xref: /freebsd/sys/i386/i386/vm_machdep.c (revision 38f0b757fd84d17d0fc24739a7cda160c4516d81)
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/sysent.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 XEN
93 #include <xen/hypervisor.h>
94 #endif
95 #ifdef PC98
96 #include <pc98/cbus/cbus.h>
97 #else
98 #include <isa/isareg.h>
99 #endif
100 
101 #ifdef XBOX
102 #include <machine/xbox.h>
103 #endif
104 
105 #ifndef NSFBUFS
106 #define	NSFBUFS		(512 + maxusers * 16)
107 #endif
108 
109 _Static_assert(OFFSETOF_CURTHREAD == offsetof(struct pcpu, pc_curthread),
110     "OFFSETOF_CURTHREAD does not correspond with offset of pc_curthread.");
111 _Static_assert(OFFSETOF_CURPCB == offsetof(struct pcpu, pc_curpcb),
112     "OFFSETOF_CURPCB does not correspond with offset of pc_curpcb.");
113 
114 static void	cpu_reset_real(void);
115 #ifdef SMP
116 static void	cpu_reset_proxy(void);
117 static u_int	cpu_reset_proxyid;
118 static volatile u_int	cpu_reset_proxy_active;
119 #endif
120 
121 static int nsfbufs;
122 static int nsfbufspeak;
123 static int nsfbufsused;
124 
125 SYSCTL_INT(_kern_ipc, OID_AUTO, nsfbufs, CTLFLAG_RDTUN, &nsfbufs, 0,
126     "Maximum number of sendfile(2) sf_bufs available");
127 SYSCTL_INT(_kern_ipc, OID_AUTO, nsfbufspeak, CTLFLAG_RD, &nsfbufspeak, 0,
128     "Number of sendfile(2) sf_bufs at peak usage");
129 SYSCTL_INT(_kern_ipc, OID_AUTO, nsfbufsused, CTLFLAG_RD, &nsfbufsused, 0,
130     "Number of sendfile(2) sf_bufs in use");
131 
132 static void	sf_buf_init(void *arg);
133 SYSINIT(sock_sf, SI_SUB_MBUF, SI_ORDER_ANY, sf_buf_init, NULL);
134 
135 LIST_HEAD(sf_head, sf_buf);
136 
137 /*
138  * A hash table of active sendfile(2) buffers
139  */
140 static struct sf_head *sf_buf_active;
141 static u_long sf_buf_hashmask;
142 
143 #define	SF_BUF_HASH(m)	(((m) - vm_page_array) & sf_buf_hashmask)
144 
145 static TAILQ_HEAD(, sf_buf) sf_buf_freelist;
146 static u_int	sf_buf_alloc_want;
147 
148 /*
149  * A lock used to synchronize access to the hash table and free list
150  */
151 static struct mtx sf_buf_lock;
152 
153 extern int	_ucodesel, _udatasel;
154 
155 /*
156  * Finish a fork operation, with process p2 nearly set up.
157  * Copy and update the pcb, set up the stack so that the child
158  * ready to run and return to user mode.
159  */
160 void
161 cpu_fork(td1, p2, td2, flags)
162 	register struct thread *td1;
163 	register struct proc *p2;
164 	struct thread *td2;
165 	int flags;
166 {
167 	register struct proc *p1;
168 	struct pcb *pcb2;
169 	struct mdproc *mdp2;
170 
171 	p1 = td1->td_proc;
172 	if ((flags & RFPROC) == 0) {
173 		if ((flags & RFMEM) == 0) {
174 			/* unshare user LDT */
175 			struct mdproc *mdp1 = &p1->p_md;
176 			struct proc_ldt *pldt, *pldt1;
177 
178 			mtx_lock_spin(&dt_lock);
179 			if ((pldt1 = mdp1->md_ldt) != NULL &&
180 			    pldt1->ldt_refcnt > 1) {
181 				pldt = user_ldt_alloc(mdp1, pldt1->ldt_len);
182 				if (pldt == NULL)
183 					panic("could not copy LDT");
184 				mdp1->md_ldt = pldt;
185 				set_user_ldt(mdp1);
186 				user_ldt_deref(pldt1);
187 			} else
188 				mtx_unlock_spin(&dt_lock);
189 		}
190 		return;
191 	}
192 
193 	/* Ensure that td1's pcb is up to date. */
194 	if (td1 == curthread)
195 		td1->td_pcb->pcb_gs = rgs();
196 #ifdef DEV_NPX
197 	critical_enter();
198 	if (PCPU_GET(fpcurthread) == td1)
199 		npxsave(td1->td_pcb->pcb_save);
200 	critical_exit();
201 #endif
202 
203 	/* Point the pcb to the top of the stack */
204 	pcb2 = (struct pcb *)(td2->td_kstack +
205 	    td2->td_kstack_pages * PAGE_SIZE) - 1;
206 	td2->td_pcb = pcb2;
207 
208 	/* Copy td1's pcb */
209 	bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
210 
211 	/* Properly initialize pcb_save */
212 	pcb2->pcb_save = &pcb2->pcb_user_save;
213 
214 	/* Point mdproc and then copy over td1's contents */
215 	mdp2 = &p2->p_md;
216 	bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
217 
218 	/*
219 	 * Create a new fresh stack for the new process.
220 	 * Copy the trap frame for the return to user mode as if from a
221 	 * syscall.  This copies most of the user mode register values.
222 	 * The -16 is so we can expand the trapframe if we go to vm86.
223 	 */
224 	td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb - 16) - 1;
225 	bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
226 
227 	td2->td_frame->tf_eax = 0;		/* Child returns zero */
228 	td2->td_frame->tf_eflags &= ~PSL_C;	/* success */
229 	td2->td_frame->tf_edx = 1;
230 
231 	/*
232 	 * If the parent process has the trap bit set (i.e. a debugger had
233 	 * single stepped the process to the system call), we need to clear
234 	 * the trap flag from the new frame unless the debugger had set PF_FORK
235 	 * on the parent.  Otherwise, the child will receive a (likely
236 	 * unexpected) SIGTRAP when it executes the first instruction after
237 	 * returning  to userland.
238 	 */
239 	if ((p1->p_pfsflags & PF_FORK) == 0)
240 		td2->td_frame->tf_eflags &= ~PSL_T;
241 
242 	/*
243 	 * Set registers for trampoline to user mode.  Leave space for the
244 	 * return address on stack.  These are the kernel mode register values.
245 	 */
246 #ifdef PAE
247 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdpt);
248 #else
249 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdir);
250 #endif
251 	pcb2->pcb_edi = 0;
252 	pcb2->pcb_esi = (int)fork_return;	/* fork_trampoline argument */
253 	pcb2->pcb_ebp = 0;
254 	pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *);
255 	pcb2->pcb_ebx = (int)td2;		/* fork_trampoline argument */
256 	pcb2->pcb_eip = (int)fork_trampoline;
257 	pcb2->pcb_psl = PSL_KERNEL;		/* ints disabled */
258 	/*-
259 	 * pcb2->pcb_dr*:	cloned above.
260 	 * pcb2->pcb_savefpu:	cloned above.
261 	 * pcb2->pcb_flags:	cloned above.
262 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
263 	 * pcb2->pcb_gs:	cloned above.
264 	 * pcb2->pcb_ext:	cleared below.
265 	 */
266 
267 	/*
268 	 * XXX don't copy the i/o pages.  this should probably be fixed.
269 	 */
270 	pcb2->pcb_ext = 0;
271 
272 	/* Copy the LDT, if necessary. */
273 	mtx_lock_spin(&dt_lock);
274 	if (mdp2->md_ldt != NULL) {
275 		if (flags & RFMEM) {
276 			mdp2->md_ldt->ldt_refcnt++;
277 		} else {
278 			mdp2->md_ldt = user_ldt_alloc(mdp2,
279 			    mdp2->md_ldt->ldt_len);
280 			if (mdp2->md_ldt == NULL)
281 				panic("could not copy LDT");
282 		}
283 	}
284 	mtx_unlock_spin(&dt_lock);
285 
286 	/* Setup to release spin count in fork_exit(). */
287 	td2->td_md.md_spinlock_count = 1;
288 	/*
289 	 * XXX XEN need to check on PSL_USER is handled
290 	 */
291 	td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
292 	/*
293 	 * Now, cpu_switch() can schedule the new process.
294 	 * pcb_esp is loaded pointing to the cpu_switch() stack frame
295 	 * containing the return address when exiting cpu_switch.
296 	 * This will normally be to fork_trampoline(), which will have
297 	 * %ebx loaded with the new proc's pointer.  fork_trampoline()
298 	 * will set up a stack to call fork_return(p, frame); to complete
299 	 * the return to user-mode.
300 	 */
301 }
302 
303 /*
304  * Intercept the return address from a freshly forked process that has NOT
305  * been scheduled yet.
306  *
307  * This is needed to make kernel threads stay in kernel mode.
308  */
309 void
310 cpu_set_fork_handler(td, func, arg)
311 	struct thread *td;
312 	void (*func)(void *);
313 	void *arg;
314 {
315 	/*
316 	 * Note that the trap frame follows the args, so the function
317 	 * is really called like this:  func(arg, frame);
318 	 */
319 	td->td_pcb->pcb_esi = (int) func;	/* function */
320 	td->td_pcb->pcb_ebx = (int) arg;	/* first arg */
321 }
322 
323 void
324 cpu_exit(struct thread *td)
325 {
326 
327 	/*
328 	 * If this process has a custom LDT, release it.  Reset pc->pcb_gs
329 	 * and %gs before we free it in case they refer to an LDT entry.
330 	 */
331 	mtx_lock_spin(&dt_lock);
332 	if (td->td_proc->p_md.md_ldt) {
333 		td->td_pcb->pcb_gs = _udatasel;
334 		load_gs(_udatasel);
335 		user_ldt_free(td);
336 	} else
337 		mtx_unlock_spin(&dt_lock);
338 }
339 
340 void
341 cpu_thread_exit(struct thread *td)
342 {
343 
344 #ifdef DEV_NPX
345 	critical_enter();
346 	if (td == PCPU_GET(fpcurthread))
347 		npxdrop();
348 	critical_exit();
349 #endif
350 
351 	/* Disable any hardware breakpoints. */
352 	if (td->td_pcb->pcb_flags & PCB_DBREGS) {
353 		reset_dbregs();
354 		td->td_pcb->pcb_flags &= ~PCB_DBREGS;
355 	}
356 }
357 
358 void
359 cpu_thread_clean(struct thread *td)
360 {
361 	struct pcb *pcb;
362 
363 	pcb = td->td_pcb;
364 	if (pcb->pcb_ext != NULL) {
365 		/* if (pcb->pcb_ext->ext_refcount-- == 1) ?? */
366 		/*
367 		 * XXX do we need to move the TSS off the allocated pages
368 		 * before freeing them?  (not done here)
369 		 */
370 		kmem_free(kernel_arena, (vm_offset_t)pcb->pcb_ext,
371 		    ctob(IOPAGES + 1));
372 		pcb->pcb_ext = NULL;
373 	}
374 }
375 
376 void
377 cpu_thread_swapin(struct thread *td)
378 {
379 }
380 
381 void
382 cpu_thread_swapout(struct thread *td)
383 {
384 }
385 
386 void
387 cpu_thread_alloc(struct thread *td)
388 {
389 
390 	td->td_pcb = (struct pcb *)(td->td_kstack +
391 	    td->td_kstack_pages * PAGE_SIZE) - 1;
392 	td->td_frame = (struct trapframe *)((caddr_t)td->td_pcb - 16) - 1;
393 	td->td_pcb->pcb_ext = NULL;
394 	td->td_pcb->pcb_save = &td->td_pcb->pcb_user_save;
395 }
396 
397 void
398 cpu_thread_free(struct thread *td)
399 {
400 
401 	cpu_thread_clean(td);
402 }
403 
404 void
405 cpu_set_syscall_retval(struct thread *td, int error)
406 {
407 
408 	switch (error) {
409 	case 0:
410 		td->td_frame->tf_eax = td->td_retval[0];
411 		td->td_frame->tf_edx = td->td_retval[1];
412 		td->td_frame->tf_eflags &= ~PSL_C;
413 		break;
414 
415 	case ERESTART:
416 		/*
417 		 * Reconstruct pc, assuming lcall $X,y is 7 bytes, int
418 		 * 0x80 is 2 bytes. We saved this in tf_err.
419 		 */
420 		td->td_frame->tf_eip -= td->td_frame->tf_err;
421 		break;
422 
423 	case EJUSTRETURN:
424 		break;
425 
426 	default:
427 		if (td->td_proc->p_sysent->sv_errsize) {
428 			if (error >= td->td_proc->p_sysent->sv_errsize)
429 				error = -1;	/* XXX */
430 			else
431 				error = td->td_proc->p_sysent->sv_errtbl[error];
432 		}
433 		td->td_frame->tf_eax = error;
434 		td->td_frame->tf_eflags |= PSL_C;
435 		break;
436 	}
437 }
438 
439 /*
440  * Initialize machine state (pcb and trap frame) for a new thread about to
441  * upcall. Put enough state in the new thread's PCB to get it to go back
442  * userret(), where we can intercept it again to set the return (upcall)
443  * Address and stack, along with those from upcals that are from other sources
444  * such as those generated in thread_userret() itself.
445  */
446 void
447 cpu_set_upcall(struct thread *td, struct thread *td0)
448 {
449 	struct pcb *pcb2;
450 
451 	/* Point the pcb to the top of the stack. */
452 	pcb2 = td->td_pcb;
453 
454 	/*
455 	 * Copy the upcall pcb.  This loads kernel regs.
456 	 * Those not loaded individually below get their default
457 	 * values here.
458 	 */
459 	bcopy(td0->td_pcb, pcb2, sizeof(*pcb2));
460 	pcb2->pcb_flags &= ~(PCB_NPXINITDONE | PCB_NPXUSERINITDONE |
461 	    PCB_KERNNPX);
462 	pcb2->pcb_save = &pcb2->pcb_user_save;
463 
464 	/*
465 	 * Create a new fresh stack for the new thread.
466 	 */
467 	bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe));
468 
469 	/* If the current thread has the trap bit set (i.e. a debugger had
470 	 * single stepped the process to the system call), we need to clear
471 	 * the trap flag from the new frame. Otherwise, the new thread will
472 	 * receive a (likely unexpected) SIGTRAP when it executes the first
473 	 * instruction after returning to userland.
474 	 */
475 	td->td_frame->tf_eflags &= ~PSL_T;
476 
477 	/*
478 	 * Set registers for trampoline to user mode.  Leave space for the
479 	 * return address on stack.  These are the kernel mode register values.
480 	 */
481 	pcb2->pcb_edi = 0;
482 	pcb2->pcb_esi = (int)fork_return;		    /* trampoline arg */
483 	pcb2->pcb_ebp = 0;
484 	pcb2->pcb_esp = (int)td->td_frame - sizeof(void *); /* trampoline arg */
485 	pcb2->pcb_ebx = (int)td;			    /* trampoline arg */
486 	pcb2->pcb_eip = (int)fork_trampoline;
487 	pcb2->pcb_psl &= ~(PSL_I);	/* interrupts must be disabled */
488 	pcb2->pcb_gs = rgs();
489 	/*
490 	 * If we didn't copy the pcb, we'd need to do the following registers:
491 	 * pcb2->pcb_cr3:	cloned above.
492 	 * pcb2->pcb_dr*:	cloned above.
493 	 * pcb2->pcb_savefpu:	cloned above.
494 	 * pcb2->pcb_flags:	cloned above.
495 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
496 	 * pcb2->pcb_gs:	cloned above.
497 	 * pcb2->pcb_ext:	cleared below.
498 	 */
499 	pcb2->pcb_ext = NULL;
500 
501 	/* Setup to release spin count in fork_exit(). */
502 	td->td_md.md_spinlock_count = 1;
503 	td->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
504 }
505 
506 /*
507  * Set that machine state for performing an upcall that has to
508  * be done in thread_userret() so that those upcalls generated
509  * in thread_userret() itself can be done as well.
510  */
511 void
512 cpu_set_upcall_kse(struct thread *td, void (*entry)(void *), void *arg,
513 	stack_t *stack)
514 {
515 
516 	/*
517 	 * Do any extra cleaning that needs to be done.
518 	 * The thread may have optional components
519 	 * that are not present in a fresh thread.
520 	 * This may be a recycled thread so make it look
521 	 * as though it's newly allocated.
522 	 */
523 	cpu_thread_clean(td);
524 
525 	/*
526 	 * Set the trap frame to point at the beginning of the uts
527 	 * function.
528 	 */
529 	td->td_frame->tf_ebp = 0;
530 	td->td_frame->tf_esp =
531 	    (((int)stack->ss_sp + stack->ss_size - 4) & ~0x0f) - 4;
532 	td->td_frame->tf_eip = (int)entry;
533 
534 	/*
535 	 * Pass the address of the mailbox for this kse to the uts
536 	 * function as a parameter on the stack.
537 	 */
538 	suword((void *)(td->td_frame->tf_esp + sizeof(void *)),
539 	    (int)arg);
540 }
541 
542 int
543 cpu_set_user_tls(struct thread *td, void *tls_base)
544 {
545 	struct segment_descriptor sd;
546 	uint32_t base;
547 
548 	/*
549 	 * Construct a descriptor and store it in the pcb for
550 	 * the next context switch.  Also store it in the gdt
551 	 * so that the load of tf_fs into %fs will activate it
552 	 * at return to userland.
553 	 */
554 	base = (uint32_t)tls_base;
555 	sd.sd_lobase = base & 0xffffff;
556 	sd.sd_hibase = (base >> 24) & 0xff;
557 	sd.sd_lolimit = 0xffff;	/* 4GB limit, wraps around */
558 	sd.sd_hilimit = 0xf;
559 	sd.sd_type  = SDT_MEMRWA;
560 	sd.sd_dpl   = SEL_UPL;
561 	sd.sd_p     = 1;
562 	sd.sd_xx    = 0;
563 	sd.sd_def32 = 1;
564 	sd.sd_gran  = 1;
565 	critical_enter();
566 	/* set %gs */
567 	td->td_pcb->pcb_gsd = sd;
568 	if (td == curthread) {
569 		PCPU_GET(fsgs_gdt)[1] = sd;
570 		load_gs(GSEL(GUGS_SEL, SEL_UPL));
571 	}
572 	critical_exit();
573 	return (0);
574 }
575 
576 /*
577  * Convert kernel VA to physical address
578  */
579 vm_paddr_t
580 kvtop(void *addr)
581 {
582 	vm_paddr_t pa;
583 
584 	pa = pmap_kextract((vm_offset_t)addr);
585 	if (pa == 0)
586 		panic("kvtop: zero page frame");
587 	return (pa);
588 }
589 
590 #ifdef SMP
591 static void
592 cpu_reset_proxy()
593 {
594 	cpuset_t tcrp;
595 
596 	cpu_reset_proxy_active = 1;
597 	while (cpu_reset_proxy_active == 1)
598 		;	/* Wait for other cpu to see that we've started */
599 	CPU_SETOF(cpu_reset_proxyid, &tcrp);
600 	stop_cpus(tcrp);
601 	printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
602 	DELAY(1000000);
603 	cpu_reset_real();
604 }
605 #endif
606 
607 void
608 cpu_reset()
609 {
610 #ifdef XBOX
611 	if (arch_i386_is_xbox) {
612 		/* Kick the PIC16L, it can reboot the box */
613 		pic16l_reboot();
614 		for (;;);
615 	}
616 #endif
617 
618 #ifdef SMP
619 	cpuset_t map;
620 	u_int cnt;
621 
622 	if (smp_active) {
623 		map = all_cpus;
624 		CPU_CLR(PCPU_GET(cpuid), &map);
625 		CPU_NAND(&map, &stopped_cpus);
626 		if (!CPU_EMPTY(&map)) {
627 			printf("cpu_reset: Stopping other CPUs\n");
628 			stop_cpus(map);
629 		}
630 
631 		if (PCPU_GET(cpuid) != 0) {
632 			cpu_reset_proxyid = PCPU_GET(cpuid);
633 			cpustop_restartfunc = cpu_reset_proxy;
634 			cpu_reset_proxy_active = 0;
635 			printf("cpu_reset: Restarting BSP\n");
636 
637 			/* Restart CPU #0. */
638 			/* XXX: restart_cpus(1 << 0); */
639 			CPU_SETOF(0, &started_cpus);
640 			wmb();
641 
642 			cnt = 0;
643 			while (cpu_reset_proxy_active == 0 && cnt < 10000000)
644 				cnt++;	/* Wait for BSP to announce restart */
645 			if (cpu_reset_proxy_active == 0)
646 				printf("cpu_reset: Failed to restart BSP\n");
647 			enable_intr();
648 			cpu_reset_proxy_active = 2;
649 
650 			while (1);
651 			/* NOTREACHED */
652 		}
653 
654 		DELAY(1000000);
655 	}
656 #endif
657 	cpu_reset_real();
658 	/* NOTREACHED */
659 }
660 
661 static void
662 cpu_reset_real()
663 {
664 	struct region_descriptor null_idt;
665 #ifndef PC98
666 	int b;
667 #endif
668 
669 	disable_intr();
670 #ifdef XEN
671 	if (smp_processor_id() == 0)
672 		HYPERVISOR_shutdown(SHUTDOWN_reboot);
673 	else
674 		HYPERVISOR_shutdown(SHUTDOWN_poweroff);
675 #endif
676 #ifdef CPU_ELAN
677 	if (elan_mmcr != NULL)
678 		elan_mmcr->RESCFG = 1;
679 #endif
680 
681 	if (cpu == CPU_GEODE1100) {
682 		/* Attempt Geode's own reset */
683 		outl(0xcf8, 0x80009044ul);
684 		outl(0xcfc, 0xf);
685 	}
686 
687 #ifdef PC98
688 	/*
689 	 * Attempt to do a CPU reset via CPU reset port.
690 	 */
691 	if ((inb(0x35) & 0xa0) != 0xa0) {
692 		outb(0x37, 0x0f);		/* SHUT0 = 0. */
693 		outb(0x37, 0x0b);		/* SHUT1 = 0. */
694 	}
695 	outb(0xf0, 0x00);		/* Reset. */
696 #else
697 #if !defined(BROKEN_KEYBOARD_RESET)
698 	/*
699 	 * Attempt to do a CPU reset via the keyboard controller,
700 	 * do not turn off GateA20, as any machine that fails
701 	 * to do the reset here would then end up in no man's land.
702 	 */
703 	outb(IO_KBD + 4, 0xFE);
704 	DELAY(500000);	/* wait 0.5 sec to see if that did it */
705 #endif
706 
707 	/*
708 	 * Attempt to force a reset via the Reset Control register at
709 	 * I/O port 0xcf9.  Bit 2 forces a system reset when it
710 	 * transitions from 0 to 1.  Bit 1 selects the type of reset
711 	 * to attempt: 0 selects a "soft" reset, and 1 selects a
712 	 * "hard" reset.  We try a "hard" reset.  The first write sets
713 	 * bit 1 to select a "hard" reset and clears bit 2.  The
714 	 * second write forces a 0 -> 1 transition in bit 2 to trigger
715 	 * a reset.
716 	 */
717 	outb(0xcf9, 0x2);
718 	outb(0xcf9, 0x6);
719 	DELAY(500000);  /* wait 0.5 sec to see if that did it */
720 
721 	/*
722 	 * Attempt to force a reset via the Fast A20 and Init register
723 	 * at I/O port 0x92.  Bit 1 serves as an alternate A20 gate.
724 	 * Bit 0 asserts INIT# when set to 1.  We are careful to only
725 	 * preserve bit 1 while setting bit 0.  We also must clear bit
726 	 * 0 before setting it if it isn't already clear.
727 	 */
728 	b = inb(0x92);
729 	if (b != 0xff) {
730 		if ((b & 0x1) != 0)
731 			outb(0x92, b & 0xfe);
732 		outb(0x92, b | 0x1);
733 		DELAY(500000);  /* wait 0.5 sec to see if that did it */
734 	}
735 #endif /* PC98 */
736 
737 	printf("No known reset method worked, attempting CPU shutdown\n");
738 	DELAY(1000000); /* wait 1 sec for printf to complete */
739 
740 	/* Wipe the IDT. */
741 	null_idt.rd_limit = 0;
742 	null_idt.rd_base = 0;
743 	lidt(&null_idt);
744 
745 	/* "good night, sweet prince .... <THUNK!>" */
746 	breakpoint();
747 
748 	/* NOTREACHED */
749 	while(1);
750 }
751 
752 /*
753  * Allocate a pool of sf_bufs (sendfile(2) or "super-fast" if you prefer. :-))
754  */
755 static void
756 sf_buf_init(void *arg)
757 {
758 	struct sf_buf *sf_bufs;
759 	vm_offset_t sf_base;
760 	int i;
761 
762 	nsfbufs = NSFBUFS;
763 	TUNABLE_INT_FETCH("kern.ipc.nsfbufs", &nsfbufs);
764 
765 	sf_buf_active = hashinit(nsfbufs, M_TEMP, &sf_buf_hashmask);
766 	TAILQ_INIT(&sf_buf_freelist);
767 	sf_base = kva_alloc(nsfbufs * PAGE_SIZE);
768 	sf_bufs = malloc(nsfbufs * sizeof(struct sf_buf), M_TEMP,
769 	    M_NOWAIT | M_ZERO);
770 	for (i = 0; i < nsfbufs; i++) {
771 		sf_bufs[i].kva = sf_base + i * PAGE_SIZE;
772 		TAILQ_INSERT_TAIL(&sf_buf_freelist, &sf_bufs[i], free_entry);
773 	}
774 	sf_buf_alloc_want = 0;
775 	mtx_init(&sf_buf_lock, "sf_buf", NULL, MTX_DEF);
776 }
777 
778 /*
779  * Invalidate the cache lines that may belong to the page, if
780  * (possibly old) mapping of the page by sf buffer exists.  Returns
781  * TRUE when mapping was found and cache invalidated.
782  */
783 boolean_t
784 sf_buf_invalidate_cache(vm_page_t m)
785 {
786 	struct sf_head *hash_list;
787 	struct sf_buf *sf;
788 	boolean_t ret;
789 
790 	hash_list = &sf_buf_active[SF_BUF_HASH(m)];
791 	ret = FALSE;
792 	mtx_lock(&sf_buf_lock);
793 	LIST_FOREACH(sf, hash_list, list_entry) {
794 		if (sf->m == m) {
795 			/*
796 			 * Use pmap_qenter to update the pte for
797 			 * existing mapping, in particular, the PAT
798 			 * settings are recalculated.
799 			 */
800 			pmap_qenter(sf->kva, &m, 1);
801 			pmap_invalidate_cache_range(sf->kva, sf->kva +
802 			    PAGE_SIZE);
803 			ret = TRUE;
804 			break;
805 		}
806 	}
807 	mtx_unlock(&sf_buf_lock);
808 	return (ret);
809 }
810 
811 /*
812  * Get an sf_buf from the freelist.  May block if none are available.
813  */
814 struct sf_buf *
815 sf_buf_alloc(struct vm_page *m, int flags)
816 {
817 	pt_entry_t opte, *ptep;
818 	struct sf_head *hash_list;
819 	struct sf_buf *sf;
820 #ifdef SMP
821 	cpuset_t other_cpus;
822 	u_int cpuid;
823 #endif
824 	int error;
825 
826 	KASSERT(curthread->td_pinned > 0 || (flags & SFB_CPUPRIVATE) == 0,
827 	    ("sf_buf_alloc(SFB_CPUPRIVATE): curthread not pinned"));
828 	hash_list = &sf_buf_active[SF_BUF_HASH(m)];
829 	mtx_lock(&sf_buf_lock);
830 	LIST_FOREACH(sf, hash_list, list_entry) {
831 		if (sf->m == m) {
832 			sf->ref_count++;
833 			if (sf->ref_count == 1) {
834 				TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
835 				nsfbufsused++;
836 				nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
837 			}
838 #ifdef SMP
839 			goto shootdown;
840 #else
841 			goto done;
842 #endif
843 		}
844 	}
845 	while ((sf = TAILQ_FIRST(&sf_buf_freelist)) == NULL) {
846 		if (flags & SFB_NOWAIT)
847 			goto done;
848 		sf_buf_alloc_want++;
849 		SFSTAT_INC(sf_allocwait);
850 		error = msleep(&sf_buf_freelist, &sf_buf_lock,
851 		    (flags & SFB_CATCH) ? PCATCH | PVM : PVM, "sfbufa", 0);
852 		sf_buf_alloc_want--;
853 
854 		/*
855 		 * If we got a signal, don't risk going back to sleep.
856 		 */
857 		if (error)
858 			goto done;
859 	}
860 	TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
861 	if (sf->m != NULL)
862 		LIST_REMOVE(sf, list_entry);
863 	LIST_INSERT_HEAD(hash_list, sf, list_entry);
864 	sf->ref_count = 1;
865 	sf->m = m;
866 	nsfbufsused++;
867 	nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
868 
869 	/*
870 	 * Update the sf_buf's virtual-to-physical mapping, flushing the
871 	 * virtual address from the TLB.  Since the reference count for
872 	 * the sf_buf's old mapping was zero, that mapping is not
873 	 * currently in use.  Consequently, there is no need to exchange
874 	 * the old and new PTEs atomically, even under PAE.
875 	 */
876 	ptep = vtopte(sf->kva);
877 	opte = *ptep;
878 #ifdef XEN
879        PT_SET_MA(sf->kva, xpmap_ptom(VM_PAGE_TO_PHYS(m)) | pgeflag
880 	   | PG_RW | PG_V | pmap_cache_bits(m->md.pat_mode, 0));
881 #else
882 	*ptep = VM_PAGE_TO_PHYS(m) | pgeflag | PG_RW | PG_V |
883 	    pmap_cache_bits(m->md.pat_mode, 0);
884 #endif
885 
886 	/*
887 	 * Avoid unnecessary TLB invalidations: If the sf_buf's old
888 	 * virtual-to-physical mapping was not used, then any processor
889 	 * that has invalidated the sf_buf's virtual address from its TLB
890 	 * since the last used mapping need not invalidate again.
891 	 */
892 #ifdef SMP
893 	if ((opte & (PG_V | PG_A)) ==  (PG_V | PG_A))
894 		CPU_ZERO(&sf->cpumask);
895 shootdown:
896 	sched_pin();
897 	cpuid = PCPU_GET(cpuid);
898 	if (!CPU_ISSET(cpuid, &sf->cpumask)) {
899 		CPU_SET(cpuid, &sf->cpumask);
900 		invlpg(sf->kva);
901 	}
902 	if ((flags & SFB_CPUPRIVATE) == 0) {
903 		other_cpus = all_cpus;
904 		CPU_CLR(cpuid, &other_cpus);
905 		CPU_NAND(&other_cpus, &sf->cpumask);
906 		if (!CPU_EMPTY(&other_cpus)) {
907 			CPU_OR(&sf->cpumask, &other_cpus);
908 			smp_masked_invlpg(other_cpus, sf->kva);
909 		}
910 	}
911 	sched_unpin();
912 #else
913 	if ((opte & (PG_V | PG_A)) ==  (PG_V | PG_A))
914 		pmap_invalidate_page(kernel_pmap, sf->kva);
915 #endif
916 done:
917 	mtx_unlock(&sf_buf_lock);
918 	return (sf);
919 }
920 
921 /*
922  * Remove a reference from the given sf_buf, adding it to the free
923  * list when its reference count reaches zero.  A freed sf_buf still,
924  * however, retains its virtual-to-physical mapping until it is
925  * recycled or reactivated by sf_buf_alloc(9).
926  */
927 void
928 sf_buf_free(struct sf_buf *sf)
929 {
930 
931 	mtx_lock(&sf_buf_lock);
932 	sf->ref_count--;
933 	if (sf->ref_count == 0) {
934 		TAILQ_INSERT_TAIL(&sf_buf_freelist, sf, free_entry);
935 		nsfbufsused--;
936 #ifdef XEN
937 /*
938  * Xen doesn't like having dangling R/W mappings
939  */
940 		pmap_qremove(sf->kva, 1);
941 		sf->m = NULL;
942 		LIST_REMOVE(sf, list_entry);
943 #endif
944 		if (sf_buf_alloc_want > 0)
945 			wakeup(&sf_buf_freelist);
946 	}
947 	mtx_unlock(&sf_buf_lock);
948 }
949 
950 /*
951  * Software interrupt handler for queued VM system processing.
952  */
953 void
954 swi_vm(void *dummy)
955 {
956 	if (busdma_swi_pending != 0)
957 		busdma_swi();
958 }
959 
960 /*
961  * Tell whether this address is in some physical memory region.
962  * Currently used by the kernel coredump code in order to avoid
963  * dumping the ``ISA memory hole'' which could cause indefinite hangs,
964  * or other unpredictable behaviour.
965  */
966 
967 int
968 is_physical_memory(vm_paddr_t addr)
969 {
970 
971 #ifdef DEV_ISA
972 	/* The ISA ``memory hole''. */
973 	if (addr >= 0xa0000 && addr < 0x100000)
974 		return 0;
975 #endif
976 
977 	/*
978 	 * stuff other tests for known memory-mapped devices (PCI?)
979 	 * here
980 	 */
981 
982 	return 1;
983 }
984