vm_machdep.c revision 1362
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 * $Id: vm_machdep.c,v 1.18 1994/04/05 03:23:09 davidg Exp $ 42 */ 43 44#include "npx.h" 45#include "param.h" 46#include "systm.h" 47#include "proc.h" 48#include "malloc.h" 49#include "buf.h" 50#include "user.h" 51 52#include "../include/cpu.h" 53 54#include "vm/vm.h" 55#include "vm/vm_kern.h" 56 57#define b_cylin b_resid 58 59#ifndef NOBOUNCE 60 61caddr_t bouncememory; 62vm_offset_t bouncepa, bouncepaend; 63int bouncepages, bpwait; 64vm_map_t io_map; 65int bmwait, bmfreeing; 66 67#define BITS_IN_UNSIGNED (8*sizeof(unsigned)) 68int bounceallocarraysize; 69unsigned *bounceallocarray; 70int bouncefree; 71 72#define SIXTEENMEG (4096*4096) 73#define MAXBKVA 1024 74 75/* special list that can be used at interrupt time for eventual kva free */ 76struct kvasfree { 77 vm_offset_t addr; 78 vm_offset_t size; 79} kvaf[MAXBKVA]; 80 81int kvasfreecnt; 82 83vm_offset_t vm_bounce_kva(); 84/* 85 * get bounce buffer pages (count physically contiguous) 86 * (only 1 inplemented now) 87 */ 88vm_offset_t 89vm_bounce_page_find(count) 90 int count; 91{ 92 int bit; 93 int s,i; 94 95 if (count != 1) 96 panic("vm_bounce_page_find -- no support for > 1 page yet!!!"); 97 98 s = splbio(); 99retry: 100 for (i = 0; i < bounceallocarraysize; i++) { 101 if (bounceallocarray[i] != 0xffffffff) { 102 if (bit = ffs(~bounceallocarray[i])) { 103 bounceallocarray[i] |= 1 << (bit - 1) ; 104 bouncefree -= count; 105 splx(s); 106 return bouncepa + (i * BITS_IN_UNSIGNED + (bit - 1)) * NBPG; 107 } 108 } 109 } 110 bpwait = 1; 111 tsleep((caddr_t) &bounceallocarray, PRIBIO, "bncwai", 0); 112 goto retry; 113} 114 115void 116vm_bounce_kva_free(addr, size, now) 117 vm_offset_t addr; 118 vm_offset_t size; 119 int now; 120{ 121 int s = splbio(); 122 kvaf[kvasfreecnt].addr = addr; 123 kvaf[kvasfreecnt++].size = size; 124 if( now) 125 vm_bounce_kva(0,0); 126 else 127 wakeup((caddr_t) io_map); 128 splx(s); 129} 130 131/* 132 * free count bounce buffer pages 133 */ 134void 135vm_bounce_page_free(pa, count) 136 vm_offset_t pa; 137 int count; 138{ 139 int allocindex; 140 int index; 141 int bit; 142 143 if (count != 1) 144 panic("vm_bounce_page_free -- no support for > 1 page yet!!!\n"); 145 146 index = (pa - bouncepa) / NBPG; 147 148 if ((index < 0) || (index >= bouncepages)) 149 panic("vm_bounce_page_free -- bad index\n"); 150 151 allocindex = index / BITS_IN_UNSIGNED; 152 bit = index % BITS_IN_UNSIGNED; 153 154 bounceallocarray[allocindex] &= ~(1 << bit); 155 156 bouncefree += count; 157 if (bpwait) { 158 bpwait = 0; 159 wakeup((caddr_t) &bounceallocarray); 160 } 161} 162 163/* 164 * allocate count bounce buffer kva pages 165 */ 166vm_offset_t 167vm_bounce_kva(count, waitok) 168 int count; 169 int waitok; 170{ 171 int tofree; 172 int i; 173 int startfree; 174 vm_offset_t kva = 0; 175 int s = splbio(); 176 int size = count; 177 startfree = 0; 178more: 179 if (!bmfreeing && (tofree = kvasfreecnt)) { 180 bmfreeing = 1; 181 for (i = startfree; i < kvasfreecnt; i++) { 182 /* 183 * if we have a kva of the right size, no sense 184 * in freeing/reallocating... 185 * might affect fragmentation short term, but 186 * as long as the amount of io_map is 187 * significantly more than the maximum transfer 188 * size, I don't think that it is a problem. 189 */ 190 pmap_remove(kernel_pmap, 191 kvaf[i].addr, kvaf[i].addr + kvaf[i].size); 192 if( size && !kva && kvaf[i].size == size) { 193 kva = kvaf[i].addr; 194 } else { 195 kmem_free_wakeup(io_map, kvaf[i].addr, 196 kvaf[i].size); 197 } 198 } 199 if (kvasfreecnt != tofree) { 200 startfree = i; 201 bmfreeing = 0; 202 goto more; 203 } 204 kvasfreecnt = 0; 205 bmfreeing = 0; 206 } 207 208 if( size == 0) { 209 splx(s); 210 return NULL; 211 } 212 213 if (!kva && !(kva = kmem_alloc_pageable(io_map, size))) { 214 if( !waitok) { 215 splx(s); 216 return NULL; 217 } 218 bmwait = 1; 219 tsleep((caddr_t) io_map, PRIBIO, "bmwait", 0); 220 goto more; 221 } 222 splx(s); 223 224 return kva; 225} 226 227/* 228 * do the things necessary to the struct buf to implement 229 * bounce buffers... inserted before the disk sort 230 */ 231void 232vm_bounce_alloc(bp) 233 struct buf *bp; 234{ 235 int countvmpg; 236 vm_offset_t vastart, vaend; 237 vm_offset_t vapstart, vapend; 238 vm_offset_t va, kva; 239 vm_offset_t pa; 240 int dobounceflag = 0; 241 int bounceindex; 242 int i; 243 int s; 244 245 if (bouncepages == 0) 246 return; 247 248 if (bp->b_bufsize < bp->b_bcount) { 249 printf("vm_bounce_alloc: b_bufsize(%d) < b_bcount(%d) !!!!\n", 250 bp->b_bufsize, bp->b_bcount); 251 bp->b_bufsize = bp->b_bcount; 252 } 253 254 vastart = (vm_offset_t) bp->b_un.b_addr; 255 vaend = (vm_offset_t) bp->b_un.b_addr + bp->b_bufsize; 256 257 vapstart = i386_trunc_page(vastart); 258 vapend = i386_round_page(vaend); 259 countvmpg = (vapend - vapstart) / NBPG; 260 261/* 262 * if any page is above 16MB, then go into bounce-buffer mode 263 */ 264 va = vapstart; 265 for (i = 0; i < countvmpg; i++) { 266 pa = pmap_kextract(va); 267 if (pa >= SIXTEENMEG) 268 ++dobounceflag; 269 va += NBPG; 270 } 271 if (dobounceflag == 0) 272 return; 273 274 if (bouncepages < dobounceflag) 275 panic("Not enough bounce buffers!!!"); 276 277/* 278 * allocate a replacement kva for b_addr 279 */ 280 kva = vm_bounce_kva(countvmpg*NBPG, 1); 281 va = vapstart; 282 for (i = 0; i < countvmpg; i++) { 283 pa = pmap_kextract(va); 284 if (pa >= SIXTEENMEG) { 285 /* 286 * allocate a replacement page 287 */ 288 vm_offset_t bpa = vm_bounce_page_find(1); 289 pmap_kenter(kva + (NBPG * i), bpa); 290 /* 291 * if we are writing, the copy the data into the page 292 */ 293 if ((bp->b_flags & B_READ) == 0) { 294 pmap_update(); 295 bcopy((caddr_t) va, (caddr_t) kva + (NBPG * i), NBPG); 296 } 297 } else { 298 /* 299 * use original page 300 */ 301 pmap_kenter(kva + (NBPG * i), pa); 302 } 303 va += NBPG; 304 } 305 pmap_update(); 306 307/* 308 * flag the buffer as being bounced 309 */ 310 bp->b_flags |= B_BOUNCE; 311/* 312 * save the original buffer kva 313 */ 314 bp->b_savekva = bp->b_un.b_addr; 315/* 316 * put our new kva into the buffer (offset by original offset) 317 */ 318 bp->b_un.b_addr = (caddr_t) (((vm_offset_t) kva) | 319 ((vm_offset_t) bp->b_savekva & (NBPG - 1))); 320 return; 321} 322 323/* 324 * hook into biodone to free bounce buffer 325 */ 326void 327vm_bounce_free(bp) 328 struct buf *bp; 329{ 330 int i; 331 vm_offset_t origkva, bouncekva; 332 vm_offset_t vastart, vaend; 333 vm_offset_t vapstart, vapend; 334 int countbounce = 0; 335 vm_offset_t firstbouncepa = 0; 336 int firstbounceindex; 337 int countvmpg; 338 vm_offset_t bcount; 339 int s; 340 341/* 342 * if this isn't a bounced buffer, then just return 343 */ 344 if ((bp->b_flags & B_BOUNCE) == 0) 345 return; 346 347 origkva = (vm_offset_t) bp->b_savekva; 348 bouncekva = (vm_offset_t) bp->b_un.b_addr; 349 350 vastart = bouncekva; 351 vaend = bouncekva + bp->b_bufsize; 352 bcount = bp->b_bufsize; 353 354 vapstart = i386_trunc_page(vastart); 355 vapend = i386_round_page(vaend); 356 357 countvmpg = (vapend - vapstart) / NBPG; 358 359/* 360 * check every page in the kva space for b_addr 361 */ 362 for (i = 0; i < countvmpg; i++) { 363 vm_offset_t mybouncepa; 364 vm_offset_t copycount; 365 366 copycount = i386_round_page(bouncekva + 1) - bouncekva; 367 mybouncepa = pmap_kextract(i386_trunc_page(bouncekva)); 368 369/* 370 * if this is a bounced pa, then process as one 371 */ 372 if ((mybouncepa >= bouncepa) && (mybouncepa < bouncepaend)) { 373 if (copycount > bcount) 374 copycount = bcount; 375/* 376 * if this is a read, then copy from bounce buffer into original buffer 377 */ 378 if (bp->b_flags & B_READ) 379 bcopy((caddr_t) bouncekva, (caddr_t) origkva, copycount); 380/* 381 * free the bounce allocation 382 */ 383 vm_bounce_page_free(i386_trunc_page(mybouncepa), 1); 384 } 385 386 origkva += copycount; 387 bouncekva += copycount; 388 bcount -= copycount; 389 } 390 391/* 392 * add the old kva into the "to free" list 393 */ 394 bouncekva = i386_trunc_page((vm_offset_t) bp->b_un.b_addr); 395 vm_bounce_kva_free( bouncekva, countvmpg*NBPG, 0); 396 if (bmwait) { 397 /* 398 * if anyone is waiting on the bounce-map, then wakeup 399 */ 400 wakeup((caddr_t) io_map); 401 bmwait = 0; 402 } 403 404 bp->b_un.b_addr = bp->b_savekva; 405 bp->b_savekva = 0; 406 bp->b_flags &= ~B_BOUNCE; 407 408 return; 409} 410 411#endif /* NOBOUNCE */ 412 413/* 414 * init the bounce buffer system 415 */ 416void 417vm_bounce_init() 418{ 419 vm_offset_t minaddr, maxaddr; 420 421 io_map = kmem_suballoc(kernel_map, &minaddr, &maxaddr, MAXBKVA * NBPG, FALSE); 422 kvasfreecnt = 0; 423 424#ifndef NOBOUNCE 425 if (bouncepages == 0) 426 return; 427 428 bounceallocarraysize = (bouncepages + BITS_IN_UNSIGNED - 1) / BITS_IN_UNSIGNED; 429 bounceallocarray = malloc(bounceallocarraysize * sizeof(unsigned), M_TEMP, M_NOWAIT); 430 431 if (!bounceallocarray) 432 panic("Cannot allocate bounce resource array\n"); 433 434 bzero(bounceallocarray, bounceallocarraysize * sizeof(long)); 435 436 437 bouncepa = pmap_kextract((vm_offset_t) bouncememory); 438 bouncepaend = bouncepa + bouncepages * NBPG; 439 bouncefree = bouncepages; 440#endif 441 442} 443 444 445static void 446cldiskvamerge( kvanew, orig1, orig1cnt, orig2, orig2cnt) 447 vm_offset_t kvanew; 448 vm_offset_t orig1, orig1cnt; 449 vm_offset_t orig2, orig2cnt; 450{ 451 int i; 452 vm_offset_t pa; 453/* 454 * enter the transfer physical addresses into the new kva 455 */ 456 for(i=0;i<orig1cnt;i++) { 457 vm_offset_t pa; 458 pa = pmap_kextract((caddr_t) orig1 + i * PAGE_SIZE); 459 pmap_kenter(kvanew + i * PAGE_SIZE, pa); 460 } 461 462 for(i=0;i<orig2cnt;i++) { 463 vm_offset_t pa; 464 pa = pmap_kextract((caddr_t) orig2 + i * PAGE_SIZE); 465 pmap_kenter(kvanew + (i + orig1cnt) * PAGE_SIZE, pa); 466 } 467 pmap_update(); 468} 469 470void 471cldisksort(struct buf *dp, struct buf *bp, vm_offset_t maxio) 472{ 473 register struct buf *ap, *newbp; 474 int i, trycount=0; 475 vm_offset_t orig1pages, orig2pages; 476 vm_offset_t orig1begin, orig2begin; 477 vm_offset_t kvanew, kvaorig; 478 479 /* 480 * If nothing on the activity queue, then 481 * we become the only thing. 482 */ 483 ap = dp->b_actf; 484 if(ap == NULL) { 485 dp->b_actf = bp; 486 dp->b_actl = bp; 487 bp->av_forw = NULL; 488 return; 489 } 490 491 /* 492 * If we lie after the first (currently active) 493 * request, then we must locate the second request list 494 * and add ourselves to it. 495 */ 496 497 if (bp->b_cylin < ap->b_cylin) { 498 while (ap->av_forw) { 499 /* 500 * Check for an ``inversion'' in the 501 * normally ascending cylinder numbers, 502 * indicating the start of the second request list. 503 */ 504 if (ap->av_forw->b_cylin < ap->b_cylin) { 505 /* 506 * Search the second request list 507 * for the first request at a larger 508 * cylinder number. We go before that; 509 * if there is no such request, we go at end. 510 */ 511 do { 512 if (bp->b_cylin < ap->av_forw->b_cylin) 513 goto insert; 514 ap = ap->av_forw; 515 } while (ap->av_forw); 516 goto insert; /* after last */ 517 } 518 ap = ap->av_forw; 519 } 520 /* 521 * No inversions... we will go after the last, and 522 * be the first request in the second request list. 523 */ 524 goto insert; 525 } 526 /* 527 * Request is at/after the current request... 528 * sort in the first request list. 529 */ 530 while (ap->av_forw) { 531 /* 532 * We want to go after the current request 533 * if there is an inversion after it (i.e. it is 534 * the end of the first request list), or if 535 * the next request is a larger cylinder than our request. 536 */ 537 if (ap->av_forw->b_cylin < ap->b_cylin || 538 bp->b_cylin < ap->av_forw->b_cylin ) 539 goto insert; 540 ap = ap->av_forw; 541 } 542 543insert: 544 /* 545 * we currently only cluster I/O transfers that are at page-aligned 546 * kvas and transfers that are multiples of page lengths. 547 */ 548 if(((bp->b_bcount & PAGE_MASK) == 0) && 549 (((vm_offset_t) bp->b_un.b_addr & PAGE_MASK) == 0)) { 550 /* 551 * merge with previous? 552 * conditions: 553 * 1) We reside physically immediately after the previous block. 554 * 2) The previous block is not first on the device queue because 555 * such a block might be active. 556 * 3) The mode of the two I/Os is identical. 557 * 4) The previous kva is page aligned and the previous transfer 558 * is a multiple of a page in length. 559 * 5) And the total I/O size would be below the maximum. 560 */ 561 if( (ap->b_blkno + (ap->b_bcount / DEV_BSIZE) == bp->b_blkno) && 562 (dp->b_actf != ap) && 563 ((ap->b_flags & ~B_CLUSTER) == bp->b_flags) && 564 ((ap->b_bcount & PAGE_MASK) == 0) && 565 (((vm_offset_t) ap->b_un.b_addr & PAGE_MASK) == 0) && 566 (ap->b_bcount + bp->b_bcount < maxio)) { 567 568 orig1begin = (vm_offset_t) ap->b_un.b_addr; 569 orig1pages = ap->b_bcount / PAGE_SIZE; 570 571 orig2begin = (vm_offset_t) bp->b_un.b_addr; 572 orig2pages = bp->b_bcount / PAGE_SIZE; 573 /* 574 * see if we can allocate a kva, if we cannot, the don't 575 * cluster. 576 */ 577 kvanew = vm_bounce_kva( PAGE_SIZE * (orig1pages + orig2pages), 0); 578 if( !kvanew) { 579 goto nocluster; 580 } 581 582 583 if( (ap->b_flags & B_CLUSTER) == 0) { 584 585 /* 586 * get a physical buf pointer 587 */ 588 newbp = (struct buf *)trypbuf(); 589 if( !newbp) { 590 vm_bounce_kva_free( kvanew, PAGE_SIZE * (orig1pages + orig2pages), 1); 591 goto nocluster; 592 } 593 594 cldiskvamerge( kvanew, orig1begin, orig1pages, orig2begin, orig2pages); 595 596 /* 597 * build the new bp to be handed off to the device 598 */ 599 600 *newbp = *ap; 601 newbp->b_flags |= B_CLUSTER; 602 newbp->b_un.b_addr = (caddr_t) kvanew; 603 newbp->b_bcount += bp->b_bcount; 604 newbp->b_bufsize = newbp->b_bcount; 605 newbp->b_clusterf = ap; 606 newbp->b_clusterl = bp; 607 608 /* 609 * enter the new bp onto the device queue 610 */ 611 if( ap->av_forw) 612 ap->av_forw->av_back = newbp; 613 else 614 dp->b_actl = newbp; 615 616 if( dp->b_actf != ap ) 617 ap->av_back->av_forw = newbp; 618 else 619 dp->b_actf = newbp; 620 621 /* 622 * enter the previous bps onto the cluster queue 623 */ 624 ap->av_forw = bp; 625 bp->av_back = ap; 626 627 ap->av_back = NULL; 628 bp->av_forw = NULL; 629 630 } else { 631 vm_offset_t addr; 632 633 cldiskvamerge( kvanew, orig1begin, orig1pages, orig2begin, orig2pages); 634 /* 635 * free the old kva 636 */ 637 vm_bounce_kva_free( orig1begin, ap->b_bufsize, 0); 638 639 ap->b_un.b_addr = (caddr_t) kvanew; 640 641 ap->b_clusterl->av_forw = bp; 642 bp->av_forw = NULL; 643 bp->av_back = ap->b_clusterl; 644 ap->b_clusterl = bp; 645 646 ap->b_bcount += bp->b_bcount; 647 ap->b_bufsize = ap->b_bcount; 648 } 649 return; 650 /* 651 * merge with next? 652 * conditions: 653 * 1) We reside physically before the next block. 654 * 3) The mode of the two I/Os is identical. 655 * 4) The next kva is page aligned and the next transfer 656 * is a multiple of a page in length. 657 * 5) And the total I/O size would be below the maximum. 658 */ 659 } else if( ap->av_forw && 660 (bp->b_blkno + (bp->b_bcount / DEV_BSIZE) == ap->av_forw->b_blkno) && 661 (bp->b_flags == (ap->av_forw->b_flags & ~B_CLUSTER)) && 662 ((ap->av_forw->b_bcount & PAGE_MASK) == 0) && 663 (((vm_offset_t) ap->av_forw->b_un.b_addr & PAGE_MASK) == 0) && 664 (ap->av_forw->b_bcount + bp->b_bcount < maxio)) { 665 666 orig1begin = (vm_offset_t) bp->b_un.b_addr; 667 orig1pages = bp->b_bcount / PAGE_SIZE; 668 669 orig2begin = (vm_offset_t) ap->av_forw->b_un.b_addr; 670 orig2pages = ap->av_forw->b_bcount / PAGE_SIZE; 671 672 /* 673 * see if we can allocate a kva, if we cannot, the don't 674 * cluster. 675 */ 676 kvanew = vm_bounce_kva( PAGE_SIZE * (orig1pages + orig2pages), 0); 677 if( !kvanew) { 678 goto nocluster; 679 } 680 681 682 /* 683 * if next isn't a cluster we need to create one 684 */ 685 if( (ap->av_forw->b_flags & B_CLUSTER) == 0) { 686 687 /* 688 * get a physical buf pointer 689 */ 690 newbp = (struct buf *)trypbuf(); 691 if( !newbp) { 692 vm_bounce_kva_free( kvanew, PAGE_SIZE * (orig1pages + orig2pages), 1); 693 goto nocluster; 694 } 695 696 cldiskvamerge( kvanew, orig1begin, orig1pages, orig2begin, orig2pages); 697 698 pmap_update(); 699 700 ap = ap->av_forw; 701 *newbp = *ap; 702 newbp->b_flags |= B_CLUSTER; 703 newbp->b_un.b_addr = (caddr_t) kvanew; 704 newbp->b_blkno = bp->b_blkno; 705 newbp->b_bcount += bp->b_bcount; 706 newbp->b_bufsize = newbp->b_bcount; 707 newbp->b_clusterf = bp; 708 newbp->b_clusterl = ap; 709 710 if( ap->av_forw) 711 ap->av_forw->av_back = newbp; 712 else 713 dp->b_actl = newbp; 714 715 if( dp->b_actf != ap ) 716 ap->av_back->av_forw = newbp; 717 else 718 dp->b_actf = newbp; 719 720 bp->av_forw = ap; 721 ap->av_back = bp; 722 723 bp->av_back = NULL; 724 ap->av_forw = NULL; 725 } else { 726 vm_offset_t addr; 727 728 cldiskvamerge( kvanew, orig1begin, orig1pages, orig2begin, orig2pages); 729 ap = ap->av_forw; 730 vm_bounce_kva_free( orig2begin, ap->b_bufsize, 0); 731 732 ap->b_un.b_addr = (caddr_t) kvanew; 733 bp->av_forw = ap->b_clusterf; 734 ap->b_clusterf->av_back = bp; 735 ap->b_clusterf = bp; 736 bp->av_back = NULL; 737 738 ap->b_blkno = bp->b_blkno; 739 ap->b_bcount += bp->b_bcount; 740 ap->b_bufsize = ap->b_bcount; 741 742 } 743 return; 744 } 745 } 746 /* 747 * don't merge 748 */ 749nocluster: 750 bp->av_forw = ap->av_forw; 751 if( bp->av_forw) 752 bp->av_forw->av_back = bp; 753 else 754 dp->b_actl = bp; 755 756 ap->av_forw = bp; 757 bp->av_back = ap; 758} 759 760 761/* 762 * Finish a fork operation, with process p2 nearly set up. 763 * Copy and update the kernel stack and pcb, making the child 764 * ready to run, and marking it so that it can return differently 765 * than the parent. Returns 1 in the child process, 0 in the parent. 766 * We currently double-map the user area so that the stack is at the same 767 * address in each process; in the future we will probably relocate 768 * the frame pointers on the stack after copying. 769 */ 770int 771cpu_fork(p1, p2) 772 register struct proc *p1, *p2; 773{ 774 register struct user *up = p2->p_addr; 775 int foo, offset, addr, i; 776 extern char kstack[]; 777 extern int mvesp(); 778 779 /* 780 * Copy pcb and stack from proc p1 to p2. 781 * We do this as cheaply as possible, copying only the active 782 * part of the stack. The stack and pcb need to agree; 783 * this is tricky, as the final pcb is constructed by savectx, 784 * but its frame isn't yet on the stack when the stack is copied. 785 * swtch compensates for this when the child eventually runs. 786 * This should be done differently, with a single call 787 * that copies and updates the pcb+stack, 788 * replacing the bcopy and savectx. 789 */ 790 p2->p_addr->u_pcb = p1->p_addr->u_pcb; 791 offset = mvesp() - (int)kstack; 792 bcopy((caddr_t)kstack + offset, (caddr_t)p2->p_addr + offset, 793 (unsigned) ctob(UPAGES) - offset); 794 p2->p_regs = p1->p_regs; 795 796 /* 797 * Wire top of address space of child to it's kstack. 798 * First, fault in a page of pte's to map it. 799 */ 800#if 0 801 addr = trunc_page((u_int)vtopte(kstack)); 802 vm_map_pageable(&p2->p_vmspace->vm_map, addr, addr+NBPG, FALSE); 803 for (i=0; i < UPAGES; i++) 804 pmap_enter(&p2->p_vmspace->vm_pmap, kstack+i*NBPG, 805 pmap_extract(kernel_pmap, ((int)p2->p_addr)+i*NBPG), 806 /* 807 * The user area has to be mapped writable because 808 * it contains the kernel stack (when CR0_WP is on 809 * on a 486 there is no user-read/kernel-write 810 * mode). It is protected from user mode access 811 * by the segment limits. 812 */ 813 VM_PROT_READ|VM_PROT_WRITE, TRUE); 814#endif 815 pmap_activate(&p2->p_vmspace->vm_pmap, &up->u_pcb); 816 817 /* 818 * 819 * Arrange for a non-local goto when the new process 820 * is started, to resume here, returning nonzero from setjmp. 821 */ 822 if (savectx(up, 1)) { 823 /* 824 * Return 1 in child. 825 */ 826 return (1); 827 } 828 return (0); 829} 830 831#ifdef notyet 832/* 833 * cpu_exit is called as the last action during exit. 834 * 835 * We change to an inactive address space and a "safe" stack, 836 * passing thru an argument to the new stack. Now, safely isolated 837 * from the resources we're shedding, we release the address space 838 * and any remaining machine-dependent resources, including the 839 * memory for the user structure and kernel stack. 840 * 841 * Next, we assign a dummy context to be written over by swtch, 842 * calling it to send this process off to oblivion. 843 * [The nullpcb allows us to minimize cost in swtch() by not having 844 * a special case]. 845 */ 846struct proc *swtch_to_inactive(); 847volatile void 848cpu_exit(p) 849 register struct proc *p; 850{ 851 static struct pcb nullpcb; /* pcb to overwrite on last swtch */ 852 853#if NNPX > 0 854 npxexit(p); 855#endif /* NNPX */ 856 857 /* move to inactive space and stack, passing arg accross */ 858 p = swtch_to_inactive(p); 859 860 /* drop per-process resources */ 861 vmspace_free(p->p_vmspace); 862 kmem_free(kernel_map, (vm_offset_t)p->p_addr, ctob(UPAGES)); 863 864 p->p_addr = (struct user *) &nullpcb; 865 splclock(); 866 swtch(); 867 /* NOTREACHED */ 868} 869#else 870void 871cpu_exit(p) 872 register struct proc *p; 873{ 874 875#if NNPX > 0 876 npxexit(p); 877#endif /* NNPX */ 878 splclock(); 879 curproc = 0; 880 swtch(); 881 /* 882 * This is to shutup the compiler, and if swtch() failed I suppose 883 * this would be a good thing. This keeps gcc happy because panic 884 * is a volatile void function as well. 885 */ 886 panic("cpu_exit"); 887} 888 889void 890cpu_wait(p) struct proc *p; { 891/* extern vm_map_t upages_map; */ 892 extern char kstack[]; 893 894 /* drop per-process resources */ 895 pmap_remove(vm_map_pmap(kernel_map), (vm_offset_t) p->p_addr, 896 ((vm_offset_t) p->p_addr) + ctob(UPAGES)); 897 kmem_free(kernel_map, (vm_offset_t)p->p_addr, ctob(UPAGES)); 898 vmspace_free(p->p_vmspace); 899} 900#endif 901 902/* 903 * Set a red zone in the kernel stack after the u. area. 904 */ 905void 906setredzone(pte, vaddr) 907 u_short *pte; 908 caddr_t vaddr; 909{ 910/* eventually do this by setting up an expand-down stack segment 911 for ss0: selector, allowing stack access down to top of u. 912 this means though that protection violations need to be handled 913 thru a double fault exception that must do an integral task 914 switch to a known good context, within which a dump can be 915 taken. a sensible scheme might be to save the initial context 916 used by sched (that has physical memory mapped 1:1 at bottom) 917 and take the dump while still in mapped mode */ 918} 919 920/* 921 * Convert kernel VA to physical address 922 */ 923u_long 924kvtop(void *addr) 925{ 926 vm_offset_t va; 927 928 va = pmap_kextract((vm_offset_t)addr); 929 if (va == 0) 930 panic("kvtop: zero page frame"); 931 return((int)va); 932} 933 934extern vm_map_t phys_map; 935 936/* 937 * Map an IO request into kernel virtual address space. 938 * 939 * All requests are (re)mapped into kernel VA space. 940 * Notice that we use b_bufsize for the size of the buffer 941 * to be mapped. b_bcount might be modified by the driver. 942 */ 943void 944vmapbuf(bp) 945 register struct buf *bp; 946{ 947 register int npf; 948 register caddr_t addr; 949 register long flags = bp->b_flags; 950 struct proc *p; 951 int off; 952 vm_offset_t kva; 953 register vm_offset_t pa; 954 955 if ((flags & B_PHYS) == 0) 956 panic("vmapbuf"); 957 addr = bp->b_saveaddr = bp->b_un.b_addr; 958 off = (int)addr & PGOFSET; 959 p = bp->b_proc; 960 npf = btoc(round_page(bp->b_bufsize + off)); 961 kva = kmem_alloc_wait(phys_map, ctob(npf)); 962 bp->b_un.b_addr = (caddr_t) (kva + off); 963 while (npf--) { 964 pa = pmap_extract(&p->p_vmspace->vm_pmap, (vm_offset_t)addr); 965 if (pa == 0) 966 panic("vmapbuf: null page frame"); 967 pmap_kenter(kva, trunc_page(pa)); 968 addr += PAGE_SIZE; 969 kva += PAGE_SIZE; 970 } 971 pmap_update(); 972} 973 974/* 975 * Free the io map PTEs associated with this IO operation. 976 * We also invalidate the TLB entries and restore the original b_addr. 977 */ 978void 979vunmapbuf(bp) 980 register struct buf *bp; 981{ 982 register int npf; 983 register caddr_t addr = bp->b_un.b_addr; 984 vm_offset_t kva; 985 986 if ((bp->b_flags & B_PHYS) == 0) 987 panic("vunmapbuf"); 988 npf = btoc(round_page(bp->b_bufsize + ((int)addr & PGOFSET))); 989 kva = (vm_offset_t)((int)addr & ~PGOFSET); 990 kmem_free_wakeup(phys_map, kva, ctob(npf)); 991 bp->b_un.b_addr = bp->b_saveaddr; 992 bp->b_saveaddr = NULL; 993} 994 995/* 996 * Force reset the processor by invalidating the entire address space! 997 */ 998void 999cpu_reset() { 1000 1001 /* force a shutdown by unmapping entire address space ! */ 1002 bzero((caddr_t) PTD, NBPG); 1003 1004 /* "good night, sweet prince .... <THUNK!>" */ 1005 tlbflush(); 1006 /* NOTREACHED */ 1007 while(1); 1008} 1009 1010/* 1011 * Grow the user stack to allow for 'sp'. This version grows the stack in 1012 * chunks of SGROWSIZ. 1013 */ 1014int 1015grow(p, sp) 1016 struct proc *p; 1017 int sp; 1018{ 1019 unsigned int nss; 1020 caddr_t v; 1021 struct vmspace *vm = p->p_vmspace; 1022 1023 if ((caddr_t)sp <= vm->vm_maxsaddr || (unsigned)sp >= (unsigned)USRSTACK) 1024 return (1); 1025 1026 nss = roundup(USRSTACK - (unsigned)sp, PAGE_SIZE); 1027 1028 if (nss > p->p_rlimit[RLIMIT_STACK].rlim_cur) 1029 return (0); 1030 1031 if (vm->vm_ssize && roundup(vm->vm_ssize << PAGE_SHIFT, 1032 SGROWSIZ) < nss) { 1033 int grow_amount; 1034 /* 1035 * If necessary, grow the VM that the stack occupies 1036 * to allow for the rlimit. This allows us to not have 1037 * to allocate all of the VM up-front in execve (which 1038 * is expensive). 1039 * Grow the VM by the amount requested rounded up to 1040 * the nearest SGROWSIZ to provide for some hysteresis. 1041 */ 1042 grow_amount = roundup((nss - (vm->vm_ssize << PAGE_SHIFT)), SGROWSIZ); 1043 v = (char *)USRSTACK - roundup(vm->vm_ssize << PAGE_SHIFT, 1044 SGROWSIZ) - grow_amount; 1045 /* 1046 * If there isn't enough room to extend by SGROWSIZ, then 1047 * just extend to the maximum size 1048 */ 1049 if (v < vm->vm_maxsaddr) { 1050 v = vm->vm_maxsaddr; 1051 grow_amount = MAXSSIZ - (vm->vm_ssize << PAGE_SHIFT); 1052 } 1053 if (vm_allocate(&vm->vm_map, (vm_offset_t *)&v, 1054 grow_amount, FALSE) != KERN_SUCCESS) { 1055 return (0); 1056 } 1057 vm->vm_ssize += grow_amount >> PAGE_SHIFT; 1058 } 1059 1060 return (1); 1061} 1062