xref: /netbsd-src/sys/uvm/uvm_swap.c (revision 7788a0781fe6ff2cce37368b4578a7ade0850cb1)
1 /*	$NetBSD: uvm_swap.c,v 1.163 2013/05/07 15:49:09 riastradh 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 1.163 2013/05/07 15:49:09 riastradh 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/buf.h>
42 #include <sys/bufq.h>
43 #include <sys/conf.h>
44 #include <sys/proc.h>
45 #include <sys/namei.h>
46 #include <sys/disklabel.h>
47 #include <sys/errno.h>
48 #include <sys/kernel.h>
49 #include <sys/vnode.h>
50 #include <sys/file.h>
51 #include <sys/vmem.h>
52 #include <sys/blist.h>
53 #include <sys/mount.h>
54 #include <sys/pool.h>
55 #include <sys/kmem.h>
56 #include <sys/syscallargs.h>
57 #include <sys/swap.h>
58 #include <sys/kauth.h>
59 #include <sys/sysctl.h>
60 #include <sys/workqueue.h>
61 
62 #include <uvm/uvm.h>
63 
64 #include <miscfs/specfs/specdev.h>
65 
66 /*
67  * uvm_swap.c: manage configuration and i/o to swap space.
68  */
69 
70 /*
71  * swap space is managed in the following way:
72  *
73  * each swap partition or file is described by a "swapdev" structure.
74  * each "swapdev" structure contains a "swapent" structure which contains
75  * information that is passed up to the user (via system calls).
76  *
77  * each swap partition is assigned a "priority" (int) which controls
78  * swap parition usage.
79  *
80  * the system maintains a global data structure describing all swap
81  * partitions/files.   there is a sorted LIST of "swappri" structures
82  * which describe "swapdev"'s at that priority.   this LIST is headed
83  * by the "swap_priority" global var.    each "swappri" contains a
84  * CIRCLEQ of "swapdev" structures at that priority.
85  *
86  * locking:
87  *  - swap_syscall_lock (krwlock_t): this lock serializes the swapctl
88  *    system call and prevents the swap priority list from changing
89  *    while we are in the middle of a system call (e.g. SWAP_STATS).
90  *  - uvm_swap_data_lock (kmutex_t): this lock protects all swap data
91  *    structures including the priority list, the swapdev structures,
92  *    and the swapmap arena.
93  *
94  * each swap device has the following info:
95  *  - swap device in use (could be disabled, preventing future use)
96  *  - swap enabled (allows new allocations on swap)
97  *  - map info in /dev/drum
98  *  - vnode pointer
99  * for swap files only:
100  *  - block size
101  *  - max byte count in buffer
102  *  - buffer
103  *
104  * userland controls and configures swap with the swapctl(2) system call.
105  * the sys_swapctl performs the following operations:
106  *  [1] SWAP_NSWAP: returns the number of swap devices currently configured
107  *  [2] SWAP_STATS: given a pointer to an array of swapent structures
108  *	(passed in via "arg") of a size passed in via "misc" ... we load
109  *	the current swap config into the array. The actual work is done
110  *	in the uvm_swap_stats() function.
111  *  [3] SWAP_ON: given a pathname in arg (could be device or file) and a
112  *	priority in "misc", start swapping on it.
113  *  [4] SWAP_OFF: as SWAP_ON, but stops swapping to a device
114  *  [5] SWAP_CTL: changes the priority of a swap device (new priority in
115  *	"misc")
116  */
117 
118 /*
119  * swapdev: describes a single swap partition/file
120  *
121  * note the following should be true:
122  * swd_inuse <= swd_nblks  [number of blocks in use is <= total blocks]
123  * swd_nblks <= swd_mapsize [because mapsize includes miniroot+disklabel]
124  */
125 struct swapdev {
126 	dev_t			swd_dev;	/* device id */
127 	int			swd_flags;	/* flags:inuse/enable/fake */
128 	int			swd_priority;	/* our priority */
129 	int			swd_nblks;	/* blocks in this device */
130 	char			*swd_path;	/* saved pathname of device */
131 	int			swd_pathlen;	/* length of pathname */
132 	int			swd_npages;	/* #pages we can use */
133 	int			swd_npginuse;	/* #pages in use */
134 	int			swd_npgbad;	/* #pages bad */
135 	int			swd_drumoffset;	/* page0 offset in drum */
136 	int			swd_drumsize;	/* #pages in drum */
137 	blist_t			swd_blist;	/* blist for this swapdev */
138 	struct vnode		*swd_vp;	/* backing vnode */
139 	CIRCLEQ_ENTRY(swapdev)	swd_next;	/* priority circleq */
140 
141 	int			swd_bsize;	/* blocksize (bytes) */
142 	int			swd_maxactive;	/* max active i/o reqs */
143 	struct bufq_state	*swd_tab;	/* buffer list */
144 	int			swd_active;	/* number of active buffers */
145 };
146 
147 /*
148  * swap device priority entry; the list is kept sorted on `spi_priority'.
149  */
150 struct swappri {
151 	int			spi_priority;     /* priority */
152 	CIRCLEQ_HEAD(spi_swapdev, swapdev)	spi_swapdev;
153 	/* circleq of swapdevs at this priority */
154 	LIST_ENTRY(swappri)	spi_swappri;      /* global list of pri's */
155 };
156 
157 /*
158  * The following two structures are used to keep track of data transfers
159  * on swap devices associated with regular files.
160  * NOTE: this code is more or less a copy of vnd.c; we use the same
161  * structure names here to ease porting..
162  */
163 struct vndxfer {
164 	struct buf	*vx_bp;		/* Pointer to parent buffer */
165 	struct swapdev	*vx_sdp;
166 	int		vx_error;
167 	int		vx_pending;	/* # of pending aux buffers */
168 	int		vx_flags;
169 #define VX_BUSY		1
170 #define VX_DEAD		2
171 };
172 
173 struct vndbuf {
174 	struct buf	vb_buf;
175 	struct vndxfer	*vb_xfer;
176 };
177 
178 /*
179  * NetBSD 1.3 swapctl(SWAP_STATS, ...) swapent structure; uses 32 bit
180  * dev_t and has no se_path[] member.
181  */
182 struct swapent13 {
183 	int32_t	se13_dev;		/* device id */
184 	int	se13_flags;		/* flags */
185 	int	se13_nblks;		/* total blocks */
186 	int	se13_inuse;		/* blocks in use */
187 	int	se13_priority;		/* priority of this device */
188 };
189 
190 /*
191  * NetBSD 5.0 swapctl(SWAP_STATS, ...) swapent structure; uses 32 bit
192  * dev_t.
193  */
194 struct swapent50 {
195 	int32_t	se50_dev;		/* device id */
196 	int	se50_flags;		/* flags */
197 	int	se50_nblks;		/* total blocks */
198 	int	se50_inuse;		/* blocks in use */
199 	int	se50_priority;		/* priority of this device */
200 	char	se50_path[PATH_MAX+1];	/* path name */
201 };
202 
203 /*
204  * We keep a of pool vndbuf's and vndxfer structures.
205  */
206 static struct pool vndxfer_pool, vndbuf_pool;
207 
208 /*
209  * local variables
210  */
211 static vmem_t *swapmap;	/* controls the mapping of /dev/drum */
212 
213 /* list of all active swap devices [by priority] */
214 LIST_HEAD(swap_priority, swappri);
215 static struct swap_priority swap_priority;
216 
217 /* locks */
218 static krwlock_t swap_syscall_lock;
219 
220 /* workqueue and use counter for swap to regular files */
221 static int sw_reg_count = 0;
222 static struct workqueue *sw_reg_workqueue;
223 
224 /* tuneables */
225 u_int uvm_swapisfull_factor = 99;
226 
227 /*
228  * prototypes
229  */
230 static struct swapdev	*swapdrum_getsdp(int);
231 
232 static struct swapdev	*swaplist_find(struct vnode *, bool);
233 static void		 swaplist_insert(struct swapdev *,
234 					 struct swappri *, int);
235 static void		 swaplist_trim(void);
236 
237 static int swap_on(struct lwp *, struct swapdev *);
238 static int swap_off(struct lwp *, struct swapdev *);
239 
240 static void uvm_swap_stats(int, struct swapent *, int, register_t *);
241 
242 static void sw_reg_strategy(struct swapdev *, struct buf *, int);
243 static void sw_reg_biodone(struct buf *);
244 static void sw_reg_iodone(struct work *wk, void *dummy);
245 static void sw_reg_start(struct swapdev *);
246 
247 static int uvm_swap_io(struct vm_page **, int, int, int);
248 
249 /*
250  * uvm_swap_init: init the swap system data structures and locks
251  *
252  * => called at boot time from init_main.c after the filesystems
253  *	are brought up (which happens after uvm_init())
254  */
255 void
256 uvm_swap_init(void)
257 {
258 	UVMHIST_FUNC("uvm_swap_init");
259 
260 	UVMHIST_CALLED(pdhist);
261 	/*
262 	 * first, init the swap list, its counter, and its lock.
263 	 * then get a handle on the vnode for /dev/drum by using
264 	 * the its dev_t number ("swapdev", from MD conf.c).
265 	 */
266 
267 	LIST_INIT(&swap_priority);
268 	uvmexp.nswapdev = 0;
269 	rw_init(&swap_syscall_lock);
270 	mutex_init(&uvm_swap_data_lock, MUTEX_DEFAULT, IPL_NONE);
271 
272 	if (bdevvp(swapdev, &swapdev_vp))
273 		panic("%s: can't get vnode for swap device", __func__);
274 	if (vn_lock(swapdev_vp, LK_EXCLUSIVE | LK_RETRY))
275 		panic("%s: can't lock swap device", __func__);
276 	if (VOP_OPEN(swapdev_vp, FREAD | FWRITE, NOCRED))
277 		panic("%s: can't open swap device", __func__);
278 	VOP_UNLOCK(swapdev_vp);
279 
280 	/*
281 	 * create swap block resource map to map /dev/drum.   the range
282 	 * from 1 to INT_MAX allows 2 gigablocks of swap space.  note
283 	 * that block 0 is reserved (used to indicate an allocation
284 	 * failure, or no allocation).
285 	 */
286 	swapmap = vmem_create("swapmap", 1, INT_MAX - 1, 1, NULL, NULL, NULL, 0,
287 	    VM_NOSLEEP, IPL_NONE);
288 	if (swapmap == 0) {
289 		panic("%s: vmem_create failed", __func__);
290 	}
291 
292 	pool_init(&vndxfer_pool, sizeof(struct vndxfer), 0, 0, 0, "swp vnx",
293 	    NULL, IPL_BIO);
294 	pool_init(&vndbuf_pool, sizeof(struct vndbuf), 0, 0, 0, "swp vnd",
295 	    NULL, IPL_BIO);
296 
297 	UVMHIST_LOG(pdhist, "<- done", 0, 0, 0, 0);
298 }
299 
300 /*
301  * swaplist functions: functions that operate on the list of swap
302  * devices on the system.
303  */
304 
305 /*
306  * swaplist_insert: insert swap device "sdp" into the global list
307  *
308  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
309  * => caller must provide a newly allocated swappri structure (we will
310  *	FREE it if we don't need it... this it to prevent allocation
311  *	blocking here while adding swap)
312  */
313 static void
314 swaplist_insert(struct swapdev *sdp, struct swappri *newspp, int priority)
315 {
316 	struct swappri *spp, *pspp;
317 	UVMHIST_FUNC("swaplist_insert"); UVMHIST_CALLED(pdhist);
318 
319 	/*
320 	 * find entry at or after which to insert the new device.
321 	 */
322 	pspp = NULL;
323 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
324 		if (priority <= spp->spi_priority)
325 			break;
326 		pspp = spp;
327 	}
328 
329 	/*
330 	 * new priority?
331 	 */
332 	if (spp == NULL || spp->spi_priority != priority) {
333 		spp = newspp;  /* use newspp! */
334 		UVMHIST_LOG(pdhist, "created new swappri = %d",
335 			    priority, 0, 0, 0);
336 
337 		spp->spi_priority = priority;
338 		CIRCLEQ_INIT(&spp->spi_swapdev);
339 
340 		if (pspp)
341 			LIST_INSERT_AFTER(pspp, spp, spi_swappri);
342 		else
343 			LIST_INSERT_HEAD(&swap_priority, spp, spi_swappri);
344 	} else {
345 	  	/* we don't need a new priority structure, free it */
346 		kmem_free(newspp, sizeof(*newspp));
347 	}
348 
349 	/*
350 	 * priority found (or created).   now insert on the priority's
351 	 * circleq list and bump the total number of swapdevs.
352 	 */
353 	sdp->swd_priority = priority;
354 	CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
355 	uvmexp.nswapdev++;
356 }
357 
358 /*
359  * swaplist_find: find and optionally remove a swap device from the
360  *	global list.
361  *
362  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
363  * => we return the swapdev we found (and removed)
364  */
365 static struct swapdev *
366 swaplist_find(struct vnode *vp, bool remove)
367 {
368 	struct swapdev *sdp;
369 	struct swappri *spp;
370 
371 	/*
372 	 * search the lists for the requested vp
373 	 */
374 
375 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
376 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
377 			if (sdp->swd_vp == vp) {
378 				if (remove) {
379 					CIRCLEQ_REMOVE(&spp->spi_swapdev,
380 					    sdp, swd_next);
381 					uvmexp.nswapdev--;
382 				}
383 				return(sdp);
384 			}
385 		}
386 	}
387 	return (NULL);
388 }
389 
390 /*
391  * swaplist_trim: scan priority list for empty priority entries and kill
392  *	them.
393  *
394  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
395  */
396 static void
397 swaplist_trim(void)
398 {
399 	struct swappri *spp, *nextspp;
400 
401 	LIST_FOREACH_SAFE(spp, &swap_priority, spi_swappri, nextspp) {
402 		if (CIRCLEQ_FIRST(&spp->spi_swapdev) !=
403 		    (void *)&spp->spi_swapdev)
404 			continue;
405 		LIST_REMOVE(spp, spi_swappri);
406 		kmem_free(spp, sizeof(*spp));
407 	}
408 }
409 
410 /*
411  * swapdrum_getsdp: given a page offset in /dev/drum, convert it back
412  *	to the "swapdev" that maps that section of the drum.
413  *
414  * => each swapdev takes one big contig chunk of the drum
415  * => caller must hold uvm_swap_data_lock
416  */
417 static struct swapdev *
418 swapdrum_getsdp(int pgno)
419 {
420 	struct swapdev *sdp;
421 	struct swappri *spp;
422 
423 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
424 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
425 			if (sdp->swd_flags & SWF_FAKE)
426 				continue;
427 			if (pgno >= sdp->swd_drumoffset &&
428 			    pgno < (sdp->swd_drumoffset + sdp->swd_drumsize)) {
429 				return sdp;
430 			}
431 		}
432 	}
433 	return NULL;
434 }
435 
436 
437 /*
438  * sys_swapctl: main entry point for swapctl(2) system call
439  * 	[with two helper functions: swap_on and swap_off]
440  */
441 int
442 sys_swapctl(struct lwp *l, const struct sys_swapctl_args *uap, register_t *retval)
443 {
444 	/* {
445 		syscallarg(int) cmd;
446 		syscallarg(void *) arg;
447 		syscallarg(int) misc;
448 	} */
449 	struct vnode *vp;
450 	struct nameidata nd;
451 	struct swappri *spp;
452 	struct swapdev *sdp;
453 	struct swapent *sep;
454 #define SWAP_PATH_MAX (PATH_MAX + 1)
455 	char	*userpath;
456 	size_t	len = 0;
457 	int	error, misc;
458 	int	priority;
459 	UVMHIST_FUNC("sys_swapctl"); UVMHIST_CALLED(pdhist);
460 
461 	/*
462 	 * we handle the non-priv NSWAP and STATS request first.
463 	 *
464 	 * SWAP_NSWAP: return number of config'd swap devices
465 	 * [can also be obtained with uvmexp sysctl]
466 	 */
467 	if (SCARG(uap, cmd) == SWAP_NSWAP) {
468 		const int nswapdev = uvmexp.nswapdev;
469 		UVMHIST_LOG(pdhist, "<- done SWAP_NSWAP=%d", nswapdev, 0, 0, 0);
470 		*retval = nswapdev;
471 		return 0;
472 	}
473 
474 	misc = SCARG(uap, misc);
475 	userpath = kmem_alloc(SWAP_PATH_MAX, KM_SLEEP);
476 
477 	/*
478 	 * ensure serialized syscall access by grabbing the swap_syscall_lock
479 	 */
480 	rw_enter(&swap_syscall_lock, RW_WRITER);
481 
482 	/*
483 	 * SWAP_STATS: get stats on current # of configured swap devs
484 	 *
485 	 * note that the swap_priority list can't change as long
486 	 * as we are holding the swap_syscall_lock.  we don't want
487 	 * to grab the uvm_swap_data_lock because we may fault&sleep during
488 	 * copyout() and we don't want to be holding that lock then!
489 	 */
490 	if (SCARG(uap, cmd) == SWAP_STATS
491 #if defined(COMPAT_50)
492 	    || SCARG(uap, cmd) == SWAP_STATS50
493 #endif
494 #if defined(COMPAT_13)
495 	    || SCARG(uap, cmd) == SWAP_STATS13
496 #endif
497 	    ) {
498 		if ((size_t)misc > (size_t)uvmexp.nswapdev)
499 			misc = uvmexp.nswapdev;
500 
501 		if (misc == 0) {
502 			error = EINVAL;
503 			goto out;
504 		}
505 		KASSERT(misc > 0);
506 #if defined(COMPAT_13)
507 		if (SCARG(uap, cmd) == SWAP_STATS13)
508 			len = sizeof(struct swapent13) * misc;
509 		else
510 #endif
511 #if defined(COMPAT_50)
512 		if (SCARG(uap, cmd) == SWAP_STATS50)
513 			len = sizeof(struct swapent50) * misc;
514 		else
515 #endif
516 			len = sizeof(struct swapent) * misc;
517 		sep = (struct swapent *)kmem_alloc(len, KM_SLEEP);
518 
519 		uvm_swap_stats(SCARG(uap, cmd), sep, misc, retval);
520 		error = copyout(sep, SCARG(uap, arg), len);
521 
522 		kmem_free(sep, len);
523 		UVMHIST_LOG(pdhist, "<- done SWAP_STATS", 0, 0, 0, 0);
524 		goto out;
525 	}
526 	if (SCARG(uap, cmd) == SWAP_GETDUMPDEV) {
527 		dev_t	*devp = (dev_t *)SCARG(uap, arg);
528 
529 		error = copyout(&dumpdev, devp, sizeof(dumpdev));
530 		goto out;
531 	}
532 
533 	/*
534 	 * all other requests require superuser privs.   verify.
535 	 */
536 	if ((error = kauth_authorize_system(l->l_cred, KAUTH_SYSTEM_SWAPCTL,
537 	    0, NULL, NULL, NULL)))
538 		goto out;
539 
540 	if (SCARG(uap, cmd) == SWAP_DUMPOFF) {
541 		/* drop the current dump device */
542 		dumpdev = NODEV;
543 		dumpcdev = NODEV;
544 		cpu_dumpconf();
545 		goto out;
546 	}
547 
548 	/*
549 	 * at this point we expect a path name in arg.   we will
550 	 * use namei() to gain a vnode reference (vref), and lock
551 	 * the vnode (VOP_LOCK).
552 	 *
553 	 * XXX: a NULL arg means use the root vnode pointer (e.g. for
554 	 * miniroot)
555 	 */
556 	if (SCARG(uap, arg) == NULL) {
557 		vp = rootvp;		/* miniroot */
558 		vref(vp);
559 		if (vn_lock(vp, LK_EXCLUSIVE)) {
560 			vrele(vp);
561 			error = EBUSY;
562 			goto out;
563 		}
564 		if (SCARG(uap, cmd) == SWAP_ON &&
565 		    copystr("miniroot", userpath, SWAP_PATH_MAX, &len))
566 			panic("swapctl: miniroot copy failed");
567 		KASSERT(len > 0);
568 	} else {
569 		struct pathbuf *pb;
570 
571 		/*
572 		 * This used to allow copying in one extra byte
573 		 * (SWAP_PATH_MAX instead of PATH_MAX) for SWAP_ON.
574 		 * This was completely pointless because if anyone
575 		 * used that extra byte namei would fail with
576 		 * ENAMETOOLONG anyway, so I've removed the excess
577 		 * logic. - dholland 20100215
578 		 */
579 
580 		error = pathbuf_copyin(SCARG(uap, arg), &pb);
581 		if (error) {
582 			goto out;
583 		}
584 		if (SCARG(uap, cmd) == SWAP_ON) {
585 			/* get a copy of the string */
586 			pathbuf_copystring(pb, userpath, SWAP_PATH_MAX);
587 			len = strlen(userpath) + 1;
588 		}
589 		NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | TRYEMULROOT, pb);
590 		if ((error = namei(&nd))) {
591 			pathbuf_destroy(pb);
592 			goto out;
593 		}
594 		vp = nd.ni_vp;
595 		pathbuf_destroy(pb);
596 	}
597 	/* note: "vp" is referenced and locked */
598 
599 	error = 0;		/* assume no error */
600 	switch(SCARG(uap, cmd)) {
601 
602 	case SWAP_DUMPDEV:
603 		if (vp->v_type != VBLK) {
604 			error = ENOTBLK;
605 			break;
606 		}
607 		if (bdevsw_lookup(vp->v_rdev)) {
608 			dumpdev = vp->v_rdev;
609 			dumpcdev = devsw_blk2chr(dumpdev);
610 		} else
611 			dumpdev = NODEV;
612 		cpu_dumpconf();
613 		break;
614 
615 	case SWAP_CTL:
616 		/*
617 		 * get new priority, remove old entry (if any) and then
618 		 * reinsert it in the correct place.  finally, prune out
619 		 * any empty priority structures.
620 		 */
621 		priority = SCARG(uap, misc);
622 		spp = kmem_alloc(sizeof(*spp), KM_SLEEP);
623 		mutex_enter(&uvm_swap_data_lock);
624 		if ((sdp = swaplist_find(vp, true)) == NULL) {
625 			error = ENOENT;
626 		} else {
627 			swaplist_insert(sdp, spp, priority);
628 			swaplist_trim();
629 		}
630 		mutex_exit(&uvm_swap_data_lock);
631 		if (error)
632 			kmem_free(spp, sizeof(*spp));
633 		break;
634 
635 	case SWAP_ON:
636 
637 		/*
638 		 * check for duplicates.   if none found, then insert a
639 		 * dummy entry on the list to prevent someone else from
640 		 * trying to enable this device while we are working on
641 		 * it.
642 		 */
643 
644 		priority = SCARG(uap, misc);
645 		sdp = kmem_zalloc(sizeof(*sdp), KM_SLEEP);
646 		spp = kmem_alloc(sizeof(*spp), KM_SLEEP);
647 		sdp->swd_flags = SWF_FAKE;
648 		sdp->swd_vp = vp;
649 		sdp->swd_dev = (vp->v_type == VBLK) ? vp->v_rdev : NODEV;
650 		bufq_alloc(&sdp->swd_tab, "disksort", BUFQ_SORT_RAWBLOCK);
651 		mutex_enter(&uvm_swap_data_lock);
652 		if (swaplist_find(vp, false) != NULL) {
653 			error = EBUSY;
654 			mutex_exit(&uvm_swap_data_lock);
655 			bufq_free(sdp->swd_tab);
656 			kmem_free(sdp, sizeof(*sdp));
657 			kmem_free(spp, sizeof(*spp));
658 			break;
659 		}
660 		swaplist_insert(sdp, spp, priority);
661 		mutex_exit(&uvm_swap_data_lock);
662 
663 		KASSERT(len > 0);
664 		sdp->swd_pathlen = len;
665 		sdp->swd_path = kmem_alloc(len, KM_SLEEP);
666 		if (copystr(userpath, sdp->swd_path, len, 0) != 0)
667 			panic("swapctl: copystr");
668 
669 		/*
670 		 * we've now got a FAKE placeholder in the swap list.
671 		 * now attempt to enable swap on it.  if we fail, undo
672 		 * what we've done and kill the fake entry we just inserted.
673 		 * if swap_on is a success, it will clear the SWF_FAKE flag
674 		 */
675 
676 		if ((error = swap_on(l, sdp)) != 0) {
677 			mutex_enter(&uvm_swap_data_lock);
678 			(void) swaplist_find(vp, true);  /* kill fake entry */
679 			swaplist_trim();
680 			mutex_exit(&uvm_swap_data_lock);
681 			bufq_free(sdp->swd_tab);
682 			kmem_free(sdp->swd_path, sdp->swd_pathlen);
683 			kmem_free(sdp, sizeof(*sdp));
684 			break;
685 		}
686 		break;
687 
688 	case SWAP_OFF:
689 		mutex_enter(&uvm_swap_data_lock);
690 		if ((sdp = swaplist_find(vp, false)) == NULL) {
691 			mutex_exit(&uvm_swap_data_lock);
692 			error = ENXIO;
693 			break;
694 		}
695 
696 		/*
697 		 * If a device isn't in use or enabled, we
698 		 * can't stop swapping from it (again).
699 		 */
700 		if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0) {
701 			mutex_exit(&uvm_swap_data_lock);
702 			error = EBUSY;
703 			break;
704 		}
705 
706 		/*
707 		 * do the real work.
708 		 */
709 		error = swap_off(l, sdp);
710 		break;
711 
712 	default:
713 		error = EINVAL;
714 	}
715 
716 	/*
717 	 * done!  release the ref gained by namei() and unlock.
718 	 */
719 	vput(vp);
720 out:
721 	rw_exit(&swap_syscall_lock);
722 	kmem_free(userpath, SWAP_PATH_MAX);
723 
724 	UVMHIST_LOG(pdhist, "<- done!  error=%d", error, 0, 0, 0);
725 	return (error);
726 }
727 
728 /*
729  * uvm_swap_stats: implements swapctl(SWAP_STATS). The function is kept
730  * away from sys_swapctl() in order to allow COMPAT_* swapctl()
731  * emulation to use it directly without going through sys_swapctl().
732  * The problem with using sys_swapctl() there is that it involves
733  * copying the swapent array to the stackgap, and this array's size
734  * is not known at build time. Hence it would not be possible to
735  * ensure it would fit in the stackgap in any case.
736  */
737 static void
738 uvm_swap_stats(int cmd, struct swapent *sep, int sec, register_t *retval)
739 {
740 	struct swappri *spp;
741 	struct swapdev *sdp;
742 	int count = 0;
743 
744 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
745 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
746 			int inuse;
747 
748 			if (sec-- <= 0)
749 				break;
750 
751 			/*
752 			 * backwards compatibility for system call.
753 			 * For NetBSD 1.3 and 5.0, we have to use
754 			 * the 32 bit dev_t.  For 5.0 and -current
755 			 * we have to add the path.
756 			 */
757 			inuse = btodb((uint64_t)sdp->swd_npginuse <<
758 			    PAGE_SHIFT);
759 
760 #if defined(COMPAT_13) || defined(COMPAT_50)
761 			if (cmd == SWAP_STATS) {
762 #endif
763 				sep->se_dev = sdp->swd_dev;
764 				sep->se_flags = sdp->swd_flags;
765 				sep->se_nblks = sdp->swd_nblks;
766 				sep->se_inuse = inuse;
767 				sep->se_priority = sdp->swd_priority;
768 				KASSERT(sdp->swd_pathlen <
769 				    sizeof(sep->se_path));
770 				strcpy(sep->se_path, sdp->swd_path);
771 				sep++;
772 #if defined(COMPAT_13)
773 			} else if (cmd == SWAP_STATS13) {
774 				struct swapent13 *sep13 =
775 				    (struct swapent13 *)sep;
776 
777 				sep13->se13_dev = sdp->swd_dev;
778 				sep13->se13_flags = sdp->swd_flags;
779 				sep13->se13_nblks = sdp->swd_nblks;
780 				sep13->se13_inuse = inuse;
781 				sep13->se13_priority = sdp->swd_priority;
782 				sep = (struct swapent *)(sep13 + 1);
783 #endif
784 #if defined(COMPAT_50)
785 			} else if (cmd == SWAP_STATS50) {
786 				struct swapent50 *sep50 =
787 				    (struct swapent50 *)sep;
788 
789 				sep50->se50_dev = sdp->swd_dev;
790 				sep50->se50_flags = sdp->swd_flags;
791 				sep50->se50_nblks = sdp->swd_nblks;
792 				sep50->se50_inuse = inuse;
793 				sep50->se50_priority = sdp->swd_priority;
794 				KASSERT(sdp->swd_pathlen <
795 				    sizeof(sep50->se50_path));
796 				strcpy(sep50->se50_path, sdp->swd_path);
797 				sep = (struct swapent *)(sep50 + 1);
798 #endif
799 #if defined(COMPAT_13) || defined(COMPAT_50)
800 			}
801 #endif
802 			count++;
803 		}
804 	}
805 	*retval = count;
806 }
807 
808 /*
809  * swap_on: attempt to enable a swapdev for swapping.   note that the
810  *	swapdev is already on the global list, but disabled (marked
811  *	SWF_FAKE).
812  *
813  * => we avoid the start of the disk (to protect disk labels)
814  * => we also avoid the miniroot, if we are swapping to root.
815  * => caller should leave uvm_swap_data_lock unlocked, we may lock it
816  *	if needed.
817  */
818 static int
819 swap_on(struct lwp *l, struct swapdev *sdp)
820 {
821 	struct vnode *vp;
822 	int error, npages, nblocks, size;
823 	long addr;
824 	vmem_addr_t result;
825 	struct vattr va;
826 	dev_t dev;
827 	UVMHIST_FUNC("swap_on"); UVMHIST_CALLED(pdhist);
828 
829 	/*
830 	 * we want to enable swapping on sdp.   the swd_vp contains
831 	 * the vnode we want (locked and ref'd), and the swd_dev
832 	 * contains the dev_t of the file, if it a block device.
833 	 */
834 
835 	vp = sdp->swd_vp;
836 	dev = sdp->swd_dev;
837 
838 	/*
839 	 * open the swap file (mostly useful for block device files to
840 	 * let device driver know what is up).
841 	 *
842 	 * we skip the open/close for root on swap because the root
843 	 * has already been opened when root was mounted (mountroot).
844 	 */
845 	if (vp != rootvp) {
846 		if ((error = VOP_OPEN(vp, FREAD|FWRITE, l->l_cred)))
847 			return (error);
848 	}
849 
850 	/* XXX this only works for block devices */
851 	UVMHIST_LOG(pdhist, "  dev=%d, major(dev)=%d", dev, major(dev), 0,0);
852 
853 	/*
854 	 * we now need to determine the size of the swap area.   for
855 	 * block specials we can call the d_psize function.
856 	 * for normal files, we must stat [get attrs].
857 	 *
858 	 * we put the result in nblks.
859 	 * for normal files, we also want the filesystem block size
860 	 * (which we get with statfs).
861 	 */
862 	switch (vp->v_type) {
863 	case VBLK:
864 		if ((nblocks = bdev_size(dev)) == -1) {
865 			error = ENXIO;
866 			goto bad;
867 		}
868 		break;
869 
870 	case VREG:
871 		if ((error = VOP_GETATTR(vp, &va, l->l_cred)))
872 			goto bad;
873 		nblocks = (int)btodb(va.va_size);
874 		sdp->swd_bsize = 1 << vp->v_mount->mnt_fs_bshift;
875 		/*
876 		 * limit the max # of outstanding I/O requests we issue
877 		 * at any one time.   take it easy on NFS servers.
878 		 */
879 		if (vp->v_tag == VT_NFS)
880 			sdp->swd_maxactive = 2; /* XXX */
881 		else
882 			sdp->swd_maxactive = 8; /* XXX */
883 		break;
884 
885 	default:
886 		error = ENXIO;
887 		goto bad;
888 	}
889 
890 	/*
891 	 * save nblocks in a safe place and convert to pages.
892 	 */
893 
894 	sdp->swd_nblks = nblocks;
895 	npages = dbtob((uint64_t)nblocks) >> PAGE_SHIFT;
896 
897 	/*
898 	 * for block special files, we want to make sure that leave
899 	 * the disklabel and bootblocks alone, so we arrange to skip
900 	 * over them (arbitrarily choosing to skip PAGE_SIZE bytes).
901 	 * note that because of this the "size" can be less than the
902 	 * actual number of blocks on the device.
903 	 */
904 	if (vp->v_type == VBLK) {
905 		/* we use pages 1 to (size - 1) [inclusive] */
906 		size = npages - 1;
907 		addr = 1;
908 	} else {
909 		/* we use pages 0 to (size - 1) [inclusive] */
910 		size = npages;
911 		addr = 0;
912 	}
913 
914 	/*
915 	 * make sure we have enough blocks for a reasonable sized swap
916 	 * area.   we want at least one page.
917 	 */
918 
919 	if (size < 1) {
920 		UVMHIST_LOG(pdhist, "  size <= 1!!", 0, 0, 0, 0);
921 		error = EINVAL;
922 		goto bad;
923 	}
924 
925 	UVMHIST_LOG(pdhist, "  dev=%x: size=%d addr=%ld\n", dev, size, addr, 0);
926 
927 	/*
928 	 * now we need to allocate an extent to manage this swap device
929 	 */
930 
931 	sdp->swd_blist = blist_create(npages);
932 	/* mark all expect the `saved' region free. */
933 	blist_free(sdp->swd_blist, addr, size);
934 
935 	/*
936 	 * if the vnode we are swapping to is the root vnode
937 	 * (i.e. we are swapping to the miniroot) then we want
938 	 * to make sure we don't overwrite it.   do a statfs to
939 	 * find its size and skip over it.
940 	 */
941 	if (vp == rootvp) {
942 		struct mount *mp;
943 		struct statvfs *sp;
944 		int rootblocks, rootpages;
945 
946 		mp = rootvnode->v_mount;
947 		sp = &mp->mnt_stat;
948 		rootblocks = sp->f_blocks * btodb(sp->f_frsize);
949 		/*
950 		 * XXX: sp->f_blocks isn't the total number of
951 		 * blocks in the filesystem, it's the number of
952 		 * data blocks.  so, our rootblocks almost
953 		 * definitely underestimates the total size
954 		 * of the filesystem - how badly depends on the
955 		 * details of the filesystem type.  there isn't
956 		 * an obvious way to deal with this cleanly
957 		 * and perfectly, so for now we just pad our
958 		 * rootblocks estimate with an extra 5 percent.
959 		 */
960 		rootblocks += (rootblocks >> 5) +
961 			(rootblocks >> 6) +
962 			(rootblocks >> 7);
963 		rootpages = round_page(dbtob(rootblocks)) >> PAGE_SHIFT;
964 		if (rootpages > size)
965 			panic("swap_on: miniroot larger than swap?");
966 
967 		if (rootpages != blist_fill(sdp->swd_blist, addr, rootpages)) {
968 			panic("swap_on: unable to preserve miniroot");
969 		}
970 
971 		size -= rootpages;
972 		printf("Preserved %d pages of miniroot ", rootpages);
973 		printf("leaving %d pages of swap\n", size);
974 	}
975 
976 	/*
977 	 * add a ref to vp to reflect usage as a swap device.
978 	 */
979 	vref(vp);
980 
981 	/*
982 	 * now add the new swapdev to the drum and enable.
983 	 */
984 	error = vmem_alloc(swapmap, npages, VM_BESTFIT | VM_SLEEP, &result);
985 	if (error != 0)
986 		panic("swapdrum_add");
987 	/*
988 	 * If this is the first regular swap create the workqueue.
989 	 * => Protected by swap_syscall_lock.
990 	 */
991 	if (vp->v_type != VBLK) {
992 		if (sw_reg_count++ == 0) {
993 			KASSERT(sw_reg_workqueue == NULL);
994 			if (workqueue_create(&sw_reg_workqueue, "swapiod",
995 			    sw_reg_iodone, NULL, PRIBIO, IPL_BIO, 0) != 0)
996 				panic("%s: workqueue_create failed", __func__);
997 		}
998 	}
999 
1000 	sdp->swd_drumoffset = (int)result;
1001 	sdp->swd_drumsize = npages;
1002 	sdp->swd_npages = size;
1003 	mutex_enter(&uvm_swap_data_lock);
1004 	sdp->swd_flags &= ~SWF_FAKE;	/* going live */
1005 	sdp->swd_flags |= (SWF_INUSE|SWF_ENABLE);
1006 	uvmexp.swpages += size;
1007 	uvmexp.swpgavail += size;
1008 	mutex_exit(&uvm_swap_data_lock);
1009 	return (0);
1010 
1011 	/*
1012 	 * failure: clean up and return error.
1013 	 */
1014 
1015 bad:
1016 	if (sdp->swd_blist) {
1017 		blist_destroy(sdp->swd_blist);
1018 	}
1019 	if (vp != rootvp) {
1020 		(void)VOP_CLOSE(vp, FREAD|FWRITE, l->l_cred);
1021 	}
1022 	return (error);
1023 }
1024 
1025 /*
1026  * swap_off: stop swapping on swapdev
1027  *
1028  * => swap data should be locked, we will unlock.
1029  */
1030 static int
1031 swap_off(struct lwp *l, struct swapdev *sdp)
1032 {
1033 	int npages = sdp->swd_npages;
1034 	int error = 0;
1035 
1036 	UVMHIST_FUNC("swap_off"); UVMHIST_CALLED(pdhist);
1037 	UVMHIST_LOG(pdhist, "  dev=%x, npages=%d", sdp->swd_dev,npages,0,0);
1038 
1039 	/* disable the swap area being removed */
1040 	sdp->swd_flags &= ~SWF_ENABLE;
1041 	uvmexp.swpgavail -= npages;
1042 	mutex_exit(&uvm_swap_data_lock);
1043 
1044 	/*
1045 	 * the idea is to find all the pages that are paged out to this
1046 	 * device, and page them all in.  in uvm, swap-backed pageable
1047 	 * memory can take two forms: aobjs and anons.  call the
1048 	 * swapoff hook for each subsystem to bring in pages.
1049 	 */
1050 
1051 	if (uao_swap_off(sdp->swd_drumoffset,
1052 			 sdp->swd_drumoffset + sdp->swd_drumsize) ||
1053 	    amap_swap_off(sdp->swd_drumoffset,
1054 			  sdp->swd_drumoffset + sdp->swd_drumsize)) {
1055 		error = ENOMEM;
1056 	} else if (sdp->swd_npginuse > sdp->swd_npgbad) {
1057 		error = EBUSY;
1058 	}
1059 
1060 	if (error) {
1061 		mutex_enter(&uvm_swap_data_lock);
1062 		sdp->swd_flags |= SWF_ENABLE;
1063 		uvmexp.swpgavail += npages;
1064 		mutex_exit(&uvm_swap_data_lock);
1065 
1066 		return error;
1067 	}
1068 
1069 	/*
1070 	 * If this is the last regular swap destroy the workqueue.
1071 	 * => Protected by swap_syscall_lock.
1072 	 */
1073 	if (sdp->swd_vp->v_type != VBLK) {
1074 		KASSERT(sw_reg_count > 0);
1075 		KASSERT(sw_reg_workqueue != NULL);
1076 		if (--sw_reg_count == 0) {
1077 			workqueue_destroy(sw_reg_workqueue);
1078 			sw_reg_workqueue = NULL;
1079 		}
1080 	}
1081 
1082 	/*
1083 	 * done with the vnode.
1084 	 * drop our ref on the vnode before calling VOP_CLOSE()
1085 	 * so that spec_close() can tell if this is the last close.
1086 	 */
1087 	vrele(sdp->swd_vp);
1088 	if (sdp->swd_vp != rootvp) {
1089 		(void) VOP_CLOSE(sdp->swd_vp, FREAD|FWRITE, l->l_cred);
1090 	}
1091 
1092 	mutex_enter(&uvm_swap_data_lock);
1093 	uvmexp.swpages -= npages;
1094 	uvmexp.swpginuse -= sdp->swd_npgbad;
1095 
1096 	if (swaplist_find(sdp->swd_vp, true) == NULL)
1097 		panic("%s: swapdev not in list", __func__);
1098 	swaplist_trim();
1099 	mutex_exit(&uvm_swap_data_lock);
1100 
1101 	/*
1102 	 * free all resources!
1103 	 */
1104 	vmem_free(swapmap, sdp->swd_drumoffset, sdp->swd_drumsize);
1105 	blist_destroy(sdp->swd_blist);
1106 	bufq_free(sdp->swd_tab);
1107 	kmem_free(sdp, sizeof(*sdp));
1108 	return (0);
1109 }
1110 
1111 /*
1112  * /dev/drum interface and i/o functions
1113  */
1114 
1115 /*
1116  * swstrategy: perform I/O on the drum
1117  *
1118  * => we must map the i/o request from the drum to the correct swapdev.
1119  */
1120 static void
1121 swstrategy(struct buf *bp)
1122 {
1123 	struct swapdev *sdp;
1124 	struct vnode *vp;
1125 	int pageno, bn;
1126 	UVMHIST_FUNC("swstrategy"); UVMHIST_CALLED(pdhist);
1127 
1128 	/*
1129 	 * convert block number to swapdev.   note that swapdev can't
1130 	 * be yanked out from under us because we are holding resources
1131 	 * in it (i.e. the blocks we are doing I/O on).
1132 	 */
1133 	pageno = dbtob((int64_t)bp->b_blkno) >> PAGE_SHIFT;
1134 	mutex_enter(&uvm_swap_data_lock);
1135 	sdp = swapdrum_getsdp(pageno);
1136 	mutex_exit(&uvm_swap_data_lock);
1137 	if (sdp == NULL) {
1138 		bp->b_error = EINVAL;
1139 		bp->b_resid = bp->b_bcount;
1140 		biodone(bp);
1141 		UVMHIST_LOG(pdhist, "  failed to get swap device", 0, 0, 0, 0);
1142 		return;
1143 	}
1144 
1145 	/*
1146 	 * convert drum page number to block number on this swapdev.
1147 	 */
1148 
1149 	pageno -= sdp->swd_drumoffset;	/* page # on swapdev */
1150 	bn = btodb((uint64_t)pageno << PAGE_SHIFT); /* convert to diskblock */
1151 
1152 	UVMHIST_LOG(pdhist, "  %s: mapoff=%x bn=%x bcount=%ld",
1153 		((bp->b_flags & B_READ) == 0) ? "write" : "read",
1154 		sdp->swd_drumoffset, bn, bp->b_bcount);
1155 
1156 	/*
1157 	 * for block devices we finish up here.
1158 	 * for regular files we have to do more work which we delegate
1159 	 * to sw_reg_strategy().
1160 	 */
1161 
1162 	vp = sdp->swd_vp;		/* swapdev vnode pointer */
1163 	switch (vp->v_type) {
1164 	default:
1165 		panic("%s: vnode type 0x%x", __func__, vp->v_type);
1166 
1167 	case VBLK:
1168 
1169 		/*
1170 		 * must convert "bp" from an I/O on /dev/drum to an I/O
1171 		 * on the swapdev (sdp).
1172 		 */
1173 		bp->b_blkno = bn;		/* swapdev block number */
1174 		bp->b_dev = sdp->swd_dev;	/* swapdev dev_t */
1175 
1176 		/*
1177 		 * if we are doing a write, we have to redirect the i/o on
1178 		 * drum's v_numoutput counter to the swapdevs.
1179 		 */
1180 		if ((bp->b_flags & B_READ) == 0) {
1181 			mutex_enter(bp->b_objlock);
1182 			vwakeup(bp);	/* kills one 'v_numoutput' on drum */
1183 			mutex_exit(bp->b_objlock);
1184 			mutex_enter(vp->v_interlock);
1185 			vp->v_numoutput++;	/* put it on swapdev */
1186 			mutex_exit(vp->v_interlock);
1187 		}
1188 
1189 		/*
1190 		 * finally plug in swapdev vnode and start I/O
1191 		 */
1192 		bp->b_vp = vp;
1193 		bp->b_objlock = vp->v_interlock;
1194 		VOP_STRATEGY(vp, bp);
1195 		return;
1196 
1197 	case VREG:
1198 		/*
1199 		 * delegate to sw_reg_strategy function.
1200 		 */
1201 		sw_reg_strategy(sdp, bp, bn);
1202 		return;
1203 	}
1204 	/* NOTREACHED */
1205 }
1206 
1207 /*
1208  * swread: the read function for the drum (just a call to physio)
1209  */
1210 /*ARGSUSED*/
1211 static int
1212 swread(dev_t dev, struct uio *uio, int ioflag)
1213 {
1214 	UVMHIST_FUNC("swread"); UVMHIST_CALLED(pdhist);
1215 
1216 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1217 	return (physio(swstrategy, NULL, dev, B_READ, minphys, uio));
1218 }
1219 
1220 /*
1221  * swwrite: the write function for the drum (just a call to physio)
1222  */
1223 /*ARGSUSED*/
1224 static int
1225 swwrite(dev_t dev, struct uio *uio, int ioflag)
1226 {
1227 	UVMHIST_FUNC("swwrite"); UVMHIST_CALLED(pdhist);
1228 
1229 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1230 	return (physio(swstrategy, NULL, dev, B_WRITE, minphys, uio));
1231 }
1232 
1233 const struct bdevsw swap_bdevsw = {
1234 	nullopen, nullclose, swstrategy, noioctl, nodump, nosize, D_OTHER,
1235 };
1236 
1237 const struct cdevsw swap_cdevsw = {
1238 	nullopen, nullclose, swread, swwrite, noioctl,
1239 	nostop, notty, nopoll, nommap, nokqfilter, D_OTHER,
1240 };
1241 
1242 /*
1243  * sw_reg_strategy: handle swap i/o to regular files
1244  */
1245 static void
1246 sw_reg_strategy(struct swapdev *sdp, struct buf *bp, int bn)
1247 {
1248 	struct vnode	*vp;
1249 	struct vndxfer	*vnx;
1250 	daddr_t		nbn;
1251 	char 		*addr;
1252 	off_t		byteoff;
1253 	int		s, off, nra, error, sz, resid;
1254 	UVMHIST_FUNC("sw_reg_strategy"); UVMHIST_CALLED(pdhist);
1255 
1256 	/*
1257 	 * allocate a vndxfer head for this transfer and point it to
1258 	 * our buffer.
1259 	 */
1260 	vnx = pool_get(&vndxfer_pool, PR_WAITOK);
1261 	vnx->vx_flags = VX_BUSY;
1262 	vnx->vx_error = 0;
1263 	vnx->vx_pending = 0;
1264 	vnx->vx_bp = bp;
1265 	vnx->vx_sdp = sdp;
1266 
1267 	/*
1268 	 * setup for main loop where we read filesystem blocks into
1269 	 * our buffer.
1270 	 */
1271 	error = 0;
1272 	bp->b_resid = bp->b_bcount;	/* nothing transfered yet! */
1273 	addr = bp->b_data;		/* current position in buffer */
1274 	byteoff = dbtob((uint64_t)bn);
1275 
1276 	for (resid = bp->b_resid; resid; resid -= sz) {
1277 		struct vndbuf	*nbp;
1278 
1279 		/*
1280 		 * translate byteoffset into block number.  return values:
1281 		 *   vp = vnode of underlying device
1282 		 *  nbn = new block number (on underlying vnode dev)
1283 		 *  nra = num blocks we can read-ahead (excludes requested
1284 		 *	block)
1285 		 */
1286 		nra = 0;
1287 		error = VOP_BMAP(sdp->swd_vp, byteoff / sdp->swd_bsize,
1288 				 	&vp, &nbn, &nra);
1289 
1290 		if (error == 0 && nbn == (daddr_t)-1) {
1291 			/*
1292 			 * this used to just set error, but that doesn't
1293 			 * do the right thing.  Instead, it causes random
1294 			 * memory errors.  The panic() should remain until
1295 			 * this condition doesn't destabilize the system.
1296 			 */
1297 #if 1
1298 			panic("%s: swap to sparse file", __func__);
1299 #else
1300 			error = EIO;	/* failure */
1301 #endif
1302 		}
1303 
1304 		/*
1305 		 * punt if there was an error or a hole in the file.
1306 		 * we must wait for any i/o ops we have already started
1307 		 * to finish before returning.
1308 		 *
1309 		 * XXX we could deal with holes here but it would be
1310 		 * a hassle (in the write case).
1311 		 */
1312 		if (error) {
1313 			s = splbio();
1314 			vnx->vx_error = error;	/* pass error up */
1315 			goto out;
1316 		}
1317 
1318 		/*
1319 		 * compute the size ("sz") of this transfer (in bytes).
1320 		 */
1321 		off = byteoff % sdp->swd_bsize;
1322 		sz = (1 + nra) * sdp->swd_bsize - off;
1323 		if (sz > resid)
1324 			sz = resid;
1325 
1326 		UVMHIST_LOG(pdhist, "sw_reg_strategy: "
1327 			    "vp %p/%p offset 0x%x/0x%x",
1328 			    sdp->swd_vp, vp, byteoff, nbn);
1329 
1330 		/*
1331 		 * now get a buf structure.   note that the vb_buf is
1332 		 * at the front of the nbp structure so that you can
1333 		 * cast pointers between the two structure easily.
1334 		 */
1335 		nbp = pool_get(&vndbuf_pool, PR_WAITOK);
1336 		buf_init(&nbp->vb_buf);
1337 		nbp->vb_buf.b_flags    = bp->b_flags;
1338 		nbp->vb_buf.b_cflags   = bp->b_cflags;
1339 		nbp->vb_buf.b_oflags   = bp->b_oflags;
1340 		nbp->vb_buf.b_bcount   = sz;
1341 		nbp->vb_buf.b_bufsize  = sz;
1342 		nbp->vb_buf.b_error    = 0;
1343 		nbp->vb_buf.b_data     = addr;
1344 		nbp->vb_buf.b_lblkno   = 0;
1345 		nbp->vb_buf.b_blkno    = nbn + btodb(off);
1346 		nbp->vb_buf.b_rawblkno = nbp->vb_buf.b_blkno;
1347 		nbp->vb_buf.b_iodone   = sw_reg_biodone;
1348 		nbp->vb_buf.b_vp       = vp;
1349 		nbp->vb_buf.b_objlock  = vp->v_interlock;
1350 		if (vp->v_type == VBLK) {
1351 			nbp->vb_buf.b_dev = vp->v_rdev;
1352 		}
1353 
1354 		nbp->vb_xfer = vnx;	/* patch it back in to vnx */
1355 
1356 		/*
1357 		 * Just sort by block number
1358 		 */
1359 		s = splbio();
1360 		if (vnx->vx_error != 0) {
1361 			buf_destroy(&nbp->vb_buf);
1362 			pool_put(&vndbuf_pool, nbp);
1363 			goto out;
1364 		}
1365 		vnx->vx_pending++;
1366 
1367 		/* sort it in and start I/O if we are not over our limit */
1368 		/* XXXAD locking */
1369 		bufq_put(sdp->swd_tab, &nbp->vb_buf);
1370 		sw_reg_start(sdp);
1371 		splx(s);
1372 
1373 		/*
1374 		 * advance to the next I/O
1375 		 */
1376 		byteoff += sz;
1377 		addr += sz;
1378 	}
1379 
1380 	s = splbio();
1381 
1382 out: /* Arrive here at splbio */
1383 	vnx->vx_flags &= ~VX_BUSY;
1384 	if (vnx->vx_pending == 0) {
1385 		error = vnx->vx_error;
1386 		pool_put(&vndxfer_pool, vnx);
1387 		bp->b_error = error;
1388 		biodone(bp);
1389 	}
1390 	splx(s);
1391 }
1392 
1393 /*
1394  * sw_reg_start: start an I/O request on the requested swapdev
1395  *
1396  * => reqs are sorted by b_rawblkno (above)
1397  */
1398 static void
1399 sw_reg_start(struct swapdev *sdp)
1400 {
1401 	struct buf	*bp;
1402 	struct vnode	*vp;
1403 	UVMHIST_FUNC("sw_reg_start"); UVMHIST_CALLED(pdhist);
1404 
1405 	/* recursion control */
1406 	if ((sdp->swd_flags & SWF_BUSY) != 0)
1407 		return;
1408 
1409 	sdp->swd_flags |= SWF_BUSY;
1410 
1411 	while (sdp->swd_active < sdp->swd_maxactive) {
1412 		bp = bufq_get(sdp->swd_tab);
1413 		if (bp == NULL)
1414 			break;
1415 		sdp->swd_active++;
1416 
1417 		UVMHIST_LOG(pdhist,
1418 		    "sw_reg_start:  bp %p vp %p blkno %p cnt %lx",
1419 		    bp, bp->b_vp, bp->b_blkno, bp->b_bcount);
1420 		vp = bp->b_vp;
1421 		KASSERT(bp->b_objlock == vp->v_interlock);
1422 		if ((bp->b_flags & B_READ) == 0) {
1423 			mutex_enter(vp->v_interlock);
1424 			vp->v_numoutput++;
1425 			mutex_exit(vp->v_interlock);
1426 		}
1427 		VOP_STRATEGY(vp, bp);
1428 	}
1429 	sdp->swd_flags &= ~SWF_BUSY;
1430 }
1431 
1432 /*
1433  * sw_reg_biodone: one of our i/o's has completed
1434  */
1435 static void
1436 sw_reg_biodone(struct buf *bp)
1437 {
1438 	workqueue_enqueue(sw_reg_workqueue, &bp->b_work, NULL);
1439 }
1440 
1441 /*
1442  * sw_reg_iodone: one of our i/o's has completed and needs post-i/o cleanup
1443  *
1444  * => note that we can recover the vndbuf struct by casting the buf ptr
1445  */
1446 static void
1447 sw_reg_iodone(struct work *wk, void *dummy)
1448 {
1449 	struct vndbuf *vbp = (void *)wk;
1450 	struct vndxfer *vnx = vbp->vb_xfer;
1451 	struct buf *pbp = vnx->vx_bp;		/* parent buffer */
1452 	struct swapdev	*sdp = vnx->vx_sdp;
1453 	int s, resid, error;
1454 	KASSERT(&vbp->vb_buf.b_work == wk);
1455 	UVMHIST_FUNC("sw_reg_iodone"); UVMHIST_CALLED(pdhist);
1456 
1457 	UVMHIST_LOG(pdhist, "  vbp=%p vp=%p blkno=%x addr=%p",
1458 	    vbp, vbp->vb_buf.b_vp, vbp->vb_buf.b_blkno, vbp->vb_buf.b_data);
1459 	UVMHIST_LOG(pdhist, "  cnt=%lx resid=%lx",
1460 	    vbp->vb_buf.b_bcount, vbp->vb_buf.b_resid, 0, 0);
1461 
1462 	/*
1463 	 * protect vbp at splbio and update.
1464 	 */
1465 
1466 	s = splbio();
1467 	resid = vbp->vb_buf.b_bcount - vbp->vb_buf.b_resid;
1468 	pbp->b_resid -= resid;
1469 	vnx->vx_pending--;
1470 
1471 	if (vbp->vb_buf.b_error != 0) {
1472 		/* pass error upward */
1473 		error = vbp->vb_buf.b_error ? vbp->vb_buf.b_error : EIO;
1474 		UVMHIST_LOG(pdhist, "  got error=%d !", error, 0, 0, 0);
1475 		vnx->vx_error = error;
1476 	}
1477 
1478 	/*
1479 	 * kill vbp structure
1480 	 */
1481 	buf_destroy(&vbp->vb_buf);
1482 	pool_put(&vndbuf_pool, vbp);
1483 
1484 	/*
1485 	 * wrap up this transaction if it has run to completion or, in
1486 	 * case of an error, when all auxiliary buffers have returned.
1487 	 */
1488 	if (vnx->vx_error != 0) {
1489 		/* pass error upward */
1490 		error = vnx->vx_error;
1491 		if ((vnx->vx_flags & VX_BUSY) == 0 && vnx->vx_pending == 0) {
1492 			pbp->b_error = error;
1493 			biodone(pbp);
1494 			pool_put(&vndxfer_pool, vnx);
1495 		}
1496 	} else if (pbp->b_resid == 0) {
1497 		KASSERT(vnx->vx_pending == 0);
1498 		if ((vnx->vx_flags & VX_BUSY) == 0) {
1499 			UVMHIST_LOG(pdhist, "  iodone error=%d !",
1500 			    pbp, vnx->vx_error, 0, 0);
1501 			biodone(pbp);
1502 			pool_put(&vndxfer_pool, vnx);
1503 		}
1504 	}
1505 
1506 	/*
1507 	 * done!   start next swapdev I/O if one is pending
1508 	 */
1509 	sdp->swd_active--;
1510 	sw_reg_start(sdp);
1511 	splx(s);
1512 }
1513 
1514 
1515 /*
1516  * uvm_swap_alloc: allocate space on swap
1517  *
1518  * => allocation is done "round robin" down the priority list, as we
1519  *	allocate in a priority we "rotate" the circle queue.
1520  * => space can be freed with uvm_swap_free
1521  * => we return the page slot number in /dev/drum (0 == invalid slot)
1522  * => we lock uvm_swap_data_lock
1523  * => XXXMRG: "LESSOK" INTERFACE NEEDED TO EXTENT SYSTEM
1524  */
1525 int
1526 uvm_swap_alloc(int *nslots /* IN/OUT */, bool lessok)
1527 {
1528 	struct swapdev *sdp;
1529 	struct swappri *spp;
1530 	UVMHIST_FUNC("uvm_swap_alloc"); UVMHIST_CALLED(pdhist);
1531 
1532 	/*
1533 	 * no swap devices configured yet?   definite failure.
1534 	 */
1535 	if (uvmexp.nswapdev < 1)
1536 		return 0;
1537 
1538 	/*
1539 	 * XXXJAK: BEGIN HACK
1540 	 *
1541 	 * blist_alloc() in subr_blist.c will panic if we try to allocate
1542 	 * too many slots.
1543 	 */
1544 	if (*nslots > BLIST_MAX_ALLOC) {
1545 		if (__predict_false(lessok == false))
1546 			return 0;
1547 		*nslots = BLIST_MAX_ALLOC;
1548 	}
1549 	/* XXXJAK: END HACK */
1550 
1551 	/*
1552 	 * lock data lock, convert slots into blocks, and enter loop
1553 	 */
1554 	mutex_enter(&uvm_swap_data_lock);
1555 
1556 ReTry:	/* XXXMRG */
1557 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
1558 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
1559 			uint64_t result;
1560 
1561 			/* if it's not enabled, then we can't swap from it */
1562 			if ((sdp->swd_flags & SWF_ENABLE) == 0)
1563 				continue;
1564 			if (sdp->swd_npginuse + *nslots > sdp->swd_npages)
1565 				continue;
1566 			result = blist_alloc(sdp->swd_blist, *nslots);
1567 			if (result == BLIST_NONE) {
1568 				continue;
1569 			}
1570 			KASSERT(result < sdp->swd_drumsize);
1571 
1572 			/*
1573 			 * successful allocation!  now rotate the circleq.
1574 			 */
1575 			CIRCLEQ_REMOVE(&spp->spi_swapdev, sdp, swd_next);
1576 			CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
1577 			sdp->swd_npginuse += *nslots;
1578 			uvmexp.swpginuse += *nslots;
1579 			mutex_exit(&uvm_swap_data_lock);
1580 			/* done!  return drum slot number */
1581 			UVMHIST_LOG(pdhist,
1582 			    "success!  returning %d slots starting at %d",
1583 			    *nslots, result + sdp->swd_drumoffset, 0, 0);
1584 			return (result + sdp->swd_drumoffset);
1585 		}
1586 	}
1587 
1588 	/* XXXMRG: BEGIN HACK */
1589 	if (*nslots > 1 && lessok) {
1590 		*nslots = 1;
1591 		/* XXXMRG: ugh!  blist should support this for us */
1592 		goto ReTry;
1593 	}
1594 	/* XXXMRG: END HACK */
1595 
1596 	mutex_exit(&uvm_swap_data_lock);
1597 	return 0;
1598 }
1599 
1600 /*
1601  * uvm_swapisfull: return true if most of available swap is allocated
1602  * and in use.  we don't count some small portion as it may be inaccessible
1603  * to us at any given moment, for example if there is lock contention or if
1604  * pages are busy.
1605  */
1606 bool
1607 uvm_swapisfull(void)
1608 {
1609 	int swpgonly;
1610 	bool rv;
1611 
1612 	mutex_enter(&uvm_swap_data_lock);
1613 	KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1614 	swpgonly = (int)((uint64_t)uvmexp.swpgonly * 100 /
1615 	    uvm_swapisfull_factor);
1616 	rv = (swpgonly >= uvmexp.swpgavail);
1617 	mutex_exit(&uvm_swap_data_lock);
1618 
1619 	return (rv);
1620 }
1621 
1622 /*
1623  * uvm_swap_markbad: keep track of swap ranges where we've had i/o errors
1624  *
1625  * => we lock uvm_swap_data_lock
1626  */
1627 void
1628 uvm_swap_markbad(int startslot, int nslots)
1629 {
1630 	struct swapdev *sdp;
1631 	UVMHIST_FUNC("uvm_swap_markbad"); UVMHIST_CALLED(pdhist);
1632 
1633 	mutex_enter(&uvm_swap_data_lock);
1634 	sdp = swapdrum_getsdp(startslot);
1635 	KASSERT(sdp != NULL);
1636 
1637 	/*
1638 	 * we just keep track of how many pages have been marked bad
1639 	 * in this device, to make everything add up in swap_off().
1640 	 * we assume here that the range of slots will all be within
1641 	 * one swap device.
1642 	 */
1643 
1644 	KASSERT(uvmexp.swpgonly >= nslots);
1645 	uvmexp.swpgonly -= nslots;
1646 	sdp->swd_npgbad += nslots;
1647 	UVMHIST_LOG(pdhist, "now %d bad", sdp->swd_npgbad, 0,0,0);
1648 	mutex_exit(&uvm_swap_data_lock);
1649 }
1650 
1651 /*
1652  * uvm_swap_free: free swap slots
1653  *
1654  * => this can be all or part of an allocation made by uvm_swap_alloc
1655  * => we lock uvm_swap_data_lock
1656  */
1657 void
1658 uvm_swap_free(int startslot, int nslots)
1659 {
1660 	struct swapdev *sdp;
1661 	UVMHIST_FUNC("uvm_swap_free"); UVMHIST_CALLED(pdhist);
1662 
1663 	UVMHIST_LOG(pdhist, "freeing %d slots starting at %d", nslots,
1664 	    startslot, 0, 0);
1665 
1666 	/*
1667 	 * ignore attempts to free the "bad" slot.
1668 	 */
1669 
1670 	if (startslot == SWSLOT_BAD) {
1671 		return;
1672 	}
1673 
1674 	/*
1675 	 * convert drum slot offset back to sdp, free the blocks
1676 	 * in the extent, and return.   must hold pri lock to do
1677 	 * lookup and access the extent.
1678 	 */
1679 
1680 	mutex_enter(&uvm_swap_data_lock);
1681 	sdp = swapdrum_getsdp(startslot);
1682 	KASSERT(uvmexp.nswapdev >= 1);
1683 	KASSERT(sdp != NULL);
1684 	KASSERT(sdp->swd_npginuse >= nslots);
1685 	blist_free(sdp->swd_blist, startslot - sdp->swd_drumoffset, nslots);
1686 	sdp->swd_npginuse -= nslots;
1687 	uvmexp.swpginuse -= nslots;
1688 	mutex_exit(&uvm_swap_data_lock);
1689 }
1690 
1691 /*
1692  * uvm_swap_put: put any number of pages into a contig place on swap
1693  *
1694  * => can be sync or async
1695  */
1696 
1697 int
1698 uvm_swap_put(int swslot, struct vm_page **ppsp, int npages, int flags)
1699 {
1700 	int error;
1701 
1702 	error = uvm_swap_io(ppsp, swslot, npages, B_WRITE |
1703 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1704 	return error;
1705 }
1706 
1707 /*
1708  * uvm_swap_get: get a single page from swap
1709  *
1710  * => usually a sync op (from fault)
1711  */
1712 
1713 int
1714 uvm_swap_get(struct vm_page *page, int swslot, int flags)
1715 {
1716 	int error;
1717 
1718 	uvmexp.nswget++;
1719 	KASSERT(flags & PGO_SYNCIO);
1720 	if (swslot == SWSLOT_BAD) {
1721 		return EIO;
1722 	}
1723 
1724 	error = uvm_swap_io(&page, swslot, 1, B_READ |
1725 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1726 	if (error == 0) {
1727 
1728 		/*
1729 		 * this page is no longer only in swap.
1730 		 */
1731 
1732 		mutex_enter(&uvm_swap_data_lock);
1733 		KASSERT(uvmexp.swpgonly > 0);
1734 		uvmexp.swpgonly--;
1735 		mutex_exit(&uvm_swap_data_lock);
1736 	}
1737 	return error;
1738 }
1739 
1740 /*
1741  * uvm_swap_io: do an i/o operation to swap
1742  */
1743 
1744 static int
1745 uvm_swap_io(struct vm_page **pps, int startslot, int npages, int flags)
1746 {
1747 	daddr_t startblk;
1748 	struct	buf *bp;
1749 	vaddr_t kva;
1750 	int	error, mapinflags;
1751 	bool write, async;
1752 	UVMHIST_FUNC("uvm_swap_io"); UVMHIST_CALLED(pdhist);
1753 
1754 	UVMHIST_LOG(pdhist, "<- called, startslot=%d, npages=%d, flags=%d",
1755 	    startslot, npages, flags, 0);
1756 
1757 	write = (flags & B_READ) == 0;
1758 	async = (flags & B_ASYNC) != 0;
1759 
1760 	/*
1761 	 * allocate a buf for the i/o.
1762 	 */
1763 
1764 	KASSERT(curlwp != uvm.pagedaemon_lwp || (write && async));
1765 	bp = getiobuf(swapdev_vp, curlwp != uvm.pagedaemon_lwp);
1766 	if (bp == NULL) {
1767 		uvm_aio_aiodone_pages(pps, npages, true, ENOMEM);
1768 		return ENOMEM;
1769 	}
1770 
1771 	/*
1772 	 * convert starting drum slot to block number
1773 	 */
1774 
1775 	startblk = btodb((uint64_t)startslot << PAGE_SHIFT);
1776 
1777 	/*
1778 	 * first, map the pages into the kernel.
1779 	 */
1780 
1781 	mapinflags = !write ?
1782 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_READ :
1783 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_WRITE;
1784 	kva = uvm_pagermapin(pps, npages, mapinflags);
1785 
1786 	/*
1787 	 * fill in the bp/sbp.   we currently route our i/o through
1788 	 * /dev/drum's vnode [swapdev_vp].
1789 	 */
1790 
1791 	bp->b_cflags = BC_BUSY | BC_NOCACHE;
1792 	bp->b_flags = (flags & (B_READ|B_ASYNC));
1793 	bp->b_proc = &proc0;	/* XXX */
1794 	bp->b_vnbufs.le_next = NOLIST;
1795 	bp->b_data = (void *)kva;
1796 	bp->b_blkno = startblk;
1797 	bp->b_bufsize = bp->b_bcount = npages << PAGE_SHIFT;
1798 
1799 	/*
1800 	 * bump v_numoutput (counter of number of active outputs).
1801 	 */
1802 
1803 	if (write) {
1804 		mutex_enter(swapdev_vp->v_interlock);
1805 		swapdev_vp->v_numoutput++;
1806 		mutex_exit(swapdev_vp->v_interlock);
1807 	}
1808 
1809 	/*
1810 	 * for async ops we must set up the iodone handler.
1811 	 */
1812 
1813 	if (async) {
1814 		bp->b_iodone = uvm_aio_biodone;
1815 		UVMHIST_LOG(pdhist, "doing async!", 0, 0, 0, 0);
1816 		if (curlwp == uvm.pagedaemon_lwp)
1817 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1818 		else
1819 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
1820 	} else {
1821 		bp->b_iodone = NULL;
1822 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1823 	}
1824 	UVMHIST_LOG(pdhist,
1825 	    "about to start io: data = %p blkno = 0x%x, bcount = %ld",
1826 	    bp->b_data, bp->b_blkno, bp->b_bcount, 0);
1827 
1828 	/*
1829 	 * now we start the I/O, and if async, return.
1830 	 */
1831 
1832 	VOP_STRATEGY(swapdev_vp, bp);
1833 	if (async)
1834 		return 0;
1835 
1836 	/*
1837 	 * must be sync i/o.   wait for it to finish
1838 	 */
1839 
1840 	error = biowait(bp);
1841 
1842 	/*
1843 	 * kill the pager mapping
1844 	 */
1845 
1846 	uvm_pagermapout(kva, npages);
1847 
1848 	/*
1849 	 * now dispose of the buf and we're done.
1850 	 */
1851 
1852 	if (write) {
1853 		mutex_enter(swapdev_vp->v_interlock);
1854 		vwakeup(bp);
1855 		mutex_exit(swapdev_vp->v_interlock);
1856 	}
1857 	putiobuf(bp);
1858 	UVMHIST_LOG(pdhist, "<- done (sync)  error=%d", error, 0, 0, 0);
1859 
1860 	return (error);
1861 }
1862