1 /* $NetBSD: uvm_glue.c,v 1.125 2008/04/24 15:35:31 ad 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 <sys/cdefs.h> 70 __KERNEL_RCSID(0, "$NetBSD: uvm_glue.c,v 1.125 2008/04/24 15:35:31 ad Exp $"); 71 72 #include "opt_coredump.h" 73 #include "opt_kgdb.h" 74 #include "opt_kstack.h" 75 #include "opt_uvmhist.h" 76 77 /* 78 * uvm_glue.c: glue functions 79 */ 80 81 #include <sys/param.h> 82 #include <sys/systm.h> 83 #include <sys/proc.h> 84 #include <sys/resourcevar.h> 85 #include <sys/buf.h> 86 #include <sys/user.h> 87 #include <sys/syncobj.h> 88 #include <sys/cpu.h> 89 #include <sys/atomic.h> 90 91 #include <uvm/uvm.h> 92 93 /* 94 * local prototypes 95 */ 96 97 static void uvm_swapout(struct lwp *); 98 static int uarea_swapin(vaddr_t); 99 100 #define UVM_NUAREA_HIWAT 20 101 #define UVM_NUAREA_LOWAT 16 102 103 #define UAREA_NEXTFREE(uarea) (*(vaddr_t *)(UAREA_TO_USER(uarea))) 104 105 /* 106 * XXXCDC: do these really belong here? 107 */ 108 109 /* 110 * uvm_kernacc: can the kernel access a region of memory 111 * 112 * - used only by /dev/kmem driver (mem.c) 113 */ 114 115 bool 116 uvm_kernacc(void *addr, size_t len, int rw) 117 { 118 bool rv; 119 vaddr_t saddr, eaddr; 120 vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE; 121 122 saddr = trunc_page((vaddr_t)addr); 123 eaddr = round_page((vaddr_t)addr + len); 124 vm_map_lock_read(kernel_map); 125 rv = uvm_map_checkprot(kernel_map, saddr, eaddr, prot); 126 vm_map_unlock_read(kernel_map); 127 128 return(rv); 129 } 130 131 #ifdef KGDB 132 /* 133 * Change protections on kernel pages from addr to addr+len 134 * (presumably so debugger can plant a breakpoint). 135 * 136 * We force the protection change at the pmap level. If we were 137 * to use vm_map_protect a change to allow writing would be lazily- 138 * applied meaning we would still take a protection fault, something 139 * we really don't want to do. It would also fragment the kernel 140 * map unnecessarily. We cannot use pmap_protect since it also won't 141 * enforce a write-enable request. Using pmap_enter is the only way 142 * we can ensure the change takes place properly. 143 */ 144 void 145 uvm_chgkprot(void *addr, size_t len, int rw) 146 { 147 vm_prot_t prot; 148 paddr_t pa; 149 vaddr_t sva, eva; 150 151 prot = rw == B_READ ? VM_PROT_READ : VM_PROT_READ|VM_PROT_WRITE; 152 eva = round_page((vaddr_t)addr + len); 153 for (sva = trunc_page((vaddr_t)addr); sva < eva; sva += PAGE_SIZE) { 154 /* 155 * Extract physical address for the page. 156 */ 157 if (pmap_extract(pmap_kernel(), sva, &pa) == false) 158 panic("%s: invalid page", __func__); 159 pmap_enter(pmap_kernel(), sva, pa, prot, PMAP_WIRED); 160 } 161 pmap_update(pmap_kernel()); 162 } 163 #endif 164 165 /* 166 * uvm_vslock: wire user memory for I/O 167 * 168 * - called from physio and sys___sysctl 169 * - XXXCDC: consider nuking this (or making it a macro?) 170 */ 171 172 int 173 uvm_vslock(struct vmspace *vs, void *addr, size_t len, vm_prot_t access_type) 174 { 175 struct vm_map *map; 176 vaddr_t start, end; 177 int error; 178 179 map = &vs->vm_map; 180 start = trunc_page((vaddr_t)addr); 181 end = round_page((vaddr_t)addr + len); 182 error = uvm_fault_wire(map, start, end, access_type, 0); 183 return error; 184 } 185 186 /* 187 * uvm_vsunlock: unwire user memory wired by uvm_vslock() 188 * 189 * - called from physio and sys___sysctl 190 * - XXXCDC: consider nuking this (or making it a macro?) 191 */ 192 193 void 194 uvm_vsunlock(struct vmspace *vs, void *addr, size_t len) 195 { 196 uvm_fault_unwire(&vs->vm_map, trunc_page((vaddr_t)addr), 197 round_page((vaddr_t)addr + len)); 198 } 199 200 /* 201 * uvm_proc_fork: fork a virtual address space 202 * 203 * - the address space is copied as per parent map's inherit values 204 */ 205 void 206 uvm_proc_fork(struct proc *p1, struct proc *p2, bool shared) 207 { 208 209 if (shared == true) { 210 p2->p_vmspace = NULL; 211 uvmspace_share(p1, p2); 212 } else { 213 p2->p_vmspace = uvmspace_fork(p1->p_vmspace); 214 } 215 216 cpu_proc_fork(p1, p2); 217 } 218 219 220 /* 221 * uvm_lwp_fork: fork a thread 222 * 223 * - a new "user" structure is allocated for the child process 224 * [filled in by MD layer...] 225 * - if specified, the child gets a new user stack described by 226 * stack and stacksize 227 * - NOTE: the kernel stack may be at a different location in the child 228 * process, and thus addresses of automatic variables may be invalid 229 * after cpu_lwp_fork returns in the child process. We do nothing here 230 * after cpu_lwp_fork returns. 231 * - XXXCDC: we need a way for this to return a failure value rather 232 * than just hang 233 */ 234 void 235 uvm_lwp_fork(struct lwp *l1, struct lwp *l2, void *stack, size_t stacksize, 236 void (*func)(void *), void *arg) 237 { 238 int error; 239 240 /* 241 * Wire down the U-area for the process, which contains the PCB 242 * and the kernel stack. Wired state is stored in l->l_flag's 243 * L_INMEM bit rather than in the vm_map_entry's wired count 244 * to prevent kernel_map fragmentation. If we reused a cached U-area, 245 * L_INMEM will already be set and we don't need to do anything. 246 * 247 * Note the kernel stack gets read/write accesses right off the bat. 248 */ 249 250 if ((l2->l_flag & LW_INMEM) == 0) { 251 vaddr_t uarea = USER_TO_UAREA(l2->l_addr); 252 253 if ((error = uarea_swapin(uarea)) != 0) 254 panic("%s: uvm_fault_wire failed: %d", __func__, error); 255 #ifdef PMAP_UAREA 256 /* Tell the pmap this is a u-area mapping */ 257 PMAP_UAREA(uarea); 258 #endif 259 l2->l_flag |= LW_INMEM; 260 } 261 262 #ifdef KSTACK_CHECK_MAGIC 263 /* 264 * fill stack with magic number 265 */ 266 kstack_setup_magic(l2); 267 #endif 268 269 /* 270 * cpu_lwp_fork() copy and update the pcb, and make the child ready 271 * to run. If this is a normal user fork, the child will exit 272 * directly to user mode via child_return() on its first time 273 * slice and will not return here. If this is a kernel thread, 274 * the specified entry point will be executed. 275 */ 276 cpu_lwp_fork(l1, l2, stack, stacksize, func, arg); 277 } 278 279 /* 280 * uvm_cpu_attach: initialize per-CPU data structures. 281 */ 282 283 void 284 uvm_cpu_attach(struct cpu_info *ci) 285 { 286 287 } 288 289 static int 290 uarea_swapin(vaddr_t addr) 291 { 292 293 return uvm_fault_wire(kernel_map, addr, addr + USPACE, 294 VM_PROT_READ | VM_PROT_WRITE, 0); 295 } 296 297 static void 298 uarea_swapout(vaddr_t addr) 299 { 300 301 uvm_fault_unwire(kernel_map, addr, addr + USPACE); 302 } 303 304 #ifndef USPACE_ALIGN 305 #define USPACE_ALIGN 0 306 #endif 307 308 static pool_cache_t uvm_uarea_cache; 309 310 static int 311 uarea_ctor(void *arg, void *obj, int flags) 312 { 313 314 KASSERT((flags & PR_WAITOK) != 0); 315 return uarea_swapin((vaddr_t)obj); 316 } 317 318 static void * 319 uarea_poolpage_alloc(struct pool *pp, int flags) 320 { 321 322 return (void *)uvm_km_alloc(kernel_map, pp->pr_alloc->pa_pagesz, 323 USPACE_ALIGN, UVM_KMF_PAGEABLE | 324 ((flags & PR_WAITOK) != 0 ? UVM_KMF_WAITVA : 325 (UVM_KMF_NOWAIT | UVM_KMF_TRYLOCK))); 326 } 327 328 static void 329 uarea_poolpage_free(struct pool *pp, void *addr) 330 { 331 332 uvm_km_free(kernel_map, (vaddr_t)addr, pp->pr_alloc->pa_pagesz, 333 UVM_KMF_PAGEABLE); 334 } 335 336 static struct pool_allocator uvm_uarea_allocator = { 337 .pa_alloc = uarea_poolpage_alloc, 338 .pa_free = uarea_poolpage_free, 339 .pa_pagesz = USPACE, 340 }; 341 342 void 343 uvm_uarea_init(void) 344 { 345 int flags = PR_NOTOUCH; 346 347 /* 348 * specify PR_NOALIGN unless the alignment provided by 349 * the backend (USPACE_ALIGN) is sufficient to provide 350 * pool page size (UPSACE) alignment. 351 */ 352 353 if ((USPACE_ALIGN == 0 && USPACE != PAGE_SIZE) || 354 (USPACE_ALIGN % USPACE) != 0) { 355 flags |= PR_NOALIGN; 356 } 357 358 uvm_uarea_cache = pool_cache_init(USPACE, USPACE_ALIGN, 0, flags, 359 "uarea", &uvm_uarea_allocator, IPL_NONE, uarea_ctor, NULL, NULL); 360 } 361 362 /* 363 * uvm_uarea_alloc: allocate a u-area 364 */ 365 366 bool 367 uvm_uarea_alloc(vaddr_t *uaddrp) 368 { 369 370 *uaddrp = (vaddr_t)pool_cache_get(uvm_uarea_cache, PR_WAITOK); 371 return true; 372 } 373 374 /* 375 * uvm_uarea_free: free a u-area 376 */ 377 378 void 379 uvm_uarea_free(vaddr_t uaddr, struct cpu_info *ci) 380 { 381 382 pool_cache_put(uvm_uarea_cache, (void *)uaddr); 383 } 384 385 /* 386 * uvm_proc_exit: exit a virtual address space 387 * 388 * - borrow proc0's address space because freeing the vmspace 389 * of the dead process may block. 390 */ 391 392 void 393 uvm_proc_exit(struct proc *p) 394 { 395 struct lwp *l = curlwp; /* XXX */ 396 struct vmspace *ovm; 397 398 KASSERT(p == l->l_proc); 399 ovm = p->p_vmspace; 400 401 /* 402 * borrow proc0's address space. 403 */ 404 pmap_deactivate(l); 405 p->p_vmspace = proc0.p_vmspace; 406 pmap_activate(l); 407 408 uvmspace_free(ovm); 409 } 410 411 void 412 uvm_lwp_exit(struct lwp *l) 413 { 414 vaddr_t va = USER_TO_UAREA(l->l_addr); 415 416 l->l_flag &= ~LW_INMEM; 417 uvm_uarea_free(va, l->l_cpu); 418 l->l_addr = NULL; 419 } 420 421 /* 422 * uvm_init_limit: init per-process VM limits 423 * 424 * - called for process 0 and then inherited by all others. 425 */ 426 427 void 428 uvm_init_limits(struct proc *p) 429 { 430 431 /* 432 * Set up the initial limits on process VM. Set the maximum 433 * resident set size to be all of (reasonably) available memory. 434 * This causes any single, large process to start random page 435 * replacement once it fills memory. 436 */ 437 438 p->p_rlimit[RLIMIT_STACK].rlim_cur = DFLSSIZ; 439 p->p_rlimit[RLIMIT_STACK].rlim_max = maxsmap; 440 p->p_rlimit[RLIMIT_DATA].rlim_cur = DFLDSIZ; 441 p->p_rlimit[RLIMIT_DATA].rlim_max = maxdmap; 442 p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(uvmexp.free); 443 } 444 445 #ifdef DEBUG 446 int enableswap = 1; 447 int swapdebug = 0; 448 #define SDB_FOLLOW 1 449 #define SDB_SWAPIN 2 450 #define SDB_SWAPOUT 4 451 #endif 452 453 /* 454 * uvm_swapin: swap in an lwp's u-area. 455 * 456 * - must be called with the LWP's swap lock held. 457 * - naturally, must not be called with l == curlwp 458 */ 459 460 void 461 uvm_swapin(struct lwp *l) 462 { 463 int error; 464 465 /* XXXSMP notyet KASSERT(mutex_owned(&l->l_swaplock)); */ 466 KASSERT(l != curlwp); 467 468 error = uarea_swapin(USER_TO_UAREA(l->l_addr)); 469 if (error) { 470 panic("%s: rewiring stack failed: %d", __func__, error); 471 } 472 473 /* 474 * Some architectures need to be notified when the user area has 475 * moved to new physical page(s) (e.g. see mips/mips/vm_machdep.c). 476 */ 477 cpu_swapin(l); 478 lwp_lock(l); 479 if (l->l_stat == LSRUN) 480 sched_enqueue(l, false); 481 l->l_flag |= LW_INMEM; 482 l->l_swtime = 0; 483 lwp_unlock(l); 484 ++uvmexp.swapins; 485 } 486 487 /* 488 * uvm_kick_scheduler: kick the scheduler into action if not running. 489 * 490 * - called when swapped out processes have been awoken. 491 */ 492 493 void 494 uvm_kick_scheduler(void) 495 { 496 497 if (uvm.swap_running == false) 498 return; 499 500 mutex_enter(&uvm_scheduler_mutex); 501 uvm.scheduler_kicked = true; 502 cv_signal(&uvm.scheduler_cv); 503 mutex_exit(&uvm_scheduler_mutex); 504 } 505 506 /* 507 * uvm_scheduler: process zero main loop 508 * 509 * - attempt to swapin every swaped-out, runnable process in order of 510 * priority. 511 * - if not enough memory, wake the pagedaemon and let it clear space. 512 */ 513 514 void 515 uvm_scheduler(void) 516 { 517 struct lwp *l, *ll; 518 int pri; 519 int ppri; 520 521 l = curlwp; 522 lwp_lock(l); 523 l->l_priority = PRI_VM; 524 l->l_class = SCHED_FIFO; 525 lwp_unlock(l); 526 527 for (;;) { 528 #ifdef DEBUG 529 mutex_enter(&uvm_scheduler_mutex); 530 while (!enableswap) 531 cv_wait(&uvm.scheduler_cv, &uvm_scheduler_mutex); 532 mutex_exit(&uvm_scheduler_mutex); 533 #endif 534 ll = NULL; /* process to choose */ 535 ppri = INT_MIN; /* its priority */ 536 537 mutex_enter(proc_lock); 538 LIST_FOREACH(l, &alllwp, l_list) { 539 /* is it a runnable swapped out process? */ 540 if (l->l_stat == LSRUN && !(l->l_flag & LW_INMEM)) { 541 pri = l->l_swtime + l->l_slptime - 542 (l->l_proc->p_nice - NZERO) * 8; 543 if (pri > ppri) { /* higher priority? */ 544 ll = l; 545 ppri = pri; 546 } 547 } 548 } 549 #ifdef DEBUG 550 if (swapdebug & SDB_FOLLOW) 551 printf("%s: running, procp %p pri %d\n", __func__, ll, 552 ppri); 553 #endif 554 /* 555 * Nothing to do, back to sleep 556 */ 557 if ((l = ll) == NULL) { 558 mutex_exit(proc_lock); 559 mutex_enter(&uvm_scheduler_mutex); 560 if (uvm.scheduler_kicked == false) 561 cv_wait(&uvm.scheduler_cv, 562 &uvm_scheduler_mutex); 563 uvm.scheduler_kicked = false; 564 mutex_exit(&uvm_scheduler_mutex); 565 continue; 566 } 567 568 /* 569 * we have found swapped out process which we would like 570 * to bring back in. 571 * 572 * XXX: this part is really bogus cuz we could deadlock 573 * on memory despite our feeble check 574 */ 575 if (uvmexp.free > atop(USPACE)) { 576 #ifdef DEBUG 577 if (swapdebug & SDB_SWAPIN) 578 printf("swapin: pid %d(%s)@%p, pri %d " 579 "free %d\n", l->l_proc->p_pid, 580 l->l_proc->p_comm, l->l_addr, ppri, 581 uvmexp.free); 582 #endif 583 mutex_enter(&l->l_swaplock); 584 mutex_exit(proc_lock); 585 uvm_swapin(l); 586 mutex_exit(&l->l_swaplock); 587 continue; 588 } else { 589 /* 590 * not enough memory, jab the pageout daemon and 591 * wait til the coast is clear 592 */ 593 mutex_exit(proc_lock); 594 #ifdef DEBUG 595 if (swapdebug & SDB_FOLLOW) 596 printf("%s: no room for pid %d(%s)," 597 " free %d\n", __func__, l->l_proc->p_pid, 598 l->l_proc->p_comm, uvmexp.free); 599 #endif 600 uvm_wait("schedpwait"); 601 #ifdef DEBUG 602 if (swapdebug & SDB_FOLLOW) 603 printf("%s: room again, free %d\n", __func__, 604 uvmexp.free); 605 #endif 606 } 607 } 608 } 609 610 /* 611 * swappable: is LWP "l" swappable? 612 */ 613 614 static bool 615 swappable(struct lwp *l) 616 { 617 618 if ((l->l_flag & (LW_INMEM|LW_RUNNING|LW_SYSTEM|LW_WEXIT)) != LW_INMEM) 619 return false; 620 if (l->l_holdcnt != 0) 621 return false; 622 if (l->l_syncobj == &rw_syncobj || l->l_syncobj == &mutex_syncobj) 623 return false; 624 return true; 625 } 626 627 /* 628 * swapout_threads: find threads that can be swapped and unwire their 629 * u-areas. 630 * 631 * - called by the pagedaemon 632 * - try and swap at least one processs 633 * - processes that are sleeping or stopped for maxslp or more seconds 634 * are swapped... otherwise the longest-sleeping or stopped process 635 * is swapped, otherwise the longest resident process... 636 */ 637 638 void 639 uvm_swapout_threads(void) 640 { 641 struct lwp *l; 642 struct lwp *outl, *outl2; 643 int outpri, outpri2; 644 int didswap = 0; 645 extern int maxslp; 646 bool gotit; 647 648 /* XXXCDC: should move off to uvmexp. or uvm., also in uvm_meter */ 649 650 #ifdef DEBUG 651 if (!enableswap) 652 return; 653 #endif 654 655 /* 656 * outl/outpri : stop/sleep thread with largest sleeptime < maxslp 657 * outl2/outpri2: the longest resident thread (its swap time) 658 */ 659 outl = outl2 = NULL; 660 outpri = outpri2 = 0; 661 662 restart: 663 mutex_enter(proc_lock); 664 LIST_FOREACH(l, &alllwp, l_list) { 665 KASSERT(l->l_proc != NULL); 666 if (!mutex_tryenter(&l->l_swaplock)) 667 continue; 668 if (!swappable(l)) { 669 mutex_exit(&l->l_swaplock); 670 continue; 671 } 672 switch (l->l_stat) { 673 case LSONPROC: 674 break; 675 676 case LSRUN: 677 if (l->l_swtime > outpri2) { 678 outl2 = l; 679 outpri2 = l->l_swtime; 680 } 681 break; 682 683 case LSSLEEP: 684 case LSSTOP: 685 if (l->l_slptime >= maxslp) { 686 mutex_exit(proc_lock); 687 uvm_swapout(l); 688 /* 689 * Locking in the wrong direction - 690 * try to prevent the LWP from exiting. 691 */ 692 gotit = mutex_tryenter(proc_lock); 693 mutex_exit(&l->l_swaplock); 694 didswap++; 695 if (!gotit) 696 goto restart; 697 continue; 698 } else if (l->l_slptime > outpri) { 699 outl = l; 700 outpri = l->l_slptime; 701 } 702 break; 703 } 704 mutex_exit(&l->l_swaplock); 705 } 706 707 /* 708 * If we didn't get rid of any real duds, toss out the next most 709 * likely sleeping/stopped or running candidate. We only do this 710 * if we are real low on memory since we don't gain much by doing 711 * it (USPACE bytes). 712 */ 713 if (didswap == 0 && uvmexp.free <= atop(round_page(USPACE))) { 714 if ((l = outl) == NULL) 715 l = outl2; 716 #ifdef DEBUG 717 if (swapdebug & SDB_SWAPOUT) 718 printf("%s: no duds, try procp %p\n", __func__, l); 719 #endif 720 if (l) { 721 mutex_enter(&l->l_swaplock); 722 mutex_exit(proc_lock); 723 if (swappable(l)) 724 uvm_swapout(l); 725 mutex_exit(&l->l_swaplock); 726 return; 727 } 728 } 729 730 mutex_exit(proc_lock); 731 } 732 733 /* 734 * uvm_swapout: swap out lwp "l" 735 * 736 * - currently "swapout" means "unwire U-area" and "pmap_collect()" 737 * the pmap. 738 * - must be called with l->l_swaplock held. 739 * - XXXCDC: should deactivate all process' private anonymous memory 740 */ 741 742 static void 743 uvm_swapout(struct lwp *l) 744 { 745 KASSERT(mutex_owned(&l->l_swaplock)); 746 747 #ifdef DEBUG 748 if (swapdebug & SDB_SWAPOUT) 749 printf("%s: lid %d.%d(%s)@%p, stat %x pri %d free %d\n", 750 __func__, l->l_proc->p_pid, l->l_lid, l->l_proc->p_comm, 751 l->l_addr, l->l_stat, l->l_slptime, uvmexp.free); 752 #endif 753 754 /* 755 * Mark it as (potentially) swapped out. 756 */ 757 lwp_lock(l); 758 if (!swappable(l)) { 759 KDASSERT(l->l_cpu != curcpu()); 760 lwp_unlock(l); 761 return; 762 } 763 l->l_flag &= ~LW_INMEM; 764 l->l_swtime = 0; 765 if (l->l_stat == LSRUN) 766 sched_dequeue(l); 767 lwp_unlock(l); 768 l->l_ru.ru_nswap++; 769 ++uvmexp.swapouts; 770 771 /* 772 * Do any machine-specific actions necessary before swapout. 773 * This can include saving floating point state, etc. 774 */ 775 cpu_swapout(l); 776 777 /* 778 * Unwire the to-be-swapped process's user struct and kernel stack. 779 */ 780 uarea_swapout(USER_TO_UAREA(l->l_addr)); 781 pmap_collect(vm_map_pmap(&l->l_proc->p_vmspace->vm_map)); 782 } 783 784 /* 785 * uvm_lwp_hold: prevent lwp "l" from being swapped out, and bring 786 * back into memory if it is currently swapped. 787 */ 788 789 void 790 uvm_lwp_hold(struct lwp *l) 791 { 792 793 if (l == curlwp) { 794 atomic_inc_uint(&l->l_holdcnt); 795 } else { 796 mutex_enter(&l->l_swaplock); 797 if (atomic_inc_uint_nv(&l->l_holdcnt) == 1 && 798 (l->l_flag & LW_INMEM) == 0) 799 uvm_swapin(l); 800 mutex_exit(&l->l_swaplock); 801 } 802 } 803 804 /* 805 * uvm_lwp_rele: release a hold on lwp "l". when the holdcount 806 * drops to zero, it's eligable to be swapped. 807 */ 808 809 void 810 uvm_lwp_rele(struct lwp *l) 811 { 812 813 KASSERT(l->l_holdcnt != 0); 814 815 atomic_dec_uint(&l->l_holdcnt); 816 } 817 818 #ifdef COREDUMP 819 /* 820 * uvm_coredump_walkmap: walk a process's map for the purpose of dumping 821 * a core file. 822 */ 823 824 int 825 uvm_coredump_walkmap(struct proc *p, void *iocookie, 826 int (*func)(struct proc *, void *, struct uvm_coredump_state *), 827 void *cookie) 828 { 829 struct uvm_coredump_state state; 830 struct vmspace *vm = p->p_vmspace; 831 struct vm_map *map = &vm->vm_map; 832 struct vm_map_entry *entry; 833 int error; 834 835 entry = NULL; 836 vm_map_lock_read(map); 837 state.end = 0; 838 for (;;) { 839 if (entry == NULL) 840 entry = map->header.next; 841 else if (!uvm_map_lookup_entry(map, state.end, &entry)) 842 entry = entry->next; 843 if (entry == &map->header) 844 break; 845 846 state.cookie = cookie; 847 if (state.end > entry->start) { 848 state.start = state.end; 849 } else { 850 state.start = entry->start; 851 } 852 state.realend = entry->end; 853 state.end = entry->end; 854 state.prot = entry->protection; 855 state.flags = 0; 856 857 /* 858 * Dump the region unless one of the following is true: 859 * 860 * (1) the region has neither object nor amap behind it 861 * (ie. it has never been accessed). 862 * 863 * (2) the region has no amap and is read-only 864 * (eg. an executable text section). 865 * 866 * (3) the region's object is a device. 867 * 868 * (4) the region is unreadable by the process. 869 */ 870 871 KASSERT(!UVM_ET_ISSUBMAP(entry)); 872 KASSERT(state.start < VM_MAXUSER_ADDRESS); 873 KASSERT(state.end <= VM_MAXUSER_ADDRESS); 874 if (entry->object.uvm_obj == NULL && 875 entry->aref.ar_amap == NULL) { 876 state.realend = state.start; 877 } else if ((entry->protection & VM_PROT_WRITE) == 0 && 878 entry->aref.ar_amap == NULL) { 879 state.realend = state.start; 880 } else if (entry->object.uvm_obj != NULL && 881 UVM_OBJ_IS_DEVICE(entry->object.uvm_obj)) { 882 state.realend = state.start; 883 } else if ((entry->protection & VM_PROT_READ) == 0) { 884 state.realend = state.start; 885 } else { 886 if (state.start >= (vaddr_t)vm->vm_maxsaddr) 887 state.flags |= UVM_COREDUMP_STACK; 888 889 /* 890 * If this an anonymous entry, only dump instantiated 891 * pages. 892 */ 893 if (entry->object.uvm_obj == NULL) { 894 vaddr_t end; 895 896 amap_lock(entry->aref.ar_amap); 897 for (end = state.start; 898 end < state.end; end += PAGE_SIZE) { 899 struct vm_anon *anon; 900 anon = amap_lookup(&entry->aref, 901 end - entry->start); 902 /* 903 * If we have already encountered an 904 * uninstantiated page, stop at the 905 * first instantied page. 906 */ 907 if (anon != NULL && 908 state.realend != state.end) { 909 state.end = end; 910 break; 911 } 912 913 /* 914 * If this page is the first 915 * uninstantiated page, mark this as 916 * the real ending point. Continue to 917 * counting uninstantiated pages. 918 */ 919 if (anon == NULL && 920 state.realend == state.end) { 921 state.realend = end; 922 } 923 } 924 amap_unlock(entry->aref.ar_amap); 925 } 926 } 927 928 929 vm_map_unlock_read(map); 930 error = (*func)(p, iocookie, &state); 931 if (error) 932 return (error); 933 vm_map_lock_read(map); 934 } 935 vm_map_unlock_read(map); 936 937 return (0); 938 } 939 #endif /* COREDUMP */ 940