xref: /netbsd-src/sys/uvm/uvm_swap.c (revision 9fb66d812c00ebfb445c0b47dea128f32aa6fe96)
1 /*	$NetBSD: uvm_swap.c,v 1.203 2021/03/13 15:29:55 skrll Exp $	*/
2 
3 /*
4  * Copyright (c) 1995, 1996, 1997, 2009 Matthew R. Green
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
21  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
22  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
23  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
24  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  * from: NetBSD: vm_swap.c,v 1.52 1997/12/02 13:47:37 pk Exp
29  * from: Id: uvm_swap.c,v 1.1.2.42 1998/02/02 20:38:06 chuck Exp
30  */
31 
32 #include <sys/cdefs.h>
33 __KERNEL_RCSID(0, "$NetBSD: uvm_swap.c,v 1.203 2021/03/13 15:29:55 skrll Exp $");
34 
35 #include "opt_uvmhist.h"
36 #include "opt_compat_netbsd.h"
37 #include "opt_ddb.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/atomic.h>
42 #include <sys/buf.h>
43 #include <sys/bufq.h>
44 #include <sys/conf.h>
45 #include <sys/cprng.h>
46 #include <sys/proc.h>
47 #include <sys/namei.h>
48 #include <sys/disklabel.h>
49 #include <sys/errno.h>
50 #include <sys/kernel.h>
51 #include <sys/vnode.h>
52 #include <sys/file.h>
53 #include <sys/vmem.h>
54 #include <sys/blist.h>
55 #include <sys/mount.h>
56 #include <sys/pool.h>
57 #include <sys/kmem.h>
58 #include <sys/syscallargs.h>
59 #include <sys/swap.h>
60 #include <sys/kauth.h>
61 #include <sys/sysctl.h>
62 #include <sys/workqueue.h>
63 
64 #include <uvm/uvm.h>
65 
66 #include <miscfs/specfs/specdev.h>
67 
68 #include <crypto/aes/aes.h>
69 #include <crypto/aes/aes_cbc.h>
70 
71 /*
72  * uvm_swap.c: manage configuration and i/o to swap space.
73  */
74 
75 /*
76  * swap space is managed in the following way:
77  *
78  * each swap partition or file is described by a "swapdev" structure.
79  * each "swapdev" structure contains a "swapent" structure which contains
80  * information that is passed up to the user (via system calls).
81  *
82  * each swap partition is assigned a "priority" (int) which controls
83  * swap partition usage.
84  *
85  * the system maintains a global data structure describing all swap
86  * partitions/files.   there is a sorted LIST of "swappri" structures
87  * which describe "swapdev"'s at that priority.   this LIST is headed
88  * by the "swap_priority" global var.    each "swappri" contains a
89  * TAILQ of "swapdev" structures at that priority.
90  *
91  * locking:
92  *  - swap_syscall_lock (krwlock_t): this lock serializes the swapctl
93  *    system call and prevents the swap priority list from changing
94  *    while we are in the middle of a system call (e.g. SWAP_STATS).
95  *  - uvm_swap_data_lock (kmutex_t): this lock protects all swap data
96  *    structures including the priority list, the swapdev structures,
97  *    and the swapmap arena.
98  *
99  * each swap device has the following info:
100  *  - swap device in use (could be disabled, preventing future use)
101  *  - swap enabled (allows new allocations on swap)
102  *  - map info in /dev/drum
103  *  - vnode pointer
104  * for swap files only:
105  *  - block size
106  *  - max byte count in buffer
107  *  - buffer
108  *
109  * userland controls and configures swap with the swapctl(2) system call.
110  * the sys_swapctl performs the following operations:
111  *  [1] SWAP_NSWAP: returns the number of swap devices currently configured
112  *  [2] SWAP_STATS: given a pointer to an array of swapent structures
113  *	(passed in via "arg") of a size passed in via "misc" ... we load
114  *	the current swap config into the array. The actual work is done
115  *	in the uvm_swap_stats() function.
116  *  [3] SWAP_ON: given a pathname in arg (could be device or file) and a
117  *	priority in "misc", start swapping on it.
118  *  [4] SWAP_OFF: as SWAP_ON, but stops swapping to a device
119  *  [5] SWAP_CTL: changes the priority of a swap device (new priority in
120  *	"misc")
121  */
122 
123 /*
124  * swapdev: describes a single swap partition/file
125  *
126  * note the following should be true:
127  * swd_inuse <= swd_nblks  [number of blocks in use is <= total blocks]
128  * swd_nblks <= swd_mapsize [because mapsize includes miniroot+disklabel]
129  */
130 struct swapdev {
131 	dev_t			swd_dev;	/* device id */
132 	int			swd_flags;	/* flags:inuse/enable/fake */
133 	int			swd_priority;	/* our priority */
134 	int			swd_nblks;	/* blocks in this device */
135 	char			*swd_path;	/* saved pathname of device */
136 	int			swd_pathlen;	/* length of pathname */
137 	int			swd_npages;	/* #pages we can use */
138 	int			swd_npginuse;	/* #pages in use */
139 	int			swd_npgbad;	/* #pages bad */
140 	int			swd_drumoffset;	/* page0 offset in drum */
141 	int			swd_drumsize;	/* #pages in drum */
142 	blist_t			swd_blist;	/* blist for this swapdev */
143 	struct vnode		*swd_vp;	/* backing vnode */
144 	TAILQ_ENTRY(swapdev)	swd_next;	/* priority tailq */
145 
146 	int			swd_bsize;	/* blocksize (bytes) */
147 	int			swd_maxactive;	/* max active i/o reqs */
148 	struct bufq_state	*swd_tab;	/* buffer list */
149 	int			swd_active;	/* number of active buffers */
150 
151 	volatile uint32_t	*swd_encmap;	/* bitmap of encrypted slots */
152 	struct aesenc		swd_enckey;	/* AES key expanded for enc */
153 	struct aesdec		swd_deckey;	/* AES key expanded for dec */
154 	bool			swd_encinit;	/* true if keys initialized */
155 };
156 
157 /*
158  * swap device priority entry; the list is kept sorted on `spi_priority'.
159  */
160 struct swappri {
161 	int			spi_priority;     /* priority */
162 	TAILQ_HEAD(spi_swapdev, swapdev)	spi_swapdev;
163 	/* tailq of swapdevs at this priority */
164 	LIST_ENTRY(swappri)	spi_swappri;      /* global list of pri's */
165 };
166 
167 /*
168  * The following two structures are used to keep track of data transfers
169  * on swap devices associated with regular files.
170  * NOTE: this code is more or less a copy of vnd.c; we use the same
171  * structure names here to ease porting..
172  */
173 struct vndxfer {
174 	struct buf	*vx_bp;		/* Pointer to parent buffer */
175 	struct swapdev	*vx_sdp;
176 	int		vx_error;
177 	int		vx_pending;	/* # of pending aux buffers */
178 	int		vx_flags;
179 #define VX_BUSY		1
180 #define VX_DEAD		2
181 };
182 
183 struct vndbuf {
184 	struct buf	vb_buf;
185 	struct vndxfer	*vb_xfer;
186 };
187 
188 /*
189  * We keep a of pool vndbuf's and vndxfer structures.
190  */
191 static struct pool vndxfer_pool, vndbuf_pool;
192 
193 /*
194  * local variables
195  */
196 static vmem_t *swapmap;	/* controls the mapping of /dev/drum */
197 
198 /* list of all active swap devices [by priority] */
199 LIST_HEAD(swap_priority, swappri);
200 static struct swap_priority swap_priority;
201 
202 /* locks */
203 static kmutex_t uvm_swap_data_lock __cacheline_aligned;
204 static krwlock_t swap_syscall_lock;
205 
206 /* workqueue and use counter for swap to regular files */
207 static int sw_reg_count = 0;
208 static struct workqueue *sw_reg_workqueue;
209 
210 /* tuneables */
211 u_int uvm_swapisfull_factor = 99;
212 bool uvm_swap_encrypt = false;
213 
214 /*
215  * prototypes
216  */
217 static struct swapdev	*swapdrum_getsdp(int);
218 
219 static struct swapdev	*swaplist_find(struct vnode *, bool);
220 static void		 swaplist_insert(struct swapdev *,
221 					 struct swappri *, int);
222 static void		 swaplist_trim(void);
223 
224 static int swap_on(struct lwp *, struct swapdev *);
225 static int swap_off(struct lwp *, struct swapdev *);
226 
227 static void sw_reg_strategy(struct swapdev *, struct buf *, int);
228 static void sw_reg_biodone(struct buf *);
229 static void sw_reg_iodone(struct work *wk, void *dummy);
230 static void sw_reg_start(struct swapdev *);
231 
232 static int uvm_swap_io(struct vm_page **, int, int, int);
233 
234 static void uvm_swap_genkey(struct swapdev *);
235 static void uvm_swap_encryptpage(struct swapdev *, void *, int);
236 static void uvm_swap_decryptpage(struct swapdev *, void *, int);
237 
238 static size_t
239 encmap_size(size_t npages)
240 {
241 	struct swapdev *sdp;
242 	const size_t bytesperword = sizeof(sdp->swd_encmap[0]);
243 	const size_t bitsperword = NBBY * bytesperword;
244 	const size_t nbits = npages; /* one bit for each page */
245 	const size_t nwords = howmany(nbits, bitsperword);
246 	const size_t nbytes = nwords * bytesperword;
247 
248 	return nbytes;
249 }
250 
251 /*
252  * uvm_swap_init: init the swap system data structures and locks
253  *
254  * => called at boot time from init_main.c after the filesystems
255  *	are brought up (which happens after uvm_init())
256  */
257 void
258 uvm_swap_init(void)
259 {
260 	UVMHIST_FUNC(__func__);
261 
262 	UVMHIST_CALLED(pdhist);
263 	/*
264 	 * first, init the swap list, its counter, and its lock.
265 	 * then get a handle on the vnode for /dev/drum by using
266 	 * the its dev_t number ("swapdev", from MD conf.c).
267 	 */
268 
269 	LIST_INIT(&swap_priority);
270 	uvmexp.nswapdev = 0;
271 	rw_init(&swap_syscall_lock);
272 	mutex_init(&uvm_swap_data_lock, MUTEX_DEFAULT, IPL_NONE);
273 
274 	if (bdevvp(swapdev, &swapdev_vp))
275 		panic("%s: can't get vnode for swap device", __func__);
276 	if (vn_lock(swapdev_vp, LK_EXCLUSIVE | LK_RETRY))
277 		panic("%s: can't lock swap device", __func__);
278 	if (VOP_OPEN(swapdev_vp, FREAD | FWRITE, NOCRED))
279 		panic("%s: can't open swap device", __func__);
280 	VOP_UNLOCK(swapdev_vp);
281 
282 	/*
283 	 * create swap block resource map to map /dev/drum.   the range
284 	 * from 1 to INT_MAX allows 2 gigablocks of swap space.  note
285 	 * that block 0 is reserved (used to indicate an allocation
286 	 * failure, or no allocation).
287 	 */
288 	swapmap = vmem_create("swapmap", 1, INT_MAX - 1, 1, NULL, NULL, NULL, 0,
289 	    VM_NOSLEEP, IPL_NONE);
290 	if (swapmap == 0) {
291 		panic("%s: vmem_create failed", __func__);
292 	}
293 
294 	pool_init(&vndxfer_pool, sizeof(struct vndxfer), 0, 0, 0, "swp vnx",
295 	    NULL, IPL_BIO);
296 	pool_init(&vndbuf_pool, sizeof(struct vndbuf), 0, 0, 0, "swp vnd",
297 	    NULL, IPL_BIO);
298 
299 	UVMHIST_LOG(pdhist, "<- done", 0, 0, 0, 0);
300 }
301 
302 /*
303  * swaplist functions: functions that operate on the list of swap
304  * devices on the system.
305  */
306 
307 /*
308  * swaplist_insert: insert swap device "sdp" into the global list
309  *
310  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
311  * => caller must provide a newly allocated swappri structure (we will
312  *	FREE it if we don't need it... this it to prevent allocation
313  *	blocking here while adding swap)
314  */
315 static void
316 swaplist_insert(struct swapdev *sdp, struct swappri *newspp, int priority)
317 {
318 	struct swappri *spp, *pspp;
319 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
320 
321 	KASSERT(rw_write_held(&swap_syscall_lock));
322 	KASSERT(mutex_owned(&uvm_swap_data_lock));
323 
324 	/*
325 	 * find entry at or after which to insert the new device.
326 	 */
327 	pspp = NULL;
328 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
329 		if (priority <= spp->spi_priority)
330 			break;
331 		pspp = spp;
332 	}
333 
334 	/*
335 	 * new priority?
336 	 */
337 	if (spp == NULL || spp->spi_priority != priority) {
338 		spp = newspp;  /* use newspp! */
339 		UVMHIST_LOG(pdhist, "created new swappri = %jd",
340 			    priority, 0, 0, 0);
341 
342 		spp->spi_priority = priority;
343 		TAILQ_INIT(&spp->spi_swapdev);
344 
345 		if (pspp)
346 			LIST_INSERT_AFTER(pspp, spp, spi_swappri);
347 		else
348 			LIST_INSERT_HEAD(&swap_priority, spp, spi_swappri);
349 	} else {
350 	  	/* we don't need a new priority structure, free it */
351 		kmem_free(newspp, sizeof(*newspp));
352 	}
353 
354 	/*
355 	 * priority found (or created).   now insert on the priority's
356 	 * tailq list and bump the total number of swapdevs.
357 	 */
358 	sdp->swd_priority = priority;
359 	TAILQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
360 	uvmexp.nswapdev++;
361 }
362 
363 /*
364  * swaplist_find: find and optionally remove a swap device from the
365  *	global list.
366  *
367  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
368  * => we return the swapdev we found (and removed)
369  */
370 static struct swapdev *
371 swaplist_find(struct vnode *vp, bool remove)
372 {
373 	struct swapdev *sdp;
374 	struct swappri *spp;
375 
376 	KASSERT(rw_lock_held(&swap_syscall_lock));
377 	KASSERT(remove ? rw_write_held(&swap_syscall_lock) : 1);
378 	KASSERT(mutex_owned(&uvm_swap_data_lock));
379 
380 	/*
381 	 * search the lists for the requested vp
382 	 */
383 
384 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
385 		TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
386 			if (sdp->swd_vp == vp) {
387 				if (remove) {
388 					TAILQ_REMOVE(&spp->spi_swapdev,
389 					    sdp, swd_next);
390 					uvmexp.nswapdev--;
391 				}
392 				return(sdp);
393 			}
394 		}
395 	}
396 	return (NULL);
397 }
398 
399 /*
400  * swaplist_trim: scan priority list for empty priority entries and kill
401  *	them.
402  *
403  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
404  */
405 static void
406 swaplist_trim(void)
407 {
408 	struct swappri *spp, *nextspp;
409 
410 	KASSERT(rw_write_held(&swap_syscall_lock));
411 	KASSERT(mutex_owned(&uvm_swap_data_lock));
412 
413 	LIST_FOREACH_SAFE(spp, &swap_priority, spi_swappri, nextspp) {
414 		if (!TAILQ_EMPTY(&spp->spi_swapdev))
415 			continue;
416 		LIST_REMOVE(spp, spi_swappri);
417 		kmem_free(spp, sizeof(*spp));
418 	}
419 }
420 
421 /*
422  * swapdrum_getsdp: given a page offset in /dev/drum, convert it back
423  *	to the "swapdev" that maps that section of the drum.
424  *
425  * => each swapdev takes one big contig chunk of the drum
426  * => caller must hold uvm_swap_data_lock
427  */
428 static struct swapdev *
429 swapdrum_getsdp(int pgno)
430 {
431 	struct swapdev *sdp;
432 	struct swappri *spp;
433 
434 	KASSERT(mutex_owned(&uvm_swap_data_lock));
435 
436 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
437 		TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
438 			if (sdp->swd_flags & SWF_FAKE)
439 				continue;
440 			if (pgno >= sdp->swd_drumoffset &&
441 			    pgno < (sdp->swd_drumoffset + sdp->swd_drumsize)) {
442 				return sdp;
443 			}
444 		}
445 	}
446 	return NULL;
447 }
448 
449 /*
450  * swapdrum_sdp_is: true iff the swap device for pgno is sdp
451  *
452  * => for use in positive assertions only; result is not stable
453  */
454 static bool __debugused
455 swapdrum_sdp_is(int pgno, struct swapdev *sdp)
456 {
457 	bool result;
458 
459 	mutex_enter(&uvm_swap_data_lock);
460 	result = swapdrum_getsdp(pgno) == sdp;
461 	mutex_exit(&uvm_swap_data_lock);
462 
463 	return result;
464 }
465 
466 void swapsys_lock(krw_t op)
467 {
468 	rw_enter(&swap_syscall_lock, op);
469 }
470 
471 void swapsys_unlock(void)
472 {
473 	rw_exit(&swap_syscall_lock);
474 }
475 
476 static void
477 swapent_cvt(struct swapent *se, const struct swapdev *sdp, int inuse)
478 {
479 	se->se_dev = sdp->swd_dev;
480 	se->se_flags = sdp->swd_flags;
481 	se->se_nblks = sdp->swd_nblks;
482 	se->se_inuse = inuse;
483 	se->se_priority = sdp->swd_priority;
484 	KASSERT(sdp->swd_pathlen < sizeof(se->se_path));
485 	strcpy(se->se_path, sdp->swd_path);
486 }
487 
488 int (*uvm_swap_stats13)(const struct sys_swapctl_args *, register_t *) =
489     (void *)enosys;
490 int (*uvm_swap_stats50)(const struct sys_swapctl_args *, register_t *) =
491     (void *)enosys;
492 
493 /*
494  * sys_swapctl: main entry point for swapctl(2) system call
495  * 	[with two helper functions: swap_on and swap_off]
496  */
497 int
498 sys_swapctl(struct lwp *l, const struct sys_swapctl_args *uap, register_t *retval)
499 {
500 	/* {
501 		syscallarg(int) cmd;
502 		syscallarg(void *) arg;
503 		syscallarg(int) misc;
504 	} */
505 	struct vnode *vp;
506 	struct nameidata nd;
507 	struct swappri *spp;
508 	struct swapdev *sdp;
509 #define SWAP_PATH_MAX (PATH_MAX + 1)
510 	char	*userpath;
511 	size_t	len = 0;
512 	int	error;
513 	int	priority;
514 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
515 
516 	/*
517 	 * we handle the non-priv NSWAP and STATS request first.
518 	 *
519 	 * SWAP_NSWAP: return number of config'd swap devices
520 	 * [can also be obtained with uvmexp sysctl]
521 	 */
522 	if (SCARG(uap, cmd) == SWAP_NSWAP) {
523 		const int nswapdev = uvmexp.nswapdev;
524 		UVMHIST_LOG(pdhist, "<- done SWAP_NSWAP=%jd", nswapdev,
525 		    0, 0, 0);
526 		*retval = nswapdev;
527 		return 0;
528 	}
529 
530 	userpath = kmem_alloc(SWAP_PATH_MAX, KM_SLEEP);
531 
532 	/*
533 	 * ensure serialized syscall access by grabbing the swap_syscall_lock
534 	 */
535 	rw_enter(&swap_syscall_lock, RW_WRITER);
536 
537 	/*
538 	 * SWAP_STATS: get stats on current # of configured swap devs
539 	 *
540 	 * note that the swap_priority list can't change as long
541 	 * as we are holding the swap_syscall_lock.  we don't want
542 	 * to grab the uvm_swap_data_lock because we may fault&sleep during
543 	 * copyout() and we don't want to be holding that lock then!
544 	 */
545 	switch (SCARG(uap, cmd)) {
546 	case SWAP_STATS13:
547 		error = (*uvm_swap_stats13)(uap, retval);
548 		goto out;
549 	case SWAP_STATS50:
550 		error = (*uvm_swap_stats50)(uap, retval);
551 		goto out;
552 	case SWAP_STATS:
553 		error = uvm_swap_stats(SCARG(uap, arg), SCARG(uap, misc),
554 		    NULL, sizeof(struct swapent), retval);
555 		UVMHIST_LOG(pdhist, "<- done SWAP_STATS", 0, 0, 0, 0);
556 		goto out;
557 
558 	case SWAP_GETDUMPDEV:
559 		error = copyout(&dumpdev, SCARG(uap, arg), sizeof(dumpdev));
560 		goto out;
561 	default:
562 		break;
563 	}
564 
565 	/*
566 	 * all other requests require superuser privs.   verify.
567 	 */
568 	if ((error = kauth_authorize_system(l->l_cred, KAUTH_SYSTEM_SWAPCTL,
569 	    0, NULL, NULL, NULL)))
570 		goto out;
571 
572 	if (SCARG(uap, cmd) == SWAP_DUMPOFF) {
573 		/* drop the current dump device */
574 		dumpdev = NODEV;
575 		dumpcdev = NODEV;
576 		cpu_dumpconf();
577 		goto out;
578 	}
579 
580 	/*
581 	 * at this point we expect a path name in arg.   we will
582 	 * use namei() to gain a vnode reference (vref), and lock
583 	 * the vnode (VOP_LOCK).
584 	 *
585 	 * XXX: a NULL arg means use the root vnode pointer (e.g. for
586 	 * miniroot)
587 	 */
588 	if (SCARG(uap, arg) == NULL) {
589 		vp = rootvp;		/* miniroot */
590 		vref(vp);
591 		if (vn_lock(vp, LK_EXCLUSIVE)) {
592 			vrele(vp);
593 			error = EBUSY;
594 			goto out;
595 		}
596 		if (SCARG(uap, cmd) == SWAP_ON &&
597 		    copystr("miniroot", userpath, SWAP_PATH_MAX, &len))
598 			panic("swapctl: miniroot copy failed");
599 	} else {
600 		struct pathbuf *pb;
601 
602 		/*
603 		 * This used to allow copying in one extra byte
604 		 * (SWAP_PATH_MAX instead of PATH_MAX) for SWAP_ON.
605 		 * This was completely pointless because if anyone
606 		 * used that extra byte namei would fail with
607 		 * ENAMETOOLONG anyway, so I've removed the excess
608 		 * logic. - dholland 20100215
609 		 */
610 
611 		error = pathbuf_copyin(SCARG(uap, arg), &pb);
612 		if (error) {
613 			goto out;
614 		}
615 		if (SCARG(uap, cmd) == SWAP_ON) {
616 			/* get a copy of the string */
617 			pathbuf_copystring(pb, userpath, SWAP_PATH_MAX);
618 			len = strlen(userpath) + 1;
619 		}
620 		NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | TRYEMULROOT, pb);
621 		if ((error = namei(&nd))) {
622 			pathbuf_destroy(pb);
623 			goto out;
624 		}
625 		vp = nd.ni_vp;
626 		pathbuf_destroy(pb);
627 	}
628 	/* note: "vp" is referenced and locked */
629 
630 	error = 0;		/* assume no error */
631 	switch(SCARG(uap, cmd)) {
632 
633 	case SWAP_DUMPDEV:
634 		if (vp->v_type != VBLK) {
635 			error = ENOTBLK;
636 			break;
637 		}
638 		if (bdevsw_lookup(vp->v_rdev)) {
639 			dumpdev = vp->v_rdev;
640 			dumpcdev = devsw_blk2chr(dumpdev);
641 		} else
642 			dumpdev = NODEV;
643 		cpu_dumpconf();
644 		break;
645 
646 	case SWAP_CTL:
647 		/*
648 		 * get new priority, remove old entry (if any) and then
649 		 * reinsert it in the correct place.  finally, prune out
650 		 * any empty priority structures.
651 		 */
652 		priority = SCARG(uap, misc);
653 		spp = kmem_alloc(sizeof(*spp), KM_SLEEP);
654 		mutex_enter(&uvm_swap_data_lock);
655 		if ((sdp = swaplist_find(vp, true)) == NULL) {
656 			error = ENOENT;
657 		} else {
658 			swaplist_insert(sdp, spp, priority);
659 			swaplist_trim();
660 		}
661 		mutex_exit(&uvm_swap_data_lock);
662 		if (error)
663 			kmem_free(spp, sizeof(*spp));
664 		break;
665 
666 	case SWAP_ON:
667 
668 		/*
669 		 * check for duplicates.   if none found, then insert a
670 		 * dummy entry on the list to prevent someone else from
671 		 * trying to enable this device while we are working on
672 		 * it.
673 		 */
674 
675 		priority = SCARG(uap, misc);
676 		sdp = kmem_zalloc(sizeof(*sdp), KM_SLEEP);
677 		spp = kmem_alloc(sizeof(*spp), KM_SLEEP);
678 		sdp->swd_flags = SWF_FAKE;
679 		sdp->swd_vp = vp;
680 		sdp->swd_dev = (vp->v_type == VBLK) ? vp->v_rdev : NODEV;
681 		bufq_alloc(&sdp->swd_tab, "disksort", BUFQ_SORT_RAWBLOCK);
682 		mutex_enter(&uvm_swap_data_lock);
683 		if (swaplist_find(vp, false) != NULL) {
684 			error = EBUSY;
685 			mutex_exit(&uvm_swap_data_lock);
686 			bufq_free(sdp->swd_tab);
687 			kmem_free(sdp, sizeof(*sdp));
688 			kmem_free(spp, sizeof(*spp));
689 			break;
690 		}
691 		swaplist_insert(sdp, spp, priority);
692 		mutex_exit(&uvm_swap_data_lock);
693 
694 		KASSERT(len > 0);
695 		sdp->swd_pathlen = len;
696 		sdp->swd_path = kmem_alloc(len, KM_SLEEP);
697 		if (copystr(userpath, sdp->swd_path, len, 0) != 0)
698 			panic("swapctl: copystr");
699 
700 		/*
701 		 * we've now got a FAKE placeholder in the swap list.
702 		 * now attempt to enable swap on it.  if we fail, undo
703 		 * what we've done and kill the fake entry we just inserted.
704 		 * if swap_on is a success, it will clear the SWF_FAKE flag
705 		 */
706 
707 		if ((error = swap_on(l, sdp)) != 0) {
708 			mutex_enter(&uvm_swap_data_lock);
709 			(void) swaplist_find(vp, true);  /* kill fake entry */
710 			swaplist_trim();
711 			mutex_exit(&uvm_swap_data_lock);
712 			bufq_free(sdp->swd_tab);
713 			kmem_free(sdp->swd_path, sdp->swd_pathlen);
714 			kmem_free(sdp, sizeof(*sdp));
715 			break;
716 		}
717 		break;
718 
719 	case SWAP_OFF:
720 		mutex_enter(&uvm_swap_data_lock);
721 		if ((sdp = swaplist_find(vp, false)) == NULL) {
722 			mutex_exit(&uvm_swap_data_lock);
723 			error = ENXIO;
724 			break;
725 		}
726 
727 		/*
728 		 * If a device isn't in use or enabled, we
729 		 * can't stop swapping from it (again).
730 		 */
731 		if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0) {
732 			mutex_exit(&uvm_swap_data_lock);
733 			error = EBUSY;
734 			break;
735 		}
736 
737 		/*
738 		 * do the real work.
739 		 */
740 		error = swap_off(l, sdp);
741 		break;
742 
743 	default:
744 		error = EINVAL;
745 	}
746 
747 	/*
748 	 * done!  release the ref gained by namei() and unlock.
749 	 */
750 	vput(vp);
751 out:
752 	rw_exit(&swap_syscall_lock);
753 	kmem_free(userpath, SWAP_PATH_MAX);
754 
755 	UVMHIST_LOG(pdhist, "<- done!  error=%jd", error, 0, 0, 0);
756 	return (error);
757 }
758 
759 /*
760  * uvm_swap_stats: implements swapctl(SWAP_STATS). The function is kept
761  * away from sys_swapctl() in order to allow COMPAT_* swapctl()
762  * emulation to use it directly without going through sys_swapctl().
763  * The problem with using sys_swapctl() there is that it involves
764  * copying the swapent array to the stackgap, and this array's size
765  * is not known at build time. Hence it would not be possible to
766  * ensure it would fit in the stackgap in any case.
767  */
768 int
769 uvm_swap_stats(char *ptr, int misc,
770     void (*f)(void *, const struct swapent *), size_t len,
771     register_t *retval)
772 {
773 	struct swappri *spp;
774 	struct swapdev *sdp;
775 	struct swapent sep;
776 	int count = 0;
777 	int error;
778 
779 	KASSERT(len <= sizeof(sep));
780 	if (len == 0)
781 		return ENOSYS;
782 
783 	if (misc < 0)
784 		return EINVAL;
785 
786 	if (misc == 0 || uvmexp.nswapdev == 0)
787 		return 0;
788 
789 	/* Make sure userland cannot exhaust kernel memory */
790 	if ((size_t)misc > (size_t)uvmexp.nswapdev)
791 		misc = uvmexp.nswapdev;
792 
793 	KASSERT(rw_lock_held(&swap_syscall_lock));
794 
795 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
796 		TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
797 			int inuse;
798 
799 			if (misc-- <= 0)
800 				break;
801 
802 			inuse = btodb((uint64_t)sdp->swd_npginuse <<
803 			    PAGE_SHIFT);
804 
805 			memset(&sep, 0, sizeof(sep));
806 			swapent_cvt(&sep, sdp, inuse);
807 			if (f)
808 				(*f)(&sep, &sep);
809 			if ((error = copyout(&sep, ptr, len)) != 0)
810 				return error;
811 			ptr += len;
812 			count++;
813 		}
814 	}
815 	*retval = count;
816 	return 0;
817 }
818 
819 /*
820  * swap_on: attempt to enable a swapdev for swapping.   note that the
821  *	swapdev is already on the global list, but disabled (marked
822  *	SWF_FAKE).
823  *
824  * => we avoid the start of the disk (to protect disk labels)
825  * => we also avoid the miniroot, if we are swapping to root.
826  * => caller should leave uvm_swap_data_lock unlocked, we may lock it
827  *	if needed.
828  */
829 static int
830 swap_on(struct lwp *l, struct swapdev *sdp)
831 {
832 	struct vnode *vp;
833 	int error, npages, nblocks, size;
834 	long addr;
835 	vmem_addr_t result;
836 	struct vattr va;
837 	dev_t dev;
838 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
839 
840 	/*
841 	 * we want to enable swapping on sdp.   the swd_vp contains
842 	 * the vnode we want (locked and ref'd), and the swd_dev
843 	 * contains the dev_t of the file, if it a block device.
844 	 */
845 
846 	vp = sdp->swd_vp;
847 	dev = sdp->swd_dev;
848 
849 	/*
850 	 * open the swap file (mostly useful for block device files to
851 	 * let device driver know what is up).
852 	 *
853 	 * we skip the open/close for root on swap because the root
854 	 * has already been opened when root was mounted (mountroot).
855 	 */
856 	if (vp != rootvp) {
857 		if ((error = VOP_OPEN(vp, FREAD|FWRITE, l->l_cred)))
858 			return (error);
859 	}
860 
861 	/* XXX this only works for block devices */
862 	UVMHIST_LOG(pdhist, "  dev=%jd, major(dev)=%jd", dev, major(dev), 0, 0);
863 
864 	/*
865 	 * we now need to determine the size of the swap area.   for
866 	 * block specials we can call the d_psize function.
867 	 * for normal files, we must stat [get attrs].
868 	 *
869 	 * we put the result in nblks.
870 	 * for normal files, we also want the filesystem block size
871 	 * (which we get with statfs).
872 	 */
873 	switch (vp->v_type) {
874 	case VBLK:
875 		if ((nblocks = bdev_size(dev)) == -1) {
876 			error = ENXIO;
877 			goto bad;
878 		}
879 		break;
880 
881 	case VREG:
882 		if ((error = VOP_GETATTR(vp, &va, l->l_cred)))
883 			goto bad;
884 		nblocks = (int)btodb(va.va_size);
885 		sdp->swd_bsize = 1 << vp->v_mount->mnt_fs_bshift;
886 		/*
887 		 * limit the max # of outstanding I/O requests we issue
888 		 * at any one time.   take it easy on NFS servers.
889 		 */
890 		if (vp->v_tag == VT_NFS)
891 			sdp->swd_maxactive = 2; /* XXX */
892 		else
893 			sdp->swd_maxactive = 8; /* XXX */
894 		break;
895 
896 	default:
897 		error = ENXIO;
898 		goto bad;
899 	}
900 
901 	/*
902 	 * save nblocks in a safe place and convert to pages.
903 	 */
904 
905 	sdp->swd_nblks = nblocks;
906 	npages = dbtob((uint64_t)nblocks) >> PAGE_SHIFT;
907 
908 	/*
909 	 * for block special files, we want to make sure that leave
910 	 * the disklabel and bootblocks alone, so we arrange to skip
911 	 * over them (arbitrarily choosing to skip PAGE_SIZE bytes).
912 	 * note that because of this the "size" can be less than the
913 	 * actual number of blocks on the device.
914 	 */
915 	if (vp->v_type == VBLK) {
916 		/* we use pages 1 to (size - 1) [inclusive] */
917 		size = npages - 1;
918 		addr = 1;
919 	} else {
920 		/* we use pages 0 to (size - 1) [inclusive] */
921 		size = npages;
922 		addr = 0;
923 	}
924 
925 	/*
926 	 * make sure we have enough blocks for a reasonable sized swap
927 	 * area.   we want at least one page.
928 	 */
929 
930 	if (size < 1) {
931 		UVMHIST_LOG(pdhist, "  size <= 1!!", 0, 0, 0, 0);
932 		error = EINVAL;
933 		goto bad;
934 	}
935 
936 	UVMHIST_LOG(pdhist, "  dev=%#jx: size=%jd addr=%jd", dev, size, addr, 0);
937 
938 	/*
939 	 * now we need to allocate an extent to manage this swap device
940 	 */
941 
942 	sdp->swd_blist = blist_create(npages);
943 	/* mark all expect the `saved' region free. */
944 	blist_free(sdp->swd_blist, addr, size);
945 
946 	/*
947 	 * allocate space to for swap encryption state and mark the
948 	 * keys uninitialized so we generate them lazily
949 	 */
950 	sdp->swd_encmap = kmem_zalloc(encmap_size(npages), KM_SLEEP);
951 	sdp->swd_encinit = false;
952 
953 	/*
954 	 * if the vnode we are swapping to is the root vnode
955 	 * (i.e. we are swapping to the miniroot) then we want
956 	 * to make sure we don't overwrite it.   do a statfs to
957 	 * find its size and skip over it.
958 	 */
959 	if (vp == rootvp) {
960 		struct mount *mp;
961 		struct statvfs *sp;
962 		int rootblocks, rootpages;
963 
964 		mp = rootvnode->v_mount;
965 		sp = &mp->mnt_stat;
966 		rootblocks = sp->f_blocks * btodb(sp->f_frsize);
967 		/*
968 		 * XXX: sp->f_blocks isn't the total number of
969 		 * blocks in the filesystem, it's the number of
970 		 * data blocks.  so, our rootblocks almost
971 		 * definitely underestimates the total size
972 		 * of the filesystem - how badly depends on the
973 		 * details of the filesystem type.  there isn't
974 		 * an obvious way to deal with this cleanly
975 		 * and perfectly, so for now we just pad our
976 		 * rootblocks estimate with an extra 5 percent.
977 		 */
978 		rootblocks += (rootblocks >> 5) +
979 			(rootblocks >> 6) +
980 			(rootblocks >> 7);
981 		rootpages = round_page(dbtob(rootblocks)) >> PAGE_SHIFT;
982 		if (rootpages > size)
983 			panic("swap_on: miniroot larger than swap?");
984 
985 		if (rootpages != blist_fill(sdp->swd_blist, addr, rootpages)) {
986 			panic("swap_on: unable to preserve miniroot");
987 		}
988 
989 		size -= rootpages;
990 		printf("Preserved %d pages of miniroot ", rootpages);
991 		printf("leaving %d pages of swap\n", size);
992 	}
993 
994 	/*
995 	 * add a ref to vp to reflect usage as a swap device.
996 	 */
997 	vref(vp);
998 
999 	/*
1000 	 * now add the new swapdev to the drum and enable.
1001 	 */
1002 	error = vmem_alloc(swapmap, npages, VM_BESTFIT | VM_SLEEP, &result);
1003 	if (error != 0)
1004 		panic("swapdrum_add");
1005 	/*
1006 	 * If this is the first regular swap create the workqueue.
1007 	 * => Protected by swap_syscall_lock.
1008 	 */
1009 	if (vp->v_type != VBLK) {
1010 		if (sw_reg_count++ == 0) {
1011 			KASSERT(sw_reg_workqueue == NULL);
1012 			if (workqueue_create(&sw_reg_workqueue, "swapiod",
1013 			    sw_reg_iodone, NULL, PRIBIO, IPL_BIO, 0) != 0)
1014 				panic("%s: workqueue_create failed", __func__);
1015 		}
1016 	}
1017 
1018 	sdp->swd_drumoffset = (int)result;
1019 	sdp->swd_drumsize = npages;
1020 	sdp->swd_npages = size;
1021 	mutex_enter(&uvm_swap_data_lock);
1022 	sdp->swd_flags &= ~SWF_FAKE;	/* going live */
1023 	sdp->swd_flags |= (SWF_INUSE|SWF_ENABLE);
1024 	uvmexp.swpages += size;
1025 	uvmexp.swpgavail += size;
1026 	mutex_exit(&uvm_swap_data_lock);
1027 	return (0);
1028 
1029 	/*
1030 	 * failure: clean up and return error.
1031 	 */
1032 
1033 bad:
1034 	if (sdp->swd_blist) {
1035 		blist_destroy(sdp->swd_blist);
1036 	}
1037 	if (vp != rootvp) {
1038 		(void)VOP_CLOSE(vp, FREAD|FWRITE, l->l_cred);
1039 	}
1040 	return (error);
1041 }
1042 
1043 /*
1044  * swap_off: stop swapping on swapdev
1045  *
1046  * => swap data should be locked, we will unlock.
1047  */
1048 static int
1049 swap_off(struct lwp *l, struct swapdev *sdp)
1050 {
1051 	int npages = sdp->swd_npages;
1052 	int error = 0;
1053 
1054 	UVMHIST_FUNC(__func__);
1055 	UVMHIST_CALLARGS(pdhist, "  dev=%#jx, npages=%jd", sdp->swd_dev,npages, 0, 0);
1056 
1057 	KASSERT(rw_write_held(&swap_syscall_lock));
1058 	KASSERT(mutex_owned(&uvm_swap_data_lock));
1059 
1060 	/* disable the swap area being removed */
1061 	sdp->swd_flags &= ~SWF_ENABLE;
1062 	uvmexp.swpgavail -= npages;
1063 	mutex_exit(&uvm_swap_data_lock);
1064 
1065 	/*
1066 	 * the idea is to find all the pages that are paged out to this
1067 	 * device, and page them all in.  in uvm, swap-backed pageable
1068 	 * memory can take two forms: aobjs and anons.  call the
1069 	 * swapoff hook for each subsystem to bring in pages.
1070 	 */
1071 
1072 	if (uao_swap_off(sdp->swd_drumoffset,
1073 			 sdp->swd_drumoffset + sdp->swd_drumsize) ||
1074 	    amap_swap_off(sdp->swd_drumoffset,
1075 			  sdp->swd_drumoffset + sdp->swd_drumsize)) {
1076 		error = ENOMEM;
1077 	} else if (sdp->swd_npginuse > sdp->swd_npgbad) {
1078 		error = EBUSY;
1079 	}
1080 
1081 	if (error) {
1082 		mutex_enter(&uvm_swap_data_lock);
1083 		sdp->swd_flags |= SWF_ENABLE;
1084 		uvmexp.swpgavail += npages;
1085 		mutex_exit(&uvm_swap_data_lock);
1086 
1087 		return error;
1088 	}
1089 
1090 	/*
1091 	 * If this is the last regular swap destroy the workqueue.
1092 	 * => Protected by swap_syscall_lock.
1093 	 */
1094 	if (sdp->swd_vp->v_type != VBLK) {
1095 		KASSERT(sw_reg_count > 0);
1096 		KASSERT(sw_reg_workqueue != NULL);
1097 		if (--sw_reg_count == 0) {
1098 			workqueue_destroy(sw_reg_workqueue);
1099 			sw_reg_workqueue = NULL;
1100 		}
1101 	}
1102 
1103 	/*
1104 	 * done with the vnode.
1105 	 * drop our ref on the vnode before calling VOP_CLOSE()
1106 	 * so that spec_close() can tell if this is the last close.
1107 	 */
1108 	vrele(sdp->swd_vp);
1109 	if (sdp->swd_vp != rootvp) {
1110 		(void) VOP_CLOSE(sdp->swd_vp, FREAD|FWRITE, l->l_cred);
1111 	}
1112 
1113 	mutex_enter(&uvm_swap_data_lock);
1114 	uvmexp.swpages -= npages;
1115 	uvmexp.swpginuse -= sdp->swd_npgbad;
1116 
1117 	if (swaplist_find(sdp->swd_vp, true) == NULL)
1118 		panic("%s: swapdev not in list", __func__);
1119 	swaplist_trim();
1120 	mutex_exit(&uvm_swap_data_lock);
1121 
1122 	/*
1123 	 * free all resources!
1124 	 */
1125 	vmem_free(swapmap, sdp->swd_drumoffset, sdp->swd_drumsize);
1126 	blist_destroy(sdp->swd_blist);
1127 	bufq_free(sdp->swd_tab);
1128 	kmem_free(__UNVOLATILE(sdp->swd_encmap),
1129 	    encmap_size(sdp->swd_drumsize));
1130 	explicit_memset(&sdp->swd_enckey, 0, sizeof sdp->swd_enckey);
1131 	explicit_memset(&sdp->swd_deckey, 0, sizeof sdp->swd_deckey);
1132 	kmem_free(sdp, sizeof(*sdp));
1133 	return (0);
1134 }
1135 
1136 void
1137 uvm_swap_shutdown(struct lwp *l)
1138 {
1139 	struct swapdev *sdp;
1140 	struct swappri *spp;
1141 	struct vnode *vp;
1142 	int error;
1143 
1144 	printf("turning off swap...");
1145 	rw_enter(&swap_syscall_lock, RW_WRITER);
1146 	mutex_enter(&uvm_swap_data_lock);
1147 again:
1148 	LIST_FOREACH(spp, &swap_priority, spi_swappri)
1149 		TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
1150 			if (sdp->swd_flags & SWF_FAKE)
1151 				continue;
1152 			if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0)
1153 				continue;
1154 #ifdef DEBUG
1155 			printf("\nturning off swap on %s...", sdp->swd_path);
1156 #endif
1157 			/* Have to lock and reference vnode for swap_off(). */
1158 			vn_lock(vp = sdp->swd_vp, LK_EXCLUSIVE|LK_RETRY);
1159 			vref(vp);
1160 			error = swap_off(l, sdp);
1161 			vput(vp);
1162 			mutex_enter(&uvm_swap_data_lock);
1163 			if (error) {
1164 				printf("stopping swap on %s failed "
1165 				    "with error %d\n", sdp->swd_path, error);
1166 				TAILQ_REMOVE(&spp->spi_swapdev, sdp, swd_next);
1167 				uvmexp.nswapdev--;
1168 				swaplist_trim();
1169 			}
1170 			goto again;
1171 		}
1172 	printf(" done\n");
1173 	mutex_exit(&uvm_swap_data_lock);
1174 	rw_exit(&swap_syscall_lock);
1175 }
1176 
1177 
1178 /*
1179  * /dev/drum interface and i/o functions
1180  */
1181 
1182 /*
1183  * swstrategy: perform I/O on the drum
1184  *
1185  * => we must map the i/o request from the drum to the correct swapdev.
1186  */
1187 static void
1188 swstrategy(struct buf *bp)
1189 {
1190 	struct swapdev *sdp;
1191 	struct vnode *vp;
1192 	int pageno, bn;
1193 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
1194 
1195 	/*
1196 	 * convert block number to swapdev.   note that swapdev can't
1197 	 * be yanked out from under us because we are holding resources
1198 	 * in it (i.e. the blocks we are doing I/O on).
1199 	 */
1200 	pageno = dbtob((int64_t)bp->b_blkno) >> PAGE_SHIFT;
1201 	mutex_enter(&uvm_swap_data_lock);
1202 	sdp = swapdrum_getsdp(pageno);
1203 	mutex_exit(&uvm_swap_data_lock);
1204 	if (sdp == NULL) {
1205 		bp->b_error = EINVAL;
1206 		bp->b_resid = bp->b_bcount;
1207 		biodone(bp);
1208 		UVMHIST_LOG(pdhist, "  failed to get swap device", 0, 0, 0, 0);
1209 		return;
1210 	}
1211 
1212 	/*
1213 	 * convert drum page number to block number on this swapdev.
1214 	 */
1215 
1216 	pageno -= sdp->swd_drumoffset;	/* page # on swapdev */
1217 	bn = btodb((uint64_t)pageno << PAGE_SHIFT); /* convert to diskblock */
1218 
1219 	UVMHIST_LOG(pdhist, "  Rd/Wr (0/1) %jd: mapoff=%#jx bn=%#jx bcount=%jd",
1220 		((bp->b_flags & B_READ) == 0) ? 1 : 0,
1221 		sdp->swd_drumoffset, bn, bp->b_bcount);
1222 
1223 	/*
1224 	 * for block devices we finish up here.
1225 	 * for regular files we have to do more work which we delegate
1226 	 * to sw_reg_strategy().
1227 	 */
1228 
1229 	vp = sdp->swd_vp;		/* swapdev vnode pointer */
1230 	switch (vp->v_type) {
1231 	default:
1232 		panic("%s: vnode type 0x%x", __func__, vp->v_type);
1233 
1234 	case VBLK:
1235 
1236 		/*
1237 		 * must convert "bp" from an I/O on /dev/drum to an I/O
1238 		 * on the swapdev (sdp).
1239 		 */
1240 		bp->b_blkno = bn;		/* swapdev block number */
1241 		bp->b_dev = sdp->swd_dev;	/* swapdev dev_t */
1242 
1243 		/*
1244 		 * if we are doing a write, we have to redirect the i/o on
1245 		 * drum's v_numoutput counter to the swapdevs.
1246 		 */
1247 		if ((bp->b_flags & B_READ) == 0) {
1248 			mutex_enter(bp->b_objlock);
1249 			vwakeup(bp);	/* kills one 'v_numoutput' on drum */
1250 			mutex_exit(bp->b_objlock);
1251 			mutex_enter(vp->v_interlock);
1252 			vp->v_numoutput++;	/* put it on swapdev */
1253 			mutex_exit(vp->v_interlock);
1254 		}
1255 
1256 		/*
1257 		 * finally plug in swapdev vnode and start I/O
1258 		 */
1259 		bp->b_vp = vp;
1260 		bp->b_objlock = vp->v_interlock;
1261 		VOP_STRATEGY(vp, bp);
1262 		return;
1263 
1264 	case VREG:
1265 		/*
1266 		 * delegate to sw_reg_strategy function.
1267 		 */
1268 		sw_reg_strategy(sdp, bp, bn);
1269 		return;
1270 	}
1271 	/* NOTREACHED */
1272 }
1273 
1274 /*
1275  * swread: the read function for the drum (just a call to physio)
1276  */
1277 /*ARGSUSED*/
1278 static int
1279 swread(dev_t dev, struct uio *uio, int ioflag)
1280 {
1281 	UVMHIST_FUNC(__func__);
1282 	UVMHIST_CALLARGS(pdhist, "  dev=%#jx offset=%#jx", dev, uio->uio_offset, 0, 0);
1283 
1284 	return (physio(swstrategy, NULL, dev, B_READ, minphys, uio));
1285 }
1286 
1287 /*
1288  * swwrite: the write function for the drum (just a call to physio)
1289  */
1290 /*ARGSUSED*/
1291 static int
1292 swwrite(dev_t dev, struct uio *uio, int ioflag)
1293 {
1294 	UVMHIST_FUNC(__func__);
1295 	UVMHIST_CALLARGS(pdhist, "  dev=%#jx offset=%#jx", dev, uio->uio_offset, 0, 0);
1296 
1297 	return (physio(swstrategy, NULL, dev, B_WRITE, minphys, uio));
1298 }
1299 
1300 const struct bdevsw swap_bdevsw = {
1301 	.d_open = nullopen,
1302 	.d_close = nullclose,
1303 	.d_strategy = swstrategy,
1304 	.d_ioctl = noioctl,
1305 	.d_dump = nodump,
1306 	.d_psize = nosize,
1307 	.d_discard = nodiscard,
1308 	.d_flag = D_OTHER
1309 };
1310 
1311 const struct cdevsw swap_cdevsw = {
1312 	.d_open = nullopen,
1313 	.d_close = nullclose,
1314 	.d_read = swread,
1315 	.d_write = swwrite,
1316 	.d_ioctl = noioctl,
1317 	.d_stop = nostop,
1318 	.d_tty = notty,
1319 	.d_poll = nopoll,
1320 	.d_mmap = nommap,
1321 	.d_kqfilter = nokqfilter,
1322 	.d_discard = nodiscard,
1323 	.d_flag = D_OTHER,
1324 };
1325 
1326 /*
1327  * sw_reg_strategy: handle swap i/o to regular files
1328  */
1329 static void
1330 sw_reg_strategy(struct swapdev *sdp, struct buf *bp, int bn)
1331 {
1332 	struct vnode	*vp;
1333 	struct vndxfer	*vnx;
1334 	daddr_t		nbn;
1335 	char 		*addr;
1336 	off_t		byteoff;
1337 	int		s, off, nra, error, sz, resid;
1338 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
1339 
1340 	/*
1341 	 * allocate a vndxfer head for this transfer and point it to
1342 	 * our buffer.
1343 	 */
1344 	vnx = pool_get(&vndxfer_pool, PR_WAITOK);
1345 	vnx->vx_flags = VX_BUSY;
1346 	vnx->vx_error = 0;
1347 	vnx->vx_pending = 0;
1348 	vnx->vx_bp = bp;
1349 	vnx->vx_sdp = sdp;
1350 
1351 	/*
1352 	 * setup for main loop where we read filesystem blocks into
1353 	 * our buffer.
1354 	 */
1355 	error = 0;
1356 	bp->b_resid = bp->b_bcount;	/* nothing transferred yet! */
1357 	addr = bp->b_data;		/* current position in buffer */
1358 	byteoff = dbtob((uint64_t)bn);
1359 
1360 	for (resid = bp->b_resid; resid; resid -= sz) {
1361 		struct vndbuf	*nbp;
1362 
1363 		/*
1364 		 * translate byteoffset into block number.  return values:
1365 		 *   vp = vnode of underlying device
1366 		 *  nbn = new block number (on underlying vnode dev)
1367 		 *  nra = num blocks we can read-ahead (excludes requested
1368 		 *	block)
1369 		 */
1370 		nra = 0;
1371 		error = VOP_BMAP(sdp->swd_vp, byteoff / sdp->swd_bsize,
1372 				 	&vp, &nbn, &nra);
1373 
1374 		if (error == 0 && nbn == (daddr_t)-1) {
1375 			/*
1376 			 * this used to just set error, but that doesn't
1377 			 * do the right thing.  Instead, it causes random
1378 			 * memory errors.  The panic() should remain until
1379 			 * this condition doesn't destabilize the system.
1380 			 */
1381 #if 1
1382 			panic("%s: swap to sparse file", __func__);
1383 #else
1384 			error = EIO;	/* failure */
1385 #endif
1386 		}
1387 
1388 		/*
1389 		 * punt if there was an error or a hole in the file.
1390 		 * we must wait for any i/o ops we have already started
1391 		 * to finish before returning.
1392 		 *
1393 		 * XXX we could deal with holes here but it would be
1394 		 * a hassle (in the write case).
1395 		 */
1396 		if (error) {
1397 			s = splbio();
1398 			vnx->vx_error = error;	/* pass error up */
1399 			goto out;
1400 		}
1401 
1402 		/*
1403 		 * compute the size ("sz") of this transfer (in bytes).
1404 		 */
1405 		off = byteoff % sdp->swd_bsize;
1406 		sz = (1 + nra) * sdp->swd_bsize - off;
1407 		if (sz > resid)
1408 			sz = resid;
1409 
1410 		UVMHIST_LOG(pdhist, "sw_reg_strategy: "
1411 		    "vp %#jx/%#jx offset %#jx/%#jx",
1412 		    (uintptr_t)sdp->swd_vp, (uintptr_t)vp, byteoff, nbn);
1413 
1414 		/*
1415 		 * now get a buf structure.   note that the vb_buf is
1416 		 * at the front of the nbp structure so that you can
1417 		 * cast pointers between the two structure easily.
1418 		 */
1419 		nbp = pool_get(&vndbuf_pool, PR_WAITOK);
1420 		buf_init(&nbp->vb_buf);
1421 		nbp->vb_buf.b_flags    = bp->b_flags;
1422 		nbp->vb_buf.b_cflags   = bp->b_cflags;
1423 		nbp->vb_buf.b_oflags   = bp->b_oflags;
1424 		nbp->vb_buf.b_bcount   = sz;
1425 		nbp->vb_buf.b_bufsize  = sz;
1426 		nbp->vb_buf.b_error    = 0;
1427 		nbp->vb_buf.b_data     = addr;
1428 		nbp->vb_buf.b_lblkno   = 0;
1429 		nbp->vb_buf.b_blkno    = nbn + btodb(off);
1430 		nbp->vb_buf.b_rawblkno = nbp->vb_buf.b_blkno;
1431 		nbp->vb_buf.b_iodone   = sw_reg_biodone;
1432 		nbp->vb_buf.b_vp       = vp;
1433 		nbp->vb_buf.b_objlock  = vp->v_interlock;
1434 		if (vp->v_type == VBLK) {
1435 			nbp->vb_buf.b_dev = vp->v_rdev;
1436 		}
1437 
1438 		nbp->vb_xfer = vnx;	/* patch it back in to vnx */
1439 
1440 		/*
1441 		 * Just sort by block number
1442 		 */
1443 		s = splbio();
1444 		if (vnx->vx_error != 0) {
1445 			buf_destroy(&nbp->vb_buf);
1446 			pool_put(&vndbuf_pool, nbp);
1447 			goto out;
1448 		}
1449 		vnx->vx_pending++;
1450 
1451 		/* sort it in and start I/O if we are not over our limit */
1452 		/* XXXAD locking */
1453 		bufq_put(sdp->swd_tab, &nbp->vb_buf);
1454 		sw_reg_start(sdp);
1455 		splx(s);
1456 
1457 		/*
1458 		 * advance to the next I/O
1459 		 */
1460 		byteoff += sz;
1461 		addr += sz;
1462 	}
1463 
1464 	s = splbio();
1465 
1466 out: /* Arrive here at splbio */
1467 	vnx->vx_flags &= ~VX_BUSY;
1468 	if (vnx->vx_pending == 0) {
1469 		error = vnx->vx_error;
1470 		pool_put(&vndxfer_pool, vnx);
1471 		bp->b_error = error;
1472 		biodone(bp);
1473 	}
1474 	splx(s);
1475 }
1476 
1477 /*
1478  * sw_reg_start: start an I/O request on the requested swapdev
1479  *
1480  * => reqs are sorted by b_rawblkno (above)
1481  */
1482 static void
1483 sw_reg_start(struct swapdev *sdp)
1484 {
1485 	struct buf	*bp;
1486 	struct vnode	*vp;
1487 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
1488 
1489 	/* recursion control */
1490 	if ((sdp->swd_flags & SWF_BUSY) != 0)
1491 		return;
1492 
1493 	sdp->swd_flags |= SWF_BUSY;
1494 
1495 	while (sdp->swd_active < sdp->swd_maxactive) {
1496 		bp = bufq_get(sdp->swd_tab);
1497 		if (bp == NULL)
1498 			break;
1499 		sdp->swd_active++;
1500 
1501 		UVMHIST_LOG(pdhist,
1502 		    "sw_reg_start:  bp %#jx vp %#jx blkno %#jx cnt %#jx",
1503 		    (uintptr_t)bp, (uintptr_t)bp->b_vp, (uintptr_t)bp->b_blkno,
1504 		    bp->b_bcount);
1505 		vp = bp->b_vp;
1506 		KASSERT(bp->b_objlock == vp->v_interlock);
1507 		if ((bp->b_flags & B_READ) == 0) {
1508 			mutex_enter(vp->v_interlock);
1509 			vp->v_numoutput++;
1510 			mutex_exit(vp->v_interlock);
1511 		}
1512 		VOP_STRATEGY(vp, bp);
1513 	}
1514 	sdp->swd_flags &= ~SWF_BUSY;
1515 }
1516 
1517 /*
1518  * sw_reg_biodone: one of our i/o's has completed
1519  */
1520 static void
1521 sw_reg_biodone(struct buf *bp)
1522 {
1523 	workqueue_enqueue(sw_reg_workqueue, &bp->b_work, NULL);
1524 }
1525 
1526 /*
1527  * sw_reg_iodone: one of our i/o's has completed and needs post-i/o cleanup
1528  *
1529  * => note that we can recover the vndbuf struct by casting the buf ptr
1530  */
1531 static void
1532 sw_reg_iodone(struct work *wk, void *dummy)
1533 {
1534 	struct vndbuf *vbp = (void *)wk;
1535 	struct vndxfer *vnx = vbp->vb_xfer;
1536 	struct buf *pbp = vnx->vx_bp;		/* parent buffer */
1537 	struct swapdev	*sdp = vnx->vx_sdp;
1538 	int s, resid, error;
1539 	KASSERT(&vbp->vb_buf.b_work == wk);
1540 	UVMHIST_FUNC(__func__);
1541 	UVMHIST_CALLARGS(pdhist, "  vbp=%#jx vp=%#jx blkno=%#jx addr=%#jx",
1542 	    (uintptr_t)vbp, (uintptr_t)vbp->vb_buf.b_vp, vbp->vb_buf.b_blkno,
1543 	    (uintptr_t)vbp->vb_buf.b_data);
1544 	UVMHIST_LOG(pdhist, "  cnt=%#jx resid=%#jx",
1545 	    vbp->vb_buf.b_bcount, vbp->vb_buf.b_resid, 0, 0);
1546 
1547 	/*
1548 	 * protect vbp at splbio and update.
1549 	 */
1550 
1551 	s = splbio();
1552 	resid = vbp->vb_buf.b_bcount - vbp->vb_buf.b_resid;
1553 	pbp->b_resid -= resid;
1554 	vnx->vx_pending--;
1555 
1556 	if (vbp->vb_buf.b_error != 0) {
1557 		/* pass error upward */
1558 		error = vbp->vb_buf.b_error ? vbp->vb_buf.b_error : EIO;
1559 		UVMHIST_LOG(pdhist, "  got error=%jd !", error, 0, 0, 0);
1560 		vnx->vx_error = error;
1561 	}
1562 
1563 	/*
1564 	 * kill vbp structure
1565 	 */
1566 	buf_destroy(&vbp->vb_buf);
1567 	pool_put(&vndbuf_pool, vbp);
1568 
1569 	/*
1570 	 * wrap up this transaction if it has run to completion or, in
1571 	 * case of an error, when all auxiliary buffers have returned.
1572 	 */
1573 	if (vnx->vx_error != 0) {
1574 		/* pass error upward */
1575 		error = vnx->vx_error;
1576 		if ((vnx->vx_flags & VX_BUSY) == 0 && vnx->vx_pending == 0) {
1577 			pbp->b_error = error;
1578 			biodone(pbp);
1579 			pool_put(&vndxfer_pool, vnx);
1580 		}
1581 	} else if (pbp->b_resid == 0) {
1582 		KASSERT(vnx->vx_pending == 0);
1583 		if ((vnx->vx_flags & VX_BUSY) == 0) {
1584 			UVMHIST_LOG(pdhist, "  iodone, pbp=%#jx error=%jd !",
1585 			    (uintptr_t)pbp, vnx->vx_error, 0, 0);
1586 			biodone(pbp);
1587 			pool_put(&vndxfer_pool, vnx);
1588 		}
1589 	}
1590 
1591 	/*
1592 	 * done!   start next swapdev I/O if one is pending
1593 	 */
1594 	sdp->swd_active--;
1595 	sw_reg_start(sdp);
1596 	splx(s);
1597 }
1598 
1599 
1600 /*
1601  * uvm_swap_alloc: allocate space on swap
1602  *
1603  * => allocation is done "round robin" down the priority list, as we
1604  *	allocate in a priority we "rotate" the circle queue.
1605  * => space can be freed with uvm_swap_free
1606  * => we return the page slot number in /dev/drum (0 == invalid slot)
1607  * => we lock uvm_swap_data_lock
1608  * => XXXMRG: "LESSOK" INTERFACE NEEDED TO EXTENT SYSTEM
1609  */
1610 int
1611 uvm_swap_alloc(int *nslots /* IN/OUT */, bool lessok)
1612 {
1613 	struct swapdev *sdp;
1614 	struct swappri *spp;
1615 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
1616 
1617 	/*
1618 	 * no swap devices configured yet?   definite failure.
1619 	 */
1620 	if (uvmexp.nswapdev < 1)
1621 		return 0;
1622 
1623 	/*
1624 	 * XXXJAK: BEGIN HACK
1625 	 *
1626 	 * blist_alloc() in subr_blist.c will panic if we try to allocate
1627 	 * too many slots.
1628 	 */
1629 	if (*nslots > BLIST_MAX_ALLOC) {
1630 		if (__predict_false(lessok == false))
1631 			return 0;
1632 		*nslots = BLIST_MAX_ALLOC;
1633 	}
1634 	/* XXXJAK: END HACK */
1635 
1636 	/*
1637 	 * lock data lock, convert slots into blocks, and enter loop
1638 	 */
1639 	mutex_enter(&uvm_swap_data_lock);
1640 
1641 ReTry:	/* XXXMRG */
1642 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
1643 		TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
1644 			uint64_t result;
1645 
1646 			/* if it's not enabled, then we can't swap from it */
1647 			if ((sdp->swd_flags & SWF_ENABLE) == 0)
1648 				continue;
1649 			if (sdp->swd_npginuse + *nslots > sdp->swd_npages)
1650 				continue;
1651 			result = blist_alloc(sdp->swd_blist, *nslots);
1652 			if (result == BLIST_NONE) {
1653 				continue;
1654 			}
1655 			KASSERT(result < sdp->swd_drumsize);
1656 
1657 			/*
1658 			 * successful allocation!  now rotate the tailq.
1659 			 */
1660 			TAILQ_REMOVE(&spp->spi_swapdev, sdp, swd_next);
1661 			TAILQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
1662 			sdp->swd_npginuse += *nslots;
1663 			uvmexp.swpginuse += *nslots;
1664 			mutex_exit(&uvm_swap_data_lock);
1665 			/* done!  return drum slot number */
1666 			UVMHIST_LOG(pdhist,
1667 			    "success!  returning %jd slots starting at %jd",
1668 			    *nslots, result + sdp->swd_drumoffset, 0, 0);
1669 			return (result + sdp->swd_drumoffset);
1670 		}
1671 	}
1672 
1673 	/* XXXMRG: BEGIN HACK */
1674 	if (*nslots > 1 && lessok) {
1675 		*nslots = 1;
1676 		/* XXXMRG: ugh!  blist should support this for us */
1677 		goto ReTry;
1678 	}
1679 	/* XXXMRG: END HACK */
1680 
1681 	mutex_exit(&uvm_swap_data_lock);
1682 	return 0;
1683 }
1684 
1685 /*
1686  * uvm_swapisfull: return true if most of available swap is allocated
1687  * and in use.  we don't count some small portion as it may be inaccessible
1688  * to us at any given moment, for example if there is lock contention or if
1689  * pages are busy.
1690  */
1691 bool
1692 uvm_swapisfull(void)
1693 {
1694 	int swpgonly;
1695 	bool rv;
1696 
1697 	if (uvmexp.swpages == 0) {
1698 		return true;
1699 	}
1700 
1701 	mutex_enter(&uvm_swap_data_lock);
1702 	KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1703 	swpgonly = (int)((uint64_t)uvmexp.swpgonly * 100 /
1704 	    uvm_swapisfull_factor);
1705 	rv = (swpgonly >= uvmexp.swpgavail);
1706 	mutex_exit(&uvm_swap_data_lock);
1707 
1708 	return (rv);
1709 }
1710 
1711 /*
1712  * uvm_swap_markbad: keep track of swap ranges where we've had i/o errors
1713  *
1714  * => we lock uvm_swap_data_lock
1715  */
1716 void
1717 uvm_swap_markbad(int startslot, int nslots)
1718 {
1719 	struct swapdev *sdp;
1720 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
1721 
1722 	mutex_enter(&uvm_swap_data_lock);
1723 	sdp = swapdrum_getsdp(startslot);
1724 	KASSERT(sdp != NULL);
1725 
1726 	/*
1727 	 * we just keep track of how many pages have been marked bad
1728 	 * in this device, to make everything add up in swap_off().
1729 	 * we assume here that the range of slots will all be within
1730 	 * one swap device.
1731 	 */
1732 
1733 	KASSERT(uvmexp.swpgonly >= nslots);
1734 	atomic_add_int(&uvmexp.swpgonly, -nslots);
1735 	sdp->swd_npgbad += nslots;
1736 	UVMHIST_LOG(pdhist, "now %jd bad", sdp->swd_npgbad, 0,0,0);
1737 	mutex_exit(&uvm_swap_data_lock);
1738 }
1739 
1740 /*
1741  * uvm_swap_free: free swap slots
1742  *
1743  * => this can be all or part of an allocation made by uvm_swap_alloc
1744  * => we lock uvm_swap_data_lock
1745  */
1746 void
1747 uvm_swap_free(int startslot, int nslots)
1748 {
1749 	struct swapdev *sdp;
1750 	UVMHIST_FUNC(__func__);
1751 	UVMHIST_CALLARGS(pdhist, "freeing %jd slots starting at %jd", nslots,
1752 	    startslot, 0, 0);
1753 
1754 	/*
1755 	 * ignore attempts to free the "bad" slot.
1756 	 */
1757 
1758 	if (startslot == SWSLOT_BAD) {
1759 		return;
1760 	}
1761 
1762 	/*
1763 	 * convert drum slot offset back to sdp, free the blocks
1764 	 * in the extent, and return.   must hold pri lock to do
1765 	 * lookup and access the extent.
1766 	 */
1767 
1768 	mutex_enter(&uvm_swap_data_lock);
1769 	sdp = swapdrum_getsdp(startslot);
1770 	KASSERT(uvmexp.nswapdev >= 1);
1771 	KASSERT(sdp != NULL);
1772 	KASSERT(sdp->swd_npginuse >= nslots);
1773 	blist_free(sdp->swd_blist, startslot - sdp->swd_drumoffset, nslots);
1774 	sdp->swd_npginuse -= nslots;
1775 	uvmexp.swpginuse -= nslots;
1776 	mutex_exit(&uvm_swap_data_lock);
1777 }
1778 
1779 /*
1780  * uvm_swap_put: put any number of pages into a contig place on swap
1781  *
1782  * => can be sync or async
1783  */
1784 
1785 int
1786 uvm_swap_put(int swslot, struct vm_page **ppsp, int npages, int flags)
1787 {
1788 	int error;
1789 
1790 	error = uvm_swap_io(ppsp, swslot, npages, B_WRITE |
1791 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1792 	return error;
1793 }
1794 
1795 /*
1796  * uvm_swap_get: get a single page from swap
1797  *
1798  * => usually a sync op (from fault)
1799  */
1800 
1801 int
1802 uvm_swap_get(struct vm_page *page, int swslot, int flags)
1803 {
1804 	int error;
1805 
1806 	atomic_inc_uint(&uvmexp.nswget);
1807 	KASSERT(flags & PGO_SYNCIO);
1808 	if (swslot == SWSLOT_BAD) {
1809 		return EIO;
1810 	}
1811 
1812 	error = uvm_swap_io(&page, swslot, 1, B_READ |
1813 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1814 	if (error == 0) {
1815 
1816 		/*
1817 		 * this page is no longer only in swap.
1818 		 */
1819 
1820 		KASSERT(uvmexp.swpgonly > 0);
1821 		atomic_dec_uint(&uvmexp.swpgonly);
1822 	}
1823 	return error;
1824 }
1825 
1826 /*
1827  * uvm_swap_io: do an i/o operation to swap
1828  */
1829 
1830 static int
1831 uvm_swap_io(struct vm_page **pps, int startslot, int npages, int flags)
1832 {
1833 	daddr_t startblk;
1834 	struct	buf *bp;
1835 	vaddr_t kva;
1836 	int	error, mapinflags;
1837 	bool write, async, swap_encrypt;
1838 	UVMHIST_FUNC(__func__);
1839 	UVMHIST_CALLARGS(pdhist, "<- called, startslot=%jd, npages=%jd, flags=%#jx",
1840 	    startslot, npages, flags, 0);
1841 
1842 	write = (flags & B_READ) == 0;
1843 	async = (flags & B_ASYNC) != 0;
1844 	swap_encrypt = atomic_load_relaxed(&uvm_swap_encrypt);
1845 
1846 	/*
1847 	 * allocate a buf for the i/o.
1848 	 */
1849 
1850 	KASSERT(curlwp != uvm.pagedaemon_lwp || (write && async));
1851 	bp = getiobuf(swapdev_vp, curlwp != uvm.pagedaemon_lwp);
1852 	if (bp == NULL) {
1853 		uvm_aio_aiodone_pages(pps, npages, true, ENOMEM);
1854 		return ENOMEM;
1855 	}
1856 
1857 	/*
1858 	 * convert starting drum slot to block number
1859 	 */
1860 
1861 	startblk = btodb((uint64_t)startslot << PAGE_SHIFT);
1862 
1863 	/*
1864 	 * first, map the pages into the kernel.
1865 	 */
1866 
1867 	mapinflags = !write ?
1868 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_READ :
1869 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_WRITE;
1870 	if (write && swap_encrypt)	/* need to encrypt in-place */
1871 		mapinflags |= UVMPAGER_MAPIN_READ;
1872 	kva = uvm_pagermapin(pps, npages, mapinflags);
1873 
1874 	/*
1875 	 * encrypt writes in place if requested
1876 	 */
1877 
1878 	if (write) do {
1879 		struct swapdev *sdp;
1880 		int i;
1881 
1882 		/*
1883 		 * Get the swapdev so we can discriminate on the
1884 		 * encryption state.  There may or may not be an
1885 		 * encryption key generated; we may or may not be asked
1886 		 * to encrypt swap.
1887 		 *
1888 		 * 1. NO KEY, NO ENCRYPTION: Nothing to do.
1889 		 *
1890 		 * 2. NO KEY, BUT ENCRYPTION: Generate a key, encrypt,
1891 		 *    and mark the slots encrypted.
1892 		 *
1893 		 * 3. KEY, BUT NO ENCRYPTION: The slots may already be
1894 		 *    marked encrypted from a past life.  Mark them not
1895 		 *    encrypted.
1896 		 *
1897 		 * 4. KEY, ENCRYPTION: Encrypt and mark the slots
1898 		 *    encrypted.
1899 		 */
1900 		mutex_enter(&uvm_swap_data_lock);
1901 		sdp = swapdrum_getsdp(startslot);
1902 		if (!sdp->swd_encinit) {
1903 			if (!swap_encrypt) {
1904 				mutex_exit(&uvm_swap_data_lock);
1905 				break;
1906 			}
1907 			uvm_swap_genkey(sdp);
1908 		}
1909 		KASSERT(sdp->swd_encinit);
1910 		mutex_exit(&uvm_swap_data_lock);
1911 
1912 		for (i = 0; i < npages; i++) {
1913 			int s = startslot + i;
1914 			KDASSERT(swapdrum_sdp_is(s, sdp));
1915 			KASSERT(s >= sdp->swd_drumoffset);
1916 			s -= sdp->swd_drumoffset;
1917 			KASSERT(s < sdp->swd_drumsize);
1918 
1919 			if (swap_encrypt) {
1920 				uvm_swap_encryptpage(sdp,
1921 				    (void *)(kva + (vsize_t)i*PAGE_SIZE), s);
1922 				atomic_or_32(&sdp->swd_encmap[s/32],
1923 				    __BIT(s%32));
1924 			} else {
1925 				atomic_and_32(&sdp->swd_encmap[s/32],
1926 				    ~__BIT(s%32));
1927 			}
1928 		}
1929 	} while (0);
1930 
1931 	/*
1932 	 * fill in the bp/sbp.   we currently route our i/o through
1933 	 * /dev/drum's vnode [swapdev_vp].
1934 	 */
1935 
1936 	bp->b_cflags = BC_BUSY | BC_NOCACHE;
1937 	bp->b_flags = (flags & (B_READ|B_ASYNC));
1938 	bp->b_proc = &proc0;	/* XXX */
1939 	bp->b_vnbufs.le_next = NOLIST;
1940 	bp->b_data = (void *)kva;
1941 	bp->b_blkno = startblk;
1942 	bp->b_bufsize = bp->b_bcount = npages << PAGE_SHIFT;
1943 
1944 	/*
1945 	 * bump v_numoutput (counter of number of active outputs).
1946 	 */
1947 
1948 	if (write) {
1949 		mutex_enter(swapdev_vp->v_interlock);
1950 		swapdev_vp->v_numoutput++;
1951 		mutex_exit(swapdev_vp->v_interlock);
1952 	}
1953 
1954 	/*
1955 	 * for async ops we must set up the iodone handler.
1956 	 */
1957 
1958 	if (async) {
1959 		bp->b_iodone = uvm_aio_aiodone;
1960 		UVMHIST_LOG(pdhist, "doing async!", 0, 0, 0, 0);
1961 		if (curlwp == uvm.pagedaemon_lwp)
1962 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1963 		else
1964 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
1965 	} else {
1966 		bp->b_iodone = NULL;
1967 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1968 	}
1969 	UVMHIST_LOG(pdhist,
1970 	    "about to start io: data = %#jx blkno = %#jx, bcount = %jd",
1971 	    (uintptr_t)bp->b_data, bp->b_blkno, bp->b_bcount, 0);
1972 
1973 	/*
1974 	 * now we start the I/O, and if async, return.
1975 	 */
1976 
1977 	VOP_STRATEGY(swapdev_vp, bp);
1978 	if (async) {
1979 		/*
1980 		 * Reads are always synchronous; if this changes, we
1981 		 * need to add an asynchronous path for decryption.
1982 		 */
1983 		KASSERT(write);
1984 		return 0;
1985 	}
1986 
1987 	/*
1988 	 * must be sync i/o.   wait for it to finish
1989 	 */
1990 
1991 	error = biowait(bp);
1992 	if (error)
1993 		goto out;
1994 
1995 	/*
1996 	 * decrypt reads in place if needed
1997 	 */
1998 
1999 	if (!write) do {
2000 		struct swapdev *sdp;
2001 		bool encinit;
2002 		int i;
2003 
2004 		/*
2005 		 * Get the sdp.  Everything about it except the encinit
2006 		 * bit, saying whether the encryption key is
2007 		 * initialized or not, and the encrypted bit for each
2008 		 * page, is stable until all swap pages have been
2009 		 * released and the device is removed.
2010 		 */
2011 		mutex_enter(&uvm_swap_data_lock);
2012 		sdp = swapdrum_getsdp(startslot);
2013 		encinit = sdp->swd_encinit;
2014 		mutex_exit(&uvm_swap_data_lock);
2015 
2016 		if (!encinit)
2017 			/*
2018 			 * If there's no encryption key, there's no way
2019 			 * any of these slots can be encrypted, so
2020 			 * nothing to do here.
2021 			 */
2022 			break;
2023 		for (i = 0; i < npages; i++) {
2024 			int s = startslot + i;
2025 			KDASSERT(swapdrum_sdp_is(s, sdp));
2026 			KASSERT(s >= sdp->swd_drumoffset);
2027 			s -= sdp->swd_drumoffset;
2028 			KASSERT(s < sdp->swd_drumsize);
2029 			if ((atomic_load_relaxed(&sdp->swd_encmap[s/32]) &
2030 				__BIT(s%32)) == 0)
2031 				continue;
2032 			uvm_swap_decryptpage(sdp,
2033 			    (void *)(kva + (vsize_t)i*PAGE_SIZE), s);
2034 		}
2035 	} while (0);
2036 out:
2037 	/*
2038 	 * kill the pager mapping
2039 	 */
2040 
2041 	uvm_pagermapout(kva, npages);
2042 
2043 	/*
2044 	 * now dispose of the buf and we're done.
2045 	 */
2046 
2047 	if (write) {
2048 		mutex_enter(swapdev_vp->v_interlock);
2049 		vwakeup(bp);
2050 		mutex_exit(swapdev_vp->v_interlock);
2051 	}
2052 	putiobuf(bp);
2053 	UVMHIST_LOG(pdhist, "<- done (sync)  error=%jd", error, 0, 0, 0);
2054 
2055 	return (error);
2056 }
2057 
2058 /*
2059  * uvm_swap_genkey(sdp)
2060  *
2061  *	Generate a key for swap encryption.
2062  */
2063 static void
2064 uvm_swap_genkey(struct swapdev *sdp)
2065 {
2066 	uint8_t key[32];
2067 
2068 	KASSERT(!sdp->swd_encinit);
2069 
2070 	cprng_strong(kern_cprng, key, sizeof key, 0);
2071 	aes_setenckey256(&sdp->swd_enckey, key);
2072 	aes_setdeckey256(&sdp->swd_deckey, key);
2073 	explicit_memset(key, 0, sizeof key);
2074 
2075 	sdp->swd_encinit = true;
2076 }
2077 
2078 /*
2079  * uvm_swap_encryptpage(sdp, kva, slot)
2080  *
2081  *	Encrypt one page of data at kva for the specified slot number
2082  *	in the swap device.
2083  */
2084 static void
2085 uvm_swap_encryptpage(struct swapdev *sdp, void *kva, int slot)
2086 {
2087 	uint8_t preiv[16] __aligned(16) = {0}, iv[16] __aligned(16);
2088 
2089 	/* iv := AES_k(le32enc(slot) || 0^96) */
2090 	le32enc(preiv, slot);
2091 	aes_enc(&sdp->swd_enckey, (const void *)preiv, iv, AES_256_NROUNDS);
2092 
2093 	/* *kva := AES-CBC_k(iv, *kva) */
2094 	aes_cbc_enc(&sdp->swd_enckey, kva, kva, PAGE_SIZE, iv,
2095 	    AES_256_NROUNDS);
2096 
2097 	explicit_memset(&iv, 0, sizeof iv);
2098 }
2099 
2100 /*
2101  * uvm_swap_decryptpage(sdp, kva, slot)
2102  *
2103  *	Decrypt one page of data at kva for the specified slot number
2104  *	in the swap device.
2105  */
2106 static void
2107 uvm_swap_decryptpage(struct swapdev *sdp, void *kva, int slot)
2108 {
2109 	uint8_t preiv[16] __aligned(16) = {0}, iv[16] __aligned(16);
2110 
2111 	/* iv := AES_k(le32enc(slot) || 0^96) */
2112 	le32enc(preiv, slot);
2113 	aes_enc(&sdp->swd_enckey, (const void *)preiv, iv, AES_256_NROUNDS);
2114 
2115 	/* *kva := AES-CBC^{-1}_k(iv, *kva) */
2116 	aes_cbc_dec(&sdp->swd_deckey, kva, kva, PAGE_SIZE, iv,
2117 	    AES_256_NROUNDS);
2118 
2119 	explicit_memset(&iv, 0, sizeof iv);
2120 }
2121 
2122 SYSCTL_SETUP(sysctl_uvmswap_setup, "sysctl uvmswap setup")
2123 {
2124 
2125 	sysctl_createv(clog, 0, NULL, NULL,
2126 	    CTLFLAG_PERMANENT|CTLFLAG_READWRITE, CTLTYPE_BOOL, "swap_encrypt",
2127 	    SYSCTL_DESCR("Encrypt data when swapped out to disk"),
2128 	    NULL, 0, &uvm_swap_encrypt, 0,
2129 	    CTL_VM, CTL_CREATE, CTL_EOL);
2130 }
2131