xref: /netbsd-src/sys/uvm/uvm_swap.c (revision bf1e9b32e27832f0c493206710fb8b58a980838a)
1 /*	$NetBSD: uvm_swap.c,v 1.94 2005/06/27 02:29:32 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 1.94 2005/06/27 02:29:32 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 
63 #include <uvm/uvm.h>
64 
65 #include <miscfs/specfs/specdev.h>
66 
67 /*
68  * uvm_swap.c: manage configuration and i/o to swap space.
69  */
70 
71 /*
72  * swap space is managed in the following way:
73  *
74  * each swap partition or file is described by a "swapdev" structure.
75  * each "swapdev" structure contains a "swapent" structure which contains
76  * information that is passed up to the user (via system calls).
77  *
78  * each swap partition is assigned a "priority" (int) which controls
79  * swap parition usage.
80  *
81  * the system maintains a global data structure describing all swap
82  * partitions/files.   there is a sorted LIST of "swappri" structures
83  * which describe "swapdev"'s at that priority.   this LIST is headed
84  * by the "swap_priority" global var.    each "swappri" contains a
85  * CIRCLEQ of "swapdev" structures at that priority.
86  *
87  * locking:
88  *  - swap_syscall_lock (sleep lock): this lock serializes the swapctl
89  *    system call and prevents the swap priority list from changing
90  *    while we are in the middle of a system call (e.g. SWAP_STATS).
91  *  - uvm.swap_data_lock (simple_lock): this lock protects all swap data
92  *    structures including the priority list, the swapdev structures,
93  *    and the swapmap extent.
94  *
95  * each swap device has the following info:
96  *  - swap device in use (could be disabled, preventing future use)
97  *  - swap enabled (allows new allocations on swap)
98  *  - map info in /dev/drum
99  *  - vnode pointer
100  * for swap files only:
101  *  - block size
102  *  - max byte count in buffer
103  *  - buffer
104  *
105  * userland controls and configures swap with the swapctl(2) system call.
106  * the sys_swapctl performs the following operations:
107  *  [1] SWAP_NSWAP: returns the number of swap devices currently configured
108  *  [2] SWAP_STATS: given a pointer to an array of swapent structures
109  *	(passed in via "arg") of a size passed in via "misc" ... we load
110  *	the current swap config into the array. The actual work is done
111  *	in the uvm_swap_stats(9) function.
112  *  [3] SWAP_ON: given a pathname in arg (could be device or file) and a
113  *	priority in "misc", start swapping on it.
114  *  [4] SWAP_OFF: as SWAP_ON, but stops swapping to a device
115  *  [5] SWAP_CTL: changes the priority of a swap device (new priority in
116  *	"misc")
117  */
118 
119 /*
120  * swapdev: describes a single swap partition/file
121  *
122  * note the following should be true:
123  * swd_inuse <= swd_nblks  [number of blocks in use is <= total blocks]
124  * swd_nblks <= swd_mapsize [because mapsize includes miniroot+disklabel]
125  */
126 struct swapdev {
127 	struct oswapent swd_ose;
128 #define	swd_dev		swd_ose.ose_dev		/* device id */
129 #define	swd_flags	swd_ose.ose_flags	/* flags:inuse/enable/fake */
130 #define	swd_priority	swd_ose.ose_priority	/* our priority */
131 	/* also: swd_ose.ose_nblks, swd_ose.ose_inuse */
132 	char			*swd_path;	/* saved pathname of device */
133 	int			swd_pathlen;	/* length of pathname */
134 	int			swd_npages;	/* #pages we can use */
135 	int			swd_npginuse;	/* #pages in use */
136 	int			swd_npgbad;	/* #pages bad */
137 	int			swd_drumoffset;	/* page0 offset in drum */
138 	int			swd_drumsize;	/* #pages in drum */
139 	blist_t			swd_blist;	/* blist for this swapdev */
140 	struct vnode		*swd_vp;	/* backing vnode */
141 	CIRCLEQ_ENTRY(swapdev)	swd_next;	/* priority circleq */
142 
143 	int			swd_bsize;	/* blocksize (bytes) */
144 	int			swd_maxactive;	/* max active i/o reqs */
145 	struct bufq_state	swd_tab;	/* buffer list */
146 	int			swd_active;	/* number of active buffers */
147 };
148 
149 /*
150  * swap device priority entry; the list is kept sorted on `spi_priority'.
151  */
152 struct swappri {
153 	int			spi_priority;     /* priority */
154 	CIRCLEQ_HEAD(spi_swapdev, swapdev)	spi_swapdev;
155 	/* circleq of swapdevs at this priority */
156 	LIST_ENTRY(swappri)	spi_swappri;      /* global list of pri's */
157 };
158 
159 /*
160  * The following two structures are used to keep track of data transfers
161  * on swap devices associated with regular files.
162  * NOTE: this code is more or less a copy of vnd.c; we use the same
163  * structure names here to ease porting..
164  */
165 struct vndxfer {
166 	struct buf	*vx_bp;		/* Pointer to parent buffer */
167 	struct swapdev	*vx_sdp;
168 	int		vx_error;
169 	int		vx_pending;	/* # of pending aux buffers */
170 	int		vx_flags;
171 #define VX_BUSY		1
172 #define VX_DEAD		2
173 };
174 
175 struct vndbuf {
176 	struct buf	vb_buf;
177 	struct vndxfer	*vb_xfer;
178 };
179 
180 
181 /*
182  * We keep a of pool vndbuf's and vndxfer structures.
183  */
184 POOL_INIT(vndxfer_pool, sizeof(struct vndxfer), 0, 0, 0, "swp vnx", NULL);
185 POOL_INIT(vndbuf_pool, sizeof(struct vndbuf), 0, 0, 0, "swp vnd", NULL);
186 
187 #define	getvndxfer(vnx)	do {						\
188 	int sp = splbio();						\
189 	vnx = pool_get(&vndxfer_pool, PR_WAITOK);			\
190 	splx(sp);							\
191 } while (/*CONSTCOND*/ 0)
192 
193 #define putvndxfer(vnx) {						\
194 	pool_put(&vndxfer_pool, (void *)(vnx));				\
195 }
196 
197 #define	getvndbuf(vbp)	do {						\
198 	int sp = splbio();						\
199 	vbp = pool_get(&vndbuf_pool, PR_WAITOK);			\
200 	splx(sp);							\
201 } while (/*CONSTCOND*/ 0)
202 
203 #define putvndbuf(vbp) {						\
204 	pool_put(&vndbuf_pool, (void *)(vbp));				\
205 }
206 
207 /*
208  * local variables
209  */
210 static struct extent *swapmap;		/* controls the mapping of /dev/drum */
211 
212 MALLOC_DEFINE(M_VMSWAP, "VM swap", "VM swap structures");
213 
214 /* list of all active swap devices [by priority] */
215 LIST_HEAD(swap_priority, swappri);
216 static struct swap_priority swap_priority;
217 
218 /* locks */
219 struct lock swap_syscall_lock;
220 
221 /*
222  * prototypes
223  */
224 static struct swapdev	*swapdrum_getsdp(int);
225 
226 static struct swapdev	*swaplist_find(struct vnode *, int);
227 static void		 swaplist_insert(struct swapdev *,
228 					 struct swappri *, int);
229 static void		 swaplist_trim(void);
230 
231 static int swap_on(struct proc *, struct swapdev *);
232 static int swap_off(struct proc *, struct swapdev *);
233 
234 static void sw_reg_strategy(struct swapdev *, struct buf *, int);
235 static void sw_reg_iodone(struct buf *);
236 static void sw_reg_start(struct swapdev *);
237 
238 static int uvm_swap_io(struct vm_page **, int, int, int);
239 
240 /*
241  * uvm_swap_init: init the swap system data structures and locks
242  *
243  * => called at boot time from init_main.c after the filesystems
244  *	are brought up (which happens after uvm_init())
245  */
246 void
247 uvm_swap_init(void)
248 {
249 	UVMHIST_FUNC("uvm_swap_init");
250 
251 	UVMHIST_CALLED(pdhist);
252 	/*
253 	 * first, init the swap list, its counter, and its lock.
254 	 * then get a handle on the vnode for /dev/drum by using
255 	 * the its dev_t number ("swapdev", from MD conf.c).
256 	 */
257 
258 	LIST_INIT(&swap_priority);
259 	uvmexp.nswapdev = 0;
260 	lockinit(&swap_syscall_lock, PVM, "swapsys", 0, 0);
261 	simple_lock_init(&uvm.swap_data_lock);
262 
263 	if (bdevvp(swapdev, &swapdev_vp))
264 		panic("uvm_swap_init: can't get vnode for swap device");
265 
266 	/*
267 	 * create swap block resource map to map /dev/drum.   the range
268 	 * from 1 to INT_MAX allows 2 gigablocks of swap space.  note
269 	 * that block 0 is reserved (used to indicate an allocation
270 	 * failure, or no allocation).
271 	 */
272 	swapmap = extent_create("swapmap", 1, INT_MAX,
273 				M_VMSWAP, 0, 0, EX_NOWAIT);
274 	if (swapmap == 0)
275 		panic("uvm_swap_init: extent_create failed");
276 
277 	/*
278 	 * done!
279 	 */
280 	UVMHIST_LOG(pdhist, "<- done", 0, 0, 0, 0);
281 }
282 
283 /*
284  * swaplist functions: functions that operate on the list of swap
285  * devices on the system.
286  */
287 
288 /*
289  * swaplist_insert: insert swap device "sdp" into the global list
290  *
291  * => caller must hold both swap_syscall_lock and uvm.swap_data_lock
292  * => caller must provide a newly malloc'd swappri structure (we will
293  *	FREE it if we don't need it... this it to prevent malloc blocking
294  *	here while adding swap)
295  */
296 static void
297 swaplist_insert(struct swapdev *sdp, struct swappri *newspp, int priority)
298 {
299 	struct swappri *spp, *pspp;
300 	UVMHIST_FUNC("swaplist_insert"); UVMHIST_CALLED(pdhist);
301 
302 	/*
303 	 * find entry at or after which to insert the new device.
304 	 */
305 	pspp = NULL;
306 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
307 		if (priority <= spp->spi_priority)
308 			break;
309 		pspp = spp;
310 	}
311 
312 	/*
313 	 * new priority?
314 	 */
315 	if (spp == NULL || spp->spi_priority != priority) {
316 		spp = newspp;  /* use newspp! */
317 		UVMHIST_LOG(pdhist, "created new swappri = %d",
318 			    priority, 0, 0, 0);
319 
320 		spp->spi_priority = priority;
321 		CIRCLEQ_INIT(&spp->spi_swapdev);
322 
323 		if (pspp)
324 			LIST_INSERT_AFTER(pspp, spp, spi_swappri);
325 		else
326 			LIST_INSERT_HEAD(&swap_priority, spp, spi_swappri);
327 	} else {
328 	  	/* we don't need a new priority structure, free it */
329 		FREE(newspp, M_VMSWAP);
330 	}
331 
332 	/*
333 	 * priority found (or created).   now insert on the priority's
334 	 * circleq list and bump the total number of swapdevs.
335 	 */
336 	sdp->swd_priority = priority;
337 	CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
338 	uvmexp.nswapdev++;
339 }
340 
341 /*
342  * swaplist_find: find and optionally remove a swap device from the
343  *	global list.
344  *
345  * => caller must hold both swap_syscall_lock and uvm.swap_data_lock
346  * => we return the swapdev we found (and removed)
347  */
348 static struct swapdev *
349 swaplist_find(struct vnode *vp, boolean_t remove)
350 {
351 	struct swapdev *sdp;
352 	struct swappri *spp;
353 
354 	/*
355 	 * search the lists for the requested vp
356 	 */
357 
358 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
359 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
360 			if (sdp->swd_vp == vp) {
361 				if (remove) {
362 					CIRCLEQ_REMOVE(&spp->spi_swapdev,
363 					    sdp, swd_next);
364 					uvmexp.nswapdev--;
365 				}
366 				return(sdp);
367 			}
368 		}
369 	}
370 	return (NULL);
371 }
372 
373 
374 /*
375  * swaplist_trim: scan priority list for empty priority entries and kill
376  *	them.
377  *
378  * => caller must hold both swap_syscall_lock and uvm.swap_data_lock
379  */
380 static void
381 swaplist_trim(void)
382 {
383 	struct swappri *spp, *nextspp;
384 
385 	for (spp = LIST_FIRST(&swap_priority); spp != NULL; spp = nextspp) {
386 		nextspp = LIST_NEXT(spp, spi_swappri);
387 		if (CIRCLEQ_FIRST(&spp->spi_swapdev) !=
388 		    (void *)&spp->spi_swapdev)
389 			continue;
390 		LIST_REMOVE(spp, spi_swappri);
391 		free(spp, M_VMSWAP);
392 	}
393 }
394 
395 /*
396  * swapdrum_getsdp: given a page offset in /dev/drum, convert it back
397  *	to the "swapdev" that maps that section of the drum.
398  *
399  * => each swapdev takes one big contig chunk of the drum
400  * => caller must hold uvm.swap_data_lock
401  */
402 static struct swapdev *
403 swapdrum_getsdp(int pgno)
404 {
405 	struct swapdev *sdp;
406 	struct swappri *spp;
407 
408 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
409 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
410 			if (sdp->swd_flags & SWF_FAKE)
411 				continue;
412 			if (pgno >= sdp->swd_drumoffset &&
413 			    pgno < (sdp->swd_drumoffset + sdp->swd_drumsize)) {
414 				return sdp;
415 			}
416 		}
417 	}
418 	return NULL;
419 }
420 
421 
422 /*
423  * sys_swapctl: main entry point for swapctl(2) system call
424  * 	[with two helper functions: swap_on and swap_off]
425  */
426 int
427 sys_swapctl(struct lwp *l, void *v, register_t *retval)
428 {
429 	struct sys_swapctl_args /* {
430 		syscallarg(int) cmd;
431 		syscallarg(void *) arg;
432 		syscallarg(int) misc;
433 	} */ *uap = (struct sys_swapctl_args *)v;
434 	struct proc *p = l->l_proc;
435 	struct vnode *vp;
436 	struct nameidata nd;
437 	struct swappri *spp;
438 	struct swapdev *sdp;
439 	struct swapent *sep;
440 	char	userpath[PATH_MAX + 1];
441 	size_t	len;
442 	int	error, misc;
443 	int	priority;
444 	UVMHIST_FUNC("sys_swapctl"); UVMHIST_CALLED(pdhist);
445 
446 	misc = SCARG(uap, misc);
447 
448 	/*
449 	 * ensure serialized syscall access by grabbing the swap_syscall_lock
450 	 */
451 	lockmgr(&swap_syscall_lock, LK_EXCLUSIVE, NULL);
452 
453 	/*
454 	 * we handle the non-priv NSWAP and STATS request first.
455 	 *
456 	 * SWAP_NSWAP: return number of config'd swap devices
457 	 * [can also be obtained with uvmexp sysctl]
458 	 */
459 	if (SCARG(uap, cmd) == SWAP_NSWAP) {
460 		UVMHIST_LOG(pdhist, "<- done SWAP_NSWAP=%d", uvmexp.nswapdev,
461 		    0, 0, 0);
462 		*retval = uvmexp.nswapdev;
463 		error = 0;
464 		goto out;
465 	}
466 
467 	/*
468 	 * SWAP_STATS: get stats on current # of configured swap devs
469 	 *
470 	 * note that the swap_priority list can't change as long
471 	 * as we are holding the swap_syscall_lock.  we don't want
472 	 * to grab the uvm.swap_data_lock because we may fault&sleep during
473 	 * copyout() and we don't want to be holding that lock then!
474 	 */
475 	if (SCARG(uap, cmd) == SWAP_STATS
476 #if defined(COMPAT_13)
477 	    || SCARG(uap, cmd) == SWAP_OSTATS
478 #endif
479 	    ) {
480 		if ((size_t)misc > (size_t)uvmexp.nswapdev)
481 			misc = uvmexp.nswapdev;
482 #if defined(COMPAT_13)
483 		if (SCARG(uap, cmd) == SWAP_OSTATS)
484 			len = sizeof(struct oswapent) * misc;
485 		else
486 #endif
487 			len = sizeof(struct swapent) * misc;
488 		sep = (struct swapent *)malloc(len, M_TEMP, M_WAITOK);
489 
490 		uvm_swap_stats(SCARG(uap, cmd), sep, misc, retval);
491 		error = copyout(sep, SCARG(uap, arg), len);
492 
493 		free(sep, M_TEMP);
494 		UVMHIST_LOG(pdhist, "<- done SWAP_STATS", 0, 0, 0, 0);
495 		goto out;
496 	}
497 	if (SCARG(uap, cmd) == SWAP_GETDUMPDEV) {
498 		dev_t	*devp = (dev_t *)SCARG(uap, arg);
499 
500 		error = copyout(&dumpdev, devp, sizeof(dumpdev));
501 		goto out;
502 	}
503 
504 	/*
505 	 * all other requests require superuser privs.   verify.
506 	 */
507 	if ((error = suser(p->p_ucred, &p->p_acflag)))
508 		goto out;
509 
510 	/*
511 	 * at this point we expect a path name in arg.   we will
512 	 * use namei() to gain a vnode reference (vref), and lock
513 	 * the vnode (VOP_LOCK).
514 	 *
515 	 * XXX: a NULL arg means use the root vnode pointer (e.g. for
516 	 * miniroot)
517 	 */
518 	if (SCARG(uap, arg) == NULL) {
519 		vp = rootvp;		/* miniroot */
520 		if (vget(vp, LK_EXCLUSIVE)) {
521 			error = EBUSY;
522 			goto out;
523 		}
524 		if (SCARG(uap, cmd) == SWAP_ON &&
525 		    copystr("miniroot", userpath, sizeof userpath, &len))
526 			panic("swapctl: miniroot copy failed");
527 	} else {
528 		int	space;
529 		char	*where;
530 
531 		if (SCARG(uap, cmd) == SWAP_ON) {
532 			if ((error = copyinstr(SCARG(uap, arg), userpath,
533 			    sizeof userpath, &len)))
534 				goto out;
535 			space = UIO_SYSSPACE;
536 			where = userpath;
537 		} else {
538 			space = UIO_USERSPACE;
539 			where = (char *)SCARG(uap, arg);
540 		}
541 		NDINIT(&nd, LOOKUP, FOLLOW|LOCKLEAF, space, where, p);
542 		if ((error = namei(&nd)))
543 			goto out;
544 		vp = nd.ni_vp;
545 	}
546 	/* note: "vp" is referenced and locked */
547 
548 	error = 0;		/* assume no error */
549 	switch(SCARG(uap, cmd)) {
550 
551 	case SWAP_DUMPDEV:
552 		if (vp->v_type != VBLK) {
553 			error = ENOTBLK;
554 			break;
555 		}
556 		dumpdev = vp->v_rdev;
557 		cpu_dumpconf();
558 		break;
559 
560 	case SWAP_CTL:
561 		/*
562 		 * get new priority, remove old entry (if any) and then
563 		 * reinsert it in the correct place.  finally, prune out
564 		 * any empty priority structures.
565 		 */
566 		priority = SCARG(uap, misc);
567 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
568 		simple_lock(&uvm.swap_data_lock);
569 		if ((sdp = swaplist_find(vp, 1)) == NULL) {
570 			error = ENOENT;
571 		} else {
572 			swaplist_insert(sdp, spp, priority);
573 			swaplist_trim();
574 		}
575 		simple_unlock(&uvm.swap_data_lock);
576 		if (error)
577 			free(spp, M_VMSWAP);
578 		break;
579 
580 	case SWAP_ON:
581 
582 		/*
583 		 * check for duplicates.   if none found, then insert a
584 		 * dummy entry on the list to prevent someone else from
585 		 * trying to enable this device while we are working on
586 		 * it.
587 		 */
588 
589 		priority = SCARG(uap, misc);
590 		sdp = malloc(sizeof *sdp, M_VMSWAP, M_WAITOK);
591 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
592 		memset(sdp, 0, sizeof(*sdp));
593 		sdp->swd_flags = SWF_FAKE;
594 		sdp->swd_vp = vp;
595 		sdp->swd_dev = (vp->v_type == VBLK) ? vp->v_rdev : NODEV;
596 		bufq_alloc(&sdp->swd_tab, BUFQ_DISKSORT|BUFQ_SORT_RAWBLOCK);
597 		simple_lock(&uvm.swap_data_lock);
598 		if (swaplist_find(vp, 0) != NULL) {
599 			error = EBUSY;
600 			simple_unlock(&uvm.swap_data_lock);
601 			bufq_free(&sdp->swd_tab);
602 			free(sdp, M_VMSWAP);
603 			free(spp, M_VMSWAP);
604 			break;
605 		}
606 		swaplist_insert(sdp, spp, priority);
607 		simple_unlock(&uvm.swap_data_lock);
608 
609 		sdp->swd_pathlen = len;
610 		sdp->swd_path = malloc(sdp->swd_pathlen, M_VMSWAP, M_WAITOK);
611 		if (copystr(userpath, sdp->swd_path, sdp->swd_pathlen, 0) != 0)
612 			panic("swapctl: copystr");
613 
614 		/*
615 		 * we've now got a FAKE placeholder in the swap list.
616 		 * now attempt to enable swap on it.  if we fail, undo
617 		 * what we've done and kill the fake entry we just inserted.
618 		 * if swap_on is a success, it will clear the SWF_FAKE flag
619 		 */
620 
621 		if ((error = swap_on(p, sdp)) != 0) {
622 			simple_lock(&uvm.swap_data_lock);
623 			(void) swaplist_find(vp, 1);  /* kill fake entry */
624 			swaplist_trim();
625 			simple_unlock(&uvm.swap_data_lock);
626 			bufq_free(&sdp->swd_tab);
627 			free(sdp->swd_path, M_VMSWAP);
628 			free(sdp, M_VMSWAP);
629 			break;
630 		}
631 		break;
632 
633 	case SWAP_OFF:
634 		simple_lock(&uvm.swap_data_lock);
635 		if ((sdp = swaplist_find(vp, 0)) == NULL) {
636 			simple_unlock(&uvm.swap_data_lock);
637 			error = ENXIO;
638 			break;
639 		}
640 
641 		/*
642 		 * If a device isn't in use or enabled, we
643 		 * can't stop swapping from it (again).
644 		 */
645 		if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0) {
646 			simple_unlock(&uvm.swap_data_lock);
647 			error = EBUSY;
648 			break;
649 		}
650 
651 		/*
652 		 * do the real work.
653 		 */
654 		error = swap_off(p, sdp);
655 		break;
656 
657 	default:
658 		error = EINVAL;
659 	}
660 
661 	/*
662 	 * done!  release the ref gained by namei() and unlock.
663 	 */
664 	vput(vp);
665 
666 out:
667 	lockmgr(&swap_syscall_lock, LK_RELEASE, NULL);
668 
669 	UVMHIST_LOG(pdhist, "<- done!  error=%d", error, 0, 0, 0);
670 	return (error);
671 }
672 
673 /*
674  * swap_stats: implements swapctl(SWAP_STATS). The function is kept
675  * away from sys_swapctl() in order to allow COMPAT_* swapctl()
676  * emulation to use it directly without going through sys_swapctl().
677  * The problem with using sys_swapctl() there is that it involves
678  * copying the swapent array to the stackgap, and this array's size
679  * is not known at build time. Hence it would not be possible to
680  * ensure it would fit in the stackgap in any case.
681  */
682 void
683 uvm_swap_stats(int cmd, struct swapent *sep, int sec, register_t *retval)
684 {
685 	struct swappri *spp;
686 	struct swapdev *sdp;
687 	int count = 0;
688 
689 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
690 		for (sdp = CIRCLEQ_FIRST(&spp->spi_swapdev);
691 		     sdp != (void *)&spp->spi_swapdev && sec-- > 0;
692 		     sdp = CIRCLEQ_NEXT(sdp, swd_next)) {
693 		  	/*
694 			 * backwards compatibility for system call.
695 			 * note that we use 'struct oswapent' as an
696 			 * overlay into both 'struct swapdev' and
697 			 * the userland 'struct swapent', as we
698 			 * want to retain backwards compatibility
699 			 * with NetBSD 1.3.
700 			 */
701 			sdp->swd_ose.ose_inuse =
702 			    btodb((u_int64_t)sdp->swd_npginuse <<
703 			    PAGE_SHIFT);
704 			(void)memcpy(sep, &sdp->swd_ose,
705 			    sizeof(struct oswapent));
706 
707 			/* now copy out the path if necessary */
708 #if defined(COMPAT_13)
709 			if (cmd == SWAP_STATS)
710 #endif
711 				(void)memcpy(&sep->se_path, sdp->swd_path,
712 				    sdp->swd_pathlen);
713 
714 			count++;
715 #if defined(COMPAT_13)
716 			if (cmd == SWAP_OSTATS)
717 				sep = (struct swapent *)
718 				    ((struct oswapent *)sep + 1);
719 			else
720 #endif
721 				sep++;
722 		}
723 	}
724 
725 	*retval = count;
726 	return;
727 }
728 
729 /*
730  * swap_on: attempt to enable a swapdev for swapping.   note that the
731  *	swapdev is already on the global list, but disabled (marked
732  *	SWF_FAKE).
733  *
734  * => we avoid the start of the disk (to protect disk labels)
735  * => we also avoid the miniroot, if we are swapping to root.
736  * => caller should leave uvm.swap_data_lock unlocked, we may lock it
737  *	if needed.
738  */
739 static int
740 swap_on(struct proc *p, struct swapdev *sdp)
741 {
742 	struct vnode *vp;
743 	int error, npages, nblocks, size;
744 	long addr;
745 	u_long result;
746 	struct vattr va;
747 #ifdef NFS
748 	extern int (**nfsv2_vnodeop_p)(void *);
749 #endif /* NFS */
750 	const struct bdevsw *bdev;
751 	dev_t dev;
752 	UVMHIST_FUNC("swap_on"); UVMHIST_CALLED(pdhist);
753 
754 	/*
755 	 * we want to enable swapping on sdp.   the swd_vp contains
756 	 * the vnode we want (locked and ref'd), and the swd_dev
757 	 * contains the dev_t of the file, if it a block device.
758 	 */
759 
760 	vp = sdp->swd_vp;
761 	dev = sdp->swd_dev;
762 
763 	/*
764 	 * open the swap file (mostly useful for block device files to
765 	 * let device driver know what is up).
766 	 *
767 	 * we skip the open/close for root on swap because the root
768 	 * has already been opened when root was mounted (mountroot).
769 	 */
770 	if (vp != rootvp) {
771 		if ((error = VOP_OPEN(vp, FREAD|FWRITE, p->p_ucred, p)))
772 			return (error);
773 	}
774 
775 	/* XXX this only works for block devices */
776 	UVMHIST_LOG(pdhist, "  dev=%d, major(dev)=%d", dev, major(dev), 0,0);
777 
778 	/*
779 	 * we now need to determine the size of the swap area.   for
780 	 * block specials we can call the d_psize function.
781 	 * for normal files, we must stat [get attrs].
782 	 *
783 	 * we put the result in nblks.
784 	 * for normal files, we also want the filesystem block size
785 	 * (which we get with statfs).
786 	 */
787 	switch (vp->v_type) {
788 	case VBLK:
789 		bdev = bdevsw_lookup(dev);
790 		if (bdev == NULL || bdev->d_psize == NULL ||
791 		    (nblocks = (*bdev->d_psize)(dev)) == -1) {
792 			error = ENXIO;
793 			goto bad;
794 		}
795 		break;
796 
797 	case VREG:
798 		if ((error = VOP_GETATTR(vp, &va, p->p_ucred, p)))
799 			goto bad;
800 		nblocks = (int)btodb(va.va_size);
801 		if ((error =
802 		     VFS_STATVFS(vp->v_mount, &vp->v_mount->mnt_stat, p)) != 0)
803 			goto bad;
804 
805 		sdp->swd_bsize = vp->v_mount->mnt_stat.f_iosize;
806 		/*
807 		 * limit the max # of outstanding I/O requests we issue
808 		 * at any one time.   take it easy on NFS servers.
809 		 */
810 #ifdef NFS
811 		if (vp->v_op == nfsv2_vnodeop_p)
812 			sdp->swd_maxactive = 2; /* XXX */
813 		else
814 #endif /* NFS */
815 			sdp->swd_maxactive = 8; /* XXX */
816 		break;
817 
818 	default:
819 		error = ENXIO;
820 		goto bad;
821 	}
822 
823 	/*
824 	 * save nblocks in a safe place and convert to pages.
825 	 */
826 
827 	sdp->swd_ose.ose_nblks = nblocks;
828 	npages = dbtob((u_int64_t)nblocks) >> PAGE_SHIFT;
829 
830 	/*
831 	 * for block special files, we want to make sure that leave
832 	 * the disklabel and bootblocks alone, so we arrange to skip
833 	 * over them (arbitrarily choosing to skip PAGE_SIZE bytes).
834 	 * note that because of this the "size" can be less than the
835 	 * actual number of blocks on the device.
836 	 */
837 	if (vp->v_type == VBLK) {
838 		/* we use pages 1 to (size - 1) [inclusive] */
839 		size = npages - 1;
840 		addr = 1;
841 	} else {
842 		/* we use pages 0 to (size - 1) [inclusive] */
843 		size = npages;
844 		addr = 0;
845 	}
846 
847 	/*
848 	 * make sure we have enough blocks for a reasonable sized swap
849 	 * area.   we want at least one page.
850 	 */
851 
852 	if (size < 1) {
853 		UVMHIST_LOG(pdhist, "  size <= 1!!", 0, 0, 0, 0);
854 		error = EINVAL;
855 		goto bad;
856 	}
857 
858 	UVMHIST_LOG(pdhist, "  dev=%x: size=%d addr=%ld\n", dev, size, addr, 0);
859 
860 	/*
861 	 * now we need to allocate an extent to manage this swap device
862 	 */
863 
864 	sdp->swd_blist = blist_create(npages);
865 	/* mark all expect the `saved' region free. */
866 	blist_free(sdp->swd_blist, addr, size);
867 
868 	/*
869 	 * if the vnode we are swapping to is the root vnode
870 	 * (i.e. we are swapping to the miniroot) then we want
871 	 * to make sure we don't overwrite it.   do a statfs to
872 	 * find its size and skip over it.
873 	 */
874 	if (vp == rootvp) {
875 		struct mount *mp;
876 		struct statvfs *sp;
877 		int rootblocks, rootpages;
878 
879 		mp = rootvnode->v_mount;
880 		sp = &mp->mnt_stat;
881 		rootblocks = sp->f_blocks * btodb(sp->f_frsize);
882 		/*
883 		 * XXX: sp->f_blocks isn't the total number of
884 		 * blocks in the filesystem, it's the number of
885 		 * data blocks.  so, our rootblocks almost
886 		 * definitely underestimates the total size
887 		 * of the filesystem - how badly depends on the
888 		 * details of the filesystem type.  there isn't
889 		 * an obvious way to deal with this cleanly
890 		 * and perfectly, so for now we just pad our
891 		 * rootblocks estimate with an extra 5 percent.
892 		 */
893 		rootblocks += (rootblocks >> 5) +
894 			(rootblocks >> 6) +
895 			(rootblocks >> 7);
896 		rootpages = round_page(dbtob(rootblocks)) >> PAGE_SHIFT;
897 		if (rootpages > size)
898 			panic("swap_on: miniroot larger than swap?");
899 
900 		if (rootpages != blist_fill(sdp->swd_blist, addr, rootpages)) {
901 			panic("swap_on: unable to preserve miniroot");
902 		}
903 
904 		size -= rootpages;
905 		printf("Preserved %d pages of miniroot ", rootpages);
906 		printf("leaving %d pages of swap\n", size);
907 	}
908 
909 	/*
910 	 * add a ref to vp to reflect usage as a swap device.
911 	 */
912 	vref(vp);
913 
914 	/*
915 	 * now add the new swapdev to the drum and enable.
916 	 */
917 	if (extent_alloc(swapmap, npages, EX_NOALIGN, EX_NOBOUNDARY,
918 	    EX_WAITOK, &result))
919 		panic("swapdrum_add");
920 
921 	sdp->swd_drumoffset = (int)result;
922 	sdp->swd_drumsize = npages;
923 	sdp->swd_npages = size;
924 	simple_lock(&uvm.swap_data_lock);
925 	sdp->swd_flags &= ~SWF_FAKE;	/* going live */
926 	sdp->swd_flags |= (SWF_INUSE|SWF_ENABLE);
927 	uvmexp.swpages += size;
928 	uvmexp.swpgavail += size;
929 	simple_unlock(&uvm.swap_data_lock);
930 	return (0);
931 
932 	/*
933 	 * failure: clean up and return error.
934 	 */
935 
936 bad:
937 	if (sdp->swd_blist) {
938 		blist_destroy(sdp->swd_blist);
939 	}
940 	if (vp != rootvp) {
941 		(void)VOP_CLOSE(vp, FREAD|FWRITE, p->p_ucred, p);
942 	}
943 	return (error);
944 }
945 
946 /*
947  * swap_off: stop swapping on swapdev
948  *
949  * => swap data should be locked, we will unlock.
950  */
951 static int
952 swap_off(struct proc *p, struct swapdev *sdp)
953 {
954 	int npages = sdp->swd_npages;
955 	int error = 0;
956 
957 	UVMHIST_FUNC("swap_off"); UVMHIST_CALLED(pdhist);
958 	UVMHIST_LOG(pdhist, "  dev=%x, npages=%d", sdp->swd_dev,npages,0,0);
959 
960 	/* disable the swap area being removed */
961 	sdp->swd_flags &= ~SWF_ENABLE;
962 	uvmexp.swpgavail -= npages;
963 	simple_unlock(&uvm.swap_data_lock);
964 
965 	/*
966 	 * the idea is to find all the pages that are paged out to this
967 	 * device, and page them all in.  in uvm, swap-backed pageable
968 	 * memory can take two forms: aobjs and anons.  call the
969 	 * swapoff hook for each subsystem to bring in pages.
970 	 */
971 
972 	if (uao_swap_off(sdp->swd_drumoffset,
973 			 sdp->swd_drumoffset + sdp->swd_drumsize) ||
974 	    amap_swap_off(sdp->swd_drumoffset,
975 			  sdp->swd_drumoffset + sdp->swd_drumsize)) {
976 		error = ENOMEM;
977 	} else if (sdp->swd_npginuse > sdp->swd_npgbad) {
978 		error = EBUSY;
979 	}
980 
981 	if (error) {
982 		simple_lock(&uvm.swap_data_lock);
983 		sdp->swd_flags |= SWF_ENABLE;
984 		uvmexp.swpgavail += npages;
985 		simple_unlock(&uvm.swap_data_lock);
986 
987 		return error;
988 	}
989 
990 	/*
991 	 * done with the vnode.
992 	 * drop our ref on the vnode before calling VOP_CLOSE()
993 	 * so that spec_close() can tell if this is the last close.
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 	simple_lock(&uvm.swap_data_lock);
1001 	uvmexp.swpages -= npages;
1002 	uvmexp.swpginuse -= sdp->swd_npgbad;
1003 
1004 	if (swaplist_find(sdp->swd_vp, 1) == NULL)
1005 		panic("swap_off: swapdev not in list");
1006 	swaplist_trim();
1007 	simple_unlock(&uvm.swap_data_lock);
1008 
1009 	/*
1010 	 * free all resources!
1011 	 */
1012 	extent_free(swapmap, sdp->swd_drumoffset, sdp->swd_drumsize,
1013 		    EX_WAITOK);
1014 	blist_destroy(sdp->swd_blist);
1015 	bufq_free(&sdp->swd_tab);
1016 	free(sdp, M_VMSWAP);
1017 	return (0);
1018 }
1019 
1020 /*
1021  * /dev/drum interface and i/o functions
1022  */
1023 
1024 /*
1025  * swstrategy: perform I/O on the drum
1026  *
1027  * => we must map the i/o request from the drum to the correct swapdev.
1028  */
1029 static void
1030 swstrategy(struct buf *bp)
1031 {
1032 	struct swapdev *sdp;
1033 	struct vnode *vp;
1034 	int s, pageno, bn;
1035 	UVMHIST_FUNC("swstrategy"); UVMHIST_CALLED(pdhist);
1036 
1037 	/*
1038 	 * convert block number to swapdev.   note that swapdev can't
1039 	 * be yanked out from under us because we are holding resources
1040 	 * in it (i.e. the blocks we are doing I/O on).
1041 	 */
1042 	pageno = dbtob((int64_t)bp->b_blkno) >> PAGE_SHIFT;
1043 	simple_lock(&uvm.swap_data_lock);
1044 	sdp = swapdrum_getsdp(pageno);
1045 	simple_unlock(&uvm.swap_data_lock);
1046 	if (sdp == NULL) {
1047 		bp->b_error = EINVAL;
1048 		bp->b_flags |= B_ERROR;
1049 		biodone(bp);
1050 		UVMHIST_LOG(pdhist, "  failed to get swap device", 0, 0, 0, 0);
1051 		return;
1052 	}
1053 
1054 	/*
1055 	 * convert drum page number to block number on this swapdev.
1056 	 */
1057 
1058 	pageno -= sdp->swd_drumoffset;	/* page # on swapdev */
1059 	bn = btodb((u_int64_t)pageno << PAGE_SHIFT); /* convert to diskblock */
1060 
1061 	UVMHIST_LOG(pdhist, "  %s: mapoff=%x bn=%x bcount=%ld",
1062 		((bp->b_flags & B_READ) == 0) ? "write" : "read",
1063 		sdp->swd_drumoffset, bn, bp->b_bcount);
1064 
1065 	/*
1066 	 * for block devices we finish up here.
1067 	 * for regular files we have to do more work which we delegate
1068 	 * to sw_reg_strategy().
1069 	 */
1070 
1071 	switch (sdp->swd_vp->v_type) {
1072 	default:
1073 		panic("swstrategy: vnode type 0x%x", sdp->swd_vp->v_type);
1074 
1075 	case VBLK:
1076 
1077 		/*
1078 		 * must convert "bp" from an I/O on /dev/drum to an I/O
1079 		 * on the swapdev (sdp).
1080 		 */
1081 		s = splbio();
1082 		bp->b_blkno = bn;		/* swapdev block number */
1083 		vp = sdp->swd_vp;		/* swapdev vnode pointer */
1084 		bp->b_dev = sdp->swd_dev;	/* swapdev dev_t */
1085 
1086 		/*
1087 		 * if we are doing a write, we have to redirect the i/o on
1088 		 * drum's v_numoutput counter to the swapdevs.
1089 		 */
1090 		if ((bp->b_flags & B_READ) == 0) {
1091 			vwakeup(bp);	/* kills one 'v_numoutput' on drum */
1092 			V_INCR_NUMOUTPUT(vp);	/* put it on swapdev */
1093 		}
1094 
1095 		/*
1096 		 * finally plug in swapdev vnode and start I/O
1097 		 */
1098 		bp->b_vp = vp;
1099 		splx(s);
1100 		VOP_STRATEGY(vp, bp);
1101 		return;
1102 
1103 	case VREG:
1104 		/*
1105 		 * delegate to sw_reg_strategy function.
1106 		 */
1107 		sw_reg_strategy(sdp, bp, bn);
1108 		return;
1109 	}
1110 	/* NOTREACHED */
1111 }
1112 
1113 /*
1114  * swread: the read function for the drum (just a call to physio)
1115  */
1116 /*ARGSUSED*/
1117 static int
1118 swread(dev_t dev, struct uio *uio, int ioflag)
1119 {
1120 	UVMHIST_FUNC("swread"); UVMHIST_CALLED(pdhist);
1121 
1122 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1123 	return (physio(swstrategy, NULL, dev, B_READ, minphys, uio));
1124 }
1125 
1126 /*
1127  * swwrite: the write function for the drum (just a call to physio)
1128  */
1129 /*ARGSUSED*/
1130 static int
1131 swwrite(dev_t dev, struct uio *uio, int ioflag)
1132 {
1133 	UVMHIST_FUNC("swwrite"); UVMHIST_CALLED(pdhist);
1134 
1135 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1136 	return (physio(swstrategy, NULL, dev, B_WRITE, minphys, uio));
1137 }
1138 
1139 const struct bdevsw swap_bdevsw = {
1140 	noopen, noclose, swstrategy, noioctl, nodump, nosize,
1141 };
1142 
1143 const struct cdevsw swap_cdevsw = {
1144 	nullopen, nullclose, swread, swwrite, noioctl,
1145 	    nostop, notty, nopoll, nommap, nokqfilter
1146 };
1147 
1148 /*
1149  * sw_reg_strategy: handle swap i/o to regular files
1150  */
1151 static void
1152 sw_reg_strategy(struct swapdev *sdp, struct buf *bp, int bn)
1153 {
1154 	struct vnode	*vp;
1155 	struct vndxfer	*vnx;
1156 	daddr_t		nbn;
1157 	caddr_t		addr;
1158 	off_t		byteoff;
1159 	int		s, off, nra, error, sz, resid;
1160 	UVMHIST_FUNC("sw_reg_strategy"); UVMHIST_CALLED(pdhist);
1161 
1162 	/*
1163 	 * allocate a vndxfer head for this transfer and point it to
1164 	 * our buffer.
1165 	 */
1166 	getvndxfer(vnx);
1167 	vnx->vx_flags = VX_BUSY;
1168 	vnx->vx_error = 0;
1169 	vnx->vx_pending = 0;
1170 	vnx->vx_bp = bp;
1171 	vnx->vx_sdp = sdp;
1172 
1173 	/*
1174 	 * setup for main loop where we read filesystem blocks into
1175 	 * our buffer.
1176 	 */
1177 	error = 0;
1178 	bp->b_resid = bp->b_bcount;	/* nothing transfered yet! */
1179 	addr = bp->b_data;		/* current position in buffer */
1180 	byteoff = dbtob((u_int64_t)bn);
1181 
1182 	for (resid = bp->b_resid; resid; resid -= sz) {
1183 		struct vndbuf	*nbp;
1184 
1185 		/*
1186 		 * translate byteoffset into block number.  return values:
1187 		 *   vp = vnode of underlying device
1188 		 *  nbn = new block number (on underlying vnode dev)
1189 		 *  nra = num blocks we can read-ahead (excludes requested
1190 		 *	block)
1191 		 */
1192 		nra = 0;
1193 		error = VOP_BMAP(sdp->swd_vp, byteoff / sdp->swd_bsize,
1194 				 	&vp, &nbn, &nra);
1195 
1196 		if (error == 0 && nbn == (daddr_t)-1) {
1197 			/*
1198 			 * this used to just set error, but that doesn't
1199 			 * do the right thing.  Instead, it causes random
1200 			 * memory errors.  The panic() should remain until
1201 			 * this condition doesn't destabilize the system.
1202 			 */
1203 #if 1
1204 			panic("sw_reg_strategy: swap to sparse file");
1205 #else
1206 			error = EIO;	/* failure */
1207 #endif
1208 		}
1209 
1210 		/*
1211 		 * punt if there was an error or a hole in the file.
1212 		 * we must wait for any i/o ops we have already started
1213 		 * to finish before returning.
1214 		 *
1215 		 * XXX we could deal with holes here but it would be
1216 		 * a hassle (in the write case).
1217 		 */
1218 		if (error) {
1219 			s = splbio();
1220 			vnx->vx_error = error;	/* pass error up */
1221 			goto out;
1222 		}
1223 
1224 		/*
1225 		 * compute the size ("sz") of this transfer (in bytes).
1226 		 */
1227 		off = byteoff % sdp->swd_bsize;
1228 		sz = (1 + nra) * sdp->swd_bsize - off;
1229 		if (sz > resid)
1230 			sz = resid;
1231 
1232 		UVMHIST_LOG(pdhist, "sw_reg_strategy: "
1233 			    "vp %p/%p offset 0x%x/0x%x",
1234 			    sdp->swd_vp, vp, byteoff, nbn);
1235 
1236 		/*
1237 		 * now get a buf structure.   note that the vb_buf is
1238 		 * at the front of the nbp structure so that you can
1239 		 * cast pointers between the two structure easily.
1240 		 */
1241 		getvndbuf(nbp);
1242 		BUF_INIT(&nbp->vb_buf);
1243 		nbp->vb_buf.b_flags    = bp->b_flags | B_CALL;
1244 		nbp->vb_buf.b_bcount   = sz;
1245 		nbp->vb_buf.b_bufsize  = sz;
1246 		nbp->vb_buf.b_error    = 0;
1247 		nbp->vb_buf.b_data     = addr;
1248 		nbp->vb_buf.b_lblkno   = 0;
1249 		nbp->vb_buf.b_blkno    = nbn + btodb(off);
1250 		nbp->vb_buf.b_rawblkno = nbp->vb_buf.b_blkno;
1251 		nbp->vb_buf.b_iodone   = sw_reg_iodone;
1252 		nbp->vb_buf.b_vp       = vp;
1253 		if (vp->v_type == VBLK) {
1254 			nbp->vb_buf.b_dev = vp->v_rdev;
1255 		}
1256 
1257 		nbp->vb_xfer = vnx;	/* patch it back in to vnx */
1258 
1259 		/*
1260 		 * Just sort by block number
1261 		 */
1262 		s = splbio();
1263 		if (vnx->vx_error != 0) {
1264 			putvndbuf(nbp);
1265 			goto out;
1266 		}
1267 		vnx->vx_pending++;
1268 
1269 		/* sort it in and start I/O if we are not over our limit */
1270 		BUFQ_PUT(&sdp->swd_tab, &nbp->vb_buf);
1271 		sw_reg_start(sdp);
1272 		splx(s);
1273 
1274 		/*
1275 		 * advance to the next I/O
1276 		 */
1277 		byteoff += sz;
1278 		addr += sz;
1279 	}
1280 
1281 	s = splbio();
1282 
1283 out: /* Arrive here at splbio */
1284 	vnx->vx_flags &= ~VX_BUSY;
1285 	if (vnx->vx_pending == 0) {
1286 		if (vnx->vx_error != 0) {
1287 			bp->b_error = vnx->vx_error;
1288 			bp->b_flags |= B_ERROR;
1289 		}
1290 		putvndxfer(vnx);
1291 		biodone(bp);
1292 	}
1293 	splx(s);
1294 }
1295 
1296 /*
1297  * sw_reg_start: start an I/O request on the requested swapdev
1298  *
1299  * => reqs are sorted by b_rawblkno (above)
1300  */
1301 static void
1302 sw_reg_start(struct swapdev *sdp)
1303 {
1304 	struct buf	*bp;
1305 	UVMHIST_FUNC("sw_reg_start"); UVMHIST_CALLED(pdhist);
1306 
1307 	/* recursion control */
1308 	if ((sdp->swd_flags & SWF_BUSY) != 0)
1309 		return;
1310 
1311 	sdp->swd_flags |= SWF_BUSY;
1312 
1313 	while (sdp->swd_active < sdp->swd_maxactive) {
1314 		bp = BUFQ_GET(&sdp->swd_tab);
1315 		if (bp == NULL)
1316 			break;
1317 		sdp->swd_active++;
1318 
1319 		UVMHIST_LOG(pdhist,
1320 		    "sw_reg_start:  bp %p vp %p blkno %p cnt %lx",
1321 		    bp, bp->b_vp, bp->b_blkno, bp->b_bcount);
1322 		if ((bp->b_flags & B_READ) == 0)
1323 			V_INCR_NUMOUTPUT(bp->b_vp);
1324 
1325 		VOP_STRATEGY(bp->b_vp, bp);
1326 	}
1327 	sdp->swd_flags &= ~SWF_BUSY;
1328 }
1329 
1330 /*
1331  * sw_reg_iodone: one of our i/o's has completed and needs post-i/o cleanup
1332  *
1333  * => note that we can recover the vndbuf struct by casting the buf ptr
1334  */
1335 static void
1336 sw_reg_iodone(struct buf *bp)
1337 {
1338 	struct vndbuf *vbp = (struct vndbuf *) bp;
1339 	struct vndxfer *vnx = vbp->vb_xfer;
1340 	struct buf *pbp = vnx->vx_bp;		/* parent buffer */
1341 	struct swapdev	*sdp = vnx->vx_sdp;
1342 	int s, resid, error;
1343 	UVMHIST_FUNC("sw_reg_iodone"); UVMHIST_CALLED(pdhist);
1344 
1345 	UVMHIST_LOG(pdhist, "  vbp=%p vp=%p blkno=%x addr=%p",
1346 	    vbp, vbp->vb_buf.b_vp, vbp->vb_buf.b_blkno, vbp->vb_buf.b_data);
1347 	UVMHIST_LOG(pdhist, "  cnt=%lx resid=%lx",
1348 	    vbp->vb_buf.b_bcount, vbp->vb_buf.b_resid, 0, 0);
1349 
1350 	/*
1351 	 * protect vbp at splbio and update.
1352 	 */
1353 
1354 	s = splbio();
1355 	resid = vbp->vb_buf.b_bcount - vbp->vb_buf.b_resid;
1356 	pbp->b_resid -= resid;
1357 	vnx->vx_pending--;
1358 
1359 	if (vbp->vb_buf.b_flags & B_ERROR) {
1360 		/* pass error upward */
1361 		error = vbp->vb_buf.b_error ? vbp->vb_buf.b_error : EIO;
1362 		UVMHIST_LOG(pdhist, "  got error=%d !", error, 0, 0, 0);
1363 		vnx->vx_error = error;
1364 	}
1365 
1366 	/*
1367 	 * kill vbp structure
1368 	 */
1369 	putvndbuf(vbp);
1370 
1371 	/*
1372 	 * wrap up this transaction if it has run to completion or, in
1373 	 * case of an error, when all auxiliary buffers have returned.
1374 	 */
1375 	if (vnx->vx_error != 0) {
1376 		/* pass error upward */
1377 		pbp->b_flags |= B_ERROR;
1378 		pbp->b_error = vnx->vx_error;
1379 		if ((vnx->vx_flags & VX_BUSY) == 0 && vnx->vx_pending == 0) {
1380 			putvndxfer(vnx);
1381 			biodone(pbp);
1382 		}
1383 	} else if (pbp->b_resid == 0) {
1384 		KASSERT(vnx->vx_pending == 0);
1385 		if ((vnx->vx_flags & VX_BUSY) == 0) {
1386 			UVMHIST_LOG(pdhist, "  iodone error=%d !",
1387 			    pbp, vnx->vx_error, 0, 0);
1388 			putvndxfer(vnx);
1389 			biodone(pbp);
1390 		}
1391 	}
1392 
1393 	/*
1394 	 * done!   start next swapdev I/O if one is pending
1395 	 */
1396 	sdp->swd_active--;
1397 	sw_reg_start(sdp);
1398 	splx(s);
1399 }
1400 
1401 
1402 /*
1403  * uvm_swap_alloc: allocate space on swap
1404  *
1405  * => allocation is done "round robin" down the priority list, as we
1406  *	allocate in a priority we "rotate" the circle queue.
1407  * => space can be freed with uvm_swap_free
1408  * => we return the page slot number in /dev/drum (0 == invalid slot)
1409  * => we lock uvm.swap_data_lock
1410  * => XXXMRG: "LESSOK" INTERFACE NEEDED TO EXTENT SYSTEM
1411  */
1412 int
1413 uvm_swap_alloc(int *nslots /* IN/OUT */, boolean_t lessok)
1414 {
1415 	struct swapdev *sdp;
1416 	struct swappri *spp;
1417 	UVMHIST_FUNC("uvm_swap_alloc"); UVMHIST_CALLED(pdhist);
1418 
1419 	/*
1420 	 * no swap devices configured yet?   definite failure.
1421 	 */
1422 	if (uvmexp.nswapdev < 1)
1423 		return 0;
1424 
1425 	/*
1426 	 * lock data lock, convert slots into blocks, and enter loop
1427 	 */
1428 	simple_lock(&uvm.swap_data_lock);
1429 
1430 ReTry:	/* XXXMRG */
1431 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
1432 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
1433 			uint64_t result;
1434 
1435 			/* if it's not enabled, then we can't swap from it */
1436 			if ((sdp->swd_flags & SWF_ENABLE) == 0)
1437 				continue;
1438 			if (sdp->swd_npginuse + *nslots > sdp->swd_npages)
1439 				continue;
1440 			result = blist_alloc(sdp->swd_blist, *nslots);
1441 			if (result == BLIST_NONE) {
1442 				continue;
1443 			}
1444 			KASSERT(result < sdp->swd_drumsize);
1445 
1446 			/*
1447 			 * successful allocation!  now rotate the circleq.
1448 			 */
1449 			CIRCLEQ_REMOVE(&spp->spi_swapdev, sdp, swd_next);
1450 			CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
1451 			sdp->swd_npginuse += *nslots;
1452 			uvmexp.swpginuse += *nslots;
1453 			simple_unlock(&uvm.swap_data_lock);
1454 			/* done!  return drum slot number */
1455 			UVMHIST_LOG(pdhist,
1456 			    "success!  returning %d slots starting at %d",
1457 			    *nslots, result + sdp->swd_drumoffset, 0, 0);
1458 			return (result + sdp->swd_drumoffset);
1459 		}
1460 	}
1461 
1462 	/* XXXMRG: BEGIN HACK */
1463 	if (*nslots > 1 && lessok) {
1464 		*nslots = 1;
1465 		/* XXXMRG: ugh!  blist should support this for us */
1466 		goto ReTry;
1467 	}
1468 	/* XXXMRG: END HACK */
1469 
1470 	simple_unlock(&uvm.swap_data_lock);
1471 	return 0;
1472 }
1473 
1474 boolean_t
1475 uvm_swapisfull(void)
1476 {
1477 	boolean_t rv;
1478 
1479 	simple_lock(&uvm.swap_data_lock);
1480 	KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1481 	rv = (uvmexp.swpgonly >= uvmexp.swpgavail);
1482 	simple_unlock(&uvm.swap_data_lock);
1483 
1484 	return (rv);
1485 }
1486 
1487 /*
1488  * uvm_swap_markbad: keep track of swap ranges where we've had i/o errors
1489  *
1490  * => we lock uvm.swap_data_lock
1491  */
1492 void
1493 uvm_swap_markbad(int startslot, int nslots)
1494 {
1495 	struct swapdev *sdp;
1496 	UVMHIST_FUNC("uvm_swap_markbad"); UVMHIST_CALLED(pdhist);
1497 
1498 	simple_lock(&uvm.swap_data_lock);
1499 	sdp = swapdrum_getsdp(startslot);
1500 	KASSERT(sdp != NULL);
1501 
1502 	/*
1503 	 * we just keep track of how many pages have been marked bad
1504 	 * in this device, to make everything add up in swap_off().
1505 	 * we assume here that the range of slots will all be within
1506 	 * one swap device.
1507 	 */
1508 
1509 	KASSERT(uvmexp.swpgonly >= nslots);
1510 	uvmexp.swpgonly -= nslots;
1511 	sdp->swd_npgbad += nslots;
1512 	UVMHIST_LOG(pdhist, "now %d bad", sdp->swd_npgbad, 0,0,0);
1513 	simple_unlock(&uvm.swap_data_lock);
1514 }
1515 
1516 /*
1517  * uvm_swap_free: free swap slots
1518  *
1519  * => this can be all or part of an allocation made by uvm_swap_alloc
1520  * => we lock uvm.swap_data_lock
1521  */
1522 void
1523 uvm_swap_free(int startslot, int nslots)
1524 {
1525 	struct swapdev *sdp;
1526 	UVMHIST_FUNC("uvm_swap_free"); UVMHIST_CALLED(pdhist);
1527 
1528 	UVMHIST_LOG(pdhist, "freeing %d slots starting at %d", nslots,
1529 	    startslot, 0, 0);
1530 
1531 	/*
1532 	 * ignore attempts to free the "bad" slot.
1533 	 */
1534 
1535 	if (startslot == SWSLOT_BAD) {
1536 		return;
1537 	}
1538 
1539 	/*
1540 	 * convert drum slot offset back to sdp, free the blocks
1541 	 * in the extent, and return.   must hold pri lock to do
1542 	 * lookup and access the extent.
1543 	 */
1544 
1545 	simple_lock(&uvm.swap_data_lock);
1546 	sdp = swapdrum_getsdp(startslot);
1547 	KASSERT(uvmexp.nswapdev >= 1);
1548 	KASSERT(sdp != NULL);
1549 	KASSERT(sdp->swd_npginuse >= nslots);
1550 	blist_free(sdp->swd_blist, startslot - sdp->swd_drumoffset, nslots);
1551 	sdp->swd_npginuse -= nslots;
1552 	uvmexp.swpginuse -= nslots;
1553 	simple_unlock(&uvm.swap_data_lock);
1554 }
1555 
1556 /*
1557  * uvm_swap_put: put any number of pages into a contig place on swap
1558  *
1559  * => can be sync or async
1560  */
1561 
1562 int
1563 uvm_swap_put(int swslot, struct vm_page **ppsp, int npages, int flags)
1564 {
1565 	int error;
1566 
1567 	error = uvm_swap_io(ppsp, swslot, npages, B_WRITE |
1568 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1569 	return error;
1570 }
1571 
1572 /*
1573  * uvm_swap_get: get a single page from swap
1574  *
1575  * => usually a sync op (from fault)
1576  */
1577 
1578 int
1579 uvm_swap_get(struct vm_page *page, int swslot, int flags)
1580 {
1581 	int error;
1582 
1583 	uvmexp.nswget++;
1584 	KASSERT(flags & PGO_SYNCIO);
1585 	if (swslot == SWSLOT_BAD) {
1586 		return EIO;
1587 	}
1588 
1589 	error = uvm_swap_io(&page, swslot, 1, B_READ |
1590 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1591 	if (error == 0) {
1592 
1593 		/*
1594 		 * this page is no longer only in swap.
1595 		 */
1596 
1597 		simple_lock(&uvm.swap_data_lock);
1598 		KASSERT(uvmexp.swpgonly > 0);
1599 		uvmexp.swpgonly--;
1600 		simple_unlock(&uvm.swap_data_lock);
1601 	}
1602 	return error;
1603 }
1604 
1605 /*
1606  * uvm_swap_io: do an i/o operation to swap
1607  */
1608 
1609 static int
1610 uvm_swap_io(struct vm_page **pps, int startslot, int npages, int flags)
1611 {
1612 	daddr_t startblk;
1613 	struct	buf *bp;
1614 	vaddr_t kva;
1615 	int	error, s, mapinflags;
1616 	boolean_t write, async;
1617 	UVMHIST_FUNC("uvm_swap_io"); UVMHIST_CALLED(pdhist);
1618 
1619 	UVMHIST_LOG(pdhist, "<- called, startslot=%d, npages=%d, flags=%d",
1620 	    startslot, npages, flags, 0);
1621 
1622 	write = (flags & B_READ) == 0;
1623 	async = (flags & B_ASYNC) != 0;
1624 
1625 	/*
1626 	 * convert starting drum slot to block number
1627 	 */
1628 
1629 	startblk = btodb((u_int64_t)startslot << PAGE_SHIFT);
1630 
1631 	/*
1632 	 * first, map the pages into the kernel.
1633 	 */
1634 
1635 	mapinflags = !write ?
1636 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_READ :
1637 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_WRITE;
1638 	kva = uvm_pagermapin(pps, npages, mapinflags);
1639 
1640 	/*
1641 	 * now allocate a buf for the i/o.
1642 	 */
1643 
1644 	s = splbio();
1645 	bp = pool_get(&bufpool, PR_WAITOK);
1646 	splx(s);
1647 
1648 	/*
1649 	 * fill in the bp/sbp.   we currently route our i/o through
1650 	 * /dev/drum's vnode [swapdev_vp].
1651 	 */
1652 
1653 	BUF_INIT(bp);
1654 	bp->b_flags = B_BUSY | B_NOCACHE | (flags & (B_READ|B_ASYNC));
1655 	bp->b_proc = &proc0;	/* XXX */
1656 	bp->b_vnbufs.le_next = NOLIST;
1657 	bp->b_data = (caddr_t)kva;
1658 	bp->b_blkno = startblk;
1659 	bp->b_vp = swapdev_vp;
1660 	bp->b_bufsize = bp->b_bcount = npages << PAGE_SHIFT;
1661 
1662 	/*
1663 	 * bump v_numoutput (counter of number of active outputs).
1664 	 */
1665 
1666 	if (write) {
1667 		s = splbio();
1668 		V_INCR_NUMOUTPUT(swapdev_vp);
1669 		splx(s);
1670 	}
1671 
1672 	/*
1673 	 * for async ops we must set up the iodone handler.
1674 	 */
1675 
1676 	if (async) {
1677 		bp->b_flags |= B_CALL;
1678 		bp->b_iodone = uvm_aio_biodone;
1679 		UVMHIST_LOG(pdhist, "doing async!", 0, 0, 0, 0);
1680 		if (curproc == uvm.pagedaemon_proc)
1681 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1682 		else
1683 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
1684 	} else {
1685 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
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(swapdev_vp, 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