xref: /netbsd-src/sys/uvm/uvm_swap.c (revision e6c7e151de239c49d2e38720a061ed9d1fa99309)
1 /*	$NetBSD: uvm_swap.c,v 1.186 2020/02/18 20:23:17 chs 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 1.186 2020/02/18 20:23:17 chs 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/atomic.h>
42 #include <sys/buf.h>
43 #include <sys/bufq.h>
44 #include <sys/conf.h>
45 #include <sys/proc.h>
46 #include <sys/namei.h>
47 #include <sys/disklabel.h>
48 #include <sys/errno.h>
49 #include <sys/kernel.h>
50 #include <sys/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/kmem.h>
57 #include <sys/syscallargs.h>
58 #include <sys/swap.h>
59 #include <sys/kauth.h>
60 #include <sys/sysctl.h>
61 #include <sys/workqueue.h>
62 
63 #include <uvm/uvm.h>
64 
65 #include <miscfs/specfs/specdev.h>
66 
67 /*
68  * uvm_swap.c: manage configuration and i/o to swap space.
69  */
70 
71 /*
72  * swap space is managed in the following way:
73  *
74  * each swap partition or file is described by a "swapdev" structure.
75  * each "swapdev" structure contains a "swapent" structure which contains
76  * information that is passed up to the user (via system calls).
77  *
78  * each swap partition is assigned a "priority" (int) which controls
79  * swap parition usage.
80  *
81  * the system maintains a global data structure describing all swap
82  * partitions/files.   there is a sorted LIST of "swappri" structures
83  * which describe "swapdev"'s at that priority.   this LIST is headed
84  * by the "swap_priority" global var.    each "swappri" contains a
85  * TAILQ of "swapdev" structures at that priority.
86  *
87  * locking:
88  *  - swap_syscall_lock (krwlock_t): this lock serializes the swapctl
89  *    system call and prevents the swap priority list from changing
90  *    while we are in the middle of a system call (e.g. SWAP_STATS).
91  *  - uvm_swap_data_lock (kmutex_t): this lock protects all swap data
92  *    structures including the priority list, the swapdev structures,
93  *    and the swapmap arena.
94  *
95  * each swap device has the following info:
96  *  - swap device in use (could be disabled, preventing future use)
97  *  - swap enabled (allows new allocations on swap)
98  *  - map info in /dev/drum
99  *  - vnode pointer
100  * for swap files only:
101  *  - block size
102  *  - max byte count in buffer
103  *  - buffer
104  *
105  * userland controls and configures swap with the swapctl(2) system call.
106  * the sys_swapctl performs the following operations:
107  *  [1] SWAP_NSWAP: returns the number of swap devices currently configured
108  *  [2] SWAP_STATS: given a pointer to an array of swapent structures
109  *	(passed in via "arg") of a size passed in via "misc" ... we load
110  *	the current swap config into the array. The actual work is done
111  *	in the uvm_swap_stats() 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 	dev_t			swd_dev;	/* device id */
128 	int			swd_flags;	/* flags:inuse/enable/fake */
129 	int			swd_priority;	/* our priority */
130 	int			swd_nblks;	/* blocks in this device */
131 	char			*swd_path;	/* saved pathname of device */
132 	int			swd_pathlen;	/* length of pathname */
133 	int			swd_npages;	/* #pages we can use */
134 	int			swd_npginuse;	/* #pages in use */
135 	int			swd_npgbad;	/* #pages bad */
136 	int			swd_drumoffset;	/* page0 offset in drum */
137 	int			swd_drumsize;	/* #pages in drum */
138 	blist_t			swd_blist;	/* blist for this swapdev */
139 	struct vnode		*swd_vp;	/* backing vnode */
140 	TAILQ_ENTRY(swapdev)	swd_next;	/* priority tailq */
141 
142 	int			swd_bsize;	/* blocksize (bytes) */
143 	int			swd_maxactive;	/* max active i/o reqs */
144 	struct bufq_state	*swd_tab;	/* buffer list */
145 	int			swd_active;	/* number of active buffers */
146 };
147 
148 /*
149  * swap device priority entry; the list is kept sorted on `spi_priority'.
150  */
151 struct swappri {
152 	int			spi_priority;     /* priority */
153 	TAILQ_HEAD(spi_swapdev, swapdev)	spi_swapdev;
154 	/* tailq of swapdevs at this priority */
155 	LIST_ENTRY(swappri)	spi_swappri;      /* global list of pri's */
156 };
157 
158 /*
159  * The following two structures are used to keep track of data transfers
160  * on swap devices associated with regular files.
161  * NOTE: this code is more or less a copy of vnd.c; we use the same
162  * structure names here to ease porting..
163  */
164 struct vndxfer {
165 	struct buf	*vx_bp;		/* Pointer to parent buffer */
166 	struct swapdev	*vx_sdp;
167 	int		vx_error;
168 	int		vx_pending;	/* # of pending aux buffers */
169 	int		vx_flags;
170 #define VX_BUSY		1
171 #define VX_DEAD		2
172 };
173 
174 struct vndbuf {
175 	struct buf	vb_buf;
176 	struct vndxfer	*vb_xfer;
177 };
178 
179 /*
180  * We keep a of pool vndbuf's and vndxfer structures.
181  */
182 static struct pool vndxfer_pool, vndbuf_pool;
183 
184 /*
185  * local variables
186  */
187 static vmem_t *swapmap;	/* controls the mapping of /dev/drum */
188 
189 /* list of all active swap devices [by priority] */
190 LIST_HEAD(swap_priority, swappri);
191 static struct swap_priority swap_priority;
192 
193 /* locks */
194 static kmutex_t uvm_swap_data_lock __cacheline_aligned;
195 static krwlock_t swap_syscall_lock;
196 
197 /* workqueue and use counter for swap to regular files */
198 static int sw_reg_count = 0;
199 static struct workqueue *sw_reg_workqueue;
200 
201 /* tuneables */
202 u_int uvm_swapisfull_factor = 99;
203 
204 /*
205  * prototypes
206  */
207 static struct swapdev	*swapdrum_getsdp(int);
208 
209 static struct swapdev	*swaplist_find(struct vnode *, bool);
210 static void		 swaplist_insert(struct swapdev *,
211 					 struct swappri *, int);
212 static void		 swaplist_trim(void);
213 
214 static int swap_on(struct lwp *, struct swapdev *);
215 static int swap_off(struct lwp *, struct swapdev *);
216 
217 static void sw_reg_strategy(struct swapdev *, struct buf *, int);
218 static void sw_reg_biodone(struct buf *);
219 static void sw_reg_iodone(struct work *wk, void *dummy);
220 static void sw_reg_start(struct swapdev *);
221 
222 static int uvm_swap_io(struct vm_page **, int, int, int);
223 
224 /*
225  * uvm_swap_init: init the swap system data structures and locks
226  *
227  * => called at boot time from init_main.c after the filesystems
228  *	are brought up (which happens after uvm_init())
229  */
230 void
231 uvm_swap_init(void)
232 {
233 	UVMHIST_FUNC("uvm_swap_init");
234 
235 	UVMHIST_CALLED(pdhist);
236 	/*
237 	 * first, init the swap list, its counter, and its lock.
238 	 * then get a handle on the vnode for /dev/drum by using
239 	 * the its dev_t number ("swapdev", from MD conf.c).
240 	 */
241 
242 	LIST_INIT(&swap_priority);
243 	uvmexp.nswapdev = 0;
244 	rw_init(&swap_syscall_lock);
245 	mutex_init(&uvm_swap_data_lock, MUTEX_DEFAULT, IPL_NONE);
246 
247 	if (bdevvp(swapdev, &swapdev_vp))
248 		panic("%s: can't get vnode for swap device", __func__);
249 	if (vn_lock(swapdev_vp, LK_EXCLUSIVE | LK_RETRY))
250 		panic("%s: can't lock swap device", __func__);
251 	if (VOP_OPEN(swapdev_vp, FREAD | FWRITE, NOCRED))
252 		panic("%s: can't open swap device", __func__);
253 	VOP_UNLOCK(swapdev_vp);
254 
255 	/*
256 	 * create swap block resource map to map /dev/drum.   the range
257 	 * from 1 to INT_MAX allows 2 gigablocks of swap space.  note
258 	 * that block 0 is reserved (used to indicate an allocation
259 	 * failure, or no allocation).
260 	 */
261 	swapmap = vmem_create("swapmap", 1, INT_MAX - 1, 1, NULL, NULL, NULL, 0,
262 	    VM_NOSLEEP, IPL_NONE);
263 	if (swapmap == 0) {
264 		panic("%s: vmem_create failed", __func__);
265 	}
266 
267 	pool_init(&vndxfer_pool, sizeof(struct vndxfer), 0, 0, 0, "swp vnx",
268 	    NULL, IPL_BIO);
269 	pool_init(&vndbuf_pool, sizeof(struct vndbuf), 0, 0, 0, "swp vnd",
270 	    NULL, IPL_BIO);
271 
272 	UVMHIST_LOG(pdhist, "<- done", 0, 0, 0, 0);
273 }
274 
275 /*
276  * swaplist functions: functions that operate on the list of swap
277  * devices on the system.
278  */
279 
280 /*
281  * swaplist_insert: insert swap device "sdp" into the global list
282  *
283  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
284  * => caller must provide a newly allocated swappri structure (we will
285  *	FREE it if we don't need it... this it to prevent allocation
286  *	blocking here while adding swap)
287  */
288 static void
289 swaplist_insert(struct swapdev *sdp, struct swappri *newspp, int priority)
290 {
291 	struct swappri *spp, *pspp;
292 	UVMHIST_FUNC("swaplist_insert"); UVMHIST_CALLED(pdhist);
293 
294 	/*
295 	 * find entry at or after which to insert the new device.
296 	 */
297 	pspp = NULL;
298 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
299 		if (priority <= spp->spi_priority)
300 			break;
301 		pspp = spp;
302 	}
303 
304 	/*
305 	 * new priority?
306 	 */
307 	if (spp == NULL || spp->spi_priority != priority) {
308 		spp = newspp;  /* use newspp! */
309 		UVMHIST_LOG(pdhist, "created new swappri = %jd",
310 			    priority, 0, 0, 0);
311 
312 		spp->spi_priority = priority;
313 		TAILQ_INIT(&spp->spi_swapdev);
314 
315 		if (pspp)
316 			LIST_INSERT_AFTER(pspp, spp, spi_swappri);
317 		else
318 			LIST_INSERT_HEAD(&swap_priority, spp, spi_swappri);
319 	} else {
320 	  	/* we don't need a new priority structure, free it */
321 		kmem_free(newspp, sizeof(*newspp));
322 	}
323 
324 	/*
325 	 * priority found (or created).   now insert on the priority's
326 	 * tailq list and bump the total number of swapdevs.
327 	 */
328 	sdp->swd_priority = priority;
329 	TAILQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
330 	uvmexp.nswapdev++;
331 }
332 
333 /*
334  * swaplist_find: find and optionally remove a swap device from the
335  *	global list.
336  *
337  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
338  * => we return the swapdev we found (and removed)
339  */
340 static struct swapdev *
341 swaplist_find(struct vnode *vp, bool remove)
342 {
343 	struct swapdev *sdp;
344 	struct swappri *spp;
345 
346 	/*
347 	 * search the lists for the requested vp
348 	 */
349 
350 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
351 		TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
352 			if (sdp->swd_vp == vp) {
353 				if (remove) {
354 					TAILQ_REMOVE(&spp->spi_swapdev,
355 					    sdp, swd_next);
356 					uvmexp.nswapdev--;
357 				}
358 				return(sdp);
359 			}
360 		}
361 	}
362 	return (NULL);
363 }
364 
365 /*
366  * swaplist_trim: scan priority list for empty priority entries and kill
367  *	them.
368  *
369  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
370  */
371 static void
372 swaplist_trim(void)
373 {
374 	struct swappri *spp, *nextspp;
375 
376 	LIST_FOREACH_SAFE(spp, &swap_priority, spi_swappri, nextspp) {
377 		if (!TAILQ_EMPTY(&spp->spi_swapdev))
378 			continue;
379 		LIST_REMOVE(spp, spi_swappri);
380 		kmem_free(spp, sizeof(*spp));
381 	}
382 }
383 
384 /*
385  * swapdrum_getsdp: given a page offset in /dev/drum, convert it back
386  *	to the "swapdev" that maps that section of the drum.
387  *
388  * => each swapdev takes one big contig chunk of the drum
389  * => caller must hold uvm_swap_data_lock
390  */
391 static struct swapdev *
392 swapdrum_getsdp(int pgno)
393 {
394 	struct swapdev *sdp;
395 	struct swappri *spp;
396 
397 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
398 		TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
399 			if (sdp->swd_flags & SWF_FAKE)
400 				continue;
401 			if (pgno >= sdp->swd_drumoffset &&
402 			    pgno < (sdp->swd_drumoffset + sdp->swd_drumsize)) {
403 				return sdp;
404 			}
405 		}
406 	}
407 	return NULL;
408 }
409 
410 void swapsys_lock(krw_t op)
411 {
412 	rw_enter(&swap_syscall_lock, op);
413 }
414 
415 void swapsys_unlock(void)
416 {
417 	rw_exit(&swap_syscall_lock);
418 }
419 
420 static void
421 swapent_cvt(struct swapent *se, const struct swapdev *sdp, int inuse)
422 {
423 	se->se_dev = sdp->swd_dev;
424 	se->se_flags = sdp->swd_flags;
425 	se->se_nblks = sdp->swd_nblks;
426 	se->se_inuse = inuse;
427 	se->se_priority = sdp->swd_priority;
428 	KASSERT(sdp->swd_pathlen < sizeof(se->se_path));
429 	strcpy(se->se_path, sdp->swd_path);
430 }
431 
432 int (*uvm_swap_stats13)(const struct sys_swapctl_args *, register_t *) =
433     (void *)enosys;
434 int (*uvm_swap_stats50)(const struct sys_swapctl_args *, register_t *) =
435     (void *)enosys;
436 
437 /*
438  * sys_swapctl: main entry point for swapctl(2) system call
439  * 	[with two helper functions: swap_on and swap_off]
440  */
441 int
442 sys_swapctl(struct lwp *l, const struct sys_swapctl_args *uap, register_t *retval)
443 {
444 	/* {
445 		syscallarg(int) cmd;
446 		syscallarg(void *) arg;
447 		syscallarg(int) misc;
448 	} */
449 	struct vnode *vp;
450 	struct nameidata nd;
451 	struct swappri *spp;
452 	struct swapdev *sdp;
453 #define SWAP_PATH_MAX (PATH_MAX + 1)
454 	char	*userpath;
455 	size_t	len = 0;
456 	int	error;
457 	int	priority;
458 	UVMHIST_FUNC("sys_swapctl"); UVMHIST_CALLED(pdhist);
459 
460 	/*
461 	 * we handle the non-priv NSWAP and STATS request first.
462 	 *
463 	 * SWAP_NSWAP: return number of config'd swap devices
464 	 * [can also be obtained with uvmexp sysctl]
465 	 */
466 	if (SCARG(uap, cmd) == SWAP_NSWAP) {
467 		const int nswapdev = uvmexp.nswapdev;
468 		UVMHIST_LOG(pdhist, "<- done SWAP_NSWAP=%jd", nswapdev,
469 		    0, 0, 0);
470 		*retval = nswapdev;
471 		return 0;
472 	}
473 
474 	userpath = kmem_alloc(SWAP_PATH_MAX, KM_SLEEP);
475 
476 	/*
477 	 * ensure serialized syscall access by grabbing the swap_syscall_lock
478 	 */
479 	rw_enter(&swap_syscall_lock, RW_WRITER);
480 
481 	/*
482 	 * SWAP_STATS: get stats on current # of configured swap devs
483 	 *
484 	 * note that the swap_priority list can't change as long
485 	 * as we are holding the swap_syscall_lock.  we don't want
486 	 * to grab the uvm_swap_data_lock because we may fault&sleep during
487 	 * copyout() and we don't want to be holding that lock then!
488 	 */
489 	switch (SCARG(uap, cmd)) {
490 	case SWAP_STATS13:
491 		error = (*uvm_swap_stats13)(uap, retval);
492 		goto out;
493 	case SWAP_STATS50:
494 		error = (*uvm_swap_stats50)(uap, retval);
495 		goto out;
496 	case SWAP_STATS:
497 		error = uvm_swap_stats(SCARG(uap, arg), SCARG(uap, misc),
498 		    NULL, sizeof(struct swapent), retval);
499 		UVMHIST_LOG(pdhist, "<- done SWAP_STATS", 0, 0, 0, 0);
500 		goto out;
501 
502 	case SWAP_GETDUMPDEV:
503 		error = copyout(&dumpdev, SCARG(uap, arg), sizeof(dumpdev));
504 		goto out;
505 	default:
506 		break;
507 	}
508 
509 	/*
510 	 * all other requests require superuser privs.   verify.
511 	 */
512 	if ((error = kauth_authorize_system(l->l_cred, KAUTH_SYSTEM_SWAPCTL,
513 	    0, NULL, NULL, NULL)))
514 		goto out;
515 
516 	if (SCARG(uap, cmd) == SWAP_DUMPOFF) {
517 		/* drop the current dump device */
518 		dumpdev = NODEV;
519 		dumpcdev = NODEV;
520 		cpu_dumpconf();
521 		goto out;
522 	}
523 
524 	/*
525 	 * at this point we expect a path name in arg.   we will
526 	 * use namei() to gain a vnode reference (vref), and lock
527 	 * the vnode (VOP_LOCK).
528 	 *
529 	 * XXX: a NULL arg means use the root vnode pointer (e.g. for
530 	 * miniroot)
531 	 */
532 	if (SCARG(uap, arg) == NULL) {
533 		vp = rootvp;		/* miniroot */
534 		vref(vp);
535 		if (vn_lock(vp, LK_EXCLUSIVE)) {
536 			vrele(vp);
537 			error = EBUSY;
538 			goto out;
539 		}
540 		if (SCARG(uap, cmd) == SWAP_ON &&
541 		    copystr("miniroot", userpath, SWAP_PATH_MAX, &len))
542 			panic("swapctl: miniroot copy failed");
543 	} else {
544 		struct pathbuf *pb;
545 
546 		/*
547 		 * This used to allow copying in one extra byte
548 		 * (SWAP_PATH_MAX instead of PATH_MAX) for SWAP_ON.
549 		 * This was completely pointless because if anyone
550 		 * used that extra byte namei would fail with
551 		 * ENAMETOOLONG anyway, so I've removed the excess
552 		 * logic. - dholland 20100215
553 		 */
554 
555 		error = pathbuf_copyin(SCARG(uap, arg), &pb);
556 		if (error) {
557 			goto out;
558 		}
559 		if (SCARG(uap, cmd) == SWAP_ON) {
560 			/* get a copy of the string */
561 			pathbuf_copystring(pb, userpath, SWAP_PATH_MAX);
562 			len = strlen(userpath) + 1;
563 		}
564 		NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | TRYEMULROOT, pb);
565 		if ((error = namei(&nd))) {
566 			pathbuf_destroy(pb);
567 			goto out;
568 		}
569 		vp = nd.ni_vp;
570 		pathbuf_destroy(pb);
571 	}
572 	/* note: "vp" is referenced and locked */
573 
574 	error = 0;		/* assume no error */
575 	switch(SCARG(uap, cmd)) {
576 
577 	case SWAP_DUMPDEV:
578 		if (vp->v_type != VBLK) {
579 			error = ENOTBLK;
580 			break;
581 		}
582 		if (bdevsw_lookup(vp->v_rdev)) {
583 			dumpdev = vp->v_rdev;
584 			dumpcdev = devsw_blk2chr(dumpdev);
585 		} else
586 			dumpdev = NODEV;
587 		cpu_dumpconf();
588 		break;
589 
590 	case SWAP_CTL:
591 		/*
592 		 * get new priority, remove old entry (if any) and then
593 		 * reinsert it in the correct place.  finally, prune out
594 		 * any empty priority structures.
595 		 */
596 		priority = SCARG(uap, misc);
597 		spp = kmem_alloc(sizeof(*spp), KM_SLEEP);
598 		mutex_enter(&uvm_swap_data_lock);
599 		if ((sdp = swaplist_find(vp, true)) == NULL) {
600 			error = ENOENT;
601 		} else {
602 			swaplist_insert(sdp, spp, priority);
603 			swaplist_trim();
604 		}
605 		mutex_exit(&uvm_swap_data_lock);
606 		if (error)
607 			kmem_free(spp, sizeof(*spp));
608 		break;
609 
610 	case SWAP_ON:
611 
612 		/*
613 		 * check for duplicates.   if none found, then insert a
614 		 * dummy entry on the list to prevent someone else from
615 		 * trying to enable this device while we are working on
616 		 * it.
617 		 */
618 
619 		priority = SCARG(uap, misc);
620 		sdp = kmem_zalloc(sizeof(*sdp), KM_SLEEP);
621 		spp = kmem_alloc(sizeof(*spp), KM_SLEEP);
622 		sdp->swd_flags = SWF_FAKE;
623 		sdp->swd_vp = vp;
624 		sdp->swd_dev = (vp->v_type == VBLK) ? vp->v_rdev : NODEV;
625 		bufq_alloc(&sdp->swd_tab, "disksort", BUFQ_SORT_RAWBLOCK);
626 		mutex_enter(&uvm_swap_data_lock);
627 		if (swaplist_find(vp, false) != NULL) {
628 			error = EBUSY;
629 			mutex_exit(&uvm_swap_data_lock);
630 			bufq_free(sdp->swd_tab);
631 			kmem_free(sdp, sizeof(*sdp));
632 			kmem_free(spp, sizeof(*spp));
633 			break;
634 		}
635 		swaplist_insert(sdp, spp, priority);
636 		mutex_exit(&uvm_swap_data_lock);
637 
638 		KASSERT(len > 0);
639 		sdp->swd_pathlen = len;
640 		sdp->swd_path = kmem_alloc(len, KM_SLEEP);
641 		if (copystr(userpath, sdp->swd_path, len, 0) != 0)
642 			panic("swapctl: copystr");
643 
644 		/*
645 		 * we've now got a FAKE placeholder in the swap list.
646 		 * now attempt to enable swap on it.  if we fail, undo
647 		 * what we've done and kill the fake entry we just inserted.
648 		 * if swap_on is a success, it will clear the SWF_FAKE flag
649 		 */
650 
651 		if ((error = swap_on(l, sdp)) != 0) {
652 			mutex_enter(&uvm_swap_data_lock);
653 			(void) swaplist_find(vp, true);  /* kill fake entry */
654 			swaplist_trim();
655 			mutex_exit(&uvm_swap_data_lock);
656 			bufq_free(sdp->swd_tab);
657 			kmem_free(sdp->swd_path, sdp->swd_pathlen);
658 			kmem_free(sdp, sizeof(*sdp));
659 			break;
660 		}
661 		break;
662 
663 	case SWAP_OFF:
664 		mutex_enter(&uvm_swap_data_lock);
665 		if ((sdp = swaplist_find(vp, false)) == NULL) {
666 			mutex_exit(&uvm_swap_data_lock);
667 			error = ENXIO;
668 			break;
669 		}
670 
671 		/*
672 		 * If a device isn't in use or enabled, we
673 		 * can't stop swapping from it (again).
674 		 */
675 		if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0) {
676 			mutex_exit(&uvm_swap_data_lock);
677 			error = EBUSY;
678 			break;
679 		}
680 
681 		/*
682 		 * do the real work.
683 		 */
684 		error = swap_off(l, sdp);
685 		break;
686 
687 	default:
688 		error = EINVAL;
689 	}
690 
691 	/*
692 	 * done!  release the ref gained by namei() and unlock.
693 	 */
694 	vput(vp);
695 out:
696 	rw_exit(&swap_syscall_lock);
697 	kmem_free(userpath, SWAP_PATH_MAX);
698 
699 	UVMHIST_LOG(pdhist, "<- done!  error=%jd", error, 0, 0, 0);
700 	return (error);
701 }
702 
703 /*
704  * uvm_swap_stats: implements swapctl(SWAP_STATS). The function is kept
705  * away from sys_swapctl() in order to allow COMPAT_* swapctl()
706  * emulation to use it directly without going through sys_swapctl().
707  * The problem with using sys_swapctl() there is that it involves
708  * copying the swapent array to the stackgap, and this array's size
709  * is not known at build time. Hence it would not be possible to
710  * ensure it would fit in the stackgap in any case.
711  */
712 int
713 uvm_swap_stats(char *ptr, int misc,
714     void (*f)(void *, const struct swapent *), size_t len,
715     register_t *retval)
716 {
717 	struct swappri *spp;
718 	struct swapdev *sdp;
719 	struct swapent sep;
720 	int count = 0;
721 	int error;
722 
723 	KASSERT(len <= sizeof(sep));
724 	if (len == 0)
725 		return ENOSYS;
726 
727 	if (misc < 0)
728 		return EINVAL;
729 
730 	if (misc == 0 || uvmexp.nswapdev == 0)
731 		return 0;
732 
733 	/* Make sure userland cannot exhaust kernel memory */
734 	if ((size_t)misc > (size_t)uvmexp.nswapdev)
735 		misc = uvmexp.nswapdev;
736 
737 	KASSERT(rw_lock_held(&swap_syscall_lock));
738 
739 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
740 		TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
741 			int inuse;
742 
743 			if (misc-- <= 0)
744 				break;
745 
746 			inuse = btodb((uint64_t)sdp->swd_npginuse <<
747 			    PAGE_SHIFT);
748 
749 			memset(&sep, 0, sizeof(sep));
750 			swapent_cvt(&sep, sdp, inuse);
751 			if (f)
752 				(*f)(&sep, &sep);
753 			if ((error = copyout(&sep, ptr, len)) != 0)
754 				return error;
755 			ptr += len;
756 			count++;
757 		}
758 	}
759 	*retval = count;
760 	return 0;
761 }
762 
763 /*
764  * swap_on: attempt to enable a swapdev for swapping.   note that the
765  *	swapdev is already on the global list, but disabled (marked
766  *	SWF_FAKE).
767  *
768  * => we avoid the start of the disk (to protect disk labels)
769  * => we also avoid the miniroot, if we are swapping to root.
770  * => caller should leave uvm_swap_data_lock unlocked, we may lock it
771  *	if needed.
772  */
773 static int
774 swap_on(struct lwp *l, struct swapdev *sdp)
775 {
776 	struct vnode *vp;
777 	int error, npages, nblocks, size;
778 	long addr;
779 	vmem_addr_t result;
780 	struct vattr va;
781 	dev_t dev;
782 	UVMHIST_FUNC("swap_on"); UVMHIST_CALLED(pdhist);
783 
784 	/*
785 	 * we want to enable swapping on sdp.   the swd_vp contains
786 	 * the vnode we want (locked and ref'd), and the swd_dev
787 	 * contains the dev_t of the file, if it a block device.
788 	 */
789 
790 	vp = sdp->swd_vp;
791 	dev = sdp->swd_dev;
792 
793 	/*
794 	 * open the swap file (mostly useful for block device files to
795 	 * let device driver know what is up).
796 	 *
797 	 * we skip the open/close for root on swap because the root
798 	 * has already been opened when root was mounted (mountroot).
799 	 */
800 	if (vp != rootvp) {
801 		if ((error = VOP_OPEN(vp, FREAD|FWRITE, l->l_cred)))
802 			return (error);
803 	}
804 
805 	/* XXX this only works for block devices */
806 	UVMHIST_LOG(pdhist, "  dev=%jd, major(dev)=%jd", dev, major(dev), 0, 0);
807 
808 	/*
809 	 * we now need to determine the size of the swap area.   for
810 	 * block specials we can call the d_psize function.
811 	 * for normal files, we must stat [get attrs].
812 	 *
813 	 * we put the result in nblks.
814 	 * for normal files, we also want the filesystem block size
815 	 * (which we get with statfs).
816 	 */
817 	switch (vp->v_type) {
818 	case VBLK:
819 		if ((nblocks = bdev_size(dev)) == -1) {
820 			error = ENXIO;
821 			goto bad;
822 		}
823 		break;
824 
825 	case VREG:
826 		if ((error = VOP_GETATTR(vp, &va, l->l_cred)))
827 			goto bad;
828 		nblocks = (int)btodb(va.va_size);
829 		sdp->swd_bsize = 1 << vp->v_mount->mnt_fs_bshift;
830 		/*
831 		 * limit the max # of outstanding I/O requests we issue
832 		 * at any one time.   take it easy on NFS servers.
833 		 */
834 		if (vp->v_tag == VT_NFS)
835 			sdp->swd_maxactive = 2; /* XXX */
836 		else
837 			sdp->swd_maxactive = 8; /* XXX */
838 		break;
839 
840 	default:
841 		error = ENXIO;
842 		goto bad;
843 	}
844 
845 	/*
846 	 * save nblocks in a safe place and convert to pages.
847 	 */
848 
849 	sdp->swd_nblks = nblocks;
850 	npages = dbtob((uint64_t)nblocks) >> PAGE_SHIFT;
851 
852 	/*
853 	 * for block special files, we want to make sure that leave
854 	 * the disklabel and bootblocks alone, so we arrange to skip
855 	 * over them (arbitrarily choosing to skip PAGE_SIZE bytes).
856 	 * note that because of this the "size" can be less than the
857 	 * actual number of blocks on the device.
858 	 */
859 	if (vp->v_type == VBLK) {
860 		/* we use pages 1 to (size - 1) [inclusive] */
861 		size = npages - 1;
862 		addr = 1;
863 	} else {
864 		/* we use pages 0 to (size - 1) [inclusive] */
865 		size = npages;
866 		addr = 0;
867 	}
868 
869 	/*
870 	 * make sure we have enough blocks for a reasonable sized swap
871 	 * area.   we want at least one page.
872 	 */
873 
874 	if (size < 1) {
875 		UVMHIST_LOG(pdhist, "  size <= 1!!", 0, 0, 0, 0);
876 		error = EINVAL;
877 		goto bad;
878 	}
879 
880 	UVMHIST_LOG(pdhist, "  dev=%jx: size=%jd addr=%jd", dev, size, addr, 0);
881 
882 	/*
883 	 * now we need to allocate an extent to manage this swap device
884 	 */
885 
886 	sdp->swd_blist = blist_create(npages);
887 	/* mark all expect the `saved' region free. */
888 	blist_free(sdp->swd_blist, addr, size);
889 
890 	/*
891 	 * if the vnode we are swapping to is the root vnode
892 	 * (i.e. we are swapping to the miniroot) then we want
893 	 * to make sure we don't overwrite it.   do a statfs to
894 	 * find its size and skip over it.
895 	 */
896 	if (vp == rootvp) {
897 		struct mount *mp;
898 		struct statvfs *sp;
899 		int rootblocks, rootpages;
900 
901 		mp = rootvnode->v_mount;
902 		sp = &mp->mnt_stat;
903 		rootblocks = sp->f_blocks * btodb(sp->f_frsize);
904 		/*
905 		 * XXX: sp->f_blocks isn't the total number of
906 		 * blocks in the filesystem, it's the number of
907 		 * data blocks.  so, our rootblocks almost
908 		 * definitely underestimates the total size
909 		 * of the filesystem - how badly depends on the
910 		 * details of the filesystem type.  there isn't
911 		 * an obvious way to deal with this cleanly
912 		 * and perfectly, so for now we just pad our
913 		 * rootblocks estimate with an extra 5 percent.
914 		 */
915 		rootblocks += (rootblocks >> 5) +
916 			(rootblocks >> 6) +
917 			(rootblocks >> 7);
918 		rootpages = round_page(dbtob(rootblocks)) >> PAGE_SHIFT;
919 		if (rootpages > size)
920 			panic("swap_on: miniroot larger than swap?");
921 
922 		if (rootpages != blist_fill(sdp->swd_blist, addr, rootpages)) {
923 			panic("swap_on: unable to preserve miniroot");
924 		}
925 
926 		size -= rootpages;
927 		printf("Preserved %d pages of miniroot ", rootpages);
928 		printf("leaving %d pages of swap\n", size);
929 	}
930 
931 	/*
932 	 * add a ref to vp to reflect usage as a swap device.
933 	 */
934 	vref(vp);
935 
936 	/*
937 	 * now add the new swapdev to the drum and enable.
938 	 */
939 	error = vmem_alloc(swapmap, npages, VM_BESTFIT | VM_SLEEP, &result);
940 	if (error != 0)
941 		panic("swapdrum_add");
942 	/*
943 	 * If this is the first regular swap create the workqueue.
944 	 * => Protected by swap_syscall_lock.
945 	 */
946 	if (vp->v_type != VBLK) {
947 		if (sw_reg_count++ == 0) {
948 			KASSERT(sw_reg_workqueue == NULL);
949 			if (workqueue_create(&sw_reg_workqueue, "swapiod",
950 			    sw_reg_iodone, NULL, PRIBIO, IPL_BIO, 0) != 0)
951 				panic("%s: workqueue_create failed", __func__);
952 		}
953 	}
954 
955 	sdp->swd_drumoffset = (int)result;
956 	sdp->swd_drumsize = npages;
957 	sdp->swd_npages = size;
958 	mutex_enter(&uvm_swap_data_lock);
959 	sdp->swd_flags &= ~SWF_FAKE;	/* going live */
960 	sdp->swd_flags |= (SWF_INUSE|SWF_ENABLE);
961 	uvmexp.swpages += size;
962 	uvmexp.swpgavail += size;
963 	mutex_exit(&uvm_swap_data_lock);
964 	return (0);
965 
966 	/*
967 	 * failure: clean up and return error.
968 	 */
969 
970 bad:
971 	if (sdp->swd_blist) {
972 		blist_destroy(sdp->swd_blist);
973 	}
974 	if (vp != rootvp) {
975 		(void)VOP_CLOSE(vp, FREAD|FWRITE, l->l_cred);
976 	}
977 	return (error);
978 }
979 
980 /*
981  * swap_off: stop swapping on swapdev
982  *
983  * => swap data should be locked, we will unlock.
984  */
985 static int
986 swap_off(struct lwp *l, struct swapdev *sdp)
987 {
988 	int npages = sdp->swd_npages;
989 	int error = 0;
990 
991 	UVMHIST_FUNC("swap_off"); UVMHIST_CALLED(pdhist);
992 	UVMHIST_LOG(pdhist, "  dev=%jx, npages=%jd", sdp->swd_dev,npages, 0, 0);
993 
994 	/* disable the swap area being removed */
995 	sdp->swd_flags &= ~SWF_ENABLE;
996 	uvmexp.swpgavail -= npages;
997 	mutex_exit(&uvm_swap_data_lock);
998 
999 	/*
1000 	 * the idea is to find all the pages that are paged out to this
1001 	 * device, and page them all in.  in uvm, swap-backed pageable
1002 	 * memory can take two forms: aobjs and anons.  call the
1003 	 * swapoff hook for each subsystem to bring in pages.
1004 	 */
1005 
1006 	if (uao_swap_off(sdp->swd_drumoffset,
1007 			 sdp->swd_drumoffset + sdp->swd_drumsize) ||
1008 	    amap_swap_off(sdp->swd_drumoffset,
1009 			  sdp->swd_drumoffset + sdp->swd_drumsize)) {
1010 		error = ENOMEM;
1011 	} else if (sdp->swd_npginuse > sdp->swd_npgbad) {
1012 		error = EBUSY;
1013 	}
1014 
1015 	if (error) {
1016 		mutex_enter(&uvm_swap_data_lock);
1017 		sdp->swd_flags |= SWF_ENABLE;
1018 		uvmexp.swpgavail += npages;
1019 		mutex_exit(&uvm_swap_data_lock);
1020 
1021 		return error;
1022 	}
1023 
1024 	/*
1025 	 * If this is the last regular swap destroy the workqueue.
1026 	 * => Protected by swap_syscall_lock.
1027 	 */
1028 	if (sdp->swd_vp->v_type != VBLK) {
1029 		KASSERT(sw_reg_count > 0);
1030 		KASSERT(sw_reg_workqueue != NULL);
1031 		if (--sw_reg_count == 0) {
1032 			workqueue_destroy(sw_reg_workqueue);
1033 			sw_reg_workqueue = NULL;
1034 		}
1035 	}
1036 
1037 	/*
1038 	 * done with the vnode.
1039 	 * drop our ref on the vnode before calling VOP_CLOSE()
1040 	 * so that spec_close() can tell if this is the last close.
1041 	 */
1042 	vrele(sdp->swd_vp);
1043 	if (sdp->swd_vp != rootvp) {
1044 		(void) VOP_CLOSE(sdp->swd_vp, FREAD|FWRITE, l->l_cred);
1045 	}
1046 
1047 	mutex_enter(&uvm_swap_data_lock);
1048 	uvmexp.swpages -= npages;
1049 	uvmexp.swpginuse -= sdp->swd_npgbad;
1050 
1051 	if (swaplist_find(sdp->swd_vp, true) == NULL)
1052 		panic("%s: swapdev not in list", __func__);
1053 	swaplist_trim();
1054 	mutex_exit(&uvm_swap_data_lock);
1055 
1056 	/*
1057 	 * free all resources!
1058 	 */
1059 	vmem_free(swapmap, sdp->swd_drumoffset, sdp->swd_drumsize);
1060 	blist_destroy(sdp->swd_blist);
1061 	bufq_free(sdp->swd_tab);
1062 	kmem_free(sdp, sizeof(*sdp));
1063 	return (0);
1064 }
1065 
1066 void
1067 uvm_swap_shutdown(struct lwp *l)
1068 {
1069 	struct swapdev *sdp;
1070 	struct swappri *spp;
1071 	struct vnode *vp;
1072 	int error;
1073 
1074 	printf("turning off swap...");
1075 	rw_enter(&swap_syscall_lock, RW_WRITER);
1076 	mutex_enter(&uvm_swap_data_lock);
1077 again:
1078 	LIST_FOREACH(spp, &swap_priority, spi_swappri)
1079 		TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
1080 			if (sdp->swd_flags & SWF_FAKE)
1081 				continue;
1082 			if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0)
1083 				continue;
1084 #ifdef DEBUG
1085 			printf("\nturning off swap on %s...",
1086 			    sdp->swd_path);
1087 #endif
1088 			if (vn_lock(vp = sdp->swd_vp, LK_EXCLUSIVE)) {
1089 				error = EBUSY;
1090 				vp = NULL;
1091 			} else
1092 				error = 0;
1093 			if (!error) {
1094 				error = swap_off(l, sdp);
1095 				mutex_enter(&uvm_swap_data_lock);
1096 			}
1097 			if (error) {
1098 				printf("stopping swap on %s failed "
1099 				    "with error %d\n", sdp->swd_path, error);
1100 				TAILQ_REMOVE(&spp->spi_swapdev, sdp,
1101 				    swd_next);
1102 				uvmexp.nswapdev--;
1103 				swaplist_trim();
1104 				if (vp)
1105 					vput(vp);
1106 			}
1107 			goto again;
1108 		}
1109 	printf(" done\n");
1110 	mutex_exit(&uvm_swap_data_lock);
1111 	rw_exit(&swap_syscall_lock);
1112 }
1113 
1114 
1115 /*
1116  * /dev/drum interface and i/o functions
1117  */
1118 
1119 /*
1120  * swstrategy: perform I/O on the drum
1121  *
1122  * => we must map the i/o request from the drum to the correct swapdev.
1123  */
1124 static void
1125 swstrategy(struct buf *bp)
1126 {
1127 	struct swapdev *sdp;
1128 	struct vnode *vp;
1129 	int pageno, bn;
1130 	UVMHIST_FUNC("swstrategy"); UVMHIST_CALLED(pdhist);
1131 
1132 	/*
1133 	 * convert block number to swapdev.   note that swapdev can't
1134 	 * be yanked out from under us because we are holding resources
1135 	 * in it (i.e. the blocks we are doing I/O on).
1136 	 */
1137 	pageno = dbtob((int64_t)bp->b_blkno) >> PAGE_SHIFT;
1138 	mutex_enter(&uvm_swap_data_lock);
1139 	sdp = swapdrum_getsdp(pageno);
1140 	mutex_exit(&uvm_swap_data_lock);
1141 	if (sdp == NULL) {
1142 		bp->b_error = EINVAL;
1143 		bp->b_resid = bp->b_bcount;
1144 		biodone(bp);
1145 		UVMHIST_LOG(pdhist, "  failed to get swap device", 0, 0, 0, 0);
1146 		return;
1147 	}
1148 
1149 	/*
1150 	 * convert drum page number to block number on this swapdev.
1151 	 */
1152 
1153 	pageno -= sdp->swd_drumoffset;	/* page # on swapdev */
1154 	bn = btodb((uint64_t)pageno << PAGE_SHIFT); /* convert to diskblock */
1155 
1156 	UVMHIST_LOG(pdhist, "  Rd/Wr (0/1) %jd: mapoff=%jx bn=%jx bcount=%jd",
1157 		((bp->b_flags & B_READ) == 0) ? 1 : 0,
1158 		sdp->swd_drumoffset, bn, bp->b_bcount);
1159 
1160 	/*
1161 	 * for block devices we finish up here.
1162 	 * for regular files we have to do more work which we delegate
1163 	 * to sw_reg_strategy().
1164 	 */
1165 
1166 	vp = sdp->swd_vp;		/* swapdev vnode pointer */
1167 	switch (vp->v_type) {
1168 	default:
1169 		panic("%s: vnode type 0x%x", __func__, vp->v_type);
1170 
1171 	case VBLK:
1172 
1173 		/*
1174 		 * must convert "bp" from an I/O on /dev/drum to an I/O
1175 		 * on the swapdev (sdp).
1176 		 */
1177 		bp->b_blkno = bn;		/* swapdev block number */
1178 		bp->b_dev = sdp->swd_dev;	/* swapdev dev_t */
1179 
1180 		/*
1181 		 * if we are doing a write, we have to redirect the i/o on
1182 		 * drum's v_numoutput counter to the swapdevs.
1183 		 */
1184 		if ((bp->b_flags & B_READ) == 0) {
1185 			mutex_enter(bp->b_objlock);
1186 			vwakeup(bp);	/* kills one 'v_numoutput' on drum */
1187 			mutex_exit(bp->b_objlock);
1188 			mutex_enter(vp->v_interlock);
1189 			vp->v_numoutput++;	/* put it on swapdev */
1190 			mutex_exit(vp->v_interlock);
1191 		}
1192 
1193 		/*
1194 		 * finally plug in swapdev vnode and start I/O
1195 		 */
1196 		bp->b_vp = vp;
1197 		bp->b_objlock = vp->v_interlock;
1198 		VOP_STRATEGY(vp, bp);
1199 		return;
1200 
1201 	case VREG:
1202 		/*
1203 		 * delegate to sw_reg_strategy function.
1204 		 */
1205 		sw_reg_strategy(sdp, bp, bn);
1206 		return;
1207 	}
1208 	/* NOTREACHED */
1209 }
1210 
1211 /*
1212  * swread: the read function for the drum (just a call to physio)
1213  */
1214 /*ARGSUSED*/
1215 static int
1216 swread(dev_t dev, struct uio *uio, int ioflag)
1217 {
1218 	UVMHIST_FUNC("swread"); UVMHIST_CALLED(pdhist);
1219 
1220 	UVMHIST_LOG(pdhist, "  dev=%jx offset=%jx", dev, uio->uio_offset, 0, 0);
1221 	return (physio(swstrategy, NULL, dev, B_READ, minphys, uio));
1222 }
1223 
1224 /*
1225  * swwrite: the write function for the drum (just a call to physio)
1226  */
1227 /*ARGSUSED*/
1228 static int
1229 swwrite(dev_t dev, struct uio *uio, int ioflag)
1230 {
1231 	UVMHIST_FUNC("swwrite"); UVMHIST_CALLED(pdhist);
1232 
1233 	UVMHIST_LOG(pdhist, "  dev=%jx offset=%jx", dev, uio->uio_offset, 0, 0);
1234 	return (physio(swstrategy, NULL, dev, B_WRITE, minphys, uio));
1235 }
1236 
1237 const struct bdevsw swap_bdevsw = {
1238 	.d_open = nullopen,
1239 	.d_close = nullclose,
1240 	.d_strategy = swstrategy,
1241 	.d_ioctl = noioctl,
1242 	.d_dump = nodump,
1243 	.d_psize = nosize,
1244 	.d_discard = nodiscard,
1245 	.d_flag = D_OTHER
1246 };
1247 
1248 const struct cdevsw swap_cdevsw = {
1249 	.d_open = nullopen,
1250 	.d_close = nullclose,
1251 	.d_read = swread,
1252 	.d_write = swwrite,
1253 	.d_ioctl = noioctl,
1254 	.d_stop = nostop,
1255 	.d_tty = notty,
1256 	.d_poll = nopoll,
1257 	.d_mmap = nommap,
1258 	.d_kqfilter = nokqfilter,
1259 	.d_discard = nodiscard,
1260 	.d_flag = D_OTHER,
1261 };
1262 
1263 /*
1264  * sw_reg_strategy: handle swap i/o to regular files
1265  */
1266 static void
1267 sw_reg_strategy(struct swapdev *sdp, struct buf *bp, int bn)
1268 {
1269 	struct vnode	*vp;
1270 	struct vndxfer	*vnx;
1271 	daddr_t		nbn;
1272 	char 		*addr;
1273 	off_t		byteoff;
1274 	int		s, off, nra, error, sz, resid;
1275 	UVMHIST_FUNC("sw_reg_strategy"); UVMHIST_CALLED(pdhist);
1276 
1277 	/*
1278 	 * allocate a vndxfer head for this transfer and point it to
1279 	 * our buffer.
1280 	 */
1281 	vnx = pool_get(&vndxfer_pool, PR_WAITOK);
1282 	vnx->vx_flags = VX_BUSY;
1283 	vnx->vx_error = 0;
1284 	vnx->vx_pending = 0;
1285 	vnx->vx_bp = bp;
1286 	vnx->vx_sdp = sdp;
1287 
1288 	/*
1289 	 * setup for main loop where we read filesystem blocks into
1290 	 * our buffer.
1291 	 */
1292 	error = 0;
1293 	bp->b_resid = bp->b_bcount;	/* nothing transferred yet! */
1294 	addr = bp->b_data;		/* current position in buffer */
1295 	byteoff = dbtob((uint64_t)bn);
1296 
1297 	for (resid = bp->b_resid; resid; resid -= sz) {
1298 		struct vndbuf	*nbp;
1299 
1300 		/*
1301 		 * translate byteoffset into block number.  return values:
1302 		 *   vp = vnode of underlying device
1303 		 *  nbn = new block number (on underlying vnode dev)
1304 		 *  nra = num blocks we can read-ahead (excludes requested
1305 		 *	block)
1306 		 */
1307 		nra = 0;
1308 		error = VOP_BMAP(sdp->swd_vp, byteoff / sdp->swd_bsize,
1309 				 	&vp, &nbn, &nra);
1310 
1311 		if (error == 0 && nbn == (daddr_t)-1) {
1312 			/*
1313 			 * this used to just set error, but that doesn't
1314 			 * do the right thing.  Instead, it causes random
1315 			 * memory errors.  The panic() should remain until
1316 			 * this condition doesn't destabilize the system.
1317 			 */
1318 #if 1
1319 			panic("%s: swap to sparse file", __func__);
1320 #else
1321 			error = EIO;	/* failure */
1322 #endif
1323 		}
1324 
1325 		/*
1326 		 * punt if there was an error or a hole in the file.
1327 		 * we must wait for any i/o ops we have already started
1328 		 * to finish before returning.
1329 		 *
1330 		 * XXX we could deal with holes here but it would be
1331 		 * a hassle (in the write case).
1332 		 */
1333 		if (error) {
1334 			s = splbio();
1335 			vnx->vx_error = error;	/* pass error up */
1336 			goto out;
1337 		}
1338 
1339 		/*
1340 		 * compute the size ("sz") of this transfer (in bytes).
1341 		 */
1342 		off = byteoff % sdp->swd_bsize;
1343 		sz = (1 + nra) * sdp->swd_bsize - off;
1344 		if (sz > resid)
1345 			sz = resid;
1346 
1347 		UVMHIST_LOG(pdhist, "sw_reg_strategy: "
1348 		    "vp %#jx/%#jx offset 0x%jx/0x%jx",
1349 		    (uintptr_t)sdp->swd_vp, (uintptr_t)vp, byteoff, nbn);
1350 
1351 		/*
1352 		 * now get a buf structure.   note that the vb_buf is
1353 		 * at the front of the nbp structure so that you can
1354 		 * cast pointers between the two structure easily.
1355 		 */
1356 		nbp = pool_get(&vndbuf_pool, PR_WAITOK);
1357 		buf_init(&nbp->vb_buf);
1358 		nbp->vb_buf.b_flags    = bp->b_flags;
1359 		nbp->vb_buf.b_cflags   = bp->b_cflags;
1360 		nbp->vb_buf.b_oflags   = bp->b_oflags;
1361 		nbp->vb_buf.b_bcount   = sz;
1362 		nbp->vb_buf.b_bufsize  = sz;
1363 		nbp->vb_buf.b_error    = 0;
1364 		nbp->vb_buf.b_data     = addr;
1365 		nbp->vb_buf.b_lblkno   = 0;
1366 		nbp->vb_buf.b_blkno    = nbn + btodb(off);
1367 		nbp->vb_buf.b_rawblkno = nbp->vb_buf.b_blkno;
1368 		nbp->vb_buf.b_iodone   = sw_reg_biodone;
1369 		nbp->vb_buf.b_vp       = vp;
1370 		nbp->vb_buf.b_objlock  = vp->v_interlock;
1371 		if (vp->v_type == VBLK) {
1372 			nbp->vb_buf.b_dev = vp->v_rdev;
1373 		}
1374 
1375 		nbp->vb_xfer = vnx;	/* patch it back in to vnx */
1376 
1377 		/*
1378 		 * Just sort by block number
1379 		 */
1380 		s = splbio();
1381 		if (vnx->vx_error != 0) {
1382 			buf_destroy(&nbp->vb_buf);
1383 			pool_put(&vndbuf_pool, nbp);
1384 			goto out;
1385 		}
1386 		vnx->vx_pending++;
1387 
1388 		/* sort it in and start I/O if we are not over our limit */
1389 		/* XXXAD locking */
1390 		bufq_put(sdp->swd_tab, &nbp->vb_buf);
1391 		sw_reg_start(sdp);
1392 		splx(s);
1393 
1394 		/*
1395 		 * advance to the next I/O
1396 		 */
1397 		byteoff += sz;
1398 		addr += sz;
1399 	}
1400 
1401 	s = splbio();
1402 
1403 out: /* Arrive here at splbio */
1404 	vnx->vx_flags &= ~VX_BUSY;
1405 	if (vnx->vx_pending == 0) {
1406 		error = vnx->vx_error;
1407 		pool_put(&vndxfer_pool, vnx);
1408 		bp->b_error = error;
1409 		biodone(bp);
1410 	}
1411 	splx(s);
1412 }
1413 
1414 /*
1415  * sw_reg_start: start an I/O request on the requested swapdev
1416  *
1417  * => reqs are sorted by b_rawblkno (above)
1418  */
1419 static void
1420 sw_reg_start(struct swapdev *sdp)
1421 {
1422 	struct buf	*bp;
1423 	struct vnode	*vp;
1424 	UVMHIST_FUNC("sw_reg_start"); UVMHIST_CALLED(pdhist);
1425 
1426 	/* recursion control */
1427 	if ((sdp->swd_flags & SWF_BUSY) != 0)
1428 		return;
1429 
1430 	sdp->swd_flags |= SWF_BUSY;
1431 
1432 	while (sdp->swd_active < sdp->swd_maxactive) {
1433 		bp = bufq_get(sdp->swd_tab);
1434 		if (bp == NULL)
1435 			break;
1436 		sdp->swd_active++;
1437 
1438 		UVMHIST_LOG(pdhist,
1439 		    "sw_reg_start:  bp %#jx vp %#jx blkno %#jx cnt %jx",
1440 		    (uintptr_t)bp, (uintptr_t)bp->b_vp, (uintptr_t)bp->b_blkno,
1441 		    bp->b_bcount);
1442 		vp = bp->b_vp;
1443 		KASSERT(bp->b_objlock == vp->v_interlock);
1444 		if ((bp->b_flags & B_READ) == 0) {
1445 			mutex_enter(vp->v_interlock);
1446 			vp->v_numoutput++;
1447 			mutex_exit(vp->v_interlock);
1448 		}
1449 		VOP_STRATEGY(vp, bp);
1450 	}
1451 	sdp->swd_flags &= ~SWF_BUSY;
1452 }
1453 
1454 /*
1455  * sw_reg_biodone: one of our i/o's has completed
1456  */
1457 static void
1458 sw_reg_biodone(struct buf *bp)
1459 {
1460 	workqueue_enqueue(sw_reg_workqueue, &bp->b_work, NULL);
1461 }
1462 
1463 /*
1464  * sw_reg_iodone: one of our i/o's has completed and needs post-i/o cleanup
1465  *
1466  * => note that we can recover the vndbuf struct by casting the buf ptr
1467  */
1468 static void
1469 sw_reg_iodone(struct work *wk, void *dummy)
1470 {
1471 	struct vndbuf *vbp = (void *)wk;
1472 	struct vndxfer *vnx = vbp->vb_xfer;
1473 	struct buf *pbp = vnx->vx_bp;		/* parent buffer */
1474 	struct swapdev	*sdp = vnx->vx_sdp;
1475 	int s, resid, error;
1476 	KASSERT(&vbp->vb_buf.b_work == wk);
1477 	UVMHIST_FUNC("sw_reg_iodone"); UVMHIST_CALLED(pdhist);
1478 
1479 	UVMHIST_LOG(pdhist, "  vbp=%#jx vp=%#jx blkno=%jx addr=%#jx",
1480 	    (uintptr_t)vbp, (uintptr_t)vbp->vb_buf.b_vp, vbp->vb_buf.b_blkno,
1481 	    (uintptr_t)vbp->vb_buf.b_data);
1482 	UVMHIST_LOG(pdhist, "  cnt=%jx resid=%jx",
1483 	    vbp->vb_buf.b_bcount, vbp->vb_buf.b_resid, 0, 0);
1484 
1485 	/*
1486 	 * protect vbp at splbio and update.
1487 	 */
1488 
1489 	s = splbio();
1490 	resid = vbp->vb_buf.b_bcount - vbp->vb_buf.b_resid;
1491 	pbp->b_resid -= resid;
1492 	vnx->vx_pending--;
1493 
1494 	if (vbp->vb_buf.b_error != 0) {
1495 		/* pass error upward */
1496 		error = vbp->vb_buf.b_error ? vbp->vb_buf.b_error : EIO;
1497 		UVMHIST_LOG(pdhist, "  got error=%jd !", error, 0, 0, 0);
1498 		vnx->vx_error = error;
1499 	}
1500 
1501 	/*
1502 	 * kill vbp structure
1503 	 */
1504 	buf_destroy(&vbp->vb_buf);
1505 	pool_put(&vndbuf_pool, vbp);
1506 
1507 	/*
1508 	 * wrap up this transaction if it has run to completion or, in
1509 	 * case of an error, when all auxiliary buffers have returned.
1510 	 */
1511 	if (vnx->vx_error != 0) {
1512 		/* pass error upward */
1513 		error = vnx->vx_error;
1514 		if ((vnx->vx_flags & VX_BUSY) == 0 && vnx->vx_pending == 0) {
1515 			pbp->b_error = error;
1516 			biodone(pbp);
1517 			pool_put(&vndxfer_pool, vnx);
1518 		}
1519 	} else if (pbp->b_resid == 0) {
1520 		KASSERT(vnx->vx_pending == 0);
1521 		if ((vnx->vx_flags & VX_BUSY) == 0) {
1522 			UVMHIST_LOG(pdhist, "  iodone, pbp=%#jx error=%jd !",
1523 			    (uintptr_t)pbp, vnx->vx_error, 0, 0);
1524 			biodone(pbp);
1525 			pool_put(&vndxfer_pool, vnx);
1526 		}
1527 	}
1528 
1529 	/*
1530 	 * done!   start next swapdev I/O if one is pending
1531 	 */
1532 	sdp->swd_active--;
1533 	sw_reg_start(sdp);
1534 	splx(s);
1535 }
1536 
1537 
1538 /*
1539  * uvm_swap_alloc: allocate space on swap
1540  *
1541  * => allocation is done "round robin" down the priority list, as we
1542  *	allocate in a priority we "rotate" the circle queue.
1543  * => space can be freed with uvm_swap_free
1544  * => we return the page slot number in /dev/drum (0 == invalid slot)
1545  * => we lock uvm_swap_data_lock
1546  * => XXXMRG: "LESSOK" INTERFACE NEEDED TO EXTENT SYSTEM
1547  */
1548 int
1549 uvm_swap_alloc(int *nslots /* IN/OUT */, bool lessok)
1550 {
1551 	struct swapdev *sdp;
1552 	struct swappri *spp;
1553 	UVMHIST_FUNC("uvm_swap_alloc"); UVMHIST_CALLED(pdhist);
1554 
1555 	/*
1556 	 * no swap devices configured yet?   definite failure.
1557 	 */
1558 	if (uvmexp.nswapdev < 1)
1559 		return 0;
1560 
1561 	/*
1562 	 * XXXJAK: BEGIN HACK
1563 	 *
1564 	 * blist_alloc() in subr_blist.c will panic if we try to allocate
1565 	 * too many slots.
1566 	 */
1567 	if (*nslots > BLIST_MAX_ALLOC) {
1568 		if (__predict_false(lessok == false))
1569 			return 0;
1570 		*nslots = BLIST_MAX_ALLOC;
1571 	}
1572 	/* XXXJAK: END HACK */
1573 
1574 	/*
1575 	 * lock data lock, convert slots into blocks, and enter loop
1576 	 */
1577 	mutex_enter(&uvm_swap_data_lock);
1578 
1579 ReTry:	/* XXXMRG */
1580 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
1581 		TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
1582 			uint64_t result;
1583 
1584 			/* if it's not enabled, then we can't swap from it */
1585 			if ((sdp->swd_flags & SWF_ENABLE) == 0)
1586 				continue;
1587 			if (sdp->swd_npginuse + *nslots > sdp->swd_npages)
1588 				continue;
1589 			result = blist_alloc(sdp->swd_blist, *nslots);
1590 			if (result == BLIST_NONE) {
1591 				continue;
1592 			}
1593 			KASSERT(result < sdp->swd_drumsize);
1594 
1595 			/*
1596 			 * successful allocation!  now rotate the tailq.
1597 			 */
1598 			TAILQ_REMOVE(&spp->spi_swapdev, sdp, swd_next);
1599 			TAILQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
1600 			sdp->swd_npginuse += *nslots;
1601 			uvmexp.swpginuse += *nslots;
1602 			mutex_exit(&uvm_swap_data_lock);
1603 			/* done!  return drum slot number */
1604 			UVMHIST_LOG(pdhist,
1605 			    "success!  returning %jd slots starting at %jd",
1606 			    *nslots, result + sdp->swd_drumoffset, 0, 0);
1607 			return (result + sdp->swd_drumoffset);
1608 		}
1609 	}
1610 
1611 	/* XXXMRG: BEGIN HACK */
1612 	if (*nslots > 1 && lessok) {
1613 		*nslots = 1;
1614 		/* XXXMRG: ugh!  blist should support this for us */
1615 		goto ReTry;
1616 	}
1617 	/* XXXMRG: END HACK */
1618 
1619 	mutex_exit(&uvm_swap_data_lock);
1620 	return 0;
1621 }
1622 
1623 /*
1624  * uvm_swapisfull: return true if most of available swap is allocated
1625  * and in use.  we don't count some small portion as it may be inaccessible
1626  * to us at any given moment, for example if there is lock contention or if
1627  * pages are busy.
1628  */
1629 bool
1630 uvm_swapisfull(void)
1631 {
1632 	int swpgonly;
1633 	bool rv;
1634 
1635 	mutex_enter(&uvm_swap_data_lock);
1636 	KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1637 	swpgonly = (int)((uint64_t)uvmexp.swpgonly * 100 /
1638 	    uvm_swapisfull_factor);
1639 	rv = (swpgonly >= uvmexp.swpgavail);
1640 	mutex_exit(&uvm_swap_data_lock);
1641 
1642 	return (rv);
1643 }
1644 
1645 /*
1646  * uvm_swap_markbad: keep track of swap ranges where we've had i/o errors
1647  *
1648  * => we lock uvm_swap_data_lock
1649  */
1650 void
1651 uvm_swap_markbad(int startslot, int nslots)
1652 {
1653 	struct swapdev *sdp;
1654 	UVMHIST_FUNC("uvm_swap_markbad"); UVMHIST_CALLED(pdhist);
1655 
1656 	mutex_enter(&uvm_swap_data_lock);
1657 	sdp = swapdrum_getsdp(startslot);
1658 	KASSERT(sdp != NULL);
1659 
1660 	/*
1661 	 * we just keep track of how many pages have been marked bad
1662 	 * in this device, to make everything add up in swap_off().
1663 	 * we assume here that the range of slots will all be within
1664 	 * one swap device.
1665 	 */
1666 
1667 	KASSERT(uvmexp.swpgonly >= nslots);
1668 	atomic_add_int(&uvmexp.swpgonly, -nslots);
1669 	sdp->swd_npgbad += nslots;
1670 	UVMHIST_LOG(pdhist, "now %jd bad", sdp->swd_npgbad, 0,0,0);
1671 	mutex_exit(&uvm_swap_data_lock);
1672 }
1673 
1674 /*
1675  * uvm_swap_free: free swap slots
1676  *
1677  * => this can be all or part of an allocation made by uvm_swap_alloc
1678  * => we lock uvm_swap_data_lock
1679  */
1680 void
1681 uvm_swap_free(int startslot, int nslots)
1682 {
1683 	struct swapdev *sdp;
1684 	UVMHIST_FUNC("uvm_swap_free"); UVMHIST_CALLED(pdhist);
1685 
1686 	UVMHIST_LOG(pdhist, "freeing %jd slots starting at %jd", nslots,
1687 	    startslot, 0, 0);
1688 
1689 	/*
1690 	 * ignore attempts to free the "bad" slot.
1691 	 */
1692 
1693 	if (startslot == SWSLOT_BAD) {
1694 		return;
1695 	}
1696 
1697 	/*
1698 	 * convert drum slot offset back to sdp, free the blocks
1699 	 * in the extent, and return.   must hold pri lock to do
1700 	 * lookup and access the extent.
1701 	 */
1702 
1703 	mutex_enter(&uvm_swap_data_lock);
1704 	sdp = swapdrum_getsdp(startslot);
1705 	KASSERT(uvmexp.nswapdev >= 1);
1706 	KASSERT(sdp != NULL);
1707 	KASSERT(sdp->swd_npginuse >= nslots);
1708 	blist_free(sdp->swd_blist, startslot - sdp->swd_drumoffset, nslots);
1709 	sdp->swd_npginuse -= nslots;
1710 	uvmexp.swpginuse -= nslots;
1711 	mutex_exit(&uvm_swap_data_lock);
1712 }
1713 
1714 /*
1715  * uvm_swap_put: put any number of pages into a contig place on swap
1716  *
1717  * => can be sync or async
1718  */
1719 
1720 int
1721 uvm_swap_put(int swslot, struct vm_page **ppsp, int npages, int flags)
1722 {
1723 	int error;
1724 
1725 	error = uvm_swap_io(ppsp, swslot, npages, B_WRITE |
1726 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1727 	return error;
1728 }
1729 
1730 /*
1731  * uvm_swap_get: get a single page from swap
1732  *
1733  * => usually a sync op (from fault)
1734  */
1735 
1736 int
1737 uvm_swap_get(struct vm_page *page, int swslot, int flags)
1738 {
1739 	int error;
1740 
1741 	atomic_inc_uint(&uvmexp.nswget);
1742 	KASSERT(flags & PGO_SYNCIO);
1743 	if (swslot == SWSLOT_BAD) {
1744 		return EIO;
1745 	}
1746 
1747 	error = uvm_swap_io(&page, swslot, 1, B_READ |
1748 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1749 	if (error == 0) {
1750 
1751 		/*
1752 		 * this page is no longer only in swap.
1753 		 */
1754 
1755 		KASSERT(uvmexp.swpgonly > 0);
1756 		atomic_dec_uint(&uvmexp.swpgonly);
1757 	}
1758 	return error;
1759 }
1760 
1761 /*
1762  * uvm_swap_io: do an i/o operation to swap
1763  */
1764 
1765 static int
1766 uvm_swap_io(struct vm_page **pps, int startslot, int npages, int flags)
1767 {
1768 	daddr_t startblk;
1769 	struct	buf *bp;
1770 	vaddr_t kva;
1771 	int	error, mapinflags;
1772 	bool write, async;
1773 	UVMHIST_FUNC("uvm_swap_io"); UVMHIST_CALLED(pdhist);
1774 
1775 	UVMHIST_LOG(pdhist, "<- called, startslot=%jd, npages=%jd, flags=%jd",
1776 	    startslot, npages, flags, 0);
1777 
1778 	write = (flags & B_READ) == 0;
1779 	async = (flags & B_ASYNC) != 0;
1780 
1781 	/*
1782 	 * allocate a buf for the i/o.
1783 	 */
1784 
1785 	KASSERT(curlwp != uvm.pagedaemon_lwp || (write && async));
1786 	bp = getiobuf(swapdev_vp, curlwp != uvm.pagedaemon_lwp);
1787 	if (bp == NULL) {
1788 		uvm_aio_aiodone_pages(pps, npages, true, ENOMEM);
1789 		return ENOMEM;
1790 	}
1791 
1792 	/*
1793 	 * convert starting drum slot to block number
1794 	 */
1795 
1796 	startblk = btodb((uint64_t)startslot << PAGE_SHIFT);
1797 
1798 	/*
1799 	 * first, map the pages into the kernel.
1800 	 */
1801 
1802 	mapinflags = !write ?
1803 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_READ :
1804 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_WRITE;
1805 	kva = uvm_pagermapin(pps, npages, mapinflags);
1806 
1807 	/*
1808 	 * fill in the bp/sbp.   we currently route our i/o through
1809 	 * /dev/drum's vnode [swapdev_vp].
1810 	 */
1811 
1812 	bp->b_cflags = BC_BUSY | BC_NOCACHE;
1813 	bp->b_flags = (flags & (B_READ|B_ASYNC));
1814 	bp->b_proc = &proc0;	/* XXX */
1815 	bp->b_vnbufs.le_next = NOLIST;
1816 	bp->b_data = (void *)kva;
1817 	bp->b_blkno = startblk;
1818 	bp->b_bufsize = bp->b_bcount = npages << PAGE_SHIFT;
1819 
1820 	/*
1821 	 * bump v_numoutput (counter of number of active outputs).
1822 	 */
1823 
1824 	if (write) {
1825 		mutex_enter(swapdev_vp->v_interlock);
1826 		swapdev_vp->v_numoutput++;
1827 		mutex_exit(swapdev_vp->v_interlock);
1828 	}
1829 
1830 	/*
1831 	 * for async ops we must set up the iodone handler.
1832 	 */
1833 
1834 	if (async) {
1835 		bp->b_iodone = uvm_aio_aiodone;
1836 		UVMHIST_LOG(pdhist, "doing async!", 0, 0, 0, 0);
1837 		if (curlwp == uvm.pagedaemon_lwp)
1838 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1839 		else
1840 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
1841 	} else {
1842 		bp->b_iodone = NULL;
1843 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1844 	}
1845 	UVMHIST_LOG(pdhist,
1846 	    "about to start io: data = %#jx blkno = 0x%jx, bcount = %jd",
1847 	    (uintptr_t)bp->b_data, bp->b_blkno, bp->b_bcount, 0);
1848 
1849 	/*
1850 	 * now we start the I/O, and if async, return.
1851 	 */
1852 
1853 	VOP_STRATEGY(swapdev_vp, bp);
1854 	if (async)
1855 		return 0;
1856 
1857 	/*
1858 	 * must be sync i/o.   wait for it to finish
1859 	 */
1860 
1861 	error = biowait(bp);
1862 
1863 	/*
1864 	 * kill the pager mapping
1865 	 */
1866 
1867 	uvm_pagermapout(kva, npages);
1868 
1869 	/*
1870 	 * now dispose of the buf and we're done.
1871 	 */
1872 
1873 	if (write) {
1874 		mutex_enter(swapdev_vp->v_interlock);
1875 		vwakeup(bp);
1876 		mutex_exit(swapdev_vp->v_interlock);
1877 	}
1878 	putiobuf(bp);
1879 	UVMHIST_LOG(pdhist, "<- done (sync)  error=%jd", error, 0, 0, 0);
1880 
1881 	return (error);
1882 }
1883