xref: /freebsd/sys/i386/i386/vm_machdep.c (revision c0b9f4fe659b6839541970eb5675e57f4d814969)
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 			pldt = mdp1->md_ldt;
162 			if (pldt && pldt->ldt_refcnt > 1) {
163 				pldt = user_ldt_alloc(mdp1, pldt->ldt_len);
164 				if (pldt == NULL)
165 					panic("could not copy LDT");
166 				mdp1->md_ldt = pldt;
167 				set_user_ldt(mdp1);
168 				user_ldt_free(td1);
169 			}
170 		}
171 		return;
172 	}
173 
174 	/* Ensure that p1's pcb is up to date. */
175 #ifdef DEV_NPX
176 	if (td1 == curthread)
177 		td1->td_pcb->pcb_gs = rgs();
178 	savecrit = intr_disable();
179 	if (PCPU_GET(fpcurthread) == td1)
180 		npxsave(&td1->td_pcb->pcb_save);
181 	intr_restore(savecrit);
182 #endif
183 
184 	/* Point the pcb to the top of the stack */
185 	pcb2 = (struct pcb *)(td2->td_kstack +
186 	    td2->td_kstack_pages * PAGE_SIZE) - 1;
187 	td2->td_pcb = pcb2;
188 
189 	/* Copy p1's pcb */
190 	bcopy(td1->td_pcb, pcb2, sizeof(*pcb2));
191 
192 	/* Point mdproc and then copy over td1's contents */
193 	mdp2 = &p2->p_md;
194 	bcopy(&p1->p_md, mdp2, sizeof(*mdp2));
195 
196 	/*
197 	 * Create a new fresh stack for the new process.
198 	 * Copy the trap frame for the return to user mode as if from a
199 	 * syscall.  This copies most of the user mode register values.
200 	 * The -16 is so we can expand the trapframe if we go to vm86.
201 	 */
202 	td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb - 16) - 1;
203 	bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe));
204 
205 	td2->td_frame->tf_eax = 0;		/* Child returns zero */
206 	td2->td_frame->tf_eflags &= ~PSL_C;	/* success */
207 	td2->td_frame->tf_edx = 1;
208 
209 	/*
210 	 * If the parent process has the trap bit set (i.e. a debugger had
211 	 * single stepped the process to the system call), we need to clear
212 	 * the trap flag from the new frame unless the debugger had set PF_FORK
213 	 * on the parent.  Otherwise, the child will receive a (likely
214 	 * unexpected) SIGTRAP when it executes the first instruction after
215 	 * returning  to userland.
216 	 */
217 	if ((p1->p_pfsflags & PF_FORK) == 0)
218 		td2->td_frame->tf_eflags &= ~PSL_T;
219 
220 	/*
221 	 * Set registers for trampoline to user mode.  Leave space for the
222 	 * return address on stack.  These are the kernel mode register values.
223 	 */
224 #ifdef PAE
225 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdpt);
226 #else
227 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdir);
228 #endif
229 	pcb2->pcb_edi = 0;
230 	pcb2->pcb_esi = (int)fork_return;	/* fork_trampoline argument */
231 	pcb2->pcb_ebp = 0;
232 	pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *);
233 	pcb2->pcb_ebx = (int)td2;		/* fork_trampoline argument */
234 	pcb2->pcb_eip = (int)fork_trampoline;
235 	pcb2->pcb_psl = PSL_KERNEL;		/* ints disabled */
236 	pcb2->pcb_gs = rgs();
237 	/*-
238 	 * pcb2->pcb_dr*:	cloned above.
239 	 * pcb2->pcb_savefpu:	cloned above.
240 	 * pcb2->pcb_flags:	cloned above.
241 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
242 	 * pcb2->pcb_gs:	cloned above.
243 	 * pcb2->pcb_ext:	cleared below.
244 	 */
245 
246 	/*
247 	 * XXX don't copy the i/o pages.  this should probably be fixed.
248 	 */
249 	pcb2->pcb_ext = 0;
250 
251 	/* Copy the LDT, if necessary. */
252 	mtx_lock_spin(&sched_lock);
253 	if (mdp2->md_ldt != NULL) {
254 		if (flags & RFMEM) {
255 			mdp2->md_ldt->ldt_refcnt++;
256 		} else {
257 			mdp2->md_ldt = user_ldt_alloc(mdp2,
258 			    mdp2->md_ldt->ldt_len);
259 			if (mdp2->md_ldt == NULL)
260 				panic("could not copy LDT");
261 		}
262 	}
263 	mtx_unlock_spin(&sched_lock);
264 
265 	/* Setup to release sched_lock in fork_exit(). */
266 	td2->td_md.md_spinlock_count = 1;
267 	td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
268 
269 	/*
270 	 * Now, cpu_switch() can schedule the new process.
271 	 * pcb_esp is loaded pointing to the cpu_switch() stack frame
272 	 * containing the return address when exiting cpu_switch.
273 	 * This will normally be to fork_trampoline(), which will have
274 	 * %ebx loaded with the new proc's pointer.  fork_trampoline()
275 	 * will set up a stack to call fork_return(p, frame); to complete
276 	 * the return to user-mode.
277 	 */
278 }
279 
280 /*
281  * Intercept the return address from a freshly forked process that has NOT
282  * been scheduled yet.
283  *
284  * This is needed to make kernel threads stay in kernel mode.
285  */
286 void
287 cpu_set_fork_handler(td, func, arg)
288 	struct thread *td;
289 	void (*func)(void *);
290 	void *arg;
291 {
292 	/*
293 	 * Note that the trap frame follows the args, so the function
294 	 * is really called like this:  func(arg, frame);
295 	 */
296 	td->td_pcb->pcb_esi = (int) func;	/* function */
297 	td->td_pcb->pcb_ebx = (int) arg;	/* first arg */
298 }
299 
300 void
301 cpu_exit(struct thread *td)
302 {
303 
304 	/*
305 	 * If this process has a custom LDT, release it.  Reset pc->pcb_gs
306 	 * and %gs before we free it in case they refer to an LDT entry.
307 	 */
308 	if (td->td_proc->p_md.md_ldt) {
309 		td->td_pcb->pcb_gs = _udatasel;
310 		load_gs(_udatasel);
311 		user_ldt_free(td);
312 	}
313 }
314 
315 void
316 cpu_thread_exit(struct thread *td)
317 {
318 
319 #ifdef DEV_NPX
320 	if (td == PCPU_GET(fpcurthread))
321 		npxdrop();
322 #endif
323 
324 	/* Disable any hardware breakpoints. */
325 	if (td->td_pcb->pcb_flags & PCB_DBREGS) {
326 		reset_dbregs();
327 		td->td_pcb->pcb_flags &= ~PCB_DBREGS;
328 	}
329 }
330 
331 void
332 cpu_thread_clean(struct thread *td)
333 {
334 	struct pcb *pcb;
335 
336 	pcb = td->td_pcb;
337 	if (pcb->pcb_ext != NULL) {
338 		/* XXXKSE  XXXSMP  not SMP SAFE.. what locks do we have? */
339 		/* if (pcb->pcb_ext->ext_refcount-- == 1) ?? */
340 		/*
341 		 * XXX do we need to move the TSS off the allocated pages
342 		 * before freeing them?  (not done here)
343 		 */
344 		kmem_free(kernel_map, (vm_offset_t)pcb->pcb_ext,
345 		    ctob(IOPAGES + 1));
346 		pcb->pcb_ext = NULL;
347 	}
348 }
349 
350 void
351 cpu_thread_swapin(struct thread *td)
352 {
353 }
354 
355 void
356 cpu_thread_swapout(struct thread *td)
357 {
358 }
359 
360 void
361 cpu_thread_setup(struct thread *td)
362 {
363 
364 	td->td_pcb = (struct pcb *)(td->td_kstack +
365 	    td->td_kstack_pages * PAGE_SIZE) - 1;
366 	td->td_frame = (struct trapframe *)((caddr_t)td->td_pcb - 16) - 1;
367 	td->td_pcb->pcb_ext = NULL;
368 }
369 
370 /*
371  * Initialize machine state (pcb and trap frame) for a new thread about to
372  * upcall. Put enough state in the new thread's PCB to get it to go back
373  * userret(), where we can intercept it again to set the return (upcall)
374  * Address and stack, along with those from upcals that are from other sources
375  * such as those generated in thread_userret() itself.
376  */
377 void
378 cpu_set_upcall(struct thread *td, struct thread *td0)
379 {
380 	struct pcb *pcb2;
381 
382 	/* Point the pcb to the top of the stack. */
383 	pcb2 = td->td_pcb;
384 
385 	/*
386 	 * Copy the upcall pcb.  This loads kernel regs.
387 	 * Those not loaded individually below get their default
388 	 * values here.
389 	 *
390 	 * XXXKSE It might be a good idea to simply skip this as
391 	 * the values of the other registers may be unimportant.
392 	 * This would remove any requirement for knowing the KSE
393 	 * at this time (see the matching comment below for
394 	 * more analysis) (need a good safe default).
395 	 */
396 	bcopy(td0->td_pcb, pcb2, sizeof(*pcb2));
397 	pcb2->pcb_flags &= ~(PCB_NPXTRAP|PCB_NPXINITDONE);
398 
399 	/*
400 	 * Create a new fresh stack for the new thread.
401 	 * The -16 is so we can expand the trapframe if we go to vm86.
402 	 * Don't forget to set this stack value into whatever supplies
403 	 * the address for the fault handlers.
404 	 * The contexts are filled in at the time we actually DO the
405 	 * upcall as only then do we know which KSE we got.
406 	 */
407 	bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe));
408 
409 	/*
410 	 * Set registers for trampoline to user mode.  Leave space for the
411 	 * return address on stack.  These are the kernel mode register values.
412 	 */
413 #ifdef PAE
414 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(td->td_proc->p_vmspace)->pm_pdpt);
415 #else
416 	pcb2->pcb_cr3 = vtophys(vmspace_pmap(td->td_proc->p_vmspace)->pm_pdir);
417 #endif
418 	pcb2->pcb_edi = 0;
419 	pcb2->pcb_esi = (int)fork_return;		    /* trampoline arg */
420 	pcb2->pcb_ebp = 0;
421 	pcb2->pcb_esp = (int)td->td_frame - sizeof(void *); /* trampoline arg */
422 	pcb2->pcb_ebx = (int)td;			    /* trampoline arg */
423 	pcb2->pcb_eip = (int)fork_trampoline;
424 	pcb2->pcb_psl &= ~(PSL_I);	/* interrupts must be disabled */
425 	pcb2->pcb_gs = rgs();
426 	/*
427 	 * If we didn't copy the pcb, we'd need to do the following registers:
428 	 * pcb2->pcb_dr*:	cloned above.
429 	 * pcb2->pcb_savefpu:	cloned above.
430 	 * pcb2->pcb_flags:	cloned above.
431 	 * pcb2->pcb_onfault:	cloned above (always NULL here?).
432 	 * pcb2->pcb_gs:	cloned above.  XXXKSE ???
433 	 * pcb2->pcb_ext:	cleared below.
434 	 */
435 	pcb2->pcb_ext = NULL;
436 
437 	/* Setup to release sched_lock in fork_exit(). */
438 	td->td_md.md_spinlock_count = 1;
439 	td->td_md.md_saved_flags = PSL_KERNEL | PSL_I;
440 }
441 
442 /*
443  * Set that machine state for performing an upcall that has to
444  * be done in thread_userret() so that those upcalls generated
445  * in thread_userret() itself can be done as well.
446  */
447 void
448 cpu_set_upcall_kse(struct thread *td, void (*entry)(void *), void *arg,
449 	stack_t *stack)
450 {
451 
452 	/*
453 	 * Do any extra cleaning that needs to be done.
454 	 * The thread may have optional components
455 	 * that are not present in a fresh thread.
456 	 * This may be a recycled thread so make it look
457 	 * as though it's newly allocated.
458 	 */
459 	cpu_thread_clean(td);
460 
461 	/*
462 	 * Set the trap frame to point at the beginning of the uts
463 	 * function.
464 	 */
465 	td->td_frame->tf_ebp = 0;
466 	td->td_frame->tf_esp =
467 	    (int)stack->ss_sp + stack->ss_size - 16;
468 	td->td_frame->tf_eip = (int)entry;
469 
470 	/*
471 	 * Pass the address of the mailbox for this kse to the uts
472 	 * function as a parameter on the stack.
473 	 */
474 	suword((void *)(td->td_frame->tf_esp + sizeof(void *)),
475 	    (int)arg);
476 }
477 
478 int
479 cpu_set_user_tls(struct thread *td, void *tls_base)
480 {
481 	struct segment_descriptor sd;
482 	uint32_t base;
483 
484 	/*
485 	 * Construct a descriptor and store it in the pcb for
486 	 * the next context switch.  Also store it in the gdt
487 	 * so that the load of tf_fs into %fs will activate it
488 	 * at return to userland.
489 	 */
490 	base = (uint32_t)tls_base;
491 	sd.sd_lobase = base & 0xffffff;
492 	sd.sd_hibase = (base >> 24) & 0xff;
493 	sd.sd_lolimit = 0xffff;	/* 4GB limit, wraps around */
494 	sd.sd_hilimit = 0xf;
495 	sd.sd_type  = SDT_MEMRWA;
496 	sd.sd_dpl   = SEL_UPL;
497 	sd.sd_p     = 1;
498 	sd.sd_xx    = 0;
499 	sd.sd_def32 = 1;
500 	sd.sd_gran  = 1;
501 	critical_enter();
502 	/* set %gs */
503 	td->td_pcb->pcb_gsd = sd;
504 	if (td == curthread) {
505 		PCPU_GET(fsgs_gdt)[1] = sd;
506 		load_gs(GSEL(GUGS_SEL, SEL_UPL));
507 	}
508 	critical_exit();
509 	return (0);
510 }
511 
512 /*
513  * Convert kernel VA to physical address
514  */
515 vm_paddr_t
516 kvtop(void *addr)
517 {
518 	vm_paddr_t pa;
519 
520 	pa = pmap_kextract((vm_offset_t)addr);
521 	if (pa == 0)
522 		panic("kvtop: zero page frame");
523 	return (pa);
524 }
525 
526 #ifdef SMP
527 static void
528 cpu_reset_proxy()
529 {
530 
531 	cpu_reset_proxy_active = 1;
532 	while (cpu_reset_proxy_active == 1)
533 		;	/* Wait for other cpu to see that we've started */
534 	stop_cpus((1<<cpu_reset_proxyid));
535 	printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
536 	DELAY(1000000);
537 	cpu_reset_real();
538 }
539 #endif
540 
541 void
542 cpu_reset()
543 {
544 #ifdef XBOX
545 	if (arch_i386_is_xbox) {
546 		/* Kick the PIC16L, it can reboot the box */
547 		pic16l_reboot();
548 		for (;;);
549 	}
550 #endif
551 
552 #ifdef SMP
553 	u_int cnt, map;
554 
555 	if (smp_active) {
556 		map = PCPU_GET(other_cpus) & ~stopped_cpus;
557 		if (map != 0) {
558 			printf("cpu_reset: Stopping other CPUs\n");
559 			stop_cpus(map);
560 		}
561 
562 		if (PCPU_GET(cpuid) != 0) {
563 			cpu_reset_proxyid = PCPU_GET(cpuid);
564 			cpustop_restartfunc = cpu_reset_proxy;
565 			cpu_reset_proxy_active = 0;
566 			printf("cpu_reset: Restarting BSP\n");
567 
568 			/* Restart CPU #0. */
569 			/* XXX: restart_cpus(1 << 0); */
570 			atomic_store_rel_int(&started_cpus, (1 << 0));
571 
572 			cnt = 0;
573 			while (cpu_reset_proxy_active == 0 && cnt < 10000000)
574 				cnt++;	/* Wait for BSP to announce restart */
575 			if (cpu_reset_proxy_active == 0)
576 				printf("cpu_reset: Failed to restart BSP\n");
577 			enable_intr();
578 			cpu_reset_proxy_active = 2;
579 
580 			while (1);
581 			/* NOTREACHED */
582 		}
583 
584 		DELAY(1000000);
585 	}
586 #endif
587 	cpu_reset_real();
588 	/* NOTREACHED */
589 }
590 
591 static void
592 cpu_reset_real()
593 {
594 
595 #ifdef CPU_ELAN
596 	if (elan_mmcr != NULL)
597 		elan_mmcr->RESCFG = 1;
598 #endif
599 
600 	if (cpu == CPU_GEODE1100) {
601 		/* Attempt Geode's own reset */
602 		outl(0xcf8, 0x80009044ul);
603 		outl(0xcfc, 0xf);
604 	}
605 
606 #ifdef PC98
607 	/*
608 	 * Attempt to do a CPU reset via CPU reset port.
609 	 */
610 	disable_intr();
611 	if ((inb(0x35) & 0xa0) != 0xa0) {
612 		outb(0x37, 0x0f);		/* SHUT0 = 0. */
613 		outb(0x37, 0x0b);		/* SHUT1 = 0. */
614 	}
615 	outb(0xf0, 0x00);		/* Reset. */
616 #else
617 #if !defined(BROKEN_KEYBOARD_RESET)
618 	/*
619 	 * Attempt to do a CPU reset via the keyboard controller,
620 	 * do not turn off GateA20, as any machine that fails
621 	 * to do the reset here would then end up in no man's land.
622 	 */
623 	outb(IO_KBD + 4, 0xFE);
624 	DELAY(500000);	/* wait 0.5 sec to see if that did it */
625 	printf("Keyboard reset did not work, attempting CPU shutdown\n");
626 	DELAY(1000000);	/* wait 1 sec for printf to complete */
627 #endif
628 #endif /* PC98 */
629 
630 	/* Force a shutdown by unmapping entire address space. */
631 	bzero((caddr_t)PTD, NBPTD);
632 
633 	/* "good night, sweet prince .... <THUNK!>" */
634 	invltlb();
635 	/* NOTREACHED */
636 	while(1);
637 }
638 
639 /*
640  * Allocate a pool of sf_bufs (sendfile(2) or "super-fast" if you prefer. :-))
641  */
642 static void
643 sf_buf_init(void *arg)
644 {
645 	struct sf_buf *sf_bufs;
646 	vm_offset_t sf_base;
647 	int i;
648 
649 	nsfbufs = NSFBUFS;
650 	TUNABLE_INT_FETCH("kern.ipc.nsfbufs", &nsfbufs);
651 
652 	sf_buf_active = hashinit(nsfbufs, M_TEMP, &sf_buf_hashmask);
653 	TAILQ_INIT(&sf_buf_freelist);
654 	sf_base = kmem_alloc_nofault(kernel_map, nsfbufs * PAGE_SIZE);
655 	sf_bufs = malloc(nsfbufs * sizeof(struct sf_buf), M_TEMP,
656 	    M_NOWAIT | M_ZERO);
657 	for (i = 0; i < nsfbufs; i++) {
658 		sf_bufs[i].kva = sf_base + i * PAGE_SIZE;
659 		TAILQ_INSERT_TAIL(&sf_buf_freelist, &sf_bufs[i], free_entry);
660 	}
661 	sf_buf_alloc_want = 0;
662 	mtx_init(&sf_buf_lock, "sf_buf", NULL, MTX_DEF);
663 }
664 
665 /*
666  * Get an sf_buf from the freelist. Will block if none are available.
667  */
668 struct sf_buf *
669 sf_buf_alloc(struct vm_page *m, int flags)
670 {
671 	pt_entry_t opte, *ptep;
672 	struct sf_head *hash_list;
673 	struct sf_buf *sf;
674 #ifdef SMP
675 	cpumask_t cpumask, other_cpus;
676 #endif
677 	int error;
678 
679 	KASSERT(curthread->td_pinned > 0 || (flags & SFB_CPUPRIVATE) == 0,
680 	    ("sf_buf_alloc(SFB_CPUPRIVATE): curthread not pinned"));
681 	hash_list = &sf_buf_active[SF_BUF_HASH(m)];
682 	mtx_lock(&sf_buf_lock);
683 	LIST_FOREACH(sf, hash_list, list_entry) {
684 		if (sf->m == m) {
685 			sf->ref_count++;
686 			if (sf->ref_count == 1) {
687 				TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
688 				nsfbufsused++;
689 				nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
690 			}
691 #ifdef SMP
692 			goto shootdown;
693 #else
694 			goto done;
695 #endif
696 		}
697 	}
698 	while ((sf = TAILQ_FIRST(&sf_buf_freelist)) == NULL) {
699 		if (flags & SFB_NOWAIT)
700 			goto done;
701 		sf_buf_alloc_want++;
702 		mbstat.sf_allocwait++;
703 		error = msleep(&sf_buf_freelist, &sf_buf_lock,
704 		    (flags & SFB_CATCH) ? PCATCH | PVM : PVM, "sfbufa", 0);
705 		sf_buf_alloc_want--;
706 
707 		/*
708 		 * If we got a signal, don't risk going back to sleep.
709 		 */
710 		if (error)
711 			goto done;
712 	}
713 	TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
714 	if (sf->m != NULL)
715 		LIST_REMOVE(sf, list_entry);
716 	LIST_INSERT_HEAD(hash_list, sf, list_entry);
717 	sf->ref_count = 1;
718 	sf->m = m;
719 	nsfbufsused++;
720 	nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
721 
722 	/*
723 	 * Update the sf_buf's virtual-to-physical mapping, flushing the
724 	 * virtual address from the TLB.  Since the reference count for
725 	 * the sf_buf's old mapping was zero, that mapping is not
726 	 * currently in use.  Consequently, there is no need to exchange
727 	 * the old and new PTEs atomically, even under PAE.
728 	 */
729 	ptep = vtopte(sf->kva);
730 	opte = *ptep;
731 	*ptep = VM_PAGE_TO_PHYS(m) | pgeflag | PG_RW | PG_V;
732 
733 	/*
734 	 * Avoid unnecessary TLB invalidations: If the sf_buf's old
735 	 * virtual-to-physical mapping was not used, then any processor
736 	 * that has invalidated the sf_buf's virtual address from its TLB
737 	 * since the last used mapping need not invalidate again.
738 	 */
739 #ifdef SMP
740 	if ((opte & (PG_V | PG_A)) ==  (PG_V | PG_A))
741 		sf->cpumask = 0;
742 shootdown:
743 	sched_pin();
744 	cpumask = PCPU_GET(cpumask);
745 	if ((sf->cpumask & cpumask) == 0) {
746 		sf->cpumask |= cpumask;
747 		invlpg(sf->kva);
748 	}
749 	if ((flags & SFB_CPUPRIVATE) == 0) {
750 		other_cpus = PCPU_GET(other_cpus) & ~sf->cpumask;
751 		if (other_cpus != 0) {
752 			sf->cpumask |= other_cpus;
753 			mtx_lock_spin(&smp_ipi_mtx);
754 			smp_masked_invlpg(other_cpus, sf->kva);
755 			mtx_unlock_spin(&smp_ipi_mtx);
756 		}
757 	}
758 	sched_unpin();
759 #else
760 	if ((opte & (PG_V | PG_A)) ==  (PG_V | PG_A))
761 		pmap_invalidate_page(kernel_pmap, sf->kva);
762 #endif
763 done:
764 	mtx_unlock(&sf_buf_lock);
765 	return (sf);
766 }
767 
768 /*
769  * Remove a reference from the given sf_buf, adding it to the free
770  * list when its reference count reaches zero.  A freed sf_buf still,
771  * however, retains its virtual-to-physical mapping until it is
772  * recycled or reactivated by sf_buf_alloc(9).
773  */
774 void
775 sf_buf_free(struct sf_buf *sf)
776 {
777 
778 	mtx_lock(&sf_buf_lock);
779 	sf->ref_count--;
780 	if (sf->ref_count == 0) {
781 		TAILQ_INSERT_TAIL(&sf_buf_freelist, sf, free_entry);
782 		nsfbufsused--;
783 		if (sf_buf_alloc_want > 0)
784 			wakeup_one(&sf_buf_freelist);
785 	}
786 	mtx_unlock(&sf_buf_lock);
787 }
788 
789 /*
790  * Software interrupt handler for queued VM system processing.
791  */
792 void
793 swi_vm(void *dummy)
794 {
795 	if (busdma_swi_pending != 0)
796 		busdma_swi();
797 }
798 
799 /*
800  * Tell whether this address is in some physical memory region.
801  * Currently used by the kernel coredump code in order to avoid
802  * dumping the ``ISA memory hole'' which could cause indefinite hangs,
803  * or other unpredictable behaviour.
804  */
805 
806 int
807 is_physical_memory(vm_paddr_t addr)
808 {
809 
810 #ifdef DEV_ISA
811 	/* The ISA ``memory hole''. */
812 	if (addr >= 0xa0000 && addr < 0x100000)
813 		return 0;
814 #endif
815 
816 	/*
817 	 * stuff other tests for known memory-mapped devices (PCI?)
818 	 * here
819 	 */
820 
821 	return 1;
822 }
823