1 /* $NetBSD: uvm_glue.c,v 1.46 2001/04/21 17:38:24 thorpej Exp $ */ 2 3 /* 4 * Copyright (c) 1997 Charles D. Cranor and Washington University. 5 * Copyright (c) 1991, 1993, The Regents of the University of California. 6 * 7 * All rights reserved. 8 * 9 * This code is derived from software contributed to Berkeley by 10 * The Mach Operating System project at Carnegie-Mellon University. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. All advertising materials mentioning features or use of this software 21 * must display the following acknowledgement: 22 * This product includes software developed by Charles D. Cranor, 23 * Washington University, the University of California, Berkeley and 24 * its contributors. 25 * 4. Neither the name of the University nor the names of its contributors 26 * may be used to endorse or promote products derived from this software 27 * without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 39 * SUCH DAMAGE. 40 * 41 * @(#)vm_glue.c 8.6 (Berkeley) 1/5/94 42 * from: Id: uvm_glue.c,v 1.1.2.8 1998/02/07 01:16:54 chs Exp 43 * 44 * 45 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 46 * All rights reserved. 47 * 48 * Permission to use, copy, modify and distribute this software and 49 * its documentation is hereby granted, provided that both the copyright 50 * notice and this permission notice appear in all copies of the 51 * software, derivative works or modified versions, and any portions 52 * thereof, and that both notices appear in supporting documentation. 53 * 54 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 55 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 56 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 57 * 58 * Carnegie Mellon requests users of this software to return to 59 * 60 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 61 * School of Computer Science 62 * Carnegie Mellon University 63 * Pittsburgh PA 15213-3890 64 * 65 * any improvements or extensions that they make and grant Carnegie the 66 * rights to redistribute these changes. 67 */ 68 69 #include "opt_uvmhist.h" 70 #include "opt_sysv.h" 71 72 /* 73 * uvm_glue.c: glue functions 74 */ 75 76 #include <sys/param.h> 77 #include <sys/systm.h> 78 #include <sys/proc.h> 79 #include <sys/resourcevar.h> 80 #include <sys/buf.h> 81 #include <sys/user.h> 82 #ifdef SYSVSHM 83 #include <sys/shm.h> 84 #endif 85 86 #include <uvm/uvm.h> 87 88 #include <machine/cpu.h> 89 90 /* 91 * local prototypes 92 */ 93 94 static void uvm_swapout __P((struct proc *)); 95 96 /* 97 * XXXCDC: do these really belong here? 98 */ 99 100 int readbuffers = 0; /* allow KGDB to read kern buffer pool */ 101 /* XXX: see uvm_kernacc */ 102 103 104 /* 105 * uvm_kernacc: can the kernel access a region of memory 106 * 107 * - called from malloc [DIAGNOSTIC], and /dev/kmem driver (mem.c) 108 */ 109 110 boolean_t 111 uvm_kernacc(addr, len, rw) 112 caddr_t addr; 113 size_t len; 114 int rw; 115 { 116 boolean_t rv; 117 vaddr_t saddr, eaddr; 118 vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE; 119 120 saddr = trunc_page((vaddr_t)addr); 121 eaddr = round_page((vaddr_t)addr + len); 122 vm_map_lock_read(kernel_map); 123 rv = uvm_map_checkprot(kernel_map, saddr, eaddr, prot); 124 vm_map_unlock_read(kernel_map); 125 126 /* 127 * XXX there are still some things (e.g. the buffer cache) that 128 * are managed behind the VM system's back so even though an 129 * address is accessible in the mind of the VM system, there may 130 * not be physical pages where the VM thinks there is. This can 131 * lead to bogus allocation of pages in the kernel address space 132 * or worse, inconsistencies at the pmap level. We only worry 133 * about the buffer cache for now. 134 */ 135 if (!readbuffers && rv && (eaddr > (vaddr_t)buffers && 136 saddr < (vaddr_t)buffers + MAXBSIZE * nbuf)) 137 rv = FALSE; 138 return(rv); 139 } 140 141 /* 142 * uvm_useracc: can the user access it? 143 * 144 * - called from physio() and sys___sysctl(). 145 */ 146 147 boolean_t 148 uvm_useracc(addr, len, rw) 149 caddr_t addr; 150 size_t len; 151 int rw; 152 { 153 vm_map_t map; 154 boolean_t rv; 155 vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE; 156 157 /* XXX curproc */ 158 map = &curproc->p_vmspace->vm_map; 159 160 vm_map_lock_read(map); 161 rv = uvm_map_checkprot(map, trunc_page((vaddr_t)addr), 162 round_page((vaddr_t)addr + len), prot); 163 vm_map_unlock_read(map); 164 165 return(rv); 166 } 167 168 #ifdef KGDB 169 /* 170 * Change protections on kernel pages from addr to addr+len 171 * (presumably so debugger can plant a breakpoint). 172 * 173 * We force the protection change at the pmap level. If we were 174 * to use vm_map_protect a change to allow writing would be lazily- 175 * applied meaning we would still take a protection fault, something 176 * we really don't want to do. It would also fragment the kernel 177 * map unnecessarily. We cannot use pmap_protect since it also won't 178 * enforce a write-enable request. Using pmap_enter is the only way 179 * we can ensure the change takes place properly. 180 */ 181 void 182 uvm_chgkprot(addr, len, rw) 183 caddr_t addr; 184 size_t len; 185 int rw; 186 { 187 vm_prot_t prot; 188 paddr_t pa; 189 vaddr_t sva, eva; 190 191 prot = rw == B_READ ? VM_PROT_READ : VM_PROT_READ|VM_PROT_WRITE; 192 eva = round_page((vaddr_t)addr + len); 193 for (sva = trunc_page((vaddr_t)addr); sva < eva; sva += PAGE_SIZE) { 194 /* 195 * Extract physical address for the page. 196 * We use a cheezy hack to differentiate physical 197 * page 0 from an invalid mapping, not that it 198 * really matters... 199 */ 200 if (pmap_extract(pmap_kernel(), sva, &pa) == FALSE) 201 panic("chgkprot: invalid page"); 202 pmap_enter(pmap_kernel(), sva, pa, prot, PMAP_WIRED); 203 } 204 } 205 #endif 206 207 /* 208 * vslock: wire user memory for I/O 209 * 210 * - called from physio and sys___sysctl 211 * - XXXCDC: consider nuking this (or making it a macro?) 212 */ 213 214 int 215 uvm_vslock(p, addr, len, access_type) 216 struct proc *p; 217 caddr_t addr; 218 size_t len; 219 vm_prot_t access_type; 220 { 221 vm_map_t map; 222 vaddr_t start, end; 223 int error; 224 225 map = &p->p_vmspace->vm_map; 226 start = trunc_page((vaddr_t)addr); 227 end = round_page((vaddr_t)addr + len); 228 error = uvm_fault_wire(map, start, end, access_type); 229 return error; 230 } 231 232 /* 233 * vslock: wire user memory for I/O 234 * 235 * - called from physio and sys___sysctl 236 * - XXXCDC: consider nuking this (or making it a macro?) 237 */ 238 239 void 240 uvm_vsunlock(p, addr, len) 241 struct proc *p; 242 caddr_t addr; 243 size_t len; 244 { 245 uvm_fault_unwire(&p->p_vmspace->vm_map, trunc_page((vaddr_t)addr), 246 round_page((vaddr_t)addr + len)); 247 } 248 249 /* 250 * uvm_fork: fork a virtual address space 251 * 252 * - the address space is copied as per parent map's inherit values 253 * - a new "user" structure is allocated for the child process 254 * [filled in by MD layer...] 255 * - if specified, the child gets a new user stack described by 256 * stack and stacksize 257 * - NOTE: the kernel stack may be at a different location in the child 258 * process, and thus addresses of automatic variables may be invalid 259 * after cpu_fork returns in the child process. We do nothing here 260 * after cpu_fork returns. 261 * - XXXCDC: we need a way for this to return a failure value rather 262 * than just hang 263 */ 264 void 265 uvm_fork(p1, p2, shared, stack, stacksize, func, arg) 266 struct proc *p1, *p2; 267 boolean_t shared; 268 void *stack; 269 size_t stacksize; 270 void (*func) __P((void *)); 271 void *arg; 272 { 273 struct user *up = p2->p_addr; 274 int error; 275 276 if (shared == TRUE) { 277 p2->p_vmspace = NULL; 278 uvmspace_share(p1, p2); /* share vmspace */ 279 } else 280 p2->p_vmspace = uvmspace_fork(p1->p_vmspace); /* fork vmspace */ 281 282 /* 283 * Wire down the U-area for the process, which contains the PCB 284 * and the kernel stack. Wired state is stored in p->p_flag's 285 * P_INMEM bit rather than in the vm_map_entry's wired count 286 * to prevent kernel_map fragmentation. 287 * 288 * Note the kernel stack gets read/write accesses right off 289 * the bat. 290 */ 291 error = uvm_fault_wire(kernel_map, (vaddr_t)up, 292 (vaddr_t)up + USPACE, VM_PROT_READ | VM_PROT_WRITE); 293 if (error) 294 panic("uvm_fork: uvm_fault_wire failed: %d", error); 295 296 /* 297 * p_stats currently points at a field in the user struct. Copy 298 * parts of p_stats, and zero out the rest. 299 */ 300 p2->p_stats = &up->u_stats; 301 memset(&up->u_stats.pstat_startzero, 0, 302 ((caddr_t)&up->u_stats.pstat_endzero - 303 (caddr_t)&up->u_stats.pstat_startzero)); 304 memcpy(&up->u_stats.pstat_startcopy, &p1->p_stats->pstat_startcopy, 305 ((caddr_t)&up->u_stats.pstat_endcopy - 306 (caddr_t)&up->u_stats.pstat_startcopy)); 307 308 /* 309 * cpu_fork() copy and update the pcb, and make the child ready 310 * to run. If this is a normal user fork, the child will exit 311 * directly to user mode via child_return() on its first time 312 * slice and will not return here. If this is a kernel thread, 313 * the specified entry point will be executed. 314 */ 315 cpu_fork(p1, p2, stack, stacksize, func, arg); 316 } 317 318 /* 319 * uvm_exit: exit a virtual address space 320 * 321 * - the process passed to us is a dead (pre-zombie) process; we 322 * are running on a different context now (the reaper). 323 * - we must run in a separate thread because freeing the vmspace 324 * of the dead process may block. 325 */ 326 void 327 uvm_exit(p) 328 struct proc *p; 329 { 330 vaddr_t va = (vaddr_t)p->p_addr; 331 332 uvmspace_free(p->p_vmspace); 333 p->p_flag &= ~P_INMEM; 334 uvm_fault_unwire(kernel_map, va, va + USPACE); 335 uvm_km_free(kernel_map, va, USPACE); 336 p->p_addr = NULL; 337 } 338 339 /* 340 * uvm_init_limit: init per-process VM limits 341 * 342 * - called for process 0 and then inherited by all others. 343 */ 344 void 345 uvm_init_limits(p) 346 struct proc *p; 347 { 348 349 /* 350 * Set up the initial limits on process VM. Set the maximum 351 * resident set size to be all of (reasonably) available memory. 352 * This causes any single, large process to start random page 353 * replacement once it fills memory. 354 */ 355 356 p->p_rlimit[RLIMIT_STACK].rlim_cur = DFLSSIZ; 357 p->p_rlimit[RLIMIT_STACK].rlim_max = MAXSSIZ; 358 p->p_rlimit[RLIMIT_DATA].rlim_cur = DFLDSIZ; 359 p->p_rlimit[RLIMIT_DATA].rlim_max = MAXDSIZ; 360 p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(uvmexp.free); 361 } 362 363 #ifdef DEBUG 364 int enableswap = 1; 365 int swapdebug = 0; 366 #define SDB_FOLLOW 1 367 #define SDB_SWAPIN 2 368 #define SDB_SWAPOUT 4 369 #endif 370 371 /* 372 * uvm_swapin: swap in a process's u-area. 373 */ 374 375 void 376 uvm_swapin(p) 377 struct proc *p; 378 { 379 vaddr_t addr; 380 int s; 381 382 addr = (vaddr_t)p->p_addr; 383 /* make P_INMEM true */ 384 uvm_fault_wire(kernel_map, addr, addr + USPACE, 385 VM_PROT_READ | VM_PROT_WRITE); 386 387 /* 388 * Some architectures need to be notified when the user area has 389 * moved to new physical page(s) (e.g. see mips/mips/vm_machdep.c). 390 */ 391 cpu_swapin(p); 392 SCHED_LOCK(s); 393 if (p->p_stat == SRUN) 394 setrunqueue(p); 395 p->p_flag |= P_INMEM; 396 SCHED_UNLOCK(s); 397 p->p_swtime = 0; 398 ++uvmexp.swapins; 399 } 400 401 /* 402 * uvm_scheduler: process zero main loop 403 * 404 * - attempt to swapin every swaped-out, runnable process in order of 405 * priority. 406 * - if not enough memory, wake the pagedaemon and let it clear space. 407 */ 408 409 void 410 uvm_scheduler() 411 { 412 struct proc *p; 413 int pri; 414 struct proc *pp; 415 int ppri; 416 417 loop: 418 #ifdef DEBUG 419 while (!enableswap) 420 tsleep(&proc0, PVM, "noswap", 0); 421 #endif 422 pp = NULL; /* process to choose */ 423 ppri = INT_MIN; /* its priority */ 424 proclist_lock_read(); 425 LIST_FOREACH(p, &allproc, p_list) { 426 427 /* is it a runnable swapped out process? */ 428 if (p->p_stat == SRUN && (p->p_flag & P_INMEM) == 0) { 429 pri = p->p_swtime + p->p_slptime - 430 (p->p_nice - NZERO) * 8; 431 if (pri > ppri) { /* higher priority? remember it. */ 432 pp = p; 433 ppri = pri; 434 } 435 } 436 } 437 /* 438 * XXXSMP: possible unlock/sleep race between here and the 439 * "scheduler" tsleep below.. 440 */ 441 proclist_unlock_read(); 442 443 #ifdef DEBUG 444 if (swapdebug & SDB_FOLLOW) 445 printf("scheduler: running, procp %p pri %d\n", pp, ppri); 446 #endif 447 /* 448 * Nothing to do, back to sleep 449 */ 450 if ((p = pp) == NULL) { 451 tsleep(&proc0, PVM, "scheduler", 0); 452 goto loop; 453 } 454 455 /* 456 * we have found swapped out process which we would like to bring 457 * back in. 458 * 459 * XXX: this part is really bogus cuz we could deadlock on memory 460 * despite our feeble check 461 */ 462 if (uvmexp.free > atop(USPACE)) { 463 #ifdef DEBUG 464 if (swapdebug & SDB_SWAPIN) 465 printf("swapin: pid %d(%s)@%p, pri %d free %d\n", 466 p->p_pid, p->p_comm, p->p_addr, ppri, uvmexp.free); 467 #endif 468 uvm_swapin(p); 469 goto loop; 470 } 471 /* 472 * not enough memory, jab the pageout daemon and wait til the coast 473 * is clear 474 */ 475 #ifdef DEBUG 476 if (swapdebug & SDB_FOLLOW) 477 printf("scheduler: no room for pid %d(%s), free %d\n", 478 p->p_pid, p->p_comm, uvmexp.free); 479 #endif 480 uvm_wait("schedpwait"); 481 #ifdef DEBUG 482 if (swapdebug & SDB_FOLLOW) 483 printf("scheduler: room again, free %d\n", uvmexp.free); 484 #endif 485 goto loop; 486 } 487 488 /* 489 * swappable: is process "p" swappable? 490 */ 491 492 #define swappable(p) \ 493 (((p)->p_flag & (P_SYSTEM | P_INMEM | P_WEXIT)) == P_INMEM && \ 494 (p)->p_holdcnt == 0) 495 496 /* 497 * swapout_threads: find threads that can be swapped and unwire their 498 * u-areas. 499 * 500 * - called by the pagedaemon 501 * - try and swap at least one processs 502 * - processes that are sleeping or stopped for maxslp or more seconds 503 * are swapped... otherwise the longest-sleeping or stopped process 504 * is swapped, otherwise the longest resident process... 505 */ 506 void 507 uvm_swapout_threads() 508 { 509 struct proc *p; 510 struct proc *outp, *outp2; 511 int outpri, outpri2; 512 int didswap = 0; 513 extern int maxslp; 514 /* XXXCDC: should move off to uvmexp. or uvm., also in uvm_meter */ 515 516 #ifdef DEBUG 517 if (!enableswap) 518 return; 519 #endif 520 521 /* 522 * outp/outpri : stop/sleep process with largest sleeptime < maxslp 523 * outp2/outpri2: the longest resident process (its swap time) 524 */ 525 outp = outp2 = NULL; 526 outpri = outpri2 = 0; 527 proclist_lock_read(); 528 LIST_FOREACH(p, &allproc, p_list) { 529 if (!swappable(p)) 530 continue; 531 switch (p->p_stat) { 532 case SRUN: 533 case SONPROC: 534 if (p->p_swtime > outpri2) { 535 outp2 = p; 536 outpri2 = p->p_swtime; 537 } 538 continue; 539 540 case SSLEEP: 541 case SSTOP: 542 if (p->p_slptime >= maxslp) { 543 uvm_swapout(p); 544 didswap++; 545 } else if (p->p_slptime > outpri) { 546 outp = p; 547 outpri = p->p_slptime; 548 } 549 continue; 550 } 551 } 552 proclist_unlock_read(); 553 554 /* 555 * If we didn't get rid of any real duds, toss out the next most 556 * likely sleeping/stopped or running candidate. We only do this 557 * if we are real low on memory since we don't gain much by doing 558 * it (USPACE bytes). 559 */ 560 if (didswap == 0 && uvmexp.free <= atop(round_page(USPACE))) { 561 if ((p = outp) == NULL) 562 p = outp2; 563 #ifdef DEBUG 564 if (swapdebug & SDB_SWAPOUT) 565 printf("swapout_threads: no duds, try procp %p\n", p); 566 #endif 567 if (p) 568 uvm_swapout(p); 569 } 570 } 571 572 /* 573 * uvm_swapout: swap out process "p" 574 * 575 * - currently "swapout" means "unwire U-area" and "pmap_collect()" 576 * the pmap. 577 * - XXXCDC: should deactivate all process' private anonymous memory 578 */ 579 580 static void 581 uvm_swapout(p) 582 struct proc *p; 583 { 584 vaddr_t addr; 585 int s; 586 587 #ifdef DEBUG 588 if (swapdebug & SDB_SWAPOUT) 589 printf("swapout: pid %d(%s)@%p, stat %x pri %d free %d\n", 590 p->p_pid, p->p_comm, p->p_addr, p->p_stat, 591 p->p_slptime, uvmexp.free); 592 #endif 593 594 /* 595 * Do any machine-specific actions necessary before swapout. 596 * This can include saving floating point state, etc. 597 */ 598 cpu_swapout(p); 599 600 /* 601 * Mark it as (potentially) swapped out. 602 */ 603 SCHED_LOCK(s); 604 p->p_flag &= ~P_INMEM; 605 if (p->p_stat == SRUN) 606 remrunqueue(p); 607 SCHED_UNLOCK(s); 608 p->p_swtime = 0; 609 ++uvmexp.swapouts; 610 611 /* 612 * Unwire the to-be-swapped process's user struct and kernel stack. 613 */ 614 addr = (vaddr_t)p->p_addr; 615 uvm_fault_unwire(kernel_map, addr, addr + USPACE); /* !P_INMEM */ 616 pmap_collect(vm_map_pmap(&p->p_vmspace->vm_map)); 617 } 618 619