xref: /freebsd/sys/i386/i386/vm_machdep.c (revision aa64588d28258aef88cc33b8043112e8856948d0)
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 <x86/isa/isa.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 void	cpu_reset_real(void);
110 #ifdef SMP
111 static void	cpu_reset_proxy(void);
112 static u_int	cpu_reset_proxyid;
113 static volatile u_int	cpu_reset_proxy_active;
114 #endif
115 static void	sf_buf_init(void *arg);
116 SYSINIT(sock_sf, SI_SUB_MBUF, SI_ORDER_ANY, sf_buf_init, NULL);
117 
118 LIST_HEAD(sf_head, sf_buf);
119 
120 /*
121  * A hash table of active sendfile(2) buffers
122  */
123 static struct sf_head *sf_buf_active;
124 static u_long sf_buf_hashmask;
125 
126 #define	SF_BUF_HASH(m)	(((m) - vm_page_array) & sf_buf_hashmask)
127 
128 static TAILQ_HEAD(, sf_buf) sf_buf_freelist;
129 static u_int	sf_buf_alloc_want;
130 
131 /*
132  * A lock used to synchronize access to the hash table and free list
133  */
134 static struct mtx sf_buf_lock;
135 
136 extern int	_ucodesel, _udatasel;
137 
138 /*
139  * Finish a fork operation, with process p2 nearly set up.
140  * Copy and update the pcb, set up the stack so that the child
141  * ready to run and return to user mode.
142  */
143 void
144 cpu_fork(td1, p2, td2, flags)
145 	register struct thread *td1;
146 	register struct proc *p2;
147 	struct thread *td2;
148 	int flags;
149 {
150 	register struct proc *p1;
151 	struct pcb *pcb2;
152 	struct mdproc *mdp2;
153 #ifdef DEV_NPX
154 	register_t savecrit;
155 #endif
156 
157 	p1 = td1->td_proc;
158 	if ((flags & RFPROC) == 0) {
159 		if ((flags & RFMEM) == 0) {
160 			/* unshare user LDT */
161 			struct mdproc *mdp1 = &p1->p_md;
162 			struct proc_ldt *pldt, *pldt1;
163 
164 			mtx_lock_spin(&dt_lock);
165 			if ((pldt1 = mdp1->md_ldt) != NULL &&
166 			    pldt1->ldt_refcnt > 1) {
167 				pldt = user_ldt_alloc(mdp1, pldt1->ldt_len);
168 				if (pldt == NULL)
169 					panic("could not copy LDT");
170 				mdp1->md_ldt = pldt;
171 				set_user_ldt(mdp1);
172 				user_ldt_deref(pldt1);
173 			} else
174 				mtx_unlock_spin(&dt_lock);
175 		}
176 		return;
177 	}
178 
179 	/* Ensure that td1's pcb is up to date. */
180 	if (td1 == curthread)
181 		td1->td_pcb->pcb_gs = rgs();
182 #ifdef DEV_NPX
183 	savecrit = intr_disable();
184 	if (PCPU_GET(fpcurthread) == td1)
185 		npxsave(td1->td_pcb->pcb_save);
186 	intr_restore(savecrit);
187 #endif
188 
189 	/* Point the pcb to the top of the stack */
190 	pcb2 = (struct pcb *)(td2->td_kstack +
191 	    td2->td_kstack_pages * PAGE_SIZE) - 1;
192 	td2->td_pcb = pcb2;
193 
194 	/* Copy td1's pcb */
195 	bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
196 
197 	/* Properly initialize pcb_save */
198 	pcb2->pcb_save = &pcb2->pcb_user_save;
199 
200 	/* Point mdproc and then copy over td1's contents */
201 	mdp2 = &p2->p_md;
202 	bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
203 
204 	/*
205 	 * Create a new fresh stack for the new process.
206 	 * Copy the trap frame for the return to user mode as if from a
207 	 * syscall.  This copies most of the user mode register values.
208 	 * The -16 is so we can expand the trapframe if we go to vm86.
209 	 */
210 	td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb - 16) - 1;
211 	bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
212 
213 	td2->td_frame->tf_eax = 0;		/* Child returns zero */
214 	td2->td_frame->tf_eflags &= ~PSL_C;	/* success */
215 	td2->td_frame->tf_edx = 1;
216 
217 	/*
218 	 * If the parent process has the trap bit set (i.e. a debugger had
219 	 * single stepped the process to the system call), we need to clear
220 	 * the trap flag from the new frame unless the debugger had set PF_FORK
221 	 * on the parent.  Otherwise, the child will receive a (likely
222 	 * unexpected) SIGTRAP when it executes the first instruction after
223 	 * returning  to userland.
224 	 */
225 	if ((p1->p_pfsflags & PF_FORK) == 0)
226 		td2->td_frame->tf_eflags &= ~PSL_T;
227 
228 	/*
229 	 * Set registers for trampoline to user mode.  Leave space for the
230 	 * return address on stack.  These are the kernel mode register values.
231 	 */
232 #ifdef PAE
233 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdpt);
234 #else
235 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdir);
236 #endif
237 	pcb2->pcb_edi = 0;
238 	pcb2->pcb_esi = (int)fork_return;	/* fork_trampoline argument */
239 	pcb2->pcb_ebp = 0;
240 	pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *);
241 	pcb2->pcb_ebx = (int)td2;		/* fork_trampoline argument */
242 	pcb2->pcb_eip = (int)fork_trampoline;
243 	pcb2->pcb_psl = PSL_KERNEL;		/* ints disabled */
244 	/*-
245 	 * pcb2->pcb_dr*:	cloned above.
246 	 * pcb2->pcb_savefpu:	cloned above.
247 	 * pcb2->pcb_flags:	cloned above.
248 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
249 	 * pcb2->pcb_gs:	cloned above.
250 	 * pcb2->pcb_ext:	cleared below.
251 	 */
252 
253 	/*
254 	 * XXX don't copy the i/o pages.  this should probably be fixed.
255 	 */
256 	pcb2->pcb_ext = 0;
257 
258 	/* Copy the LDT, if necessary. */
259 	mtx_lock_spin(&dt_lock);
260 	if (mdp2->md_ldt != NULL) {
261 		if (flags & RFMEM) {
262 			mdp2->md_ldt->ldt_refcnt++;
263 		} else {
264 			mdp2->md_ldt = user_ldt_alloc(mdp2,
265 			    mdp2->md_ldt->ldt_len);
266 			if (mdp2->md_ldt == NULL)
267 				panic("could not copy LDT");
268 		}
269 	}
270 	mtx_unlock_spin(&dt_lock);
271 
272 	/* Setup to release spin count in fork_exit(). */
273 	td2->td_md.md_spinlock_count = 1;
274 	/*
275 	 * XXX XEN need to check on PSL_USER is handled
276 	 */
277 	td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
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 	td->td_pcb->pcb_save = &td->td_pcb->pcb_user_save;
379 }
380 
381 void
382 cpu_thread_free(struct thread *td)
383 {
384 
385 	cpu_thread_clean(td);
386 }
387 
388 void
389 cpu_set_syscall_retval(struct thread *td, int error)
390 {
391 
392 	switch (error) {
393 	case 0:
394 		td->td_frame->tf_eax = td->td_retval[0];
395 		td->td_frame->tf_edx = td->td_retval[1];
396 		td->td_frame->tf_eflags &= ~PSL_C;
397 		break;
398 
399 	case ERESTART:
400 		/*
401 		 * Reconstruct pc, assuming lcall $X,y is 7 bytes, int
402 		 * 0x80 is 2 bytes. We saved this in tf_err.
403 		 */
404 		td->td_frame->tf_eip -= td->td_frame->tf_err;
405 		break;
406 
407 	case EJUSTRETURN:
408 		break;
409 
410 	default:
411 		if (td->td_proc->p_sysent->sv_errsize) {
412 			if (error >= td->td_proc->p_sysent->sv_errsize)
413 				error = -1;	/* XXX */
414 			else
415 				error = td->td_proc->p_sysent->sv_errtbl[error];
416 		}
417 		td->td_frame->tf_eax = error;
418 		td->td_frame->tf_eflags |= PSL_C;
419 		break;
420 	}
421 }
422 
423 /*
424  * Initialize machine state (pcb and trap frame) for a new thread about to
425  * upcall. Put enough state in the new thread's PCB to get it to go back
426  * userret(), where we can intercept it again to set the return (upcall)
427  * Address and stack, along with those from upcals that are from other sources
428  * such as those generated in thread_userret() itself.
429  */
430 void
431 cpu_set_upcall(struct thread *td, struct thread *td0)
432 {
433 	struct pcb *pcb2;
434 
435 	/* Point the pcb to the top of the stack. */
436 	pcb2 = td->td_pcb;
437 
438 	/*
439 	 * Copy the upcall pcb.  This loads kernel regs.
440 	 * Those not loaded individually below get their default
441 	 * values here.
442 	 */
443 	bcopy(td0->td_pcb, pcb2, sizeof(*pcb2));
444 	pcb2->pcb_flags &= ~(PCB_NPXTRAP|PCB_NPXINITDONE|PCB_NPXUSERINITDONE);
445 	pcb2->pcb_save = &pcb2->pcb_user_save;
446 
447 	/*
448 	 * Create a new fresh stack for the new thread.
449 	 */
450 	bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe));
451 
452 	/* If the current thread has the trap bit set (i.e. a debugger had
453 	 * single stepped the process to the system call), we need to clear
454 	 * the trap flag from the new frame. Otherwise, the new thread will
455 	 * receive a (likely unexpected) SIGTRAP when it executes the first
456 	 * instruction after returning to userland.
457 	 */
458 	td->td_frame->tf_eflags &= ~PSL_T;
459 
460 	/*
461 	 * Set registers for trampoline to user mode.  Leave space for the
462 	 * return address on stack.  These are the kernel mode register values.
463 	 */
464 	pcb2->pcb_edi = 0;
465 	pcb2->pcb_esi = (int)fork_return;		    /* trampoline arg */
466 	pcb2->pcb_ebp = 0;
467 	pcb2->pcb_esp = (int)td->td_frame - sizeof(void *); /* trampoline arg */
468 	pcb2->pcb_ebx = (int)td;			    /* trampoline arg */
469 	pcb2->pcb_eip = (int)fork_trampoline;
470 	pcb2->pcb_psl &= ~(PSL_I);	/* interrupts must be disabled */
471 	pcb2->pcb_gs = rgs();
472 	/*
473 	 * If we didn't copy the pcb, we'd need to do the following registers:
474 	 * pcb2->pcb_cr3:	cloned above.
475 	 * pcb2->pcb_dr*:	cloned above.
476 	 * pcb2->pcb_savefpu:	cloned above.
477 	 * pcb2->pcb_flags:	cloned above.
478 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
479 	 * pcb2->pcb_gs:	cloned above.
480 	 * pcb2->pcb_ext:	cleared below.
481 	 */
482 	pcb2->pcb_ext = NULL;
483 
484 	/* Setup to release spin count in fork_exit(). */
485 	td->td_md.md_spinlock_count = 1;
486 	td->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
487 }
488 
489 /*
490  * Set that machine state for performing an upcall that has to
491  * be done in thread_userret() so that those upcalls generated
492  * in thread_userret() itself can be done as well.
493  */
494 void
495 cpu_set_upcall_kse(struct thread *td, void (*entry)(void *), void *arg,
496 	stack_t *stack)
497 {
498 
499 	/*
500 	 * Do any extra cleaning that needs to be done.
501 	 * The thread may have optional components
502 	 * that are not present in a fresh thread.
503 	 * This may be a recycled thread so make it look
504 	 * as though it's newly allocated.
505 	 */
506 	cpu_thread_clean(td);
507 
508 	/*
509 	 * Set the trap frame to point at the beginning of the uts
510 	 * function.
511 	 */
512 	td->td_frame->tf_ebp = 0;
513 	td->td_frame->tf_esp =
514 	    (((int)stack->ss_sp + stack->ss_size - 4) & ~0x0f) - 4;
515 	td->td_frame->tf_eip = (int)entry;
516 
517 	/*
518 	 * Pass the address of the mailbox for this kse to the uts
519 	 * function as a parameter on the stack.
520 	 */
521 	suword((void *)(td->td_frame->tf_esp + sizeof(void *)),
522 	    (int)arg);
523 }
524 
525 int
526 cpu_set_user_tls(struct thread *td, void *tls_base)
527 {
528 	struct segment_descriptor sd;
529 	uint32_t base;
530 
531 	/*
532 	 * Construct a descriptor and store it in the pcb for
533 	 * the next context switch.  Also store it in the gdt
534 	 * so that the load of tf_fs into %fs will activate it
535 	 * at return to userland.
536 	 */
537 	base = (uint32_t)tls_base;
538 	sd.sd_lobase = base & 0xffffff;
539 	sd.sd_hibase = (base >> 24) & 0xff;
540 	sd.sd_lolimit = 0xffff;	/* 4GB limit, wraps around */
541 	sd.sd_hilimit = 0xf;
542 	sd.sd_type  = SDT_MEMRWA;
543 	sd.sd_dpl   = SEL_UPL;
544 	sd.sd_p     = 1;
545 	sd.sd_xx    = 0;
546 	sd.sd_def32 = 1;
547 	sd.sd_gran  = 1;
548 	critical_enter();
549 	/* set %gs */
550 	td->td_pcb->pcb_gsd = sd;
551 	if (td == curthread) {
552 		PCPU_GET(fsgs_gdt)[1] = sd;
553 		load_gs(GSEL(GUGS_SEL, SEL_UPL));
554 	}
555 	critical_exit();
556 	return (0);
557 }
558 
559 /*
560  * Convert kernel VA to physical address
561  */
562 vm_paddr_t
563 kvtop(void *addr)
564 {
565 	vm_paddr_t pa;
566 
567 	pa = pmap_kextract((vm_offset_t)addr);
568 	if (pa == 0)
569 		panic("kvtop: zero page frame");
570 	return (pa);
571 }
572 
573 #ifdef SMP
574 static void
575 cpu_reset_proxy()
576 {
577 
578 	cpu_reset_proxy_active = 1;
579 	while (cpu_reset_proxy_active == 1)
580 		;	/* Wait for other cpu to see that we've started */
581 	stop_cpus((1<<cpu_reset_proxyid));
582 	printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
583 	DELAY(1000000);
584 	cpu_reset_real();
585 }
586 #endif
587 
588 void
589 cpu_reset()
590 {
591 #ifdef XBOX
592 	if (arch_i386_is_xbox) {
593 		/* Kick the PIC16L, it can reboot the box */
594 		pic16l_reboot();
595 		for (;;);
596 	}
597 #endif
598 
599 #ifdef SMP
600 	u_int cnt, map;
601 
602 	if (smp_active) {
603 		map = PCPU_GET(other_cpus) & ~stopped_cpus;
604 		if (map != 0) {
605 			printf("cpu_reset: Stopping other CPUs\n");
606 			stop_cpus(map);
607 		}
608 
609 		if (PCPU_GET(cpuid) != 0) {
610 			cpu_reset_proxyid = PCPU_GET(cpuid);
611 			cpustop_restartfunc = cpu_reset_proxy;
612 			cpu_reset_proxy_active = 0;
613 			printf("cpu_reset: Restarting BSP\n");
614 
615 			/* Restart CPU #0. */
616 			/* XXX: restart_cpus(1 << 0); */
617 			atomic_store_rel_int(&started_cpus, (1 << 0));
618 
619 			cnt = 0;
620 			while (cpu_reset_proxy_active == 0 && cnt < 10000000)
621 				cnt++;	/* Wait for BSP to announce restart */
622 			if (cpu_reset_proxy_active == 0)
623 				printf("cpu_reset: Failed to restart BSP\n");
624 			enable_intr();
625 			cpu_reset_proxy_active = 2;
626 
627 			while (1);
628 			/* NOTREACHED */
629 		}
630 
631 		DELAY(1000000);
632 	}
633 #endif
634 	cpu_reset_real();
635 	/* NOTREACHED */
636 }
637 
638 static void
639 cpu_reset_real()
640 {
641 	struct region_descriptor null_idt;
642 #ifndef PC98
643 	int b;
644 #endif
645 
646 	disable_intr();
647 #ifdef XEN
648 	if (smp_processor_id() == 0)
649 		HYPERVISOR_shutdown(SHUTDOWN_reboot);
650 	else
651 		HYPERVISOR_shutdown(SHUTDOWN_poweroff);
652 #endif
653 #ifdef CPU_ELAN
654 	if (elan_mmcr != NULL)
655 		elan_mmcr->RESCFG = 1;
656 #endif
657 
658 	if (cpu == CPU_GEODE1100) {
659 		/* Attempt Geode's own reset */
660 		outl(0xcf8, 0x80009044ul);
661 		outl(0xcfc, 0xf);
662 	}
663 
664 #ifdef PC98
665 	/*
666 	 * Attempt to do a CPU reset via CPU reset port.
667 	 */
668 	if ((inb(0x35) & 0xa0) != 0xa0) {
669 		outb(0x37, 0x0f);		/* SHUT0 = 0. */
670 		outb(0x37, 0x0b);		/* SHUT1 = 0. */
671 	}
672 	outb(0xf0, 0x00);		/* Reset. */
673 #else
674 #if !defined(BROKEN_KEYBOARD_RESET)
675 	/*
676 	 * Attempt to do a CPU reset via the keyboard controller,
677 	 * do not turn off GateA20, as any machine that fails
678 	 * to do the reset here would then end up in no man's land.
679 	 */
680 	outb(IO_KBD + 4, 0xFE);
681 	DELAY(500000);	/* wait 0.5 sec to see if that did it */
682 #endif
683 
684 	/*
685 	 * Attempt to force a reset via the Reset Control register at
686 	 * I/O port 0xcf9.  Bit 2 forces a system reset when it
687 	 * transitions from 0 to 1.  Bit 1 selects the type of reset
688 	 * to attempt: 0 selects a "soft" reset, and 1 selects a
689 	 * "hard" reset.  We try a "hard" reset.  The first write sets
690 	 * bit 1 to select a "hard" reset and clears bit 2.  The
691 	 * second write forces a 0 -> 1 transition in bit 2 to trigger
692 	 * a reset.
693 	 */
694 	outb(0xcf9, 0x2);
695 	outb(0xcf9, 0x6);
696 	DELAY(500000);  /* wait 0.5 sec to see if that did it */
697 
698 	/*
699 	 * Attempt to force a reset via the Fast A20 and Init register
700 	 * at I/O port 0x92.  Bit 1 serves as an alternate A20 gate.
701 	 * Bit 0 asserts INIT# when set to 1.  We are careful to only
702 	 * preserve bit 1 while setting bit 0.  We also must clear bit
703 	 * 0 before setting it if it isn't already clear.
704 	 */
705 	b = inb(0x92);
706 	if (b != 0xff) {
707 		if ((b & 0x1) != 0)
708 			outb(0x92, b & 0xfe);
709 		outb(0x92, b | 0x1);
710 		DELAY(500000);  /* wait 0.5 sec to see if that did it */
711 	}
712 #endif /* PC98 */
713 
714 	printf("No known reset method worked, attempting CPU shutdown\n");
715 	DELAY(1000000); /* wait 1 sec for printf to complete */
716 
717 	/* Wipe the IDT. */
718 	null_idt.rd_limit = 0;
719 	null_idt.rd_base = 0;
720 	lidt(&null_idt);
721 
722 	/* "good night, sweet prince .... <THUNK!>" */
723 	breakpoint();
724 
725 	/* NOTREACHED */
726 	while(1);
727 }
728 
729 /*
730  * Allocate a pool of sf_bufs (sendfile(2) or "super-fast" if you prefer. :-))
731  */
732 static void
733 sf_buf_init(void *arg)
734 {
735 	struct sf_buf *sf_bufs;
736 	vm_offset_t sf_base;
737 	int i;
738 
739 	nsfbufs = NSFBUFS;
740 	TUNABLE_INT_FETCH("kern.ipc.nsfbufs", &nsfbufs);
741 
742 	sf_buf_active = hashinit(nsfbufs, M_TEMP, &sf_buf_hashmask);
743 	TAILQ_INIT(&sf_buf_freelist);
744 	sf_base = kmem_alloc_nofault(kernel_map, nsfbufs * PAGE_SIZE);
745 	sf_bufs = malloc(nsfbufs * sizeof(struct sf_buf), M_TEMP,
746 	    M_NOWAIT | M_ZERO);
747 	for (i = 0; i < nsfbufs; i++) {
748 		sf_bufs[i].kva = sf_base + i * PAGE_SIZE;
749 		TAILQ_INSERT_TAIL(&sf_buf_freelist, &sf_bufs[i], free_entry);
750 	}
751 	sf_buf_alloc_want = 0;
752 	mtx_init(&sf_buf_lock, "sf_buf", NULL, MTX_DEF);
753 }
754 
755 /*
756  * Invalidate the cache lines that may belong to the page, if
757  * (possibly old) mapping of the page by sf buffer exists.  Returns
758  * TRUE when mapping was found and cache invalidated.
759  */
760 boolean_t
761 sf_buf_invalidate_cache(vm_page_t m)
762 {
763 	struct sf_head *hash_list;
764 	struct sf_buf *sf;
765 	boolean_t ret;
766 
767 	hash_list = &sf_buf_active[SF_BUF_HASH(m)];
768 	ret = FALSE;
769 	mtx_lock(&sf_buf_lock);
770 	LIST_FOREACH(sf, hash_list, list_entry) {
771 		if (sf->m == m) {
772 			/*
773 			 * Use pmap_qenter to update the pte for
774 			 * existing mapping, in particular, the PAT
775 			 * settings are recalculated.
776 			 */
777 			pmap_qenter(sf->kva, &m, 1);
778 			pmap_invalidate_cache_range(sf->kva, sf->kva +
779 			    PAGE_SIZE);
780 			ret = TRUE;
781 			break;
782 		}
783 	}
784 	mtx_unlock(&sf_buf_lock);
785 	return (ret);
786 }
787 
788 /*
789  * Get an sf_buf from the freelist.  May block if none are available.
790  */
791 struct sf_buf *
792 sf_buf_alloc(struct vm_page *m, int flags)
793 {
794 	pt_entry_t opte, *ptep;
795 	struct sf_head *hash_list;
796 	struct sf_buf *sf;
797 #ifdef SMP
798 	cpumask_t cpumask, other_cpus;
799 #endif
800 	int error;
801 
802 	KASSERT(curthread->td_pinned > 0 || (flags & SFB_CPUPRIVATE) == 0,
803 	    ("sf_buf_alloc(SFB_CPUPRIVATE): curthread not pinned"));
804 	hash_list = &sf_buf_active[SF_BUF_HASH(m)];
805 	mtx_lock(&sf_buf_lock);
806 	LIST_FOREACH(sf, hash_list, list_entry) {
807 		if (sf->m == m) {
808 			sf->ref_count++;
809 			if (sf->ref_count == 1) {
810 				TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
811 				nsfbufsused++;
812 				nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
813 			}
814 #ifdef SMP
815 			goto shootdown;
816 #else
817 			goto done;
818 #endif
819 		}
820 	}
821 	while ((sf = TAILQ_FIRST(&sf_buf_freelist)) == NULL) {
822 		if (flags & SFB_NOWAIT)
823 			goto done;
824 		sf_buf_alloc_want++;
825 		mbstat.sf_allocwait++;
826 		error = msleep(&sf_buf_freelist, &sf_buf_lock,
827 		    (flags & SFB_CATCH) ? PCATCH | PVM : PVM, "sfbufa", 0);
828 		sf_buf_alloc_want--;
829 
830 		/*
831 		 * If we got a signal, don't risk going back to sleep.
832 		 */
833 		if (error)
834 			goto done;
835 	}
836 	TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
837 	if (sf->m != NULL)
838 		LIST_REMOVE(sf, list_entry);
839 	LIST_INSERT_HEAD(hash_list, sf, list_entry);
840 	sf->ref_count = 1;
841 	sf->m = m;
842 	nsfbufsused++;
843 	nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
844 
845 	/*
846 	 * Update the sf_buf's virtual-to-physical mapping, flushing the
847 	 * virtual address from the TLB.  Since the reference count for
848 	 * the sf_buf's old mapping was zero, that mapping is not
849 	 * currently in use.  Consequently, there is no need to exchange
850 	 * the old and new PTEs atomically, even under PAE.
851 	 */
852 	ptep = vtopte(sf->kva);
853 	opte = *ptep;
854 #ifdef XEN
855        PT_SET_MA(sf->kva, xpmap_ptom(VM_PAGE_TO_PHYS(m)) | pgeflag
856 	   | PG_RW | PG_V | pmap_cache_bits(m->md.pat_mode, 0));
857 #else
858 	*ptep = VM_PAGE_TO_PHYS(m) | pgeflag | PG_RW | PG_V |
859 	    pmap_cache_bits(m->md.pat_mode, 0);
860 #endif
861 
862 	/*
863 	 * Avoid unnecessary TLB invalidations: If the sf_buf's old
864 	 * virtual-to-physical mapping was not used, then any processor
865 	 * that has invalidated the sf_buf's virtual address from its TLB
866 	 * since the last used mapping need not invalidate again.
867 	 */
868 #ifdef SMP
869 	if ((opte & (PG_V | PG_A)) ==  (PG_V | PG_A))
870 		sf->cpumask = 0;
871 shootdown:
872 	sched_pin();
873 	cpumask = PCPU_GET(cpumask);
874 	if ((sf->cpumask & cpumask) == 0) {
875 		sf->cpumask |= cpumask;
876 		invlpg(sf->kva);
877 	}
878 	if ((flags & SFB_CPUPRIVATE) == 0) {
879 		other_cpus = PCPU_GET(other_cpus) & ~sf->cpumask;
880 		if (other_cpus != 0) {
881 			sf->cpumask |= other_cpus;
882 			smp_masked_invlpg(other_cpus, sf->kva);
883 		}
884 	}
885 	sched_unpin();
886 #else
887 	if ((opte & (PG_V | PG_A)) ==  (PG_V | PG_A))
888 		pmap_invalidate_page(kernel_pmap, sf->kva);
889 #endif
890 done:
891 	mtx_unlock(&sf_buf_lock);
892 	return (sf);
893 }
894 
895 /*
896  * Remove a reference from the given sf_buf, adding it to the free
897  * list when its reference count reaches zero.  A freed sf_buf still,
898  * however, retains its virtual-to-physical mapping until it is
899  * recycled or reactivated by sf_buf_alloc(9).
900  */
901 void
902 sf_buf_free(struct sf_buf *sf)
903 {
904 
905 	mtx_lock(&sf_buf_lock);
906 	sf->ref_count--;
907 	if (sf->ref_count == 0) {
908 		TAILQ_INSERT_TAIL(&sf_buf_freelist, sf, free_entry);
909 		nsfbufsused--;
910 #ifdef XEN
911 /*
912  * Xen doesn't like having dangling R/W mappings
913  */
914 		pmap_qremove(sf->kva, 1);
915 		sf->m = NULL;
916 		LIST_REMOVE(sf, list_entry);
917 #endif
918 		if (sf_buf_alloc_want > 0)
919 			wakeup_one(&sf_buf_freelist);
920 	}
921 	mtx_unlock(&sf_buf_lock);
922 }
923 
924 /*
925  * Software interrupt handler for queued VM system processing.
926  */
927 void
928 swi_vm(void *dummy)
929 {
930 	if (busdma_swi_pending != 0)
931 		busdma_swi();
932 }
933 
934 /*
935  * Tell whether this address is in some physical memory region.
936  * Currently used by the kernel coredump code in order to avoid
937  * dumping the ``ISA memory hole'' which could cause indefinite hangs,
938  * or other unpredictable behaviour.
939  */
940 
941 int
942 is_physical_memory(vm_paddr_t addr)
943 {
944 
945 #ifdef DEV_ISA
946 	/* The ISA ``memory hole''. */
947 	if (addr >= 0xa0000 && addr < 0x100000)
948 		return 0;
949 #endif
950 
951 	/*
952 	 * stuff other tests for known memory-mapped devices (PCI?)
953 	 * here
954 	 */
955 
956 	return 1;
957 }
958