1 /* $NetBSD: uvm_glue.c,v 1.134 2008/11/19 18:36:10 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 1.134 2008/11/19 18:36:10 ad Exp $"); 71 72 #include "opt_kgdb.h" 73 #include "opt_kstack.h" 74 #include "opt_uvmhist.h" 75 76 /* 77 * uvm_glue.c: glue functions 78 */ 79 80 #include <sys/param.h> 81 #include <sys/systm.h> 82 #include <sys/proc.h> 83 #include <sys/resourcevar.h> 84 #include <sys/buf.h> 85 #include <sys/user.h> 86 #include <sys/syncobj.h> 87 #include <sys/cpu.h> 88 #include <sys/atomic.h> 89 90 #include <uvm/uvm.h> 91 92 /* 93 * local prototypes 94 */ 95 96 static void uvm_swapout(struct lwp *); 97 static int uarea_swapin(vaddr_t); 98 99 /* 100 * XXXCDC: do these really belong here? 101 */ 102 103 /* 104 * uvm_kernacc: can the kernel access a region of memory 105 * 106 * - used only by /dev/kmem driver (mem.c) 107 */ 108 109 bool 110 uvm_kernacc(void *addr, size_t len, int rw) 111 { 112 bool rv; 113 vaddr_t saddr, eaddr; 114 vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE; 115 116 saddr = trunc_page((vaddr_t)addr); 117 eaddr = round_page((vaddr_t)addr + len); 118 vm_map_lock_read(kernel_map); 119 rv = uvm_map_checkprot(kernel_map, saddr, eaddr, prot); 120 vm_map_unlock_read(kernel_map); 121 122 return(rv); 123 } 124 125 #ifdef KGDB 126 /* 127 * Change protections on kernel pages from addr to addr+len 128 * (presumably so debugger can plant a breakpoint). 129 * 130 * We force the protection change at the pmap level. If we were 131 * to use vm_map_protect a change to allow writing would be lazily- 132 * applied meaning we would still take a protection fault, something 133 * we really don't want to do. It would also fragment the kernel 134 * map unnecessarily. We cannot use pmap_protect since it also won't 135 * enforce a write-enable request. Using pmap_enter is the only way 136 * we can ensure the change takes place properly. 137 */ 138 void 139 uvm_chgkprot(void *addr, size_t len, int rw) 140 { 141 vm_prot_t prot; 142 paddr_t pa; 143 vaddr_t sva, eva; 144 145 prot = rw == B_READ ? VM_PROT_READ : VM_PROT_READ|VM_PROT_WRITE; 146 eva = round_page((vaddr_t)addr + len); 147 for (sva = trunc_page((vaddr_t)addr); sva < eva; sva += PAGE_SIZE) { 148 /* 149 * Extract physical address for the page. 150 */ 151 if (pmap_extract(pmap_kernel(), sva, &pa) == false) 152 panic("%s: invalid page", __func__); 153 pmap_enter(pmap_kernel(), sva, pa, prot, PMAP_WIRED); 154 } 155 pmap_update(pmap_kernel()); 156 } 157 #endif 158 159 /* 160 * uvm_vslock: wire user memory for I/O 161 * 162 * - called from physio and sys___sysctl 163 * - XXXCDC: consider nuking this (or making it a macro?) 164 */ 165 166 int 167 uvm_vslock(struct vmspace *vs, void *addr, size_t len, vm_prot_t access_type) 168 { 169 struct vm_map *map; 170 vaddr_t start, end; 171 int error; 172 173 map = &vs->vm_map; 174 start = trunc_page((vaddr_t)addr); 175 end = round_page((vaddr_t)addr + len); 176 error = uvm_fault_wire(map, start, end, access_type, 0); 177 return error; 178 } 179 180 /* 181 * uvm_vsunlock: unwire user memory wired by uvm_vslock() 182 * 183 * - called from physio and sys___sysctl 184 * - XXXCDC: consider nuking this (or making it a macro?) 185 */ 186 187 void 188 uvm_vsunlock(struct vmspace *vs, void *addr, size_t len) 189 { 190 uvm_fault_unwire(&vs->vm_map, trunc_page((vaddr_t)addr), 191 round_page((vaddr_t)addr + len)); 192 } 193 194 /* 195 * uvm_proc_fork: fork a virtual address space 196 * 197 * - the address space is copied as per parent map's inherit values 198 */ 199 void 200 uvm_proc_fork(struct proc *p1, struct proc *p2, bool shared) 201 { 202 203 if (shared == true) { 204 p2->p_vmspace = NULL; 205 uvmspace_share(p1, p2); 206 } else { 207 p2->p_vmspace = uvmspace_fork(p1->p_vmspace); 208 } 209 210 cpu_proc_fork(p1, p2); 211 } 212 213 214 /* 215 * uvm_lwp_fork: fork a thread 216 * 217 * - a new "user" structure is allocated for the child process 218 * [filled in by MD layer...] 219 * - if specified, the child gets a new user stack described by 220 * stack and stacksize 221 * - NOTE: the kernel stack may be at a different location in the child 222 * process, and thus addresses of automatic variables may be invalid 223 * after cpu_lwp_fork returns in the child process. We do nothing here 224 * after cpu_lwp_fork returns. 225 * - XXXCDC: we need a way for this to return a failure value rather 226 * than just hang 227 */ 228 void 229 uvm_lwp_fork(struct lwp *l1, struct lwp *l2, void *stack, size_t stacksize, 230 void (*func)(void *), void *arg) 231 { 232 int error; 233 234 /* 235 * Wire down the U-area for the process, which contains the PCB 236 * and the kernel stack. Wired state is stored in l->l_flag's 237 * L_INMEM bit rather than in the vm_map_entry's wired count 238 * to prevent kernel_map fragmentation. If we reused a cached U-area, 239 * L_INMEM will already be set and we don't need to do anything. 240 * 241 * Note the kernel stack gets read/write accesses right off the bat. 242 */ 243 244 if ((l2->l_flag & LW_INMEM) == 0) { 245 vaddr_t uarea = USER_TO_UAREA(l2->l_addr); 246 247 if ((error = uarea_swapin(uarea)) != 0) 248 panic("%s: uvm_fault_wire failed: %d", __func__, error); 249 #ifdef PMAP_UAREA 250 /* Tell the pmap this is a u-area mapping */ 251 PMAP_UAREA(uarea); 252 #endif 253 l2->l_flag |= LW_INMEM; 254 } 255 256 #ifdef KSTACK_CHECK_MAGIC 257 /* 258 * fill stack with magic number 259 */ 260 kstack_setup_magic(l2); 261 #endif 262 263 /* 264 * cpu_lwp_fork() copy and update the pcb, and make the child ready 265 * to run. If this is a normal user fork, the child will exit 266 * directly to user mode via child_return() on its first time 267 * slice and will not return here. If this is a kernel thread, 268 * the specified entry point will be executed. 269 */ 270 cpu_lwp_fork(l1, l2, stack, stacksize, func, arg); 271 } 272 273 static int 274 uarea_swapin(vaddr_t addr) 275 { 276 277 return uvm_fault_wire(kernel_map, addr, addr + USPACE, 278 VM_PROT_READ | VM_PROT_WRITE, 0); 279 } 280 281 static void 282 uarea_swapout(vaddr_t addr) 283 { 284 285 uvm_fault_unwire(kernel_map, addr, addr + USPACE); 286 } 287 288 #ifndef USPACE_ALIGN 289 #define USPACE_ALIGN 0 290 #endif 291 292 static pool_cache_t uvm_uarea_cache; 293 294 static int 295 uarea_ctor(void *arg, void *obj, int flags) 296 { 297 298 KASSERT((flags & PR_WAITOK) != 0); 299 return uarea_swapin((vaddr_t)obj); 300 } 301 302 static void * 303 uarea_poolpage_alloc(struct pool *pp, int flags) 304 { 305 306 return (void *)uvm_km_alloc(kernel_map, pp->pr_alloc->pa_pagesz, 307 USPACE_ALIGN, UVM_KMF_PAGEABLE | 308 ((flags & PR_WAITOK) != 0 ? UVM_KMF_WAITVA : 309 (UVM_KMF_NOWAIT | UVM_KMF_TRYLOCK))); 310 } 311 312 static void 313 uarea_poolpage_free(struct pool *pp, void *addr) 314 { 315 316 uvm_km_free(kernel_map, (vaddr_t)addr, pp->pr_alloc->pa_pagesz, 317 UVM_KMF_PAGEABLE); 318 } 319 320 static struct pool_allocator uvm_uarea_allocator = { 321 .pa_alloc = uarea_poolpage_alloc, 322 .pa_free = uarea_poolpage_free, 323 .pa_pagesz = USPACE, 324 }; 325 326 void 327 uvm_uarea_init(void) 328 { 329 int flags = PR_NOTOUCH; 330 331 /* 332 * specify PR_NOALIGN unless the alignment provided by 333 * the backend (USPACE_ALIGN) is sufficient to provide 334 * pool page size (UPSACE) alignment. 335 */ 336 337 if ((USPACE_ALIGN == 0 && USPACE != PAGE_SIZE) || 338 (USPACE_ALIGN % USPACE) != 0) { 339 flags |= PR_NOALIGN; 340 } 341 342 uvm_uarea_cache = pool_cache_init(USPACE, USPACE_ALIGN, 0, flags, 343 "uarea", &uvm_uarea_allocator, IPL_NONE, uarea_ctor, NULL, NULL); 344 } 345 346 /* 347 * uvm_uarea_alloc: allocate a u-area 348 */ 349 350 bool 351 uvm_uarea_alloc(vaddr_t *uaddrp) 352 { 353 354 *uaddrp = (vaddr_t)pool_cache_get(uvm_uarea_cache, PR_WAITOK); 355 return true; 356 } 357 358 /* 359 * uvm_uarea_free: free a u-area 360 */ 361 362 void 363 uvm_uarea_free(vaddr_t uaddr, struct cpu_info *ci) 364 { 365 366 pool_cache_put(uvm_uarea_cache, (void *)uaddr); 367 } 368 369 /* 370 * uvm_proc_exit: exit a virtual address space 371 * 372 * - borrow proc0's address space because freeing the vmspace 373 * of the dead process may block. 374 */ 375 376 void 377 uvm_proc_exit(struct proc *p) 378 { 379 struct lwp *l = curlwp; /* XXX */ 380 struct vmspace *ovm; 381 382 KASSERT(p == l->l_proc); 383 ovm = p->p_vmspace; 384 385 /* 386 * borrow proc0's address space. 387 */ 388 KPREEMPT_DISABLE(l); 389 pmap_deactivate(l); 390 p->p_vmspace = proc0.p_vmspace; 391 pmap_activate(l); 392 KPREEMPT_ENABLE(l); 393 394 uvmspace_free(ovm); 395 } 396 397 void 398 uvm_lwp_exit(struct lwp *l) 399 { 400 vaddr_t va = USER_TO_UAREA(l->l_addr); 401 402 l->l_flag &= ~LW_INMEM; 403 uvm_uarea_free(va, l->l_cpu); 404 l->l_addr = NULL; 405 } 406 407 /* 408 * uvm_init_limit: init per-process VM limits 409 * 410 * - called for process 0 and then inherited by all others. 411 */ 412 413 void 414 uvm_init_limits(struct proc *p) 415 { 416 417 /* 418 * Set up the initial limits on process VM. Set the maximum 419 * resident set size to be all of (reasonably) available memory. 420 * This causes any single, large process to start random page 421 * replacement once it fills memory. 422 */ 423 424 p->p_rlimit[RLIMIT_STACK].rlim_cur = DFLSSIZ; 425 p->p_rlimit[RLIMIT_STACK].rlim_max = maxsmap; 426 p->p_rlimit[RLIMIT_DATA].rlim_cur = DFLDSIZ; 427 p->p_rlimit[RLIMIT_DATA].rlim_max = maxdmap; 428 p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(uvmexp.free); 429 } 430 431 #ifdef DEBUG 432 int enableswap = 1; 433 int swapdebug = 0; 434 #define SDB_FOLLOW 1 435 #define SDB_SWAPIN 2 436 #define SDB_SWAPOUT 4 437 #endif 438 439 /* 440 * uvm_swapin: swap in an lwp's u-area. 441 * 442 * - must be called with the LWP's swap lock held. 443 * - naturally, must not be called with l == curlwp 444 */ 445 446 void 447 uvm_swapin(struct lwp *l) 448 { 449 int error; 450 451 /* XXXSMP notyet KASSERT(mutex_owned(&l->l_swaplock)); */ 452 KASSERT(l != curlwp); 453 454 error = uarea_swapin(USER_TO_UAREA(l->l_addr)); 455 if (error) { 456 panic("%s: rewiring stack failed: %d", __func__, error); 457 } 458 459 /* 460 * Some architectures need to be notified when the user area has 461 * moved to new physical page(s) (e.g. see mips/mips/vm_machdep.c). 462 */ 463 cpu_swapin(l); 464 lwp_lock(l); 465 if (l->l_stat == LSRUN) 466 sched_enqueue(l, false); 467 l->l_flag |= LW_INMEM; 468 l->l_swtime = 0; 469 lwp_unlock(l); 470 ++uvmexp.swapins; 471 } 472 473 /* 474 * uvm_kick_scheduler: kick the scheduler into action if not running. 475 * 476 * - called when swapped out processes have been awoken. 477 */ 478 479 void 480 uvm_kick_scheduler(void) 481 { 482 483 if (uvm.swap_running == false) 484 return; 485 486 mutex_enter(&uvm_scheduler_mutex); 487 uvm.scheduler_kicked = true; 488 cv_signal(&uvm.scheduler_cv); 489 mutex_exit(&uvm_scheduler_mutex); 490 } 491 492 /* 493 * uvm_scheduler: process zero main loop 494 * 495 * - attempt to swapin every swaped-out, runnable process in order of 496 * priority. 497 * - if not enough memory, wake the pagedaemon and let it clear space. 498 */ 499 500 void 501 uvm_scheduler(void) 502 { 503 struct lwp *l, *ll; 504 int pri; 505 int ppri; 506 507 l = curlwp; 508 lwp_lock(l); 509 l->l_priority = PRI_VM; 510 l->l_class = SCHED_FIFO; 511 lwp_unlock(l); 512 513 for (;;) { 514 #ifdef DEBUG 515 mutex_enter(&uvm_scheduler_mutex); 516 while (!enableswap) 517 cv_wait(&uvm.scheduler_cv, &uvm_scheduler_mutex); 518 mutex_exit(&uvm_scheduler_mutex); 519 #endif 520 ll = NULL; /* process to choose */ 521 ppri = INT_MIN; /* its priority */ 522 523 mutex_enter(proc_lock); 524 LIST_FOREACH(l, &alllwp, l_list) { 525 /* is it a runnable swapped out process? */ 526 if (l->l_stat == LSRUN && !(l->l_flag & LW_INMEM)) { 527 pri = l->l_swtime + l->l_slptime - 528 (l->l_proc->p_nice - NZERO) * 8; 529 if (pri > ppri) { /* higher priority? */ 530 ll = l; 531 ppri = pri; 532 } 533 } 534 } 535 #ifdef DEBUG 536 if (swapdebug & SDB_FOLLOW) 537 printf("%s: running, procp %p pri %d\n", __func__, ll, 538 ppri); 539 #endif 540 /* 541 * Nothing to do, back to sleep 542 */ 543 if ((l = ll) == NULL) { 544 mutex_exit(proc_lock); 545 mutex_enter(&uvm_scheduler_mutex); 546 if (uvm.scheduler_kicked == false) 547 cv_wait(&uvm.scheduler_cv, 548 &uvm_scheduler_mutex); 549 uvm.scheduler_kicked = false; 550 mutex_exit(&uvm_scheduler_mutex); 551 continue; 552 } 553 554 /* 555 * we have found swapped out process which we would like 556 * to bring back in. 557 * 558 * XXX: this part is really bogus cuz we could deadlock 559 * on memory despite our feeble check 560 */ 561 if (uvmexp.free > atop(USPACE)) { 562 #ifdef DEBUG 563 if (swapdebug & SDB_SWAPIN) 564 printf("swapin: pid %d(%s)@%p, pri %d " 565 "free %d\n", l->l_proc->p_pid, 566 l->l_proc->p_comm, l->l_addr, ppri, 567 uvmexp.free); 568 #endif 569 mutex_enter(&l->l_swaplock); 570 mutex_exit(proc_lock); 571 uvm_swapin(l); 572 mutex_exit(&l->l_swaplock); 573 continue; 574 } else { 575 /* 576 * not enough memory, jab the pageout daemon and 577 * wait til the coast is clear 578 */ 579 mutex_exit(proc_lock); 580 #ifdef DEBUG 581 if (swapdebug & SDB_FOLLOW) 582 printf("%s: no room for pid %d(%s)," 583 " free %d\n", __func__, l->l_proc->p_pid, 584 l->l_proc->p_comm, uvmexp.free); 585 #endif 586 uvm_wait("schedpwait"); 587 #ifdef DEBUG 588 if (swapdebug & SDB_FOLLOW) 589 printf("%s: room again, free %d\n", __func__, 590 uvmexp.free); 591 #endif 592 } 593 } 594 } 595 596 /* 597 * swappable: is LWP "l" swappable? 598 */ 599 600 static bool 601 swappable(struct lwp *l) 602 { 603 604 if ((l->l_flag & (LW_INMEM|LW_SYSTEM|LW_WEXIT)) != LW_INMEM) 605 return false; 606 if ((l->l_pflag & LP_RUNNING) != 0) 607 return false; 608 if (l->l_holdcnt != 0) 609 return false; 610 if (l->l_class != SCHED_OTHER) 611 return false; 612 if (l->l_syncobj == &rw_syncobj || l->l_syncobj == &mutex_syncobj) 613 return false; 614 if (l->l_proc->p_stat != SACTIVE && l->l_proc->p_stat != SSTOP) 615 return false; 616 return true; 617 } 618 619 /* 620 * swapout_threads: find threads that can be swapped and unwire their 621 * u-areas. 622 * 623 * - called by the pagedaemon 624 * - try and swap at least one processs 625 * - processes that are sleeping or stopped for maxslp or more seconds 626 * are swapped... otherwise the longest-sleeping or stopped process 627 * is swapped, otherwise the longest resident process... 628 */ 629 630 void 631 uvm_swapout_threads(void) 632 { 633 struct lwp *l; 634 struct lwp *outl, *outl2; 635 int outpri, outpri2; 636 int didswap = 0; 637 extern int maxslp; 638 bool gotit; 639 640 /* XXXCDC: should move off to uvmexp. or uvm., also in uvm_meter */ 641 642 #ifdef DEBUG 643 if (!enableswap) 644 return; 645 #endif 646 647 /* 648 * outl/outpri : stop/sleep thread with largest sleeptime < maxslp 649 * outl2/outpri2: the longest resident thread (its swap time) 650 */ 651 outl = outl2 = NULL; 652 outpri = outpri2 = 0; 653 654 restart: 655 mutex_enter(proc_lock); 656 LIST_FOREACH(l, &alllwp, l_list) { 657 KASSERT(l->l_proc != NULL); 658 if (!mutex_tryenter(&l->l_swaplock)) 659 continue; 660 if (!swappable(l)) { 661 mutex_exit(&l->l_swaplock); 662 continue; 663 } 664 switch (l->l_stat) { 665 case LSONPROC: 666 break; 667 668 case LSRUN: 669 if (l->l_swtime > outpri2) { 670 outl2 = l; 671 outpri2 = l->l_swtime; 672 } 673 break; 674 675 case LSSLEEP: 676 case LSSTOP: 677 if (l->l_slptime >= maxslp) { 678 mutex_exit(proc_lock); 679 uvm_swapout(l); 680 /* 681 * Locking in the wrong direction - 682 * try to prevent the LWP from exiting. 683 */ 684 gotit = mutex_tryenter(proc_lock); 685 mutex_exit(&l->l_swaplock); 686 didswap++; 687 if (!gotit) 688 goto restart; 689 continue; 690 } else if (l->l_slptime > outpri) { 691 outl = l; 692 outpri = l->l_slptime; 693 } 694 break; 695 } 696 mutex_exit(&l->l_swaplock); 697 } 698 699 /* 700 * If we didn't get rid of any real duds, toss out the next most 701 * likely sleeping/stopped or running candidate. We only do this 702 * if we are real low on memory since we don't gain much by doing 703 * it (USPACE bytes). 704 */ 705 if (didswap == 0 && uvmexp.free <= atop(round_page(USPACE))) { 706 if ((l = outl) == NULL) 707 l = outl2; 708 #ifdef DEBUG 709 if (swapdebug & SDB_SWAPOUT) 710 printf("%s: no duds, try procp %p\n", __func__, l); 711 #endif 712 if (l) { 713 mutex_enter(&l->l_swaplock); 714 mutex_exit(proc_lock); 715 if (swappable(l)) 716 uvm_swapout(l); 717 mutex_exit(&l->l_swaplock); 718 return; 719 } 720 } 721 722 mutex_exit(proc_lock); 723 } 724 725 /* 726 * uvm_swapout: swap out lwp "l" 727 * 728 * - currently "swapout" means "unwire U-area" and "pmap_collect()" 729 * the pmap. 730 * - must be called with l->l_swaplock held. 731 * - XXXCDC: should deactivate all process' private anonymous memory 732 */ 733 734 static void 735 uvm_swapout(struct lwp *l) 736 { 737 struct vm_map *map; 738 739 KASSERT(mutex_owned(&l->l_swaplock)); 740 741 #ifdef DEBUG 742 if (swapdebug & SDB_SWAPOUT) 743 printf("%s: lid %d.%d(%s)@%p, stat %x pri %d free %d\n", 744 __func__, l->l_proc->p_pid, l->l_lid, l->l_proc->p_comm, 745 l->l_addr, l->l_stat, l->l_slptime, uvmexp.free); 746 #endif 747 748 /* 749 * Mark it as (potentially) swapped out. 750 */ 751 lwp_lock(l); 752 if (!swappable(l)) { 753 KDASSERT(l->l_cpu != curcpu()); 754 lwp_unlock(l); 755 return; 756 } 757 l->l_flag &= ~LW_INMEM; 758 l->l_swtime = 0; 759 if (l->l_stat == LSRUN) 760 sched_dequeue(l); 761 lwp_unlock(l); 762 l->l_ru.ru_nswap++; 763 ++uvmexp.swapouts; 764 765 /* 766 * Do any machine-specific actions necessary before swapout. 767 * This can include saving floating point state, etc. 768 */ 769 cpu_swapout(l); 770 771 /* 772 * Unwire the to-be-swapped process's user struct and kernel stack. 773 */ 774 uarea_swapout(USER_TO_UAREA(l->l_addr)); 775 map = &l->l_proc->p_vmspace->vm_map; 776 if (vm_map_lock_try(map)) { 777 pmap_collect(vm_map_pmap(map)); 778 vm_map_unlock(map); 779 } 780 } 781 782 /* 783 * uvm_lwp_hold: prevent lwp "l" from being swapped out, and bring 784 * back into memory if it is currently swapped. 785 */ 786 787 void 788 uvm_lwp_hold(struct lwp *l) 789 { 790 791 if (l == curlwp) { 792 atomic_inc_uint(&l->l_holdcnt); 793 } else { 794 mutex_enter(&l->l_swaplock); 795 if (atomic_inc_uint_nv(&l->l_holdcnt) == 1 && 796 (l->l_flag & LW_INMEM) == 0) 797 uvm_swapin(l); 798 mutex_exit(&l->l_swaplock); 799 } 800 } 801 802 /* 803 * uvm_lwp_rele: release a hold on lwp "l". when the holdcount 804 * drops to zero, it's eligable to be swapped. 805 */ 806 807 void 808 uvm_lwp_rele(struct lwp *l) 809 { 810 811 KASSERT(l->l_holdcnt != 0); 812 813 atomic_dec_uint(&l->l_holdcnt); 814 } 815