xref: /netbsd-src/sys/uvm/uvm_swap.c (revision ce2c90c7c172d95d2402a5b3d96d8f8e6d138a21)
1 /*	$NetBSD: uvm_swap.c,v 1.108 2006/10/12 04:35:40 thorpej 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 1.108 2006/10/12 04:35:40 thorpej 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/bufq.h>
46 #include <sys/conf.h>
47 #include <sys/proc.h>
48 #include <sys/namei.h>
49 #include <sys/disklabel.h>
50 #include <sys/errno.h>
51 #include <sys/kernel.h>
52 #include <sys/malloc.h>
53 #include <sys/vnode.h>
54 #include <sys/file.h>
55 #include <sys/extent.h>
56 #include <sys/blist.h>
57 #include <sys/mount.h>
58 #include <sys/pool.h>
59 #include <sys/sa.h>
60 #include <sys/syscallargs.h>
61 #include <sys/swap.h>
62 #include <sys/kauth.h>
63 
64 #include <uvm/uvm.h>
65 
66 #include <miscfs/specfs/specdev.h>
67 
68 /*
69  * uvm_swap.c: manage configuration and i/o to swap space.
70  */
71 
72 /*
73  * swap space is managed in the following way:
74  *
75  * each swap partition or file is described by a "swapdev" structure.
76  * each "swapdev" structure contains a "swapent" structure which contains
77  * information that is passed up to the user (via system calls).
78  *
79  * each swap partition is assigned a "priority" (int) which controls
80  * swap parition usage.
81  *
82  * the system maintains a global data structure describing all swap
83  * partitions/files.   there is a sorted LIST of "swappri" structures
84  * which describe "swapdev"'s at that priority.   this LIST is headed
85  * by the "swap_priority" global var.    each "swappri" contains a
86  * CIRCLEQ of "swapdev" structures at that priority.
87  *
88  * locking:
89  *  - swap_syscall_lock (sleep lock): this lock serializes the swapctl
90  *    system call and prevents the swap priority list from changing
91  *    while we are in the middle of a system call (e.g. SWAP_STATS).
92  *  - uvm.swap_data_lock (simple_lock): this lock protects all swap data
93  *    structures including the priority list, the swapdev structures,
94  *    and the swapmap extent.
95  *
96  * each swap device has the following info:
97  *  - swap device in use (could be disabled, preventing future use)
98  *  - swap enabled (allows new allocations on swap)
99  *  - map info in /dev/drum
100  *  - vnode pointer
101  * for swap files only:
102  *  - block size
103  *  - max byte count in buffer
104  *  - buffer
105  *
106  * userland controls and configures swap with the swapctl(2) system call.
107  * the sys_swapctl performs the following operations:
108  *  [1] SWAP_NSWAP: returns the number of swap devices currently configured
109  *  [2] SWAP_STATS: given a pointer to an array of swapent structures
110  *	(passed in via "arg") of a size passed in via "misc" ... we load
111  *	the current swap config into the array. The actual work is done
112  *	in the uvm_swap_stats(9) function.
113  *  [3] SWAP_ON: given a pathname in arg (could be device or file) and a
114  *	priority in "misc", start swapping on it.
115  *  [4] SWAP_OFF: as SWAP_ON, but stops swapping to a device
116  *  [5] SWAP_CTL: changes the priority of a swap device (new priority in
117  *	"misc")
118  */
119 
120 /*
121  * swapdev: describes a single swap partition/file
122  *
123  * note the following should be true:
124  * swd_inuse <= swd_nblks  [number of blocks in use is <= total blocks]
125  * swd_nblks <= swd_mapsize [because mapsize includes miniroot+disklabel]
126  */
127 struct swapdev {
128 	struct oswapent swd_ose;
129 #define	swd_dev		swd_ose.ose_dev		/* device id */
130 #define	swd_flags	swd_ose.ose_flags	/* flags:inuse/enable/fake */
131 #define	swd_priority	swd_ose.ose_priority	/* our priority */
132 	/* also: swd_ose.ose_nblks, swd_ose.ose_inuse */
133 	char			*swd_path;	/* saved pathname of device */
134 	int			swd_pathlen;	/* length of pathname */
135 	int			swd_npages;	/* #pages we can use */
136 	int			swd_npginuse;	/* #pages in use */
137 	int			swd_npgbad;	/* #pages bad */
138 	int			swd_drumoffset;	/* page0 offset in drum */
139 	int			swd_drumsize;	/* #pages in drum */
140 	blist_t			swd_blist;	/* blist for this swapdev */
141 	struct vnode		*swd_vp;	/* backing vnode */
142 	CIRCLEQ_ENTRY(swapdev)	swd_next;	/* priority circleq */
143 
144 	int			swd_bsize;	/* blocksize (bytes) */
145 	int			swd_maxactive;	/* max active i/o reqs */
146 	struct bufq_state	*swd_tab;	/* buffer list */
147 	int			swd_active;	/* number of active buffers */
148 };
149 
150 /*
151  * swap device priority entry; the list is kept sorted on `spi_priority'.
152  */
153 struct swappri {
154 	int			spi_priority;     /* priority */
155 	CIRCLEQ_HEAD(spi_swapdev, swapdev)	spi_swapdev;
156 	/* circleq of swapdevs at this priority */
157 	LIST_ENTRY(swappri)	spi_swappri;      /* global list of pri's */
158 };
159 
160 /*
161  * The following two structures are used to keep track of data transfers
162  * on swap devices associated with regular files.
163  * NOTE: this code is more or less a copy of vnd.c; we use the same
164  * structure names here to ease porting..
165  */
166 struct vndxfer {
167 	struct buf	*vx_bp;		/* Pointer to parent buffer */
168 	struct swapdev	*vx_sdp;
169 	int		vx_error;
170 	int		vx_pending;	/* # of pending aux buffers */
171 	int		vx_flags;
172 #define VX_BUSY		1
173 #define VX_DEAD		2
174 };
175 
176 struct vndbuf {
177 	struct buf	vb_buf;
178 	struct vndxfer	*vb_xfer;
179 };
180 
181 
182 /*
183  * We keep a of pool vndbuf's and vndxfer structures.
184  */
185 POOL_INIT(vndxfer_pool, sizeof(struct vndxfer), 0, 0, 0, "swp vnx", NULL);
186 POOL_INIT(vndbuf_pool, sizeof(struct vndbuf), 0, 0, 0, "swp vnd", NULL);
187 
188 #define	getvndxfer(vnx)	do {						\
189 	int sp = splbio();						\
190 	vnx = pool_get(&vndxfer_pool, PR_WAITOK);			\
191 	splx(sp);							\
192 } while (/*CONSTCOND*/ 0)
193 
194 #define putvndxfer(vnx) {						\
195 	pool_put(&vndxfer_pool, (void *)(vnx));				\
196 }
197 
198 #define	getvndbuf(vbp)	do {						\
199 	int sp = splbio();						\
200 	vbp = pool_get(&vndbuf_pool, PR_WAITOK);			\
201 	splx(sp);							\
202 } while (/*CONSTCOND*/ 0)
203 
204 #define putvndbuf(vbp) {						\
205 	pool_put(&vndbuf_pool, (void *)(vbp));				\
206 }
207 
208 /*
209  * local variables
210  */
211 static struct extent *swapmap;		/* controls the mapping of /dev/drum */
212 
213 MALLOC_DEFINE(M_VMSWAP, "VM swap", "VM swap structures");
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 static struct lock swap_syscall_lock;
221 
222 /*
223  * prototypes
224  */
225 static struct swapdev	*swapdrum_getsdp(int);
226 
227 static struct swapdev	*swaplist_find(struct vnode *, int);
228 static void		 swaplist_insert(struct swapdev *,
229 					 struct swappri *, int);
230 static void		 swaplist_trim(void);
231 
232 static int swap_on(struct lwp *, struct swapdev *);
233 static int swap_off(struct lwp *, struct swapdev *);
234 
235 static void uvm_swap_stats_locked(int, struct swapent *, int, register_t *);
236 
237 static void sw_reg_strategy(struct swapdev *, struct buf *, int);
238 static void sw_reg_iodone(struct buf *);
239 static void sw_reg_start(struct swapdev *);
240 
241 static int uvm_swap_io(struct vm_page **, int, int, int);
242 
243 /*
244  * uvm_swap_init: init the swap system data structures and locks
245  *
246  * => called at boot time from init_main.c after the filesystems
247  *	are brought up (which happens after uvm_init())
248  */
249 void
250 uvm_swap_init(void)
251 {
252 	UVMHIST_FUNC("uvm_swap_init");
253 
254 	UVMHIST_CALLED(pdhist);
255 	/*
256 	 * first, init the swap list, its counter, and its lock.
257 	 * then get a handle on the vnode for /dev/drum by using
258 	 * the its dev_t number ("swapdev", from MD conf.c).
259 	 */
260 
261 	LIST_INIT(&swap_priority);
262 	uvmexp.nswapdev = 0;
263 	lockinit(&swap_syscall_lock, PVM, "swapsys", 0, 0);
264 	simple_lock_init(&uvm.swap_data_lock);
265 
266 	if (bdevvp(swapdev, &swapdev_vp))
267 		panic("uvm_swap_init: can't get vnode for swap device");
268 
269 	/*
270 	 * create swap block resource map to map /dev/drum.   the range
271 	 * from 1 to INT_MAX allows 2 gigablocks of swap space.  note
272 	 * that block 0 is reserved (used to indicate an allocation
273 	 * failure, or no allocation).
274 	 */
275 	swapmap = extent_create("swapmap", 1, INT_MAX,
276 				M_VMSWAP, 0, 0, EX_NOWAIT);
277 	if (swapmap == 0)
278 		panic("uvm_swap_init: extent_create failed");
279 
280 	/*
281 	 * done!
282 	 */
283 	UVMHIST_LOG(pdhist, "<- done", 0, 0, 0, 0);
284 }
285 
286 /*
287  * swaplist functions: functions that operate on the list of swap
288  * devices on the system.
289  */
290 
291 /*
292  * swaplist_insert: insert swap device "sdp" into the global list
293  *
294  * => caller must hold both swap_syscall_lock and uvm.swap_data_lock
295  * => caller must provide a newly malloc'd swappri structure (we will
296  *	FREE it if we don't need it... this it to prevent malloc blocking
297  *	here while adding swap)
298  */
299 static void
300 swaplist_insert(struct swapdev *sdp, struct swappri *newspp, int priority)
301 {
302 	struct swappri *spp, *pspp;
303 	UVMHIST_FUNC("swaplist_insert"); UVMHIST_CALLED(pdhist);
304 
305 	/*
306 	 * find entry at or after which to insert the new device.
307 	 */
308 	pspp = NULL;
309 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
310 		if (priority <= spp->spi_priority)
311 			break;
312 		pspp = spp;
313 	}
314 
315 	/*
316 	 * new priority?
317 	 */
318 	if (spp == NULL || spp->spi_priority != priority) {
319 		spp = newspp;  /* use newspp! */
320 		UVMHIST_LOG(pdhist, "created new swappri = %d",
321 			    priority, 0, 0, 0);
322 
323 		spp->spi_priority = priority;
324 		CIRCLEQ_INIT(&spp->spi_swapdev);
325 
326 		if (pspp)
327 			LIST_INSERT_AFTER(pspp, spp, spi_swappri);
328 		else
329 			LIST_INSERT_HEAD(&swap_priority, spp, spi_swappri);
330 	} else {
331 	  	/* we don't need a new priority structure, free it */
332 		FREE(newspp, M_VMSWAP);
333 	}
334 
335 	/*
336 	 * priority found (or created).   now insert on the priority's
337 	 * circleq list and bump the total number of swapdevs.
338 	 */
339 	sdp->swd_priority = priority;
340 	CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
341 	uvmexp.nswapdev++;
342 }
343 
344 /*
345  * swaplist_find: find and optionally remove a swap device from the
346  *	global list.
347  *
348  * => caller must hold both swap_syscall_lock and uvm.swap_data_lock
349  * => we return the swapdev we found (and removed)
350  */
351 static struct swapdev *
352 swaplist_find(struct vnode *vp, boolean_t remove)
353 {
354 	struct swapdev *sdp;
355 	struct swappri *spp;
356 
357 	/*
358 	 * search the lists for the requested vp
359 	 */
360 
361 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
362 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
363 			if (sdp->swd_vp == vp) {
364 				if (remove) {
365 					CIRCLEQ_REMOVE(&spp->spi_swapdev,
366 					    sdp, swd_next);
367 					uvmexp.nswapdev--;
368 				}
369 				return(sdp);
370 			}
371 		}
372 	}
373 	return (NULL);
374 }
375 
376 
377 /*
378  * swaplist_trim: scan priority list for empty priority entries and kill
379  *	them.
380  *
381  * => caller must hold both swap_syscall_lock and uvm.swap_data_lock
382  */
383 static void
384 swaplist_trim(void)
385 {
386 	struct swappri *spp, *nextspp;
387 
388 	for (spp = LIST_FIRST(&swap_priority); spp != NULL; spp = nextspp) {
389 		nextspp = LIST_NEXT(spp, spi_swappri);
390 		if (CIRCLEQ_FIRST(&spp->spi_swapdev) !=
391 		    (void *)&spp->spi_swapdev)
392 			continue;
393 		LIST_REMOVE(spp, spi_swappri);
394 		free(spp, M_VMSWAP);
395 	}
396 }
397 
398 /*
399  * swapdrum_getsdp: given a page offset in /dev/drum, convert it back
400  *	to the "swapdev" that maps that section of the drum.
401  *
402  * => each swapdev takes one big contig chunk of the drum
403  * => caller must hold uvm.swap_data_lock
404  */
405 static struct swapdev *
406 swapdrum_getsdp(int pgno)
407 {
408 	struct swapdev *sdp;
409 	struct swappri *spp;
410 
411 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
412 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
413 			if (sdp->swd_flags & SWF_FAKE)
414 				continue;
415 			if (pgno >= sdp->swd_drumoffset &&
416 			    pgno < (sdp->swd_drumoffset + sdp->swd_drumsize)) {
417 				return sdp;
418 			}
419 		}
420 	}
421 	return NULL;
422 }
423 
424 
425 /*
426  * sys_swapctl: main entry point for swapctl(2) system call
427  * 	[with two helper functions: swap_on and swap_off]
428  */
429 int
430 sys_swapctl(struct lwp *l, void *v, register_t *retval)
431 {
432 	struct sys_swapctl_args /* {
433 		syscallarg(int) cmd;
434 		syscallarg(void *) arg;
435 		syscallarg(int) misc;
436 	} */ *uap = (struct sys_swapctl_args *)v;
437 	struct vnode *vp;
438 	struct nameidata nd;
439 	struct swappri *spp;
440 	struct swapdev *sdp;
441 	struct swapent *sep;
442 #define SWAP_PATH_MAX (PATH_MAX + 1)
443 	char	*userpath;
444 	size_t	len;
445 	int	error, misc;
446 	int	priority;
447 	UVMHIST_FUNC("sys_swapctl"); UVMHIST_CALLED(pdhist);
448 
449 	misc = SCARG(uap, misc);
450 
451 	/*
452 	 * ensure serialized syscall access by grabbing the swap_syscall_lock
453 	 */
454 	lockmgr(&swap_syscall_lock, LK_EXCLUSIVE, NULL);
455 
456 	userpath = malloc(SWAP_PATH_MAX, M_TEMP, M_WAITOK);
457 	/*
458 	 * we handle the non-priv NSWAP and STATS request first.
459 	 *
460 	 * SWAP_NSWAP: return number of config'd swap devices
461 	 * [can also be obtained with uvmexp sysctl]
462 	 */
463 	if (SCARG(uap, cmd) == SWAP_NSWAP) {
464 		UVMHIST_LOG(pdhist, "<- done SWAP_NSWAP=%d", uvmexp.nswapdev,
465 		    0, 0, 0);
466 		*retval = uvmexp.nswapdev;
467 		error = 0;
468 		goto out;
469 	}
470 
471 	/*
472 	 * SWAP_STATS: get stats on current # of configured swap devs
473 	 *
474 	 * note that the swap_priority list can't change as long
475 	 * as we are holding the swap_syscall_lock.  we don't want
476 	 * to grab the uvm.swap_data_lock because we may fault&sleep during
477 	 * copyout() and we don't want to be holding that lock then!
478 	 */
479 	if (SCARG(uap, cmd) == SWAP_STATS
480 #if defined(COMPAT_13)
481 	    || SCARG(uap, cmd) == SWAP_OSTATS
482 #endif
483 	    ) {
484 		if ((size_t)misc > (size_t)uvmexp.nswapdev)
485 			misc = uvmexp.nswapdev;
486 #if defined(COMPAT_13)
487 		if (SCARG(uap, cmd) == SWAP_OSTATS)
488 			len = sizeof(struct oswapent) * misc;
489 		else
490 #endif
491 			len = sizeof(struct swapent) * misc;
492 		sep = (struct swapent *)malloc(len, M_TEMP, M_WAITOK);
493 
494 		uvm_swap_stats_locked(SCARG(uap, cmd), sep, misc, retval);
495 		error = copyout(sep, SCARG(uap, arg), len);
496 
497 		free(sep, M_TEMP);
498 		UVMHIST_LOG(pdhist, "<- done SWAP_STATS", 0, 0, 0, 0);
499 		goto out;
500 	}
501 	if (SCARG(uap, cmd) == SWAP_GETDUMPDEV) {
502 		dev_t	*devp = (dev_t *)SCARG(uap, arg);
503 
504 		error = copyout(&dumpdev, devp, sizeof(dumpdev));
505 		goto out;
506 	}
507 
508 	/*
509 	 * all other requests require superuser privs.   verify.
510 	 */
511 	if ((error = kauth_authorize_system(l->l_cred, KAUTH_SYSTEM_SWAPCTL,
512 	    0, NULL, NULL, NULL)))
513 		goto out;
514 
515 	if (SCARG(uap, cmd) == SWAP_DUMPOFF) {
516 		/* drop the current dump device */
517 		dumpdev = NODEV;
518 		cpu_dumpconf();
519 		goto out;
520 	}
521 
522 	/*
523 	 * at this point we expect a path name in arg.   we will
524 	 * use namei() to gain a vnode reference (vref), and lock
525 	 * the vnode (VOP_LOCK).
526 	 *
527 	 * XXX: a NULL arg means use the root vnode pointer (e.g. for
528 	 * miniroot)
529 	 */
530 	if (SCARG(uap, arg) == NULL) {
531 		vp = rootvp;		/* miniroot */
532 		if (vget(vp, LK_EXCLUSIVE)) {
533 			error = EBUSY;
534 			goto out;
535 		}
536 		if (SCARG(uap, cmd) == SWAP_ON &&
537 		    copystr("miniroot", userpath, SWAP_PATH_MAX, &len))
538 			panic("swapctl: miniroot copy failed");
539 	} else {
540 		int	space;
541 		char	*where;
542 
543 		if (SCARG(uap, cmd) == SWAP_ON) {
544 			if ((error = copyinstr(SCARG(uap, arg), userpath,
545 			    SWAP_PATH_MAX, &len)))
546 				goto out;
547 			space = UIO_SYSSPACE;
548 			where = userpath;
549 		} else {
550 			space = UIO_USERSPACE;
551 			where = (char *)SCARG(uap, arg);
552 		}
553 		NDINIT(&nd, LOOKUP, FOLLOW|LOCKLEAF, space, where, l);
554 		if ((error = namei(&nd)))
555 			goto out;
556 		vp = nd.ni_vp;
557 	}
558 	/* note: "vp" is referenced and locked */
559 
560 	error = 0;		/* assume no error */
561 	switch(SCARG(uap, cmd)) {
562 
563 	case SWAP_DUMPDEV:
564 		if (vp->v_type != VBLK) {
565 			error = ENOTBLK;
566 			break;
567 		}
568 		dumpdev = vp->v_rdev;
569 		cpu_dumpconf();
570 		break;
571 
572 	case SWAP_CTL:
573 		/*
574 		 * get new priority, remove old entry (if any) and then
575 		 * reinsert it in the correct place.  finally, prune out
576 		 * any empty priority structures.
577 		 */
578 		priority = SCARG(uap, misc);
579 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
580 		simple_lock(&uvm.swap_data_lock);
581 		if ((sdp = swaplist_find(vp, 1)) == NULL) {
582 			error = ENOENT;
583 		} else {
584 			swaplist_insert(sdp, spp, priority);
585 			swaplist_trim();
586 		}
587 		simple_unlock(&uvm.swap_data_lock);
588 		if (error)
589 			free(spp, M_VMSWAP);
590 		break;
591 
592 	case SWAP_ON:
593 
594 		/*
595 		 * check for duplicates.   if none found, then insert a
596 		 * dummy entry on the list to prevent someone else from
597 		 * trying to enable this device while we are working on
598 		 * it.
599 		 */
600 
601 		priority = SCARG(uap, misc);
602 		sdp = malloc(sizeof *sdp, M_VMSWAP, M_WAITOK);
603 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
604 		memset(sdp, 0, sizeof(*sdp));
605 		sdp->swd_flags = SWF_FAKE;
606 		sdp->swd_vp = vp;
607 		sdp->swd_dev = (vp->v_type == VBLK) ? vp->v_rdev : NODEV;
608 		bufq_alloc(&sdp->swd_tab, "disksort", BUFQ_SORT_RAWBLOCK);
609 		simple_lock(&uvm.swap_data_lock);
610 		if (swaplist_find(vp, 0) != NULL) {
611 			error = EBUSY;
612 			simple_unlock(&uvm.swap_data_lock);
613 			bufq_free(sdp->swd_tab);
614 			free(sdp, M_VMSWAP);
615 			free(spp, M_VMSWAP);
616 			break;
617 		}
618 		swaplist_insert(sdp, spp, priority);
619 		simple_unlock(&uvm.swap_data_lock);
620 
621 		sdp->swd_pathlen = len;
622 		sdp->swd_path = malloc(sdp->swd_pathlen, M_VMSWAP, M_WAITOK);
623 		if (copystr(userpath, sdp->swd_path, sdp->swd_pathlen, 0) != 0)
624 			panic("swapctl: copystr");
625 
626 		/*
627 		 * we've now got a FAKE placeholder in the swap list.
628 		 * now attempt to enable swap on it.  if we fail, undo
629 		 * what we've done and kill the fake entry we just inserted.
630 		 * if swap_on is a success, it will clear the SWF_FAKE flag
631 		 */
632 
633 		if ((error = swap_on(l, sdp)) != 0) {
634 			simple_lock(&uvm.swap_data_lock);
635 			(void) swaplist_find(vp, 1);  /* kill fake entry */
636 			swaplist_trim();
637 			simple_unlock(&uvm.swap_data_lock);
638 			bufq_free(sdp->swd_tab);
639 			free(sdp->swd_path, M_VMSWAP);
640 			free(sdp, M_VMSWAP);
641 			break;
642 		}
643 		break;
644 
645 	case SWAP_OFF:
646 		simple_lock(&uvm.swap_data_lock);
647 		if ((sdp = swaplist_find(vp, 0)) == NULL) {
648 			simple_unlock(&uvm.swap_data_lock);
649 			error = ENXIO;
650 			break;
651 		}
652 
653 		/*
654 		 * If a device isn't in use or enabled, we
655 		 * can't stop swapping from it (again).
656 		 */
657 		if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0) {
658 			simple_unlock(&uvm.swap_data_lock);
659 			error = EBUSY;
660 			break;
661 		}
662 
663 		/*
664 		 * do the real work.
665 		 */
666 		error = swap_off(l, sdp);
667 		break;
668 
669 	default:
670 		error = EINVAL;
671 	}
672 
673 	/*
674 	 * done!  release the ref gained by namei() and unlock.
675 	 */
676 	vput(vp);
677 
678 out:
679 	free(userpath, M_TEMP);
680 	lockmgr(&swap_syscall_lock, LK_RELEASE, NULL);
681 
682 	UVMHIST_LOG(pdhist, "<- done!  error=%d", error, 0, 0, 0);
683 	return (error);
684 }
685 
686 /*
687  * swap_stats: implements swapctl(SWAP_STATS). The function is kept
688  * away from sys_swapctl() in order to allow COMPAT_* swapctl()
689  * emulation to use it directly without going through sys_swapctl().
690  * The problem with using sys_swapctl() there is that it involves
691  * copying the swapent array to the stackgap, and this array's size
692  * is not known at build time. Hence it would not be possible to
693  * ensure it would fit in the stackgap in any case.
694  */
695 void
696 uvm_swap_stats(int cmd, struct swapent *sep, int sec, register_t *retval)
697 {
698 
699 	lockmgr(&swap_syscall_lock, LK_EXCLUSIVE, NULL);
700 	uvm_swap_stats_locked(cmd, sep, sec, retval);
701 	lockmgr(&swap_syscall_lock, LK_RELEASE, NULL);
702 }
703 
704 static void
705 uvm_swap_stats_locked(int cmd, struct swapent *sep, int sec, register_t *retval)
706 {
707 	struct swappri *spp;
708 	struct swapdev *sdp;
709 	int count = 0;
710 
711 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
712 		for (sdp = CIRCLEQ_FIRST(&spp->spi_swapdev);
713 		     sdp != (void *)&spp->spi_swapdev && sec-- > 0;
714 		     sdp = CIRCLEQ_NEXT(sdp, swd_next)) {
715 		  	/*
716 			 * backwards compatibility for system call.
717 			 * note that we use 'struct oswapent' as an
718 			 * overlay into both 'struct swapdev' and
719 			 * the userland 'struct swapent', as we
720 			 * want to retain backwards compatibility
721 			 * with NetBSD 1.3.
722 			 */
723 			sdp->swd_ose.ose_inuse =
724 			    btodb((uint64_t)sdp->swd_npginuse <<
725 			    PAGE_SHIFT);
726 			(void)memcpy(sep, &sdp->swd_ose,
727 			    sizeof(struct oswapent));
728 
729 			/* now copy out the path if necessary */
730 #if !defined(COMPAT_13)
731 			(void) cmd;
732 #endif
733 #if defined(COMPAT_13)
734 			if (cmd == SWAP_STATS)
735 #endif
736 				(void)memcpy(&sep->se_path, sdp->swd_path,
737 				    sdp->swd_pathlen);
738 
739 			count++;
740 #if defined(COMPAT_13)
741 			if (cmd == SWAP_OSTATS)
742 				sep = (struct swapent *)
743 				    ((struct oswapent *)sep + 1);
744 			else
745 #endif
746 				sep++;
747 		}
748 	}
749 
750 	*retval = count;
751 	return;
752 }
753 
754 /*
755  * swap_on: attempt to enable a swapdev for swapping.   note that the
756  *	swapdev is already on the global list, but disabled (marked
757  *	SWF_FAKE).
758  *
759  * => we avoid the start of the disk (to protect disk labels)
760  * => we also avoid the miniroot, if we are swapping to root.
761  * => caller should leave uvm.swap_data_lock unlocked, we may lock it
762  *	if needed.
763  */
764 static int
765 swap_on(struct lwp *l, struct swapdev *sdp)
766 {
767 	struct vnode *vp;
768 	int error, npages, nblocks, size;
769 	long addr;
770 	u_long result;
771 	struct vattr va;
772 #ifdef NFS
773 	extern int (**nfsv2_vnodeop_p)(void *);
774 #endif /* NFS */
775 	const struct bdevsw *bdev;
776 	dev_t dev;
777 	UVMHIST_FUNC("swap_on"); UVMHIST_CALLED(pdhist);
778 
779 	/*
780 	 * we want to enable swapping on sdp.   the swd_vp contains
781 	 * the vnode we want (locked and ref'd), and the swd_dev
782 	 * contains the dev_t of the file, if it a block device.
783 	 */
784 
785 	vp = sdp->swd_vp;
786 	dev = sdp->swd_dev;
787 
788 	/*
789 	 * open the swap file (mostly useful for block device files to
790 	 * let device driver know what is up).
791 	 *
792 	 * we skip the open/close for root on swap because the root
793 	 * has already been opened when root was mounted (mountroot).
794 	 */
795 	if (vp != rootvp) {
796 		if ((error = VOP_OPEN(vp, FREAD|FWRITE, l->l_cred, l)))
797 			return (error);
798 	}
799 
800 	/* XXX this only works for block devices */
801 	UVMHIST_LOG(pdhist, "  dev=%d, major(dev)=%d", dev, major(dev), 0,0);
802 
803 	/*
804 	 * we now need to determine the size of the swap area.   for
805 	 * block specials we can call the d_psize function.
806 	 * for normal files, we must stat [get attrs].
807 	 *
808 	 * we put the result in nblks.
809 	 * for normal files, we also want the filesystem block size
810 	 * (which we get with statfs).
811 	 */
812 	switch (vp->v_type) {
813 	case VBLK:
814 		bdev = bdevsw_lookup(dev);
815 		if (bdev == NULL || bdev->d_psize == NULL ||
816 		    (nblocks = (*bdev->d_psize)(dev)) == -1) {
817 			error = ENXIO;
818 			goto bad;
819 		}
820 		break;
821 
822 	case VREG:
823 		if ((error = VOP_GETATTR(vp, &va, l->l_cred, l)))
824 			goto bad;
825 		nblocks = (int)btodb(va.va_size);
826 		if ((error =
827 		     VFS_STATVFS(vp->v_mount, &vp->v_mount->mnt_stat, l)) != 0)
828 			goto bad;
829 
830 		sdp->swd_bsize = vp->v_mount->mnt_stat.f_iosize;
831 		/*
832 		 * limit the max # of outstanding I/O requests we issue
833 		 * at any one time.   take it easy on NFS servers.
834 		 */
835 #ifdef NFS
836 		if (vp->v_op == nfsv2_vnodeop_p)
837 			sdp->swd_maxactive = 2; /* XXX */
838 		else
839 #endif /* NFS */
840 			sdp->swd_maxactive = 8; /* XXX */
841 		break;
842 
843 	default:
844 		error = ENXIO;
845 		goto bad;
846 	}
847 
848 	/*
849 	 * save nblocks in a safe place and convert to pages.
850 	 */
851 
852 	sdp->swd_ose.ose_nblks = nblocks;
853 	npages = dbtob((uint64_t)nblocks) >> PAGE_SHIFT;
854 
855 	/*
856 	 * for block special files, we want to make sure that leave
857 	 * the disklabel and bootblocks alone, so we arrange to skip
858 	 * over them (arbitrarily choosing to skip PAGE_SIZE bytes).
859 	 * note that because of this the "size" can be less than the
860 	 * actual number of blocks on the device.
861 	 */
862 	if (vp->v_type == VBLK) {
863 		/* we use pages 1 to (size - 1) [inclusive] */
864 		size = npages - 1;
865 		addr = 1;
866 	} else {
867 		/* we use pages 0 to (size - 1) [inclusive] */
868 		size = npages;
869 		addr = 0;
870 	}
871 
872 	/*
873 	 * make sure we have enough blocks for a reasonable sized swap
874 	 * area.   we want at least one page.
875 	 */
876 
877 	if (size < 1) {
878 		UVMHIST_LOG(pdhist, "  size <= 1!!", 0, 0, 0, 0);
879 		error = EINVAL;
880 		goto bad;
881 	}
882 
883 	UVMHIST_LOG(pdhist, "  dev=%x: size=%d addr=%ld\n", dev, size, addr, 0);
884 
885 	/*
886 	 * now we need to allocate an extent to manage this swap device
887 	 */
888 
889 	sdp->swd_blist = blist_create(npages);
890 	/* mark all expect the `saved' region free. */
891 	blist_free(sdp->swd_blist, addr, size);
892 
893 	/*
894 	 * if the vnode we are swapping to is the root vnode
895 	 * (i.e. we are swapping to the miniroot) then we want
896 	 * to make sure we don't overwrite it.   do a statfs to
897 	 * find its size and skip over it.
898 	 */
899 	if (vp == rootvp) {
900 		struct mount *mp;
901 		struct statvfs *sp;
902 		int rootblocks, rootpages;
903 
904 		mp = rootvnode->v_mount;
905 		sp = &mp->mnt_stat;
906 		rootblocks = sp->f_blocks * btodb(sp->f_frsize);
907 		/*
908 		 * XXX: sp->f_blocks isn't the total number of
909 		 * blocks in the filesystem, it's the number of
910 		 * data blocks.  so, our rootblocks almost
911 		 * definitely underestimates the total size
912 		 * of the filesystem - how badly depends on the
913 		 * details of the filesystem type.  there isn't
914 		 * an obvious way to deal with this cleanly
915 		 * and perfectly, so for now we just pad our
916 		 * rootblocks estimate with an extra 5 percent.
917 		 */
918 		rootblocks += (rootblocks >> 5) +
919 			(rootblocks >> 6) +
920 			(rootblocks >> 7);
921 		rootpages = round_page(dbtob(rootblocks)) >> PAGE_SHIFT;
922 		if (rootpages > size)
923 			panic("swap_on: miniroot larger than swap?");
924 
925 		if (rootpages != blist_fill(sdp->swd_blist, addr, rootpages)) {
926 			panic("swap_on: unable to preserve miniroot");
927 		}
928 
929 		size -= rootpages;
930 		printf("Preserved %d pages of miniroot ", rootpages);
931 		printf("leaving %d pages of swap\n", size);
932 	}
933 
934 	/*
935 	 * add a ref to vp to reflect usage as a swap device.
936 	 */
937 	vref(vp);
938 
939 	/*
940 	 * now add the new swapdev to the drum and enable.
941 	 */
942 	if (extent_alloc(swapmap, npages, EX_NOALIGN, EX_NOBOUNDARY,
943 	    EX_WAITOK, &result))
944 		panic("swapdrum_add");
945 
946 	sdp->swd_drumoffset = (int)result;
947 	sdp->swd_drumsize = npages;
948 	sdp->swd_npages = size;
949 	simple_lock(&uvm.swap_data_lock);
950 	sdp->swd_flags &= ~SWF_FAKE;	/* going live */
951 	sdp->swd_flags |= (SWF_INUSE|SWF_ENABLE);
952 	uvmexp.swpages += size;
953 	uvmexp.swpgavail += size;
954 	simple_unlock(&uvm.swap_data_lock);
955 	return (0);
956 
957 	/*
958 	 * failure: clean up and return error.
959 	 */
960 
961 bad:
962 	if (sdp->swd_blist) {
963 		blist_destroy(sdp->swd_blist);
964 	}
965 	if (vp != rootvp) {
966 		(void)VOP_CLOSE(vp, FREAD|FWRITE, l->l_cred, l);
967 	}
968 	return (error);
969 }
970 
971 /*
972  * swap_off: stop swapping on swapdev
973  *
974  * => swap data should be locked, we will unlock.
975  */
976 static int
977 swap_off(struct lwp *l, struct swapdev *sdp)
978 {
979 	int npages = sdp->swd_npages;
980 	int error = 0;
981 
982 	UVMHIST_FUNC("swap_off"); UVMHIST_CALLED(pdhist);
983 	UVMHIST_LOG(pdhist, "  dev=%x, npages=%d", sdp->swd_dev,npages,0,0);
984 
985 	/* disable the swap area being removed */
986 	sdp->swd_flags &= ~SWF_ENABLE;
987 	uvmexp.swpgavail -= npages;
988 	simple_unlock(&uvm.swap_data_lock);
989 
990 	/*
991 	 * the idea is to find all the pages that are paged out to this
992 	 * device, and page them all in.  in uvm, swap-backed pageable
993 	 * memory can take two forms: aobjs and anons.  call the
994 	 * swapoff hook for each subsystem to bring in pages.
995 	 */
996 
997 	if (uao_swap_off(sdp->swd_drumoffset,
998 			 sdp->swd_drumoffset + sdp->swd_drumsize) ||
999 	    amap_swap_off(sdp->swd_drumoffset,
1000 			  sdp->swd_drumoffset + sdp->swd_drumsize)) {
1001 		error = ENOMEM;
1002 	} else if (sdp->swd_npginuse > sdp->swd_npgbad) {
1003 		error = EBUSY;
1004 	}
1005 
1006 	if (error) {
1007 		simple_lock(&uvm.swap_data_lock);
1008 		sdp->swd_flags |= SWF_ENABLE;
1009 		uvmexp.swpgavail += npages;
1010 		simple_unlock(&uvm.swap_data_lock);
1011 
1012 		return error;
1013 	}
1014 
1015 	/*
1016 	 * done with the vnode.
1017 	 * drop our ref on the vnode before calling VOP_CLOSE()
1018 	 * so that spec_close() can tell if this is the last close.
1019 	 */
1020 	vrele(sdp->swd_vp);
1021 	if (sdp->swd_vp != rootvp) {
1022 		(void) VOP_CLOSE(sdp->swd_vp, FREAD|FWRITE, l->l_cred, l);
1023 	}
1024 
1025 	simple_lock(&uvm.swap_data_lock);
1026 	uvmexp.swpages -= npages;
1027 	uvmexp.swpginuse -= sdp->swd_npgbad;
1028 
1029 	if (swaplist_find(sdp->swd_vp, 1) == NULL)
1030 		panic("swap_off: swapdev not in list");
1031 	swaplist_trim();
1032 	simple_unlock(&uvm.swap_data_lock);
1033 
1034 	/*
1035 	 * free all resources!
1036 	 */
1037 	extent_free(swapmap, sdp->swd_drumoffset, sdp->swd_drumsize,
1038 		    EX_WAITOK);
1039 	blist_destroy(sdp->swd_blist);
1040 	bufq_free(sdp->swd_tab);
1041 	free(sdp, M_VMSWAP);
1042 	return (0);
1043 }
1044 
1045 /*
1046  * /dev/drum interface and i/o functions
1047  */
1048 
1049 /*
1050  * swstrategy: perform I/O on the drum
1051  *
1052  * => we must map the i/o request from the drum to the correct swapdev.
1053  */
1054 static void
1055 swstrategy(struct buf *bp)
1056 {
1057 	struct swapdev *sdp;
1058 	struct vnode *vp;
1059 	int s, pageno, bn;
1060 	UVMHIST_FUNC("swstrategy"); UVMHIST_CALLED(pdhist);
1061 
1062 	/*
1063 	 * convert block number to swapdev.   note that swapdev can't
1064 	 * be yanked out from under us because we are holding resources
1065 	 * in it (i.e. the blocks we are doing I/O on).
1066 	 */
1067 	pageno = dbtob((int64_t)bp->b_blkno) >> PAGE_SHIFT;
1068 	simple_lock(&uvm.swap_data_lock);
1069 	sdp = swapdrum_getsdp(pageno);
1070 	simple_unlock(&uvm.swap_data_lock);
1071 	if (sdp == NULL) {
1072 		bp->b_error = EINVAL;
1073 		bp->b_flags |= B_ERROR;
1074 		biodone(bp);
1075 		UVMHIST_LOG(pdhist, "  failed to get swap device", 0, 0, 0, 0);
1076 		return;
1077 	}
1078 
1079 	/*
1080 	 * convert drum page number to block number on this swapdev.
1081 	 */
1082 
1083 	pageno -= sdp->swd_drumoffset;	/* page # on swapdev */
1084 	bn = btodb((uint64_t)pageno << PAGE_SHIFT); /* convert to diskblock */
1085 
1086 	UVMHIST_LOG(pdhist, "  %s: mapoff=%x bn=%x bcount=%ld",
1087 		((bp->b_flags & B_READ) == 0) ? "write" : "read",
1088 		sdp->swd_drumoffset, bn, bp->b_bcount);
1089 
1090 	/*
1091 	 * for block devices we finish up here.
1092 	 * for regular files we have to do more work which we delegate
1093 	 * to sw_reg_strategy().
1094 	 */
1095 
1096 	switch (sdp->swd_vp->v_type) {
1097 	default:
1098 		panic("swstrategy: vnode type 0x%x", sdp->swd_vp->v_type);
1099 
1100 	case VBLK:
1101 
1102 		/*
1103 		 * must convert "bp" from an I/O on /dev/drum to an I/O
1104 		 * on the swapdev (sdp).
1105 		 */
1106 		s = splbio();
1107 		bp->b_blkno = bn;		/* swapdev block number */
1108 		vp = sdp->swd_vp;		/* swapdev vnode pointer */
1109 		bp->b_dev = sdp->swd_dev;	/* swapdev dev_t */
1110 
1111 		/*
1112 		 * if we are doing a write, we have to redirect the i/o on
1113 		 * drum's v_numoutput counter to the swapdevs.
1114 		 */
1115 		if ((bp->b_flags & B_READ) == 0) {
1116 			vwakeup(bp);	/* kills one 'v_numoutput' on drum */
1117 			V_INCR_NUMOUTPUT(vp);	/* put it on swapdev */
1118 		}
1119 
1120 		/*
1121 		 * finally plug in swapdev vnode and start I/O
1122 		 */
1123 		bp->b_vp = vp;
1124 		splx(s);
1125 		VOP_STRATEGY(vp, bp);
1126 		return;
1127 
1128 	case VREG:
1129 		/*
1130 		 * delegate to sw_reg_strategy function.
1131 		 */
1132 		sw_reg_strategy(sdp, bp, bn);
1133 		return;
1134 	}
1135 	/* NOTREACHED */
1136 }
1137 
1138 /*
1139  * swread: the read function for the drum (just a call to physio)
1140  */
1141 /*ARGSUSED*/
1142 static int
1143 swread(dev_t dev, struct uio *uio, int ioflag __unused)
1144 {
1145 	UVMHIST_FUNC("swread"); UVMHIST_CALLED(pdhist);
1146 
1147 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1148 	return (physio(swstrategy, NULL, dev, B_READ, minphys, uio));
1149 }
1150 
1151 /*
1152  * swwrite: the write function for the drum (just a call to physio)
1153  */
1154 /*ARGSUSED*/
1155 static int
1156 swwrite(dev_t dev, struct uio *uio, int ioflag __unused)
1157 {
1158 	UVMHIST_FUNC("swwrite"); UVMHIST_CALLED(pdhist);
1159 
1160 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1161 	return (physio(swstrategy, NULL, dev, B_WRITE, minphys, uio));
1162 }
1163 
1164 const struct bdevsw swap_bdevsw = {
1165 	noopen, noclose, swstrategy, noioctl, nodump, nosize, D_OTHER,
1166 };
1167 
1168 const struct cdevsw swap_cdevsw = {
1169 	nullopen, nullclose, swread, swwrite, noioctl,
1170 	nostop, notty, nopoll, nommap, nokqfilter, D_OTHER,
1171 };
1172 
1173 /*
1174  * sw_reg_strategy: handle swap i/o to regular files
1175  */
1176 static void
1177 sw_reg_strategy(struct swapdev *sdp, struct buf *bp, int bn)
1178 {
1179 	struct vnode	*vp;
1180 	struct vndxfer	*vnx;
1181 	daddr_t		nbn;
1182 	caddr_t		addr;
1183 	off_t		byteoff;
1184 	int		s, off, nra, error, sz, resid;
1185 	UVMHIST_FUNC("sw_reg_strategy"); UVMHIST_CALLED(pdhist);
1186 
1187 	/*
1188 	 * allocate a vndxfer head for this transfer and point it to
1189 	 * our buffer.
1190 	 */
1191 	getvndxfer(vnx);
1192 	vnx->vx_flags = VX_BUSY;
1193 	vnx->vx_error = 0;
1194 	vnx->vx_pending = 0;
1195 	vnx->vx_bp = bp;
1196 	vnx->vx_sdp = sdp;
1197 
1198 	/*
1199 	 * setup for main loop where we read filesystem blocks into
1200 	 * our buffer.
1201 	 */
1202 	error = 0;
1203 	bp->b_resid = bp->b_bcount;	/* nothing transfered yet! */
1204 	addr = bp->b_data;		/* current position in buffer */
1205 	byteoff = dbtob((uint64_t)bn);
1206 
1207 	for (resid = bp->b_resid; resid; resid -= sz) {
1208 		struct vndbuf	*nbp;
1209 
1210 		/*
1211 		 * translate byteoffset into block number.  return values:
1212 		 *   vp = vnode of underlying device
1213 		 *  nbn = new block number (on underlying vnode dev)
1214 		 *  nra = num blocks we can read-ahead (excludes requested
1215 		 *	block)
1216 		 */
1217 		nra = 0;
1218 		error = VOP_BMAP(sdp->swd_vp, byteoff / sdp->swd_bsize,
1219 				 	&vp, &nbn, &nra);
1220 
1221 		if (error == 0 && nbn == (daddr_t)-1) {
1222 			/*
1223 			 * this used to just set error, but that doesn't
1224 			 * do the right thing.  Instead, it causes random
1225 			 * memory errors.  The panic() should remain until
1226 			 * this condition doesn't destabilize the system.
1227 			 */
1228 #if 1
1229 			panic("sw_reg_strategy: swap to sparse file");
1230 #else
1231 			error = EIO;	/* failure */
1232 #endif
1233 		}
1234 
1235 		/*
1236 		 * punt if there was an error or a hole in the file.
1237 		 * we must wait for any i/o ops we have already started
1238 		 * to finish before returning.
1239 		 *
1240 		 * XXX we could deal with holes here but it would be
1241 		 * a hassle (in the write case).
1242 		 */
1243 		if (error) {
1244 			s = splbio();
1245 			vnx->vx_error = error;	/* pass error up */
1246 			goto out;
1247 		}
1248 
1249 		/*
1250 		 * compute the size ("sz") of this transfer (in bytes).
1251 		 */
1252 		off = byteoff % sdp->swd_bsize;
1253 		sz = (1 + nra) * sdp->swd_bsize - off;
1254 		if (sz > resid)
1255 			sz = resid;
1256 
1257 		UVMHIST_LOG(pdhist, "sw_reg_strategy: "
1258 			    "vp %p/%p offset 0x%x/0x%x",
1259 			    sdp->swd_vp, vp, byteoff, nbn);
1260 
1261 		/*
1262 		 * now get a buf structure.   note that the vb_buf is
1263 		 * at the front of the nbp structure so that you can
1264 		 * cast pointers between the two structure easily.
1265 		 */
1266 		getvndbuf(nbp);
1267 		BUF_INIT(&nbp->vb_buf);
1268 		nbp->vb_buf.b_flags    = bp->b_flags | B_CALL;
1269 		nbp->vb_buf.b_bcount   = sz;
1270 		nbp->vb_buf.b_bufsize  = sz;
1271 		nbp->vb_buf.b_error    = 0;
1272 		nbp->vb_buf.b_data     = addr;
1273 		nbp->vb_buf.b_lblkno   = 0;
1274 		nbp->vb_buf.b_blkno    = nbn + btodb(off);
1275 		nbp->vb_buf.b_rawblkno = nbp->vb_buf.b_blkno;
1276 		nbp->vb_buf.b_iodone   = sw_reg_iodone;
1277 		nbp->vb_buf.b_vp       = vp;
1278 		if (vp->v_type == VBLK) {
1279 			nbp->vb_buf.b_dev = vp->v_rdev;
1280 		}
1281 
1282 		nbp->vb_xfer = vnx;	/* patch it back in to vnx */
1283 
1284 		/*
1285 		 * Just sort by block number
1286 		 */
1287 		s = splbio();
1288 		if (vnx->vx_error != 0) {
1289 			putvndbuf(nbp);
1290 			goto out;
1291 		}
1292 		vnx->vx_pending++;
1293 
1294 		/* sort it in and start I/O if we are not over our limit */
1295 		BUFQ_PUT(sdp->swd_tab, &nbp->vb_buf);
1296 		sw_reg_start(sdp);
1297 		splx(s);
1298 
1299 		/*
1300 		 * advance to the next I/O
1301 		 */
1302 		byteoff += sz;
1303 		addr += sz;
1304 	}
1305 
1306 	s = splbio();
1307 
1308 out: /* Arrive here at splbio */
1309 	vnx->vx_flags &= ~VX_BUSY;
1310 	if (vnx->vx_pending == 0) {
1311 		if (vnx->vx_error != 0) {
1312 			bp->b_error = vnx->vx_error;
1313 			bp->b_flags |= B_ERROR;
1314 		}
1315 		putvndxfer(vnx);
1316 		biodone(bp);
1317 	}
1318 	splx(s);
1319 }
1320 
1321 /*
1322  * sw_reg_start: start an I/O request on the requested swapdev
1323  *
1324  * => reqs are sorted by b_rawblkno (above)
1325  */
1326 static void
1327 sw_reg_start(struct swapdev *sdp)
1328 {
1329 	struct buf	*bp;
1330 	UVMHIST_FUNC("sw_reg_start"); UVMHIST_CALLED(pdhist);
1331 
1332 	/* recursion control */
1333 	if ((sdp->swd_flags & SWF_BUSY) != 0)
1334 		return;
1335 
1336 	sdp->swd_flags |= SWF_BUSY;
1337 
1338 	while (sdp->swd_active < sdp->swd_maxactive) {
1339 		bp = BUFQ_GET(sdp->swd_tab);
1340 		if (bp == NULL)
1341 			break;
1342 		sdp->swd_active++;
1343 
1344 		UVMHIST_LOG(pdhist,
1345 		    "sw_reg_start:  bp %p vp %p blkno %p cnt %lx",
1346 		    bp, bp->b_vp, bp->b_blkno, bp->b_bcount);
1347 		if ((bp->b_flags & B_READ) == 0)
1348 			V_INCR_NUMOUTPUT(bp->b_vp);
1349 
1350 		VOP_STRATEGY(bp->b_vp, bp);
1351 	}
1352 	sdp->swd_flags &= ~SWF_BUSY;
1353 }
1354 
1355 /*
1356  * sw_reg_iodone: one of our i/o's has completed and needs post-i/o cleanup
1357  *
1358  * => note that we can recover the vndbuf struct by casting the buf ptr
1359  */
1360 static void
1361 sw_reg_iodone(struct buf *bp)
1362 {
1363 	struct vndbuf *vbp = (struct vndbuf *) bp;
1364 	struct vndxfer *vnx = vbp->vb_xfer;
1365 	struct buf *pbp = vnx->vx_bp;		/* parent buffer */
1366 	struct swapdev	*sdp = vnx->vx_sdp;
1367 	int s, resid, error;
1368 	UVMHIST_FUNC("sw_reg_iodone"); UVMHIST_CALLED(pdhist);
1369 
1370 	UVMHIST_LOG(pdhist, "  vbp=%p vp=%p blkno=%x addr=%p",
1371 	    vbp, vbp->vb_buf.b_vp, vbp->vb_buf.b_blkno, vbp->vb_buf.b_data);
1372 	UVMHIST_LOG(pdhist, "  cnt=%lx resid=%lx",
1373 	    vbp->vb_buf.b_bcount, vbp->vb_buf.b_resid, 0, 0);
1374 
1375 	/*
1376 	 * protect vbp at splbio and update.
1377 	 */
1378 
1379 	s = splbio();
1380 	resid = vbp->vb_buf.b_bcount - vbp->vb_buf.b_resid;
1381 	pbp->b_resid -= resid;
1382 	vnx->vx_pending--;
1383 
1384 	if (vbp->vb_buf.b_flags & B_ERROR) {
1385 		/* pass error upward */
1386 		error = vbp->vb_buf.b_error ? vbp->vb_buf.b_error : EIO;
1387 		UVMHIST_LOG(pdhist, "  got error=%d !", error, 0, 0, 0);
1388 		vnx->vx_error = error;
1389 	}
1390 
1391 	/*
1392 	 * kill vbp structure
1393 	 */
1394 	putvndbuf(vbp);
1395 
1396 	/*
1397 	 * wrap up this transaction if it has run to completion or, in
1398 	 * case of an error, when all auxiliary buffers have returned.
1399 	 */
1400 	if (vnx->vx_error != 0) {
1401 		/* pass error upward */
1402 		pbp->b_flags |= B_ERROR;
1403 		pbp->b_error = vnx->vx_error;
1404 		if ((vnx->vx_flags & VX_BUSY) == 0 && vnx->vx_pending == 0) {
1405 			putvndxfer(vnx);
1406 			biodone(pbp);
1407 		}
1408 	} else if (pbp->b_resid == 0) {
1409 		KASSERT(vnx->vx_pending == 0);
1410 		if ((vnx->vx_flags & VX_BUSY) == 0) {
1411 			UVMHIST_LOG(pdhist, "  iodone error=%d !",
1412 			    pbp, vnx->vx_error, 0, 0);
1413 			putvndxfer(vnx);
1414 			biodone(pbp);
1415 		}
1416 	}
1417 
1418 	/*
1419 	 * done!   start next swapdev I/O if one is pending
1420 	 */
1421 	sdp->swd_active--;
1422 	sw_reg_start(sdp);
1423 	splx(s);
1424 }
1425 
1426 
1427 /*
1428  * uvm_swap_alloc: allocate space on swap
1429  *
1430  * => allocation is done "round robin" down the priority list, as we
1431  *	allocate in a priority we "rotate" the circle queue.
1432  * => space can be freed with uvm_swap_free
1433  * => we return the page slot number in /dev/drum (0 == invalid slot)
1434  * => we lock uvm.swap_data_lock
1435  * => XXXMRG: "LESSOK" INTERFACE NEEDED TO EXTENT SYSTEM
1436  */
1437 int
1438 uvm_swap_alloc(int *nslots /* IN/OUT */, boolean_t lessok)
1439 {
1440 	struct swapdev *sdp;
1441 	struct swappri *spp;
1442 	UVMHIST_FUNC("uvm_swap_alloc"); UVMHIST_CALLED(pdhist);
1443 
1444 	/*
1445 	 * no swap devices configured yet?   definite failure.
1446 	 */
1447 	if (uvmexp.nswapdev < 1)
1448 		return 0;
1449 
1450 	/*
1451 	 * lock data lock, convert slots into blocks, and enter loop
1452 	 */
1453 	simple_lock(&uvm.swap_data_lock);
1454 
1455 ReTry:	/* XXXMRG */
1456 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
1457 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
1458 			uint64_t result;
1459 
1460 			/* if it's not enabled, then we can't swap from it */
1461 			if ((sdp->swd_flags & SWF_ENABLE) == 0)
1462 				continue;
1463 			if (sdp->swd_npginuse + *nslots > sdp->swd_npages)
1464 				continue;
1465 			result = blist_alloc(sdp->swd_blist, *nslots);
1466 			if (result == BLIST_NONE) {
1467 				continue;
1468 			}
1469 			KASSERT(result < sdp->swd_drumsize);
1470 
1471 			/*
1472 			 * successful allocation!  now rotate the circleq.
1473 			 */
1474 			CIRCLEQ_REMOVE(&spp->spi_swapdev, sdp, swd_next);
1475 			CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
1476 			sdp->swd_npginuse += *nslots;
1477 			uvmexp.swpginuse += *nslots;
1478 			simple_unlock(&uvm.swap_data_lock);
1479 			/* done!  return drum slot number */
1480 			UVMHIST_LOG(pdhist,
1481 			    "success!  returning %d slots starting at %d",
1482 			    *nslots, result + sdp->swd_drumoffset, 0, 0);
1483 			return (result + sdp->swd_drumoffset);
1484 		}
1485 	}
1486 
1487 	/* XXXMRG: BEGIN HACK */
1488 	if (*nslots > 1 && lessok) {
1489 		*nslots = 1;
1490 		/* XXXMRG: ugh!  blist should support this for us */
1491 		goto ReTry;
1492 	}
1493 	/* XXXMRG: END HACK */
1494 
1495 	simple_unlock(&uvm.swap_data_lock);
1496 	return 0;
1497 }
1498 
1499 boolean_t
1500 uvm_swapisfull(void)
1501 {
1502 	boolean_t rv;
1503 
1504 	simple_lock(&uvm.swap_data_lock);
1505 	KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1506 	rv = (uvmexp.swpgonly >= uvmexp.swpgavail);
1507 	simple_unlock(&uvm.swap_data_lock);
1508 
1509 	return (rv);
1510 }
1511 
1512 /*
1513  * uvm_swap_markbad: keep track of swap ranges where we've had i/o errors
1514  *
1515  * => we lock uvm.swap_data_lock
1516  */
1517 void
1518 uvm_swap_markbad(int startslot, int nslots)
1519 {
1520 	struct swapdev *sdp;
1521 	UVMHIST_FUNC("uvm_swap_markbad"); UVMHIST_CALLED(pdhist);
1522 
1523 	simple_lock(&uvm.swap_data_lock);
1524 	sdp = swapdrum_getsdp(startslot);
1525 	KASSERT(sdp != NULL);
1526 
1527 	/*
1528 	 * we just keep track of how many pages have been marked bad
1529 	 * in this device, to make everything add up in swap_off().
1530 	 * we assume here that the range of slots will all be within
1531 	 * one swap device.
1532 	 */
1533 
1534 	KASSERT(uvmexp.swpgonly >= nslots);
1535 	uvmexp.swpgonly -= nslots;
1536 	sdp->swd_npgbad += nslots;
1537 	UVMHIST_LOG(pdhist, "now %d bad", sdp->swd_npgbad, 0,0,0);
1538 	simple_unlock(&uvm.swap_data_lock);
1539 }
1540 
1541 /*
1542  * uvm_swap_free: free swap slots
1543  *
1544  * => this can be all or part of an allocation made by uvm_swap_alloc
1545  * => we lock uvm.swap_data_lock
1546  */
1547 void
1548 uvm_swap_free(int startslot, int nslots)
1549 {
1550 	struct swapdev *sdp;
1551 	UVMHIST_FUNC("uvm_swap_free"); UVMHIST_CALLED(pdhist);
1552 
1553 	UVMHIST_LOG(pdhist, "freeing %d slots starting at %d", nslots,
1554 	    startslot, 0, 0);
1555 
1556 	/*
1557 	 * ignore attempts to free the "bad" slot.
1558 	 */
1559 
1560 	if (startslot == SWSLOT_BAD) {
1561 		return;
1562 	}
1563 
1564 	/*
1565 	 * convert drum slot offset back to sdp, free the blocks
1566 	 * in the extent, and return.   must hold pri lock to do
1567 	 * lookup and access the extent.
1568 	 */
1569 
1570 	simple_lock(&uvm.swap_data_lock);
1571 	sdp = swapdrum_getsdp(startslot);
1572 	KASSERT(uvmexp.nswapdev >= 1);
1573 	KASSERT(sdp != NULL);
1574 	KASSERT(sdp->swd_npginuse >= nslots);
1575 	blist_free(sdp->swd_blist, startslot - sdp->swd_drumoffset, nslots);
1576 	sdp->swd_npginuse -= nslots;
1577 	uvmexp.swpginuse -= nslots;
1578 	simple_unlock(&uvm.swap_data_lock);
1579 }
1580 
1581 /*
1582  * uvm_swap_put: put any number of pages into a contig place on swap
1583  *
1584  * => can be sync or async
1585  */
1586 
1587 int
1588 uvm_swap_put(int swslot, struct vm_page **ppsp, int npages, int flags)
1589 {
1590 	int error;
1591 
1592 	error = uvm_swap_io(ppsp, swslot, npages, B_WRITE |
1593 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1594 	return error;
1595 }
1596 
1597 /*
1598  * uvm_swap_get: get a single page from swap
1599  *
1600  * => usually a sync op (from fault)
1601  */
1602 
1603 int
1604 uvm_swap_get(struct vm_page *page, int swslot, int flags)
1605 {
1606 	int error;
1607 
1608 	uvmexp.nswget++;
1609 	KASSERT(flags & PGO_SYNCIO);
1610 	if (swslot == SWSLOT_BAD) {
1611 		return EIO;
1612 	}
1613 
1614 	error = uvm_swap_io(&page, swslot, 1, B_READ |
1615 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1616 	if (error == 0) {
1617 
1618 		/*
1619 		 * this page is no longer only in swap.
1620 		 */
1621 
1622 		simple_lock(&uvm.swap_data_lock);
1623 		KASSERT(uvmexp.swpgonly > 0);
1624 		uvmexp.swpgonly--;
1625 		simple_unlock(&uvm.swap_data_lock);
1626 	}
1627 	return error;
1628 }
1629 
1630 /*
1631  * uvm_swap_io: do an i/o operation to swap
1632  */
1633 
1634 static int
1635 uvm_swap_io(struct vm_page **pps, int startslot, int npages, int flags)
1636 {
1637 	daddr_t startblk;
1638 	struct	buf *bp;
1639 	vaddr_t kva;
1640 	int	error, s, mapinflags;
1641 	boolean_t write, async;
1642 	UVMHIST_FUNC("uvm_swap_io"); UVMHIST_CALLED(pdhist);
1643 
1644 	UVMHIST_LOG(pdhist, "<- called, startslot=%d, npages=%d, flags=%d",
1645 	    startslot, npages, flags, 0);
1646 
1647 	write = (flags & B_READ) == 0;
1648 	async = (flags & B_ASYNC) != 0;
1649 
1650 	/*
1651 	 * convert starting drum slot to block number
1652 	 */
1653 
1654 	startblk = btodb((uint64_t)startslot << PAGE_SHIFT);
1655 
1656 	/*
1657 	 * first, map the pages into the kernel.
1658 	 */
1659 
1660 	mapinflags = !write ?
1661 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_READ :
1662 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_WRITE;
1663 	kva = uvm_pagermapin(pps, npages, mapinflags);
1664 
1665 	/*
1666 	 * now allocate a buf for the i/o.
1667 	 */
1668 
1669 	bp = getiobuf();
1670 
1671 	/*
1672 	 * fill in the bp/sbp.   we currently route our i/o through
1673 	 * /dev/drum's vnode [swapdev_vp].
1674 	 */
1675 
1676 	bp->b_flags = B_BUSY | B_NOCACHE | (flags & (B_READ|B_ASYNC));
1677 	bp->b_proc = &proc0;	/* XXX */
1678 	bp->b_vnbufs.le_next = NOLIST;
1679 	bp->b_data = (caddr_t)kva;
1680 	bp->b_blkno = startblk;
1681 	bp->b_vp = swapdev_vp;
1682 	bp->b_bufsize = bp->b_bcount = npages << PAGE_SHIFT;
1683 
1684 	/*
1685 	 * bump v_numoutput (counter of number of active outputs).
1686 	 */
1687 
1688 	if (write) {
1689 		s = splbio();
1690 		V_INCR_NUMOUTPUT(swapdev_vp);
1691 		splx(s);
1692 	}
1693 
1694 	/*
1695 	 * for async ops we must set up the iodone handler.
1696 	 */
1697 
1698 	if (async) {
1699 		bp->b_flags |= B_CALL;
1700 		bp->b_iodone = uvm_aio_biodone;
1701 		UVMHIST_LOG(pdhist, "doing async!", 0, 0, 0, 0);
1702 		if (curproc == uvm.pagedaemon_proc)
1703 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1704 		else
1705 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
1706 	} else {
1707 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1708 	}
1709 	UVMHIST_LOG(pdhist,
1710 	    "about to start io: data = %p blkno = 0x%x, bcount = %ld",
1711 	    bp->b_data, bp->b_blkno, bp->b_bcount, 0);
1712 
1713 	/*
1714 	 * now we start the I/O, and if async, return.
1715 	 */
1716 
1717 	VOP_STRATEGY(swapdev_vp, bp);
1718 	if (async)
1719 		return 0;
1720 
1721 	/*
1722 	 * must be sync i/o.   wait for it to finish
1723 	 */
1724 
1725 	error = biowait(bp);
1726 
1727 	/*
1728 	 * kill the pager mapping
1729 	 */
1730 
1731 	uvm_pagermapout(kva, npages);
1732 
1733 	/*
1734 	 * now dispose of the buf and we're done.
1735 	 */
1736 
1737 	s = splbio();
1738 	if (write)
1739 		vwakeup(bp);
1740 	putiobuf(bp);
1741 	splx(s);
1742 	UVMHIST_LOG(pdhist, "<- done (sync)  error=%d", error, 0, 0, 0);
1743 	return (error);
1744 }
1745