vm_machdep.c revision 72746
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 * $FreeBSD: head/sys/amd64/amd64/vm_machdep.c 72746 2001-02-20 05:26:15Z jhb $ 42 */ 43 44#include "opt_npx.h" 45#include "opt_user_ldt.h" 46#ifdef PC98 47#include "opt_pc98.h" 48#endif 49#include "opt_reset.h" 50#include "opt_isa.h" 51 52#include <sys/param.h> 53#include <sys/systm.h> 54#include <sys/malloc.h> 55#include <sys/proc.h> 56#include <sys/bio.h> 57#include <sys/buf.h> 58#include <sys/vnode.h> 59#include <sys/vmmeter.h> 60#include <sys/kernel.h> 61#include <sys/ktr.h> 62#include <sys/mutex.h> 63#include <sys/sysctl.h> 64#include <sys/unistd.h> 65 66#include <machine/cpu.h> 67#include <machine/md_var.h> 68#ifdef SMP 69#include <machine/smp.h> 70#endif 71#include <machine/pcb.h> 72#include <machine/pcb_ext.h> 73#include <machine/vm86.h> 74 75#include <vm/vm.h> 76#include <vm/vm_param.h> 77#include <sys/lock.h> 78#include <vm/vm_kern.h> 79#include <vm/vm_page.h> 80#include <vm/vm_map.h> 81#include <vm/vm_extern.h> 82 83#include <sys/user.h> 84 85#ifdef PC98 86#include <pc98/pc98/pc98.h> 87#else 88#include <i386/isa/isa.h> 89#endif 90 91static void cpu_reset_real __P((void)); 92#ifdef SMP 93static void cpu_reset_proxy __P((void)); 94static u_int cpu_reset_proxyid; 95static volatile u_int cpu_reset_proxy_active; 96#endif 97extern int _ucodesel, _udatasel; 98 99/* 100 * quick version of vm_fault 101 */ 102int 103vm_fault_quick(v, prot) 104 caddr_t v; 105 int prot; 106{ 107 int r; 108 109 if (prot & VM_PROT_WRITE) 110 r = subyte(v, fubyte(v)); 111 else 112 r = fubyte(v); 113 return(r); 114} 115 116/* 117 * Finish a fork operation, with process p2 nearly set up. 118 * Copy and update the pcb, set up the stack so that the child 119 * ready to run and return to user mode. 120 */ 121void 122cpu_fork(p1, p2, flags) 123 register struct proc *p1, *p2; 124 int flags; 125{ 126 struct pcb *pcb2; 127 128 if ((flags & RFPROC) == 0) { 129#ifdef USER_LDT 130 if ((flags & RFMEM) == 0) { 131 /* unshare user LDT */ 132 struct pcb *pcb1 = &p1->p_addr->u_pcb; 133 struct pcb_ldt *pcb_ldt = pcb1->pcb_ldt; 134 if (pcb_ldt && pcb_ldt->ldt_refcnt > 1) { 135 pcb_ldt = user_ldt_alloc(pcb1,pcb_ldt->ldt_len); 136 user_ldt_free(pcb1); 137 pcb1->pcb_ldt = pcb_ldt; 138 set_user_ldt(pcb1); 139 } 140 } 141#endif 142 return; 143 } 144 145#ifdef DEV_NPX 146 /* Ensure that p1's pcb is up to date. */ 147 if (PCPU_GET(npxproc) == p1) 148 npxsave(&p1->p_addr->u_pcb.pcb_savefpu); 149#endif 150 151 /* Copy p1's pcb. */ 152 p2->p_addr->u_pcb = p1->p_addr->u_pcb; 153 pcb2 = &p2->p_addr->u_pcb; 154 155 /* 156 * Create a new fresh stack for the new process. 157 * Copy the trap frame for the return to user mode as if from a 158 * syscall. This copies the user mode register values. 159 */ 160 p2->p_md.md_regs = (struct trapframe *) 161 ((int)p2->p_addr + UPAGES * PAGE_SIZE - 16) - 1; 162 bcopy(p1->p_md.md_regs, p2->p_md.md_regs, sizeof(*p2->p_md.md_regs)); 163 164 p2->p_md.md_regs->tf_eax = 0; /* Child returns zero */ 165 p2->p_md.md_regs->tf_eflags &= ~PSL_C; /* success */ 166 p2->p_md.md_regs->tf_edx = 1; 167 168 /* 169 * Set registers for trampoline to user mode. Leave space for the 170 * return address on stack. These are the kernel mode register values. 171 */ 172 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdir); 173 pcb2->pcb_edi = 0; 174 pcb2->pcb_esi = (int)fork_return; /* fork_trampoline argument */ 175 pcb2->pcb_ebp = 0; 176 pcb2->pcb_esp = (int)p2->p_md.md_regs - sizeof(void *); 177 pcb2->pcb_ebx = (int)p2; /* fork_trampoline argument */ 178 pcb2->pcb_eip = (int)fork_trampoline; 179 /* 180 * pcb2->pcb_ldt: duplicated below, if necessary. 181 * pcb2->pcb_savefpu: cloned above. 182 * pcb2->pcb_flags: cloned above (always 0 here?). 183 * pcb2->pcb_onfault: cloned above (always NULL here?). 184 */ 185 186 /* 187 * XXX don't copy the i/o pages. this should probably be fixed. 188 */ 189 pcb2->pcb_ext = 0; 190 191#ifdef USER_LDT 192 /* Copy the LDT, if necessary. */ 193 if (pcb2->pcb_ldt != 0) { 194 if (flags & RFMEM) { 195 pcb2->pcb_ldt->ldt_refcnt++; 196 } else { 197 pcb2->pcb_ldt = user_ldt_alloc(pcb2, 198 pcb2->pcb_ldt->ldt_len); 199 } 200 } 201#endif 202 203 /* 204 * Now, cpu_switch() can schedule the new process. 205 * pcb_esp is loaded pointing to the cpu_switch() stack frame 206 * containing the return address when exiting cpu_switch. 207 * This will normally be to fork_trampoline(), which will have 208 * %ebx loaded with the new proc's pointer. fork_trampoline() 209 * will set up a stack to call fork_return(p, frame); to complete 210 * the return to user-mode. 211 */ 212} 213 214/* 215 * Intercept the return address from a freshly forked process that has NOT 216 * been scheduled yet. 217 * 218 * This is needed to make kernel threads stay in kernel mode. 219 */ 220void 221cpu_set_fork_handler(p, func, arg) 222 struct proc *p; 223 void (*func) __P((void *)); 224 void *arg; 225{ 226 /* 227 * Note that the trap frame follows the args, so the function 228 * is really called like this: func(arg, frame); 229 */ 230 p->p_addr->u_pcb.pcb_esi = (int) func; /* function */ 231 p->p_addr->u_pcb.pcb_ebx = (int) arg; /* first arg */ 232} 233 234void 235cpu_exit(p) 236 register struct proc *p; 237{ 238 struct pcb *pcb = &p->p_addr->u_pcb; 239 240#ifdef DEV_NPX 241 npxexit(p); 242#endif 243 if (pcb->pcb_ext != 0) { 244 /* 245 * XXX do we need to move the TSS off the allocated pages 246 * before freeing them? (not done here) 247 */ 248 kmem_free(kernel_map, (vm_offset_t)pcb->pcb_ext, 249 ctob(IOPAGES + 1)); 250 pcb->pcb_ext = 0; 251 } 252#ifdef USER_LDT 253 user_ldt_free(pcb); 254#endif 255 if (pcb->pcb_flags & PCB_DBREGS) { 256 /* 257 * disable all hardware breakpoints 258 */ 259 reset_dbregs(); 260 pcb->pcb_flags &= ~PCB_DBREGS; 261 } 262 mtx_lock_spin(&sched_lock); 263 mtx_unlock_flags(&Giant, MTX_NOSWITCH); 264 mtx_assert(&Giant, MA_NOTOWNED); 265 266 /* 267 * We have to wait until after releasing all locks before 268 * changing p_stat. If we block on a mutex then we will be 269 * back at SRUN when we resume and our parent will never 270 * harvest us. 271 */ 272 p->p_stat = SZOMB; 273 274 mp_fixme("assumption: p_pptr won't change at this time"); 275 wakeup(p->p_pptr); 276 277 cnt.v_swtch++; 278 cpu_throw(); 279 panic("cpu_exit"); 280} 281 282void 283cpu_wait(p) 284 struct proc *p; 285{ 286 /* drop per-process resources */ 287 pmap_dispose_proc(p); 288 289 /* and clean-out the vmspace */ 290 vmspace_free(p->p_vmspace); 291} 292 293/* 294 * Dump the machine specific header information at the start of a core dump. 295 */ 296int 297cpu_coredump(p, vp, cred) 298 struct proc *p; 299 struct vnode *vp; 300 struct ucred *cred; 301{ 302 int error; 303 caddr_t tempuser; 304 305 tempuser = malloc(ctob(UPAGES), M_TEMP, M_WAITOK | M_ZERO); 306 if (!tempuser) 307 return EINVAL; 308 309 bcopy(p->p_addr, tempuser, sizeof(struct user)); 310 bcopy(p->p_md.md_regs, 311 tempuser + ((caddr_t) p->p_md.md_regs - (caddr_t) p->p_addr), 312 sizeof(struct trapframe)); 313 314 error = vn_rdwr(UIO_WRITE, vp, (caddr_t) tempuser, 315 ctob(UPAGES), 316 (off_t)0, UIO_SYSSPACE, IO_NODELOCKED|IO_UNIT, 317 cred, (int *)NULL, p); 318 319 free(tempuser, M_TEMP); 320 321 return error; 322} 323 324#ifdef notyet 325static void 326setredzone(pte, vaddr) 327 u_short *pte; 328 caddr_t vaddr; 329{ 330/* eventually do this by setting up an expand-down stack segment 331 for ss0: selector, allowing stack access down to top of u. 332 this means though that protection violations need to be handled 333 thru a double fault exception that must do an integral task 334 switch to a known good context, within which a dump can be 335 taken. a sensible scheme might be to save the initial context 336 used by sched (that has physical memory mapped 1:1 at bottom) 337 and take the dump while still in mapped mode */ 338} 339#endif 340 341/* 342 * Convert kernel VA to physical address 343 */ 344u_long 345kvtop(void *addr) 346{ 347 vm_offset_t va; 348 349 va = pmap_kextract((vm_offset_t)addr); 350 if (va == 0) 351 panic("kvtop: zero page frame"); 352 return((int)va); 353} 354 355/* 356 * Map an IO request into kernel virtual address space. 357 * 358 * All requests are (re)mapped into kernel VA space. 359 * Notice that we use b_bufsize for the size of the buffer 360 * to be mapped. b_bcount might be modified by the driver. 361 */ 362void 363vmapbuf(bp) 364 register struct buf *bp; 365{ 366 register caddr_t addr, v, kva; 367 vm_offset_t pa; 368 369 if ((bp->b_flags & B_PHYS) == 0) 370 panic("vmapbuf"); 371 372 for (v = bp->b_saveaddr, addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 373 addr < bp->b_data + bp->b_bufsize; 374 addr += PAGE_SIZE, v += PAGE_SIZE) { 375 /* 376 * Do the vm_fault if needed; do the copy-on-write thing 377 * when reading stuff off device into memory. 378 */ 379 vm_fault_quick(addr, 380 (bp->b_iocmd == BIO_READ)?(VM_PROT_READ|VM_PROT_WRITE):VM_PROT_READ); 381 pa = trunc_page(pmap_kextract((vm_offset_t) addr)); 382 if (pa == 0) 383 panic("vmapbuf: page not present"); 384 vm_page_hold(PHYS_TO_VM_PAGE(pa)); 385 pmap_kenter((vm_offset_t) v, pa); 386 } 387 388 kva = bp->b_saveaddr; 389 bp->b_saveaddr = bp->b_data; 390 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 391} 392 393/* 394 * Free the io map PTEs associated with this IO operation. 395 * We also invalidate the TLB entries and restore the original b_addr. 396 */ 397void 398vunmapbuf(bp) 399 register struct buf *bp; 400{ 401 register caddr_t addr; 402 vm_offset_t pa; 403 404 if ((bp->b_flags & B_PHYS) == 0) 405 panic("vunmapbuf"); 406 407 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 408 addr < bp->b_data + bp->b_bufsize; 409 addr += PAGE_SIZE) { 410 pa = trunc_page(pmap_kextract((vm_offset_t) addr)); 411 pmap_kremove((vm_offset_t) addr); 412 vm_page_unhold(PHYS_TO_VM_PAGE(pa)); 413 } 414 415 bp->b_data = bp->b_saveaddr; 416} 417 418/* 419 * Force reset the processor by invalidating the entire address space! 420 */ 421 422#ifdef SMP 423static void 424cpu_reset_proxy() 425{ 426 427 cpu_reset_proxy_active = 1; 428 while (cpu_reset_proxy_active == 1) 429 ; /* Wait for other cpu to see that we've started */ 430 stop_cpus((1<<cpu_reset_proxyid)); 431 printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid); 432 DELAY(1000000); 433 cpu_reset_real(); 434} 435#endif 436 437void 438cpu_reset() 439{ 440#ifdef SMP 441 if (smp_active == 0) { 442 cpu_reset_real(); 443 /* NOTREACHED */ 444 } else { 445 446 u_int map; 447 int cnt; 448 printf("cpu_reset called on cpu#%d\n", PCPU_GET(cpuid)); 449 450 map = PCPU_GET(other_cpus) & ~ stopped_cpus; 451 452 if (map != 0) { 453 printf("cpu_reset: Stopping other CPUs\n"); 454 stop_cpus(map); /* Stop all other CPUs */ 455 } 456 457 if (PCPU_GET(cpuid) == 0) { 458 DELAY(1000000); 459 cpu_reset_real(); 460 /* NOTREACHED */ 461 } else { 462 /* We are not BSP (CPU #0) */ 463 464 cpu_reset_proxyid = PCPU_GET(cpuid); 465 cpustop_restartfunc = cpu_reset_proxy; 466 cpu_reset_proxy_active = 0; 467 printf("cpu_reset: Restarting BSP\n"); 468 started_cpus = (1<<0); /* Restart CPU #0 */ 469 470 cnt = 0; 471 while (cpu_reset_proxy_active == 0 && cnt < 10000000) 472 cnt++; /* Wait for BSP to announce restart */ 473 if (cpu_reset_proxy_active == 0) 474 printf("cpu_reset: Failed to restart BSP\n"); 475 enable_intr(); 476 cpu_reset_proxy_active = 2; 477 478 while (1); 479 /* NOTREACHED */ 480 } 481 } 482#else 483 cpu_reset_real(); 484#endif 485} 486 487static void 488cpu_reset_real() 489{ 490 491#ifdef PC98 492 /* 493 * Attempt to do a CPU reset via CPU reset port. 494 */ 495 disable_intr(); 496 if ((inb(0x35) & 0xa0) != 0xa0) { 497 outb(0x37, 0x0f); /* SHUT0 = 0. */ 498 outb(0x37, 0x0b); /* SHUT1 = 0. */ 499 } 500 outb(0xf0, 0x00); /* Reset. */ 501#else 502 /* 503 * Attempt to do a CPU reset via the keyboard controller, 504 * do not turn of the GateA20, as any machine that fails 505 * to do the reset here would then end up in no man's land. 506 */ 507 508#if !defined(BROKEN_KEYBOARD_RESET) 509 outb(IO_KBD + 4, 0xFE); 510 DELAY(500000); /* wait 0.5 sec to see if that did it */ 511 printf("Keyboard reset did not work, attempting CPU shutdown\n"); 512 DELAY(1000000); /* wait 1 sec for printf to complete */ 513#endif 514#endif /* PC98 */ 515 /* force a shutdown by unmapping entire address space ! */ 516 bzero((caddr_t) PTD, PAGE_SIZE); 517 518 /* "good night, sweet prince .... <THUNK!>" */ 519 invltlb(); 520 /* NOTREACHED */ 521 while(1); 522} 523 524int 525grow_stack(p, sp) 526 struct proc *p; 527 u_int sp; 528{ 529 int rv; 530 531 rv = vm_map_growstack (p, sp); 532 if (rv != KERN_SUCCESS) 533 return (0); 534 535 return (1); 536} 537 538SYSCTL_DECL(_vm_stats_misc); 539 540static int cnt_prezero; 541 542SYSCTL_INT(_vm_stats_misc, OID_AUTO, 543 cnt_prezero, CTLFLAG_RD, &cnt_prezero, 0, ""); 544 545/* 546 * Implement the pre-zeroed page mechanism. 547 * This routine is called from the idle loop. 548 */ 549 550#define ZIDLE_LO(v) ((v) * 2 / 3) 551#define ZIDLE_HI(v) ((v) * 4 / 5) 552 553int 554vm_page_zero_idle() 555{ 556 static int free_rover; 557 static int zero_state; 558 vm_page_t m; 559 int s; 560 561 /* 562 * Attempt to maintain approximately 1/2 of our free pages in a 563 * PG_ZERO'd state. Add some hysteresis to (attempt to) avoid 564 * generally zeroing a page when the system is near steady-state. 565 * Otherwise we might get 'flutter' during disk I/O / IPC or 566 * fast sleeps. We also do not want to be continuously zeroing 567 * pages because doing so may flush our L1 and L2 caches too much. 568 */ 569 570 if (zero_state && vm_page_zero_count >= ZIDLE_LO(cnt.v_free_count)) 571 return(0); 572 if (vm_page_zero_count >= ZIDLE_HI(cnt.v_free_count)) 573 return(0); 574 575 if (mtx_trylock(&Giant)) { 576 s = splvm(); 577 zero_state = 0; 578 m = vm_page_list_find(PQ_FREE, free_rover, FALSE); 579 if (m != NULL && (m->flags & PG_ZERO) == 0) { 580 vm_page_queues[m->queue].lcnt--; 581 TAILQ_REMOVE(&vm_page_queues[m->queue].pl, m, pageq); 582 m->queue = PQ_NONE; 583 splx(s); 584 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 585 (void)splvm(); 586 vm_page_flag_set(m, PG_ZERO); 587 m->queue = PQ_FREE + m->pc; 588 vm_page_queues[m->queue].lcnt++; 589 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, 590 pageq); 591 ++vm_page_zero_count; 592 ++cnt_prezero; 593 if (vm_page_zero_count >= ZIDLE_HI(cnt.v_free_count)) 594 zero_state = 1; 595 } 596 free_rover = (free_rover + PQ_PRIME2) & PQ_L2_MASK; 597 splx(s); 598 mtx_unlock(&Giant); 599 return (1); 600 } 601 return (0); 602} 603 604/* 605 * Software interrupt handler for queued VM system processing. 606 */ 607void 608swi_vm(void *dummy) 609{ 610 if (busdma_swi_pending != 0) 611 busdma_swi(); 612} 613 614/* 615 * Tell whether this address is in some physical memory region. 616 * Currently used by the kernel coredump code in order to avoid 617 * dumping the ``ISA memory hole'' which could cause indefinite hangs, 618 * or other unpredictable behaviour. 619 */ 620 621int 622is_physical_memory(addr) 623 vm_offset_t addr; 624{ 625 626#ifdef DEV_ISA 627 /* The ISA ``memory hole''. */ 628 if (addr >= 0xa0000 && addr < 0x100000) 629 return 0; 630#endif 631 632 /* 633 * stuff other tests for known memory-mapped devices (PCI?) 634 * here 635 */ 636 637 return 1; 638} 639