vm_machdep.c revision 71257
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 71257 2001-01-19 13:19:02Z peter $ 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 51#include <sys/param.h> 52#include <sys/systm.h> 53#include <sys/malloc.h> 54#include <sys/proc.h> 55#include <sys/bio.h> 56#include <sys/buf.h> 57#include <sys/vnode.h> 58#include <sys/vmmeter.h> 59#include <sys/kernel.h> 60#include <sys/ktr.h> 61#include <sys/mutex.h> 62#include <sys/sysctl.h> 63#include <sys/unistd.h> 64 65#include <machine/cpu.h> 66#include <machine/md_var.h> 67#ifdef SMP 68#include <machine/smp.h> 69#endif 70#include <machine/pcb.h> 71#include <machine/pcb_ext.h> 72#include <machine/vm86.h> 73 74#include <vm/vm.h> 75#include <vm/vm_param.h> 76#include <sys/lock.h> 77#include <vm/vm_kern.h> 78#include <vm/vm_page.h> 79#include <vm/vm_map.h> 80#include <vm/vm_extern.h> 81 82#include <sys/user.h> 83 84#ifdef PC98 85#include <pc98/pc98/pc98.h> 86#else 87#include <i386/isa/isa.h> 88#endif 89 90static void cpu_reset_real __P((void)); 91#ifdef SMP 92static void cpu_reset_proxy __P((void)); 93static u_int cpu_reset_proxyid; 94static volatile u_int cpu_reset_proxy_active; 95#endif 96extern int _ucodesel, _udatasel; 97 98/* 99 * quick version of vm_fault 100 */ 101int 102vm_fault_quick(v, prot) 103 caddr_t v; 104 int prot; 105{ 106 int r; 107 108 if (prot & VM_PROT_WRITE) 109 r = subyte(v, fubyte(v)); 110 else 111 r = fubyte(v); 112 return(r); 113} 114 115/* 116 * Finish a fork operation, with process p2 nearly set up. 117 * Copy and update the pcb, set up the stack so that the child 118 * ready to run and return to user mode. 119 */ 120void 121cpu_fork(p1, p2, flags) 122 register struct proc *p1, *p2; 123 int flags; 124{ 125 struct pcb *pcb2; 126 127 if ((flags & RFPROC) == 0) { 128#ifdef USER_LDT 129 if ((flags & RFMEM) == 0) { 130 /* unshare user LDT */ 131 struct pcb *pcb1 = &p1->p_addr->u_pcb; 132 struct pcb_ldt *pcb_ldt = pcb1->pcb_ldt; 133 if (pcb_ldt && pcb_ldt->ldt_refcnt > 1) { 134 pcb_ldt = user_ldt_alloc(pcb1,pcb_ldt->ldt_len); 135 user_ldt_free(pcb1); 136 pcb1->pcb_ldt = pcb_ldt; 137 set_user_ldt(pcb1); 138 } 139 } 140#endif 141 return; 142 } 143 144#ifdef DEV_NPX 145 /* Ensure that p1's pcb is up to date. */ 146 if (PCPU_GET(npxproc) == p1) 147 npxsave(&p1->p_addr->u_pcb.pcb_savefpu); 148#endif 149 150 /* Copy p1's pcb. */ 151 p2->p_addr->u_pcb = p1->p_addr->u_pcb; 152 pcb2 = &p2->p_addr->u_pcb; 153 154 /* 155 * Create a new fresh stack for the new process. 156 * Copy the trap frame for the return to user mode as if from a 157 * syscall. This copies the user mode register values. 158 */ 159 p2->p_md.md_regs = (struct trapframe *) 160 ((int)p2->p_addr + UPAGES * PAGE_SIZE - 16) - 1; 161 bcopy(p1->p_md.md_regs, p2->p_md.md_regs, sizeof(*p2->p_md.md_regs)); 162 163 /* 164 * Set registers for trampoline to user mode. Leave space for the 165 * return address on stack. These are the kernel mode register values. 166 */ 167 pcb2->pcb_cr3 = vtophys(vmspace_pmap(p2->p_vmspace)->pm_pdir); 168 pcb2->pcb_edi = 0; 169 pcb2->pcb_esi = (int)fork_return; /* fork_trampoline argument */ 170 pcb2->pcb_ebp = 0; 171 pcb2->pcb_esp = (int)p2->p_md.md_regs - sizeof(void *); 172 pcb2->pcb_ebx = (int)p2; /* fork_trampoline argument */ 173 pcb2->pcb_eip = (int)fork_trampoline; 174 /* 175 * pcb2->pcb_ldt: duplicated below, if necessary. 176 * pcb2->pcb_savefpu: cloned above. 177 * pcb2->pcb_flags: cloned above (always 0 here?). 178 * pcb2->pcb_onfault: cloned above (always NULL here?). 179 */ 180 181 pcb2->pcb_schednest = 0; 182 183 /* 184 * XXX don't copy the i/o pages. this should probably be fixed. 185 */ 186 pcb2->pcb_ext = 0; 187 188#ifdef USER_LDT 189 /* Copy the LDT, if necessary. */ 190 if (pcb2->pcb_ldt != 0) { 191 if (flags & RFMEM) { 192 pcb2->pcb_ldt->ldt_refcnt++; 193 } else { 194 pcb2->pcb_ldt = user_ldt_alloc(pcb2, 195 pcb2->pcb_ldt->ldt_len); 196 } 197 } 198#endif 199 200 /* 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. 208 */ 209} 210 211/* 212 * Intercept the return address from a freshly forked process that has NOT 213 * been scheduled yet. 214 * 215 * This is needed to make kernel threads stay in kernel mode. 216 */ 217void 218cpu_set_fork_handler(p, func, arg) 219 struct proc *p; 220 void (*func) __P((void *)); 221 void *arg; 222{ 223 /* 224 * Note that the trap frame follows the args, so the function 225 * is really called like this: func(arg, frame); 226 */ 227 p->p_addr->u_pcb.pcb_esi = (int) func; /* function */ 228 p->p_addr->u_pcb.pcb_ebx = (int) arg; /* first arg */ 229} 230 231void 232cpu_exit(p) 233 register struct proc *p; 234{ 235 struct pcb *pcb = &p->p_addr->u_pcb; 236 237#ifdef DEV_NPX 238 npxexit(p); 239#endif 240 if (pcb->pcb_ext != 0) { 241 /* 242 * XXX do we need to move the TSS off the allocated pages 243 * before freeing them? (not done here) 244 */ 245 kmem_free(kernel_map, (vm_offset_t)pcb->pcb_ext, 246 ctob(IOPAGES + 1)); 247 pcb->pcb_ext = 0; 248 } 249#ifdef USER_LDT 250 user_ldt_free(pcb); 251#endif 252 if (pcb->pcb_flags & PCB_DBREGS) { 253 /* 254 * disable all hardware breakpoints 255 */ 256 reset_dbregs(); 257 pcb->pcb_flags &= ~PCB_DBREGS; 258 } 259 mtx_enter(&sched_lock, MTX_SPIN); 260 mtx_exit(&Giant, MTX_DEF | MTX_NOSWITCH); 261 mtx_assert(&Giant, MA_NOTOWNED); 262 263 /* 264 * We have to wait until after releasing all locks before 265 * changing p_stat. If we block on a mutex then we will be 266 * back at SRUN when we resume and our parent will never 267 * harvest us. 268 */ 269 p->p_stat = SZOMB; 270 271 mp_fixme("assumption: p_pptr won't change at this time"); 272 wakeup(p->p_pptr); 273 274 cnt.v_swtch++; 275 cpu_throw(); 276 panic("cpu_exit"); 277} 278 279void 280cpu_wait(p) 281 struct proc *p; 282{ 283 /* drop per-process resources */ 284 pmap_dispose_proc(p); 285 286 /* and clean-out the vmspace */ 287 vmspace_free(p->p_vmspace); 288} 289 290/* 291 * Dump the machine specific header information at the start of a core dump. 292 */ 293int 294cpu_coredump(p, vp, cred) 295 struct proc *p; 296 struct vnode *vp; 297 struct ucred *cred; 298{ 299 int error; 300 caddr_t tempuser; 301 302 tempuser = malloc(ctob(UPAGES), M_TEMP, M_WAITOK | M_ZERO); 303 if (!tempuser) 304 return EINVAL; 305 306 bcopy(p->p_addr, tempuser, sizeof(struct user)); 307 bcopy(p->p_md.md_regs, 308 tempuser + ((caddr_t) p->p_md.md_regs - (caddr_t) p->p_addr), 309 sizeof(struct trapframe)); 310 311 error = vn_rdwr(UIO_WRITE, vp, (caddr_t) tempuser, 312 ctob(UPAGES), 313 (off_t)0, UIO_SYSSPACE, IO_NODELOCKED|IO_UNIT, 314 cred, (int *)NULL, p); 315 316 free(tempuser, M_TEMP); 317 318 return error; 319} 320 321#ifdef notyet 322static void 323setredzone(pte, vaddr) 324 u_short *pte; 325 caddr_t vaddr; 326{ 327/* eventually do this by setting up an expand-down stack segment 328 for ss0: selector, allowing stack access down to top of u. 329 this means though that protection violations need to be handled 330 thru a double fault exception that must do an integral task 331 switch to a known good context, within which a dump can be 332 taken. a sensible scheme might be to save the initial context 333 used by sched (that has physical memory mapped 1:1 at bottom) 334 and take the dump while still in mapped mode */ 335} 336#endif 337 338/* 339 * Convert kernel VA to physical address 340 */ 341u_long 342kvtop(void *addr) 343{ 344 vm_offset_t va; 345 346 va = pmap_kextract((vm_offset_t)addr); 347 if (va == 0) 348 panic("kvtop: zero page frame"); 349 return((int)va); 350} 351 352/* 353 * Map an IO request into kernel virtual address space. 354 * 355 * All requests are (re)mapped into kernel VA space. 356 * Notice that we use b_bufsize for the size of the buffer 357 * to be mapped. b_bcount might be modified by the driver. 358 */ 359void 360vmapbuf(bp) 361 register struct buf *bp; 362{ 363 register caddr_t addr, v, kva; 364 vm_offset_t pa; 365 366 if ((bp->b_flags & B_PHYS) == 0) 367 panic("vmapbuf"); 368 369 for (v = bp->b_saveaddr, addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 370 addr < bp->b_data + bp->b_bufsize; 371 addr += PAGE_SIZE, v += PAGE_SIZE) { 372 /* 373 * Do the vm_fault if needed; do the copy-on-write thing 374 * when reading stuff off device into memory. 375 */ 376 vm_fault_quick(addr, 377 (bp->b_iocmd == BIO_READ)?(VM_PROT_READ|VM_PROT_WRITE):VM_PROT_READ); 378 pa = trunc_page(pmap_kextract((vm_offset_t) addr)); 379 if (pa == 0) 380 panic("vmapbuf: page not present"); 381 vm_page_hold(PHYS_TO_VM_PAGE(pa)); 382 pmap_kenter((vm_offset_t) v, pa); 383 } 384 385 kva = bp->b_saveaddr; 386 bp->b_saveaddr = bp->b_data; 387 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 388} 389 390/* 391 * Free the io map PTEs associated with this IO operation. 392 * We also invalidate the TLB entries and restore the original b_addr. 393 */ 394void 395vunmapbuf(bp) 396 register struct buf *bp; 397{ 398 register caddr_t addr; 399 vm_offset_t pa; 400 401 if ((bp->b_flags & B_PHYS) == 0) 402 panic("vunmapbuf"); 403 404 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 405 addr < bp->b_data + bp->b_bufsize; 406 addr += PAGE_SIZE) { 407 pa = trunc_page(pmap_kextract((vm_offset_t) addr)); 408 pmap_kremove((vm_offset_t) addr); 409 vm_page_unhold(PHYS_TO_VM_PAGE(pa)); 410 } 411 412 bp->b_data = bp->b_saveaddr; 413} 414 415/* 416 * Force reset the processor by invalidating the entire address space! 417 */ 418 419#ifdef SMP 420static void 421cpu_reset_proxy() 422{ 423 424 cpu_reset_proxy_active = 1; 425 while (cpu_reset_proxy_active == 1) 426 ; /* Wait for other cpu to see that we've started */ 427 stop_cpus((1<<cpu_reset_proxyid)); 428 printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid); 429 DELAY(1000000); 430 cpu_reset_real(); 431} 432#endif 433 434void 435cpu_reset() 436{ 437#ifdef SMP 438 if (smp_active == 0) { 439 cpu_reset_real(); 440 /* NOTREACHED */ 441 } else { 442 443 u_int map; 444 int cnt; 445 printf("cpu_reset called on cpu#%d\n", PCPU_GET(cpuid)); 446 447 map = PCPU_GET(other_cpus) & ~ stopped_cpus; 448 449 if (map != 0) { 450 printf("cpu_reset: Stopping other CPUs\n"); 451 stop_cpus(map); /* Stop all other CPUs */ 452 } 453 454 if (PCPU_GET(cpuid) == 0) { 455 DELAY(1000000); 456 cpu_reset_real(); 457 /* NOTREACHED */ 458 } else { 459 /* We are not BSP (CPU #0) */ 460 461 cpu_reset_proxyid = PCPU_GET(cpuid); 462 cpustop_restartfunc = cpu_reset_proxy; 463 cpu_reset_proxy_active = 0; 464 printf("cpu_reset: Restarting BSP\n"); 465 started_cpus = (1<<0); /* Restart CPU #0 */ 466 467 cnt = 0; 468 while (cpu_reset_proxy_active == 0 && cnt < 10000000) 469 cnt++; /* Wait for BSP to announce restart */ 470 if (cpu_reset_proxy_active == 0) 471 printf("cpu_reset: Failed to restart BSP\n"); 472 enable_intr(); 473 cpu_reset_proxy_active = 2; 474 475 while (1); 476 /* NOTREACHED */ 477 } 478 } 479#else 480 cpu_reset_real(); 481#endif 482} 483 484static void 485cpu_reset_real() 486{ 487 488#ifdef PC98 489 /* 490 * Attempt to do a CPU reset via CPU reset port. 491 */ 492 disable_intr(); 493 if ((inb(0x35) & 0xa0) != 0xa0) { 494 outb(0x37, 0x0f); /* SHUT0 = 0. */ 495 outb(0x37, 0x0b); /* SHUT1 = 0. */ 496 } 497 outb(0xf0, 0x00); /* Reset. */ 498#else 499 /* 500 * Attempt to do a CPU reset via the keyboard controller, 501 * do not turn of the GateA20, as any machine that fails 502 * to do the reset here would then end up in no man's land. 503 */ 504 505#if !defined(BROKEN_KEYBOARD_RESET) 506 outb(IO_KBD + 4, 0xFE); 507 DELAY(500000); /* wait 0.5 sec to see if that did it */ 508 printf("Keyboard reset did not work, attempting CPU shutdown\n"); 509 DELAY(1000000); /* wait 1 sec for printf to complete */ 510#endif 511#endif /* PC98 */ 512 /* force a shutdown by unmapping entire address space ! */ 513 bzero((caddr_t) PTD, PAGE_SIZE); 514 515 /* "good night, sweet prince .... <THUNK!>" */ 516 invltlb(); 517 /* NOTREACHED */ 518 while(1); 519} 520 521int 522grow_stack(p, sp) 523 struct proc *p; 524 u_int sp; 525{ 526 int rv; 527 528 rv = vm_map_growstack (p, sp); 529 if (rv != KERN_SUCCESS) 530 return (0); 531 532 return (1); 533} 534 535SYSCTL_DECL(_vm_stats_misc); 536 537static int cnt_prezero; 538 539SYSCTL_INT(_vm_stats_misc, OID_AUTO, 540 cnt_prezero, CTLFLAG_RD, &cnt_prezero, 0, ""); 541 542/* 543 * Implement the pre-zeroed page mechanism. 544 * This routine is called from the idle loop. 545 */ 546 547#define ZIDLE_LO(v) ((v) * 2 / 3) 548#define ZIDLE_HI(v) ((v) * 4 / 5) 549 550int 551vm_page_zero_idle() 552{ 553 static int free_rover; 554 static int zero_state; 555 vm_page_t m; 556 int s; 557 558 /* 559 * Attempt to maintain approximately 1/2 of our free pages in a 560 * PG_ZERO'd state. Add some hysteresis to (attempt to) avoid 561 * generally zeroing a page when the system is near steady-state. 562 * Otherwise we might get 'flutter' during disk I/O / IPC or 563 * fast sleeps. We also do not want to be continuously zeroing 564 * pages because doing so may flush our L1 and L2 caches too much. 565 */ 566 567 if (zero_state && vm_page_zero_count >= ZIDLE_LO(cnt.v_free_count)) 568 return(0); 569 if (vm_page_zero_count >= ZIDLE_HI(cnt.v_free_count)) 570 return(0); 571 572 if (mtx_try_enter(&Giant, MTX_DEF)) { 573 s = splvm(); 574 zero_state = 0; 575 m = vm_page_list_find(PQ_FREE, free_rover, FALSE); 576 if (m != NULL && (m->flags & PG_ZERO) == 0) { 577 vm_page_queues[m->queue].lcnt--; 578 TAILQ_REMOVE(&vm_page_queues[m->queue].pl, m, pageq); 579 m->queue = PQ_NONE; 580 splx(s); 581 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 582 (void)splvm(); 583 vm_page_flag_set(m, PG_ZERO); 584 m->queue = PQ_FREE + m->pc; 585 vm_page_queues[m->queue].lcnt++; 586 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, 587 pageq); 588 ++vm_page_zero_count; 589 ++cnt_prezero; 590 if (vm_page_zero_count >= ZIDLE_HI(cnt.v_free_count)) 591 zero_state = 1; 592 } 593 free_rover = (free_rover + PQ_PRIME2) & PQ_L2_MASK; 594 splx(s); 595 mtx_exit(&Giant, MTX_DEF); 596 return (1); 597 } 598 return (0); 599} 600 601/* 602 * Software interrupt handler for queued VM system processing. 603 */ 604void 605swi_vm(void *dummy) 606{ 607 if (busdma_swi_pending != 0) 608 busdma_swi(); 609} 610 611/* 612 * Tell whether this address is in some physical memory region. 613 * Currently used by the kernel coredump code in order to avoid 614 * dumping the ``ISA memory hole'' which could cause indefinite hangs, 615 * or other unpredictable behaviour. 616 */ 617 618#include "isa.h" 619 620int 621is_physical_memory(addr) 622 vm_offset_t addr; 623{ 624 625#if NISA > 0 626 /* The ISA ``memory hole''. */ 627 if (addr >= 0xa0000 && addr < 0x100000) 628 return 0; 629#endif 630 631 /* 632 * stuff other tests for known memory-mapped devices (PCI?) 633 * here 634 */ 635 636 return 1; 637} 638