xref: /freebsd/sys/i386/i386/vm_machdep.c (revision 195ebc7e9e4b129de810833791a19dfb4349d6a9)
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/kernel.h>
57 #include <sys/ktr.h>
58 #include <sys/lock.h>
59 #include <sys/malloc.h>
60 #include <sys/mbuf.h>
61 #include <sys/mutex.h>
62 #include <sys/pioctl.h>
63 #include <sys/proc.h>
64 #include <sys/sf_buf.h>
65 #include <sys/smp.h>
66 #include <sys/sched.h>
67 #include <sys/sysctl.h>
68 #include <sys/unistd.h>
69 #include <sys/vnode.h>
70 #include <sys/vmmeter.h>
71 
72 #include <machine/cpu.h>
73 #include <machine/cputypes.h>
74 #include <machine/md_var.h>
75 #include <machine/pcb.h>
76 #include <machine/pcb_ext.h>
77 #include <machine/smp.h>
78 #include <machine/vm86.h>
79 
80 #ifdef CPU_ELAN
81 #include <machine/elan_mmcr.h>
82 #endif
83 
84 #include <vm/vm.h>
85 #include <vm/vm_extern.h>
86 #include <vm/vm_kern.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_param.h>
90 
91 #ifdef XEN
92 #include <xen/hypervisor.h>
93 #endif
94 #ifdef PC98
95 #include <pc98/cbus/cbus.h>
96 #else
97 #include <i386/isa/isa.h>
98 #endif
99 
100 #ifdef XBOX
101 #include <machine/xbox.h>
102 #endif
103 
104 #ifndef NSFBUFS
105 #define	NSFBUFS		(512 + maxusers * 16)
106 #endif
107 
108 static void	cpu_reset_real(void);
109 #ifdef SMP
110 static void	cpu_reset_proxy(void);
111 static u_int	cpu_reset_proxyid;
112 static volatile u_int	cpu_reset_proxy_active;
113 #endif
114 static void	sf_buf_init(void *arg);
115 SYSINIT(sock_sf, SI_SUB_MBUF, SI_ORDER_ANY, sf_buf_init, NULL);
116 
117 LIST_HEAD(sf_head, sf_buf);
118 
119 /*
120  * A hash table of active sendfile(2) buffers
121  */
122 static struct sf_head *sf_buf_active;
123 static u_long sf_buf_hashmask;
124 
125 #define	SF_BUF_HASH(m)	(((m) - vm_page_array) & sf_buf_hashmask)
126 
127 static TAILQ_HEAD(, sf_buf) sf_buf_freelist;
128 static u_int	sf_buf_alloc_want;
129 
130 /*
131  * A lock used to synchronize access to the hash table and free list
132  */
133 static struct mtx sf_buf_lock;
134 
135 extern int	_ucodesel, _udatasel;
136 
137 /*
138  * Finish a fork operation, with process p2 nearly set up.
139  * Copy and update the pcb, set up the stack so that the child
140  * ready to run and return to user mode.
141  */
142 void
143 cpu_fork(td1, p2, td2, flags)
144 	register struct thread *td1;
145 	register struct proc *p2;
146 	struct thread *td2;
147 	int flags;
148 {
149 	register struct proc *p1;
150 	struct pcb *pcb2;
151 	struct mdproc *mdp2;
152 #ifdef DEV_NPX
153 	register_t savecrit;
154 #endif
155 
156 	p1 = td1->td_proc;
157 	if ((flags & RFPROC) == 0) {
158 		if ((flags & RFMEM) == 0) {
159 			/* unshare user LDT */
160 			struct mdproc *mdp1 = &p1->p_md;
161 			struct proc_ldt *pldt, *pldt1;
162 
163 			mtx_lock_spin(&dt_lock);
164 			if ((pldt1 = mdp1->md_ldt) != NULL &&
165 			    pldt1->ldt_refcnt > 1) {
166 				pldt = user_ldt_alloc(mdp1, pldt1->ldt_len);
167 				if (pldt == NULL)
168 					panic("could not copy LDT");
169 				mdp1->md_ldt = pldt;
170 				set_user_ldt(mdp1);
171 				user_ldt_deref(pldt1);
172 			} else
173 				mtx_unlock_spin(&dt_lock);
174 		}
175 		return;
176 	}
177 
178 	/* Ensure that p1's pcb is up to date. */
179 	if (td1 == curthread)
180 		td1->td_pcb->pcb_gs = rgs();
181 #ifdef DEV_NPX
182 	savecrit = intr_disable();
183 	if (PCPU_GET(fpcurthread) == td1)
184 		npxsave(&td1->td_pcb->pcb_save);
185 	intr_restore(savecrit);
186 #endif
187 
188 	/* Point the pcb to the top of the stack */
189 	pcb2 = (struct pcb *)(td2->td_kstack +
190 	    td2->td_kstack_pages * PAGE_SIZE) - 1;
191 	td2->td_pcb = pcb2;
192 
193 	/* Copy p1's pcb */
194 	bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
195 
196 	/* Point mdproc and then copy over td1's contents */
197 	mdp2 = &p2->p_md;
198 	bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
199 
200 	/*
201 	 * Create a new fresh stack for the new process.
202 	 * Copy the trap frame for the return to user mode as if from a
203 	 * syscall.  This copies most of the user mode register values.
204 	 * The -16 is so we can expand the trapframe if we go to vm86.
205 	 */
206 	td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb - 16) - 1;
207 	bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
208 
209 	td2->td_frame->tf_eax = 0;		/* Child returns zero */
210 	td2->td_frame->tf_eflags &= ~PSL_C;	/* success */
211 	td2->td_frame->tf_edx = 1;
212 
213 	/*
214 	 * If the parent process has the trap bit set (i.e. a debugger had
215 	 * single stepped the process to the system call), we need to clear
216 	 * the trap flag from the new frame unless the debugger had set PF_FORK
217 	 * on the parent.  Otherwise, the child will receive a (likely
218 	 * unexpected) SIGTRAP when it executes the first instruction after
219 	 * returning  to userland.
220 	 */
221 	if ((p1->p_pfsflags & PF_FORK) == 0)
222 		td2->td_frame->tf_eflags &= ~PSL_T;
223 
224 	/*
225 	 * Set registers for trampoline to user mode.  Leave space for the
226 	 * return address on stack.  These are the kernel mode register values.
227 	 */
228 #ifdef PAE
229 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdpt);
230 #else
231 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdir);
232 #endif
233 	pcb2->pcb_edi = 0;
234 	pcb2->pcb_esi = (int)fork_return;	/* fork_trampoline argument */
235 	pcb2->pcb_ebp = 0;
236 	pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *);
237 	pcb2->pcb_ebx = (int)td2;		/* fork_trampoline argument */
238 	pcb2->pcb_eip = (int)fork_trampoline;
239 	pcb2->pcb_psl = PSL_KERNEL;		/* ints disabled */
240 	/*-
241 	 * pcb2->pcb_dr*:	cloned above.
242 	 * pcb2->pcb_savefpu:	cloned above.
243 	 * pcb2->pcb_flags:	cloned above.
244 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
245 	 * pcb2->pcb_gs:	cloned above.
246 	 * pcb2->pcb_ext:	cleared below.
247 	 */
248 
249 	/*
250 	 * XXX don't copy the i/o pages.  this should probably be fixed.
251 	 */
252 	pcb2->pcb_ext = 0;
253 
254 	/* Copy the LDT, if necessary. */
255 	mtx_lock_spin(&dt_lock);
256 	if (mdp2->md_ldt != NULL) {
257 		if (flags & RFMEM) {
258 			mdp2->md_ldt->ldt_refcnt++;
259 		} else {
260 			mdp2->md_ldt = user_ldt_alloc(mdp2,
261 			    mdp2->md_ldt->ldt_len);
262 			if (mdp2->md_ldt == NULL)
263 				panic("could not copy LDT");
264 		}
265 	}
266 	mtx_unlock_spin(&dt_lock);
267 
268 	/* Setup to release spin count in fork_exit(). */
269 	td2->td_md.md_spinlock_count = 1;
270 	/*
271 	 * XXX XEN need to check on PSL_USER is handled
272 	 */
273 #ifdef XEN
274 	td2->td_md.md_saved_flags = 0;
275 #else
276 	td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
277 #endif
278 	/*
279 	 * Now, cpu_switch() can schedule the new process.
280 	 * pcb_esp is loaded pointing to the cpu_switch() stack frame
281 	 * containing the return address when exiting cpu_switch.
282 	 * This will normally be to fork_trampoline(), which will have
283 	 * %ebx loaded with the new proc's pointer.  fork_trampoline()
284 	 * will set up a stack to call fork_return(p, frame); to complete
285 	 * the return to user-mode.
286 	 */
287 }
288 
289 /*
290  * Intercept the return address from a freshly forked process that has NOT
291  * been scheduled yet.
292  *
293  * This is needed to make kernel threads stay in kernel mode.
294  */
295 void
296 cpu_set_fork_handler(td, func, arg)
297 	struct thread *td;
298 	void (*func)(void *);
299 	void *arg;
300 {
301 	/*
302 	 * Note that the trap frame follows the args, so the function
303 	 * is really called like this:  func(arg, frame);
304 	 */
305 	td->td_pcb->pcb_esi = (int) func;	/* function */
306 	td->td_pcb->pcb_ebx = (int) arg;	/* first arg */
307 }
308 
309 void
310 cpu_exit(struct thread *td)
311 {
312 
313 	/*
314 	 * If this process has a custom LDT, release it.  Reset pc->pcb_gs
315 	 * and %gs before we free it in case they refer to an LDT entry.
316 	 */
317 	mtx_lock_spin(&dt_lock);
318 	if (td->td_proc->p_md.md_ldt) {
319 		td->td_pcb->pcb_gs = _udatasel;
320 		load_gs(_udatasel);
321 		user_ldt_free(td);
322 	} else
323 		mtx_unlock_spin(&dt_lock);
324 }
325 
326 void
327 cpu_thread_exit(struct thread *td)
328 {
329 
330 #ifdef DEV_NPX
331 	if (td == PCPU_GET(fpcurthread))
332 		npxdrop();
333 #endif
334 
335 	/* Disable any hardware breakpoints. */
336 	if (td->td_pcb->pcb_flags & PCB_DBREGS) {
337 		reset_dbregs();
338 		td->td_pcb->pcb_flags &= ~PCB_DBREGS;
339 	}
340 }
341 
342 void
343 cpu_thread_clean(struct thread *td)
344 {
345 	struct pcb *pcb;
346 
347 	pcb = td->td_pcb;
348 	if (pcb->pcb_ext != NULL) {
349 		/* if (pcb->pcb_ext->ext_refcount-- == 1) ?? */
350 		/*
351 		 * XXX do we need to move the TSS off the allocated pages
352 		 * before freeing them?  (not done here)
353 		 */
354 		kmem_free(kernel_map, (vm_offset_t)pcb->pcb_ext,
355 		    ctob(IOPAGES + 1));
356 		pcb->pcb_ext = NULL;
357 	}
358 }
359 
360 void
361 cpu_thread_swapin(struct thread *td)
362 {
363 }
364 
365 void
366 cpu_thread_swapout(struct thread *td)
367 {
368 }
369 
370 void
371 cpu_thread_alloc(struct thread *td)
372 {
373 
374 	td->td_pcb = (struct pcb *)(td->td_kstack +
375 	    td->td_kstack_pages * PAGE_SIZE) - 1;
376 	td->td_frame = (struct trapframe *)((caddr_t)td->td_pcb - 16) - 1;
377 	td->td_pcb->pcb_ext = NULL;
378 }
379 
380 void
381 cpu_thread_free(struct thread *td)
382 {
383 
384 	cpu_thread_clean(td);
385 }
386 
387 /*
388  * Initialize machine state (pcb and trap frame) for a new thread about to
389  * upcall. Put enough state in the new thread's PCB to get it to go back
390  * userret(), where we can intercept it again to set the return (upcall)
391  * Address and stack, along with those from upcals that are from other sources
392  * such as those generated in thread_userret() itself.
393  */
394 void
395 cpu_set_upcall(struct thread *td, struct thread *td0)
396 {
397 	struct pcb *pcb2;
398 
399 	/* Point the pcb to the top of the stack. */
400 	pcb2 = td->td_pcb;
401 
402 	/*
403 	 * Copy the upcall pcb.  This loads kernel regs.
404 	 * Those not loaded individually below get their default
405 	 * values here.
406 	 */
407 	bcopy(td0->td_pcb, pcb2, sizeof(*pcb2));
408 	pcb2->pcb_flags &= ~(PCB_NPXTRAP|PCB_NPXINITDONE);
409 
410 	/*
411 	 * Create a new fresh stack for the new thread.
412 	 */
413 	bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe));
414 
415 	/* If the current thread has the trap bit set (i.e. a debugger had
416 	 * single stepped the process to the system call), we need to clear
417 	 * the trap flag from the new frame. Otherwise, the new thread will
418 	 * receive a (likely unexpected) SIGTRAP when it executes the first
419 	 * instruction after returning to userland.
420 	 */
421 	td->td_frame->tf_eflags &= ~PSL_T;
422 
423 	/*
424 	 * Set registers for trampoline to user mode.  Leave space for the
425 	 * return address on stack.  These are the kernel mode register values.
426 	 */
427 	pcb2->pcb_edi = 0;
428 	pcb2->pcb_esi = (int)fork_return;		    /* trampoline arg */
429 	pcb2->pcb_ebp = 0;
430 	pcb2->pcb_esp = (int)td->td_frame - sizeof(void *); /* trampoline arg */
431 	pcb2->pcb_ebx = (int)td;			    /* trampoline arg */
432 	pcb2->pcb_eip = (int)fork_trampoline;
433 	pcb2->pcb_psl &= ~(PSL_I);	/* interrupts must be disabled */
434 	pcb2->pcb_gs = rgs();
435 	/*
436 	 * If we didn't copy the pcb, we'd need to do the following registers:
437 	 * pcb2->pcb_cr3:	cloned above.
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.
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 #ifdef XEN
450 	td->td_md.md_saved_flags = 0;
451 #else
452 	td->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
453 #endif
454 }
455 
456 /*
457  * Set that machine state for performing an upcall that has to
458  * be done in thread_userret() so that those upcalls generated
459  * in thread_userret() itself can be done as well.
460  */
461 void
462 cpu_set_upcall_kse(struct thread *td, void (*entry)(void *), void *arg,
463 	stack_t *stack)
464 {
465 
466 	/*
467 	 * Do any extra cleaning that needs to be done.
468 	 * The thread may have optional components
469 	 * that are not present in a fresh thread.
470 	 * This may be a recycled thread so make it look
471 	 * as though it's newly allocated.
472 	 */
473 	cpu_thread_clean(td);
474 
475 	/*
476 	 * Set the trap frame to point at the beginning of the uts
477 	 * function.
478 	 */
479 	td->td_frame->tf_ebp = 0;
480 	td->td_frame->tf_esp =
481 	    (((int)stack->ss_sp + stack->ss_size - 4) & ~0x0f) - 4;
482 	td->td_frame->tf_eip = (int)entry;
483 
484 	/*
485 	 * Pass the address of the mailbox for this kse to the uts
486 	 * function as a parameter on the stack.
487 	 */
488 	suword((void *)(td->td_frame->tf_esp + sizeof(void *)),
489 	    (int)arg);
490 }
491 
492 int
493 cpu_set_user_tls(struct thread *td, void *tls_base)
494 {
495 	struct segment_descriptor sd;
496 	uint32_t base;
497 
498 	/*
499 	 * Construct a descriptor and store it in the pcb for
500 	 * the next context switch.  Also store it in the gdt
501 	 * so that the load of tf_fs into %fs will activate it
502 	 * at return to userland.
503 	 */
504 	base = (uint32_t)tls_base;
505 	sd.sd_lobase = base & 0xffffff;
506 	sd.sd_hibase = (base >> 24) & 0xff;
507 	sd.sd_lolimit = 0xffff;	/* 4GB limit, wraps around */
508 	sd.sd_hilimit = 0xf;
509 	sd.sd_type  = SDT_MEMRWA;
510 	sd.sd_dpl   = SEL_UPL;
511 	sd.sd_p     = 1;
512 	sd.sd_xx    = 0;
513 	sd.sd_def32 = 1;
514 	sd.sd_gran  = 1;
515 	critical_enter();
516 	/* set %gs */
517 	td->td_pcb->pcb_gsd = sd;
518 	if (td == curthread) {
519 		PCPU_GET(fsgs_gdt)[1] = sd;
520 		load_gs(GSEL(GUGS_SEL, SEL_UPL));
521 	}
522 	critical_exit();
523 	return (0);
524 }
525 
526 /*
527  * Convert kernel VA to physical address
528  */
529 vm_paddr_t
530 kvtop(void *addr)
531 {
532 	vm_paddr_t pa;
533 
534 	pa = pmap_kextract((vm_offset_t)addr);
535 	if (pa == 0)
536 		panic("kvtop: zero page frame");
537 	return (pa);
538 }
539 
540 #ifdef SMP
541 static void
542 cpu_reset_proxy()
543 {
544 
545 	cpu_reset_proxy_active = 1;
546 	while (cpu_reset_proxy_active == 1)
547 		;	/* Wait for other cpu to see that we've started */
548 	stop_cpus((1<<cpu_reset_proxyid));
549 	printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
550 	DELAY(1000000);
551 	cpu_reset_real();
552 }
553 #endif
554 
555 void
556 cpu_reset()
557 {
558 #ifdef XBOX
559 	if (arch_i386_is_xbox) {
560 		/* Kick the PIC16L, it can reboot the box */
561 		pic16l_reboot();
562 		for (;;);
563 	}
564 #endif
565 
566 #ifdef SMP
567 	u_int cnt, map;
568 
569 	if (smp_active) {
570 		map = PCPU_GET(other_cpus) & ~stopped_cpus;
571 		if (map != 0) {
572 			printf("cpu_reset: Stopping other CPUs\n");
573 			stop_cpus(map);
574 		}
575 
576 		if (PCPU_GET(cpuid) != 0) {
577 			cpu_reset_proxyid = PCPU_GET(cpuid);
578 			cpustop_restartfunc = cpu_reset_proxy;
579 			cpu_reset_proxy_active = 0;
580 			printf("cpu_reset: Restarting BSP\n");
581 
582 			/* Restart CPU #0. */
583 			/* XXX: restart_cpus(1 << 0); */
584 			atomic_store_rel_int(&started_cpus, (1 << 0));
585 
586 			cnt = 0;
587 			while (cpu_reset_proxy_active == 0 && cnt < 10000000)
588 				cnt++;	/* Wait for BSP to announce restart */
589 			if (cpu_reset_proxy_active == 0)
590 				printf("cpu_reset: Failed to restart BSP\n");
591 			enable_intr();
592 			cpu_reset_proxy_active = 2;
593 
594 			while (1);
595 			/* NOTREACHED */
596 		}
597 
598 		DELAY(1000000);
599 	}
600 #endif
601 	cpu_reset_real();
602 	/* NOTREACHED */
603 }
604 
605 static void
606 cpu_reset_real()
607 {
608 	struct region_descriptor null_idt;
609 #ifndef PC98
610 	int b;
611 #endif
612 
613 	disable_intr();
614 #ifdef XEN
615 	if (smp_processor_id() == 0)
616 		HYPERVISOR_shutdown(SHUTDOWN_reboot);
617 	else
618 		HYPERVISOR_shutdown(SHUTDOWN_poweroff);
619 #endif
620 #ifdef CPU_ELAN
621 	if (elan_mmcr != NULL)
622 		elan_mmcr->RESCFG = 1;
623 #endif
624 
625 	if (cpu == CPU_GEODE1100) {
626 		/* Attempt Geode's own reset */
627 		outl(0xcf8, 0x80009044ul);
628 		outl(0xcfc, 0xf);
629 	}
630 
631 #ifdef PC98
632 	/*
633 	 * Attempt to do a CPU reset via CPU reset port.
634 	 */
635 	if ((inb(0x35) & 0xa0) != 0xa0) {
636 		outb(0x37, 0x0f);		/* SHUT0 = 0. */
637 		outb(0x37, 0x0b);		/* SHUT1 = 0. */
638 	}
639 	outb(0xf0, 0x00);		/* Reset. */
640 #else
641 #if !defined(BROKEN_KEYBOARD_RESET)
642 	/*
643 	 * Attempt to do a CPU reset via the keyboard controller,
644 	 * do not turn off GateA20, as any machine that fails
645 	 * to do the reset here would then end up in no man's land.
646 	 */
647 	outb(IO_KBD + 4, 0xFE);
648 	DELAY(500000);	/* wait 0.5 sec to see if that did it */
649 #endif
650 
651 	/*
652 	 * Attempt to force a reset via the Reset Control register at
653 	 * I/O port 0xcf9.  Bit 2 forces a system reset when it
654 	 * transitions from 0 to 1.  Bit 1 selects the type of reset
655 	 * to attempt: 0 selects a "soft" reset, and 1 selects a
656 	 * "hard" reset.  We try a "hard" reset.  The first write sets
657 	 * bit 1 to select a "hard" reset and clears bit 2.  The
658 	 * second write forces a 0 -> 1 transition in bit 2 to trigger
659 	 * a reset.
660 	 */
661 	outb(0xcf9, 0x2);
662 	outb(0xcf9, 0x6);
663 	DELAY(500000);  /* wait 0.5 sec to see if that did it */
664 
665 	/*
666 	 * Attempt to force a reset via the Fast A20 and Init register
667 	 * at I/O port 0x92.  Bit 1 serves as an alternate A20 gate.
668 	 * Bit 0 asserts INIT# when set to 1.  We are careful to only
669 	 * preserve bit 1 while setting bit 0.  We also must clear bit
670 	 * 0 before setting it if it isn't already clear.
671 	 */
672 	b = inb(0x92);
673 	if (b != 0xff) {
674 		if ((b & 0x1) != 0)
675 			outb(0x92, b & 0xfe);
676 		outb(0x92, b | 0x1);
677 		DELAY(500000);  /* wait 0.5 sec to see if that did it */
678 	}
679 #endif /* PC98 */
680 
681 	printf("No known reset method worked, attempting CPU shutdown\n");
682 	DELAY(1000000); /* wait 1 sec for printf to complete */
683 
684 	/* Wipe the IDT. */
685 	null_idt.rd_limit = 0;
686 	null_idt.rd_base = 0;
687 	lidt(&null_idt);
688 
689 	/* "good night, sweet prince .... <THUNK!>" */
690 	breakpoint();
691 
692 	/* NOTREACHED */
693 	while(1);
694 }
695 
696 /*
697  * Allocate a pool of sf_bufs (sendfile(2) or "super-fast" if you prefer. :-))
698  */
699 static void
700 sf_buf_init(void *arg)
701 {
702 	struct sf_buf *sf_bufs;
703 	vm_offset_t sf_base;
704 	int i;
705 
706 	nsfbufs = NSFBUFS;
707 	TUNABLE_INT_FETCH("kern.ipc.nsfbufs", &nsfbufs);
708 
709 	sf_buf_active = hashinit(nsfbufs, M_TEMP, &sf_buf_hashmask);
710 	TAILQ_INIT(&sf_buf_freelist);
711 	sf_base = kmem_alloc_nofault(kernel_map, nsfbufs * PAGE_SIZE);
712 	sf_bufs = malloc(nsfbufs * sizeof(struct sf_buf), M_TEMP,
713 	    M_NOWAIT | M_ZERO);
714 	for (i = 0; i < nsfbufs; i++) {
715 		sf_bufs[i].kva = sf_base + i * PAGE_SIZE;
716 		TAILQ_INSERT_TAIL(&sf_buf_freelist, &sf_bufs[i], free_entry);
717 	}
718 	sf_buf_alloc_want = 0;
719 	mtx_init(&sf_buf_lock, "sf_buf", NULL, MTX_DEF);
720 }
721 
722 /*
723  * Get an sf_buf from the freelist.  May block if none are available.
724  */
725 struct sf_buf *
726 sf_buf_alloc(struct vm_page *m, int flags)
727 {
728 	pt_entry_t opte, *ptep;
729 	struct sf_head *hash_list;
730 	struct sf_buf *sf;
731 #ifdef SMP
732 	cpumask_t cpumask, other_cpus;
733 #endif
734 	int error;
735 
736 	KASSERT(curthread->td_pinned > 0 || (flags & SFB_CPUPRIVATE) == 0,
737 	    ("sf_buf_alloc(SFB_CPUPRIVATE): curthread not pinned"));
738 	hash_list = &sf_buf_active[SF_BUF_HASH(m)];
739 	mtx_lock(&sf_buf_lock);
740 	LIST_FOREACH(sf, hash_list, list_entry) {
741 		if (sf->m == m) {
742 			sf->ref_count++;
743 			if (sf->ref_count == 1) {
744 				TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
745 				nsfbufsused++;
746 				nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
747 			}
748 #ifdef SMP
749 			goto shootdown;
750 #else
751 			goto done;
752 #endif
753 		}
754 	}
755 	while ((sf = TAILQ_FIRST(&sf_buf_freelist)) == NULL) {
756 		if (flags & SFB_NOWAIT)
757 			goto done;
758 		sf_buf_alloc_want++;
759 		mbstat.sf_allocwait++;
760 		error = msleep(&sf_buf_freelist, &sf_buf_lock,
761 		    (flags & SFB_CATCH) ? PCATCH | PVM : PVM, "sfbufa", 0);
762 		sf_buf_alloc_want--;
763 
764 		/*
765 		 * If we got a signal, don't risk going back to sleep.
766 		 */
767 		if (error)
768 			goto done;
769 	}
770 	TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
771 	if (sf->m != NULL)
772 		LIST_REMOVE(sf, list_entry);
773 	LIST_INSERT_HEAD(hash_list, sf, list_entry);
774 	sf->ref_count = 1;
775 	sf->m = m;
776 	nsfbufsused++;
777 	nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
778 
779 	/*
780 	 * Update the sf_buf's virtual-to-physical mapping, flushing the
781 	 * virtual address from the TLB.  Since the reference count for
782 	 * the sf_buf's old mapping was zero, that mapping is not
783 	 * currently in use.  Consequently, there is no need to exchange
784 	 * the old and new PTEs atomically, even under PAE.
785 	 */
786 	ptep = vtopte(sf->kva);
787 	opte = *ptep;
788 #ifdef XEN
789        PT_SET_MA(sf->kva, xpmap_ptom(VM_PAGE_TO_PHYS(m)) | pgeflag
790 	   | PG_RW | PG_V);
791 #else
792 	*ptep = VM_PAGE_TO_PHYS(m) | pgeflag | PG_RW | PG_V;
793 #endif
794 
795 	/*
796 	 * Avoid unnecessary TLB invalidations: If the sf_buf's old
797 	 * virtual-to-physical mapping was not used, then any processor
798 	 * that has invalidated the sf_buf's virtual address from its TLB
799 	 * since the last used mapping need not invalidate again.
800 	 */
801 #ifdef SMP
802 	if ((opte & (PG_V | PG_A)) ==  (PG_V | PG_A))
803 		sf->cpumask = 0;
804 shootdown:
805 	sched_pin();
806 	cpumask = PCPU_GET(cpumask);
807 	if ((sf->cpumask & cpumask) == 0) {
808 		sf->cpumask |= cpumask;
809 		invlpg(sf->kva);
810 	}
811 	if ((flags & SFB_CPUPRIVATE) == 0) {
812 		other_cpus = PCPU_GET(other_cpus) & ~sf->cpumask;
813 		if (other_cpus != 0) {
814 			sf->cpumask |= other_cpus;
815 			smp_masked_invlpg(other_cpus, sf->kva);
816 		}
817 	}
818 	sched_unpin();
819 #else
820 	if ((opte & (PG_V | PG_A)) ==  (PG_V | PG_A))
821 		pmap_invalidate_page(kernel_pmap, sf->kva);
822 #endif
823 done:
824 	mtx_unlock(&sf_buf_lock);
825 	return (sf);
826 }
827 
828 /*
829  * Remove a reference from the given sf_buf, adding it to the free
830  * list when its reference count reaches zero.  A freed sf_buf still,
831  * however, retains its virtual-to-physical mapping until it is
832  * recycled or reactivated by sf_buf_alloc(9).
833  */
834 void
835 sf_buf_free(struct sf_buf *sf)
836 {
837 
838 	mtx_lock(&sf_buf_lock);
839 	sf->ref_count--;
840 	if (sf->ref_count == 0) {
841 		TAILQ_INSERT_TAIL(&sf_buf_freelist, sf, free_entry);
842 		nsfbufsused--;
843 #ifdef XEN
844 /*
845  * Xen doesn't like having dangling R/W mappings
846  */
847 		pmap_qremove(sf->kva, 1);
848 		sf->m = NULL;
849 		LIST_REMOVE(sf, list_entry);
850 #endif
851 		if (sf_buf_alloc_want > 0)
852 			wakeup_one(&sf_buf_freelist);
853 	}
854 	mtx_unlock(&sf_buf_lock);
855 }
856 
857 /*
858  * Software interrupt handler for queued VM system processing.
859  */
860 void
861 swi_vm(void *dummy)
862 {
863 	if (busdma_swi_pending != 0)
864 		busdma_swi();
865 }
866 
867 /*
868  * Tell whether this address is in some physical memory region.
869  * Currently used by the kernel coredump code in order to avoid
870  * dumping the ``ISA memory hole'' which could cause indefinite hangs,
871  * or other unpredictable behaviour.
872  */
873 
874 int
875 is_physical_memory(vm_paddr_t addr)
876 {
877 
878 #ifdef DEV_ISA
879 	/* The ISA ``memory hole''. */
880 	if (addr >= 0xa0000 && addr < 0x100000)
881 		return 0;
882 #endif
883 
884 	/*
885 	 * stuff other tests for known memory-mapped devices (PCI?)
886 	 * here
887 	 */
888 
889 	return 1;
890 }
891