1 /* 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * %sccs.include.redist.c% 9 * 10 * @(#)vm_glue.c 7.21 (Berkeley) 04/28/93 11 * 12 * 13 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 14 * All rights reserved. 15 * 16 * Permission to use, copy, modify and distribute this software and 17 * its documentation is hereby granted, provided that both the copyright 18 * notice and this permission notice appear in all copies of the 19 * software, derivative works or modified versions, and any portions 20 * thereof, and that both notices appear in supporting documentation. 21 * 22 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 23 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 24 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 25 * 26 * Carnegie Mellon requests users of this software to return to 27 * 28 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 29 * School of Computer Science 30 * Carnegie Mellon University 31 * Pittsburgh PA 15213-3890 32 * 33 * any improvements or extensions that they make and grant Carnegie the 34 * rights to redistribute these changes. 35 */ 36 37 #include <sys/param.h> 38 #include <sys/systm.h> 39 #include <sys/proc.h> 40 #include <sys/resourcevar.h> 41 #include <sys/buf.h> 42 #include <sys/user.h> 43 44 #include <vm/vm.h> 45 #include <vm/vm_page.h> 46 #include <vm/vm_kern.h> 47 48 int avefree = 0; /* XXX */ 49 unsigned maxdmap = MAXDSIZ; /* XXX */ 50 int readbuffers = 0; /* XXX allow kgdb to read kernel buffer pool */ 51 52 int 53 kernacc(addr, len, rw) 54 caddr_t addr; 55 int len, rw; 56 { 57 boolean_t rv; 58 vm_offset_t saddr, eaddr; 59 vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE; 60 61 saddr = trunc_page(addr); 62 eaddr = round_page(addr+len-1); 63 rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot); 64 /* 65 * XXX there are still some things (e.g. the buffer cache) that 66 * are managed behind the VM system's back so even though an 67 * address is accessible in the mind of the VM system, there may 68 * not be physical pages where the VM thinks there is. This can 69 * lead to bogus allocation of pages in the kernel address space 70 * or worse, inconsistencies at the pmap level. We only worry 71 * about the buffer cache for now. 72 */ 73 if (!readbuffers && rv && (eaddr > (vm_offset_t)buffers && 74 saddr < (vm_offset_t)buffers + MAXBSIZE * nbuf)) 75 rv = FALSE; 76 return(rv == TRUE); 77 } 78 79 int 80 useracc(addr, len, rw) 81 caddr_t addr; 82 int len, rw; 83 { 84 boolean_t rv; 85 vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE; 86 87 rv = vm_map_check_protection(&curproc->p_vmspace->vm_map, 88 trunc_page(addr), round_page(addr+len-1), prot); 89 return(rv == TRUE); 90 } 91 92 #ifdef KGDB 93 /* 94 * Change protections on kernel pages from addr to addr+len 95 * (presumably so debugger can plant a breakpoint). 96 * 97 * We force the protection change at the pmap level. If we were 98 * to use vm_map_protect a change to allow writing would be lazily- 99 * applied meaning we would still take a protection fault, something 100 * we really don't want to do. It would also fragment the kernel 101 * map unnecessarily. We cannot use pmap_protect since it also won't 102 * enforce a write-enable request. Using pmap_enter is the only way 103 * we can ensure the change takes place properly. 104 */ 105 void 106 chgkprot(addr, len, rw) 107 register caddr_t addr; 108 int len, rw; 109 { 110 vm_prot_t prot; 111 vm_offset_t pa, sva, eva; 112 113 prot = rw == B_READ ? VM_PROT_READ : VM_PROT_READ|VM_PROT_WRITE; 114 eva = round_page(addr + len - 1); 115 for (sva = trunc_page(addr); sva < eva; sva += PAGE_SIZE) { 116 /* 117 * Extract physical address for the page. 118 * We use a cheezy hack to differentiate physical 119 * page 0 from an invalid mapping, not that it 120 * really matters... 121 */ 122 pa = pmap_extract(kernel_pmap, sva|1); 123 if (pa == 0) 124 panic("chgkprot: invalid page"); 125 pmap_enter(kernel_pmap, sva, pa&~1, prot, TRUE); 126 } 127 } 128 #endif 129 130 void 131 vslock(addr, len) 132 caddr_t addr; 133 u_int len; 134 { 135 vm_map_pageable(&curproc->p_vmspace->vm_map, trunc_page(addr), 136 round_page(addr+len-1), FALSE); 137 } 138 139 void 140 vsunlock(addr, len, dirtied) 141 caddr_t addr; 142 u_int len; 143 int dirtied; 144 { 145 #ifdef lint 146 dirtied++; 147 #endif lint 148 vm_map_pageable(&curproc->p_vmspace->vm_map, trunc_page(addr), 149 round_page(addr+len-1), TRUE); 150 } 151 152 /* 153 * Implement fork's actions on an address space. 154 * Here we arrange for the address space to be copied or referenced, 155 * allocate a user struct (pcb and kernel stack), then call the 156 * machine-dependent layer to fill those in and make the new process 157 * ready to run. 158 * NOTE: the kernel stack may be at a different location in the child 159 * process, and thus addresses of automatic variables may be invalid 160 * after cpu_fork returns in the child process. We do nothing here 161 * after cpu_fork returns. 162 */ 163 int 164 vm_fork(p1, p2, isvfork) 165 register struct proc *p1, *p2; 166 int isvfork; 167 { 168 register struct user *up; 169 vm_offset_t addr; 170 171 #ifdef i386 172 /* 173 * avoid copying any of the parent's pagetables or other per-process 174 * objects that reside in the map by marking all of them non-inheritable 175 */ 176 (void)vm_map_inherit(&p1->p_vmspace->vm_map, 177 UPT_MIN_ADDRESS-UPAGES*NBPG, VM_MAX_ADDRESS, VM_INHERIT_NONE); 178 #endif 179 p2->p_vmspace = vmspace_fork(p1->p_vmspace); 180 181 #ifdef SYSVSHM 182 if (p1->p_vmspace->vm_shm) 183 shmfork(p1, p2, isvfork); 184 #endif 185 186 #ifndef i386 187 /* 188 * Allocate a wired-down (for now) pcb and kernel stack for the process 189 */ 190 addr = kmem_alloc_pageable(kernel_map, ctob(UPAGES)); 191 vm_map_pageable(kernel_map, addr, addr + ctob(UPAGES), FALSE); 192 #else 193 /* XXX somehow, on 386, ocassionally pageout removes active, wired down kstack, 194 and pagetables, WITHOUT going thru vm_page_unwire! Why this appears to work is 195 not yet clear, yet it does... */ 196 addr = kmem_alloc(kernel_map, ctob(UPAGES)); 197 #endif 198 up = (struct user *)addr; 199 p2->p_addr = up; 200 201 /* 202 * p_stats and p_sigacts currently point at fields 203 * in the user struct but not at &u, instead at p_addr. 204 * Copy p_sigacts and parts of p_stats; zero the rest 205 * of p_stats (statistics). 206 */ 207 p2->p_stats = &up->u_stats; 208 p2->p_sigacts = &up->u_sigacts; 209 up->u_sigacts = *p1->p_sigacts; 210 bzero(&up->u_stats.pstat_startzero, 211 (unsigned) ((caddr_t)&up->u_stats.pstat_endzero - 212 (caddr_t)&up->u_stats.pstat_startzero)); 213 bcopy(&p1->p_stats->pstat_startcopy, &up->u_stats.pstat_startcopy, 214 ((caddr_t)&up->u_stats.pstat_endcopy - 215 (caddr_t)&up->u_stats.pstat_startcopy)); 216 217 #ifdef i386 218 { u_int addr = UPT_MIN_ADDRESS - UPAGES*NBPG; struct vm_map *vp; 219 220 vp = &p2->p_vmspace->vm_map; 221 (void)vm_deallocate(vp, addr, UPT_MAX_ADDRESS - addr); 222 (void)vm_allocate(vp, &addr, UPT_MAX_ADDRESS - addr, FALSE); 223 (void)vm_map_inherit(vp, addr, UPT_MAX_ADDRESS, VM_INHERIT_NONE); 224 } 225 #endif 226 /* 227 * cpu_fork will copy and update the kernel stack and pcb, 228 * and make the child ready to run. It marks the child 229 * so that it can return differently than the parent. 230 * It returns twice, once in the parent process and 231 * once in the child. 232 */ 233 return (cpu_fork(p1, p2)); 234 } 235 236 /* 237 * Set default limits for VM system. 238 * Called for proc 0, and then inherited by all others. 239 */ 240 void 241 vm_init_limits(p) 242 register struct proc *p; 243 { 244 245 /* 246 * Set up the initial limits on process VM. 247 * Set the maximum resident set size to be all 248 * of (reasonably) available memory. This causes 249 * any single, large process to start random page 250 * replacement once it fills memory. 251 */ 252 p->p_rlimit[RLIMIT_STACK].rlim_cur = DFLSSIZ; 253 p->p_rlimit[RLIMIT_STACK].rlim_max = MAXSSIZ; 254 p->p_rlimit[RLIMIT_DATA].rlim_cur = DFLDSIZ; 255 p->p_rlimit[RLIMIT_DATA].rlim_max = MAXDSIZ; 256 p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(cnt.v_free_count); 257 } 258 259 #include <vm/vm_pageout.h> 260 261 #ifdef DEBUG 262 int enableswap = 1; 263 int swapdebug = 0; 264 #define SDB_FOLLOW 1 265 #define SDB_SWAPIN 2 266 #define SDB_SWAPOUT 4 267 #endif 268 269 /* 270 * Brutally simple: 271 * 1. Attempt to swapin every swaped-out, runnable process in 272 * order of priority. 273 * 2. If not enough memory, wake the pageout daemon and let it 274 * clear some space. 275 */ 276 void 277 sched() 278 { 279 register struct proc *p; 280 register int pri; 281 struct proc *pp; 282 int ppri; 283 vm_offset_t addr; 284 vm_size_t size; 285 286 loop: 287 #ifdef DEBUG 288 while (!enableswap) 289 sleep((caddr_t)&proc0, PVM); 290 #endif 291 pp = NULL; 292 ppri = INT_MIN; 293 for (p = (struct proc *)allproc; p != NULL; p = p->p_nxt) { 294 if (p->p_stat == SRUN && (p->p_flag & SLOAD) == 0) { 295 pri = p->p_time + p->p_slptime - p->p_nice * 8; 296 if (pri > ppri) { 297 pp = p; 298 ppri = pri; 299 } 300 } 301 } 302 #ifdef DEBUG 303 if (swapdebug & SDB_FOLLOW) 304 printf("sched: running, procp %x pri %d\n", pp, ppri); 305 #endif 306 /* 307 * Nothing to do, back to sleep 308 */ 309 if ((p = pp) == NULL) { 310 sleep((caddr_t)&proc0, PVM); 311 goto loop; 312 } 313 314 /* 315 * We would like to bring someone in. 316 * This part is really bogus cuz we could deadlock on memory 317 * despite our feeble check. 318 */ 319 size = round_page(ctob(UPAGES)); 320 addr = (vm_offset_t) p->p_addr; 321 if (cnt.v_free_count > atop(size)) { 322 #ifdef DEBUG 323 if (swapdebug & SDB_SWAPIN) 324 printf("swapin: pid %d(%s)@%x, pri %d free %d\n", 325 p->p_pid, p->p_comm, p->p_addr, 326 ppri, cnt.v_free_count); 327 #endif 328 vm_map_pageable(kernel_map, addr, addr+size, FALSE); 329 (void) splstatclock(); 330 if (p->p_stat == SRUN) 331 setrq(p); 332 p->p_flag |= SLOAD; 333 (void) spl0(); 334 p->p_time = 0; 335 goto loop; 336 } 337 /* 338 * Not enough memory, jab the pageout daemon and wait til the 339 * coast is clear. 340 */ 341 #ifdef DEBUG 342 if (swapdebug & SDB_FOLLOW) 343 printf("sched: no room for pid %d(%s), free %d\n", 344 p->p_pid, p->p_comm, cnt.v_free_count); 345 #endif 346 (void) splhigh(); 347 VM_WAIT; 348 (void) spl0(); 349 #ifdef DEBUG 350 if (swapdebug & SDB_FOLLOW) 351 printf("sched: room again, free %d\n", cnt.v_free_count); 352 #endif 353 goto loop; 354 } 355 356 #define swappable(p) \ 357 (((p)->p_flag & (SSYS|SLOAD|SKEEP|SWEXIT|SPHYSIO)) == SLOAD) 358 359 /* 360 * Swapout is driven by the pageout daemon. Very simple, we find eligible 361 * procs and unwire their u-areas. We try to always "swap" at least one 362 * process in case we need the room for a swapin. 363 * If any procs have been sleeping/stopped for at least maxslp seconds, 364 * they are swapped. Else, we swap the longest-sleeping or stopped process, 365 * if any, otherwise the longest-resident process. 366 */ 367 void 368 swapout_threads() 369 { 370 register struct proc *p; 371 struct proc *outp, *outp2; 372 int outpri, outpri2; 373 int didswap = 0; 374 extern int maxslp; 375 376 #ifdef DEBUG 377 if (!enableswap) 378 return; 379 #endif 380 outp = outp2 = NULL; 381 outpri = outpri2 = 0; 382 for (p = (struct proc *)allproc; p != NULL; p = p->p_nxt) { 383 if (!swappable(p)) 384 continue; 385 switch (p->p_stat) { 386 case SRUN: 387 if (p->p_time > outpri2) { 388 outp2 = p; 389 outpri2 = p->p_time; 390 } 391 continue; 392 393 case SSLEEP: 394 case SSTOP: 395 if (p->p_slptime > maxslp) { 396 swapout(p); 397 didswap++; 398 } else if (p->p_slptime > outpri) { 399 outp = p; 400 outpri = p->p_slptime; 401 } 402 continue; 403 } 404 } 405 /* 406 * If we didn't get rid of any real duds, toss out the next most 407 * likely sleeping/stopped or running candidate. We only do this 408 * if we are real low on memory since we don't gain much by doing 409 * it (UPAGES pages). 410 */ 411 if (didswap == 0 && 412 cnt.v_free_count <= atop(round_page(ctob(UPAGES)))) { 413 if ((p = outp) == 0) 414 p = outp2; 415 #ifdef DEBUG 416 if (swapdebug & SDB_SWAPOUT) 417 printf("swapout_threads: no duds, try procp %x\n", p); 418 #endif 419 if (p) 420 swapout(p); 421 } 422 } 423 424 void 425 swapout(p) 426 register struct proc *p; 427 { 428 vm_offset_t addr; 429 vm_size_t size; 430 431 #ifdef DEBUG 432 if (swapdebug & SDB_SWAPOUT) 433 printf("swapout: pid %d(%s)@%x, stat %x pri %d free %d\n", 434 p->p_pid, p->p_comm, p->p_addr, p->p_stat, 435 p->p_slptime, cnt.v_free_count); 436 #endif 437 size = round_page(ctob(UPAGES)); 438 addr = (vm_offset_t) p->p_addr; 439 #if defined(hp300) || defined(luna68k) 440 /* 441 * Ugh! u-area is double mapped to a fixed address behind the 442 * back of the VM system and accesses are usually through that 443 * address rather than the per-process address. Hence reference 444 * and modify information are recorded at the fixed address and 445 * lost at context switch time. We assume the u-struct and 446 * kernel stack are always accessed/modified and force it to be so. 447 */ 448 { 449 register int i; 450 volatile long tmp; 451 452 for (i = 0; i < UPAGES; i++) { 453 tmp = *(long *)addr; *(long *)addr = tmp; 454 addr += NBPG; 455 } 456 addr = (vm_offset_t) p->p_addr; 457 } 458 #endif 459 #ifdef mips 460 /* 461 * Be sure to save the floating point coprocessor state before 462 * paging out the u-struct. 463 */ 464 { 465 extern struct proc *machFPCurProcPtr; 466 467 if (p == machFPCurProcPtr) { 468 MachSaveCurFPState(p); 469 machFPCurProcPtr = (struct proc *)0; 470 } 471 } 472 #endif 473 #ifndef i386 /* temporary measure till we find spontaineous unwire of kstack */ 474 vm_map_pageable(kernel_map, addr, addr+size, TRUE); 475 pmap_collect(vm_map_pmap(&p->p_vmspace->vm_map)); 476 #endif 477 (void) splhigh(); 478 p->p_flag &= ~SLOAD; 479 if (p->p_stat == SRUN) 480 remrq(p); 481 (void) spl0(); 482 p->p_time = 0; 483 } 484 485 /* 486 * The rest of these routines fake thread handling 487 */ 488 489 void 490 assert_wait(event, ruptible) 491 int event; 492 boolean_t ruptible; 493 { 494 #ifdef lint 495 ruptible++; 496 #endif 497 curproc->p_thread = event; 498 } 499 500 void 501 thread_block() 502 { 503 int s = splhigh(); 504 505 if (curproc->p_thread) 506 sleep((caddr_t)curproc->p_thread, PVM); 507 splx(s); 508 } 509 510 void 511 thread_sleep(event, lock, ruptible) 512 int event; 513 simple_lock_t lock; 514 boolean_t ruptible; 515 { 516 #ifdef lint 517 ruptible++; 518 #endif 519 int s = splhigh(); 520 521 curproc->p_thread = event; 522 simple_unlock(lock); 523 if (curproc->p_thread) 524 sleep((caddr_t)event, PVM); 525 splx(s); 526 } 527 528 void 529 thread_wakeup(event) 530 int event; 531 { 532 int s = splhigh(); 533 534 wakeup((caddr_t)event); 535 splx(s); 536 } 537 538 /* 539 * DEBUG stuff 540 */ 541 542 int indent = 0; 543 544 #include <machine/stdarg.h> /* see subr_prf.c */ 545 546 /*ARGSUSED2*/ 547 void 548 #if __STDC__ 549 iprintf(const char *fmt, ...) 550 #else 551 iprintf(fmt /* , va_alist */) 552 char *fmt; 553 /* va_dcl */ 554 #endif 555 { 556 register int i; 557 va_list ap; 558 559 for (i = indent; i >= 8; i -= 8) 560 printf("\t"); 561 while (--i >= 0) 562 printf(" "); 563 va_start(ap, fmt); 564 printf("%r", fmt, ap); 565 va_end(ap); 566 } 567