xref: /netbsd-src/sys/uvm/uvm_swap.c (revision 4e6df137e8e14049b5a701d249962c480449c141)
1 /*	$NetBSD: uvm_swap.c,v 1.150 2010/03/02 21:32:29 pooka Exp $	*/
2 
3 /*
4  * Copyright (c) 1995, 1996, 1997, 2009 Matthew R. Green
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
21  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
22  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
23  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
24  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  * from: NetBSD: vm_swap.c,v 1.52 1997/12/02 13:47:37 pk Exp
29  * from: Id: uvm_swap.c,v 1.1.2.42 1998/02/02 20:38:06 chuck Exp
30  */
31 
32 #include <sys/cdefs.h>
33 __KERNEL_RCSID(0, "$NetBSD: uvm_swap.c,v 1.150 2010/03/02 21:32:29 pooka Exp $");
34 
35 #include "opt_uvmhist.h"
36 #include "opt_compat_netbsd.h"
37 #include "opt_ddb.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/buf.h>
42 #include <sys/bufq.h>
43 #include <sys/conf.h>
44 #include <sys/proc.h>
45 #include <sys/namei.h>
46 #include <sys/disklabel.h>
47 #include <sys/errno.h>
48 #include <sys/kernel.h>
49 #include <sys/malloc.h>
50 #include <sys/vnode.h>
51 #include <sys/file.h>
52 #include <sys/vmem.h>
53 #include <sys/blist.h>
54 #include <sys/mount.h>
55 #include <sys/pool.h>
56 #include <sys/syscallargs.h>
57 #include <sys/swap.h>
58 #include <sys/kauth.h>
59 #include <sys/sysctl.h>
60 #include <sys/workqueue.h>
61 
62 #include <uvm/uvm.h>
63 
64 #include <miscfs/specfs/specdev.h>
65 
66 /*
67  * uvm_swap.c: manage configuration and i/o to swap space.
68  */
69 
70 /*
71  * swap space is managed in the following way:
72  *
73  * each swap partition or file is described by a "swapdev" structure.
74  * each "swapdev" structure contains a "swapent" structure which contains
75  * information that is passed up to the user (via system calls).
76  *
77  * each swap partition is assigned a "priority" (int) which controls
78  * swap parition usage.
79  *
80  * the system maintains a global data structure describing all swap
81  * partitions/files.   there is a sorted LIST of "swappri" structures
82  * which describe "swapdev"'s at that priority.   this LIST is headed
83  * by the "swap_priority" global var.    each "swappri" contains a
84  * CIRCLEQ of "swapdev" structures at that priority.
85  *
86  * locking:
87  *  - swap_syscall_lock (krwlock_t): this lock serializes the swapctl
88  *    system call and prevents the swap priority list from changing
89  *    while we are in the middle of a system call (e.g. SWAP_STATS).
90  *  - uvm_swap_data_lock (kmutex_t): this lock protects all swap data
91  *    structures including the priority list, the swapdev structures,
92  *    and the swapmap arena.
93  *
94  * each swap device has the following info:
95  *  - swap device in use (could be disabled, preventing future use)
96  *  - swap enabled (allows new allocations on swap)
97  *  - map info in /dev/drum
98  *  - vnode pointer
99  * for swap files only:
100  *  - block size
101  *  - max byte count in buffer
102  *  - buffer
103  *
104  * userland controls and configures swap with the swapctl(2) system call.
105  * the sys_swapctl performs the following operations:
106  *  [1] SWAP_NSWAP: returns the number of swap devices currently configured
107  *  [2] SWAP_STATS: given a pointer to an array of swapent structures
108  *	(passed in via "arg") of a size passed in via "misc" ... we load
109  *	the current swap config into the array. The actual work is done
110  *	in the uvm_swap_stats(9) function.
111  *  [3] SWAP_ON: given a pathname in arg (could be device or file) and a
112  *	priority in "misc", start swapping on it.
113  *  [4] SWAP_OFF: as SWAP_ON, but stops swapping to a device
114  *  [5] SWAP_CTL: changes the priority of a swap device (new priority in
115  *	"misc")
116  */
117 
118 /*
119  * swapdev: describes a single swap partition/file
120  *
121  * note the following should be true:
122  * swd_inuse <= swd_nblks  [number of blocks in use is <= total blocks]
123  * swd_nblks <= swd_mapsize [because mapsize includes miniroot+disklabel]
124  */
125 struct swapdev {
126 	dev_t			swd_dev;	/* device id */
127 	int			swd_flags;	/* flags:inuse/enable/fake */
128 	int			swd_priority;	/* our priority */
129 	int			swd_nblks;	/* blocks in this device */
130 	char			*swd_path;	/* saved pathname of device */
131 	int			swd_pathlen;	/* length of pathname */
132 	int			swd_npages;	/* #pages we can use */
133 	int			swd_npginuse;	/* #pages in use */
134 	int			swd_npgbad;	/* #pages bad */
135 	int			swd_drumoffset;	/* page0 offset in drum */
136 	int			swd_drumsize;	/* #pages in drum */
137 	blist_t			swd_blist;	/* blist for this swapdev */
138 	struct vnode		*swd_vp;	/* backing vnode */
139 	CIRCLEQ_ENTRY(swapdev)	swd_next;	/* priority circleq */
140 
141 	int			swd_bsize;	/* blocksize (bytes) */
142 	int			swd_maxactive;	/* max active i/o reqs */
143 	struct bufq_state	*swd_tab;	/* buffer list */
144 	int			swd_active;	/* number of active buffers */
145 };
146 
147 /*
148  * swap device priority entry; the list is kept sorted on `spi_priority'.
149  */
150 struct swappri {
151 	int			spi_priority;     /* priority */
152 	CIRCLEQ_HEAD(spi_swapdev, swapdev)	spi_swapdev;
153 	/* circleq of swapdevs at this priority */
154 	LIST_ENTRY(swappri)	spi_swappri;      /* global list of pri's */
155 };
156 
157 /*
158  * The following two structures are used to keep track of data transfers
159  * on swap devices associated with regular files.
160  * NOTE: this code is more or less a copy of vnd.c; we use the same
161  * structure names here to ease porting..
162  */
163 struct vndxfer {
164 	struct buf	*vx_bp;		/* Pointer to parent buffer */
165 	struct swapdev	*vx_sdp;
166 	int		vx_error;
167 	int		vx_pending;	/* # of pending aux buffers */
168 	int		vx_flags;
169 #define VX_BUSY		1
170 #define VX_DEAD		2
171 };
172 
173 struct vndbuf {
174 	struct buf	vb_buf;
175 	struct vndxfer	*vb_xfer;
176 };
177 
178 /*
179  * NetBSD 1.3 swapctl(SWAP_STATS, ...) swapent structure; uses 32 bit
180  * dev_t and has no se_path[] member.
181  */
182 struct swapent13 {
183 	int32_t	se13_dev;		/* device id */
184 	int	se13_flags;		/* flags */
185 	int	se13_nblks;		/* total blocks */
186 	int	se13_inuse;		/* blocks in use */
187 	int	se13_priority;		/* priority of this device */
188 };
189 
190 /*
191  * NetBSD 5.0 swapctl(SWAP_STATS, ...) swapent structure; uses 32 bit
192  * dev_t.
193  */
194 struct swapent50 {
195 	int32_t	se50_dev;		/* device id */
196 	int	se50_flags;		/* flags */
197 	int	se50_nblks;		/* total blocks */
198 	int	se50_inuse;		/* blocks in use */
199 	int	se50_priority;		/* priority of this device */
200 	char	se50_path[PATH_MAX+1];	/* path name */
201 };
202 
203 /*
204  * We keep a of pool vndbuf's and vndxfer structures.
205  */
206 static struct pool vndxfer_pool, vndbuf_pool;
207 
208 /*
209  * local variables
210  */
211 MALLOC_DEFINE(M_VMSWAP, "VM swap", "VM swap structures");
212 static vmem_t *swapmap;	/* controls the mapping of /dev/drum */
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 static krwlock_t swap_syscall_lock;
220 
221 /* workqueue and use counter for swap to regular files */
222 static int sw_reg_count = 0;
223 static struct workqueue *sw_reg_workqueue;
224 
225 /* tuneables */
226 u_int uvm_swapisfull_factor = 99;
227 
228 /*
229  * prototypes
230  */
231 static struct swapdev	*swapdrum_getsdp(int);
232 
233 static struct swapdev	*swaplist_find(struct vnode *, bool);
234 static void		 swaplist_insert(struct swapdev *,
235 					 struct swappri *, int);
236 static void		 swaplist_trim(void);
237 
238 static int swap_on(struct lwp *, struct swapdev *);
239 static int swap_off(struct lwp *, struct swapdev *);
240 
241 static void uvm_swap_stats_locked(int, struct swapent *, int, register_t *);
242 
243 static void sw_reg_strategy(struct swapdev *, struct buf *, int);
244 static void sw_reg_biodone(struct buf *);
245 static void sw_reg_iodone(struct work *wk, void *dummy);
246 static void sw_reg_start(struct swapdev *);
247 
248 static int uvm_swap_io(struct vm_page **, int, int, int);
249 
250 /*
251  * uvm_swap_init: init the swap system data structures and locks
252  *
253  * => called at boot time from init_main.c after the filesystems
254  *	are brought up (which happens after uvm_init())
255  */
256 void
257 uvm_swap_init(void)
258 {
259 	UVMHIST_FUNC("uvm_swap_init");
260 
261 	UVMHIST_CALLED(pdhist);
262 	/*
263 	 * first, init the swap list, its counter, and its lock.
264 	 * then get a handle on the vnode for /dev/drum by using
265 	 * the its dev_t number ("swapdev", from MD conf.c).
266 	 */
267 
268 	LIST_INIT(&swap_priority);
269 	uvmexp.nswapdev = 0;
270 	rw_init(&swap_syscall_lock);
271 	mutex_init(&uvm_swap_data_lock, MUTEX_DEFAULT, IPL_NONE);
272 
273 	if (bdevvp(swapdev, &swapdev_vp))
274 		panic("%s: can't get vnode for swap device", __func__);
275 	if (vn_lock(swapdev_vp, LK_EXCLUSIVE | LK_RETRY))
276 		panic("%s: can't lock swap device", __func__);
277 	if (VOP_OPEN(swapdev_vp, FREAD | FWRITE, NOCRED))
278 		panic("%s: can't open swap device", __func__);
279 	VOP_UNLOCK(swapdev_vp, 0);
280 
281 	/*
282 	 * create swap block resource map to map /dev/drum.   the range
283 	 * from 1 to INT_MAX allows 2 gigablocks of swap space.  note
284 	 * that block 0 is reserved (used to indicate an allocation
285 	 * failure, or no allocation).
286 	 */
287 	swapmap = vmem_create("swapmap", 1, INT_MAX - 1, 1, NULL, NULL, NULL, 0,
288 	    VM_NOSLEEP, IPL_NONE);
289 	if (swapmap == 0) {
290 		panic("%s: vmem_create failed", __func__);
291 	}
292 
293 	pool_init(&vndxfer_pool, sizeof(struct vndxfer), 0, 0, 0, "swp vnx",
294 	    NULL, IPL_BIO);
295 	pool_init(&vndbuf_pool, sizeof(struct vndbuf), 0, 0, 0, "swp vnd",
296 	    NULL, IPL_BIO);
297 
298 	UVMHIST_LOG(pdhist, "<- done", 0, 0, 0, 0);
299 }
300 
301 /*
302  * swaplist functions: functions that operate on the list of swap
303  * devices on the system.
304  */
305 
306 /*
307  * swaplist_insert: insert swap device "sdp" into the global list
308  *
309  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
310  * => caller must provide a newly malloc'd swappri structure (we will
311  *	FREE it if we don't need it... this it to prevent malloc blocking
312  *	here while adding swap)
313  */
314 static void
315 swaplist_insert(struct swapdev *sdp, struct swappri *newspp, int priority)
316 {
317 	struct swappri *spp, *pspp;
318 	UVMHIST_FUNC("swaplist_insert"); UVMHIST_CALLED(pdhist);
319 
320 	/*
321 	 * find entry at or after which to insert the new device.
322 	 */
323 	pspp = NULL;
324 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
325 		if (priority <= spp->spi_priority)
326 			break;
327 		pspp = spp;
328 	}
329 
330 	/*
331 	 * new priority?
332 	 */
333 	if (spp == NULL || spp->spi_priority != priority) {
334 		spp = newspp;  /* use newspp! */
335 		UVMHIST_LOG(pdhist, "created new swappri = %d",
336 			    priority, 0, 0, 0);
337 
338 		spp->spi_priority = priority;
339 		CIRCLEQ_INIT(&spp->spi_swapdev);
340 
341 		if (pspp)
342 			LIST_INSERT_AFTER(pspp, spp, spi_swappri);
343 		else
344 			LIST_INSERT_HEAD(&swap_priority, spp, spi_swappri);
345 	} else {
346 	  	/* we don't need a new priority structure, free it */
347 		free(newspp, M_VMSWAP);
348 	}
349 
350 	/*
351 	 * priority found (or created).   now insert on the priority's
352 	 * circleq list and bump the total number of swapdevs.
353 	 */
354 	sdp->swd_priority = priority;
355 	CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
356 	uvmexp.nswapdev++;
357 }
358 
359 /*
360  * swaplist_find: find and optionally remove a swap device from the
361  *	global list.
362  *
363  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
364  * => we return the swapdev we found (and removed)
365  */
366 static struct swapdev *
367 swaplist_find(struct vnode *vp, bool remove)
368 {
369 	struct swapdev *sdp;
370 	struct swappri *spp;
371 
372 	/*
373 	 * search the lists for the requested vp
374 	 */
375 
376 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
377 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
378 			if (sdp->swd_vp == vp) {
379 				if (remove) {
380 					CIRCLEQ_REMOVE(&spp->spi_swapdev,
381 					    sdp, swd_next);
382 					uvmexp.nswapdev--;
383 				}
384 				return(sdp);
385 			}
386 		}
387 	}
388 	return (NULL);
389 }
390 
391 /*
392  * swaplist_trim: scan priority list for empty priority entries and kill
393  *	them.
394  *
395  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
396  */
397 static void
398 swaplist_trim(void)
399 {
400 	struct swappri *spp, *nextspp;
401 
402 	for (spp = LIST_FIRST(&swap_priority); spp != NULL; spp = nextspp) {
403 		nextspp = LIST_NEXT(spp, spi_swappri);
404 		if (CIRCLEQ_FIRST(&spp->spi_swapdev) !=
405 		    (void *)&spp->spi_swapdev)
406 			continue;
407 		LIST_REMOVE(spp, spi_swappri);
408 		free(spp, M_VMSWAP);
409 	}
410 }
411 
412 /*
413  * swapdrum_getsdp: given a page offset in /dev/drum, convert it back
414  *	to the "swapdev" that maps that section of the drum.
415  *
416  * => each swapdev takes one big contig chunk of the drum
417  * => caller must hold uvm_swap_data_lock
418  */
419 static struct swapdev *
420 swapdrum_getsdp(int pgno)
421 {
422 	struct swapdev *sdp;
423 	struct swappri *spp;
424 
425 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
426 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
427 			if (sdp->swd_flags & SWF_FAKE)
428 				continue;
429 			if (pgno >= sdp->swd_drumoffset &&
430 			    pgno < (sdp->swd_drumoffset + sdp->swd_drumsize)) {
431 				return sdp;
432 			}
433 		}
434 	}
435 	return NULL;
436 }
437 
438 
439 /*
440  * sys_swapctl: main entry point for swapctl(2) system call
441  * 	[with two helper functions: swap_on and swap_off]
442  */
443 int
444 sys_swapctl(struct lwp *l, const struct sys_swapctl_args *uap, register_t *retval)
445 {
446 	/* {
447 		syscallarg(int) cmd;
448 		syscallarg(void *) arg;
449 		syscallarg(int) misc;
450 	} */
451 	struct vnode *vp;
452 	struct nameidata nd;
453 	struct swappri *spp;
454 	struct swapdev *sdp;
455 	struct swapent *sep;
456 #define SWAP_PATH_MAX (PATH_MAX + 1)
457 	char	*userpath;
458 	size_t	len;
459 	int	error, misc;
460 	int	priority;
461 	UVMHIST_FUNC("sys_swapctl"); UVMHIST_CALLED(pdhist);
462 
463 	misc = SCARG(uap, misc);
464 
465 	/*
466 	 * ensure serialized syscall access by grabbing the swap_syscall_lock
467 	 */
468 	rw_enter(&swap_syscall_lock, RW_WRITER);
469 
470 	userpath = malloc(SWAP_PATH_MAX, M_TEMP, M_WAITOK);
471 	/*
472 	 * we handle the non-priv NSWAP and STATS request first.
473 	 *
474 	 * SWAP_NSWAP: return number of config'd swap devices
475 	 * [can also be obtained with uvmexp sysctl]
476 	 */
477 	if (SCARG(uap, cmd) == SWAP_NSWAP) {
478 		UVMHIST_LOG(pdhist, "<- done SWAP_NSWAP=%d", uvmexp.nswapdev,
479 		    0, 0, 0);
480 		*retval = uvmexp.nswapdev;
481 		error = 0;
482 		goto out;
483 	}
484 
485 	/*
486 	 * SWAP_STATS: get stats on current # of configured swap devs
487 	 *
488 	 * note that the swap_priority list can't change as long
489 	 * as we are holding the swap_syscall_lock.  we don't want
490 	 * to grab the uvm_swap_data_lock because we may fault&sleep during
491 	 * copyout() and we don't want to be holding that lock then!
492 	 */
493 	if (SCARG(uap, cmd) == SWAP_STATS
494 #if defined(COMPAT_50)
495 	    || SCARG(uap, cmd) == SWAP_STATS50
496 #endif
497 #if defined(COMPAT_13)
498 	    || SCARG(uap, cmd) == SWAP_STATS13
499 #endif
500 	    ) {
501 		if ((size_t)misc > (size_t)uvmexp.nswapdev)
502 			misc = uvmexp.nswapdev;
503 #if defined(COMPAT_13)
504 		if (SCARG(uap, cmd) == SWAP_STATS13)
505 			len = sizeof(struct swapent13) * misc;
506 		else
507 #endif
508 #if defined(COMPAT_50)
509 		if (SCARG(uap, cmd) == SWAP_STATS50)
510 			len = sizeof(struct swapent50) * misc;
511 		else
512 #endif
513 			len = sizeof(struct swapent) * misc;
514 		sep = (struct swapent *)malloc(len, M_TEMP, M_WAITOK);
515 
516 		uvm_swap_stats_locked(SCARG(uap, cmd), sep, misc, retval);
517 		error = copyout(sep, SCARG(uap, arg), len);
518 
519 		free(sep, M_TEMP);
520 		UVMHIST_LOG(pdhist, "<- done SWAP_STATS", 0, 0, 0, 0);
521 		goto out;
522 	}
523 	if (SCARG(uap, cmd) == SWAP_GETDUMPDEV) {
524 		dev_t	*devp = (dev_t *)SCARG(uap, arg);
525 
526 		error = copyout(&dumpdev, devp, sizeof(dumpdev));
527 		goto out;
528 	}
529 
530 	/*
531 	 * all other requests require superuser privs.   verify.
532 	 */
533 	if ((error = kauth_authorize_system(l->l_cred, KAUTH_SYSTEM_SWAPCTL,
534 	    0, NULL, NULL, NULL)))
535 		goto out;
536 
537 	if (SCARG(uap, cmd) == SWAP_DUMPOFF) {
538 		/* drop the current dump device */
539 		dumpdev = NODEV;
540 		dumpcdev = NODEV;
541 		cpu_dumpconf();
542 		goto out;
543 	}
544 
545 	/*
546 	 * at this point we expect a path name in arg.   we will
547 	 * use namei() to gain a vnode reference (vref), and lock
548 	 * the vnode (VOP_LOCK).
549 	 *
550 	 * XXX: a NULL arg means use the root vnode pointer (e.g. for
551 	 * miniroot)
552 	 */
553 	if (SCARG(uap, arg) == NULL) {
554 		vp = rootvp;		/* miniroot */
555 		if (vget(vp, LK_EXCLUSIVE)) {
556 			error = EBUSY;
557 			goto out;
558 		}
559 		if (SCARG(uap, cmd) == SWAP_ON &&
560 		    copystr("miniroot", userpath, SWAP_PATH_MAX, &len))
561 			panic("swapctl: miniroot copy failed");
562 	} else {
563 		int	space;
564 		char	*where;
565 
566 		if (SCARG(uap, cmd) == SWAP_ON) {
567 			if ((error = copyinstr(SCARG(uap, arg), userpath,
568 			    SWAP_PATH_MAX, &len)))
569 				goto out;
570 			space = UIO_SYSSPACE;
571 			where = userpath;
572 		} else {
573 			space = UIO_USERSPACE;
574 			where = (char *)SCARG(uap, arg);
575 		}
576 		NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | TRYEMULROOT,
577 		    space, where);
578 		if ((error = namei(&nd)))
579 			goto out;
580 		vp = nd.ni_vp;
581 	}
582 	/* note: "vp" is referenced and locked */
583 
584 	error = 0;		/* assume no error */
585 	switch(SCARG(uap, cmd)) {
586 
587 	case SWAP_DUMPDEV:
588 		if (vp->v_type != VBLK) {
589 			error = ENOTBLK;
590 			break;
591 		}
592 		if (bdevsw_lookup(vp->v_rdev)) {
593 			dumpdev = vp->v_rdev;
594 			dumpcdev = devsw_blk2chr(dumpdev);
595 		} else
596 			dumpdev = NODEV;
597 		cpu_dumpconf();
598 		break;
599 
600 	case SWAP_CTL:
601 		/*
602 		 * get new priority, remove old entry (if any) and then
603 		 * reinsert it in the correct place.  finally, prune out
604 		 * any empty priority structures.
605 		 */
606 		priority = SCARG(uap, misc);
607 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
608 		mutex_enter(&uvm_swap_data_lock);
609 		if ((sdp = swaplist_find(vp, true)) == NULL) {
610 			error = ENOENT;
611 		} else {
612 			swaplist_insert(sdp, spp, priority);
613 			swaplist_trim();
614 		}
615 		mutex_exit(&uvm_swap_data_lock);
616 		if (error)
617 			free(spp, M_VMSWAP);
618 		break;
619 
620 	case SWAP_ON:
621 
622 		/*
623 		 * check for duplicates.   if none found, then insert a
624 		 * dummy entry on the list to prevent someone else from
625 		 * trying to enable this device while we are working on
626 		 * it.
627 		 */
628 
629 		priority = SCARG(uap, misc);
630 		sdp = malloc(sizeof *sdp, M_VMSWAP, M_WAITOK);
631 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
632 		memset(sdp, 0, sizeof(*sdp));
633 		sdp->swd_flags = SWF_FAKE;
634 		sdp->swd_vp = vp;
635 		sdp->swd_dev = (vp->v_type == VBLK) ? vp->v_rdev : NODEV;
636 		bufq_alloc(&sdp->swd_tab, "disksort", BUFQ_SORT_RAWBLOCK);
637 		mutex_enter(&uvm_swap_data_lock);
638 		if (swaplist_find(vp, false) != NULL) {
639 			error = EBUSY;
640 			mutex_exit(&uvm_swap_data_lock);
641 			bufq_free(sdp->swd_tab);
642 			free(sdp, M_VMSWAP);
643 			free(spp, M_VMSWAP);
644 			break;
645 		}
646 		swaplist_insert(sdp, spp, priority);
647 		mutex_exit(&uvm_swap_data_lock);
648 
649 		sdp->swd_pathlen = len;
650 		sdp->swd_path = malloc(sdp->swd_pathlen, M_VMSWAP, M_WAITOK);
651 		if (copystr(userpath, sdp->swd_path, sdp->swd_pathlen, 0) != 0)
652 			panic("swapctl: copystr");
653 
654 		/*
655 		 * we've now got a FAKE placeholder in the swap list.
656 		 * now attempt to enable swap on it.  if we fail, undo
657 		 * what we've done and kill the fake entry we just inserted.
658 		 * if swap_on is a success, it will clear the SWF_FAKE flag
659 		 */
660 
661 		if ((error = swap_on(l, sdp)) != 0) {
662 			mutex_enter(&uvm_swap_data_lock);
663 			(void) swaplist_find(vp, true);  /* kill fake entry */
664 			swaplist_trim();
665 			mutex_exit(&uvm_swap_data_lock);
666 			bufq_free(sdp->swd_tab);
667 			free(sdp->swd_path, M_VMSWAP);
668 			free(sdp, M_VMSWAP);
669 			break;
670 		}
671 		break;
672 
673 	case SWAP_OFF:
674 		mutex_enter(&uvm_swap_data_lock);
675 		if ((sdp = swaplist_find(vp, false)) == NULL) {
676 			mutex_exit(&uvm_swap_data_lock);
677 			error = ENXIO;
678 			break;
679 		}
680 
681 		/*
682 		 * If a device isn't in use or enabled, we
683 		 * can't stop swapping from it (again).
684 		 */
685 		if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0) {
686 			mutex_exit(&uvm_swap_data_lock);
687 			error = EBUSY;
688 			break;
689 		}
690 
691 		/*
692 		 * do the real work.
693 		 */
694 		error = swap_off(l, sdp);
695 		break;
696 
697 	default:
698 		error = EINVAL;
699 	}
700 
701 	/*
702 	 * done!  release the ref gained by namei() and unlock.
703 	 */
704 	vput(vp);
705 
706 out:
707 	free(userpath, M_TEMP);
708 	rw_exit(&swap_syscall_lock);
709 
710 	UVMHIST_LOG(pdhist, "<- done!  error=%d", error, 0, 0, 0);
711 	return (error);
712 }
713 
714 /*
715  * swap_stats: implements swapctl(SWAP_STATS). The function is kept
716  * away from sys_swapctl() in order to allow COMPAT_* swapctl()
717  * emulation to use it directly without going through sys_swapctl().
718  * The problem with using sys_swapctl() there is that it involves
719  * copying the swapent array to the stackgap, and this array's size
720  * is not known at build time. Hence it would not be possible to
721  * ensure it would fit in the stackgap in any case.
722  */
723 void
724 uvm_swap_stats(int cmd, struct swapent *sep, int sec, register_t *retval)
725 {
726 
727 	rw_enter(&swap_syscall_lock, RW_READER);
728 	uvm_swap_stats_locked(cmd, sep, sec, retval);
729 	rw_exit(&swap_syscall_lock);
730 }
731 
732 static void
733 uvm_swap_stats_locked(int cmd, struct swapent *sep, int sec, register_t *retval)
734 {
735 	struct swappri *spp;
736 	struct swapdev *sdp;
737 	int count = 0;
738 
739 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
740 		for (sdp = CIRCLEQ_FIRST(&spp->spi_swapdev);
741 		     sdp != (void *)&spp->spi_swapdev && sec-- > 0;
742 		     sdp = CIRCLEQ_NEXT(sdp, swd_next)) {
743 			int inuse;
744 
745 		  	/*
746 			 * backwards compatibility for system call.
747 			 * For NetBSD 1.3 and 5.0, we have to use
748 			 * the 32 bit dev_t.  For 5.0 and -current
749 			 * we have to add the path.
750 			 */
751 			inuse = btodb((uint64_t)sdp->swd_npginuse <<
752 			    PAGE_SHIFT);
753 
754 #if defined(COMPAT_13) || defined(COMPAT_50)
755 			if (cmd == SWAP_STATS) {
756 #endif
757 				sep->se_dev = sdp->swd_dev;
758 				sep->se_flags = sdp->swd_flags;
759 				sep->se_nblks = sdp->swd_nblks;
760 				sep->se_inuse = inuse;
761 				sep->se_priority = sdp->swd_priority;
762 				memcpy(&sep->se_path, sdp->swd_path,
763 				       sizeof sep->se_path);
764 				sep++;
765 #if defined(COMPAT_13)
766 			} else if (cmd == SWAP_STATS13) {
767 				struct swapent13 *sep13 =
768 				    (struct swapent13 *)sep;
769 
770 				sep13->se13_dev = sdp->swd_dev;
771 				sep13->se13_flags = sdp->swd_flags;
772 				sep13->se13_nblks = sdp->swd_nblks;
773 				sep13->se13_inuse = inuse;
774 				sep13->se13_priority = sdp->swd_priority;
775 				sep = (struct swapent *)(sep13 + 1);
776 #endif
777 #if defined(COMPAT_50)
778 			} else if (cmd == SWAP_STATS50) {
779 				struct swapent50 *sep50 =
780 				    (struct swapent50 *)sep;
781 
782 				sep50->se50_dev = sdp->swd_dev;
783 				sep50->se50_flags = sdp->swd_flags;
784 				sep50->se50_nblks = sdp->swd_nblks;
785 				sep50->se50_inuse = inuse;
786 				sep50->se50_priority = sdp->swd_priority;
787 				memcpy(&sep50->se50_path, sdp->swd_path,
788 				       sizeof sep50->se50_path);
789 				sep = (struct swapent *)(sep50 + 1);
790 #endif
791 #if defined(COMPAT_13) || defined(COMPAT_50)
792 			}
793 #endif
794 			count++;
795 		}
796 	}
797 
798 	*retval = count;
799 	return;
800 }
801 
802 /*
803  * swap_on: attempt to enable a swapdev for swapping.   note that the
804  *	swapdev is already on the global list, but disabled (marked
805  *	SWF_FAKE).
806  *
807  * => we avoid the start of the disk (to protect disk labels)
808  * => we also avoid the miniroot, if we are swapping to root.
809  * => caller should leave uvm_swap_data_lock unlocked, we may lock it
810  *	if needed.
811  */
812 static int
813 swap_on(struct lwp *l, struct swapdev *sdp)
814 {
815 	struct vnode *vp;
816 	int error, npages, nblocks, size;
817 	long addr;
818 	u_long result;
819 	struct vattr va;
820 	const struct bdevsw *bdev;
821 	dev_t dev;
822 	UVMHIST_FUNC("swap_on"); UVMHIST_CALLED(pdhist);
823 
824 	/*
825 	 * we want to enable swapping on sdp.   the swd_vp contains
826 	 * the vnode we want (locked and ref'd), and the swd_dev
827 	 * contains the dev_t of the file, if it a block device.
828 	 */
829 
830 	vp = sdp->swd_vp;
831 	dev = sdp->swd_dev;
832 
833 	/*
834 	 * open the swap file (mostly useful for block device files to
835 	 * let device driver know what is up).
836 	 *
837 	 * we skip the open/close for root on swap because the root
838 	 * has already been opened when root was mounted (mountroot).
839 	 */
840 	if (vp != rootvp) {
841 		if ((error = VOP_OPEN(vp, FREAD|FWRITE, l->l_cred)))
842 			return (error);
843 	}
844 
845 	/* XXX this only works for block devices */
846 	UVMHIST_LOG(pdhist, "  dev=%d, major(dev)=%d", dev, major(dev), 0,0);
847 
848 	/*
849 	 * we now need to determine the size of the swap area.   for
850 	 * block specials we can call the d_psize function.
851 	 * for normal files, we must stat [get attrs].
852 	 *
853 	 * we put the result in nblks.
854 	 * for normal files, we also want the filesystem block size
855 	 * (which we get with statfs).
856 	 */
857 	switch (vp->v_type) {
858 	case VBLK:
859 		bdev = bdevsw_lookup(dev);
860 		if (bdev == NULL || bdev->d_psize == NULL ||
861 		    (nblocks = (*bdev->d_psize)(dev)) == -1) {
862 			error = ENXIO;
863 			goto bad;
864 		}
865 		break;
866 
867 	case VREG:
868 		if ((error = VOP_GETATTR(vp, &va, l->l_cred)))
869 			goto bad;
870 		nblocks = (int)btodb(va.va_size);
871 		sdp->swd_bsize = 1 << vp->v_mount->mnt_fs_bshift;
872 		/*
873 		 * limit the max # of outstanding I/O requests we issue
874 		 * at any one time.   take it easy on NFS servers.
875 		 */
876 		if (vp->v_tag == VT_NFS)
877 			sdp->swd_maxactive = 2; /* XXX */
878 		else
879 			sdp->swd_maxactive = 8; /* XXX */
880 		break;
881 
882 	default:
883 		error = ENXIO;
884 		goto bad;
885 	}
886 
887 	/*
888 	 * save nblocks in a safe place and convert to pages.
889 	 */
890 
891 	sdp->swd_nblks = nblocks;
892 	npages = dbtob((uint64_t)nblocks) >> PAGE_SHIFT;
893 
894 	/*
895 	 * for block special files, we want to make sure that leave
896 	 * the disklabel and bootblocks alone, so we arrange to skip
897 	 * over them (arbitrarily choosing to skip PAGE_SIZE bytes).
898 	 * note that because of this the "size" can be less than the
899 	 * actual number of blocks on the device.
900 	 */
901 	if (vp->v_type == VBLK) {
902 		/* we use pages 1 to (size - 1) [inclusive] */
903 		size = npages - 1;
904 		addr = 1;
905 	} else {
906 		/* we use pages 0 to (size - 1) [inclusive] */
907 		size = npages;
908 		addr = 0;
909 	}
910 
911 	/*
912 	 * make sure we have enough blocks for a reasonable sized swap
913 	 * area.   we want at least one page.
914 	 */
915 
916 	if (size < 1) {
917 		UVMHIST_LOG(pdhist, "  size <= 1!!", 0, 0, 0, 0);
918 		error = EINVAL;
919 		goto bad;
920 	}
921 
922 	UVMHIST_LOG(pdhist, "  dev=%x: size=%d addr=%ld\n", dev, size, addr, 0);
923 
924 	/*
925 	 * now we need to allocate an extent to manage this swap device
926 	 */
927 
928 	sdp->swd_blist = blist_create(npages);
929 	/* mark all expect the `saved' region free. */
930 	blist_free(sdp->swd_blist, addr, size);
931 
932 	/*
933 	 * if the vnode we are swapping to is the root vnode
934 	 * (i.e. we are swapping to the miniroot) then we want
935 	 * to make sure we don't overwrite it.   do a statfs to
936 	 * find its size and skip over it.
937 	 */
938 	if (vp == rootvp) {
939 		struct mount *mp;
940 		struct statvfs *sp;
941 		int rootblocks, rootpages;
942 
943 		mp = rootvnode->v_mount;
944 		sp = &mp->mnt_stat;
945 		rootblocks = sp->f_blocks * btodb(sp->f_frsize);
946 		/*
947 		 * XXX: sp->f_blocks isn't the total number of
948 		 * blocks in the filesystem, it's the number of
949 		 * data blocks.  so, our rootblocks almost
950 		 * definitely underestimates the total size
951 		 * of the filesystem - how badly depends on the
952 		 * details of the filesystem type.  there isn't
953 		 * an obvious way to deal with this cleanly
954 		 * and perfectly, so for now we just pad our
955 		 * rootblocks estimate with an extra 5 percent.
956 		 */
957 		rootblocks += (rootblocks >> 5) +
958 			(rootblocks >> 6) +
959 			(rootblocks >> 7);
960 		rootpages = round_page(dbtob(rootblocks)) >> PAGE_SHIFT;
961 		if (rootpages > size)
962 			panic("swap_on: miniroot larger than swap?");
963 
964 		if (rootpages != blist_fill(sdp->swd_blist, addr, rootpages)) {
965 			panic("swap_on: unable to preserve miniroot");
966 		}
967 
968 		size -= rootpages;
969 		printf("Preserved %d pages of miniroot ", rootpages);
970 		printf("leaving %d pages of swap\n", size);
971 	}
972 
973 	/*
974 	 * add a ref to vp to reflect usage as a swap device.
975 	 */
976 	vref(vp);
977 
978 	/*
979 	 * now add the new swapdev to the drum and enable.
980 	 */
981 	result = vmem_alloc(swapmap, npages, VM_BESTFIT | VM_SLEEP);
982 	if (result == 0)
983 		panic("swapdrum_add");
984 	/*
985 	 * If this is the first regular swap create the workqueue.
986 	 * => Protected by swap_syscall_lock.
987 	 */
988 	if (vp->v_type != VBLK) {
989 		if (sw_reg_count++ == 0) {
990 			KASSERT(sw_reg_workqueue == NULL);
991 			if (workqueue_create(&sw_reg_workqueue, "swapiod",
992 			    sw_reg_iodone, NULL, PRIBIO, IPL_BIO, 0) != 0)
993 				panic("%s: workqueue_create failed", __func__);
994 		}
995 	}
996 
997 	sdp->swd_drumoffset = (int)result;
998 	sdp->swd_drumsize = npages;
999 	sdp->swd_npages = size;
1000 	mutex_enter(&uvm_swap_data_lock);
1001 	sdp->swd_flags &= ~SWF_FAKE;	/* going live */
1002 	sdp->swd_flags |= (SWF_INUSE|SWF_ENABLE);
1003 	uvmexp.swpages += size;
1004 	uvmexp.swpgavail += size;
1005 	mutex_exit(&uvm_swap_data_lock);
1006 	return (0);
1007 
1008 	/*
1009 	 * failure: clean up and return error.
1010 	 */
1011 
1012 bad:
1013 	if (sdp->swd_blist) {
1014 		blist_destroy(sdp->swd_blist);
1015 	}
1016 	if (vp != rootvp) {
1017 		(void)VOP_CLOSE(vp, FREAD|FWRITE, l->l_cred);
1018 	}
1019 	return (error);
1020 }
1021 
1022 /*
1023  * swap_off: stop swapping on swapdev
1024  *
1025  * => swap data should be locked, we will unlock.
1026  */
1027 static int
1028 swap_off(struct lwp *l, struct swapdev *sdp)
1029 {
1030 	int npages = sdp->swd_npages;
1031 	int error = 0;
1032 
1033 	UVMHIST_FUNC("swap_off"); UVMHIST_CALLED(pdhist);
1034 	UVMHIST_LOG(pdhist, "  dev=%x, npages=%d", sdp->swd_dev,npages,0,0);
1035 
1036 	/* disable the swap area being removed */
1037 	sdp->swd_flags &= ~SWF_ENABLE;
1038 	uvmexp.swpgavail -= npages;
1039 	mutex_exit(&uvm_swap_data_lock);
1040 
1041 	/*
1042 	 * the idea is to find all the pages that are paged out to this
1043 	 * device, and page them all in.  in uvm, swap-backed pageable
1044 	 * memory can take two forms: aobjs and anons.  call the
1045 	 * swapoff hook for each subsystem to bring in pages.
1046 	 */
1047 
1048 	if (uao_swap_off(sdp->swd_drumoffset,
1049 			 sdp->swd_drumoffset + sdp->swd_drumsize) ||
1050 	    amap_swap_off(sdp->swd_drumoffset,
1051 			  sdp->swd_drumoffset + sdp->swd_drumsize)) {
1052 		error = ENOMEM;
1053 	} else if (sdp->swd_npginuse > sdp->swd_npgbad) {
1054 		error = EBUSY;
1055 	}
1056 
1057 	if (error) {
1058 		mutex_enter(&uvm_swap_data_lock);
1059 		sdp->swd_flags |= SWF_ENABLE;
1060 		uvmexp.swpgavail += npages;
1061 		mutex_exit(&uvm_swap_data_lock);
1062 
1063 		return error;
1064 	}
1065 
1066 	/*
1067 	 * If this is the last regular swap destroy the workqueue.
1068 	 * => Protected by swap_syscall_lock.
1069 	 */
1070 	if (sdp->swd_vp->v_type != VBLK) {
1071 		KASSERT(sw_reg_count > 0);
1072 		KASSERT(sw_reg_workqueue != NULL);
1073 		if (--sw_reg_count == 0) {
1074 			workqueue_destroy(sw_reg_workqueue);
1075 			sw_reg_workqueue = NULL;
1076 		}
1077 	}
1078 
1079 	/*
1080 	 * done with the vnode.
1081 	 * drop our ref on the vnode before calling VOP_CLOSE()
1082 	 * so that spec_close() can tell if this is the last close.
1083 	 */
1084 	vrele(sdp->swd_vp);
1085 	if (sdp->swd_vp != rootvp) {
1086 		(void) VOP_CLOSE(sdp->swd_vp, FREAD|FWRITE, l->l_cred);
1087 	}
1088 
1089 	mutex_enter(&uvm_swap_data_lock);
1090 	uvmexp.swpages -= npages;
1091 	uvmexp.swpginuse -= sdp->swd_npgbad;
1092 
1093 	if (swaplist_find(sdp->swd_vp, true) == NULL)
1094 		panic("%s: swapdev not in list", __func__);
1095 	swaplist_trim();
1096 	mutex_exit(&uvm_swap_data_lock);
1097 
1098 	/*
1099 	 * free all resources!
1100 	 */
1101 	vmem_free(swapmap, sdp->swd_drumoffset, sdp->swd_drumsize);
1102 	blist_destroy(sdp->swd_blist);
1103 	bufq_free(sdp->swd_tab);
1104 	free(sdp, M_VMSWAP);
1105 	return (0);
1106 }
1107 
1108 /*
1109  * /dev/drum interface and i/o functions
1110  */
1111 
1112 /*
1113  * swstrategy: perform I/O on the drum
1114  *
1115  * => we must map the i/o request from the drum to the correct swapdev.
1116  */
1117 static void
1118 swstrategy(struct buf *bp)
1119 {
1120 	struct swapdev *sdp;
1121 	struct vnode *vp;
1122 	int pageno, bn;
1123 	UVMHIST_FUNC("swstrategy"); UVMHIST_CALLED(pdhist);
1124 
1125 	/*
1126 	 * convert block number to swapdev.   note that swapdev can't
1127 	 * be yanked out from under us because we are holding resources
1128 	 * in it (i.e. the blocks we are doing I/O on).
1129 	 */
1130 	pageno = dbtob((int64_t)bp->b_blkno) >> PAGE_SHIFT;
1131 	mutex_enter(&uvm_swap_data_lock);
1132 	sdp = swapdrum_getsdp(pageno);
1133 	mutex_exit(&uvm_swap_data_lock);
1134 	if (sdp == NULL) {
1135 		bp->b_error = EINVAL;
1136 		biodone(bp);
1137 		UVMHIST_LOG(pdhist, "  failed to get swap device", 0, 0, 0, 0);
1138 		return;
1139 	}
1140 
1141 	/*
1142 	 * convert drum page number to block number on this swapdev.
1143 	 */
1144 
1145 	pageno -= sdp->swd_drumoffset;	/* page # on swapdev */
1146 	bn = btodb((uint64_t)pageno << PAGE_SHIFT); /* convert to diskblock */
1147 
1148 	UVMHIST_LOG(pdhist, "  %s: mapoff=%x bn=%x bcount=%ld",
1149 		((bp->b_flags & B_READ) == 0) ? "write" : "read",
1150 		sdp->swd_drumoffset, bn, bp->b_bcount);
1151 
1152 	/*
1153 	 * for block devices we finish up here.
1154 	 * for regular files we have to do more work which we delegate
1155 	 * to sw_reg_strategy().
1156 	 */
1157 
1158 	vp = sdp->swd_vp;		/* swapdev vnode pointer */
1159 	switch (vp->v_type) {
1160 	default:
1161 		panic("%s: vnode type 0x%x", __func__, vp->v_type);
1162 
1163 	case VBLK:
1164 
1165 		/*
1166 		 * must convert "bp" from an I/O on /dev/drum to an I/O
1167 		 * on the swapdev (sdp).
1168 		 */
1169 		bp->b_blkno = bn;		/* swapdev block number */
1170 		bp->b_dev = sdp->swd_dev;	/* swapdev dev_t */
1171 
1172 		/*
1173 		 * if we are doing a write, we have to redirect the i/o on
1174 		 * drum's v_numoutput counter to the swapdevs.
1175 		 */
1176 		if ((bp->b_flags & B_READ) == 0) {
1177 			mutex_enter(bp->b_objlock);
1178 			vwakeup(bp);	/* kills one 'v_numoutput' on drum */
1179 			mutex_exit(bp->b_objlock);
1180 			mutex_enter(&vp->v_interlock);
1181 			vp->v_numoutput++;	/* put it on swapdev */
1182 			mutex_exit(&vp->v_interlock);
1183 		}
1184 
1185 		/*
1186 		 * finally plug in swapdev vnode and start I/O
1187 		 */
1188 		bp->b_vp = vp;
1189 		bp->b_objlock = &vp->v_interlock;
1190 		VOP_STRATEGY(vp, bp);
1191 		return;
1192 
1193 	case VREG:
1194 		/*
1195 		 * delegate to sw_reg_strategy function.
1196 		 */
1197 		sw_reg_strategy(sdp, bp, bn);
1198 		return;
1199 	}
1200 	/* NOTREACHED */
1201 }
1202 
1203 /*
1204  * swread: the read function for the drum (just a call to physio)
1205  */
1206 /*ARGSUSED*/
1207 static int
1208 swread(dev_t dev, struct uio *uio, int ioflag)
1209 {
1210 	UVMHIST_FUNC("swread"); UVMHIST_CALLED(pdhist);
1211 
1212 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1213 	return (physio(swstrategy, NULL, dev, B_READ, minphys, uio));
1214 }
1215 
1216 /*
1217  * swwrite: the write function for the drum (just a call to physio)
1218  */
1219 /*ARGSUSED*/
1220 static int
1221 swwrite(dev_t dev, struct uio *uio, int ioflag)
1222 {
1223 	UVMHIST_FUNC("swwrite"); UVMHIST_CALLED(pdhist);
1224 
1225 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1226 	return (physio(swstrategy, NULL, dev, B_WRITE, minphys, uio));
1227 }
1228 
1229 const struct bdevsw swap_bdevsw = {
1230 	nullopen, nullclose, swstrategy, noioctl, nodump, nosize, D_OTHER,
1231 };
1232 
1233 const struct cdevsw swap_cdevsw = {
1234 	nullopen, nullclose, swread, swwrite, noioctl,
1235 	nostop, notty, nopoll, nommap, nokqfilter, D_OTHER,
1236 };
1237 
1238 /*
1239  * sw_reg_strategy: handle swap i/o to regular files
1240  */
1241 static void
1242 sw_reg_strategy(struct swapdev *sdp, struct buf *bp, int bn)
1243 {
1244 	struct vnode	*vp;
1245 	struct vndxfer	*vnx;
1246 	daddr_t		nbn;
1247 	char 		*addr;
1248 	off_t		byteoff;
1249 	int		s, off, nra, error, sz, resid;
1250 	UVMHIST_FUNC("sw_reg_strategy"); UVMHIST_CALLED(pdhist);
1251 
1252 	/*
1253 	 * allocate a vndxfer head for this transfer and point it to
1254 	 * our buffer.
1255 	 */
1256 	vnx = pool_get(&vndxfer_pool, PR_WAITOK);
1257 	vnx->vx_flags = VX_BUSY;
1258 	vnx->vx_error = 0;
1259 	vnx->vx_pending = 0;
1260 	vnx->vx_bp = bp;
1261 	vnx->vx_sdp = sdp;
1262 
1263 	/*
1264 	 * setup for main loop where we read filesystem blocks into
1265 	 * our buffer.
1266 	 */
1267 	error = 0;
1268 	bp->b_resid = bp->b_bcount;	/* nothing transfered yet! */
1269 	addr = bp->b_data;		/* current position in buffer */
1270 	byteoff = dbtob((uint64_t)bn);
1271 
1272 	for (resid = bp->b_resid; resid; resid -= sz) {
1273 		struct vndbuf	*nbp;
1274 
1275 		/*
1276 		 * translate byteoffset into block number.  return values:
1277 		 *   vp = vnode of underlying device
1278 		 *  nbn = new block number (on underlying vnode dev)
1279 		 *  nra = num blocks we can read-ahead (excludes requested
1280 		 *	block)
1281 		 */
1282 		nra = 0;
1283 		error = VOP_BMAP(sdp->swd_vp, byteoff / sdp->swd_bsize,
1284 				 	&vp, &nbn, &nra);
1285 
1286 		if (error == 0 && nbn == (daddr_t)-1) {
1287 			/*
1288 			 * this used to just set error, but that doesn't
1289 			 * do the right thing.  Instead, it causes random
1290 			 * memory errors.  The panic() should remain until
1291 			 * this condition doesn't destabilize the system.
1292 			 */
1293 #if 1
1294 			panic("%s: swap to sparse file", __func__);
1295 #else
1296 			error = EIO;	/* failure */
1297 #endif
1298 		}
1299 
1300 		/*
1301 		 * punt if there was an error or a hole in the file.
1302 		 * we must wait for any i/o ops we have already started
1303 		 * to finish before returning.
1304 		 *
1305 		 * XXX we could deal with holes here but it would be
1306 		 * a hassle (in the write case).
1307 		 */
1308 		if (error) {
1309 			s = splbio();
1310 			vnx->vx_error = error;	/* pass error up */
1311 			goto out;
1312 		}
1313 
1314 		/*
1315 		 * compute the size ("sz") of this transfer (in bytes).
1316 		 */
1317 		off = byteoff % sdp->swd_bsize;
1318 		sz = (1 + nra) * sdp->swd_bsize - off;
1319 		if (sz > resid)
1320 			sz = resid;
1321 
1322 		UVMHIST_LOG(pdhist, "sw_reg_strategy: "
1323 			    "vp %p/%p offset 0x%x/0x%x",
1324 			    sdp->swd_vp, vp, byteoff, nbn);
1325 
1326 		/*
1327 		 * now get a buf structure.   note that the vb_buf is
1328 		 * at the front of the nbp structure so that you can
1329 		 * cast pointers between the two structure easily.
1330 		 */
1331 		nbp = pool_get(&vndbuf_pool, PR_WAITOK);
1332 		buf_init(&nbp->vb_buf);
1333 		nbp->vb_buf.b_flags    = bp->b_flags;
1334 		nbp->vb_buf.b_cflags   = bp->b_cflags;
1335 		nbp->vb_buf.b_oflags   = bp->b_oflags;
1336 		nbp->vb_buf.b_bcount   = sz;
1337 		nbp->vb_buf.b_bufsize  = sz;
1338 		nbp->vb_buf.b_error    = 0;
1339 		nbp->vb_buf.b_data     = addr;
1340 		nbp->vb_buf.b_lblkno   = 0;
1341 		nbp->vb_buf.b_blkno    = nbn + btodb(off);
1342 		nbp->vb_buf.b_rawblkno = nbp->vb_buf.b_blkno;
1343 		nbp->vb_buf.b_iodone   = sw_reg_biodone;
1344 		nbp->vb_buf.b_vp       = vp;
1345 		nbp->vb_buf.b_objlock  = &vp->v_interlock;
1346 		if (vp->v_type == VBLK) {
1347 			nbp->vb_buf.b_dev = vp->v_rdev;
1348 		}
1349 
1350 		nbp->vb_xfer = vnx;	/* patch it back in to vnx */
1351 
1352 		/*
1353 		 * Just sort by block number
1354 		 */
1355 		s = splbio();
1356 		if (vnx->vx_error != 0) {
1357 			buf_destroy(&nbp->vb_buf);
1358 			pool_put(&vndbuf_pool, nbp);
1359 			goto out;
1360 		}
1361 		vnx->vx_pending++;
1362 
1363 		/* sort it in and start I/O if we are not over our limit */
1364 		/* XXXAD locking */
1365 		bufq_put(sdp->swd_tab, &nbp->vb_buf);
1366 		sw_reg_start(sdp);
1367 		splx(s);
1368 
1369 		/*
1370 		 * advance to the next I/O
1371 		 */
1372 		byteoff += sz;
1373 		addr += sz;
1374 	}
1375 
1376 	s = splbio();
1377 
1378 out: /* Arrive here at splbio */
1379 	vnx->vx_flags &= ~VX_BUSY;
1380 	if (vnx->vx_pending == 0) {
1381 		error = vnx->vx_error;
1382 		pool_put(&vndxfer_pool, vnx);
1383 		bp->b_error = error;
1384 		biodone(bp);
1385 	}
1386 	splx(s);
1387 }
1388 
1389 /*
1390  * sw_reg_start: start an I/O request on the requested swapdev
1391  *
1392  * => reqs are sorted by b_rawblkno (above)
1393  */
1394 static void
1395 sw_reg_start(struct swapdev *sdp)
1396 {
1397 	struct buf	*bp;
1398 	struct vnode	*vp;
1399 	UVMHIST_FUNC("sw_reg_start"); UVMHIST_CALLED(pdhist);
1400 
1401 	/* recursion control */
1402 	if ((sdp->swd_flags & SWF_BUSY) != 0)
1403 		return;
1404 
1405 	sdp->swd_flags |= SWF_BUSY;
1406 
1407 	while (sdp->swd_active < sdp->swd_maxactive) {
1408 		bp = bufq_get(sdp->swd_tab);
1409 		if (bp == NULL)
1410 			break;
1411 		sdp->swd_active++;
1412 
1413 		UVMHIST_LOG(pdhist,
1414 		    "sw_reg_start:  bp %p vp %p blkno %p cnt %lx",
1415 		    bp, bp->b_vp, bp->b_blkno, bp->b_bcount);
1416 		vp = bp->b_vp;
1417 		KASSERT(bp->b_objlock == &vp->v_interlock);
1418 		if ((bp->b_flags & B_READ) == 0) {
1419 			mutex_enter(&vp->v_interlock);
1420 			vp->v_numoutput++;
1421 			mutex_exit(&vp->v_interlock);
1422 		}
1423 		VOP_STRATEGY(vp, bp);
1424 	}
1425 	sdp->swd_flags &= ~SWF_BUSY;
1426 }
1427 
1428 /*
1429  * sw_reg_biodone: one of our i/o's has completed
1430  */
1431 static void
1432 sw_reg_biodone(struct buf *bp)
1433 {
1434 	workqueue_enqueue(sw_reg_workqueue, &bp->b_work, NULL);
1435 }
1436 
1437 /*
1438  * sw_reg_iodone: one of our i/o's has completed and needs post-i/o cleanup
1439  *
1440  * => note that we can recover the vndbuf struct by casting the buf ptr
1441  */
1442 static void
1443 sw_reg_iodone(struct work *wk, void *dummy)
1444 {
1445 	struct vndbuf *vbp = (void *)wk;
1446 	struct vndxfer *vnx = vbp->vb_xfer;
1447 	struct buf *pbp = vnx->vx_bp;		/* parent buffer */
1448 	struct swapdev	*sdp = vnx->vx_sdp;
1449 	int s, resid, error;
1450 	KASSERT(&vbp->vb_buf.b_work == wk);
1451 	UVMHIST_FUNC("sw_reg_iodone"); UVMHIST_CALLED(pdhist);
1452 
1453 	UVMHIST_LOG(pdhist, "  vbp=%p vp=%p blkno=%x addr=%p",
1454 	    vbp, vbp->vb_buf.b_vp, vbp->vb_buf.b_blkno, vbp->vb_buf.b_data);
1455 	UVMHIST_LOG(pdhist, "  cnt=%lx resid=%lx",
1456 	    vbp->vb_buf.b_bcount, vbp->vb_buf.b_resid, 0, 0);
1457 
1458 	/*
1459 	 * protect vbp at splbio and update.
1460 	 */
1461 
1462 	s = splbio();
1463 	resid = vbp->vb_buf.b_bcount - vbp->vb_buf.b_resid;
1464 	pbp->b_resid -= resid;
1465 	vnx->vx_pending--;
1466 
1467 	if (vbp->vb_buf.b_error != 0) {
1468 		/* pass error upward */
1469 		error = vbp->vb_buf.b_error ? vbp->vb_buf.b_error : EIO;
1470 		UVMHIST_LOG(pdhist, "  got error=%d !", error, 0, 0, 0);
1471 		vnx->vx_error = error;
1472 	}
1473 
1474 	/*
1475 	 * kill vbp structure
1476 	 */
1477 	buf_destroy(&vbp->vb_buf);
1478 	pool_put(&vndbuf_pool, vbp);
1479 
1480 	/*
1481 	 * wrap up this transaction if it has run to completion or, in
1482 	 * case of an error, when all auxiliary buffers have returned.
1483 	 */
1484 	if (vnx->vx_error != 0) {
1485 		/* pass error upward */
1486 		error = vnx->vx_error;
1487 		if ((vnx->vx_flags & VX_BUSY) == 0 && vnx->vx_pending == 0) {
1488 			pbp->b_error = error;
1489 			biodone(pbp);
1490 			pool_put(&vndxfer_pool, vnx);
1491 		}
1492 	} else if (pbp->b_resid == 0) {
1493 		KASSERT(vnx->vx_pending == 0);
1494 		if ((vnx->vx_flags & VX_BUSY) == 0) {
1495 			UVMHIST_LOG(pdhist, "  iodone error=%d !",
1496 			    pbp, vnx->vx_error, 0, 0);
1497 			biodone(pbp);
1498 			pool_put(&vndxfer_pool, vnx);
1499 		}
1500 	}
1501 
1502 	/*
1503 	 * done!   start next swapdev I/O if one is pending
1504 	 */
1505 	sdp->swd_active--;
1506 	sw_reg_start(sdp);
1507 	splx(s);
1508 }
1509 
1510 
1511 /*
1512  * uvm_swap_alloc: allocate space on swap
1513  *
1514  * => allocation is done "round robin" down the priority list, as we
1515  *	allocate in a priority we "rotate" the circle queue.
1516  * => space can be freed with uvm_swap_free
1517  * => we return the page slot number in /dev/drum (0 == invalid slot)
1518  * => we lock uvm_swap_data_lock
1519  * => XXXMRG: "LESSOK" INTERFACE NEEDED TO EXTENT SYSTEM
1520  */
1521 int
1522 uvm_swap_alloc(int *nslots /* IN/OUT */, bool lessok)
1523 {
1524 	struct swapdev *sdp;
1525 	struct swappri *spp;
1526 	UVMHIST_FUNC("uvm_swap_alloc"); UVMHIST_CALLED(pdhist);
1527 
1528 	/*
1529 	 * no swap devices configured yet?   definite failure.
1530 	 */
1531 	if (uvmexp.nswapdev < 1)
1532 		return 0;
1533 
1534 	/*
1535 	 * lock data lock, convert slots into blocks, and enter loop
1536 	 */
1537 	mutex_enter(&uvm_swap_data_lock);
1538 
1539 ReTry:	/* XXXMRG */
1540 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
1541 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
1542 			uint64_t result;
1543 
1544 			/* if it's not enabled, then we can't swap from it */
1545 			if ((sdp->swd_flags & SWF_ENABLE) == 0)
1546 				continue;
1547 			if (sdp->swd_npginuse + *nslots > sdp->swd_npages)
1548 				continue;
1549 			result = blist_alloc(sdp->swd_blist, *nslots);
1550 			if (result == BLIST_NONE) {
1551 				continue;
1552 			}
1553 			KASSERT(result < sdp->swd_drumsize);
1554 
1555 			/*
1556 			 * successful allocation!  now rotate the circleq.
1557 			 */
1558 			CIRCLEQ_REMOVE(&spp->spi_swapdev, sdp, swd_next);
1559 			CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
1560 			sdp->swd_npginuse += *nslots;
1561 			uvmexp.swpginuse += *nslots;
1562 			mutex_exit(&uvm_swap_data_lock);
1563 			/* done!  return drum slot number */
1564 			UVMHIST_LOG(pdhist,
1565 			    "success!  returning %d slots starting at %d",
1566 			    *nslots, result + sdp->swd_drumoffset, 0, 0);
1567 			return (result + sdp->swd_drumoffset);
1568 		}
1569 	}
1570 
1571 	/* XXXMRG: BEGIN HACK */
1572 	if (*nslots > 1 && lessok) {
1573 		*nslots = 1;
1574 		/* XXXMRG: ugh!  blist should support this for us */
1575 		goto ReTry;
1576 	}
1577 	/* XXXMRG: END HACK */
1578 
1579 	mutex_exit(&uvm_swap_data_lock);
1580 	return 0;
1581 }
1582 
1583 /*
1584  * uvm_swapisfull: return true if most of available swap is allocated
1585  * and in use.  we don't count some small portion as it may be inaccessible
1586  * to us at any given moment, for example if there is lock contention or if
1587  * pages are busy.
1588  */
1589 bool
1590 uvm_swapisfull(void)
1591 {
1592 	int swpgonly;
1593 	bool rv;
1594 
1595 	mutex_enter(&uvm_swap_data_lock);
1596 	KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1597 	swpgonly = (int)((uint64_t)uvmexp.swpgonly * 100 /
1598 	    uvm_swapisfull_factor);
1599 	rv = (swpgonly >= uvmexp.swpgavail);
1600 	mutex_exit(&uvm_swap_data_lock);
1601 
1602 	return (rv);
1603 }
1604 
1605 /*
1606  * uvm_swap_markbad: keep track of swap ranges where we've had i/o errors
1607  *
1608  * => we lock uvm_swap_data_lock
1609  */
1610 void
1611 uvm_swap_markbad(int startslot, int nslots)
1612 {
1613 	struct swapdev *sdp;
1614 	UVMHIST_FUNC("uvm_swap_markbad"); UVMHIST_CALLED(pdhist);
1615 
1616 	mutex_enter(&uvm_swap_data_lock);
1617 	sdp = swapdrum_getsdp(startslot);
1618 	KASSERT(sdp != NULL);
1619 
1620 	/*
1621 	 * we just keep track of how many pages have been marked bad
1622 	 * in this device, to make everything add up in swap_off().
1623 	 * we assume here that the range of slots will all be within
1624 	 * one swap device.
1625 	 */
1626 
1627 	KASSERT(uvmexp.swpgonly >= nslots);
1628 	uvmexp.swpgonly -= nslots;
1629 	sdp->swd_npgbad += nslots;
1630 	UVMHIST_LOG(pdhist, "now %d bad", sdp->swd_npgbad, 0,0,0);
1631 	mutex_exit(&uvm_swap_data_lock);
1632 }
1633 
1634 /*
1635  * uvm_swap_free: free swap slots
1636  *
1637  * => this can be all or part of an allocation made by uvm_swap_alloc
1638  * => we lock uvm_swap_data_lock
1639  */
1640 void
1641 uvm_swap_free(int startslot, int nslots)
1642 {
1643 	struct swapdev *sdp;
1644 	UVMHIST_FUNC("uvm_swap_free"); UVMHIST_CALLED(pdhist);
1645 
1646 	UVMHIST_LOG(pdhist, "freeing %d slots starting at %d", nslots,
1647 	    startslot, 0, 0);
1648 
1649 	/*
1650 	 * ignore attempts to free the "bad" slot.
1651 	 */
1652 
1653 	if (startslot == SWSLOT_BAD) {
1654 		return;
1655 	}
1656 
1657 	/*
1658 	 * convert drum slot offset back to sdp, free the blocks
1659 	 * in the extent, and return.   must hold pri lock to do
1660 	 * lookup and access the extent.
1661 	 */
1662 
1663 	mutex_enter(&uvm_swap_data_lock);
1664 	sdp = swapdrum_getsdp(startslot);
1665 	KASSERT(uvmexp.nswapdev >= 1);
1666 	KASSERT(sdp != NULL);
1667 	KASSERT(sdp->swd_npginuse >= nslots);
1668 	blist_free(sdp->swd_blist, startslot - sdp->swd_drumoffset, nslots);
1669 	sdp->swd_npginuse -= nslots;
1670 	uvmexp.swpginuse -= nslots;
1671 	mutex_exit(&uvm_swap_data_lock);
1672 }
1673 
1674 /*
1675  * uvm_swap_put: put any number of pages into a contig place on swap
1676  *
1677  * => can be sync or async
1678  */
1679 
1680 int
1681 uvm_swap_put(int swslot, struct vm_page **ppsp, int npages, int flags)
1682 {
1683 	int error;
1684 
1685 	error = uvm_swap_io(ppsp, swslot, npages, B_WRITE |
1686 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1687 	return error;
1688 }
1689 
1690 /*
1691  * uvm_swap_get: get a single page from swap
1692  *
1693  * => usually a sync op (from fault)
1694  */
1695 
1696 int
1697 uvm_swap_get(struct vm_page *page, int swslot, int flags)
1698 {
1699 	int error;
1700 
1701 	uvmexp.nswget++;
1702 	KASSERT(flags & PGO_SYNCIO);
1703 	if (swslot == SWSLOT_BAD) {
1704 		return EIO;
1705 	}
1706 
1707 	error = uvm_swap_io(&page, swslot, 1, B_READ |
1708 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1709 	if (error == 0) {
1710 
1711 		/*
1712 		 * this page is no longer only in swap.
1713 		 */
1714 
1715 		mutex_enter(&uvm_swap_data_lock);
1716 		KASSERT(uvmexp.swpgonly > 0);
1717 		uvmexp.swpgonly--;
1718 		mutex_exit(&uvm_swap_data_lock);
1719 	}
1720 	return error;
1721 }
1722 
1723 /*
1724  * uvm_swap_io: do an i/o operation to swap
1725  */
1726 
1727 static int
1728 uvm_swap_io(struct vm_page **pps, int startslot, int npages, int flags)
1729 {
1730 	daddr_t startblk;
1731 	struct	buf *bp;
1732 	vaddr_t kva;
1733 	int	error, mapinflags;
1734 	bool write, async;
1735 	UVMHIST_FUNC("uvm_swap_io"); UVMHIST_CALLED(pdhist);
1736 
1737 	UVMHIST_LOG(pdhist, "<- called, startslot=%d, npages=%d, flags=%d",
1738 	    startslot, npages, flags, 0);
1739 
1740 	write = (flags & B_READ) == 0;
1741 	async = (flags & B_ASYNC) != 0;
1742 
1743 	/*
1744 	 * allocate a buf for the i/o.
1745 	 */
1746 
1747 	KASSERT(curlwp != uvm.pagedaemon_lwp || (write && async));
1748 	bp = getiobuf(swapdev_vp, curlwp != uvm.pagedaemon_lwp);
1749 	if (bp == NULL) {
1750 		uvm_aio_aiodone_pages(pps, npages, true, ENOMEM);
1751 		return ENOMEM;
1752 	}
1753 
1754 	/*
1755 	 * convert starting drum slot to block number
1756 	 */
1757 
1758 	startblk = btodb((uint64_t)startslot << PAGE_SHIFT);
1759 
1760 	/*
1761 	 * first, map the pages into the kernel.
1762 	 */
1763 
1764 	mapinflags = !write ?
1765 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_READ :
1766 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_WRITE;
1767 	kva = uvm_pagermapin(pps, npages, mapinflags);
1768 
1769 	/*
1770 	 * fill in the bp/sbp.   we currently route our i/o through
1771 	 * /dev/drum's vnode [swapdev_vp].
1772 	 */
1773 
1774 	bp->b_cflags = BC_BUSY | BC_NOCACHE;
1775 	bp->b_flags = (flags & (B_READ|B_ASYNC));
1776 	bp->b_proc = &proc0;	/* XXX */
1777 	bp->b_vnbufs.le_next = NOLIST;
1778 	bp->b_data = (void *)kva;
1779 	bp->b_blkno = startblk;
1780 	bp->b_bufsize = bp->b_bcount = npages << PAGE_SHIFT;
1781 
1782 	/*
1783 	 * bump v_numoutput (counter of number of active outputs).
1784 	 */
1785 
1786 	if (write) {
1787 		mutex_enter(&swapdev_vp->v_interlock);
1788 		swapdev_vp->v_numoutput++;
1789 		mutex_exit(&swapdev_vp->v_interlock);
1790 	}
1791 
1792 	/*
1793 	 * for async ops we must set up the iodone handler.
1794 	 */
1795 
1796 	if (async) {
1797 		bp->b_iodone = uvm_aio_biodone;
1798 		UVMHIST_LOG(pdhist, "doing async!", 0, 0, 0, 0);
1799 		if (curlwp == uvm.pagedaemon_lwp)
1800 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1801 		else
1802 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
1803 	} else {
1804 		bp->b_iodone = NULL;
1805 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1806 	}
1807 	UVMHIST_LOG(pdhist,
1808 	    "about to start io: data = %p blkno = 0x%x, bcount = %ld",
1809 	    bp->b_data, bp->b_blkno, bp->b_bcount, 0);
1810 
1811 	/*
1812 	 * now we start the I/O, and if async, return.
1813 	 */
1814 
1815 	VOP_STRATEGY(swapdev_vp, bp);
1816 	if (async)
1817 		return 0;
1818 
1819 	/*
1820 	 * must be sync i/o.   wait for it to finish
1821 	 */
1822 
1823 	error = biowait(bp);
1824 
1825 	/*
1826 	 * kill the pager mapping
1827 	 */
1828 
1829 	uvm_pagermapout(kva, npages);
1830 
1831 	/*
1832 	 * now dispose of the buf and we're done.
1833 	 */
1834 
1835 	if (write) {
1836 		mutex_enter(&swapdev_vp->v_interlock);
1837 		vwakeup(bp);
1838 		mutex_exit(&swapdev_vp->v_interlock);
1839 	}
1840 	putiobuf(bp);
1841 	UVMHIST_LOG(pdhist, "<- done (sync)  error=%d", error, 0, 0, 0);
1842 
1843 	return (error);
1844 }
1845