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
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.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
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.
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
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 * $FreeBSD: src/sys/i386/i386/vm_machdep.c,v 1.132.2.9 2003/01/25 19:02:23 dillon Exp $
42 * $DragonFly: src/sys/platform/pc32/i386/vm_machdep.c,v 1.47 2006/10/20 17:02:19 dillon Exp $
47 #include "opt_reset.h"
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/malloc.h>
54 #include <sys/interrupt.h>
55 #include <sys/vnode.h>
56 #include <sys/vmmeter.h>
57 #include <sys/kernel.h>
58 #include <sys/sysctl.h>
59 #include <sys/unistd.h>
61 #include <machine/clock.h>
62 #include <machine/cpu.h>
63 #include <machine/md_var.h>
64 #include <machine/smp.h>
65 #include <machine/pcb.h>
66 #include <machine/pcb_ext.h>
67 #include <machine/vm86.h>
68 #include <machine/segments.h>
69 #include <machine/globaldata.h> /* npxthread */
70 #include <machine/ipl.h> /* SWI_ */
73 #include <vm/vm_param.h>
75 #include <vm/vm_kern.h>
76 #include <vm/vm_page.h>
77 #include <vm/vm_map.h>
78 #include <vm/vm_extern.h>
81 #include <sys/thread2.h>
83 #include <bus/isa/i386/isa.h>
85 static void cpu_reset_real (void);
87 static void cpu_reset_proxy (void);
88 static u_int cpu_reset_proxyid;
89 static volatile u_int cpu_reset_proxy_active;
91 extern int _ucodesel, _udatasel;
95 * Finish a fork operation, with lwp lp2 nearly set up.
96 * Copy and update the pcb, set up the stack so that the child
97 * ready to run and return to user mode.
100 cpu_fork(struct lwp *lp1, struct lwp *lp2, int flags)
104 if ((flags & RFPROC) == 0) {
105 if ((flags & RFMEM) == 0) {
106 /* unshare user LDT */
107 struct pcb *pcb1 = lp1->lwp_thread->td_pcb;
108 struct pcb_ldt *pcb_ldt = pcb1->pcb_ldt;
109 if (pcb_ldt && pcb_ldt->ldt_refcnt > 1) {
110 pcb_ldt = user_ldt_alloc(pcb1,pcb_ldt->ldt_len);
112 pcb1->pcb_ldt = pcb_ldt;
120 /* Ensure that lp1's pcb is up to date. */
121 if (mdcpu->gd_npxthread == lp1->lwp_thread)
122 npxsave(lp1->lwp_thread->td_savefpu);
126 * Copy lp1's PCB. This really only applies to the
127 * debug registers and FP state, but its faster to just copy the
128 * whole thing. Because we only save the PCB at switchout time,
129 * the register state (including pcb_gs) may not be current.
131 pcb2 = lp2->lwp_thread->td_pcb;
132 *pcb2 = *lp1->lwp_thread->td_pcb;
135 * Create a new fresh stack for the new process.
136 * Copy the trap frame for the return to user mode as if from a
137 * syscall. This copies the user mode register values. The
138 * 16 byte offset saves space for vm86, and must match
139 * common_tss.esp0 (kernel stack pointer on entry from user mode)
141 * pcb_esp must allocate an additional call-return pointer below
142 * the trap frame which will be restored by cpu_restore from
143 * PCB_EIP, and the thread's td_sp pointer must allocate an
144 * additonal two worsd below the pcb_esp call-return pointer to
145 * hold the LWKT restore function pointer and eflags.
147 * The LWKT restore function pointer must be set to cpu_restore,
148 * which is our standard heavy weight process switch-in function.
149 * YYY eventually we should shortcut fork_return and fork_trampoline
150 * to use the LWKT restore function directly so we can get rid of
151 * all the extra crap we are setting up.
153 lp2->lwp_md.md_regs = (struct trapframe *)((char *)pcb2 - 16) - 1;
154 bcopy(lp1->lwp_md.md_regs, lp2->lwp_md.md_regs, sizeof(*lp2->lwp_md.md_regs));
157 * Set registers for trampoline to user mode. Leave space for the
158 * return address on stack. These are the kernel mode register values.
160 pcb2->pcb_cr3 = vtophys(vmspace_pmap(lp2->lwp_proc->p_vmspace)->pm_pdir);
162 pcb2->pcb_esi = (int)fork_return; /* fork_trampoline argument */
164 pcb2->pcb_esp = (int)lp2->lwp_md.md_regs - sizeof(void *);
165 pcb2->pcb_ebx = (int)lp2; /* fork_trampoline argument */
166 pcb2->pcb_eip = (int)fork_trampoline;
167 lp2->lwp_thread->td_sp = (char *)(pcb2->pcb_esp - sizeof(void *));
168 *(u_int32_t *)lp2->lwp_thread->td_sp = PSL_USER;
169 lp2->lwp_thread->td_sp -= sizeof(void *);
170 *(void **)lp2->lwp_thread->td_sp = (void *)cpu_heavy_restore;
175 pcb2->pcb_gs = rgs();
178 * pcb2->pcb_ldt: duplicated below, if necessary.
179 * pcb2->pcb_savefpu: cloned above.
180 * pcb2->pcb_flags: cloned above (always 0 here?).
181 * pcb2->pcb_onfault: cloned above (always NULL here?).
185 * XXX don't copy the i/o pages. this should probably be fixed.
189 /* Copy the LDT, if necessary. */
190 if (pcb2->pcb_ldt != 0) {
192 pcb2->pcb_ldt->ldt_refcnt++;
194 pcb2->pcb_ldt = user_ldt_alloc(pcb2,
195 pcb2->pcb_ldt->ldt_len);
198 bcopy(&lp1->lwp_thread->td_tls, &lp2->lwp_thread->td_tls,
199 sizeof(lp2->lwp_thread->td_tls));
201 * Now, cpu_switch() can schedule the new process.
202 * pcb_esp is loaded pointing to the cpu_switch() stack frame
203 * containing the return address when exiting cpu_switch.
204 * This will normally be to fork_trampoline(), which will have
205 * %ebx loaded with the new proc's pointer. fork_trampoline()
206 * will set up a stack to call fork_return(p, frame); to complete
207 * the return to user-mode.
212 * Intercept the return address from a freshly forked process that has NOT
213 * been scheduled yet.
215 * This is needed to make kernel threads stay in kernel mode.
218 cpu_set_fork_handler(struct lwp *lp, void (*func)(void *), void *arg)
221 * Note that the trap frame follows the args, so the function
222 * is really called like this: func(arg, frame);
224 lp->lwp_thread->td_pcb->pcb_esi = (int) func; /* function */
225 lp->lwp_thread->td_pcb->pcb_ebx = (int) arg; /* first arg */
229 cpu_set_thread_handler(thread_t td, void (*rfunc)(void), void *func, void *arg)
231 td->td_pcb->pcb_esi = (int)func;
232 td->td_pcb->pcb_ebx = (int) arg;
233 td->td_switch = cpu_lwkt_switch;
234 td->td_sp -= sizeof(void *);
235 *(void **)td->td_sp = rfunc; /* exit function on return */
236 td->td_sp -= sizeof(void *);
237 *(void **)td->td_sp = cpu_kthread_restore;
243 struct thread *td = curthread;
252 * If we were using a private TSS do a forced-switch to ourselves
253 * to switch back to the common TSS before freeing it.
256 if ((ext = pcb->pcb_ext) != NULL) {
261 kmem_free(kernel_map, (vm_offset_t)ext, ctob(IOPAGES + 1));
264 if (pcb->pcb_flags & PCB_DBREGS) {
266 * disable all hardware breakpoints
269 pcb->pcb_flags &= ~PCB_DBREGS;
271 td->td_gd->gd_cnt.v_swtch++;
273 crit_enter_quick(td);
274 lwkt_deschedule_self(td);
275 lwkt_remove_tdallq(td);
280 * Terminate the current thread. The caller must have already acquired
281 * the thread's rwlock and placed it on a reap list or otherwise notified
282 * a reaper of its existance. We set a special assembly switch function which
283 * releases td_rwlock after it has cleaned up the MMU state and switched
286 * Must be caller from a critical section and with the thread descheduled.
289 cpu_thread_exit(void)
291 curthread->td_switch = cpu_exit_switch;
292 curthread->td_flags |= TDF_EXITING;
298 * Process Reaper. Called after the caller has acquired the thread's
299 * rwlock and removed it from the reap list.
302 cpu_proc_wait(struct proc *p)
306 /* drop per-process resources */
307 td = pmap_dispose_proc(p);
309 lwkt_free_thread(td);
313 * Dump the machine specific header information at the start of a core dump.
316 cpu_coredump(struct thread *td, struct vnode *vp, struct ucred *cred)
318 struct proc *p = td->td_proc;
323 tempuser = kmalloc(ctob(UPAGES), M_TEMP, M_WAITOK);
327 bzero(tempuser, ctob(UPAGES));
328 bcopy(p->p_addr, tempuser, sizeof(struct user));
329 bcopy(p->p_md.md_regs,
330 tempuser + ((caddr_t) p->p_md.md_regs - (caddr_t) p->p_addr),
331 sizeof(struct trapframe));
332 bcopy(p->p_thread->td_pcb, tempuser + ((char *)p->p_thread->td_pcb - (char *)p->p_addr), sizeof(struct pcb));
334 error = vn_rdwr(UIO_WRITE, vp, (caddr_t) tempuser, ctob(UPAGES),
335 (off_t)0, UIO_SYSSPACE, IO_UNIT, cred, (int *)NULL);
337 kfree(tempuser, M_TEMP);
344 setredzone(u_short *pte, caddr_t vaddr)
346 /* eventually do this by setting up an expand-down stack segment
347 for ss0: selector, allowing stack access down to top of u.
348 this means though that protection violations need to be handled
349 thru a double fault exception that must do an integral task
350 switch to a known good context, within which a dump can be
351 taken. a sensible scheme might be to save the initial context
352 used by sched (that has physical memory mapped 1:1 at bottom)
353 and take the dump while still in mapped mode */
358 * Convert kernel VA to physical address
365 pa = pmap_kextract((vm_offset_t)addr);
367 panic("kvtop: zero page frame");
372 * Force reset the processor by invalidating the entire address space!
377 cpu_reset_proxy(void)
381 cpu_reset_proxy_active = 1;
382 while (cpu_reset_proxy_active == 1)
383 ; /* Wait for other cpu to disable interupts */
384 saved_mp_lock = mp_lock;
385 mp_lock = 0; /* BSP */
386 printf("cpu_reset_proxy: Grabbed mp lock for BSP\n");
387 cpu_reset_proxy_active = 3;
388 while (cpu_reset_proxy_active == 3)
389 ; /* Wait for other cpu to enable interrupts */
390 stop_cpus((1<<cpu_reset_proxyid));
391 printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
401 if (smp_active_mask == 1) {
407 printf("cpu_reset called on cpu#%d\n",mycpu->gd_cpuid);
409 map = mycpu->gd_other_cpus & ~stopped_cpus & smp_active_mask;
412 printf("cpu_reset: Stopping other CPUs\n");
413 stop_cpus(map); /* Stop all other CPUs */
416 if (mycpu->gd_cpuid == 0) {
421 /* We are not BSP (CPU #0) */
423 cpu_reset_proxyid = mycpu->gd_cpuid;
424 cpustop_restartfunc = cpu_reset_proxy;
425 printf("cpu_reset: Restarting BSP\n");
426 started_cpus = (1<<0); /* Restart CPU #0 */
429 while (cpu_reset_proxy_active == 0 && cnt < 10000000)
430 cnt++; /* Wait for BSP to announce restart */
431 if (cpu_reset_proxy_active == 0)
432 printf("cpu_reset: Failed to restart BSP\n");
433 __asm __volatile("cli" : : : "memory");
434 cpu_reset_proxy_active = 2;
436 while (cpu_reset_proxy_active == 2 && cnt < 10000000)
437 cnt++; /* Do nothing */
438 if (cpu_reset_proxy_active == 2) {
439 printf("cpu_reset: BSP did not grab mp lock\n");
440 cpu_reset_real(); /* XXX: Bogus ? */
442 cpu_reset_proxy_active = 4;
443 __asm __volatile("sti" : : : "memory");
457 * Attempt to do a CPU reset via the keyboard controller,
458 * do not turn of the GateA20, as any machine that fails
459 * to do the reset here would then end up in no man's land.
462 #if !defined(BROKEN_KEYBOARD_RESET)
463 outb(IO_KBD + 4, 0xFE);
464 DELAY(500000); /* wait 0.5 sec to see if that did it */
465 printf("Keyboard reset did not work, attempting CPU shutdown\n");
466 DELAY(1000000); /* wait 1 sec for printf to complete */
468 /* force a shutdown by unmapping entire address space ! */
469 bzero((caddr_t) PTD, PAGE_SIZE);
471 /* "good night, sweet prince .... <THUNK!>" */
478 grow_stack(struct proc *p, u_int sp)
482 rv = vm_map_growstack (p, sp);
483 if (rv != KERN_SUCCESS)
489 SYSCTL_DECL(_vm_stats_misc);
491 static int cnt_prezero;
493 SYSCTL_INT(_vm_stats_misc, OID_AUTO,
494 cnt_prezero, CTLFLAG_RD, &cnt_prezero, 0, "");
497 * Implement the pre-zeroed page mechanism.
498 * This routine is called from the idle loop.
501 #define ZIDLE_LO(v) ((v) * 2 / 3)
502 #define ZIDLE_HI(v) ((v) * 4 / 5)
505 vm_page_zero_idle(void)
507 static int free_rover;
508 static int zero_state;
512 * Attempt to maintain approximately 1/2 of our free pages in a
513 * PG_ZERO'd state. Add some hysteresis to (attempt to) avoid
514 * generally zeroing a page when the system is near steady-state.
515 * Otherwise we might get 'flutter' during disk I/O / IPC or
516 * fast sleeps. We also do not want to be continuously zeroing
517 * pages because doing so may flush our L1 and L2 caches too much.
520 if (zero_state && vm_page_zero_count >= ZIDLE_LO(vmstats.v_free_count))
522 if (vm_page_zero_count >= ZIDLE_HI(vmstats.v_free_count))
529 __asm __volatile("sti" : : : "memory");
531 m = vm_page_list_find(PQ_FREE, free_rover, FALSE);
532 if (m != NULL && (m->flags & PG_ZERO) == 0) {
533 vm_page_queues[m->queue].lcnt--;
534 TAILQ_REMOVE(&vm_page_queues[m->queue].pl, m, pageq);
537 pmap_zero_page(VM_PAGE_TO_PHYS(m));
539 vm_page_flag_set(m, PG_ZERO);
540 m->queue = PQ_FREE + m->pc;
541 vm_page_queues[m->queue].lcnt++;
542 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m,
544 ++vm_page_zero_count;
546 if (vm_page_zero_count >= ZIDLE_HI(vmstats.v_free_count))
549 free_rover = (free_rover + PQ_PRIME2) & PQ_L2_MASK;
551 __asm __volatile("cli" : : : "memory");
560 * We have to enable interrupts for a moment if the try_mplock fails
561 * in order to potentially take an IPI. XXX this should be in
564 __asm __volatile("sti; nop; cli" : : : "memory");
569 swi_vm(void *arg, void *frame)
571 if (busdma_swi_pending != 0)
576 swi_vm_setup(void *arg)
578 register_swi(SWI_VM, swi_vm, NULL, "swi_vm", NULL);
581 SYSINIT(vm_setup, SI_SUB_CPU, SI_ORDER_ANY, swi_vm_setup, NULL);
585 * Tell whether this address is in some physical memory region.
586 * Currently used by the kernel coredump code in order to avoid
587 * dumping the ``ISA memory hole'' which could cause indefinite hangs,
588 * or other unpredictable behaviour.
592 is_physical_memory(vm_offset_t addr)
596 /* The ISA ``memory hole''. */
597 if (addr >= 0xa0000 && addr < 0x100000)
602 * stuff other tests for known memory-mapped devices (PCI?)