xref: /freebsd/sys/i386/i386/vm_machdep.c (revision 6b806d21d144c25f4fad714e1c0cf780f5e27d7e)
1 /*-
2  * Copyright (c) 1982, 1986 The Regents of the University of California.
3  * Copyright (c) 1989, 1990 William Jolitz
4  * Copyright (c) 1994 John Dyson
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * the Systems Programming Group of the University of Utah Computer
9  * Science Department, and William Jolitz.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. All advertising materials mentioning features or use of this software
20  *    must display the following acknowledgement:
21  *	This product includes software developed by the University of
22  *	California, Berkeley and its contributors.
23  * 4. Neither the name of the University nor the names of its contributors
24  *    may be used to endorse or promote products derived from this software
25  *    without specific prior written permission.
26  *
27  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
28  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
29  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
30  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
31  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
32  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
33  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
34  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
35  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
36  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37  * SUCH DAMAGE.
38  *
39  *	from: @(#)vm_machdep.c	7.3 (Berkeley) 5/13/91
40  *	Utah $Hdr: vm_machdep.c 1.16.1.1 89/06/23$
41  */
42 
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
45 
46 #include "opt_isa.h"
47 #include "opt_npx.h"
48 #ifdef PC98
49 #include "opt_pc98.h"
50 #endif
51 #include "opt_reset.h"
52 #include "opt_cpu.h"
53 
54 #include <sys/param.h>
55 #include <sys/systm.h>
56 #include <sys/bio.h>
57 #include <sys/buf.h>
58 #include <sys/kse.h>
59 #include <sys/kernel.h>
60 #include <sys/ktr.h>
61 #include <sys/lock.h>
62 #include <sys/malloc.h>
63 #include <sys/mbuf.h>
64 #include <sys/mutex.h>
65 #include <sys/pioctl.h>
66 #include <sys/proc.h>
67 #include <sys/sf_buf.h>
68 #include <sys/smp.h>
69 #include <sys/sysctl.h>
70 #include <sys/unistd.h>
71 #include <sys/vnode.h>
72 #include <sys/vmmeter.h>
73 
74 #include <machine/cpu.h>
75 #include <machine/cputypes.h>
76 #include <machine/md_var.h>
77 #include <machine/pcb.h>
78 #include <machine/pcb_ext.h>
79 #include <machine/vm86.h>
80 
81 #ifdef CPU_ELAN
82 #include <machine/elan_mmcr.h>
83 #endif
84 
85 #include <vm/vm.h>
86 #include <vm/vm_extern.h>
87 #include <vm/vm_kern.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_param.h>
91 
92 #ifdef PC98
93 #include <pc98/pc98/pc98.h>
94 #else
95 #include <i386/isa/isa.h>
96 #endif
97 
98 #ifndef NSFBUFS
99 #define	NSFBUFS		(512 + maxusers * 16)
100 #endif
101 
102 static void	cpu_reset_real(void);
103 #ifdef SMP
104 static void	cpu_reset_proxy(void);
105 static u_int	cpu_reset_proxyid;
106 static volatile u_int	cpu_reset_proxy_active;
107 #endif
108 static void	sf_buf_init(void *arg);
109 SYSINIT(sock_sf, SI_SUB_MBUF, SI_ORDER_ANY, sf_buf_init, NULL)
110 
111 LIST_HEAD(sf_head, sf_buf);
112 
113 /*
114  * A hash table of active sendfile(2) buffers
115  */
116 static struct sf_head *sf_buf_active;
117 static u_long sf_buf_hashmask;
118 
119 #define	SF_BUF_HASH(m)	(((m) - vm_page_array) & sf_buf_hashmask)
120 
121 static TAILQ_HEAD(, sf_buf) sf_buf_freelist;
122 static u_int	sf_buf_alloc_want;
123 
124 /*
125  * A lock used to synchronize access to the hash table and free list
126  */
127 static struct mtx sf_buf_lock;
128 
129 extern int	_ucodesel, _udatasel;
130 
131 /*
132  * Finish a fork operation, with process p2 nearly set up.
133  * Copy and update the pcb, set up the stack so that the child
134  * ready to run and return to user mode.
135  */
136 void
137 cpu_fork(td1, p2, td2, flags)
138 	register struct thread *td1;
139 	register struct proc *p2;
140 	struct thread *td2;
141 	int flags;
142 {
143 	register struct proc *p1;
144 	struct pcb *pcb2;
145 	struct mdproc *mdp2;
146 #ifdef DEV_NPX
147 	register_t savecrit;
148 #endif
149 
150 	p1 = td1->td_proc;
151 	if ((flags & RFPROC) == 0) {
152 		if ((flags & RFMEM) == 0) {
153 			/* unshare user LDT */
154 			struct mdproc *mdp1 = &p1->p_md;
155 			struct proc_ldt *pldt = mdp1->md_ldt;
156 			if (pldt && pldt->ldt_refcnt > 1) {
157 				pldt = user_ldt_alloc(mdp1, pldt->ldt_len);
158 				if (pldt == NULL)
159 					panic("could not copy LDT");
160 				mdp1->md_ldt = pldt;
161 				set_user_ldt(mdp1);
162 				user_ldt_free(td1);
163 			}
164 		}
165 		return;
166 	}
167 
168 	/* Ensure that p1's pcb is up to date. */
169 #ifdef DEV_NPX
170 	if (td1 == curthread)
171 		td1->td_pcb->pcb_gs = rgs();
172 	savecrit = intr_disable();
173 	if (PCPU_GET(fpcurthread) == td1)
174 		npxsave(&td1->td_pcb->pcb_save);
175 	intr_restore(savecrit);
176 #endif
177 
178 	/* Point the pcb to the top of the stack */
179 	pcb2 = (struct pcb *)(td2->td_kstack +
180 	    td2->td_kstack_pages * PAGE_SIZE) - 1;
181 	td2->td_pcb = pcb2;
182 
183 	/* Copy p1's pcb */
184 	bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
185 
186 	/* Point mdproc and then copy over td1's contents */
187 	mdp2 = &p2->p_md;
188 	bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
189 
190 	/*
191 	 * Create a new fresh stack for the new process.
192 	 * Copy the trap frame for the return to user mode as if from a
193 	 * syscall.  This copies most of the user mode register values.
194 	 * The -16 is so we can expand the trapframe if we go to vm86.
195 	 */
196 	td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb - 16) - 1;
197 	bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
198 
199 	td2->td_frame->tf_eax = 0;		/* Child returns zero */
200 	td2->td_frame->tf_eflags &= ~PSL_C;	/* success */
201 	td2->td_frame->tf_edx = 1;
202 
203 	/*
204 	 * If the parent process has the trap bit set (i.e. a debugger had
205 	 * single stepped the process to the system call), we need to clear
206 	 * the trap flag from the new frame unless the debugger had set PF_FORK
207 	 * on the parent.  Otherwise, the child will receive a (likely
208 	 * unexpected) SIGTRAP when it executes the first instruction after
209 	 * returning  to userland.
210 	 */
211 	if ((p1->p_pfsflags & PF_FORK) == 0)
212 		td2->td_frame->tf_eflags &= ~PSL_T;
213 
214 	/*
215 	 * Set registers for trampoline to user mode.  Leave space for the
216 	 * return address on stack.  These are the kernel mode register values.
217 	 */
218 #ifdef PAE
219 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdpt);
220 #else
221 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdir);
222 #endif
223 	pcb2->pcb_edi = 0;
224 	pcb2->pcb_esi = (int)fork_return;	/* fork_trampoline argument */
225 	pcb2->pcb_ebp = 0;
226 	pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *);
227 	pcb2->pcb_ebx = (int)td2;		/* fork_trampoline argument */
228 	pcb2->pcb_eip = (int)fork_trampoline;
229 	pcb2->pcb_psl = PSL_KERNEL;		/* ints disabled */
230 	pcb2->pcb_gs = rgs();
231 	/*-
232 	 * pcb2->pcb_dr*:	cloned above.
233 	 * pcb2->pcb_savefpu:	cloned above.
234 	 * pcb2->pcb_flags:	cloned above.
235 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
236 	 * pcb2->pcb_gs:	cloned above.
237 	 * pcb2->pcb_ext:	cleared below.
238 	 */
239 
240 	/*
241 	 * XXX don't copy the i/o pages.  this should probably be fixed.
242 	 */
243 	pcb2->pcb_ext = 0;
244 
245 	/* Copy the LDT, if necessary. */
246 	mtx_lock_spin(&sched_lock);
247 	if (mdp2->md_ldt != NULL) {
248 		if (flags & RFMEM) {
249 			mdp2->md_ldt->ldt_refcnt++;
250 		} else {
251 			mdp2->md_ldt = user_ldt_alloc(mdp2,
252 			    mdp2->md_ldt->ldt_len);
253 			if (mdp2->md_ldt == NULL)
254 				panic("could not copy LDT");
255 		}
256 	}
257 	mtx_unlock_spin(&sched_lock);
258 
259 	/*
260 	 * Now, cpu_switch() can schedule the new process.
261 	 * pcb_esp is loaded pointing to the cpu_switch() stack frame
262 	 * containing the return address when exiting cpu_switch.
263 	 * This will normally be to fork_trampoline(), which will have
264 	 * %ebx loaded with the new proc's pointer.  fork_trampoline()
265 	 * will set up a stack to call fork_return(p, frame); to complete
266 	 * the return to user-mode.
267 	 */
268 }
269 
270 /*
271  * Intercept the return address from a freshly forked process that has NOT
272  * been scheduled yet.
273  *
274  * This is needed to make kernel threads stay in kernel mode.
275  */
276 void
277 cpu_set_fork_handler(td, func, arg)
278 	struct thread *td;
279 	void (*func)(void *);
280 	void *arg;
281 {
282 	/*
283 	 * Note that the trap frame follows the args, so the function
284 	 * is really called like this:  func(arg, frame);
285 	 */
286 	td->td_pcb->pcb_esi = (int) func;	/* function */
287 	td->td_pcb->pcb_ebx = (int) arg;	/* first arg */
288 }
289 
290 void
291 cpu_exit(struct thread *td)
292 {
293 	struct mdproc *mdp;
294 
295 	/* Reset pc->pcb_gs and %gs before possibly invalidating it. */
296 	mdp = &td->td_proc->p_md;
297 	if (mdp->md_ldt) {
298 		td->td_pcb->pcb_gs = _udatasel;
299 		load_gs(_udatasel);
300 		user_ldt_free(td);
301 	}
302 }
303 
304 void
305 cpu_thread_exit(struct thread *td)
306 {
307 	struct pcb *pcb = td->td_pcb;
308 
309 #ifdef DEV_NPX
310 	if (td == PCPU_GET(fpcurthread))
311 		npxdrop();
312 #endif
313 	if (pcb->pcb_flags & PCB_DBREGS) {
314 		/* disable all hardware breakpoints */
315 		reset_dbregs();
316 		pcb->pcb_flags &= ~PCB_DBREGS;
317 	}
318 }
319 
320 void
321 cpu_thread_clean(struct thread *td)
322 {
323 	struct pcb *pcb;
324 
325 	pcb = td->td_pcb;
326 	if (pcb->pcb_ext != 0) {
327 		/* XXXKSE  XXXSMP  not SMP SAFE.. what locks do we have? */
328 		/* if (pcb->pcb_ext->ext_refcount-- == 1) ?? */
329 		/*
330 		 * XXX do we need to move the TSS off the allocated pages
331 		 * before freeing them?  (not done here)
332 		 */
333 		kmem_free(kernel_map, (vm_offset_t)pcb->pcb_ext,
334 		    ctob(IOPAGES + 1));
335 		pcb->pcb_ext = 0;
336 	}
337 }
338 
339 void
340 cpu_thread_swapin(struct thread *td)
341 {
342 }
343 
344 void
345 cpu_thread_swapout(struct thread *td)
346 {
347 }
348 
349 void
350 cpu_thread_setup(struct thread *td)
351 {
352 
353 	td->td_pcb = (struct pcb *)(td->td_kstack +
354 	    td->td_kstack_pages * PAGE_SIZE) - 1;
355 	td->td_frame = (struct trapframe *)((caddr_t)td->td_pcb - 16) - 1;
356 	td->td_pcb->pcb_ext = NULL;
357 }
358 
359 /*
360  * Initialize machine state (pcb and trap frame) for a new thread about to
361  * upcall. Put enough state in the new thread's PCB to get it to go back
362  * userret(), where we can intercept it again to set the return (upcall)
363  * Address and stack, along with those from upcals that are from other sources
364  * such as those generated in thread_userret() itself.
365  */
366 void
367 cpu_set_upcall(struct thread *td, struct thread *td0)
368 {
369 	struct pcb *pcb2;
370 
371 	/* Point the pcb to the top of the stack. */
372 	pcb2 = td->td_pcb;
373 
374 	/*
375 	 * Copy the upcall pcb.  This loads kernel regs.
376 	 * Those not loaded individually below get their default
377 	 * values here.
378 	 *
379 	 * XXXKSE It might be a good idea to simply skip this as
380 	 * the values of the other registers may be unimportant.
381 	 * This would remove any requirement for knowing the KSE
382 	 * at this time (see the matching comment below for
383 	 * more analysis) (need a good safe default).
384 	 */
385 	bcopy(td0->td_pcb, pcb2, sizeof(*pcb2));
386 	pcb2->pcb_flags &= ~(PCB_NPXTRAP|PCB_NPXINITDONE);
387 
388 	/*
389 	 * Create a new fresh stack for the new thread.
390 	 * The -16 is so we can expand the trapframe if we go to vm86.
391 	 * Don't forget to set this stack value into whatever supplies
392 	 * the address for the fault handlers.
393 	 * The contexts are filled in at the time we actually DO the
394 	 * upcall as only then do we know which KSE we got.
395 	 */
396 	bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe));
397 
398 	/*
399 	 * Set registers for trampoline to user mode.  Leave space for the
400 	 * return address on stack.  These are the kernel mode register values.
401 	 */
402 #ifdef PAE
403 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(td->td_proc->p_vmspace)->pm_pdpt);
404 #else
405 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(td->td_proc->p_vmspace)->pm_pdir);
406 #endif
407 	pcb2->pcb_edi = 0;
408 	pcb2->pcb_esi = (int)fork_return;		    /* trampoline arg */
409 	pcb2->pcb_ebp = 0;
410 	pcb2->pcb_esp = (int)td->td_frame - sizeof(void *); /* trampoline arg */
411 	pcb2->pcb_ebx = (int)td;			    /* trampoline arg */
412 	pcb2->pcb_eip = (int)fork_trampoline;
413 	pcb2->pcb_psl &= ~(PSL_I);	/* interrupts must be disabled */
414 	pcb2->pcb_gs = rgs();
415 	/*
416 	 * If we didn't copy the pcb, we'd need to do the following registers:
417 	 * pcb2->pcb_dr*:	cloned above.
418 	 * pcb2->pcb_savefpu:	cloned above.
419 	 * pcb2->pcb_flags:	cloned above.
420 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
421 	 * pcb2->pcb_gs:	cloned above.  XXXKSE ???
422 	 * pcb2->pcb_ext:	cleared below.
423 	 */
424 	pcb2->pcb_ext = NULL;
425 }
426 
427 /*
428  * Set that machine state for performing an upcall that has to
429  * be done in thread_userret() so that those upcalls generated
430  * in thread_userret() itself can be done as well.
431  */
432 void
433 cpu_set_upcall_kse(struct thread *td, struct kse_upcall *ku)
434 {
435 
436 	/*
437 	 * Do any extra cleaning that needs to be done.
438 	 * The thread may have optional components
439 	 * that are not present in a fresh thread.
440 	 * This may be a recycled thread so make it look
441 	 * as though it's newly allocated.
442 	 */
443 	cpu_thread_clean(td);
444 
445 	/*
446 	 * Set the trap frame to point at the beginning of the uts
447 	 * function.
448 	 */
449 	td->td_frame->tf_ebp = 0;
450 	td->td_frame->tf_esp =
451 	    (int)ku->ku_stack.ss_sp + ku->ku_stack.ss_size - 16;
452 	td->td_frame->tf_eip = (int)ku->ku_func;
453 
454 	/*
455 	 * Pass the address of the mailbox for this kse to the uts
456 	 * function as a parameter on the stack.
457 	 */
458 	suword((void *)(td->td_frame->tf_esp + sizeof(void *)),
459 	    (int)ku->ku_mailbox);
460 }
461 
462 /*
463  * Convert kernel VA to physical address
464  */
465 vm_paddr_t
466 kvtop(void *addr)
467 {
468 	vm_paddr_t pa;
469 
470 	pa = pmap_kextract((vm_offset_t)addr);
471 	if (pa == 0)
472 		panic("kvtop: zero page frame");
473 	return (pa);
474 }
475 
476 #ifdef SMP
477 static void
478 cpu_reset_proxy()
479 {
480 
481 	cpu_reset_proxy_active = 1;
482 	while (cpu_reset_proxy_active == 1)
483 		;	/* Wait for other cpu to see that we've started */
484 	stop_cpus((1<<cpu_reset_proxyid));
485 	printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
486 	DELAY(1000000);
487 	cpu_reset_real();
488 }
489 #endif
490 
491 void
492 cpu_reset()
493 {
494 #ifdef SMP
495 	u_int cnt, map;
496 
497 	if (smp_active) {
498 		map = PCPU_GET(other_cpus) & ~stopped_cpus;
499 		if (map != 0) {
500 			printf("cpu_reset: Stopping other CPUs\n");
501 			stop_cpus(map);
502 		}
503 
504 		if (PCPU_GET(cpuid) != 0) {
505 			cpu_reset_proxyid = PCPU_GET(cpuid);
506 			cpustop_restartfunc = cpu_reset_proxy;
507 			cpu_reset_proxy_active = 0;
508 			printf("cpu_reset: Restarting BSP\n");
509 			started_cpus = (1<<0);		/* Restart CPU #0 */
510 
511 			cnt = 0;
512 			while (cpu_reset_proxy_active == 0 && cnt < 10000000)
513 				cnt++;	/* Wait for BSP to announce restart */
514 			if (cpu_reset_proxy_active == 0)
515 				printf("cpu_reset: Failed to restart BSP\n");
516 			enable_intr();
517 			cpu_reset_proxy_active = 2;
518 
519 			while (1);
520 			/* NOTREACHED */
521 		}
522 
523 		DELAY(1000000);
524 	}
525 #endif
526 	cpu_reset_real();
527 	/* NOTREACHED */
528 }
529 
530 static void
531 cpu_reset_real()
532 {
533 
534 #ifdef CPU_ELAN
535 	if (elan_mmcr != NULL)
536 		elan_mmcr->RESCFG = 1;
537 #endif
538 
539 	if (cpu == CPU_GEODE1100) {
540 		/* Attempt Geode's own reset */
541 		outl(0xcf8, 0x80009044ul);
542 		outl(0xcfc, 0xf);
543 	}
544 
545 #ifdef PC98
546 	/*
547 	 * Attempt to do a CPU reset via CPU reset port.
548 	 */
549 	disable_intr();
550 	if ((inb(0x35) & 0xa0) != 0xa0) {
551 		outb(0x37, 0x0f);		/* SHUT0 = 0. */
552 		outb(0x37, 0x0b);		/* SHUT1 = 0. */
553 	}
554 	outb(0xf0, 0x00);		/* Reset. */
555 #else
556 #if !defined(BROKEN_KEYBOARD_RESET)
557 	/*
558 	 * Attempt to do a CPU reset via the keyboard controller,
559 	 * do not turn off GateA20, as any machine that fails
560 	 * to do the reset here would then end up in no man's land.
561 	 */
562 	outb(IO_KBD + 4, 0xFE);
563 	DELAY(500000);	/* wait 0.5 sec to see if that did it */
564 	printf("Keyboard reset did not work, attempting CPU shutdown\n");
565 	DELAY(1000000);	/* wait 1 sec for printf to complete */
566 #endif
567 #endif /* PC98 */
568 
569 	/* Force a shutdown by unmapping entire address space. */
570 	bzero((caddr_t)PTD, NBPTD);
571 
572 	/* "good night, sweet prince .... <THUNK!>" */
573 	invltlb();
574 	/* NOTREACHED */
575 	while(1);
576 }
577 
578 /*
579  * Allocate a pool of sf_bufs (sendfile(2) or "super-fast" if you prefer. :-))
580  */
581 static void
582 sf_buf_init(void *arg)
583 {
584 	struct sf_buf *sf_bufs;
585 	vm_offset_t sf_base;
586 	int i;
587 
588 	nsfbufs = NSFBUFS;
589 	TUNABLE_INT_FETCH("kern.ipc.nsfbufs", &nsfbufs);
590 
591 	sf_buf_active = hashinit(nsfbufs, M_TEMP, &sf_buf_hashmask);
592 	TAILQ_INIT(&sf_buf_freelist);
593 	sf_base = kmem_alloc_nofault(kernel_map, nsfbufs * PAGE_SIZE);
594 	sf_bufs = malloc(nsfbufs * sizeof(struct sf_buf), M_TEMP,
595 	    M_NOWAIT | M_ZERO);
596 	for (i = 0; i < nsfbufs; i++) {
597 		sf_bufs[i].kva = sf_base + i * PAGE_SIZE;
598 		TAILQ_INSERT_TAIL(&sf_buf_freelist, &sf_bufs[i], free_entry);
599 	}
600 	sf_buf_alloc_want = 0;
601 	mtx_init(&sf_buf_lock, "sf_buf", NULL, MTX_DEF);
602 }
603 
604 /*
605  * Get an sf_buf from the freelist. Will block if none are available.
606  */
607 struct sf_buf *
608 sf_buf_alloc(struct vm_page *m, int flags)
609 {
610 	pt_entry_t opte, *ptep;
611 	struct sf_head *hash_list;
612 	struct sf_buf *sf;
613 	int error;
614 
615 	hash_list = &sf_buf_active[SF_BUF_HASH(m)];
616 	mtx_lock(&sf_buf_lock);
617 	LIST_FOREACH(sf, hash_list, list_entry) {
618 		if (sf->m == m) {
619 			sf->ref_count++;
620 			if (sf->ref_count == 1) {
621 				TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
622 				nsfbufsused++;
623 				nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
624 			}
625 			goto done;
626 		}
627 	}
628 	while ((sf = TAILQ_FIRST(&sf_buf_freelist)) == NULL) {
629 		if (flags & SFB_NOWAIT)
630 			goto done;
631 		sf_buf_alloc_want++;
632 		mbstat.sf_allocwait++;
633 		error = msleep(&sf_buf_freelist, &sf_buf_lock,
634 		    (flags & SFB_CATCH) ? PCATCH | PVM : PVM, "sfbufa", 0);
635 		sf_buf_alloc_want--;
636 
637 		/*
638 		 * If we got a signal, don't risk going back to sleep.
639 		 */
640 		if (error)
641 			goto done;
642 	}
643 	TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
644 	if (sf->m != NULL)
645 		LIST_REMOVE(sf, list_entry);
646 	LIST_INSERT_HEAD(hash_list, sf, list_entry);
647 	sf->ref_count = 1;
648 	sf->m = m;
649 	nsfbufsused++;
650 	nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
651 
652 	/*
653 	 * Update the sf_buf's virtual-to-physical mapping, flushing the
654 	 * virtual address from the TLB only if the PTE implies that the old
655 	 * mapping has been used.  Since the reference count for the sf_buf's
656 	 * old mapping was zero, that mapping is not currently in use.
657 	 * Consequently, there is no need to exchange the old and new PTEs
658 	 * atomically, even under PAE.
659 	 */
660 	ptep = vtopte(sf->kva);
661 	opte = *ptep;
662 	*ptep = VM_PAGE_TO_PHYS(m) | pgeflag | PG_RW | PG_V;
663 	if ((opte & (PG_A | PG_V)) == (PG_A | PG_V))
664 		pmap_invalidate_page(kernel_pmap, sf->kva);
665 done:
666 	mtx_unlock(&sf_buf_lock);
667 	return (sf);
668 }
669 
670 /*
671  * Remove a reference from the given sf_buf, adding it to the free
672  * list when its reference count reaches zero.  A freed sf_buf still,
673  * however, retains its virtual-to-physical mapping until it is
674  * recycled or reactivated by sf_buf_alloc(9).
675  */
676 void
677 sf_buf_free(struct sf_buf *sf)
678 {
679 
680 	mtx_lock(&sf_buf_lock);
681 	sf->ref_count--;
682 	if (sf->ref_count == 0) {
683 		TAILQ_INSERT_TAIL(&sf_buf_freelist, sf, free_entry);
684 		nsfbufsused--;
685 		if (sf_buf_alloc_want > 0)
686 			wakeup_one(&sf_buf_freelist);
687 	}
688 	mtx_unlock(&sf_buf_lock);
689 }
690 
691 /*
692  * Software interrupt handler for queued VM system processing.
693  */
694 void
695 swi_vm(void *dummy)
696 {
697 	if (busdma_swi_pending != 0)
698 		busdma_swi();
699 }
700 
701 /*
702  * Tell whether this address is in some physical memory region.
703  * Currently used by the kernel coredump code in order to avoid
704  * dumping the ``ISA memory hole'' which could cause indefinite hangs,
705  * or other unpredictable behaviour.
706  */
707 
708 int
709 is_physical_memory(vm_paddr_t addr)
710 {
711 
712 #ifdef DEV_ISA
713 	/* The ISA ``memory hole''. */
714 	if (addr >= 0xa0000 && addr < 0x100000)
715 		return 0;
716 #endif
717 
718 	/*
719 	 * stuff other tests for known memory-mapped devices (PCI?)
720 	 * here
721 	 */
722 
723 	return 1;
724 }
725