xref: /freebsd/sys/i386/i386/vm_machdep.c (revision 1e413cf93298b5b97441a21d9a50fdcd0ee9945e)
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 #include "opt_reset.h"
49 #include "opt_cpu.h"
50 #include "opt_xbox.h"
51 
52 #include <sys/param.h>
53 #include <sys/systm.h>
54 #include <sys/bio.h>
55 #include <sys/buf.h>
56 #include <sys/kse.h>
57 #include <sys/kernel.h>
58 #include <sys/ktr.h>
59 #include <sys/lock.h>
60 #include <sys/malloc.h>
61 #include <sys/mbuf.h>
62 #include <sys/mutex.h>
63 #include <sys/pioctl.h>
64 #include <sys/proc.h>
65 #include <sys/sf_buf.h>
66 #include <sys/smp.h>
67 #include <sys/sched.h>
68 #include <sys/sysctl.h>
69 #include <sys/unistd.h>
70 #include <sys/vnode.h>
71 #include <sys/vmmeter.h>
72 
73 #include <machine/cpu.h>
74 #include <machine/cputypes.h>
75 #include <machine/md_var.h>
76 #include <machine/pcb.h>
77 #include <machine/pcb_ext.h>
78 #include <machine/smp.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/cbus/cbus.h>
94 #else
95 #include <i386/isa/isa.h>
96 #endif
97 
98 #ifdef XBOX
99 #include <machine/xbox.h>
100 #endif
101 
102 #ifndef NSFBUFS
103 #define	NSFBUFS		(512 + maxusers * 16)
104 #endif
105 
106 static void	cpu_reset_real(void);
107 #ifdef SMP
108 static void	cpu_reset_proxy(void);
109 static u_int	cpu_reset_proxyid;
110 static volatile u_int	cpu_reset_proxy_active;
111 #endif
112 static void	sf_buf_init(void *arg);
113 SYSINIT(sock_sf, SI_SUB_MBUF, SI_ORDER_ANY, sf_buf_init, NULL)
114 
115 LIST_HEAD(sf_head, sf_buf);
116 
117 /*
118  * A hash table of active sendfile(2) buffers
119  */
120 static struct sf_head *sf_buf_active;
121 static u_long sf_buf_hashmask;
122 
123 #define	SF_BUF_HASH(m)	(((m) - vm_page_array) & sf_buf_hashmask)
124 
125 static TAILQ_HEAD(, sf_buf) sf_buf_freelist;
126 static u_int	sf_buf_alloc_want;
127 
128 /*
129  * A lock used to synchronize access to the hash table and free list
130  */
131 static struct mtx sf_buf_lock;
132 
133 extern int	_ucodesel, _udatasel;
134 
135 /*
136  * Finish a fork operation, with process p2 nearly set up.
137  * Copy and update the pcb, set up the stack so that the child
138  * ready to run and return to user mode.
139  */
140 void
141 cpu_fork(td1, p2, td2, flags)
142 	register struct thread *td1;
143 	register struct proc *p2;
144 	struct thread *td2;
145 	int flags;
146 {
147 	register struct proc *p1;
148 	struct pcb *pcb2;
149 	struct mdproc *mdp2;
150 #ifdef DEV_NPX
151 	register_t savecrit;
152 #endif
153 
154 	p1 = td1->td_proc;
155 	if ((flags & RFPROC) == 0) {
156 		if ((flags & RFMEM) == 0) {
157 			/* unshare user LDT */
158 			struct mdproc *mdp1 = &p1->p_md;
159 			struct proc_ldt *pldt;
160 
161 			mtx_lock_spin(&dt_lock);
162 			if ((pldt = mdp1->md_ldt) != NULL &&
163 			    pldt->ldt_refcnt > 1) {
164 				pldt = user_ldt_alloc(mdp1, pldt->ldt_len);
165 				if (pldt == NULL)
166 					panic("could not copy LDT");
167 				mdp1->md_ldt = pldt;
168 				set_user_ldt(mdp1);
169 				user_ldt_free(td1);
170 			} else
171 				mtx_unlock_spin(&dt_lock);
172 		}
173 		return;
174 	}
175 
176 	/* Ensure that p1's pcb is up to date. */
177 	if (td1 == curthread)
178 		td1->td_pcb->pcb_gs = rgs();
179 #ifdef DEV_NPX
180 	savecrit = intr_disable();
181 	if (PCPU_GET(fpcurthread) == td1)
182 		npxsave(&td1->td_pcb->pcb_save);
183 	intr_restore(savecrit);
184 #endif
185 
186 	/* Point the pcb to the top of the stack */
187 	pcb2 = (struct pcb *)(td2->td_kstack +
188 	    td2->td_kstack_pages * PAGE_SIZE) - 1;
189 	td2->td_pcb = pcb2;
190 
191 	/* Copy p1's pcb */
192 	bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
193 
194 	/* Point mdproc and then copy over td1's contents */
195 	mdp2 = &p2->p_md;
196 	bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
197 
198 	/*
199 	 * Create a new fresh stack for the new process.
200 	 * Copy the trap frame for the return to user mode as if from a
201 	 * syscall.  This copies most of the user mode register values.
202 	 * The -16 is so we can expand the trapframe if we go to vm86.
203 	 */
204 	td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb - 16) - 1;
205 	bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
206 
207 	td2->td_frame->tf_eax = 0;		/* Child returns zero */
208 	td2->td_frame->tf_eflags &= ~PSL_C;	/* success */
209 	td2->td_frame->tf_edx = 1;
210 
211 	/*
212 	 * If the parent process has the trap bit set (i.e. a debugger had
213 	 * single stepped the process to the system call), we need to clear
214 	 * the trap flag from the new frame unless the debugger had set PF_FORK
215 	 * on the parent.  Otherwise, the child will receive a (likely
216 	 * unexpected) SIGTRAP when it executes the first instruction after
217 	 * returning  to userland.
218 	 */
219 	if ((p1->p_pfsflags & PF_FORK) == 0)
220 		td2->td_frame->tf_eflags &= ~PSL_T;
221 
222 	/*
223 	 * Set registers for trampoline to user mode.  Leave space for the
224 	 * return address on stack.  These are the kernel mode register values.
225 	 */
226 #ifdef PAE
227 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdpt);
228 #else
229 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdir);
230 #endif
231 	pcb2->pcb_edi = 0;
232 	pcb2->pcb_esi = (int)fork_return;	/* fork_trampoline argument */
233 	pcb2->pcb_ebp = 0;
234 	pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *);
235 	pcb2->pcb_ebx = (int)td2;		/* fork_trampoline argument */
236 	pcb2->pcb_eip = (int)fork_trampoline;
237 	pcb2->pcb_psl = PSL_KERNEL;		/* ints disabled */
238 	/*-
239 	 * pcb2->pcb_dr*:	cloned above.
240 	 * pcb2->pcb_savefpu:	cloned above.
241 	 * pcb2->pcb_flags:	cloned above.
242 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
243 	 * pcb2->pcb_gs:	cloned above.
244 	 * pcb2->pcb_ext:	cleared below.
245 	 */
246 
247 	/*
248 	 * XXX don't copy the i/o pages.  this should probably be fixed.
249 	 */
250 	pcb2->pcb_ext = 0;
251 
252 	/* Copy the LDT, if necessary. */
253 	mtx_lock_spin(&dt_lock);
254 	if (mdp2->md_ldt != NULL) {
255 		if (flags & RFMEM) {
256 			mdp2->md_ldt->ldt_refcnt++;
257 		} else {
258 			mdp2->md_ldt = user_ldt_alloc(mdp2,
259 			    mdp2->md_ldt->ldt_len);
260 			if (mdp2->md_ldt == NULL)
261 				panic("could not copy LDT");
262 		}
263 	}
264 	mtx_unlock_spin(&dt_lock);
265 
266 	/* Setup to release spin count in fork_exit(). */
267 	td2->td_md.md_spinlock_count = 1;
268 	td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
269 
270 	/*
271 	 * Now, cpu_switch() can schedule the new process.
272 	 * pcb_esp is loaded pointing to the cpu_switch() stack frame
273 	 * containing the return address when exiting cpu_switch.
274 	 * This will normally be to fork_trampoline(), which will have
275 	 * %ebx loaded with the new proc's pointer.  fork_trampoline()
276 	 * will set up a stack to call fork_return(p, frame); to complete
277 	 * the return to user-mode.
278 	 */
279 }
280 
281 /*
282  * Intercept the return address from a freshly forked process that has NOT
283  * been scheduled yet.
284  *
285  * This is needed to make kernel threads stay in kernel mode.
286  */
287 void
288 cpu_set_fork_handler(td, func, arg)
289 	struct thread *td;
290 	void (*func)(void *);
291 	void *arg;
292 {
293 	/*
294 	 * Note that the trap frame follows the args, so the function
295 	 * is really called like this:  func(arg, frame);
296 	 */
297 	td->td_pcb->pcb_esi = (int) func;	/* function */
298 	td->td_pcb->pcb_ebx = (int) arg;	/* first arg */
299 }
300 
301 void
302 cpu_exit(struct thread *td)
303 {
304 
305 	/*
306 	 * If this process has a custom LDT, release it.  Reset pc->pcb_gs
307 	 * and %gs before we free it in case they refer to an LDT entry.
308 	 */
309 	mtx_lock_spin(&dt_lock);
310 	if (td->td_proc->p_md.md_ldt) {
311 		td->td_pcb->pcb_gs = _udatasel;
312 		load_gs(_udatasel);
313 		user_ldt_free(td);
314 	} else
315 		mtx_unlock_spin(&dt_lock);
316 }
317 
318 void
319 cpu_thread_exit(struct thread *td)
320 {
321 
322 #ifdef DEV_NPX
323 	if (td == PCPU_GET(fpcurthread))
324 		npxdrop();
325 #endif
326 
327 	/* Disable any hardware breakpoints. */
328 	if (td->td_pcb->pcb_flags & PCB_DBREGS) {
329 		reset_dbregs();
330 		td->td_pcb->pcb_flags &= ~PCB_DBREGS;
331 	}
332 }
333 
334 void
335 cpu_thread_clean(struct thread *td)
336 {
337 	struct pcb *pcb;
338 
339 	pcb = td->td_pcb;
340 	if (pcb->pcb_ext != NULL) {
341 		/* XXXKSE  XXXSMP  not SMP SAFE.. what locks do we have? */
342 		/* if (pcb->pcb_ext->ext_refcount-- == 1) ?? */
343 		/*
344 		 * XXX do we need to move the TSS off the allocated pages
345 		 * before freeing them?  (not done here)
346 		 */
347 		kmem_free(kernel_map, (vm_offset_t)pcb->pcb_ext,
348 		    ctob(IOPAGES + 1));
349 		pcb->pcb_ext = NULL;
350 	}
351 }
352 
353 void
354 cpu_thread_swapin(struct thread *td)
355 {
356 }
357 
358 void
359 cpu_thread_swapout(struct thread *td)
360 {
361 }
362 
363 void
364 cpu_thread_alloc(struct thread *td)
365 {
366 
367 	td->td_pcb = (struct pcb *)(td->td_kstack +
368 	    td->td_kstack_pages * PAGE_SIZE) - 1;
369 	td->td_frame = (struct trapframe *)((caddr_t)td->td_pcb - 16) - 1;
370 	td->td_pcb->pcb_ext = NULL;
371 }
372 
373 void
374 cpu_thread_free(struct thread *td)
375 {
376 
377 	cpu_thread_clean(td);
378 }
379 
380 /*
381  * Initialize machine state (pcb and trap frame) for a new thread about to
382  * upcall. Put enough state in the new thread's PCB to get it to go back
383  * userret(), where we can intercept it again to set the return (upcall)
384  * Address and stack, along with those from upcals that are from other sources
385  * such as those generated in thread_userret() itself.
386  */
387 void
388 cpu_set_upcall(struct thread *td, struct thread *td0)
389 {
390 	struct pcb *pcb2;
391 
392 	/* Point the pcb to the top of the stack. */
393 	pcb2 = td->td_pcb;
394 
395 	/*
396 	 * Copy the upcall pcb.  This loads kernel regs.
397 	 * Those not loaded individually below get their default
398 	 * values here.
399 	 *
400 	 * XXXKSE It might be a good idea to simply skip this as
401 	 * the values of the other registers may be unimportant.
402 	 * This would remove any requirement for knowing the KSE
403 	 * at this time (see the matching comment below for
404 	 * more analysis) (need a good safe default).
405 	 */
406 	bcopy(td0->td_pcb, pcb2, sizeof(*pcb2));
407 	pcb2->pcb_flags &= ~(PCB_NPXTRAP|PCB_NPXINITDONE);
408 
409 	/*
410 	 * Create a new fresh stack for the new thread.
411 	 * The -16 is so we can expand the trapframe if we go to vm86.
412 	 * Don't forget to set this stack value into whatever supplies
413 	 * the address for the fault handlers.
414 	 * The contexts are filled in at the time we actually DO the
415 	 * upcall as only then do we know which KSE we got.
416 	 */
417 	bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe));
418 
419 	/*
420 	 * Set registers for trampoline to user mode.  Leave space for the
421 	 * return address on stack.  These are the kernel mode register values.
422 	 */
423 #ifdef PAE
424 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(td->td_proc->p_vmspace)->pm_pdpt);
425 #else
426 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(td->td_proc->p_vmspace)->pm_pdir);
427 #endif
428 	pcb2->pcb_edi = 0;
429 	pcb2->pcb_esi = (int)fork_return;		    /* trampoline arg */
430 	pcb2->pcb_ebp = 0;
431 	pcb2->pcb_esp = (int)td->td_frame - sizeof(void *); /* trampoline arg */
432 	pcb2->pcb_ebx = (int)td;			    /* trampoline arg */
433 	pcb2->pcb_eip = (int)fork_trampoline;
434 	pcb2->pcb_psl &= ~(PSL_I);	/* interrupts must be disabled */
435 	pcb2->pcb_gs = rgs();
436 	/*
437 	 * If we didn't copy the pcb, we'd need to do the following registers:
438 	 * pcb2->pcb_dr*:	cloned above.
439 	 * pcb2->pcb_savefpu:	cloned above.
440 	 * pcb2->pcb_flags:	cloned above.
441 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
442 	 * pcb2->pcb_gs:	cloned above.  XXXKSE ???
443 	 * pcb2->pcb_ext:	cleared below.
444 	 */
445 	pcb2->pcb_ext = NULL;
446 
447 	/* Setup to release spin count in fork_exit(). */
448 	td->td_md.md_spinlock_count = 1;
449 	td->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
450 }
451 
452 /*
453  * Set that machine state for performing an upcall that has to
454  * be done in thread_userret() so that those upcalls generated
455  * in thread_userret() itself can be done as well.
456  */
457 void
458 cpu_set_upcall_kse(struct thread *td, void (*entry)(void *), void *arg,
459 	stack_t *stack)
460 {
461 
462 	/*
463 	 * Do any extra cleaning that needs to be done.
464 	 * The thread may have optional components
465 	 * that are not present in a fresh thread.
466 	 * This may be a recycled thread so make it look
467 	 * as though it's newly allocated.
468 	 */
469 	cpu_thread_clean(td);
470 
471 	/*
472 	 * Set the trap frame to point at the beginning of the uts
473 	 * function.
474 	 */
475 	td->td_frame->tf_ebp = 0;
476 	td->td_frame->tf_esp =
477 	    (((int)stack->ss_sp + stack->ss_size - 4) & ~0x0f) - 4;
478 	td->td_frame->tf_eip = (int)entry;
479 
480 	/*
481 	 * Pass the address of the mailbox for this kse to the uts
482 	 * function as a parameter on the stack.
483 	 */
484 	suword((void *)(td->td_frame->tf_esp + sizeof(void *)),
485 	    (int)arg);
486 }
487 
488 int
489 cpu_set_user_tls(struct thread *td, void *tls_base)
490 {
491 	struct segment_descriptor sd;
492 	uint32_t base;
493 
494 	/*
495 	 * Construct a descriptor and store it in the pcb for
496 	 * the next context switch.  Also store it in the gdt
497 	 * so that the load of tf_fs into %fs will activate it
498 	 * at return to userland.
499 	 */
500 	base = (uint32_t)tls_base;
501 	sd.sd_lobase = base & 0xffffff;
502 	sd.sd_hibase = (base >> 24) & 0xff;
503 	sd.sd_lolimit = 0xffff;	/* 4GB limit, wraps around */
504 	sd.sd_hilimit = 0xf;
505 	sd.sd_type  = SDT_MEMRWA;
506 	sd.sd_dpl   = SEL_UPL;
507 	sd.sd_p     = 1;
508 	sd.sd_xx    = 0;
509 	sd.sd_def32 = 1;
510 	sd.sd_gran  = 1;
511 	critical_enter();
512 	/* set %gs */
513 	td->td_pcb->pcb_gsd = sd;
514 	if (td == curthread) {
515 		PCPU_GET(fsgs_gdt)[1] = sd;
516 		load_gs(GSEL(GUGS_SEL, SEL_UPL));
517 	}
518 	critical_exit();
519 	return (0);
520 }
521 
522 /*
523  * Convert kernel VA to physical address
524  */
525 vm_paddr_t
526 kvtop(void *addr)
527 {
528 	vm_paddr_t pa;
529 
530 	pa = pmap_kextract((vm_offset_t)addr);
531 	if (pa == 0)
532 		panic("kvtop: zero page frame");
533 	return (pa);
534 }
535 
536 #ifdef SMP
537 static void
538 cpu_reset_proxy()
539 {
540 
541 	cpu_reset_proxy_active = 1;
542 	while (cpu_reset_proxy_active == 1)
543 		;	/* Wait for other cpu to see that we've started */
544 	stop_cpus((1<<cpu_reset_proxyid));
545 	printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
546 	DELAY(1000000);
547 	cpu_reset_real();
548 }
549 #endif
550 
551 void
552 cpu_reset()
553 {
554 #ifdef XBOX
555 	if (arch_i386_is_xbox) {
556 		/* Kick the PIC16L, it can reboot the box */
557 		pic16l_reboot();
558 		for (;;);
559 	}
560 #endif
561 
562 #ifdef SMP
563 	u_int cnt, map;
564 
565 	if (smp_active) {
566 		map = PCPU_GET(other_cpus) & ~stopped_cpus;
567 		if (map != 0) {
568 			printf("cpu_reset: Stopping other CPUs\n");
569 			stop_cpus(map);
570 		}
571 
572 		if (PCPU_GET(cpuid) != 0) {
573 			cpu_reset_proxyid = PCPU_GET(cpuid);
574 			cpustop_restartfunc = cpu_reset_proxy;
575 			cpu_reset_proxy_active = 0;
576 			printf("cpu_reset: Restarting BSP\n");
577 
578 			/* Restart CPU #0. */
579 			/* XXX: restart_cpus(1 << 0); */
580 			atomic_store_rel_int(&started_cpus, (1 << 0));
581 
582 			cnt = 0;
583 			while (cpu_reset_proxy_active == 0 && cnt < 10000000)
584 				cnt++;	/* Wait for BSP to announce restart */
585 			if (cpu_reset_proxy_active == 0)
586 				printf("cpu_reset: Failed to restart BSP\n");
587 			enable_intr();
588 			cpu_reset_proxy_active = 2;
589 
590 			while (1);
591 			/* NOTREACHED */
592 		}
593 
594 		DELAY(1000000);
595 	}
596 #endif
597 	cpu_reset_real();
598 	/* NOTREACHED */
599 }
600 
601 static void
602 cpu_reset_real()
603 {
604 	struct region_descriptor null_idt;
605 #ifndef PC98
606 	int b;
607 #endif
608 
609 	disable_intr();
610 #ifdef CPU_ELAN
611 	if (elan_mmcr != NULL)
612 		elan_mmcr->RESCFG = 1;
613 #endif
614 
615 	if (cpu == CPU_GEODE1100) {
616 		/* Attempt Geode's own reset */
617 		outl(0xcf8, 0x80009044ul);
618 		outl(0xcfc, 0xf);
619 	}
620 
621 #ifdef PC98
622 	/*
623 	 * Attempt to do a CPU reset via CPU reset port.
624 	 */
625 	if ((inb(0x35) & 0xa0) != 0xa0) {
626 		outb(0x37, 0x0f);		/* SHUT0 = 0. */
627 		outb(0x37, 0x0b);		/* SHUT1 = 0. */
628 	}
629 	outb(0xf0, 0x00);		/* Reset. */
630 #else
631 #if !defined(BROKEN_KEYBOARD_RESET)
632 	/*
633 	 * Attempt to do a CPU reset via the keyboard controller,
634 	 * do not turn off GateA20, as any machine that fails
635 	 * to do the reset here would then end up in no man's land.
636 	 */
637 	outb(IO_KBD + 4, 0xFE);
638 	DELAY(500000);	/* wait 0.5 sec to see if that did it */
639 #endif
640 
641 	/*
642 	 * Attempt to force a reset via the Reset Control register at
643 	 * I/O port 0xcf9.  Bit 2 forces a system reset when it is
644 	 * written as 1.  Bit 1 selects the type of reset to attempt:
645 	 * 0 selects a "soft" reset, and 1 selects a "hard" reset.  We
646 	 * try to do a "soft" reset first, and then a "hard" reset.
647 	 */
648 	outb(0xcf9, 0x2);
649 	outb(0xcf9, 0x6);
650 	DELAY(500000);  /* wait 0.5 sec to see if that did it */
651 
652 	/*
653 	 * Attempt to force a reset via the Fast A20 and Init register
654 	 * at I/O port 0x92.  Bit 1 serves as an alternate A20 gate.
655 	 * Bit 0 asserts INIT# when set to 1.  We are careful to only
656 	 * preserve bit 1 while setting bit 0.  We also must clear bit
657 	 * 0 before setting it if it isn't already clear.
658 	 */
659 	b = inb(0x92);
660 	if (b != 0xff) {
661 		if ((b & 0x1) != 0)
662 			outb(0x92, b & 0xfe);
663 		outb(0x92, b | 0x1);
664 		DELAY(500000);  /* wait 0.5 sec to see if that did it */
665 	}
666 #endif /* PC98 */
667 
668 	printf("No known reset method worked, attempting CPU shutdown\n");
669 	DELAY(1000000); /* wait 1 sec for printf to complete */
670 
671 	/* Wipe the IDT. */
672 	null_idt.rd_limit = 0;
673 	null_idt.rd_base = 0;
674 	lidt(&null_idt);
675 
676 	/* "good night, sweet prince .... <THUNK!>" */
677 	breakpoint();
678 
679 	/* NOTREACHED */
680 	while(1);
681 }
682 
683 /*
684  * Allocate a pool of sf_bufs (sendfile(2) or "super-fast" if you prefer. :-))
685  */
686 static void
687 sf_buf_init(void *arg)
688 {
689 	struct sf_buf *sf_bufs;
690 	vm_offset_t sf_base;
691 	int i;
692 
693 	nsfbufs = NSFBUFS;
694 	TUNABLE_INT_FETCH("kern.ipc.nsfbufs", &nsfbufs);
695 
696 	sf_buf_active = hashinit(nsfbufs, M_TEMP, &sf_buf_hashmask);
697 	TAILQ_INIT(&sf_buf_freelist);
698 	sf_base = kmem_alloc_nofault(kernel_map, nsfbufs * PAGE_SIZE);
699 	sf_bufs = malloc(nsfbufs * sizeof(struct sf_buf), M_TEMP,
700 	    M_NOWAIT | M_ZERO);
701 	for (i = 0; i < nsfbufs; i++) {
702 		sf_bufs[i].kva = sf_base + i * PAGE_SIZE;
703 		TAILQ_INSERT_TAIL(&sf_buf_freelist, &sf_bufs[i], free_entry);
704 	}
705 	sf_buf_alloc_want = 0;
706 	mtx_init(&sf_buf_lock, "sf_buf", NULL, MTX_DEF);
707 }
708 
709 /*
710  * Get an sf_buf from the freelist.  May block if none are available.
711  */
712 struct sf_buf *
713 sf_buf_alloc(struct vm_page *m, int flags)
714 {
715 	pt_entry_t opte, *ptep;
716 	struct sf_head *hash_list;
717 	struct sf_buf *sf;
718 #ifdef SMP
719 	cpumask_t cpumask, other_cpus;
720 #endif
721 	int error;
722 
723 	KASSERT(curthread->td_pinned > 0 || (flags & SFB_CPUPRIVATE) == 0,
724 	    ("sf_buf_alloc(SFB_CPUPRIVATE): curthread not pinned"));
725 	hash_list = &sf_buf_active[SF_BUF_HASH(m)];
726 	mtx_lock(&sf_buf_lock);
727 	LIST_FOREACH(sf, hash_list, list_entry) {
728 		if (sf->m == m) {
729 			sf->ref_count++;
730 			if (sf->ref_count == 1) {
731 				TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
732 				nsfbufsused++;
733 				nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
734 			}
735 #ifdef SMP
736 			goto shootdown;
737 #else
738 			goto done;
739 #endif
740 		}
741 	}
742 	while ((sf = TAILQ_FIRST(&sf_buf_freelist)) == NULL) {
743 		if (flags & SFB_NOWAIT)
744 			goto done;
745 		sf_buf_alloc_want++;
746 		mbstat.sf_allocwait++;
747 		error = msleep(&sf_buf_freelist, &sf_buf_lock,
748 		    (flags & SFB_CATCH) ? PCATCH | PVM : PVM, "sfbufa", 0);
749 		sf_buf_alloc_want--;
750 
751 		/*
752 		 * If we got a signal, don't risk going back to sleep.
753 		 */
754 		if (error)
755 			goto done;
756 	}
757 	TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
758 	if (sf->m != NULL)
759 		LIST_REMOVE(sf, list_entry);
760 	LIST_INSERT_HEAD(hash_list, sf, list_entry);
761 	sf->ref_count = 1;
762 	sf->m = m;
763 	nsfbufsused++;
764 	nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
765 
766 	/*
767 	 * Update the sf_buf's virtual-to-physical mapping, flushing the
768 	 * virtual address from the TLB.  Since the reference count for
769 	 * the sf_buf's old mapping was zero, that mapping is not
770 	 * currently in use.  Consequently, there is no need to exchange
771 	 * the old and new PTEs atomically, even under PAE.
772 	 */
773 	ptep = vtopte(sf->kva);
774 	opte = *ptep;
775 	*ptep = VM_PAGE_TO_PHYS(m) | pgeflag | PG_RW | PG_V;
776 
777 	/*
778 	 * Avoid unnecessary TLB invalidations: If the sf_buf's old
779 	 * virtual-to-physical mapping was not used, then any processor
780 	 * that has invalidated the sf_buf's virtual address from its TLB
781 	 * since the last used mapping need not invalidate again.
782 	 */
783 #ifdef SMP
784 	if ((opte & (PG_V | PG_A)) ==  (PG_V | PG_A))
785 		sf->cpumask = 0;
786 shootdown:
787 	sched_pin();
788 	cpumask = PCPU_GET(cpumask);
789 	if ((sf->cpumask & cpumask) == 0) {
790 		sf->cpumask |= cpumask;
791 		invlpg(sf->kva);
792 	}
793 	if ((flags & SFB_CPUPRIVATE) == 0) {
794 		other_cpus = PCPU_GET(other_cpus) & ~sf->cpumask;
795 		if (other_cpus != 0) {
796 			sf->cpumask |= other_cpus;
797 			smp_masked_invlpg(other_cpus, sf->kva);
798 		}
799 	}
800 	sched_unpin();
801 #else
802 	if ((opte & (PG_V | PG_A)) ==  (PG_V | PG_A))
803 		pmap_invalidate_page(kernel_pmap, sf->kva);
804 #endif
805 done:
806 	mtx_unlock(&sf_buf_lock);
807 	return (sf);
808 }
809 
810 /*
811  * Remove a reference from the given sf_buf, adding it to the free
812  * list when its reference count reaches zero.  A freed sf_buf still,
813  * however, retains its virtual-to-physical mapping until it is
814  * recycled or reactivated by sf_buf_alloc(9).
815  */
816 void
817 sf_buf_free(struct sf_buf *sf)
818 {
819 
820 	mtx_lock(&sf_buf_lock);
821 	sf->ref_count--;
822 	if (sf->ref_count == 0) {
823 		TAILQ_INSERT_TAIL(&sf_buf_freelist, sf, free_entry);
824 		nsfbufsused--;
825 		if (sf_buf_alloc_want > 0)
826 			wakeup_one(&sf_buf_freelist);
827 	}
828 	mtx_unlock(&sf_buf_lock);
829 }
830 
831 /*
832  * Software interrupt handler for queued VM system processing.
833  */
834 void
835 swi_vm(void *dummy)
836 {
837 	if (busdma_swi_pending != 0)
838 		busdma_swi();
839 }
840 
841 /*
842  * Tell whether this address is in some physical memory region.
843  * Currently used by the kernel coredump code in order to avoid
844  * dumping the ``ISA memory hole'' which could cause indefinite hangs,
845  * or other unpredictable behaviour.
846  */
847 
848 int
849 is_physical_memory(vm_paddr_t addr)
850 {
851 
852 #ifdef DEV_ISA
853 	/* The ISA ``memory hole''. */
854 	if (addr >= 0xa0000 && addr < 0x100000)
855 		return 0;
856 #endif
857 
858 	/*
859 	 * stuff other tests for known memory-mapped devices (PCI?)
860 	 * here
861 	 */
862 
863 	return 1;
864 }
865