xref: /netbsd-src/sys/uvm/uvm_swap.c (revision 404fbe5fb94ca1e054339640cabb2801ce52dd30)
1 /*	$NetBSD: uvm_swap.c,v 1.142 2008/12/17 20:51:39 cegger Exp $	*/
2 
3 /*
4  * Copyright (c) 1995, 1996, 1997 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 1.142 2008/12/17 20:51:39 cegger Exp $");
34 
35 #include "fs_nfs.h"
36 #include "opt_uvmhist.h"
37 #include "opt_compat_netbsd.h"
38 #include "opt_ddb.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/buf.h>
43 #include <sys/bufq.h>
44 #include <sys/conf.h>
45 #include <sys/proc.h>
46 #include <sys/namei.h>
47 #include <sys/disklabel.h>
48 #include <sys/errno.h>
49 #include <sys/kernel.h>
50 #include <sys/malloc.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/syscallargs.h>
58 #include <sys/swap.h>
59 #include <sys/kauth.h>
60 #include <sys/sysctl.h>
61 #include <sys/workqueue.h>
62 
63 #include <uvm/uvm.h>
64 
65 #include <miscfs/specfs/specdev.h>
66 
67 /*
68  * uvm_swap.c: manage configuration and i/o to swap space.
69  */
70 
71 /*
72  * swap space is managed in the following way:
73  *
74  * each swap partition or file is described by a "swapdev" structure.
75  * each "swapdev" structure contains a "swapent" structure which contains
76  * information that is passed up to the user (via system calls).
77  *
78  * each swap partition is assigned a "priority" (int) which controls
79  * swap parition usage.
80  *
81  * the system maintains a global data structure describing all swap
82  * partitions/files.   there is a sorted LIST of "swappri" structures
83  * which describe "swapdev"'s at that priority.   this LIST is headed
84  * by the "swap_priority" global var.    each "swappri" contains a
85  * CIRCLEQ of "swapdev" structures at that priority.
86  *
87  * locking:
88  *  - swap_syscall_lock (krwlock_t): this lock serializes the swapctl
89  *    system call and prevents the swap priority list from changing
90  *    while we are in the middle of a system call (e.g. SWAP_STATS).
91  *  - uvm_swap_data_lock (kmutex_t): this lock protects all swap data
92  *    structures including the priority list, the swapdev structures,
93  *    and the swapmap arena.
94  *
95  * each swap device has the following info:
96  *  - swap device in use (could be disabled, preventing future use)
97  *  - swap enabled (allows new allocations on swap)
98  *  - map info in /dev/drum
99  *  - vnode pointer
100  * for swap files only:
101  *  - block size
102  *  - max byte count in buffer
103  *  - buffer
104  *
105  * userland controls and configures swap with the swapctl(2) system call.
106  * the sys_swapctl performs the following operations:
107  *  [1] SWAP_NSWAP: returns the number of swap devices currently configured
108  *  [2] SWAP_STATS: given a pointer to an array of swapent structures
109  *	(passed in via "arg") of a size passed in via "misc" ... we load
110  *	the current swap config into the array. The actual work is done
111  *	in the uvm_swap_stats(9) function.
112  *  [3] SWAP_ON: given a pathname in arg (could be device or file) and a
113  *	priority in "misc", start swapping on it.
114  *  [4] SWAP_OFF: as SWAP_ON, but stops swapping to a device
115  *  [5] SWAP_CTL: changes the priority of a swap device (new priority in
116  *	"misc")
117  */
118 
119 /*
120  * swapdev: describes a single swap partition/file
121  *
122  * note the following should be true:
123  * swd_inuse <= swd_nblks  [number of blocks in use is <= total blocks]
124  * swd_nblks <= swd_mapsize [because mapsize includes miniroot+disklabel]
125  */
126 struct swapdev {
127 	struct oswapent swd_ose;
128 #define	swd_dev		swd_ose.ose_dev		/* device id */
129 #define	swd_flags	swd_ose.ose_flags	/* flags:inuse/enable/fake */
130 #define	swd_priority	swd_ose.ose_priority	/* our priority */
131 	/* also: swd_ose.ose_nblks, swd_ose.ose_inuse */
132 	char			*swd_path;	/* saved pathname of device */
133 	int			swd_pathlen;	/* length of pathname */
134 	int			swd_npages;	/* #pages we can use */
135 	int			swd_npginuse;	/* #pages in use */
136 	int			swd_npgbad;	/* #pages bad */
137 	int			swd_drumoffset;	/* page0 offset in drum */
138 	int			swd_drumsize;	/* #pages in drum */
139 	blist_t			swd_blist;	/* blist for this swapdev */
140 	struct vnode		*swd_vp;	/* backing vnode */
141 	CIRCLEQ_ENTRY(swapdev)	swd_next;	/* priority circleq */
142 
143 	int			swd_bsize;	/* blocksize (bytes) */
144 	int			swd_maxactive;	/* max active i/o reqs */
145 	struct bufq_state	*swd_tab;	/* buffer list */
146 	int			swd_active;	/* number of active buffers */
147 };
148 
149 /*
150  * swap device priority entry; the list is kept sorted on `spi_priority'.
151  */
152 struct swappri {
153 	int			spi_priority;     /* priority */
154 	CIRCLEQ_HEAD(spi_swapdev, swapdev)	spi_swapdev;
155 	/* circleq of swapdevs at this priority */
156 	LIST_ENTRY(swappri)	spi_swappri;      /* global list of pri's */
157 };
158 
159 /*
160  * The following two structures are used to keep track of data transfers
161  * on swap devices associated with regular files.
162  * NOTE: this code is more or less a copy of vnd.c; we use the same
163  * structure names here to ease porting..
164  */
165 struct vndxfer {
166 	struct buf	*vx_bp;		/* Pointer to parent buffer */
167 	struct swapdev	*vx_sdp;
168 	int		vx_error;
169 	int		vx_pending;	/* # of pending aux buffers */
170 	int		vx_flags;
171 #define VX_BUSY		1
172 #define VX_DEAD		2
173 };
174 
175 struct vndbuf {
176 	struct buf	vb_buf;
177 	struct vndxfer	*vb_xfer;
178 };
179 
180 
181 /*
182  * We keep a of pool vndbuf's and vndxfer structures.
183  */
184 POOL_INIT(vndxfer_pool, sizeof(struct vndxfer), 0, 0, 0, "swp vnx", NULL,
185     IPL_BIO);
186 POOL_INIT(vndbuf_pool, sizeof(struct vndbuf), 0, 0, 0, "swp vnd", NULL,
187     IPL_BIO);
188 
189 /*
190  * local variables
191  */
192 MALLOC_DEFINE(M_VMSWAP, "VM swap", "VM swap structures");
193 static vmem_t *swapmap;	/* controls the mapping of /dev/drum */
194 
195 /* list of all active swap devices [by priority] */
196 LIST_HEAD(swap_priority, swappri);
197 static struct swap_priority swap_priority;
198 
199 /* locks */
200 static krwlock_t swap_syscall_lock;
201 
202 /* workqueue and use counter for swap to regular files */
203 static int sw_reg_count = 0;
204 static struct workqueue *sw_reg_workqueue;
205 
206 /* tuneables */
207 u_int uvm_swapisfull_factor = 99;
208 
209 /*
210  * prototypes
211  */
212 static struct swapdev	*swapdrum_getsdp(int);
213 
214 static struct swapdev	*swaplist_find(struct vnode *, bool);
215 static void		 swaplist_insert(struct swapdev *,
216 					 struct swappri *, int);
217 static void		 swaplist_trim(void);
218 
219 static int swap_on(struct lwp *, struct swapdev *);
220 static int swap_off(struct lwp *, struct swapdev *);
221 
222 static void uvm_swap_stats_locked(int, struct swapent *, int, register_t *);
223 
224 static void sw_reg_strategy(struct swapdev *, struct buf *, int);
225 static void sw_reg_biodone(struct buf *);
226 static void sw_reg_iodone(struct work *wk, void *dummy);
227 static void sw_reg_start(struct swapdev *);
228 
229 static int uvm_swap_io(struct vm_page **, int, int, int);
230 
231 /*
232  * uvm_swap_init: init the swap system data structures and locks
233  *
234  * => called at boot time from init_main.c after the filesystems
235  *	are brought up (which happens after uvm_init())
236  */
237 void
238 uvm_swap_init(void)
239 {
240 	UVMHIST_FUNC("uvm_swap_init");
241 
242 	UVMHIST_CALLED(pdhist);
243 	/*
244 	 * first, init the swap list, its counter, and its lock.
245 	 * then get a handle on the vnode for /dev/drum by using
246 	 * the its dev_t number ("swapdev", from MD conf.c).
247 	 */
248 
249 	LIST_INIT(&swap_priority);
250 	uvmexp.nswapdev = 0;
251 	rw_init(&swap_syscall_lock);
252 	cv_init(&uvm.scheduler_cv, "schedule");
253 	mutex_init(&uvm_swap_data_lock, MUTEX_DEFAULT, IPL_NONE);
254 
255 	/* XXXSMP should be at IPL_VM, but for audio interrupt handlers. */
256 	mutex_init(&uvm_scheduler_mutex, MUTEX_SPIN, IPL_SCHED);
257 
258 	if (bdevvp(swapdev, &swapdev_vp))
259 		panic("uvm_swap_init: can't get vnode for swap device");
260 	if (vn_lock(swapdev_vp, LK_EXCLUSIVE | LK_RETRY))
261 		panic("uvm_swap_init: can't lock swap device");
262 	if (VOP_OPEN(swapdev_vp, FREAD | FWRITE, NOCRED))
263 		panic("uvm_swap_init: can't open swap device");
264 	VOP_UNLOCK(swapdev_vp, 0);
265 
266 	/*
267 	 * create swap block resource map to map /dev/drum.   the range
268 	 * from 1 to INT_MAX allows 2 gigablocks of swap space.  note
269 	 * that block 0 is reserved (used to indicate an allocation
270 	 * failure, or no allocation).
271 	 */
272 	swapmap = vmem_create("swapmap", 1, INT_MAX - 1, 1, NULL, NULL, NULL, 0,
273 	    VM_NOSLEEP, IPL_NONE);
274 	if (swapmap == 0)
275 		panic("uvm_swap_init: extent_create failed");
276 
277 	/*
278 	 * done!
279 	 */
280 	uvm.swap_running = true;
281 #ifdef __SWAP_BROKEN
282 	uvm.swapout_enabled = 0;
283 #else
284 	uvm.swapout_enabled = 1;
285 #endif
286 	UVMHIST_LOG(pdhist, "<- done", 0, 0, 0, 0);
287 
288         sysctl_createv(NULL, 0, NULL, NULL,
289             CTLFLAG_READWRITE,
290             CTLTYPE_INT, "swapout",
291             SYSCTL_DESCR("Set 0 to disable swapout of kernel stacks"),
292             NULL, 0, &uvm.swapout_enabled, 0, CTL_VM, CTL_CREATE, CTL_EOL);
293 }
294 
295 /*
296  * swaplist functions: functions that operate on the list of swap
297  * devices on the system.
298  */
299 
300 /*
301  * swaplist_insert: insert swap device "sdp" into the global list
302  *
303  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
304  * => caller must provide a newly malloc'd swappri structure (we will
305  *	FREE it if we don't need it... this it to prevent malloc blocking
306  *	here while adding swap)
307  */
308 static void
309 swaplist_insert(struct swapdev *sdp, struct swappri *newspp, int priority)
310 {
311 	struct swappri *spp, *pspp;
312 	UVMHIST_FUNC("swaplist_insert"); UVMHIST_CALLED(pdhist);
313 
314 	/*
315 	 * find entry at or after which to insert the new device.
316 	 */
317 	pspp = NULL;
318 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
319 		if (priority <= spp->spi_priority)
320 			break;
321 		pspp = spp;
322 	}
323 
324 	/*
325 	 * new priority?
326 	 */
327 	if (spp == NULL || spp->spi_priority != priority) {
328 		spp = newspp;  /* use newspp! */
329 		UVMHIST_LOG(pdhist, "created new swappri = %d",
330 			    priority, 0, 0, 0);
331 
332 		spp->spi_priority = priority;
333 		CIRCLEQ_INIT(&spp->spi_swapdev);
334 
335 		if (pspp)
336 			LIST_INSERT_AFTER(pspp, spp, spi_swappri);
337 		else
338 			LIST_INSERT_HEAD(&swap_priority, spp, spi_swappri);
339 	} else {
340 	  	/* we don't need a new priority structure, free it */
341 		free(newspp, M_VMSWAP);
342 	}
343 
344 	/*
345 	 * priority found (or created).   now insert on the priority's
346 	 * circleq list and bump the total number of swapdevs.
347 	 */
348 	sdp->swd_priority = priority;
349 	CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
350 	uvmexp.nswapdev++;
351 }
352 
353 /*
354  * swaplist_find: find and optionally remove a swap device from the
355  *	global list.
356  *
357  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
358  * => we return the swapdev we found (and removed)
359  */
360 static struct swapdev *
361 swaplist_find(struct vnode *vp, bool remove)
362 {
363 	struct swapdev *sdp;
364 	struct swappri *spp;
365 
366 	/*
367 	 * search the lists for the requested vp
368 	 */
369 
370 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
371 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
372 			if (sdp->swd_vp == vp) {
373 				if (remove) {
374 					CIRCLEQ_REMOVE(&spp->spi_swapdev,
375 					    sdp, swd_next);
376 					uvmexp.nswapdev--;
377 				}
378 				return(sdp);
379 			}
380 		}
381 	}
382 	return (NULL);
383 }
384 
385 /*
386  * swaplist_trim: scan priority list for empty priority entries and kill
387  *	them.
388  *
389  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
390  */
391 static void
392 swaplist_trim(void)
393 {
394 	struct swappri *spp, *nextspp;
395 
396 	for (spp = LIST_FIRST(&swap_priority); spp != NULL; spp = nextspp) {
397 		nextspp = LIST_NEXT(spp, spi_swappri);
398 		if (CIRCLEQ_FIRST(&spp->spi_swapdev) !=
399 		    (void *)&spp->spi_swapdev)
400 			continue;
401 		LIST_REMOVE(spp, spi_swappri);
402 		free(spp, M_VMSWAP);
403 	}
404 }
405 
406 /*
407  * swapdrum_getsdp: given a page offset in /dev/drum, convert it back
408  *	to the "swapdev" that maps that section of the drum.
409  *
410  * => each swapdev takes one big contig chunk of the drum
411  * => caller must hold uvm_swap_data_lock
412  */
413 static struct swapdev *
414 swapdrum_getsdp(int pgno)
415 {
416 	struct swapdev *sdp;
417 	struct swappri *spp;
418 
419 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
420 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
421 			if (sdp->swd_flags & SWF_FAKE)
422 				continue;
423 			if (pgno >= sdp->swd_drumoffset &&
424 			    pgno < (sdp->swd_drumoffset + sdp->swd_drumsize)) {
425 				return sdp;
426 			}
427 		}
428 	}
429 	return NULL;
430 }
431 
432 
433 /*
434  * sys_swapctl: main entry point for swapctl(2) system call
435  * 	[with two helper functions: swap_on and swap_off]
436  */
437 int
438 sys_swapctl(struct lwp *l, const struct sys_swapctl_args *uap, register_t *retval)
439 {
440 	/* {
441 		syscallarg(int) cmd;
442 		syscallarg(void *) arg;
443 		syscallarg(int) misc;
444 	} */
445 	struct vnode *vp;
446 	struct nameidata nd;
447 	struct swappri *spp;
448 	struct swapdev *sdp;
449 	struct swapent *sep;
450 #define SWAP_PATH_MAX (PATH_MAX + 1)
451 	char	*userpath;
452 	size_t	len;
453 	int	error, misc;
454 	int	priority;
455 	UVMHIST_FUNC("sys_swapctl"); UVMHIST_CALLED(pdhist);
456 
457 	misc = SCARG(uap, misc);
458 
459 	/*
460 	 * ensure serialized syscall access by grabbing the swap_syscall_lock
461 	 */
462 	rw_enter(&swap_syscall_lock, RW_WRITER);
463 
464 	userpath = malloc(SWAP_PATH_MAX, M_TEMP, M_WAITOK);
465 	/*
466 	 * we handle the non-priv NSWAP and STATS request first.
467 	 *
468 	 * SWAP_NSWAP: return number of config'd swap devices
469 	 * [can also be obtained with uvmexp sysctl]
470 	 */
471 	if (SCARG(uap, cmd) == SWAP_NSWAP) {
472 		UVMHIST_LOG(pdhist, "<- done SWAP_NSWAP=%d", uvmexp.nswapdev,
473 		    0, 0, 0);
474 		*retval = uvmexp.nswapdev;
475 		error = 0;
476 		goto out;
477 	}
478 
479 	/*
480 	 * SWAP_STATS: get stats on current # of configured swap devs
481 	 *
482 	 * note that the swap_priority list can't change as long
483 	 * as we are holding the swap_syscall_lock.  we don't want
484 	 * to grab the uvm_swap_data_lock because we may fault&sleep during
485 	 * copyout() and we don't want to be holding that lock then!
486 	 */
487 	if (SCARG(uap, cmd) == SWAP_STATS
488 #if defined(COMPAT_13)
489 	    || SCARG(uap, cmd) == SWAP_OSTATS
490 #endif
491 	    ) {
492 		if ((size_t)misc > (size_t)uvmexp.nswapdev)
493 			misc = uvmexp.nswapdev;
494 #if defined(COMPAT_13)
495 		if (SCARG(uap, cmd) == SWAP_OSTATS)
496 			len = sizeof(struct oswapent) * misc;
497 		else
498 #endif
499 			len = sizeof(struct swapent) * misc;
500 		sep = (struct swapent *)malloc(len, M_TEMP, M_WAITOK);
501 
502 		uvm_swap_stats_locked(SCARG(uap, cmd), sep, misc, retval);
503 		error = copyout(sep, SCARG(uap, arg), len);
504 
505 		free(sep, M_TEMP);
506 		UVMHIST_LOG(pdhist, "<- done SWAP_STATS", 0, 0, 0, 0);
507 		goto out;
508 	}
509 	if (SCARG(uap, cmd) == SWAP_GETDUMPDEV) {
510 		dev_t	*devp = (dev_t *)SCARG(uap, arg);
511 
512 		error = copyout(&dumpdev, devp, sizeof(dumpdev));
513 		goto out;
514 	}
515 
516 	/*
517 	 * all other requests require superuser privs.   verify.
518 	 */
519 	if ((error = kauth_authorize_system(l->l_cred, KAUTH_SYSTEM_SWAPCTL,
520 	    0, NULL, NULL, NULL)))
521 		goto out;
522 
523 	if (SCARG(uap, cmd) == SWAP_DUMPOFF) {
524 		/* drop the current dump device */
525 		dumpdev = NODEV;
526 		dumpcdev = NODEV;
527 		cpu_dumpconf();
528 		goto out;
529 	}
530 
531 	/*
532 	 * at this point we expect a path name in arg.   we will
533 	 * use namei() to gain a vnode reference (vref), and lock
534 	 * the vnode (VOP_LOCK).
535 	 *
536 	 * XXX: a NULL arg means use the root vnode pointer (e.g. for
537 	 * miniroot)
538 	 */
539 	if (SCARG(uap, arg) == NULL) {
540 		vp = rootvp;		/* miniroot */
541 		if (vget(vp, LK_EXCLUSIVE)) {
542 			error = EBUSY;
543 			goto out;
544 		}
545 		if (SCARG(uap, cmd) == SWAP_ON &&
546 		    copystr("miniroot", userpath, SWAP_PATH_MAX, &len))
547 			panic("swapctl: miniroot copy failed");
548 	} else {
549 		int	space;
550 		char	*where;
551 
552 		if (SCARG(uap, cmd) == SWAP_ON) {
553 			if ((error = copyinstr(SCARG(uap, arg), userpath,
554 			    SWAP_PATH_MAX, &len)))
555 				goto out;
556 			space = UIO_SYSSPACE;
557 			where = userpath;
558 		} else {
559 			space = UIO_USERSPACE;
560 			where = (char *)SCARG(uap, arg);
561 		}
562 		NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | TRYEMULROOT,
563 		    space, where);
564 		if ((error = namei(&nd)))
565 			goto out;
566 		vp = nd.ni_vp;
567 	}
568 	/* note: "vp" is referenced and locked */
569 
570 	error = 0;		/* assume no error */
571 	switch(SCARG(uap, cmd)) {
572 
573 	case SWAP_DUMPDEV:
574 		if (vp->v_type != VBLK) {
575 			error = ENOTBLK;
576 			break;
577 		}
578 		if (bdevsw_lookup(vp->v_rdev)) {
579 			dumpdev = vp->v_rdev;
580 			dumpcdev = devsw_blk2chr(dumpdev);
581 		} else
582 			dumpdev = NODEV;
583 		cpu_dumpconf();
584 		break;
585 
586 	case SWAP_CTL:
587 		/*
588 		 * get new priority, remove old entry (if any) and then
589 		 * reinsert it in the correct place.  finally, prune out
590 		 * any empty priority structures.
591 		 */
592 		priority = SCARG(uap, misc);
593 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
594 		mutex_enter(&uvm_swap_data_lock);
595 		if ((sdp = swaplist_find(vp, true)) == NULL) {
596 			error = ENOENT;
597 		} else {
598 			swaplist_insert(sdp, spp, priority);
599 			swaplist_trim();
600 		}
601 		mutex_exit(&uvm_swap_data_lock);
602 		if (error)
603 			free(spp, M_VMSWAP);
604 		break;
605 
606 	case SWAP_ON:
607 
608 		/*
609 		 * check for duplicates.   if none found, then insert a
610 		 * dummy entry on the list to prevent someone else from
611 		 * trying to enable this device while we are working on
612 		 * it.
613 		 */
614 
615 		priority = SCARG(uap, misc);
616 		sdp = malloc(sizeof *sdp, M_VMSWAP, M_WAITOK);
617 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
618 		memset(sdp, 0, sizeof(*sdp));
619 		sdp->swd_flags = SWF_FAKE;
620 		sdp->swd_vp = vp;
621 		sdp->swd_dev = (vp->v_type == VBLK) ? vp->v_rdev : NODEV;
622 		bufq_alloc(&sdp->swd_tab, "disksort", BUFQ_SORT_RAWBLOCK);
623 		mutex_enter(&uvm_swap_data_lock);
624 		if (swaplist_find(vp, false) != NULL) {
625 			error = EBUSY;
626 			mutex_exit(&uvm_swap_data_lock);
627 			bufq_free(sdp->swd_tab);
628 			free(sdp, M_VMSWAP);
629 			free(spp, M_VMSWAP);
630 			break;
631 		}
632 		swaplist_insert(sdp, spp, priority);
633 		mutex_exit(&uvm_swap_data_lock);
634 
635 		sdp->swd_pathlen = len;
636 		sdp->swd_path = malloc(sdp->swd_pathlen, M_VMSWAP, M_WAITOK);
637 		if (copystr(userpath, sdp->swd_path, sdp->swd_pathlen, 0) != 0)
638 			panic("swapctl: copystr");
639 
640 		/*
641 		 * we've now got a FAKE placeholder in the swap list.
642 		 * now attempt to enable swap on it.  if we fail, undo
643 		 * what we've done and kill the fake entry we just inserted.
644 		 * if swap_on is a success, it will clear the SWF_FAKE flag
645 		 */
646 
647 		if ((error = swap_on(l, sdp)) != 0) {
648 			mutex_enter(&uvm_swap_data_lock);
649 			(void) swaplist_find(vp, true);  /* kill fake entry */
650 			swaplist_trim();
651 			mutex_exit(&uvm_swap_data_lock);
652 			bufq_free(sdp->swd_tab);
653 			free(sdp->swd_path, M_VMSWAP);
654 			free(sdp, M_VMSWAP);
655 			break;
656 		}
657 		break;
658 
659 	case SWAP_OFF:
660 		mutex_enter(&uvm_swap_data_lock);
661 		if ((sdp = swaplist_find(vp, false)) == NULL) {
662 			mutex_exit(&uvm_swap_data_lock);
663 			error = ENXIO;
664 			break;
665 		}
666 
667 		/*
668 		 * If a device isn't in use or enabled, we
669 		 * can't stop swapping from it (again).
670 		 */
671 		if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0) {
672 			mutex_exit(&uvm_swap_data_lock);
673 			error = EBUSY;
674 			break;
675 		}
676 
677 		/*
678 		 * do the real work.
679 		 */
680 		error = swap_off(l, sdp);
681 		break;
682 
683 	default:
684 		error = EINVAL;
685 	}
686 
687 	/*
688 	 * done!  release the ref gained by namei() and unlock.
689 	 */
690 	vput(vp);
691 
692 out:
693 	free(userpath, M_TEMP);
694 	rw_exit(&swap_syscall_lock);
695 
696 	UVMHIST_LOG(pdhist, "<- done!  error=%d", error, 0, 0, 0);
697 	return (error);
698 }
699 
700 /*
701  * swap_stats: implements swapctl(SWAP_STATS). The function is kept
702  * away from sys_swapctl() in order to allow COMPAT_* swapctl()
703  * emulation to use it directly without going through sys_swapctl().
704  * The problem with using sys_swapctl() there is that it involves
705  * copying the swapent array to the stackgap, and this array's size
706  * is not known at build time. Hence it would not be possible to
707  * ensure it would fit in the stackgap in any case.
708  */
709 void
710 uvm_swap_stats(int cmd, struct swapent *sep, int sec, register_t *retval)
711 {
712 
713 	rw_enter(&swap_syscall_lock, RW_READER);
714 	uvm_swap_stats_locked(cmd, sep, sec, retval);
715 	rw_exit(&swap_syscall_lock);
716 }
717 
718 static void
719 uvm_swap_stats_locked(int cmd, struct swapent *sep, int sec, register_t *retval)
720 {
721 	struct swappri *spp;
722 	struct swapdev *sdp;
723 	int count = 0;
724 
725 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
726 		for (sdp = CIRCLEQ_FIRST(&spp->spi_swapdev);
727 		     sdp != (void *)&spp->spi_swapdev && sec-- > 0;
728 		     sdp = CIRCLEQ_NEXT(sdp, swd_next)) {
729 		  	/*
730 			 * backwards compatibility for system call.
731 			 * note that we use 'struct oswapent' as an
732 			 * overlay into both 'struct swapdev' and
733 			 * the userland 'struct swapent', as we
734 			 * want to retain backwards compatibility
735 			 * with NetBSD 1.3.
736 			 */
737 			sdp->swd_ose.ose_inuse =
738 			    btodb((uint64_t)sdp->swd_npginuse <<
739 			    PAGE_SHIFT);
740 			(void)memcpy(sep, &sdp->swd_ose,
741 			    sizeof(struct oswapent));
742 
743 			/* now copy out the path if necessary */
744 #if !defined(COMPAT_13)
745 			(void) cmd;
746 #endif
747 #if defined(COMPAT_13)
748 			if (cmd == SWAP_STATS)
749 #endif
750 				(void)memcpy(&sep->se_path, sdp->swd_path,
751 				    sdp->swd_pathlen);
752 
753 			count++;
754 #if defined(COMPAT_13)
755 			if (cmd == SWAP_OSTATS)
756 				sep = (struct swapent *)
757 				    ((struct oswapent *)sep + 1);
758 			else
759 #endif
760 				sep++;
761 		}
762 	}
763 
764 	*retval = count;
765 	return;
766 }
767 
768 /*
769  * swap_on: attempt to enable a swapdev for swapping.   note that the
770  *	swapdev is already on the global list, but disabled (marked
771  *	SWF_FAKE).
772  *
773  * => we avoid the start of the disk (to protect disk labels)
774  * => we also avoid the miniroot, if we are swapping to root.
775  * => caller should leave uvm_swap_data_lock unlocked, we may lock it
776  *	if needed.
777  */
778 static int
779 swap_on(struct lwp *l, struct swapdev *sdp)
780 {
781 	struct vnode *vp;
782 	int error, npages, nblocks, size;
783 	long addr;
784 	u_long result;
785 	struct vattr va;
786 #ifdef NFS
787 	extern int (**nfsv2_vnodeop_p)(void *);
788 #endif /* NFS */
789 	const struct bdevsw *bdev;
790 	dev_t dev;
791 	UVMHIST_FUNC("swap_on"); UVMHIST_CALLED(pdhist);
792 
793 	/*
794 	 * we want to enable swapping on sdp.   the swd_vp contains
795 	 * the vnode we want (locked and ref'd), and the swd_dev
796 	 * contains the dev_t of the file, if it a block device.
797 	 */
798 
799 	vp = sdp->swd_vp;
800 	dev = sdp->swd_dev;
801 
802 	/*
803 	 * open the swap file (mostly useful for block device files to
804 	 * let device driver know what is up).
805 	 *
806 	 * we skip the open/close for root on swap because the root
807 	 * has already been opened when root was mounted (mountroot).
808 	 */
809 	if (vp != rootvp) {
810 		if ((error = VOP_OPEN(vp, FREAD|FWRITE, l->l_cred)))
811 			return (error);
812 	}
813 
814 	/* XXX this only works for block devices */
815 	UVMHIST_LOG(pdhist, "  dev=%d, major(dev)=%d", dev, major(dev), 0,0);
816 
817 	/*
818 	 * we now need to determine the size of the swap area.   for
819 	 * block specials we can call the d_psize function.
820 	 * for normal files, we must stat [get attrs].
821 	 *
822 	 * we put the result in nblks.
823 	 * for normal files, we also want the filesystem block size
824 	 * (which we get with statfs).
825 	 */
826 	switch (vp->v_type) {
827 	case VBLK:
828 		bdev = bdevsw_lookup(dev);
829 		if (bdev == NULL || bdev->d_psize == NULL ||
830 		    (nblocks = (*bdev->d_psize)(dev)) == -1) {
831 			error = ENXIO;
832 			goto bad;
833 		}
834 		break;
835 
836 	case VREG:
837 		if ((error = VOP_GETATTR(vp, &va, l->l_cred)))
838 			goto bad;
839 		nblocks = (int)btodb(va.va_size);
840 		if ((error =
841 		     VFS_STATVFS(vp->v_mount, &vp->v_mount->mnt_stat)) != 0)
842 			goto bad;
843 
844 		sdp->swd_bsize = vp->v_mount->mnt_stat.f_iosize;
845 		/*
846 		 * limit the max # of outstanding I/O requests we issue
847 		 * at any one time.   take it easy on NFS servers.
848 		 */
849 #ifdef NFS
850 		if (vp->v_op == nfsv2_vnodeop_p)
851 			sdp->swd_maxactive = 2; /* XXX */
852 		else
853 #endif /* NFS */
854 			sdp->swd_maxactive = 8; /* XXX */
855 		break;
856 
857 	default:
858 		error = ENXIO;
859 		goto bad;
860 	}
861 
862 	/*
863 	 * save nblocks in a safe place and convert to pages.
864 	 */
865 
866 	sdp->swd_ose.ose_nblks = nblocks;
867 	npages = dbtob((uint64_t)nblocks) >> PAGE_SHIFT;
868 
869 	/*
870 	 * for block special files, we want to make sure that leave
871 	 * the disklabel and bootblocks alone, so we arrange to skip
872 	 * over them (arbitrarily choosing to skip PAGE_SIZE bytes).
873 	 * note that because of this the "size" can be less than the
874 	 * actual number of blocks on the device.
875 	 */
876 	if (vp->v_type == VBLK) {
877 		/* we use pages 1 to (size - 1) [inclusive] */
878 		size = npages - 1;
879 		addr = 1;
880 	} else {
881 		/* we use pages 0 to (size - 1) [inclusive] */
882 		size = npages;
883 		addr = 0;
884 	}
885 
886 	/*
887 	 * make sure we have enough blocks for a reasonable sized swap
888 	 * area.   we want at least one page.
889 	 */
890 
891 	if (size < 1) {
892 		UVMHIST_LOG(pdhist, "  size <= 1!!", 0, 0, 0, 0);
893 		error = EINVAL;
894 		goto bad;
895 	}
896 
897 	UVMHIST_LOG(pdhist, "  dev=%x: size=%d addr=%ld\n", dev, size, addr, 0);
898 
899 	/*
900 	 * now we need to allocate an extent to manage this swap device
901 	 */
902 
903 	sdp->swd_blist = blist_create(npages);
904 	/* mark all expect the `saved' region free. */
905 	blist_free(sdp->swd_blist, addr, size);
906 
907 	/*
908 	 * if the vnode we are swapping to is the root vnode
909 	 * (i.e. we are swapping to the miniroot) then we want
910 	 * to make sure we don't overwrite it.   do a statfs to
911 	 * find its size and skip over it.
912 	 */
913 	if (vp == rootvp) {
914 		struct mount *mp;
915 		struct statvfs *sp;
916 		int rootblocks, rootpages;
917 
918 		mp = rootvnode->v_mount;
919 		sp = &mp->mnt_stat;
920 		rootblocks = sp->f_blocks * btodb(sp->f_frsize);
921 		/*
922 		 * XXX: sp->f_blocks isn't the total number of
923 		 * blocks in the filesystem, it's the number of
924 		 * data blocks.  so, our rootblocks almost
925 		 * definitely underestimates the total size
926 		 * of the filesystem - how badly depends on the
927 		 * details of the filesystem type.  there isn't
928 		 * an obvious way to deal with this cleanly
929 		 * and perfectly, so for now we just pad our
930 		 * rootblocks estimate with an extra 5 percent.
931 		 */
932 		rootblocks += (rootblocks >> 5) +
933 			(rootblocks >> 6) +
934 			(rootblocks >> 7);
935 		rootpages = round_page(dbtob(rootblocks)) >> PAGE_SHIFT;
936 		if (rootpages > size)
937 			panic("swap_on: miniroot larger than swap?");
938 
939 		if (rootpages != blist_fill(sdp->swd_blist, addr, rootpages)) {
940 			panic("swap_on: unable to preserve miniroot");
941 		}
942 
943 		size -= rootpages;
944 		printf("Preserved %d pages of miniroot ", rootpages);
945 		printf("leaving %d pages of swap\n", size);
946 	}
947 
948 	/*
949 	 * add a ref to vp to reflect usage as a swap device.
950 	 */
951 	vref(vp);
952 
953 	/*
954 	 * now add the new swapdev to the drum and enable.
955 	 */
956 	result = vmem_alloc(swapmap, npages, VM_BESTFIT | VM_SLEEP);
957 	if (result == 0)
958 		panic("swapdrum_add");
959 	/*
960 	 * If this is the first regular swap create the workqueue.
961 	 * => Protected by swap_syscall_lock.
962 	 */
963 	if (vp->v_type != VBLK) {
964 		if (sw_reg_count++ == 0) {
965 			KASSERT(sw_reg_workqueue == NULL);
966 			if (workqueue_create(&sw_reg_workqueue, "swapiod",
967 			    sw_reg_iodone, NULL, PRIBIO, IPL_BIO, 0) != 0)
968 				panic("swap_add: workqueue_create failed");
969 		}
970 	}
971 
972 	sdp->swd_drumoffset = (int)result;
973 	sdp->swd_drumsize = npages;
974 	sdp->swd_npages = size;
975 	mutex_enter(&uvm_swap_data_lock);
976 	sdp->swd_flags &= ~SWF_FAKE;	/* going live */
977 	sdp->swd_flags |= (SWF_INUSE|SWF_ENABLE);
978 	uvmexp.swpages += size;
979 	uvmexp.swpgavail += size;
980 	mutex_exit(&uvm_swap_data_lock);
981 	return (0);
982 
983 	/*
984 	 * failure: clean up and return error.
985 	 */
986 
987 bad:
988 	if (sdp->swd_blist) {
989 		blist_destroy(sdp->swd_blist);
990 	}
991 	if (vp != rootvp) {
992 		(void)VOP_CLOSE(vp, FREAD|FWRITE, l->l_cred);
993 	}
994 	return (error);
995 }
996 
997 /*
998  * swap_off: stop swapping on swapdev
999  *
1000  * => swap data should be locked, we will unlock.
1001  */
1002 static int
1003 swap_off(struct lwp *l, struct swapdev *sdp)
1004 {
1005 	int npages = sdp->swd_npages;
1006 	int error = 0;
1007 
1008 	UVMHIST_FUNC("swap_off"); UVMHIST_CALLED(pdhist);
1009 	UVMHIST_LOG(pdhist, "  dev=%x, npages=%d", sdp->swd_dev,npages,0,0);
1010 
1011 	/* disable the swap area being removed */
1012 	sdp->swd_flags &= ~SWF_ENABLE;
1013 	uvmexp.swpgavail -= npages;
1014 	mutex_exit(&uvm_swap_data_lock);
1015 
1016 	/*
1017 	 * the idea is to find all the pages that are paged out to this
1018 	 * device, and page them all in.  in uvm, swap-backed pageable
1019 	 * memory can take two forms: aobjs and anons.  call the
1020 	 * swapoff hook for each subsystem to bring in pages.
1021 	 */
1022 
1023 	if (uao_swap_off(sdp->swd_drumoffset,
1024 			 sdp->swd_drumoffset + sdp->swd_drumsize) ||
1025 	    amap_swap_off(sdp->swd_drumoffset,
1026 			  sdp->swd_drumoffset + sdp->swd_drumsize)) {
1027 		error = ENOMEM;
1028 	} else if (sdp->swd_npginuse > sdp->swd_npgbad) {
1029 		error = EBUSY;
1030 	}
1031 
1032 	if (error) {
1033 		mutex_enter(&uvm_swap_data_lock);
1034 		sdp->swd_flags |= SWF_ENABLE;
1035 		uvmexp.swpgavail += npages;
1036 		mutex_exit(&uvm_swap_data_lock);
1037 
1038 		return error;
1039 	}
1040 
1041 	/*
1042 	 * If this is the last regular swap destroy the workqueue.
1043 	 * => Protected by swap_syscall_lock.
1044 	 */
1045 	if (sdp->swd_vp->v_type != VBLK) {
1046 		KASSERT(sw_reg_count > 0);
1047 		KASSERT(sw_reg_workqueue != NULL);
1048 		if (--sw_reg_count == 0) {
1049 			workqueue_destroy(sw_reg_workqueue);
1050 			sw_reg_workqueue = NULL;
1051 		}
1052 	}
1053 
1054 	/*
1055 	 * done with the vnode.
1056 	 * drop our ref on the vnode before calling VOP_CLOSE()
1057 	 * so that spec_close() can tell if this is the last close.
1058 	 */
1059 	vrele(sdp->swd_vp);
1060 	if (sdp->swd_vp != rootvp) {
1061 		(void) VOP_CLOSE(sdp->swd_vp, FREAD|FWRITE, l->l_cred);
1062 	}
1063 
1064 	mutex_enter(&uvm_swap_data_lock);
1065 	uvmexp.swpages -= npages;
1066 	uvmexp.swpginuse -= sdp->swd_npgbad;
1067 
1068 	if (swaplist_find(sdp->swd_vp, true) == NULL)
1069 		panic("swap_off: swapdev not in list");
1070 	swaplist_trim();
1071 	mutex_exit(&uvm_swap_data_lock);
1072 
1073 	/*
1074 	 * free all resources!
1075 	 */
1076 	vmem_free(swapmap, sdp->swd_drumoffset, sdp->swd_drumsize);
1077 	blist_destroy(sdp->swd_blist);
1078 	bufq_free(sdp->swd_tab);
1079 	free(sdp, M_VMSWAP);
1080 	return (0);
1081 }
1082 
1083 /*
1084  * /dev/drum interface and i/o functions
1085  */
1086 
1087 /*
1088  * swstrategy: perform I/O on the drum
1089  *
1090  * => we must map the i/o request from the drum to the correct swapdev.
1091  */
1092 static void
1093 swstrategy(struct buf *bp)
1094 {
1095 	struct swapdev *sdp;
1096 	struct vnode *vp;
1097 	int pageno, bn;
1098 	UVMHIST_FUNC("swstrategy"); UVMHIST_CALLED(pdhist);
1099 
1100 	/*
1101 	 * convert block number to swapdev.   note that swapdev can't
1102 	 * be yanked out from under us because we are holding resources
1103 	 * in it (i.e. the blocks we are doing I/O on).
1104 	 */
1105 	pageno = dbtob((int64_t)bp->b_blkno) >> PAGE_SHIFT;
1106 	mutex_enter(&uvm_swap_data_lock);
1107 	sdp = swapdrum_getsdp(pageno);
1108 	mutex_exit(&uvm_swap_data_lock);
1109 	if (sdp == NULL) {
1110 		bp->b_error = EINVAL;
1111 		biodone(bp);
1112 		UVMHIST_LOG(pdhist, "  failed to get swap device", 0, 0, 0, 0);
1113 		return;
1114 	}
1115 
1116 	/*
1117 	 * convert drum page number to block number on this swapdev.
1118 	 */
1119 
1120 	pageno -= sdp->swd_drumoffset;	/* page # on swapdev */
1121 	bn = btodb((uint64_t)pageno << PAGE_SHIFT); /* convert to diskblock */
1122 
1123 	UVMHIST_LOG(pdhist, "  %s: mapoff=%x bn=%x bcount=%ld",
1124 		((bp->b_flags & B_READ) == 0) ? "write" : "read",
1125 		sdp->swd_drumoffset, bn, bp->b_bcount);
1126 
1127 	/*
1128 	 * for block devices we finish up here.
1129 	 * for regular files we have to do more work which we delegate
1130 	 * to sw_reg_strategy().
1131 	 */
1132 
1133 	vp = sdp->swd_vp;		/* swapdev vnode pointer */
1134 	switch (vp->v_type) {
1135 	default:
1136 		panic("swstrategy: vnode type 0x%x", vp->v_type);
1137 
1138 	case VBLK:
1139 
1140 		/*
1141 		 * must convert "bp" from an I/O on /dev/drum to an I/O
1142 		 * on the swapdev (sdp).
1143 		 */
1144 		bp->b_blkno = bn;		/* swapdev block number */
1145 		bp->b_dev = sdp->swd_dev;	/* swapdev dev_t */
1146 
1147 		/*
1148 		 * if we are doing a write, we have to redirect the i/o on
1149 		 * drum's v_numoutput counter to the swapdevs.
1150 		 */
1151 		if ((bp->b_flags & B_READ) == 0) {
1152 			mutex_enter(bp->b_objlock);
1153 			vwakeup(bp);	/* kills one 'v_numoutput' on drum */
1154 			mutex_exit(bp->b_objlock);
1155 			mutex_enter(&vp->v_interlock);
1156 			vp->v_numoutput++;	/* put it on swapdev */
1157 			mutex_exit(&vp->v_interlock);
1158 		}
1159 
1160 		/*
1161 		 * finally plug in swapdev vnode and start I/O
1162 		 */
1163 		bp->b_vp = vp;
1164 		bp->b_objlock = &vp->v_interlock;
1165 		VOP_STRATEGY(vp, bp);
1166 		return;
1167 
1168 	case VREG:
1169 		/*
1170 		 * delegate to sw_reg_strategy function.
1171 		 */
1172 		sw_reg_strategy(sdp, bp, bn);
1173 		return;
1174 	}
1175 	/* NOTREACHED */
1176 }
1177 
1178 /*
1179  * swread: the read function for the drum (just a call to physio)
1180  */
1181 /*ARGSUSED*/
1182 static int
1183 swread(dev_t dev, struct uio *uio, int ioflag)
1184 {
1185 	UVMHIST_FUNC("swread"); UVMHIST_CALLED(pdhist);
1186 
1187 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1188 	return (physio(swstrategy, NULL, dev, B_READ, minphys, uio));
1189 }
1190 
1191 /*
1192  * swwrite: the write function for the drum (just a call to physio)
1193  */
1194 /*ARGSUSED*/
1195 static int
1196 swwrite(dev_t dev, struct uio *uio, int ioflag)
1197 {
1198 	UVMHIST_FUNC("swwrite"); UVMHIST_CALLED(pdhist);
1199 
1200 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1201 	return (physio(swstrategy, NULL, dev, B_WRITE, minphys, uio));
1202 }
1203 
1204 const struct bdevsw swap_bdevsw = {
1205 	nullopen, nullclose, swstrategy, noioctl, nodump, nosize, D_OTHER,
1206 };
1207 
1208 const struct cdevsw swap_cdevsw = {
1209 	nullopen, nullclose, swread, swwrite, noioctl,
1210 	nostop, notty, nopoll, nommap, nokqfilter, D_OTHER,
1211 };
1212 
1213 /*
1214  * sw_reg_strategy: handle swap i/o to regular files
1215  */
1216 static void
1217 sw_reg_strategy(struct swapdev *sdp, struct buf *bp, int bn)
1218 {
1219 	struct vnode	*vp;
1220 	struct vndxfer	*vnx;
1221 	daddr_t		nbn;
1222 	char 		*addr;
1223 	off_t		byteoff;
1224 	int		s, off, nra, error, sz, resid;
1225 	UVMHIST_FUNC("sw_reg_strategy"); UVMHIST_CALLED(pdhist);
1226 
1227 	/*
1228 	 * allocate a vndxfer head for this transfer and point it to
1229 	 * our buffer.
1230 	 */
1231 	vnx = pool_get(&vndxfer_pool, PR_WAITOK);
1232 	vnx->vx_flags = VX_BUSY;
1233 	vnx->vx_error = 0;
1234 	vnx->vx_pending = 0;
1235 	vnx->vx_bp = bp;
1236 	vnx->vx_sdp = sdp;
1237 
1238 	/*
1239 	 * setup for main loop where we read filesystem blocks into
1240 	 * our buffer.
1241 	 */
1242 	error = 0;
1243 	bp->b_resid = bp->b_bcount;	/* nothing transfered yet! */
1244 	addr = bp->b_data;		/* current position in buffer */
1245 	byteoff = dbtob((uint64_t)bn);
1246 
1247 	for (resid = bp->b_resid; resid; resid -= sz) {
1248 		struct vndbuf	*nbp;
1249 
1250 		/*
1251 		 * translate byteoffset into block number.  return values:
1252 		 *   vp = vnode of underlying device
1253 		 *  nbn = new block number (on underlying vnode dev)
1254 		 *  nra = num blocks we can read-ahead (excludes requested
1255 		 *	block)
1256 		 */
1257 		nra = 0;
1258 		error = VOP_BMAP(sdp->swd_vp, byteoff / sdp->swd_bsize,
1259 				 	&vp, &nbn, &nra);
1260 
1261 		if (error == 0 && nbn == (daddr_t)-1) {
1262 			/*
1263 			 * this used to just set error, but that doesn't
1264 			 * do the right thing.  Instead, it causes random
1265 			 * memory errors.  The panic() should remain until
1266 			 * this condition doesn't destabilize the system.
1267 			 */
1268 #if 1
1269 			panic("sw_reg_strategy: swap to sparse file");
1270 #else
1271 			error = EIO;	/* failure */
1272 #endif
1273 		}
1274 
1275 		/*
1276 		 * punt if there was an error or a hole in the file.
1277 		 * we must wait for any i/o ops we have already started
1278 		 * to finish before returning.
1279 		 *
1280 		 * XXX we could deal with holes here but it would be
1281 		 * a hassle (in the write case).
1282 		 */
1283 		if (error) {
1284 			s = splbio();
1285 			vnx->vx_error = error;	/* pass error up */
1286 			goto out;
1287 		}
1288 
1289 		/*
1290 		 * compute the size ("sz") of this transfer (in bytes).
1291 		 */
1292 		off = byteoff % sdp->swd_bsize;
1293 		sz = (1 + nra) * sdp->swd_bsize - off;
1294 		if (sz > resid)
1295 			sz = resid;
1296 
1297 		UVMHIST_LOG(pdhist, "sw_reg_strategy: "
1298 			    "vp %p/%p offset 0x%x/0x%x",
1299 			    sdp->swd_vp, vp, byteoff, nbn);
1300 
1301 		/*
1302 		 * now get a buf structure.   note that the vb_buf is
1303 		 * at the front of the nbp structure so that you can
1304 		 * cast pointers between the two structure easily.
1305 		 */
1306 		nbp = pool_get(&vndbuf_pool, PR_WAITOK);
1307 		buf_init(&nbp->vb_buf);
1308 		nbp->vb_buf.b_flags    = bp->b_flags;
1309 		nbp->vb_buf.b_cflags   = bp->b_cflags;
1310 		nbp->vb_buf.b_oflags   = bp->b_oflags;
1311 		nbp->vb_buf.b_bcount   = sz;
1312 		nbp->vb_buf.b_bufsize  = sz;
1313 		nbp->vb_buf.b_error    = 0;
1314 		nbp->vb_buf.b_data     = addr;
1315 		nbp->vb_buf.b_lblkno   = 0;
1316 		nbp->vb_buf.b_blkno    = nbn + btodb(off);
1317 		nbp->vb_buf.b_rawblkno = nbp->vb_buf.b_blkno;
1318 		nbp->vb_buf.b_iodone   = sw_reg_biodone;
1319 		nbp->vb_buf.b_vp       = vp;
1320 		nbp->vb_buf.b_objlock  = &vp->v_interlock;
1321 		if (vp->v_type == VBLK) {
1322 			nbp->vb_buf.b_dev = vp->v_rdev;
1323 		}
1324 
1325 		nbp->vb_xfer = vnx;	/* patch it back in to vnx */
1326 
1327 		/*
1328 		 * Just sort by block number
1329 		 */
1330 		s = splbio();
1331 		if (vnx->vx_error != 0) {
1332 			buf_destroy(&nbp->vb_buf);
1333 			pool_put(&vndbuf_pool, nbp);
1334 			goto out;
1335 		}
1336 		vnx->vx_pending++;
1337 
1338 		/* sort it in and start I/O if we are not over our limit */
1339 		/* XXXAD locking */
1340 		BUFQ_PUT(sdp->swd_tab, &nbp->vb_buf);
1341 		sw_reg_start(sdp);
1342 		splx(s);
1343 
1344 		/*
1345 		 * advance to the next I/O
1346 		 */
1347 		byteoff += sz;
1348 		addr += sz;
1349 	}
1350 
1351 	s = splbio();
1352 
1353 out: /* Arrive here at splbio */
1354 	vnx->vx_flags &= ~VX_BUSY;
1355 	if (vnx->vx_pending == 0) {
1356 		error = vnx->vx_error;
1357 		pool_put(&vndxfer_pool, vnx);
1358 		bp->b_error = error;
1359 		biodone(bp);
1360 	}
1361 	splx(s);
1362 }
1363 
1364 /*
1365  * sw_reg_start: start an I/O request on the requested swapdev
1366  *
1367  * => reqs are sorted by b_rawblkno (above)
1368  */
1369 static void
1370 sw_reg_start(struct swapdev *sdp)
1371 {
1372 	struct buf	*bp;
1373 	struct vnode	*vp;
1374 	UVMHIST_FUNC("sw_reg_start"); UVMHIST_CALLED(pdhist);
1375 
1376 	/* recursion control */
1377 	if ((sdp->swd_flags & SWF_BUSY) != 0)
1378 		return;
1379 
1380 	sdp->swd_flags |= SWF_BUSY;
1381 
1382 	while (sdp->swd_active < sdp->swd_maxactive) {
1383 		bp = BUFQ_GET(sdp->swd_tab);
1384 		if (bp == NULL)
1385 			break;
1386 		sdp->swd_active++;
1387 
1388 		UVMHIST_LOG(pdhist,
1389 		    "sw_reg_start:  bp %p vp %p blkno %p cnt %lx",
1390 		    bp, bp->b_vp, bp->b_blkno, bp->b_bcount);
1391 		vp = bp->b_vp;
1392 		KASSERT(bp->b_objlock == &vp->v_interlock);
1393 		if ((bp->b_flags & B_READ) == 0) {
1394 			mutex_enter(&vp->v_interlock);
1395 			vp->v_numoutput++;
1396 			mutex_exit(&vp->v_interlock);
1397 		}
1398 		VOP_STRATEGY(vp, bp);
1399 	}
1400 	sdp->swd_flags &= ~SWF_BUSY;
1401 }
1402 
1403 /*
1404  * sw_reg_biodone: one of our i/o's has completed
1405  */
1406 static void
1407 sw_reg_biodone(struct buf *bp)
1408 {
1409 	workqueue_enqueue(sw_reg_workqueue, &bp->b_work, NULL);
1410 }
1411 
1412 /*
1413  * sw_reg_iodone: one of our i/o's has completed and needs post-i/o cleanup
1414  *
1415  * => note that we can recover the vndbuf struct by casting the buf ptr
1416  */
1417 static void
1418 sw_reg_iodone(struct work *wk, void *dummy)
1419 {
1420 	struct vndbuf *vbp = (void *)wk;
1421 	struct vndxfer *vnx = vbp->vb_xfer;
1422 	struct buf *pbp = vnx->vx_bp;		/* parent buffer */
1423 	struct swapdev	*sdp = vnx->vx_sdp;
1424 	int s, resid, error;
1425 	KASSERT(&vbp->vb_buf.b_work == wk);
1426 	UVMHIST_FUNC("sw_reg_iodone"); UVMHIST_CALLED(pdhist);
1427 
1428 	UVMHIST_LOG(pdhist, "  vbp=%p vp=%p blkno=%x addr=%p",
1429 	    vbp, vbp->vb_buf.b_vp, vbp->vb_buf.b_blkno, vbp->vb_buf.b_data);
1430 	UVMHIST_LOG(pdhist, "  cnt=%lx resid=%lx",
1431 	    vbp->vb_buf.b_bcount, vbp->vb_buf.b_resid, 0, 0);
1432 
1433 	/*
1434 	 * protect vbp at splbio and update.
1435 	 */
1436 
1437 	s = splbio();
1438 	resid = vbp->vb_buf.b_bcount - vbp->vb_buf.b_resid;
1439 	pbp->b_resid -= resid;
1440 	vnx->vx_pending--;
1441 
1442 	if (vbp->vb_buf.b_error != 0) {
1443 		/* pass error upward */
1444 		error = vbp->vb_buf.b_error ? vbp->vb_buf.b_error : EIO;
1445 		UVMHIST_LOG(pdhist, "  got error=%d !", error, 0, 0, 0);
1446 		vnx->vx_error = error;
1447 	}
1448 
1449 	/*
1450 	 * kill vbp structure
1451 	 */
1452 	buf_destroy(&vbp->vb_buf);
1453 	pool_put(&vndbuf_pool, vbp);
1454 
1455 	/*
1456 	 * wrap up this transaction if it has run to completion or, in
1457 	 * case of an error, when all auxiliary buffers have returned.
1458 	 */
1459 	if (vnx->vx_error != 0) {
1460 		/* pass error upward */
1461 		error = vnx->vx_error;
1462 		if ((vnx->vx_flags & VX_BUSY) == 0 && vnx->vx_pending == 0) {
1463 			pbp->b_error = error;
1464 			biodone(pbp);
1465 			pool_put(&vndxfer_pool, vnx);
1466 		}
1467 	} else if (pbp->b_resid == 0) {
1468 		KASSERT(vnx->vx_pending == 0);
1469 		if ((vnx->vx_flags & VX_BUSY) == 0) {
1470 			UVMHIST_LOG(pdhist, "  iodone error=%d !",
1471 			    pbp, vnx->vx_error, 0, 0);
1472 			biodone(pbp);
1473 			pool_put(&vndxfer_pool, vnx);
1474 		}
1475 	}
1476 
1477 	/*
1478 	 * done!   start next swapdev I/O if one is pending
1479 	 */
1480 	sdp->swd_active--;
1481 	sw_reg_start(sdp);
1482 	splx(s);
1483 }
1484 
1485 
1486 /*
1487  * uvm_swap_alloc: allocate space on swap
1488  *
1489  * => allocation is done "round robin" down the priority list, as we
1490  *	allocate in a priority we "rotate" the circle queue.
1491  * => space can be freed with uvm_swap_free
1492  * => we return the page slot number in /dev/drum (0 == invalid slot)
1493  * => we lock uvm_swap_data_lock
1494  * => XXXMRG: "LESSOK" INTERFACE NEEDED TO EXTENT SYSTEM
1495  */
1496 int
1497 uvm_swap_alloc(int *nslots /* IN/OUT */, bool lessok)
1498 {
1499 	struct swapdev *sdp;
1500 	struct swappri *spp;
1501 	UVMHIST_FUNC("uvm_swap_alloc"); UVMHIST_CALLED(pdhist);
1502 
1503 	/*
1504 	 * no swap devices configured yet?   definite failure.
1505 	 */
1506 	if (uvmexp.nswapdev < 1)
1507 		return 0;
1508 
1509 	/*
1510 	 * lock data lock, convert slots into blocks, and enter loop
1511 	 */
1512 	mutex_enter(&uvm_swap_data_lock);
1513 
1514 ReTry:	/* XXXMRG */
1515 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
1516 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
1517 			uint64_t result;
1518 
1519 			/* if it's not enabled, then we can't swap from it */
1520 			if ((sdp->swd_flags & SWF_ENABLE) == 0)
1521 				continue;
1522 			if (sdp->swd_npginuse + *nslots > sdp->swd_npages)
1523 				continue;
1524 			result = blist_alloc(sdp->swd_blist, *nslots);
1525 			if (result == BLIST_NONE) {
1526 				continue;
1527 			}
1528 			KASSERT(result < sdp->swd_drumsize);
1529 
1530 			/*
1531 			 * successful allocation!  now rotate the circleq.
1532 			 */
1533 			CIRCLEQ_REMOVE(&spp->spi_swapdev, sdp, swd_next);
1534 			CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
1535 			sdp->swd_npginuse += *nslots;
1536 			uvmexp.swpginuse += *nslots;
1537 			mutex_exit(&uvm_swap_data_lock);
1538 			/* done!  return drum slot number */
1539 			UVMHIST_LOG(pdhist,
1540 			    "success!  returning %d slots starting at %d",
1541 			    *nslots, result + sdp->swd_drumoffset, 0, 0);
1542 			return (result + sdp->swd_drumoffset);
1543 		}
1544 	}
1545 
1546 	/* XXXMRG: BEGIN HACK */
1547 	if (*nslots > 1 && lessok) {
1548 		*nslots = 1;
1549 		/* XXXMRG: ugh!  blist should support this for us */
1550 		goto ReTry;
1551 	}
1552 	/* XXXMRG: END HACK */
1553 
1554 	mutex_exit(&uvm_swap_data_lock);
1555 	return 0;
1556 }
1557 
1558 /*
1559  * uvm_swapisfull: return true if most of available swap is allocated
1560  * and in use.  we don't count some small portion as it may be inaccessible
1561  * to us at any given moment, for example if there is lock contention or if
1562  * pages are busy.
1563  */
1564 bool
1565 uvm_swapisfull(void)
1566 {
1567 	int swpgonly;
1568 	bool rv;
1569 
1570 	mutex_enter(&uvm_swap_data_lock);
1571 	KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1572 	swpgonly = (int)((uint64_t)uvmexp.swpgonly * 100 /
1573 	    uvm_swapisfull_factor);
1574 	rv = (swpgonly >= uvmexp.swpgavail);
1575 	mutex_exit(&uvm_swap_data_lock);
1576 
1577 	return (rv);
1578 }
1579 
1580 /*
1581  * uvm_swap_markbad: keep track of swap ranges where we've had i/o errors
1582  *
1583  * => we lock uvm_swap_data_lock
1584  */
1585 void
1586 uvm_swap_markbad(int startslot, int nslots)
1587 {
1588 	struct swapdev *sdp;
1589 	UVMHIST_FUNC("uvm_swap_markbad"); UVMHIST_CALLED(pdhist);
1590 
1591 	mutex_enter(&uvm_swap_data_lock);
1592 	sdp = swapdrum_getsdp(startslot);
1593 	KASSERT(sdp != NULL);
1594 
1595 	/*
1596 	 * we just keep track of how many pages have been marked bad
1597 	 * in this device, to make everything add up in swap_off().
1598 	 * we assume here that the range of slots will all be within
1599 	 * one swap device.
1600 	 */
1601 
1602 	KASSERT(uvmexp.swpgonly >= nslots);
1603 	uvmexp.swpgonly -= nslots;
1604 	sdp->swd_npgbad += nslots;
1605 	UVMHIST_LOG(pdhist, "now %d bad", sdp->swd_npgbad, 0,0,0);
1606 	mutex_exit(&uvm_swap_data_lock);
1607 }
1608 
1609 /*
1610  * uvm_swap_free: free swap slots
1611  *
1612  * => this can be all or part of an allocation made by uvm_swap_alloc
1613  * => we lock uvm_swap_data_lock
1614  */
1615 void
1616 uvm_swap_free(int startslot, int nslots)
1617 {
1618 	struct swapdev *sdp;
1619 	UVMHIST_FUNC("uvm_swap_free"); UVMHIST_CALLED(pdhist);
1620 
1621 	UVMHIST_LOG(pdhist, "freeing %d slots starting at %d", nslots,
1622 	    startslot, 0, 0);
1623 
1624 	/*
1625 	 * ignore attempts to free the "bad" slot.
1626 	 */
1627 
1628 	if (startslot == SWSLOT_BAD) {
1629 		return;
1630 	}
1631 
1632 	/*
1633 	 * convert drum slot offset back to sdp, free the blocks
1634 	 * in the extent, and return.   must hold pri lock to do
1635 	 * lookup and access the extent.
1636 	 */
1637 
1638 	mutex_enter(&uvm_swap_data_lock);
1639 	sdp = swapdrum_getsdp(startslot);
1640 	KASSERT(uvmexp.nswapdev >= 1);
1641 	KASSERT(sdp != NULL);
1642 	KASSERT(sdp->swd_npginuse >= nslots);
1643 	blist_free(sdp->swd_blist, startslot - sdp->swd_drumoffset, nslots);
1644 	sdp->swd_npginuse -= nslots;
1645 	uvmexp.swpginuse -= nslots;
1646 	mutex_exit(&uvm_swap_data_lock);
1647 }
1648 
1649 /*
1650  * uvm_swap_put: put any number of pages into a contig place on swap
1651  *
1652  * => can be sync or async
1653  */
1654 
1655 int
1656 uvm_swap_put(int swslot, struct vm_page **ppsp, int npages, int flags)
1657 {
1658 	int error;
1659 
1660 	error = uvm_swap_io(ppsp, swslot, npages, B_WRITE |
1661 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1662 	return error;
1663 }
1664 
1665 /*
1666  * uvm_swap_get: get a single page from swap
1667  *
1668  * => usually a sync op (from fault)
1669  */
1670 
1671 int
1672 uvm_swap_get(struct vm_page *page, int swslot, int flags)
1673 {
1674 	int error;
1675 
1676 	uvmexp.nswget++;
1677 	KASSERT(flags & PGO_SYNCIO);
1678 	if (swslot == SWSLOT_BAD) {
1679 		return EIO;
1680 	}
1681 
1682 	error = uvm_swap_io(&page, swslot, 1, B_READ |
1683 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1684 	if (error == 0) {
1685 
1686 		/*
1687 		 * this page is no longer only in swap.
1688 		 */
1689 
1690 		mutex_enter(&uvm_swap_data_lock);
1691 		KASSERT(uvmexp.swpgonly > 0);
1692 		uvmexp.swpgonly--;
1693 		mutex_exit(&uvm_swap_data_lock);
1694 	}
1695 	return error;
1696 }
1697 
1698 /*
1699  * uvm_swap_io: do an i/o operation to swap
1700  */
1701 
1702 static int
1703 uvm_swap_io(struct vm_page **pps, int startslot, int npages, int flags)
1704 {
1705 	daddr_t startblk;
1706 	struct	buf *bp;
1707 	vaddr_t kva;
1708 	int	error, mapinflags;
1709 	bool write, async;
1710 	UVMHIST_FUNC("uvm_swap_io"); UVMHIST_CALLED(pdhist);
1711 
1712 	UVMHIST_LOG(pdhist, "<- called, startslot=%d, npages=%d, flags=%d",
1713 	    startslot, npages, flags, 0);
1714 
1715 	write = (flags & B_READ) == 0;
1716 	async = (flags & B_ASYNC) != 0;
1717 
1718 	/*
1719 	 * allocate a buf for the i/o.
1720 	 */
1721 
1722 	KASSERT(curlwp != uvm.pagedaemon_lwp || (write && async));
1723 	bp = getiobuf(swapdev_vp, curlwp != uvm.pagedaemon_lwp);
1724 	if (bp == NULL) {
1725 		uvm_aio_aiodone_pages(pps, npages, true, ENOMEM);
1726 		return ENOMEM;
1727 	}
1728 
1729 	/*
1730 	 * convert starting drum slot to block number
1731 	 */
1732 
1733 	startblk = btodb((uint64_t)startslot << PAGE_SHIFT);
1734 
1735 	/*
1736 	 * first, map the pages into the kernel.
1737 	 */
1738 
1739 	mapinflags = !write ?
1740 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_READ :
1741 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_WRITE;
1742 	kva = uvm_pagermapin(pps, npages, mapinflags);
1743 
1744 	/*
1745 	 * fill in the bp/sbp.   we currently route our i/o through
1746 	 * /dev/drum's vnode [swapdev_vp].
1747 	 */
1748 
1749 	bp->b_cflags = BC_BUSY | BC_NOCACHE;
1750 	bp->b_flags = (flags & (B_READ|B_ASYNC));
1751 	bp->b_proc = &proc0;	/* XXX */
1752 	bp->b_vnbufs.le_next = NOLIST;
1753 	bp->b_data = (void *)kva;
1754 	bp->b_blkno = startblk;
1755 	bp->b_bufsize = bp->b_bcount = npages << PAGE_SHIFT;
1756 
1757 	/*
1758 	 * bump v_numoutput (counter of number of active outputs).
1759 	 */
1760 
1761 	if (write) {
1762 		mutex_enter(&swapdev_vp->v_interlock);
1763 		swapdev_vp->v_numoutput++;
1764 		mutex_exit(&swapdev_vp->v_interlock);
1765 	}
1766 
1767 	/*
1768 	 * for async ops we must set up the iodone handler.
1769 	 */
1770 
1771 	if (async) {
1772 		bp->b_iodone = uvm_aio_biodone;
1773 		UVMHIST_LOG(pdhist, "doing async!", 0, 0, 0, 0);
1774 		if (curlwp == uvm.pagedaemon_lwp)
1775 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1776 		else
1777 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
1778 	} else {
1779 		bp->b_iodone = NULL;
1780 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1781 	}
1782 	UVMHIST_LOG(pdhist,
1783 	    "about to start io: data = %p blkno = 0x%x, bcount = %ld",
1784 	    bp->b_data, bp->b_blkno, bp->b_bcount, 0);
1785 
1786 	/*
1787 	 * now we start the I/O, and if async, return.
1788 	 */
1789 
1790 	VOP_STRATEGY(swapdev_vp, bp);
1791 	if (async)
1792 		return 0;
1793 
1794 	/*
1795 	 * must be sync i/o.   wait for it to finish
1796 	 */
1797 
1798 	error = biowait(bp);
1799 
1800 	/*
1801 	 * kill the pager mapping
1802 	 */
1803 
1804 	uvm_pagermapout(kva, npages);
1805 
1806 	/*
1807 	 * now dispose of the buf and we're done.
1808 	 */
1809 
1810 	if (write) {
1811 		mutex_enter(&swapdev_vp->v_interlock);
1812 		vwakeup(bp);
1813 		mutex_exit(&swapdev_vp->v_interlock);
1814 	}
1815 	putiobuf(bp);
1816 	UVMHIST_LOG(pdhist, "<- done (sync)  error=%d", error, 0, 0, 0);
1817 
1818 	return (error);
1819 }
1820