xref: /netbsd-src/sys/uvm/uvm_swap.c (revision 220b5c059a84c51ea44107ea8951a57ffaecdc8c)
1 /*	$NetBSD: uvm_swap.c,v 1.57 2001/11/10 07:37:01 lukem 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  * 3. The name of the author may not be used to endorse or promote products
16  *    derived from this software without specific prior written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
23  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
24  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
25  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
26  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  * from: NetBSD: vm_swap.c,v 1.52 1997/12/02 13:47:37 pk Exp
31  * from: Id: uvm_swap.c,v 1.1.2.42 1998/02/02 20:38:06 chuck Exp
32  */
33 
34 #include <sys/cdefs.h>
35 __KERNEL_RCSID(0, "$NetBSD: uvm_swap.c,v 1.57 2001/11/10 07:37:01 lukem Exp $");
36 
37 #include "fs_nfs.h"
38 #include "opt_uvmhist.h"
39 #include "opt_compat_netbsd.h"
40 #include "opt_ddb.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/buf.h>
45 #include <sys/conf.h>
46 #include <sys/proc.h>
47 #include <sys/namei.h>
48 #include <sys/disklabel.h>
49 #include <sys/errno.h>
50 #include <sys/kernel.h>
51 #include <sys/malloc.h>
52 #include <sys/vnode.h>
53 #include <sys/file.h>
54 #include <sys/extent.h>
55 #include <sys/mount.h>
56 #include <sys/pool.h>
57 #include <sys/syscallargs.h>
58 #include <sys/swap.h>
59 
60 #include <uvm/uvm.h>
61 
62 #include <miscfs/specfs/specdev.h>
63 
64 /*
65  * uvm_swap.c: manage configuration and i/o to swap space.
66  */
67 
68 /*
69  * swap space is managed in the following way:
70  *
71  * each swap partition or file is described by a "swapdev" structure.
72  * each "swapdev" structure contains a "swapent" structure which contains
73  * information that is passed up to the user (via system calls).
74  *
75  * each swap partition is assigned a "priority" (int) which controls
76  * swap parition usage.
77  *
78  * the system maintains a global data structure describing all swap
79  * partitions/files.   there is a sorted LIST of "swappri" structures
80  * which describe "swapdev"'s at that priority.   this LIST is headed
81  * by the "swap_priority" global var.    each "swappri" contains a
82  * CIRCLEQ of "swapdev" structures at that priority.
83  *
84  * locking:
85  *  - swap_syscall_lock (sleep lock): this lock serializes the swapctl
86  *    system call and prevents the swap priority list from changing
87  *    while we are in the middle of a system call (e.g. SWAP_STATS).
88  *  - uvm.swap_data_lock (simple_lock): this lock protects all swap data
89  *    structures including the priority list, the swapdev structures,
90  *    and the swapmap extent.
91  *
92  * each swap device has the following info:
93  *  - swap device in use (could be disabled, preventing future use)
94  *  - swap enabled (allows new allocations on swap)
95  *  - map info in /dev/drum
96  *  - vnode pointer
97  * for swap files only:
98  *  - block size
99  *  - max byte count in buffer
100  *  - buffer
101  *  - credentials to use when doing i/o to file
102  *
103  * userland controls and configures swap with the swapctl(2) system call.
104  * the sys_swapctl performs the following operations:
105  *  [1] SWAP_NSWAP: returns the number of swap devices currently configured
106  *  [2] SWAP_STATS: given a pointer to an array of swapent structures
107  *	(passed in via "arg") of a size passed in via "misc" ... we load
108  *	the current swap config into the array.
109  *  [3] SWAP_ON: given a pathname in arg (could be device or file) and a
110  *	priority in "misc", start swapping on it.
111  *  [4] SWAP_OFF: as SWAP_ON, but stops swapping to a device
112  *  [5] SWAP_CTL: changes the priority of a swap device (new priority in
113  *	"misc")
114  */
115 
116 /*
117  * swapdev: describes a single swap partition/file
118  *
119  * note the following should be true:
120  * swd_inuse <= swd_nblks  [number of blocks in use is <= total blocks]
121  * swd_nblks <= swd_mapsize [because mapsize includes miniroot+disklabel]
122  */
123 struct swapdev {
124 	struct oswapent swd_ose;
125 #define	swd_dev		swd_ose.ose_dev		/* device id */
126 #define	swd_flags	swd_ose.ose_flags	/* flags:inuse/enable/fake */
127 #define	swd_priority	swd_ose.ose_priority	/* our priority */
128 	/* also: swd_ose.ose_nblks, swd_ose.ose_inuse */
129 	char			*swd_path;	/* saved pathname of device */
130 	int			swd_pathlen;	/* length of pathname */
131 	int			swd_npages;	/* #pages we can use */
132 	int			swd_npginuse;	/* #pages in use */
133 	int			swd_npgbad;	/* #pages bad */
134 	int			swd_drumoffset;	/* page0 offset in drum */
135 	int			swd_drumsize;	/* #pages in drum */
136 	struct extent		*swd_ex;	/* extent for this swapdev */
137 	char			swd_exname[12];	/* name of extent above */
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 buf_queue	swd_tab;	/* buffer list */
144 	int			swd_active;	/* number of active buffers */
145 	struct ucred		*swd_cred;	/* cred for file access */
146 };
147 
148 /*
149  * swap device priority entry; the list is kept sorted on `spi_priority'.
150  */
151 struct swappri {
152 	int			spi_priority;     /* priority */
153 	CIRCLEQ_HEAD(spi_swapdev, swapdev)	spi_swapdev;
154 	/* circleq of swapdevs at this priority */
155 	LIST_ENTRY(swappri)	spi_swappri;      /* global list of pri's */
156 };
157 
158 /*
159  * The following two structures are used to keep track of data transfers
160  * on swap devices associated with regular files.
161  * NOTE: this code is more or less a copy of vnd.c; we use the same
162  * structure names here to ease porting..
163  */
164 struct vndxfer {
165 	struct buf	*vx_bp;		/* Pointer to parent buffer */
166 	struct swapdev	*vx_sdp;
167 	int		vx_error;
168 	int		vx_pending;	/* # of pending aux buffers */
169 	int		vx_flags;
170 #define VX_BUSY		1
171 #define VX_DEAD		2
172 };
173 
174 struct vndbuf {
175 	struct buf	vb_buf;
176 	struct vndxfer	*vb_xfer;
177 };
178 
179 
180 /*
181  * We keep a of pool vndbuf's and vndxfer structures.
182  */
183 static struct pool vndxfer_pool;
184 static struct pool vndbuf_pool;
185 
186 #define	getvndxfer(vnx)	do {						\
187 	int s = splbio();						\
188 	vnx = pool_get(&vndxfer_pool, PR_MALLOCOK|PR_WAITOK);		\
189 	splx(s);							\
190 } while (0)
191 
192 #define putvndxfer(vnx) {						\
193 	pool_put(&vndxfer_pool, (void *)(vnx));				\
194 }
195 
196 #define	getvndbuf(vbp)	do {						\
197 	int s = splbio();						\
198 	vbp = pool_get(&vndbuf_pool, PR_MALLOCOK|PR_WAITOK);		\
199 	splx(s);							\
200 } while (0)
201 
202 #define putvndbuf(vbp) {						\
203 	pool_put(&vndbuf_pool, (void *)(vbp));				\
204 }
205 
206 /* /dev/drum */
207 bdev_decl(sw);
208 cdev_decl(sw);
209 
210 /*
211  * local variables
212  */
213 static struct extent *swapmap;		/* controls the mapping of /dev/drum */
214 
215 /* list of all active swap devices [by priority] */
216 LIST_HEAD(swap_priority, swappri);
217 static struct swap_priority swap_priority;
218 
219 /* locks */
220 struct lock swap_syscall_lock;
221 
222 /*
223  * prototypes
224  */
225 static struct swapdev	*swapdrum_getsdp __P((int));
226 
227 static struct swapdev	*swaplist_find __P((struct vnode *, int));
228 static void		 swaplist_insert __P((struct swapdev *,
229 					     struct swappri *, int));
230 static void		 swaplist_trim __P((void));
231 
232 static int swap_on __P((struct proc *, struct swapdev *));
233 static int swap_off __P((struct proc *, struct swapdev *));
234 
235 static void sw_reg_strategy __P((struct swapdev *, struct buf *, int));
236 static void sw_reg_iodone __P((struct buf *));
237 static void sw_reg_start __P((struct swapdev *));
238 
239 static int uvm_swap_io __P((struct vm_page **, int, int, int));
240 
241 /*
242  * uvm_swap_init: init the swap system data structures and locks
243  *
244  * => called at boot time from init_main.c after the filesystems
245  *	are brought up (which happens after uvm_init())
246  */
247 void
248 uvm_swap_init()
249 {
250 	UVMHIST_FUNC("uvm_swap_init");
251 
252 	UVMHIST_CALLED(pdhist);
253 	/*
254 	 * first, init the swap list, its counter, and its lock.
255 	 * then get a handle on the vnode for /dev/drum by using
256 	 * the its dev_t number ("swapdev", from MD conf.c).
257 	 */
258 
259 	LIST_INIT(&swap_priority);
260 	uvmexp.nswapdev = 0;
261 	lockinit(&swap_syscall_lock, PVM, "swapsys", 0, 0);
262 	simple_lock_init(&uvm.swap_data_lock);
263 
264 	if (bdevvp(swapdev, &swapdev_vp))
265 		panic("uvm_swap_init: can't get vnode for swap device");
266 
267 	/*
268 	 * create swap block resource map to map /dev/drum.   the range
269 	 * from 1 to INT_MAX allows 2 gigablocks of swap space.  note
270 	 * that block 0 is reserved (used to indicate an allocation
271 	 * failure, or no allocation).
272 	 */
273 	swapmap = extent_create("swapmap", 1, INT_MAX,
274 				M_VMSWAP, 0, 0, EX_NOWAIT);
275 	if (swapmap == 0)
276 		panic("uvm_swap_init: extent_create failed");
277 
278 	/*
279 	 * allocate pools for structures used for swapping to files.
280 	 */
281 
282 	pool_init(&vndxfer_pool, sizeof(struct vndxfer), 0, 0, 0,
283 	    "swp vnx", 0, NULL, NULL, 0);
284 
285 	pool_init(&vndbuf_pool, sizeof(struct vndbuf), 0, 0, 0,
286 	    "swp vnd", 0, NULL, NULL, 0);
287 
288 	/*
289 	 * done!
290 	 */
291 	UVMHIST_LOG(pdhist, "<- done", 0, 0, 0, 0);
292 }
293 
294 /*
295  * swaplist functions: functions that operate on the list of swap
296  * devices on the system.
297  */
298 
299 /*
300  * swaplist_insert: insert swap device "sdp" into the global list
301  *
302  * => caller must hold both swap_syscall_lock and uvm.swap_data_lock
303  * => caller must provide a newly malloc'd swappri structure (we will
304  *	FREE it if we don't need it... this it to prevent malloc blocking
305  *	here while adding swap)
306  */
307 static void
308 swaplist_insert(sdp, newspp, priority)
309 	struct swapdev *sdp;
310 	struct swappri *newspp;
311 	int priority;
312 {
313 	struct swappri *spp, *pspp;
314 	UVMHIST_FUNC("swaplist_insert"); UVMHIST_CALLED(pdhist);
315 
316 	/*
317 	 * find entry at or after which to insert the new device.
318 	 */
319 	pspp = NULL;
320 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
321 		if (priority <= spp->spi_priority)
322 			break;
323 		pspp = spp;
324 	}
325 
326 	/*
327 	 * new priority?
328 	 */
329 	if (spp == NULL || spp->spi_priority != priority) {
330 		spp = newspp;  /* use newspp! */
331 		UVMHIST_LOG(pdhist, "created new swappri = %d",
332 			    priority, 0, 0, 0);
333 
334 		spp->spi_priority = priority;
335 		CIRCLEQ_INIT(&spp->spi_swapdev);
336 
337 		if (pspp)
338 			LIST_INSERT_AFTER(pspp, spp, spi_swappri);
339 		else
340 			LIST_INSERT_HEAD(&swap_priority, spp, spi_swappri);
341 	} else {
342 	  	/* we don't need a new priority structure, free it */
343 		FREE(newspp, M_VMSWAP);
344 	}
345 
346 	/*
347 	 * priority found (or created).   now insert on the priority's
348 	 * circleq list and bump the total number of swapdevs.
349 	 */
350 	sdp->swd_priority = priority;
351 	CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
352 	uvmexp.nswapdev++;
353 }
354 
355 /*
356  * swaplist_find: find and optionally remove a swap device from the
357  *	global list.
358  *
359  * => caller must hold both swap_syscall_lock and uvm.swap_data_lock
360  * => we return the swapdev we found (and removed)
361  */
362 static struct swapdev *
363 swaplist_find(vp, remove)
364 	struct vnode *vp;
365 	boolean_t remove;
366 {
367 	struct swapdev *sdp;
368 	struct swappri *spp;
369 
370 	/*
371 	 * search the lists for the requested vp
372 	 */
373 
374 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
375 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
376 			if (sdp->swd_vp == vp) {
377 				if (remove) {
378 					CIRCLEQ_REMOVE(&spp->spi_swapdev,
379 					    sdp, swd_next);
380 					uvmexp.nswapdev--;
381 				}
382 				return(sdp);
383 			}
384 		}
385 	}
386 	return (NULL);
387 }
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()
398 {
399 	struct swappri *spp, *nextspp;
400 
401 	for (spp = LIST_FIRST(&swap_priority); spp != NULL; spp = nextspp) {
402 		nextspp = LIST_NEXT(spp, spi_swappri);
403 		if (CIRCLEQ_FIRST(&spp->spi_swapdev) !=
404 		    (void *)&spp->spi_swapdev)
405 			continue;
406 		LIST_REMOVE(spp, spi_swappri);
407 		free(spp, M_VMSWAP);
408 	}
409 }
410 
411 /*
412  * swapdrum_getsdp: given a page offset in /dev/drum, convert it back
413  *	to the "swapdev" that maps that section of the drum.
414  *
415  * => each swapdev takes one big contig chunk of the drum
416  * => caller must hold uvm.swap_data_lock
417  */
418 static struct swapdev *
419 swapdrum_getsdp(pgno)
420 	int pgno;
421 {
422 	struct swapdev *sdp;
423 	struct swappri *spp;
424 
425 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
426 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
427 			if (sdp->swd_flags & SWF_FAKE)
428 				continue;
429 			if (pgno >= sdp->swd_drumoffset &&
430 			    pgno < (sdp->swd_drumoffset + sdp->swd_drumsize)) {
431 				return sdp;
432 			}
433 		}
434 	}
435 	return NULL;
436 }
437 
438 
439 /*
440  * sys_swapctl: main entry point for swapctl(2) system call
441  * 	[with two helper functions: swap_on and swap_off]
442  */
443 int
444 sys_swapctl(p, v, retval)
445 	struct proc *p;
446 	void *v;
447 	register_t *retval;
448 {
449 	struct sys_swapctl_args /* {
450 		syscallarg(int) cmd;
451 		syscallarg(void *) arg;
452 		syscallarg(int) misc;
453 	} */ *uap = (struct sys_swapctl_args *)v;
454 	struct vnode *vp;
455 	struct nameidata nd;
456 	struct swappri *spp;
457 	struct swapdev *sdp;
458 	struct swapent *sep;
459 	char	userpath[PATH_MAX + 1];
460 	size_t	len;
461 	int	count, error, misc;
462 	int	priority;
463 	UVMHIST_FUNC("sys_swapctl"); UVMHIST_CALLED(pdhist);
464 
465 	misc = SCARG(uap, misc);
466 
467 	/*
468 	 * ensure serialized syscall access by grabbing the swap_syscall_lock
469 	 */
470 	lockmgr(&swap_syscall_lock, LK_EXCLUSIVE, NULL);
471 
472 	/*
473 	 * we handle the non-priv NSWAP and STATS request first.
474 	 *
475 	 * SWAP_NSWAP: return number of config'd swap devices
476 	 * [can also be obtained with uvmexp sysctl]
477 	 */
478 	if (SCARG(uap, cmd) == SWAP_NSWAP) {
479 		UVMHIST_LOG(pdhist, "<- done SWAP_NSWAP=%d", uvmexp.nswapdev,
480 		    0, 0, 0);
481 		*retval = uvmexp.nswapdev;
482 		error = 0;
483 		goto out;
484 	}
485 
486 	/*
487 	 * SWAP_STATS: get stats on current # of configured swap devs
488 	 *
489 	 * note that the swap_priority list can't change as long
490 	 * as we are holding the swap_syscall_lock.  we don't want
491 	 * to grab the uvm.swap_data_lock because we may fault&sleep during
492 	 * copyout() and we don't want to be holding that lock then!
493 	 */
494 	if (SCARG(uap, cmd) == SWAP_STATS
495 #if defined(COMPAT_13)
496 	    || SCARG(uap, cmd) == SWAP_OSTATS
497 #endif
498 	    ) {
499 		sep = (struct swapent *)SCARG(uap, arg);
500 		count = 0;
501 
502 		LIST_FOREACH(spp, &swap_priority, spi_swappri) {
503 			for (sdp = CIRCLEQ_FIRST(&spp->spi_swapdev);
504 			     sdp != (void *)&spp->spi_swapdev && misc-- > 0;
505 			     sdp = CIRCLEQ_NEXT(sdp, swd_next)) {
506 			  	/*
507 				 * backwards compatibility for system call.
508 				 * note that we use 'struct oswapent' as an
509 				 * overlay into both 'struct swapdev' and
510 				 * the userland 'struct swapent', as we
511 				 * want to retain backwards compatibility
512 				 * with NetBSD 1.3.
513 				 */
514 				sdp->swd_ose.ose_inuse =
515 				    btodb((u_int64_t)sdp->swd_npginuse <<
516 				    PAGE_SHIFT);
517 				error = copyout(&sdp->swd_ose, sep,
518 						sizeof(struct oswapent));
519 
520 				/* now copy out the path if necessary */
521 #if defined(COMPAT_13)
522 				if (error == 0 && SCARG(uap, cmd) == SWAP_STATS)
523 #else
524 				if (error == 0)
525 #endif
526 					error = copyout(sdp->swd_path,
527 					    &sep->se_path, sdp->swd_pathlen);
528 
529 				if (error)
530 					goto out;
531 				count++;
532 #if defined(COMPAT_13)
533 				if (SCARG(uap, cmd) == SWAP_OSTATS)
534 					sep = (struct swapent *)
535 					    ((struct oswapent *)sep + 1);
536 				else
537 #endif
538 					sep++;
539 			}
540 		}
541 
542 		UVMHIST_LOG(pdhist, "<- done SWAP_STATS", 0, 0, 0, 0);
543 
544 		*retval = count;
545 		error = 0;
546 		goto out;
547 	}
548 	if (SCARG(uap, cmd) == SWAP_GETDUMPDEV) {
549 		dev_t	*devp = (dev_t *)SCARG(uap, arg);
550 
551 		error = copyout(&dumpdev, devp, sizeof(dumpdev));
552 		goto out;
553 	}
554 
555 	/*
556 	 * all other requests require superuser privs.   verify.
557 	 */
558 	if ((error = suser(p->p_ucred, &p->p_acflag)))
559 		goto out;
560 
561 	/*
562 	 * at this point we expect a path name in arg.   we will
563 	 * use namei() to gain a vnode reference (vref), and lock
564 	 * the vnode (VOP_LOCK).
565 	 *
566 	 * XXX: a NULL arg means use the root vnode pointer (e.g. for
567 	 * miniroot)
568 	 */
569 	if (SCARG(uap, arg) == NULL) {
570 		vp = rootvp;		/* miniroot */
571 		if (vget(vp, LK_EXCLUSIVE)) {
572 			error = EBUSY;
573 			goto out;
574 		}
575 		if (SCARG(uap, cmd) == SWAP_ON &&
576 		    copystr("miniroot", userpath, sizeof userpath, &len))
577 			panic("swapctl: miniroot copy failed");
578 	} else {
579 		int	space;
580 		char	*where;
581 
582 		if (SCARG(uap, cmd) == SWAP_ON) {
583 			if ((error = copyinstr(SCARG(uap, arg), userpath,
584 			    sizeof userpath, &len)))
585 				goto out;
586 			space = UIO_SYSSPACE;
587 			where = userpath;
588 		} else {
589 			space = UIO_USERSPACE;
590 			where = (char *)SCARG(uap, arg);
591 		}
592 		NDINIT(&nd, LOOKUP, FOLLOW|LOCKLEAF, space, where, p);
593 		if ((error = namei(&nd)))
594 			goto out;
595 		vp = nd.ni_vp;
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 		dumpdev = vp->v_rdev;
608 		break;
609 
610 	case SWAP_CTL:
611 		/*
612 		 * get new priority, remove old entry (if any) and then
613 		 * reinsert it in the correct place.  finally, prune out
614 		 * any empty priority structures.
615 		 */
616 		priority = SCARG(uap, misc);
617 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
618 		simple_lock(&uvm.swap_data_lock);
619 		if ((sdp = swaplist_find(vp, 1)) == NULL) {
620 			error = ENOENT;
621 		} else {
622 			swaplist_insert(sdp, spp, priority);
623 			swaplist_trim();
624 		}
625 		simple_unlock(&uvm.swap_data_lock);
626 		if (error)
627 			free(spp, M_VMSWAP);
628 		break;
629 
630 	case SWAP_ON:
631 
632 		/*
633 		 * check for duplicates.   if none found, then insert a
634 		 * dummy entry on the list to prevent someone else from
635 		 * trying to enable this device while we are working on
636 		 * it.
637 		 */
638 
639 		priority = SCARG(uap, misc);
640 		sdp = malloc(sizeof *sdp, M_VMSWAP, M_WAITOK);
641 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
642 		simple_lock(&uvm.swap_data_lock);
643 		if (swaplist_find(vp, 0) != NULL) {
644 			error = EBUSY;
645 			simple_unlock(&uvm.swap_data_lock);
646 			free(sdp, M_VMSWAP);
647 			free(spp, M_VMSWAP);
648 			break;
649 		}
650 		memset(sdp, 0, sizeof(*sdp));
651 		sdp->swd_flags = SWF_FAKE;	/* placeholder only */
652 		sdp->swd_vp = vp;
653 		sdp->swd_dev = (vp->v_type == VBLK) ? vp->v_rdev : NODEV;
654 		BUFQ_INIT(&sdp->swd_tab);
655 
656 		/*
657 		 * XXX Is NFS elaboration necessary?
658 		 */
659 		if (vp->v_type == VREG) {
660 			sdp->swd_cred = crdup(p->p_ucred);
661 		}
662 
663 		swaplist_insert(sdp, spp, priority);
664 		simple_unlock(&uvm.swap_data_lock);
665 
666 		sdp->swd_pathlen = len;
667 		sdp->swd_path = malloc(sdp->swd_pathlen, M_VMSWAP, M_WAITOK);
668 		if (copystr(userpath, sdp->swd_path, sdp->swd_pathlen, 0) != 0)
669 			panic("swapctl: copystr");
670 
671 		/*
672 		 * we've now got a FAKE placeholder in the swap list.
673 		 * now attempt to enable swap on it.  if we fail, undo
674 		 * what we've done and kill the fake entry we just inserted.
675 		 * if swap_on is a success, it will clear the SWF_FAKE flag
676 		 */
677 
678 		if ((error = swap_on(p, sdp)) != 0) {
679 			simple_lock(&uvm.swap_data_lock);
680 			(void) swaplist_find(vp, 1);  /* kill fake entry */
681 			swaplist_trim();
682 			simple_unlock(&uvm.swap_data_lock);
683 			if (vp->v_type == VREG) {
684 				crfree(sdp->swd_cred);
685 			}
686 			free(sdp->swd_path, M_VMSWAP);
687 			free(sdp, M_VMSWAP);
688 			break;
689 		}
690 		break;
691 
692 	case SWAP_OFF:
693 		simple_lock(&uvm.swap_data_lock);
694 		if ((sdp = swaplist_find(vp, 0)) == NULL) {
695 			simple_unlock(&uvm.swap_data_lock);
696 			error = ENXIO;
697 			break;
698 		}
699 
700 		/*
701 		 * If a device isn't in use or enabled, we
702 		 * can't stop swapping from it (again).
703 		 */
704 		if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0) {
705 			simple_unlock(&uvm.swap_data_lock);
706 			error = EBUSY;
707 			break;
708 		}
709 
710 		/*
711 		 * do the real work.
712 		 */
713 		error = swap_off(p, sdp);
714 		break;
715 
716 	default:
717 		error = EINVAL;
718 	}
719 
720 	/*
721 	 * done!  release the ref gained by namei() and unlock.
722 	 */
723 	vput(vp);
724 
725 out:
726 	lockmgr(&swap_syscall_lock, LK_RELEASE, NULL);
727 
728 	UVMHIST_LOG(pdhist, "<- done!  error=%d", error, 0, 0, 0);
729 	return (error);
730 }
731 
732 /*
733  * swap_on: attempt to enable a swapdev for swapping.   note that the
734  *	swapdev is already on the global list, but disabled (marked
735  *	SWF_FAKE).
736  *
737  * => we avoid the start of the disk (to protect disk labels)
738  * => we also avoid the miniroot, if we are swapping to root.
739  * => caller should leave uvm.swap_data_lock unlocked, we may lock it
740  *	if needed.
741  */
742 static int
743 swap_on(p, sdp)
744 	struct proc *p;
745 	struct swapdev *sdp;
746 {
747 	static int count = 0;	/* static */
748 	struct vnode *vp;
749 	int error, npages, nblocks, size;
750 	long addr;
751 	u_long result;
752 	struct vattr va;
753 #ifdef NFS
754 	extern int (**nfsv2_vnodeop_p) __P((void *));
755 #endif /* NFS */
756 	dev_t dev;
757 	UVMHIST_FUNC("swap_on"); UVMHIST_CALLED(pdhist);
758 
759 	/*
760 	 * we want to enable swapping on sdp.   the swd_vp contains
761 	 * the vnode we want (locked and ref'd), and the swd_dev
762 	 * contains the dev_t of the file, if it a block device.
763 	 */
764 
765 	vp = sdp->swd_vp;
766 	dev = sdp->swd_dev;
767 
768 	/*
769 	 * open the swap file (mostly useful for block device files to
770 	 * let device driver know what is up).
771 	 *
772 	 * we skip the open/close for root on swap because the root
773 	 * has already been opened when root was mounted (mountroot).
774 	 */
775 	if (vp != rootvp) {
776 		if ((error = VOP_OPEN(vp, FREAD|FWRITE, p->p_ucred, p)))
777 			return (error);
778 	}
779 
780 	/* XXX this only works for block devices */
781 	UVMHIST_LOG(pdhist, "  dev=%d, major(dev)=%d", dev, major(dev), 0,0);
782 
783 	/*
784 	 * we now need to determine the size of the swap area.   for
785 	 * block specials we can call the d_psize function.
786 	 * for normal files, we must stat [get attrs].
787 	 *
788 	 * we put the result in nblks.
789 	 * for normal files, we also want the filesystem block size
790 	 * (which we get with statfs).
791 	 */
792 	switch (vp->v_type) {
793 	case VBLK:
794 		if (bdevsw[major(dev)].d_psize == 0 ||
795 		    (nblocks = (*bdevsw[major(dev)].d_psize)(dev)) == -1) {
796 			error = ENXIO;
797 			goto bad;
798 		}
799 		break;
800 
801 	case VREG:
802 		if ((error = VOP_GETATTR(vp, &va, p->p_ucred, p)))
803 			goto bad;
804 		nblocks = (int)btodb(va.va_size);
805 		if ((error =
806 		     VFS_STATFS(vp->v_mount, &vp->v_mount->mnt_stat, p)) != 0)
807 			goto bad;
808 
809 		sdp->swd_bsize = vp->v_mount->mnt_stat.f_iosize;
810 		/*
811 		 * limit the max # of outstanding I/O requests we issue
812 		 * at any one time.   take it easy on NFS servers.
813 		 */
814 #ifdef NFS
815 		if (vp->v_op == nfsv2_vnodeop_p)
816 			sdp->swd_maxactive = 2; /* XXX */
817 		else
818 #endif /* NFS */
819 			sdp->swd_maxactive = 8; /* XXX */
820 		break;
821 
822 	default:
823 		error = ENXIO;
824 		goto bad;
825 	}
826 
827 	/*
828 	 * save nblocks in a safe place and convert to pages.
829 	 */
830 
831 	sdp->swd_ose.ose_nblks = nblocks;
832 	npages = dbtob((u_int64_t)nblocks) >> PAGE_SHIFT;
833 
834 	/*
835 	 * for block special files, we want to make sure that leave
836 	 * the disklabel and bootblocks alone, so we arrange to skip
837 	 * over them (arbitrarily choosing to skip PAGE_SIZE bytes).
838 	 * note that because of this the "size" can be less than the
839 	 * actual number of blocks on the device.
840 	 */
841 	if (vp->v_type == VBLK) {
842 		/* we use pages 1 to (size - 1) [inclusive] */
843 		size = npages - 1;
844 		addr = 1;
845 	} else {
846 		/* we use pages 0 to (size - 1) [inclusive] */
847 		size = npages;
848 		addr = 0;
849 	}
850 
851 	/*
852 	 * make sure we have enough blocks for a reasonable sized swap
853 	 * area.   we want at least one page.
854 	 */
855 
856 	if (size < 1) {
857 		UVMHIST_LOG(pdhist, "  size <= 1!!", 0, 0, 0, 0);
858 		error = EINVAL;
859 		goto bad;
860 	}
861 
862 	UVMHIST_LOG(pdhist, "  dev=%x: size=%d addr=%ld\n", dev, size, addr, 0);
863 
864 	/*
865 	 * now we need to allocate an extent to manage this swap device
866 	 */
867 	snprintf(sdp->swd_exname, sizeof(sdp->swd_exname), "swap0x%04x",
868 	    count++);
869 
870 	/* note that extent_create's 3rd arg is inclusive, thus "- 1" */
871 	sdp->swd_ex = extent_create(sdp->swd_exname, 0, npages - 1, M_VMSWAP,
872 				    0, 0, EX_WAITOK);
873 	/* allocate the `saved' region from the extent so it won't be used */
874 	if (addr) {
875 		if (extent_alloc_region(sdp->swd_ex, 0, addr, EX_WAITOK))
876 			panic("disklabel region");
877 	}
878 
879 	/*
880 	 * if the vnode we are swapping to is the root vnode
881 	 * (i.e. we are swapping to the miniroot) then we want
882 	 * to make sure we don't overwrite it.   do a statfs to
883 	 * find its size and skip over it.
884 	 */
885 	if (vp == rootvp) {
886 		struct mount *mp;
887 		struct statfs *sp;
888 		int rootblocks, rootpages;
889 
890 		mp = rootvnode->v_mount;
891 		sp = &mp->mnt_stat;
892 		rootblocks = sp->f_blocks * btodb(sp->f_bsize);
893 		rootpages = round_page(dbtob(rootblocks)) >> PAGE_SHIFT;
894 		if (rootpages > size)
895 			panic("swap_on: miniroot larger than swap?");
896 
897 		if (extent_alloc_region(sdp->swd_ex, addr,
898 					rootpages, EX_WAITOK))
899 			panic("swap_on: unable to preserve miniroot");
900 
901 		size -= rootpages;
902 		printf("Preserved %d pages of miniroot ", rootpages);
903 		printf("leaving %d pages of swap\n", size);
904 	}
905 
906   	/*
907 	 * try to add anons to reflect the new swap space.
908 	 */
909 
910 	error = uvm_anon_add(size);
911 	if (error) {
912 		goto bad;
913 	}
914 
915 	/*
916 	 * add a ref to vp to reflect usage as a swap device.
917 	 */
918 	vref(vp);
919 
920 	/*
921 	 * now add the new swapdev to the drum and enable.
922 	 */
923 	if (extent_alloc(swapmap, npages, EX_NOALIGN, EX_NOBOUNDARY,
924 	    EX_WAITOK, &result))
925 		panic("swapdrum_add");
926 
927 	sdp->swd_drumoffset = (int)result;
928 	sdp->swd_drumsize = npages;
929 	sdp->swd_npages = size;
930 	simple_lock(&uvm.swap_data_lock);
931 	sdp->swd_flags &= ~SWF_FAKE;	/* going live */
932 	sdp->swd_flags |= (SWF_INUSE|SWF_ENABLE);
933 	uvmexp.swpages += size;
934 	simple_unlock(&uvm.swap_data_lock);
935 	return (0);
936 
937 	/*
938 	 * failure: clean up and return error.
939 	 */
940 
941 bad:
942 	if (sdp->swd_ex) {
943 		extent_destroy(sdp->swd_ex);
944 	}
945 	if (vp != rootvp) {
946 		(void)VOP_CLOSE(vp, FREAD|FWRITE, p->p_ucred, p);
947 	}
948 	return (error);
949 }
950 
951 /*
952  * swap_off: stop swapping on swapdev
953  *
954  * => swap data should be locked, we will unlock.
955  */
956 static int
957 swap_off(p, sdp)
958 	struct proc *p;
959 	struct swapdev *sdp;
960 {
961 	UVMHIST_FUNC("swap_off"); UVMHIST_CALLED(pdhist);
962 	UVMHIST_LOG(pdhist, "  dev=%x", sdp->swd_dev,0,0,0);
963 
964 	/* disable the swap area being removed */
965 	sdp->swd_flags &= ~SWF_ENABLE;
966 	simple_unlock(&uvm.swap_data_lock);
967 
968 	/*
969 	 * the idea is to find all the pages that are paged out to this
970 	 * device, and page them all in.  in uvm, swap-backed pageable
971 	 * memory can take two forms: aobjs and anons.  call the
972 	 * swapoff hook for each subsystem to bring in pages.
973 	 */
974 
975 	if (uao_swap_off(sdp->swd_drumoffset,
976 			 sdp->swd_drumoffset + sdp->swd_drumsize) ||
977 	    anon_swap_off(sdp->swd_drumoffset,
978 			  sdp->swd_drumoffset + sdp->swd_drumsize)) {
979 
980 		simple_lock(&uvm.swap_data_lock);
981 		sdp->swd_flags |= SWF_ENABLE;
982 		simple_unlock(&uvm.swap_data_lock);
983 		return ENOMEM;
984 	}
985 	KASSERT(sdp->swd_npginuse == sdp->swd_npgbad);
986 
987 	/*
988 	 * done with the vnode and saved creds.
989 	 * drop our ref on the vnode before calling VOP_CLOSE()
990 	 * so that spec_close() can tell if this is the last close.
991 	 */
992 	if (sdp->swd_vp->v_type == VREG) {
993 		crfree(sdp->swd_cred);
994 	}
995 	vrele(sdp->swd_vp);
996 	if (sdp->swd_vp != rootvp) {
997 		(void) VOP_CLOSE(sdp->swd_vp, FREAD|FWRITE, p->p_ucred, p);
998 	}
999 
1000 	/* remove anons from the system */
1001 	uvm_anon_remove(sdp->swd_npages);
1002 
1003 	simple_lock(&uvm.swap_data_lock);
1004 	uvmexp.swpages -= sdp->swd_npages;
1005 
1006 	if (swaplist_find(sdp->swd_vp, 1) == NULL)
1007 		panic("swap_off: swapdev not in list\n");
1008 	swaplist_trim();
1009 	simple_unlock(&uvm.swap_data_lock);
1010 
1011 	/*
1012 	 * free all resources!
1013 	 */
1014 	extent_free(swapmap, sdp->swd_drumoffset, sdp->swd_drumsize,
1015 		    EX_WAITOK);
1016 	extent_destroy(sdp->swd_ex);
1017 	free(sdp, M_VMSWAP);
1018 	return (0);
1019 }
1020 
1021 /*
1022  * /dev/drum interface and i/o functions
1023  */
1024 
1025 /*
1026  * swread: the read function for the drum (just a call to physio)
1027  */
1028 /*ARGSUSED*/
1029 int
1030 swread(dev, uio, ioflag)
1031 	dev_t dev;
1032 	struct uio *uio;
1033 	int ioflag;
1034 {
1035 	UVMHIST_FUNC("swread"); UVMHIST_CALLED(pdhist);
1036 
1037 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1038 	return (physio(swstrategy, NULL, dev, B_READ, minphys, uio));
1039 }
1040 
1041 /*
1042  * swwrite: the write function for the drum (just a call to physio)
1043  */
1044 /*ARGSUSED*/
1045 int
1046 swwrite(dev, uio, ioflag)
1047 	dev_t dev;
1048 	struct uio *uio;
1049 	int ioflag;
1050 {
1051 	UVMHIST_FUNC("swwrite"); UVMHIST_CALLED(pdhist);
1052 
1053 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1054 	return (physio(swstrategy, NULL, dev, B_WRITE, minphys, uio));
1055 }
1056 
1057 /*
1058  * swstrategy: perform I/O on the drum
1059  *
1060  * => we must map the i/o request from the drum to the correct swapdev.
1061  */
1062 void
1063 swstrategy(bp)
1064 	struct buf *bp;
1065 {
1066 	struct swapdev *sdp;
1067 	struct vnode *vp;
1068 	int s, pageno, bn;
1069 	UVMHIST_FUNC("swstrategy"); UVMHIST_CALLED(pdhist);
1070 
1071 	/*
1072 	 * convert block number to swapdev.   note that swapdev can't
1073 	 * be yanked out from under us because we are holding resources
1074 	 * in it (i.e. the blocks we are doing I/O on).
1075 	 */
1076 	pageno = dbtob((int64_t)bp->b_blkno) >> PAGE_SHIFT;
1077 	simple_lock(&uvm.swap_data_lock);
1078 	sdp = swapdrum_getsdp(pageno);
1079 	simple_unlock(&uvm.swap_data_lock);
1080 	if (sdp == NULL) {
1081 		bp->b_error = EINVAL;
1082 		bp->b_flags |= B_ERROR;
1083 		biodone(bp);
1084 		UVMHIST_LOG(pdhist, "  failed to get swap device", 0, 0, 0, 0);
1085 		return;
1086 	}
1087 
1088 	/*
1089 	 * convert drum page number to block number on this swapdev.
1090 	 */
1091 
1092 	pageno -= sdp->swd_drumoffset;	/* page # on swapdev */
1093 	bn = btodb((u_int64_t)pageno << PAGE_SHIFT); /* convert to diskblock */
1094 
1095 	UVMHIST_LOG(pdhist, "  %s: mapoff=%x bn=%x bcount=%ld",
1096 		((bp->b_flags & B_READ) == 0) ? "write" : "read",
1097 		sdp->swd_drumoffset, bn, bp->b_bcount);
1098 
1099 	/*
1100 	 * for block devices we finish up here.
1101 	 * for regular files we have to do more work which we delegate
1102 	 * to sw_reg_strategy().
1103 	 */
1104 
1105 	switch (sdp->swd_vp->v_type) {
1106 	default:
1107 		panic("swstrategy: vnode type 0x%x", sdp->swd_vp->v_type);
1108 
1109 	case VBLK:
1110 
1111 		/*
1112 		 * must convert "bp" from an I/O on /dev/drum to an I/O
1113 		 * on the swapdev (sdp).
1114 		 */
1115 		s = splbio();
1116 		bp->b_blkno = bn;		/* swapdev block number */
1117 		vp = sdp->swd_vp;		/* swapdev vnode pointer */
1118 		bp->b_dev = sdp->swd_dev;	/* swapdev dev_t */
1119 
1120 		/*
1121 		 * if we are doing a write, we have to redirect the i/o on
1122 		 * drum's v_numoutput counter to the swapdevs.
1123 		 */
1124 		if ((bp->b_flags & B_READ) == 0) {
1125 			vwakeup(bp);	/* kills one 'v_numoutput' on drum */
1126 			vp->v_numoutput++;	/* put it on swapdev */
1127 		}
1128 
1129 		/*
1130 		 * finally plug in swapdev vnode and start I/O
1131 		 */
1132 		bp->b_vp = vp;
1133 		splx(s);
1134 		VOP_STRATEGY(bp);
1135 		return;
1136 
1137 	case VREG:
1138 		/*
1139 		 * delegate to sw_reg_strategy function.
1140 		 */
1141 		sw_reg_strategy(sdp, bp, bn);
1142 		return;
1143 	}
1144 	/* NOTREACHED */
1145 }
1146 
1147 /*
1148  * sw_reg_strategy: handle swap i/o to regular files
1149  */
1150 static void
1151 sw_reg_strategy(sdp, bp, bn)
1152 	struct swapdev	*sdp;
1153 	struct buf	*bp;
1154 	int		bn;
1155 {
1156 	struct vnode	*vp;
1157 	struct vndxfer	*vnx;
1158 	daddr_t		nbn;
1159 	caddr_t		addr;
1160 	off_t		byteoff;
1161 	int		s, off, nra, error, sz, resid;
1162 	UVMHIST_FUNC("sw_reg_strategy"); UVMHIST_CALLED(pdhist);
1163 
1164 	/*
1165 	 * allocate a vndxfer head for this transfer and point it to
1166 	 * our buffer.
1167 	 */
1168 	getvndxfer(vnx);
1169 	vnx->vx_flags = VX_BUSY;
1170 	vnx->vx_error = 0;
1171 	vnx->vx_pending = 0;
1172 	vnx->vx_bp = bp;
1173 	vnx->vx_sdp = sdp;
1174 
1175 	/*
1176 	 * setup for main loop where we read filesystem blocks into
1177 	 * our buffer.
1178 	 */
1179 	error = 0;
1180 	bp->b_resid = bp->b_bcount;	/* nothing transfered yet! */
1181 	addr = bp->b_data;		/* current position in buffer */
1182 	byteoff = dbtob((u_int64_t)bn);
1183 
1184 	for (resid = bp->b_resid; resid; resid -= sz) {
1185 		struct vndbuf	*nbp;
1186 
1187 		/*
1188 		 * translate byteoffset into block number.  return values:
1189 		 *   vp = vnode of underlying device
1190 		 *  nbn = new block number (on underlying vnode dev)
1191 		 *  nra = num blocks we can read-ahead (excludes requested
1192 		 *	block)
1193 		 */
1194 		nra = 0;
1195 		error = VOP_BMAP(sdp->swd_vp, byteoff / sdp->swd_bsize,
1196 				 	&vp, &nbn, &nra);
1197 
1198 		if (error == 0 && nbn == (daddr_t)-1) {
1199 			/*
1200 			 * this used to just set error, but that doesn't
1201 			 * do the right thing.  Instead, it causes random
1202 			 * memory errors.  The panic() should remain until
1203 			 * this condition doesn't destabilize the system.
1204 			 */
1205 #if 1
1206 			panic("sw_reg_strategy: swap to sparse file");
1207 #else
1208 			error = EIO;	/* failure */
1209 #endif
1210 		}
1211 
1212 		/*
1213 		 * punt if there was an error or a hole in the file.
1214 		 * we must wait for any i/o ops we have already started
1215 		 * to finish before returning.
1216 		 *
1217 		 * XXX we could deal with holes here but it would be
1218 		 * a hassle (in the write case).
1219 		 */
1220 		if (error) {
1221 			s = splbio();
1222 			vnx->vx_error = error;	/* pass error up */
1223 			goto out;
1224 		}
1225 
1226 		/*
1227 		 * compute the size ("sz") of this transfer (in bytes).
1228 		 */
1229 		off = byteoff % sdp->swd_bsize;
1230 		sz = (1 + nra) * sdp->swd_bsize - off;
1231 		if (sz > resid)
1232 			sz = resid;
1233 
1234 		UVMHIST_LOG(pdhist, "sw_reg_strategy: "
1235 			    "vp %p/%p offset 0x%x/0x%x",
1236 			    sdp->swd_vp, vp, byteoff, nbn);
1237 
1238 		/*
1239 		 * now get a buf structure.   note that the vb_buf is
1240 		 * at the front of the nbp structure so that you can
1241 		 * cast pointers between the two structure easily.
1242 		 */
1243 		getvndbuf(nbp);
1244 		nbp->vb_buf.b_flags    = bp->b_flags | B_CALL;
1245 		nbp->vb_buf.b_bcount   = sz;
1246 		nbp->vb_buf.b_bufsize  = sz;
1247 		nbp->vb_buf.b_error    = 0;
1248 		nbp->vb_buf.b_data     = addr;
1249 		nbp->vb_buf.b_lblkno   = 0;
1250 		nbp->vb_buf.b_blkno    = nbn + btodb(off);
1251 		nbp->vb_buf.b_rawblkno = nbp->vb_buf.b_blkno;
1252 		nbp->vb_buf.b_iodone   = sw_reg_iodone;
1253 		nbp->vb_buf.b_vp       = vp;
1254 		if (vp->v_type == VBLK) {
1255 			nbp->vb_buf.b_dev = vp->v_rdev;
1256 		}
1257 		LIST_INIT(&nbp->vb_buf.b_dep);
1258 
1259 		nbp->vb_xfer = vnx;	/* patch it back in to vnx */
1260 
1261 		/*
1262 		 * Just sort by block number
1263 		 */
1264 		s = splbio();
1265 		if (vnx->vx_error != 0) {
1266 			putvndbuf(nbp);
1267 			goto out;
1268 		}
1269 		vnx->vx_pending++;
1270 
1271 		/* sort it in and start I/O if we are not over our limit */
1272 		disksort_blkno(&sdp->swd_tab, &nbp->vb_buf);
1273 		sw_reg_start(sdp);
1274 		splx(s);
1275 
1276 		/*
1277 		 * advance to the next I/O
1278 		 */
1279 		byteoff += sz;
1280 		addr += sz;
1281 	}
1282 
1283 	s = splbio();
1284 
1285 out: /* Arrive here at splbio */
1286 	vnx->vx_flags &= ~VX_BUSY;
1287 	if (vnx->vx_pending == 0) {
1288 		if (vnx->vx_error != 0) {
1289 			bp->b_error = vnx->vx_error;
1290 			bp->b_flags |= B_ERROR;
1291 		}
1292 		putvndxfer(vnx);
1293 		biodone(bp);
1294 	}
1295 	splx(s);
1296 }
1297 
1298 /*
1299  * sw_reg_start: start an I/O request on the requested swapdev
1300  *
1301  * => reqs are sorted by disksort (above)
1302  */
1303 static void
1304 sw_reg_start(sdp)
1305 	struct swapdev	*sdp;
1306 {
1307 	struct buf	*bp;
1308 	UVMHIST_FUNC("sw_reg_start"); UVMHIST_CALLED(pdhist);
1309 
1310 	/* recursion control */
1311 	if ((sdp->swd_flags & SWF_BUSY) != 0)
1312 		return;
1313 
1314 	sdp->swd_flags |= SWF_BUSY;
1315 
1316 	while (sdp->swd_active < sdp->swd_maxactive) {
1317 		bp = BUFQ_FIRST(&sdp->swd_tab);
1318 		if (bp == NULL)
1319 			break;
1320 		BUFQ_REMOVE(&sdp->swd_tab, bp);
1321 		sdp->swd_active++;
1322 
1323 		UVMHIST_LOG(pdhist,
1324 		    "sw_reg_start:  bp %p vp %p blkno %p cnt %lx",
1325 		    bp, bp->b_vp, bp->b_blkno, bp->b_bcount);
1326 		if ((bp->b_flags & B_READ) == 0)
1327 			bp->b_vp->v_numoutput++;
1328 
1329 		VOP_STRATEGY(bp);
1330 	}
1331 	sdp->swd_flags &= ~SWF_BUSY;
1332 }
1333 
1334 /*
1335  * sw_reg_iodone: one of our i/o's has completed and needs post-i/o cleanup
1336  *
1337  * => note that we can recover the vndbuf struct by casting the buf ptr
1338  */
1339 static void
1340 sw_reg_iodone(bp)
1341 	struct buf *bp;
1342 {
1343 	struct vndbuf *vbp = (struct vndbuf *) bp;
1344 	struct vndxfer *vnx = vbp->vb_xfer;
1345 	struct buf *pbp = vnx->vx_bp;		/* parent buffer */
1346 	struct swapdev	*sdp = vnx->vx_sdp;
1347 	int		s, resid;
1348 	UVMHIST_FUNC("sw_reg_iodone"); UVMHIST_CALLED(pdhist);
1349 
1350 	UVMHIST_LOG(pdhist, "  vbp=%p vp=%p blkno=%x addr=%p",
1351 	    vbp, vbp->vb_buf.b_vp, vbp->vb_buf.b_blkno, vbp->vb_buf.b_data);
1352 	UVMHIST_LOG(pdhist, "  cnt=%lx resid=%lx",
1353 	    vbp->vb_buf.b_bcount, vbp->vb_buf.b_resid, 0, 0);
1354 
1355 	/*
1356 	 * protect vbp at splbio and update.
1357 	 */
1358 
1359 	s = splbio();
1360 	resid = vbp->vb_buf.b_bcount - vbp->vb_buf.b_resid;
1361 	pbp->b_resid -= resid;
1362 	vnx->vx_pending--;
1363 
1364 	if (vbp->vb_buf.b_error) {
1365 		UVMHIST_LOG(pdhist, "  got error=%d !",
1366 		    vbp->vb_buf.b_error, 0, 0, 0);
1367 
1368 		/* pass error upward */
1369 		vnx->vx_error = vbp->vb_buf.b_error;
1370 	}
1371 
1372 	/*
1373 	 * kill vbp structure
1374 	 */
1375 	putvndbuf(vbp);
1376 
1377 	/*
1378 	 * wrap up this transaction if it has run to completion or, in
1379 	 * case of an error, when all auxiliary buffers have returned.
1380 	 */
1381 	if (vnx->vx_error != 0) {
1382 		/* pass error upward */
1383 		pbp->b_flags |= B_ERROR;
1384 		pbp->b_error = vnx->vx_error;
1385 		if ((vnx->vx_flags & VX_BUSY) == 0 && vnx->vx_pending == 0) {
1386 			putvndxfer(vnx);
1387 			biodone(pbp);
1388 		}
1389 	} else if (pbp->b_resid == 0) {
1390 		KASSERT(vnx->vx_pending == 0);
1391 		if ((vnx->vx_flags & VX_BUSY) == 0) {
1392 			UVMHIST_LOG(pdhist, "  iodone error=%d !",
1393 			    pbp, vnx->vx_error, 0, 0);
1394 			putvndxfer(vnx);
1395 			biodone(pbp);
1396 		}
1397 	}
1398 
1399 	/*
1400 	 * done!   start next swapdev I/O if one is pending
1401 	 */
1402 	sdp->swd_active--;
1403 	sw_reg_start(sdp);
1404 	splx(s);
1405 }
1406 
1407 
1408 /*
1409  * uvm_swap_alloc: allocate space on swap
1410  *
1411  * => allocation is done "round robin" down the priority list, as we
1412  *	allocate in a priority we "rotate" the circle queue.
1413  * => space can be freed with uvm_swap_free
1414  * => we return the page slot number in /dev/drum (0 == invalid slot)
1415  * => we lock uvm.swap_data_lock
1416  * => XXXMRG: "LESSOK" INTERFACE NEEDED TO EXTENT SYSTEM
1417  */
1418 int
1419 uvm_swap_alloc(nslots, lessok)
1420 	int *nslots;	/* IN/OUT */
1421 	boolean_t lessok;
1422 {
1423 	struct swapdev *sdp;
1424 	struct swappri *spp;
1425 	u_long	result;
1426 	UVMHIST_FUNC("uvm_swap_alloc"); UVMHIST_CALLED(pdhist);
1427 
1428 	/*
1429 	 * no swap devices configured yet?   definite failure.
1430 	 */
1431 	if (uvmexp.nswapdev < 1)
1432 		return 0;
1433 
1434 	/*
1435 	 * lock data lock, convert slots into blocks, and enter loop
1436 	 */
1437 	simple_lock(&uvm.swap_data_lock);
1438 
1439 ReTry:	/* XXXMRG */
1440 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
1441 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
1442 			/* if it's not enabled, then we can't swap from it */
1443 			if ((sdp->swd_flags & SWF_ENABLE) == 0)
1444 				continue;
1445 			if (sdp->swd_npginuse + *nslots > sdp->swd_npages)
1446 				continue;
1447 			if (extent_alloc(sdp->swd_ex, *nslots, EX_NOALIGN,
1448 					 EX_NOBOUNDARY, EX_MALLOCOK|EX_NOWAIT,
1449 					 &result) != 0) {
1450 				continue;
1451 			}
1452 
1453 			/*
1454 			 * successful allocation!  now rotate the circleq.
1455 			 */
1456 			CIRCLEQ_REMOVE(&spp->spi_swapdev, sdp, swd_next);
1457 			CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
1458 			sdp->swd_npginuse += *nslots;
1459 			uvmexp.swpginuse += *nslots;
1460 			simple_unlock(&uvm.swap_data_lock);
1461 			/* done!  return drum slot number */
1462 			UVMHIST_LOG(pdhist,
1463 			    "success!  returning %d slots starting at %d",
1464 			    *nslots, result + sdp->swd_drumoffset, 0, 0);
1465 			return (result + sdp->swd_drumoffset);
1466 		}
1467 	}
1468 
1469 	/* XXXMRG: BEGIN HACK */
1470 	if (*nslots > 1 && lessok) {
1471 		*nslots = 1;
1472 		goto ReTry;	/* XXXMRG: ugh!  extent should support this for us */
1473 	}
1474 	/* XXXMRG: END HACK */
1475 
1476 	simple_unlock(&uvm.swap_data_lock);
1477 	return 0;
1478 }
1479 
1480 /*
1481  * uvm_swap_markbad: keep track of swap ranges where we've had i/o errors
1482  *
1483  * => we lock uvm.swap_data_lock
1484  */
1485 void
1486 uvm_swap_markbad(startslot, nslots)
1487 	int startslot;
1488 	int nslots;
1489 {
1490 	struct swapdev *sdp;
1491 	UVMHIST_FUNC("uvm_swap_markbad"); UVMHIST_CALLED(pdhist);
1492 
1493 	simple_lock(&uvm.swap_data_lock);
1494 	sdp = swapdrum_getsdp(startslot);
1495 
1496 	/*
1497 	 * we just keep track of how many pages have been marked bad
1498 	 * in this device, to make everything add up in swap_off().
1499 	 * we assume here that the range of slots will all be within
1500 	 * one swap device.
1501 	 */
1502 
1503 	sdp->swd_npgbad += nslots;
1504 	UVMHIST_LOG(pdhist, "now %d bad", sdp->swd_npgbad, 0,0,0);
1505 	simple_unlock(&uvm.swap_data_lock);
1506 }
1507 
1508 /*
1509  * uvm_swap_free: free swap slots
1510  *
1511  * => this can be all or part of an allocation made by uvm_swap_alloc
1512  * => we lock uvm.swap_data_lock
1513  */
1514 void
1515 uvm_swap_free(startslot, nslots)
1516 	int startslot;
1517 	int nslots;
1518 {
1519 	struct swapdev *sdp;
1520 	UVMHIST_FUNC("uvm_swap_free"); UVMHIST_CALLED(pdhist);
1521 
1522 	UVMHIST_LOG(pdhist, "freeing %d slots starting at %d", nslots,
1523 	    startslot, 0, 0);
1524 
1525 	/*
1526 	 * ignore attempts to free the "bad" slot.
1527 	 */
1528 
1529 	if (startslot == SWSLOT_BAD) {
1530 		return;
1531 	}
1532 
1533 	/*
1534 	 * convert drum slot offset back to sdp, free the blocks
1535 	 * in the extent, and return.   must hold pri lock to do
1536 	 * lookup and access the extent.
1537 	 */
1538 
1539 	simple_lock(&uvm.swap_data_lock);
1540 	sdp = swapdrum_getsdp(startslot);
1541 	KASSERT(uvmexp.nswapdev >= 1);
1542 	KASSERT(sdp != NULL);
1543 	KASSERT(sdp->swd_npginuse >= nslots);
1544 	if (extent_free(sdp->swd_ex, startslot - sdp->swd_drumoffset, nslots,
1545 			EX_MALLOCOK|EX_NOWAIT) != 0) {
1546 		printf("warning: resource shortage: %d pages of swap lost\n",
1547 			nslots);
1548 	}
1549 	sdp->swd_npginuse -= nslots;
1550 	uvmexp.swpginuse -= nslots;
1551 	simple_unlock(&uvm.swap_data_lock);
1552 }
1553 
1554 /*
1555  * uvm_swap_put: put any number of pages into a contig place on swap
1556  *
1557  * => can be sync or async
1558  */
1559 
1560 int
1561 uvm_swap_put(swslot, ppsp, npages, flags)
1562 	int swslot;
1563 	struct vm_page **ppsp;
1564 	int npages;
1565 	int flags;
1566 {
1567 	int error;
1568 
1569 	error = uvm_swap_io(ppsp, swslot, npages, B_WRITE |
1570 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1571 	return error;
1572 }
1573 
1574 /*
1575  * uvm_swap_get: get a single page from swap
1576  *
1577  * => usually a sync op (from fault)
1578  */
1579 
1580 int
1581 uvm_swap_get(page, swslot, flags)
1582 	struct vm_page *page;
1583 	int swslot, flags;
1584 {
1585 	int error;
1586 
1587 	uvmexp.nswget++;
1588 	KASSERT(flags & PGO_SYNCIO);
1589 	if (swslot == SWSLOT_BAD) {
1590 		return EIO;
1591 	}
1592 	error = uvm_swap_io(&page, swslot, 1, B_READ |
1593 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1594 	if (error == 0) {
1595 
1596 		/*
1597 		 * this page is no longer only in swap.
1598 		 */
1599 
1600 		simple_lock(&uvm.swap_data_lock);
1601 		KASSERT(uvmexp.swpgonly > 0);
1602 		uvmexp.swpgonly--;
1603 		simple_unlock(&uvm.swap_data_lock);
1604 	}
1605 	return error;
1606 }
1607 
1608 /*
1609  * uvm_swap_io: do an i/o operation to swap
1610  */
1611 
1612 static int
1613 uvm_swap_io(pps, startslot, npages, flags)
1614 	struct vm_page **pps;
1615 	int startslot, npages, flags;
1616 {
1617 	daddr_t startblk;
1618 	struct	buf *bp;
1619 	vaddr_t kva;
1620 	int	error, s, mapinflags;
1621 	boolean_t write, async;
1622 	UVMHIST_FUNC("uvm_swap_io"); UVMHIST_CALLED(pdhist);
1623 
1624 	UVMHIST_LOG(pdhist, "<- called, startslot=%d, npages=%d, flags=%d",
1625 	    startslot, npages, flags, 0);
1626 
1627 	write = (flags & B_READ) == 0;
1628 	async = (flags & B_ASYNC) != 0;
1629 
1630 	/*
1631 	 * convert starting drum slot to block number
1632 	 */
1633 
1634 	startblk = btodb((u_int64_t)startslot << PAGE_SHIFT);
1635 
1636 	/*
1637 	 * first, map the pages into the kernel.
1638 	 */
1639 
1640 	mapinflags = !write ?
1641 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_READ :
1642 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_WRITE;
1643 	kva = uvm_pagermapin(pps, npages, mapinflags);
1644 
1645 	/*
1646 	 * now allocate a buf for the i/o.
1647 	 */
1648 
1649 	s = splbio();
1650 	bp = pool_get(&bufpool, PR_WAITOK);
1651 	splx(s);
1652 
1653 	/*
1654 	 * fill in the bp/sbp.   we currently route our i/o through
1655 	 * /dev/drum's vnode [swapdev_vp].
1656 	 */
1657 
1658 	bp->b_flags = B_BUSY | B_NOCACHE | (flags & (B_READ|B_ASYNC));
1659 	bp->b_proc = &proc0;	/* XXX */
1660 	bp->b_vnbufs.le_next = NOLIST;
1661 	bp->b_data = (caddr_t)kva;
1662 	bp->b_blkno = startblk;
1663 	bp->b_vp = swapdev_vp;
1664 	bp->b_dev = swapdev_vp->v_rdev;
1665 	bp->b_bufsize = bp->b_bcount = npages << PAGE_SHIFT;
1666 	LIST_INIT(&bp->b_dep);
1667 
1668 	/*
1669 	 * bump v_numoutput (counter of number of active outputs).
1670 	 */
1671 
1672 	if (write) {
1673 		s = splbio();
1674 		swapdev_vp->v_numoutput++;
1675 		splx(s);
1676 	}
1677 
1678 	/*
1679 	 * for async ops we must set up the iodone handler.
1680 	 */
1681 
1682 	if (async) {
1683 		bp->b_flags |= B_CALL;
1684 		bp->b_iodone = uvm_aio_biodone;
1685 		UVMHIST_LOG(pdhist, "doing async!", 0, 0, 0, 0);
1686 	}
1687 	UVMHIST_LOG(pdhist,
1688 	    "about to start io: data = %p blkno = 0x%x, bcount = %ld",
1689 	    bp->b_data, bp->b_blkno, bp->b_bcount, 0);
1690 
1691 	/*
1692 	 * now we start the I/O, and if async, return.
1693 	 */
1694 
1695 	VOP_STRATEGY(bp);
1696 	if (async)
1697 		return 0;
1698 
1699 	/*
1700 	 * must be sync i/o.   wait for it to finish
1701 	 */
1702 
1703 	error = biowait(bp);
1704 
1705 	/*
1706 	 * kill the pager mapping
1707 	 */
1708 
1709 	uvm_pagermapout(kva, npages);
1710 
1711 	/*
1712 	 * now dispose of the buf and we're done.
1713 	 */
1714 
1715 	s = splbio();
1716 	if (write)
1717 		vwakeup(bp);
1718 	pool_put(&bufpool, bp);
1719 	splx(s);
1720 	UVMHIST_LOG(pdhist, "<- done (sync)  error=%d", error, 0, 0, 0);
1721 	return (error);
1722 }
1723