
451 	if (ubc_getsize(vp) == 0) {
452 		/*
453 		 * if the file has been effectively deleted, then
454 		 * we need to go through the UPL and invalidate any
455 		 * buffer headers we might have that reference any
456 		 * of it's pages
457 		 */
458 		for (offset = upl_offset; isize; isize -= PAGE_SIZE, offset += PAGE_SIZE) {
459 			if (vp->v_tag == VT_NFS) {
460 				/* check with nfs if page is OK to drop */
461 				error = nfs_buf_page_inval(vp, (off_t)f_offset);
462 			} else {
463 				blkno = ubc_offtoblk(vp, (off_t)f_offset);
464 				error = buf_invalblkno(vp, blkno, 0);
465 			}
466 			if (error) {
467 				if (!(flags & UPL_NOCOMMIT)) {
468 					ubc_upl_abort_range(upl, offset, PAGE_SIZE, UPL_ABORT_FREE_ON_EMPTY);
469 				}
470 				if (error_ret == 0) {
471 					error_ret = error;
472 				}
473 				result = PAGER_ERROR;
474 			} else if (!(flags & UPL_NOCOMMIT)) {
475 				ubc_upl_commit_range(upl, offset, PAGE_SIZE, UPL_COMMIT_FREE_ON_EMPTY);
476 			}
477 			f_offset += PAGE_SIZE;
478 		}
479 		goto out;
480 	}
481 
482 	offset = upl_offset;
483 	pg_index = base_index;
484 
485 	while (isize) {
486 		int  xsize;
487 		int  num_of_pages;
488 
489 		if (!upl_page_present(pl, pg_index)) {
490 			/*
491 			 * we asked for RET_ONLY_DIRTY, so it's possible
492 			 * to get back empty slots in the UPL
493 			 * just skip over them
494 			 */
495 			f_offset += PAGE_SIZE;
496 			offset   += PAGE_SIZE;
497 			isize    -= PAGE_SIZE;
498 			pg_index++;
499 
500 			continue;
501 		}
502 		if (!upl_dirty_page(pl, pg_index)) {
503 			/*
504 			 * if the page is not dirty and reached here it is
505 			 * marked precious or it is due to invalidation in
506 			 * memory_object_lock request as part of truncation
507 			 * We also get here from vm_object_terminate()
508 			 * So all you need to do in these
509 			 * cases is to invalidate incore buffer if it is there
510 			 * Note we must not sleep here if the buffer is busy - that is
511 			 * a lock inversion which causes deadlock.
512 			 */
513 			if (vp->v_tag == VT_NFS) {
514 				/* check with nfs if page is OK to drop */
515 				error = nfs_buf_page_inval(vp, (off_t)f_offset);
516 			} else {
517 				blkno = ubc_offtoblk(vp, (off_t)f_offset);
518 				error = buf_invalblkno(vp, blkno, 0);
519 			}
520 			if (error) {
521 				if (!(flags & UPL_NOCOMMIT)) {
522 					ubc_upl_abort_range(upl, offset, PAGE_SIZE, UPL_ABORT_FREE_ON_EMPTY);
523 				}
524 				if (error_ret == 0) {
525 					error_ret = error;
526 				}
527 				result = PAGER_ERROR;
528 			} else if (!(flags & UPL_NOCOMMIT)) {
529 				ubc_upl_commit_range(upl, offset, PAGE_SIZE, UPL_COMMIT_FREE_ON_EMPTY);
530 			}
531 			f_offset += PAGE_SIZE;
532 			offset   += PAGE_SIZE;
533 			isize    -= PAGE_SIZE;
534 			pg_index++;
535 
536 			continue;
537 		}
538 		num_of_pages = 1;
539 		xsize = isize - PAGE_SIZE;
540 
541 		while (xsize) {
542 			if (!upl_dirty_page(pl, pg_index + num_of_pages)) {
543 				break;
544 			}
545 			num_of_pages++;
546 			xsize -= PAGE_SIZE;
547 		}
548 		xsize = num_of_pages * PAGE_SIZE;
549 
550 		KDBG_RELEASE(VMDBG_CODE(DBG_VM_VNODE_PAGEOUT) | DBG_FUNC_START,
551 		    xsize, (int)f_offset);
552 
553 		if ((error = VNOP_PAGEOUT(vp, upl, offset, (off_t)f_offset,
554 		    xsize, flags, ctx))) {
555 			if (error_ret == 0) {
556 				error_ret = error;
557 			}
558 			result = PAGER_ERROR;
559 		}
560 		KDBG_RELEASE(VMDBG_CODE(DBG_VM_VNODE_PAGEOUT) | DBG_FUNC_END,
561 		    xsize, error);
562 
563 		f_offset += xsize;
564 		offset   += xsize;
565 		isize    -= xsize;
566 		pg_index += num_of_pages;
567 	}
568 out:
569 	vnode_put_from_pager(vp);
570 
571 	if (errorp) {
572 		*errorp = error_ret;
573 	}
574 
575 	return result;
576 }
577 
578 static uint64_t
ktriage_encode_v_tag_and_error(vnode_t vp,int error)579 ktriage_encode_v_tag_and_error(vnode_t vp, int error)
580 {
581 	uint64_t tag = vp->v_tag;
582 	uint32_t ktriage_error = (uint32_t)error;
583 	return tag << 32 | ktriage_error;
584 }
585 
586 
587 pager_return_t
vnode_pagein(struct vnode * vp,upl_t upl,upl_offset_t upl_offset,vm_object_offset_t f_offset,upl_size_t size,int flags,int * errorp)588 vnode_pagein(
589 	struct vnode            *vp,
590 	upl_t                   upl,
591 	upl_offset_t            upl_offset,
592 	vm_object_offset_t      f_offset,
593 	upl_size_t              size,
594 	int                     flags,
595 	int                     *errorp)
596 {
597 	upl_page_info_t *pl;
598 	int             result = PAGER_SUCCESS;
599 	int             error = 0;
600 	int             pages_in_upl;
601 	int             start_pg;
602 	int             last_pg;
603 	int             first_pg;
604 	int             xsize;
605 	int             must_commit = 1;
606 	int             ignore_valid_page_check = 0;
607 
608 	if (flags & UPL_NOCOMMIT) {
609 		must_commit = 0;
610 	}
611 
612 	if (flags & UPL_IGNORE_VALID_PAGE_CHECK) {
613 		ignore_valid_page_check = 1;
614 	}
615 
616 	/*
617 	 * This call is non-blocking and does not ever fail but it can
618 	 * only be made when there is other explicit synchronization
619 	 * with reclaiming of the vnode which, in this path, is provided
620 	 * by the paging in progress counter.
621 	 *
622 	 * In addition, this may also be entered via vm_swapfile_io
623 	 * where the existing iocount provides the necessary synchronization.
624 	 * Ideally we would not take an additional iocount here in the cases
625 	 * where an explcit iocount has already been taken but this call
626 	 * doesn't cause a deadlock as other forms of vnode_get* might if
627 	 * this thread has already taken an iocount.
628 	 */
629 	error = vnode_getalways_from_pager(vp);
630 	if (error != 0) {
631 		/* This can't happen */
632 		panic("vnode_getalways returned %d for vp %p", error, vp);
633 	}
634 
635 	if (UBCINFOEXISTS(vp) == 0) {
636 		result = PAGER_ERROR;
637 		error  = PAGER_ERROR;
638 
639 		if (upl && must_commit) {
640 			ubc_upl_abort_range(upl, upl_offset, size, UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_ERROR);
641 		}
642 
643 		ktriage_record(thread_tid(current_thread()), KDBG_TRIAGE_EVENTID(KDBG_TRIAGE_SUBSYS_VM, KDBG_TRIAGE_RESERVED, KDBG_TRIAGE_VM_VNODEPAGEIN_NO_UBCINFO),
644 		    ktriage_encode_v_tag_and_error(vp, 0) /* arg */);
645 		goto out;
646 	}
647 	if (upl == (upl_t)NULL) {
648 		flags &= ~UPL_NOCOMMIT;
649 
650 		if (size > MAX_UPL_SIZE_BYTES) {
651 			result = PAGER_ERROR;
652 			error  = PAGER_ERROR;
653 			goto out;
654 		}
655 		if (vp->v_mount->mnt_vtable->vfc_vfsflags & VFC_VFSVNOP_PAGEINV2) {
656 			/*
657 			 * filesystem has requested the new form of VNOP_PAGEIN for file
658 			 * backed objects... we will not grab the UPL befofe calling VNOP_PAGEIN...
659 			 * it is the fileystem's responsibility to grab the range we're denoting
660 			 * via 'f_offset' and 'size' into a UPL... this allows the filesystem to first
661 			 * take any locks it needs, before effectively locking the pages into a UPL...
662 			 * so we pass a NULL into the filesystem instead of a UPL pointer... the 'upl_offset'
663 			 * is used to identify the "must have" page in the extent... the filesystem is free
664 			 * to clip the extent to better fit the underlying FS blocksize if it desires as
665 			 * long as it continues to include the "must have" page... 'f_offset' + 'upl_offset'
666 			 * identifies that page
667 			 */
668 			if ((error = VNOP_PAGEIN(vp, NULL, upl_offset, (off_t)f_offset,
669 			    size, flags, vfs_context_current()))) {
670 				set_thread_pagein_error(current_thread(), error);
671 				ktriage_record(thread_tid(current_thread()), KDBG_TRIAGE_EVENTID(KDBG_TRIAGE_SUBSYS_VM, KDBG_TRIAGE_RESERVED, KDBG_TRIAGE_VM_VNODEPAGEIN_FSPAGEIN_FAIL),
672 				    ktriage_encode_v_tag_and_error(vp, error) /* arg */);
673 				result = PAGER_ERROR;
674 				error  = PAGER_ERROR;
675 			}
676 			goto out;
677 		}
678 		ubc_create_upl_kernel(vp, f_offset, size, &upl, &pl, UPL_UBC_PAGEIN | UPL_RET_ONLY_ABSENT, VM_KERN_MEMORY_FILE);
679 
680 		if (upl == (upl_t)NULL) {
681 			result =  PAGER_ABSENT;
682 			error = PAGER_ABSENT;
683 			ktriage_record(thread_tid(current_thread()), KDBG_TRIAGE_EVENTID(KDBG_TRIAGE_SUBSYS_VM, KDBG_TRIAGE_RESERVED, KDBG_TRIAGE_VM_VNODEPAGEIN_NO_UPL),
684 			    ktriage_encode_v_tag_and_error(vp, 0) /* arg */);
685 			goto out;
686 		}
687 		ubc_upl_range_needed(upl, upl_offset / PAGE_SIZE, 1);
688 
689 		upl_offset = 0;
690 		first_pg = 0;
691 
692 		/*
693 		 * if we get here, we've created the upl and
694 		 * are responsible for commiting/aborting it
695 		 * regardless of what the caller has passed in
696 		 */
697 		must_commit = 1;
698 	} else {
699 		pl = ubc_upl_pageinfo(upl);
700 		first_pg = upl_offset / PAGE_SIZE;
701 	}
702 	pages_in_upl = size / PAGE_SIZE;
703 	DTRACE_VM2(pgpgin, int, pages_in_upl, (uint64_t *), NULL);
704 
705 	/*
706 	 * before we start marching forward, we must make sure we end on
707 	 * a present page, otherwise we will be working with a freed
708 	 * upl
709 	 */
710 	for (last_pg = pages_in_upl - 1; last_pg >= first_pg; last_pg--) {
711 		if (upl_page_present(pl, last_pg)) {
712 			break;
713 		}
714 		if (last_pg == first_pg) {
715 			/*
716 			 * empty UPL, no pages are present
717 			 */
718 			if (must_commit) {
719 				ubc_upl_abort_range(upl, upl_offset, size, UPL_ABORT_FREE_ON_EMPTY);
720 			}
721 			goto out;
722 		}
723 	}
724 	pages_in_upl = last_pg + 1;
725 	last_pg = first_pg;
726 
727 	while (last_pg < pages_in_upl) {
728 		/*
729 		 * skip over missing pages...
730 		 */
731 		for (; last_pg < pages_in_upl; last_pg++) {
732 			if (upl_page_present(pl, last_pg)) {
733 				break;
734 			}
735 		}
736 
737 		if (ignore_valid_page_check == 1) {
738 			start_pg = last_pg;
739 		} else {
740 			/*
741 			 * skip over 'valid' pages... we don't want to issue I/O for these
742 			 */
743 			for (start_pg = last_pg; last_pg < pages_in_upl; last_pg++) {
744 				if (!upl_valid_page(pl, last_pg)) {
745 					break;
746 				}
747 			}
748 		}
749 
750 		if (last_pg > start_pg) {
751 			/*
752 			 * we've found a range of valid pages
753 			 * if we've got COMMIT responsibility
754 			 * commit this range of pages back to the
755 			 * cache unchanged
756 			 */
757 			xsize = (last_pg - start_pg) * PAGE_SIZE;
758 
759 			if (must_commit) {
760 				ubc_upl_abort_range(upl, start_pg * PAGE_SIZE, xsize, UPL_ABORT_FREE_ON_EMPTY);
761 			}
762 		}
763 		if (last_pg == pages_in_upl) {
764 			/*
765 			 * we're done... all pages that were present
766 			 * have either had I/O issued on them or
767 			 * were aborted unchanged...
768 			 */
769 			break;
770 		}
771 
772 		if (!upl_page_present(pl, last_pg)) {
773 			/*
774 			 * we found a range of valid pages
775 			 * terminated by a missing page...
776 			 * bump index to the next page and continue on
777 			 */
778 			last_pg++;
779 			continue;
780 		}
781 		/*
782 		 * scan from the found invalid page looking for a valid
783 		 * or non-present page before the end of the upl is reached, if we
784 		 * find one, then it will be the last page of the request to
785 		 * 'cluster_io'
786 		 */
787 		for (start_pg = last_pg; last_pg < pages_in_upl; last_pg++) {
788 			if ((!ignore_valid_page_check && upl_valid_page(pl, last_pg)) || !upl_page_present(pl, last_pg)) {
789 				break;
790 			}
791 		}
792 		if (last_pg > start_pg) {
793 			int xoff;
794 			xsize = (last_pg - start_pg) * PAGE_SIZE;
795 			xoff  = start_pg * PAGE_SIZE;
796 
797 			if ((error = VNOP_PAGEIN(vp, upl, (upl_offset_t) xoff,
798 			    (off_t)f_offset + xoff,
799 			    xsize, flags, vfs_context_current()))) {
800 				/*
801 				 * Usually this UPL will be aborted/committed by the lower cluster layer.
802 				 *
803 				 * a)	In the case of decmpfs, however, we may return an error (EAGAIN) to avoid
804 				 *	a deadlock with another thread already inflating the file.
805 				 *
806 				 * b)	In the case of content protection, EPERM is a valid error and we should respect it.
807 				 *
808 				 * In those cases, we must take care of our UPL at this layer itself.
809 				 */
810 				if (must_commit) {
811 					if (error == EAGAIN) {
812 						ubc_upl_abort_range(upl, (upl_offset_t) xoff, xsize, UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_RESTART);
813 					}
814 					if (error == EPERM) {
815 						ubc_upl_abort_range(upl, (upl_offset_t) xoff, xsize, UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_ERROR);
816 					}
817 				}
818 				set_thread_pagein_error(current_thread(), error);
819 				ktriage_record(thread_tid(current_thread()), KDBG_TRIAGE_EVENTID(KDBG_TRIAGE_SUBSYS_VM, KDBG_TRIAGE_RESERVED, KDBG_TRIAGE_VM_VNODEPAGEIN_FSPAGEIN_FAIL),
820 				    ktriage_encode_v_tag_and_error(vp, error) /* arg */);
821 				result = PAGER_ERROR;
822 				error  = PAGER_ERROR;
823 			}
824 		}
825 	}
826 out:
827 	vnode_put_from_pager(vp);
828 
829 	if (errorp) {
830 		*errorp = result;
831 	}
832 
833 	return error;
834 }
835