vm_machdep.c revision 43758
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 * $Id: vm_machdep.c,v 1.117 1999/02/08 00:37:35 dillon Exp $ 42 */ 43 44#include "npx.h" 45#include "opt_user_ldt.h" 46#include "opt_vm86.h" 47#ifdef PC98 48#include "opt_pc98.h" 49#endif 50 51#include <sys/param.h> 52#include <sys/systm.h> 53#include <sys/proc.h> 54#include <sys/malloc.h> 55#include <sys/buf.h> 56#include <sys/vnode.h> 57#include <sys/vmmeter.h> 58#include <sys/kernel.h> 59#include <sys/sysctl.h> 60 61#include <machine/clock.h> 62#include <machine/cpu.h> 63#include <machine/md_var.h> 64#ifdef SMP 65#include <machine/smp.h> 66#endif 67#ifdef VM86 68#include <machine/pcb_ext.h> 69#include <machine/vm86.h> 70#endif 71 72#include <vm/vm.h> 73#include <vm/vm_param.h> 74#include <vm/vm_prot.h> 75#include <sys/lock.h> 76#include <vm/vm_kern.h> 77#include <vm/vm_page.h> 78#include <vm/vm_map.h> 79#include <vm/vm_extern.h> 80 81#include <sys/user.h> 82 83#ifdef PC98 84#include <pc98/pc98/pc98.h> 85#else 86#include <i386/isa/isa.h> 87#endif 88 89static void cpu_reset_real __P((void)); 90#ifdef SMP 91static void cpu_reset_proxy __P((void)); 92static u_int cpu_reset_proxyid; 93static volatile u_int cpu_reset_proxy_active; 94#endif 95 96/* 97 * quick version of vm_fault 98 */ 99void 100vm_fault_quick(v, prot) 101 caddr_t v; 102 int prot; 103{ 104 if (prot & VM_PROT_WRITE) 105 subyte(v, fubyte(v)); 106 else 107 fubyte(v); 108} 109 110/* 111 * Finish a fork operation, with process p2 nearly set up. 112 * Copy and update the pcb, set up the stack so that the child 113 * ready to run and return to user mode. 114 */ 115void 116cpu_fork(p1, p2) 117 register struct proc *p1, *p2; 118{ 119 struct pcb *pcb2 = &p2->p_addr->u_pcb; 120 121#if NNPX > 0 122 /* Ensure that p1's pcb is up to date. */ 123 if (npxproc == p1) 124 npxsave(&p1->p_addr->u_pcb.pcb_savefpu); 125#endif 126 127 /* Copy p1's pcb. */ 128 p2->p_addr->u_pcb = p1->p_addr->u_pcb; 129 130 /* 131 * Create a new fresh stack for the new process. 132 * Copy the trap frame for the return to user mode as if from a 133 * syscall. This copies the user mode register values. 134 */ 135 p2->p_md.md_regs = (struct trapframe *) 136#ifdef VM86 137 ((int)p2->p_addr + UPAGES * PAGE_SIZE - 16) - 1; 138#else 139 ((int)p2->p_addr + UPAGES * PAGE_SIZE) - 1; 140#endif /* VM86 */ 141 *p2->p_md.md_regs = *p1->p_md.md_regs; 142 143 /* 144 * Set registers for trampoline to user mode. Leave space for the 145 * return address on stack. These are the kernel mode register values. 146 */ 147 pcb2->pcb_cr3 = vtophys(p2->p_vmspace->vm_pmap.pm_pdir); 148 pcb2->pcb_edi = p2->p_md.md_regs->tf_edi; 149 pcb2->pcb_esi = (int)fork_return; 150 pcb2->pcb_ebp = p2->p_md.md_regs->tf_ebp; 151 pcb2->pcb_esp = (int)p2->p_md.md_regs - sizeof(void *); 152 pcb2->pcb_ebx = (int)p2; 153 pcb2->pcb_eip = (int)fork_trampoline; 154 /* 155 * pcb2->pcb_ldt: duplicated below, if necessary. 156 * pcb2->pcb_ldt_len: cloned above. 157 * pcb2->pcb_savefpu: cloned above. 158 * pcb2->pcb_flags: cloned above (always 0 here?). 159 * pcb2->pcb_onfault: cloned above (always NULL here?). 160 */ 161 162#ifdef SMP 163 pcb2->pcb_mpnest = 1; 164#endif 165#ifdef VM86 166 /* 167 * XXX don't copy the i/o pages. this should probably be fixed. 168 */ 169 pcb2->pcb_ext = 0; 170#endif 171 172#ifdef USER_LDT 173 /* Copy the LDT, if necessary. */ 174 if (pcb2->pcb_ldt != 0) { 175 union descriptor *new_ldt; 176 size_t len = pcb2->pcb_ldt_len * sizeof(union descriptor); 177 178 new_ldt = (union descriptor *)kmem_alloc(kernel_map, len); 179 bcopy(pcb2->pcb_ldt, new_ldt, len); 180 pcb2->pcb_ldt = (caddr_t)new_ldt; 181 } 182#endif 183 184 /* 185 * Now, cpu_switch() can schedule the new process. 186 * pcb_esp is loaded pointing to the cpu_switch() stack frame 187 * containing the return address when exiting cpu_switch. 188 * This will normally be to proc_trampoline(), which will have 189 * %ebx loaded with the new proc's pointer. proc_trampoline() 190 * will set up a stack to call fork_return(p, frame); to complete 191 * the return to user-mode. 192 */ 193} 194 195/* 196 * Intercept the return address from a freshly forked process that has NOT 197 * been scheduled yet. 198 * 199 * This is needed to make kernel threads stay in kernel mode. 200 */ 201void 202cpu_set_fork_handler(p, func, arg) 203 struct proc *p; 204 void (*func) __P((const void *)); 205 const void *arg; 206{ 207 /* 208 * Note that the trap frame follows the args, so the function 209 * is really called like this: func(arg, frame); 210 */ 211 p->p_addr->u_pcb.pcb_esi = (int) func; /* function */ 212 p->p_addr->u_pcb.pcb_ebx = (int) arg; /* first arg */ 213} 214 215void 216cpu_exit(p) 217 register struct proc *p; 218{ 219#if defined(USER_LDT) || defined(VM86) 220 struct pcb *pcb = &p->p_addr->u_pcb; 221#endif 222 223#if NNPX > 0 224 npxexit(p); 225#endif /* NNPX */ 226#ifdef VM86 227 if (pcb->pcb_ext != 0) { 228 /* 229 * XXX do we need to move the TSS off the allocated pages 230 * before freeing them? (not done here) 231 */ 232 kmem_free(kernel_map, (vm_offset_t)pcb->pcb_ext, 233 ctob(IOPAGES + 1)); 234 pcb->pcb_ext = 0; 235 } 236#endif 237#ifdef USER_LDT 238 if (pcb->pcb_ldt != 0) { 239 if (pcb == curpcb) { 240 lldt(_default_ldt); 241 currentldt = _default_ldt; 242 } 243 kmem_free(kernel_map, (vm_offset_t)pcb->pcb_ldt, 244 pcb->pcb_ldt_len * sizeof(union descriptor)); 245 pcb->pcb_ldt_len = (int)pcb->pcb_ldt = 0; 246 } 247#endif 248 cnt.v_swtch++; 249 cpu_switch(p); 250 panic("cpu_exit"); 251} 252 253void 254cpu_wait(p) 255 struct proc *p; 256{ 257 /* drop per-process resources */ 258 pmap_dispose_proc(p); 259 260 /* and clean-out the vmspace */ 261 vmspace_free(p->p_vmspace); 262} 263 264/* 265 * Dump the machine specific header information at the start of a core dump. 266 */ 267int 268cpu_coredump(p, vp, cred) 269 struct proc *p; 270 struct vnode *vp; 271 struct ucred *cred; 272{ 273 int error; 274 caddr_t tempuser; 275 276 tempuser = malloc(ctob(UPAGES), M_TEMP, M_WAITOK); 277 if (!tempuser) 278 return EINVAL; 279 280 bzero(tempuser, ctob(UPAGES)); 281 bcopy(p->p_addr, tempuser, sizeof(struct user)); 282 bcopy(p->p_md.md_regs, 283 tempuser + ((caddr_t) p->p_md.md_regs - (caddr_t) p->p_addr), 284 sizeof(struct trapframe)); 285 286 error = vn_rdwr(UIO_WRITE, vp, (caddr_t) tempuser, 287 ctob(UPAGES), 288 (off_t)0, UIO_SYSSPACE, IO_NODELOCKED|IO_UNIT, 289 cred, (int *)NULL, p); 290 291 free(tempuser, M_TEMP); 292 293 return error; 294} 295 296#ifdef notyet 297static void 298setredzone(pte, vaddr) 299 u_short *pte; 300 caddr_t vaddr; 301{ 302/* eventually do this by setting up an expand-down stack segment 303 for ss0: selector, allowing stack access down to top of u. 304 this means though that protection violations need to be handled 305 thru a double fault exception that must do an integral task 306 switch to a known good context, within which a dump can be 307 taken. a sensible scheme might be to save the initial context 308 used by sched (that has physical memory mapped 1:1 at bottom) 309 and take the dump while still in mapped mode */ 310} 311#endif 312 313/* 314 * Convert kernel VA to physical address 315 */ 316u_long 317kvtop(void *addr) 318{ 319 vm_offset_t va; 320 321 va = pmap_kextract((vm_offset_t)addr); 322 if (va == 0) 323 panic("kvtop: zero page frame"); 324 return((int)va); 325} 326 327/* 328 * Map an IO request into kernel virtual address space. 329 * 330 * All requests are (re)mapped into kernel VA space. 331 * Notice that we use b_bufsize for the size of the buffer 332 * to be mapped. b_bcount might be modified by the driver. 333 */ 334void 335vmapbuf(bp) 336 register struct buf *bp; 337{ 338 register caddr_t addr, v, kva; 339 vm_offset_t pa; 340 341 if ((bp->b_flags & B_PHYS) == 0) 342 panic("vmapbuf"); 343 344 for (v = bp->b_saveaddr, addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 345 addr < bp->b_data + bp->b_bufsize; 346 addr += PAGE_SIZE, v += PAGE_SIZE) { 347 /* 348 * Do the vm_fault if needed; do the copy-on-write thing 349 * when reading stuff off device into memory. 350 */ 351 vm_fault_quick(addr, 352 (bp->b_flags&B_READ)?(VM_PROT_READ|VM_PROT_WRITE):VM_PROT_READ); 353 pa = trunc_page(pmap_kextract((vm_offset_t) addr)); 354 if (pa == 0) 355 panic("vmapbuf: page not present"); 356 vm_page_hold(PHYS_TO_VM_PAGE(pa)); 357 pmap_kenter((vm_offset_t) v, pa); 358 } 359 360 kva = bp->b_saveaddr; 361 bp->b_saveaddr = bp->b_data; 362 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 363} 364 365/* 366 * Free the io map PTEs associated with this IO operation. 367 * We also invalidate the TLB entries and restore the original b_addr. 368 */ 369void 370vunmapbuf(bp) 371 register struct buf *bp; 372{ 373 register caddr_t addr; 374 vm_offset_t pa; 375 376 if ((bp->b_flags & B_PHYS) == 0) 377 panic("vunmapbuf"); 378 379 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 380 addr < bp->b_data + bp->b_bufsize; 381 addr += PAGE_SIZE) { 382 pa = trunc_page(pmap_kextract((vm_offset_t) addr)); 383 pmap_kremove((vm_offset_t) addr); 384 vm_page_unhold(PHYS_TO_VM_PAGE(pa)); 385 } 386 387 bp->b_data = bp->b_saveaddr; 388} 389 390/* 391 * Force reset the processor by invalidating the entire address space! 392 */ 393 394#ifdef SMP 395static void 396cpu_reset_proxy() 397{ 398 u_int saved_mp_lock; 399 400 cpu_reset_proxy_active = 1; 401 while (cpu_reset_proxy_active == 1) 402 ; /* Wait for other cpu to disable interupts */ 403 saved_mp_lock = mp_lock; 404 mp_lock = 1; 405 printf("cpu_reset_proxy: Grabbed mp lock for BSP\n"); 406 cpu_reset_proxy_active = 3; 407 while (cpu_reset_proxy_active == 3) 408 ; /* Wait for other cpu to enable interrupts */ 409 stop_cpus((1<<cpu_reset_proxyid)); 410 printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid); 411 DELAY(1000000); 412 cpu_reset_real(); 413} 414#endif 415 416void 417cpu_reset() 418{ 419#ifdef SMP 420 if (smp_active == 0) { 421 cpu_reset_real(); 422 /* NOTREACHED */ 423 } else { 424 425 u_int map; 426 int cnt; 427 printf("cpu_reset called on cpu#%d\n",cpuid); 428 429 map = other_cpus & ~ stopped_cpus; 430 431 if (map != 0) { 432 printf("cpu_reset: Stopping other CPUs\n"); 433 stop_cpus(map); /* Stop all other CPUs */ 434 } 435 436 if (cpuid == 0) { 437 DELAY(1000000); 438 cpu_reset_real(); 439 /* NOTREACHED */ 440 } else { 441 /* We are not BSP (CPU #0) */ 442 443 cpu_reset_proxyid = cpuid; 444 cpustop_restartfunc = cpu_reset_proxy; 445 printf("cpu_reset: Restarting BSP\n"); 446 started_cpus = (1<<0); /* Restart CPU #0 */ 447 448 cnt = 0; 449 while (cpu_reset_proxy_active == 0 && cnt < 10000000) 450 cnt++; /* Wait for BSP to announce restart */ 451 if (cpu_reset_proxy_active == 0) 452 printf("cpu_reset: Failed to restart BSP\n"); 453 __asm __volatile("cli" : : : "memory"); 454 cpu_reset_proxy_active = 2; 455 cnt = 0; 456 while (cpu_reset_proxy_active == 2 && cnt < 10000000) 457 cnt++; /* Do nothing */ 458 if (cpu_reset_proxy_active == 2) { 459 printf("cpu_reset: BSP did not grab mp lock\n"); 460 cpu_reset_real(); /* XXX: Bogus ? */ 461 } 462 cpu_reset_proxy_active = 4; 463 __asm __volatile("sti" : : : "memory"); 464 while (1); 465 /* NOTREACHED */ 466 } 467 } 468#else 469 cpu_reset_real(); 470#endif 471} 472 473static void 474cpu_reset_real() 475{ 476 477#ifdef PC98 478 /* 479 * Attempt to do a CPU reset via CPU reset port. 480 */ 481 disable_intr(); 482 if ((inb(0x35) & 0xa0) != 0xa0) { 483 outb(0x37, 0x0f); /* SHUT0 = 0. */ 484 outb(0x37, 0x0b); /* SHUT1 = 0. */ 485 } 486 outb(0xf0, 0x00); /* Reset. */ 487#else 488 /* 489 * Attempt to do a CPU reset via the keyboard controller, 490 * do not turn of the GateA20, as any machine that fails 491 * to do the reset here would then end up in no man's land. 492 */ 493 494#if !defined(BROKEN_KEYBOARD_RESET) 495 outb(IO_KBD + 4, 0xFE); 496 DELAY(500000); /* wait 0.5 sec to see if that did it */ 497 printf("Keyboard reset did not work, attempting CPU shutdown\n"); 498 DELAY(1000000); /* wait 1 sec for printf to complete */ 499#endif 500#endif /* PC98 */ 501 /* force a shutdown by unmapping entire address space ! */ 502 bzero((caddr_t) PTD, PAGE_SIZE); 503 504 /* "good night, sweet prince .... <THUNK!>" */ 505 invltlb(); 506 /* NOTREACHED */ 507 while(1); 508} 509 510#ifndef VM_STACK 511/* 512 * Grow the user stack to allow for 'sp'. This version grows the stack in 513 * chunks of SGROWSIZ. 514 */ 515int 516grow(p, sp) 517 struct proc *p; 518 u_int sp; 519{ 520 unsigned int nss; 521 caddr_t v; 522 struct vmspace *vm = p->p_vmspace; 523 524 if ((caddr_t)sp <= vm->vm_maxsaddr || sp >= USRSTACK) 525 return (1); 526 527 nss = roundup(USRSTACK - sp, PAGE_SIZE); 528 529 if (nss > p->p_rlimit[RLIMIT_STACK].rlim_cur) 530 return (0); 531 532 if (vm->vm_ssize && roundup(vm->vm_ssize << PAGE_SHIFT, 533 SGROWSIZ) < nss) { 534 int grow_amount; 535 /* 536 * If necessary, grow the VM that the stack occupies 537 * to allow for the rlimit. This allows us to not have 538 * to allocate all of the VM up-front in execve (which 539 * is expensive). 540 * Grow the VM by the amount requested rounded up to 541 * the nearest SGROWSIZ to provide for some hysteresis. 542 */ 543 grow_amount = roundup((nss - (vm->vm_ssize << PAGE_SHIFT)), SGROWSIZ); 544 v = (char *)USRSTACK - roundup(vm->vm_ssize << PAGE_SHIFT, 545 SGROWSIZ) - grow_amount; 546 /* 547 * If there isn't enough room to extend by SGROWSIZ, then 548 * just extend to the maximum size 549 */ 550 if (v < vm->vm_maxsaddr) { 551 v = vm->vm_maxsaddr; 552 grow_amount = MAXSSIZ - (vm->vm_ssize << PAGE_SHIFT); 553 } 554 if ((grow_amount == 0) || (vm_map_find(&vm->vm_map, NULL, 0, (vm_offset_t *)&v, 555 grow_amount, FALSE, VM_PROT_ALL, VM_PROT_ALL, 0) != KERN_SUCCESS)) { 556 return (0); 557 } 558 vm->vm_ssize += grow_amount >> PAGE_SHIFT; 559 } 560 561 return (1); 562} 563#else 564int 565grow_stack(p, sp) 566 struct proc *p; 567 u_int sp; 568{ 569 int rv; 570 571 rv = vm_map_growstack (p, sp); 572 if (rv != KERN_SUCCESS) 573 return (0); 574 575 return (1); 576} 577#endif 578 579 580static int cnt_prezero; 581 582SYSCTL_INT(_vm_stats_misc, OID_AUTO, 583 cnt_prezero, CTLFLAG_RD, &cnt_prezero, 0, ""); 584 585/* 586 * Implement the pre-zeroed page mechanism. 587 * This routine is called from the idle loop. 588 */ 589 590#define ZIDLE_LO(v) ((v) * 2 / 3) 591#define ZIDLE_HI(v) ((v) * 4 / 5) 592 593int 594vm_page_zero_idle() 595{ 596 static int free_rover; 597 static int zero_state; 598 vm_page_t m; 599 int s; 600 601 /* 602 * Attempt to maintain approximately 1/2 of our free pages in a 603 * PG_ZERO'd state. Add some hysteresis to (attempt to) avoid 604 * generally zeroing a page when the system is near steady-state. 605 * Otherwise we might get 'flutter' during disk I/O / IPC or 606 * fast sleeps. We also do not want to be continuously zeroing 607 * pages because doing so may flush our L1 and L2 caches too much. 608 */ 609 610 if (zero_state && vm_page_zero_count >= ZIDLE_LO(cnt.v_free_count)) 611 return(0); 612 if (vm_page_zero_count >= ZIDLE_HI(cnt.v_free_count)) 613 return(0); 614 615#ifdef SMP 616 if (try_mplock()) { 617#endif 618 s = splvm(); 619 __asm __volatile("sti" : : : "memory"); 620 zero_state = 0; 621 m = vm_page_list_find(PQ_FREE, free_rover, FALSE); 622 if (m != NULL && (m->flags & PG_ZERO) == 0) { 623 --(*vm_page_queues[m->queue].lcnt); 624 TAILQ_REMOVE(vm_page_queues[m->queue].pl, m, pageq); 625 m->queue = PQ_NONE; 626 splx(s); 627#if 0 628 rel_mplock(); 629#endif 630 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 631#if 0 632 get_mplock(); 633#endif 634 (void)splvm(); 635 vm_page_flag_set(m, PG_ZERO); 636 m->queue = PQ_FREE + m->pc; 637 ++(*vm_page_queues[m->queue].lcnt); 638 TAILQ_INSERT_TAIL(vm_page_queues[m->queue].pl, m, 639 pageq); 640 ++vm_page_zero_count; 641 ++cnt_prezero; 642 if (vm_page_zero_count >= ZIDLE_HI(cnt.v_free_count)) 643 zero_state = 1; 644 } 645 free_rover = (free_rover + PQ_PRIME3) & PQ_L2_MASK; 646 splx(s); 647 __asm __volatile("cli" : : : "memory"); 648#ifdef SMP 649 rel_mplock(); 650#endif 651 return (1); 652#ifdef SMP 653 } 654#endif 655 return (0); 656} 657 658/* 659 * Software interrupt handler for queued VM system processing. 660 */ 661void 662swi_vm() 663{ 664 if (busdma_swi_pending != 0) 665 busdma_swi(); 666} 667 668/* 669 * Tell whether this address is in some physical memory region. 670 * Currently used by the kernel coredump code in order to avoid 671 * dumping the ``ISA memory hole'' which could cause indefinite hangs, 672 * or other unpredictable behaviour. 673 */ 674 675#include "isa.h" 676 677int 678is_physical_memory(addr) 679 vm_offset_t addr; 680{ 681 682#if NISA > 0 683 /* The ISA ``memory hole''. */ 684 if (addr >= 0xa0000 && addr < 0x100000) 685 return 0; 686#endif 687 688 /* 689 * stuff other tests for known memory-mapped devices (PCI?) 690 * here 691 */ 692 693 return 1; 694} 695