1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 22 /* 23 * Copyright 2006 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 #pragma ident "%Z%%M% %I% %E% SMI" 28 29 #include <sys/machsystm.h> 30 #include <sys/archsystm.h> 31 #include <sys/vm.h> 32 #include <sys/cpu.h> 33 #include <sys/atomic.h> 34 #include <sys/reboot.h> 35 #include <sys/kdi.h> 36 #include <sys/bootconf.h> 37 #include <sys/memlist_plat.h> 38 #include <sys/memlist_impl.h> 39 #include <sys/prom_plat.h> 40 #include <sys/prom_isa.h> 41 #include <sys/autoconf.h> 42 #include <sys/intreg.h> 43 #include <sys/ivintr.h> 44 #include <sys/fpu/fpusystm.h> 45 #include <sys/iommutsb.h> 46 #include <vm/vm_dep.h> 47 #include <vm/seg_dev.h> 48 #include <vm/seg_kmem.h> 49 #include <vm/seg_kpm.h> 50 #include <vm/seg_map.h> 51 #include <vm/seg_kp.h> 52 #include <sys/sysconf.h> 53 #include <vm/hat_sfmmu.h> 54 #include <sys/kobj.h> 55 #include <sys/sun4asi.h> 56 #include <sys/clconf.h> 57 #include <sys/platform_module.h> 58 #include <sys/panic.h> 59 #include <sys/cpu_sgnblk_defs.h> 60 #include <sys/clock.h> 61 #include <sys/cmn_err.h> 62 #include <sys/promif.h> 63 #include <sys/prom_debug.h> 64 #include <sys/traptrace.h> 65 #include <sys/memnode.h> 66 #include <sys/mem_cage.h> 67 #include <sys/mmu.h> 68 69 extern void setup_trap_table(void); 70 extern void cpu_intrq_setup(struct cpu *); 71 extern void cpu_intrq_register(struct cpu *); 72 extern void contig_mem_init(void); 73 extern void mach_dump_buffer_init(void); 74 extern void mach_descrip_init(void); 75 extern void mach_descrip_startup_fini(void); 76 extern void mach_memscrub(void); 77 extern void mach_fpras(void); 78 extern void mach_cpu_halt_idle(void); 79 extern void mach_hw_copy_limit(void); 80 extern void load_mach_drivers(void); 81 extern void load_tod_module(void); 82 #pragma weak load_tod_module 83 84 extern int ndata_alloc_mmfsa(struct memlist *ndata); 85 #pragma weak ndata_alloc_mmfsa 86 87 extern void cif_init(void); 88 #pragma weak cif_init 89 90 extern void parse_idprom(void); 91 extern void add_vx_handler(char *, int, void (*)(cell_t *)); 92 extern void mem_config_init(void); 93 extern void memseg_remap_init(void); 94 95 extern void mach_kpm_init(void); 96 97 /* 98 * External Data: 99 */ 100 extern int vac_size; /* cache size in bytes */ 101 extern uint_t vac_mask; /* VAC alignment consistency mask */ 102 extern uint_t vac_colors; 103 104 /* 105 * Global Data Definitions: 106 */ 107 108 /* 109 * XXX - Don't port this to new architectures 110 * A 3rd party volume manager driver (vxdm) depends on the symbol romp. 111 * 'romp' has no use with a prom with an IEEE 1275 client interface. 112 * The driver doesn't use the value, but it depends on the symbol. 113 */ 114 void *romp; /* veritas driver won't load without romp 4154976 */ 115 /* 116 * Declare these as initialized data so we can patch them. 117 */ 118 pgcnt_t physmem = 0; /* memory size in pages, patch if you want less */ 119 pgcnt_t segkpsize = 120 btop(SEGKPDEFSIZE); /* size of segkp segment in pages */ 121 uint_t segmap_percent = 12; /* Size of segmap segment */ 122 123 int use_cache = 1; /* cache not reliable (605 bugs) with MP */ 124 int vac_copyback = 1; 125 char *cache_mode = NULL; 126 int use_mix = 1; 127 int prom_debug = 0; 128 129 struct bootops *bootops = 0; /* passed in from boot in %o2 */ 130 caddr_t boot_tba; /* %tba at boot - used by kmdb */ 131 uint_t tba_taken_over = 0; 132 133 caddr_t s_text; /* start of kernel text segment */ 134 caddr_t e_text; /* end of kernel text segment */ 135 caddr_t s_data; /* start of kernel data segment */ 136 caddr_t e_data; /* end of kernel data segment */ 137 138 caddr_t modtext; /* beginning of module text */ 139 size_t modtext_sz; /* size of module text */ 140 caddr_t moddata; /* beginning of module data reserve */ 141 caddr_t e_moddata; /* end of module data reserve */ 142 143 /* 144 * End of first block of contiguous kernel in 32-bit virtual address space 145 */ 146 caddr_t econtig32; /* end of first blk of contiguous kernel */ 147 148 caddr_t ncbase; /* beginning of non-cached segment */ 149 caddr_t ncend; /* end of non-cached segment */ 150 caddr_t sdata; /* beginning of data segment */ 151 152 caddr_t extra_etva; /* beginning of unused nucleus text */ 153 pgcnt_t extra_etpg; /* number of pages of unused nucleus text */ 154 155 size_t ndata_remain_sz; /* bytes from end of data to 4MB boundary */ 156 caddr_t nalloc_base; /* beginning of nucleus allocation */ 157 caddr_t nalloc_end; /* end of nucleus allocatable memory */ 158 caddr_t valloc_base; /* beginning of kvalloc segment */ 159 160 caddr_t kmem64_base; /* base of kernel mem segment in 64-bit space */ 161 caddr_t kmem64_end; /* end of kernel mem segment in 64-bit space */ 162 163 uintptr_t shm_alignment; /* VAC address consistency modulus */ 164 struct memlist *phys_install; /* Total installed physical memory */ 165 struct memlist *phys_avail; /* Available (unreserved) physical memory */ 166 struct memlist *virt_avail; /* Available (unmapped?) virtual memory */ 167 struct memlist ndata; /* memlist of nucleus allocatable memory */ 168 int memexp_flag; /* memory expansion card flag */ 169 uint64_t ecache_flushaddr; /* physical address used for flushing E$ */ 170 pgcnt_t obp_pages; /* Physical pages used by OBP */ 171 172 /* 173 * VM data structures 174 */ 175 long page_hashsz; /* Size of page hash table (power of two) */ 176 struct page *pp_base; /* Base of system page struct array */ 177 size_t pp_sz; /* Size in bytes of page struct array */ 178 struct page **page_hash; /* Page hash table */ 179 struct seg ktextseg; /* Segment used for kernel executable image */ 180 struct seg kvalloc; /* Segment used for "valloc" mapping */ 181 struct seg kpseg; /* Segment used for pageable kernel virt mem */ 182 struct seg ktexthole; /* Segment used for nucleus text hole */ 183 struct seg kmapseg; /* Segment used for generic kernel mappings */ 184 struct seg kpmseg; /* Segment used for physical mapping */ 185 struct seg kdebugseg; /* Segment used for the kernel debugger */ 186 187 uintptr_t kpm_pp_base; /* Base of system kpm_page array */ 188 size_t kpm_pp_sz; /* Size of system kpm_page array */ 189 pgcnt_t kpm_npages; /* How many kpm pages are managed */ 190 191 struct seg *segkp = &kpseg; /* Pageable kernel virtual memory segment */ 192 struct seg *segkmap = &kmapseg; /* Kernel generic mapping segment */ 193 struct seg *segkpm = &kpmseg; /* 64bit kernel physical mapping segment */ 194 195 /* 196 * debugger pages (if allocated) 197 */ 198 struct vnode kdebugvp; 199 200 /* 201 * Segment for relocated kernel structures in 64-bit large RAM kernels 202 */ 203 struct seg kmem64; 204 205 struct memseg *memseg_base; 206 size_t memseg_sz; /* Used to translate a va to page */ 207 struct vnode unused_pages_vp; 208 209 /* 210 * VM data structures allocated early during boot. 211 */ 212 size_t pagehash_sz; 213 uint64_t memlist_sz; 214 215 char tbr_wr_addr_inited = 0; 216 217 218 /* 219 * Static Routines: 220 */ 221 static void memlist_add(uint64_t, uint64_t, struct memlist **, 222 struct memlist **); 223 static void kphysm_init(page_t *, struct memseg *, pgcnt_t, uintptr_t, 224 pgcnt_t); 225 static void kvm_init(void); 226 227 static void startup_init(void); 228 static void startup_memlist(void); 229 static void startup_modules(void); 230 static void startup_bop_gone(void); 231 static void startup_vm(void); 232 static void startup_end(void); 233 static void setup_cage_params(void); 234 static void startup_create_io_node(void); 235 236 static pgcnt_t npages; 237 static struct memlist *memlist; 238 void *memlist_end; 239 240 static pgcnt_t bop_alloc_pages; 241 static caddr_t hblk_base; 242 uint_t hblk_alloc_dynamic = 0; 243 uint_t hblk1_min = H1MIN; 244 uint_t hblk8_min; 245 246 247 /* 248 * Hooks for unsupported platforms and down-rev firmware 249 */ 250 int iam_positron(void); 251 #pragma weak iam_positron 252 static void do_prom_version_check(void); 253 static void kpm_init(void); 254 static void kpm_npages_setup(int); 255 static void kpm_memseg_init(void); 256 257 /* 258 * After receiving a thermal interrupt, this is the number of seconds 259 * to delay before shutting off the system, assuming 260 * shutdown fails. Use /etc/system to change the delay if this isn't 261 * large enough. 262 */ 263 int thermal_powerdown_delay = 1200; 264 265 /* 266 * Used to hold off page relocations into the cage until OBP has completed 267 * its boot-time handoff of its resources to the kernel. 268 */ 269 int page_relocate_ready = 0; 270 271 /* 272 * Enable some debugging messages concerning memory usage... 273 */ 274 #ifdef DEBUGGING_MEM 275 static int debugging_mem; 276 static void 277 printmemlist(char *title, struct memlist *list) 278 { 279 if (!debugging_mem) 280 return; 281 282 printf("%s\n", title); 283 284 while (list) { 285 prom_printf("\taddr = 0x%x %8x, size = 0x%x %8x\n", 286 (uint32_t)(list->address >> 32), (uint32_t)list->address, 287 (uint32_t)(list->size >> 32), (uint32_t)(list->size)); 288 list = list->next; 289 } 290 } 291 292 void 293 printmemseg(struct memseg *memseg) 294 { 295 if (!debugging_mem) 296 return; 297 298 printf("memseg\n"); 299 300 while (memseg) { 301 prom_printf("\tpage = 0x%p, epage = 0x%p, " 302 "pfn = 0x%x, epfn = 0x%x\n", 303 memseg->pages, memseg->epages, 304 memseg->pages_base, memseg->pages_end); 305 memseg = memseg->next; 306 } 307 } 308 309 #define debug_pause(str) halt((str)) 310 #define MPRINTF(str) if (debugging_mem) prom_printf((str)) 311 #define MPRINTF1(str, a) if (debugging_mem) prom_printf((str), (a)) 312 #define MPRINTF2(str, a, b) if (debugging_mem) prom_printf((str), (a), (b)) 313 #define MPRINTF3(str, a, b, c) \ 314 if (debugging_mem) prom_printf((str), (a), (b), (c)) 315 #else /* DEBUGGING_MEM */ 316 #define MPRINTF(str) 317 #define MPRINTF1(str, a) 318 #define MPRINTF2(str, a, b) 319 #define MPRINTF3(str, a, b, c) 320 #endif /* DEBUGGING_MEM */ 321 322 /* Simple message to indicate that the bootops pointer has been zeroed */ 323 #ifdef DEBUG 324 static int bootops_gone_on = 0; 325 #define BOOTOPS_GONE() \ 326 if (bootops_gone_on) \ 327 prom_printf("The bootops vec is zeroed now!\n"); 328 #else 329 #define BOOTOPS_GONE() 330 #endif /* DEBUG */ 331 332 /* 333 * Monitor pages may not be where this says they are. 334 * and the debugger may not be there either. 335 * 336 * Note that 'pages' here are *physical* pages, which are 8k on sun4u. 337 * 338 * Physical memory layout 339 * (not necessarily contiguous) 340 * (THIS IS SOMEWHAT WRONG) 341 * /-----------------------\ 342 * | monitor pages | 343 * availmem -|-----------------------| 344 * | | 345 * | page pool | 346 * | | 347 * |-----------------------| 348 * | configured tables | 349 * | buffers | 350 * firstaddr -|-----------------------| 351 * | hat data structures | 352 * |-----------------------| 353 * | kernel data, bss | 354 * |-----------------------| 355 * | interrupt stack | 356 * |-----------------------| 357 * | kernel text (RO) | 358 * |-----------------------| 359 * | trap table (4k) | 360 * |-----------------------| 361 * page 1 | panicbuf | 362 * |-----------------------| 363 * page 0 | reclaimed | 364 * |_______________________| 365 * 366 * 367 * 368 * Kernel's Virtual Memory Layout. 369 * /-----------------------\ 370 * 0xFFFFFFFF.FFFFFFFF -| |- 371 * | OBP's virtual page | 372 * | tables | 373 * 0xFFFFFFFC.00000000 -|-----------------------|- 374 * : : 375 * : : 376 * 0xFFFFFE00.00000000 -|-----------------------|- 377 * | | Ultrasparc I/II support 378 * | segkpm segment | up to 2TB of physical 379 * | (64-bit kernel ONLY) | memory, VAC has 2 colors 380 * | | 381 * 0xFFFFFA00.00000000 -|-----------------------|- 2TB segkpm alignment 382 * : : 383 * : : 384 * 0xFFFFF810.00000000 -|-----------------------|- hole_end 385 * | | ^ 386 * | UltraSPARC I/II call | | 387 * | bug requires an extra | | 388 * | 4 GB of space between | | 389 * | hole and used RAM | | 390 * | | | 391 * 0xFFFFF800.00000000 -|-----------------------|- | 392 * | | | 393 * | Virtual Address Hole | UltraSPARC 394 * | on UltraSPARC I/II | I/II * ONLY * 395 * | | | 396 * 0x00000800.00000000 -|-----------------------|- | 397 * | | | 398 * | UltraSPARC I/II call | | 399 * | bug requires an extra | | 400 * | 4 GB of space between | | 401 * | hole and used RAM | | 402 * | | v 403 * 0x000007FF.00000000 -|-----------------------|- hole_start ----- 404 * : : ^ 405 * : : | 406 * 0x00000XXX.XXXXXXXX -|-----------------------|- kmem64_end | 407 * | | | 408 * | 64-bit kernel ONLY | | 409 * | | | 410 * | kmem64 segment | | 411 * | | | 412 * | (Relocated extra HME | Approximately 413 * | block allocations, | 1 TB of virtual 414 * | memnode freelists, | address space 415 * | HME hash buckets, | | 416 * | mml_table, kpmp_table,| | 417 * | page_t array and | | 418 * | hashblock pool to | | 419 * | avoid hard-coded | | 420 * | 32-bit vaddr | | 421 * | limitations) | | 422 * | | v 423 * 0x00000700.00000000 -|-----------------------|- SYSLIMIT (kmem64_base) 424 * | | 425 * | segkmem segment | (SYSLIMIT - SYSBASE = 4TB) 426 * | | 427 * 0x00000300.00000000 -|-----------------------|- SYSBASE 428 * : : 429 * : : 430 * -|-----------------------|- 431 * | | 432 * | segmap segment | SEGMAPSIZE (1/8th physmem, 433 * | | 256G MAX) 434 * 0x000002a7.50000000 -|-----------------------|- SEGMAPBASE 435 * : : 436 * : : 437 * -|-----------------------|- 438 * | | 439 * | segkp | SEGKPSIZE (2GB) 440 * | | 441 * | | 442 * 0x000002a1.00000000 -|-----------------------|- SEGKPBASE 443 * | | 444 * 0x000002a0.00000000 -|-----------------------|- MEMSCRUBBASE 445 * | | (SEGKPBASE - 0x400000) 446 * 0x0000029F.FFE00000 -|-----------------------|- ARGSBASE 447 * | | (MEMSCRUBBASE - NCARGS) 448 * 0x0000029F.FFD80000 -|-----------------------|- PPMAPBASE 449 * | | (ARGSBASE - PPMAPSIZE) 450 * 0x0000029F.FFD00000 -|-----------------------|- PPMAP_FAST_BASE 451 * | | 452 * 0x0000029F.FF980000 -|-----------------------|- PIOMAPBASE 453 * | | 454 * 0x0000029F.FF580000 -|-----------------------|- NARG_BASE 455 * : : 456 * : : 457 * 0x00000000.FFFFFFFF -|-----------------------|- OFW_END_ADDR 458 * | | 459 * | OBP | 460 * | | 461 * 0x00000000.F0000000 -|-----------------------|- OFW_START_ADDR 462 * | kmdb | 463 * 0x00000000.EDD00000 -|-----------------------|- SEGDEBUGBASE 464 * : : 465 * : : 466 * 0x00000000.7c000000 -|-----------------------|- SYSLIMIT32 467 * | | 468 * | segkmem32 segment | (SYSLIMIT32 - SYSBASE32 = 469 * | | ~64MB) 470 * 0x00000000.78002000 -|-----------------------| 471 * | panicbuf | 472 * 0x00000000.78000000 -|-----------------------|- SYSBASE32 473 * : : 474 * : : 475 * | | 476 * |-----------------------|- econtig32 477 * | vm structures | 478 * 0x00000000.01C00000 |-----------------------|- nalloc_end 479 * | TSBs | 480 * |-----------------------|- end/nalloc_base 481 * | kernel data & bss | 482 * 0x00000000.01800000 -|-----------------------| 483 * : nucleus text hole : 484 * 0x00000000.01400000 -|-----------------------| 485 * : : 486 * |-----------------------| 487 * | module text | 488 * |-----------------------|- e_text/modtext 489 * | kernel text | 490 * |-----------------------| 491 * | trap table (48k) | 492 * 0x00000000.01000000 -|-----------------------|- KERNELBASE 493 * | reserved for trapstat |} TSTAT_TOTAL_SIZE 494 * |-----------------------| 495 * | | 496 * | invalid | 497 * | | 498 * 0x00000000.00000000 _|_______________________| 499 * 500 * 501 * 502 * 32-bit User Virtual Memory Layout. 503 * /-----------------------\ 504 * | | 505 * | invalid | 506 * | | 507 * 0xFFC00000 -|-----------------------|- USERLIMIT 508 * | user stack | 509 * : : 510 * : : 511 * : : 512 * | user data | 513 * -|-----------------------|- 514 * | user text | 515 * 0x00002000 -|-----------------------|- 516 * | invalid | 517 * 0x00000000 _|_______________________| 518 * 519 * 520 * 521 * 64-bit User Virtual Memory Layout. 522 * /-----------------------\ 523 * | | 524 * | invalid | 525 * | | 526 * 0xFFFFFFFF.80000000 -|-----------------------|- USERLIMIT 527 * | user stack | 528 * : : 529 * : : 530 * : : 531 * | user data | 532 * -|-----------------------|- 533 * | user text | 534 * 0x00000000.00100000 -|-----------------------|- 535 * | invalid | 536 * 0x00000000.00000000 _|_______________________| 537 */ 538 539 extern caddr_t ecache_init_scrub_flush_area(caddr_t alloc_base); 540 extern uint64_t ecache_flush_address(void); 541 542 #pragma weak load_platform_modules 543 #pragma weak plat_startup_memlist 544 #pragma weak ecache_init_scrub_flush_area 545 #pragma weak ecache_flush_address 546 547 548 /* 549 * By default the DR Cage is enabled for maximum OS 550 * MPSS performance. Users needing to disable the cage mechanism 551 * can set this variable to zero via /etc/system. 552 * Disabling the cage on systems supporting Dynamic Reconfiguration (DR) 553 * will result in loss of DR functionality. 554 * Platforms wishing to disable kernel Cage by default 555 * should do so in their set_platform_defaults() routine. 556 */ 557 int kernel_cage_enable = 1; 558 559 static void 560 setup_cage_params(void) 561 { 562 void (*func)(void); 563 564 func = (void (*)(void))kobj_getsymvalue("set_platform_cage_params", 0); 565 if (func != NULL) { 566 (*func)(); 567 return; 568 } 569 570 if (kernel_cage_enable == 0) { 571 return; 572 } 573 kcage_range_lock(); 574 if (kcage_range_init(phys_avail, 1) == 0) { 575 kcage_init(total_pages / 256); 576 } 577 kcage_range_unlock(); 578 579 if (kcage_on) { 580 cmn_err(CE_NOTE, "!Kernel Cage is ENABLED"); 581 } else { 582 cmn_err(CE_NOTE, "!Kernel Cage is DISABLED"); 583 } 584 585 } 586 587 /* 588 * Machine-dependent startup code 589 */ 590 void 591 startup(void) 592 { 593 startup_init(); 594 if (&startup_platform) 595 startup_platform(); 596 startup_memlist(); 597 startup_modules(); 598 setup_cage_params(); 599 startup_bop_gone(); 600 startup_vm(); 601 startup_end(); 602 } 603 604 struct regs sync_reg_buf; 605 uint64_t sync_tt; 606 607 void 608 sync_handler(void) 609 { 610 struct trap_info ti; 611 int i; 612 613 /* 614 * Prevent trying to talk to the other CPUs since they are 615 * sitting in the prom and won't reply. 616 */ 617 for (i = 0; i < NCPU; i++) { 618 if ((i != CPU->cpu_id) && CPU_XCALL_READY(i)) { 619 cpu[i]->cpu_flags &= ~CPU_READY; 620 cpu[i]->cpu_flags |= CPU_QUIESCED; 621 CPUSET_DEL(cpu_ready_set, cpu[i]->cpu_id); 622 } 623 } 624 625 /* 626 * We've managed to get here without going through the 627 * normal panic code path. Try and save some useful 628 * information. 629 */ 630 if (!panicstr && (curthread->t_panic_trap == NULL)) { 631 ti.trap_type = sync_tt; 632 ti.trap_regs = &sync_reg_buf; 633 ti.trap_addr = NULL; 634 ti.trap_mmu_fsr = 0x0; 635 636 curthread->t_panic_trap = &ti; 637 } 638 639 /* 640 * If we're re-entering the panic path, update the signature 641 * block so that the SC knows we're in the second part of panic. 642 */ 643 if (panicstr) 644 CPU_SIGNATURE(OS_SIG, SIGST_EXIT, SIGSUBST_DUMP, -1); 645 646 nopanicdebug = 1; /* do not perform debug_enter() prior to dump */ 647 panic("sync initiated"); 648 } 649 650 651 static void 652 startup_init(void) 653 { 654 /* 655 * We want to save the registers while we're still in OBP 656 * so that we know they haven't been fiddled with since. 657 * (In principle, OBP can't change them just because it 658 * makes a callback, but we'd rather not depend on that 659 * behavior.) 660 */ 661 char sync_str[] = 662 "warning @ warning off : sync " 663 "%%tl-c %%tstate h# %p x! " 664 "%%g1 h# %p x! %%g2 h# %p x! %%g3 h# %p x! " 665 "%%g4 h# %p x! %%g5 h# %p x! %%g6 h# %p x! " 666 "%%g7 h# %p x! %%o0 h# %p x! %%o1 h# %p x! " 667 "%%o2 h# %p x! %%o3 h# %p x! %%o4 h# %p x! " 668 "%%o5 h# %p x! %%o6 h# %p x! %%o7 h# %p x! " 669 "%%tl-c %%tpc h# %p x! %%tl-c %%tnpc h# %p x! " 670 "%%y h# %p l! %%tl-c %%tt h# %p x! " 671 "sync ; warning !"; 672 673 /* 674 * 20 == num of %p substrings 675 * 16 == max num of chars %p will expand to. 676 */ 677 char bp[sizeof (sync_str) + 16 * 20]; 678 679 (void) check_boot_version(BOP_GETVERSION(bootops)); 680 681 /* 682 * Initialize ptl1 stack for the 1st CPU. 683 */ 684 ptl1_init_cpu(&cpu0); 685 686 /* 687 * Initialize the address map for cache consistent mappings 688 * to random pages; must be done after vac_size is set. 689 */ 690 ppmapinit(); 691 692 /* 693 * Initialize the PROM callback handler. 694 */ 695 init_vx_handler(); 696 697 /* 698 * have prom call sync_callback() to handle the sync and 699 * save some useful information which will be stored in the 700 * core file later. 701 */ 702 (void) sprintf((char *)bp, sync_str, 703 (void *)&sync_reg_buf.r_tstate, (void *)&sync_reg_buf.r_g1, 704 (void *)&sync_reg_buf.r_g2, (void *)&sync_reg_buf.r_g3, 705 (void *)&sync_reg_buf.r_g4, (void *)&sync_reg_buf.r_g5, 706 (void *)&sync_reg_buf.r_g6, (void *)&sync_reg_buf.r_g7, 707 (void *)&sync_reg_buf.r_o0, (void *)&sync_reg_buf.r_o1, 708 (void *)&sync_reg_buf.r_o2, (void *)&sync_reg_buf.r_o3, 709 (void *)&sync_reg_buf.r_o4, (void *)&sync_reg_buf.r_o5, 710 (void *)&sync_reg_buf.r_o6, (void *)&sync_reg_buf.r_o7, 711 (void *)&sync_reg_buf.r_pc, (void *)&sync_reg_buf.r_npc, 712 (void *)&sync_reg_buf.r_y, (void *)&sync_tt); 713 prom_interpret(bp, 0, 0, 0, 0, 0); 714 add_vx_handler("sync", 1, (void (*)(cell_t *))sync_handler); 715 } 716 717 static u_longlong_t *boot_physinstalled, *boot_physavail, *boot_virtavail; 718 static size_t boot_physinstalled_len, boot_physavail_len, boot_virtavail_len; 719 720 #define IVSIZE ((MAXIVNUM * sizeof (intr_vec_t *)) + \ 721 (MAX_RSVD_IV * sizeof (intr_vec_t)) + \ 722 (MAX_RSVD_IVX * sizeof (intr_vecx_t))) 723 724 /* 725 * As OBP takes up some RAM when the system boots, pages will already be "lost" 726 * to the system and reflected in npages by the time we see it. 727 * 728 * We only want to allocate kernel structures in the 64-bit virtual address 729 * space on systems with enough RAM to make the overhead of keeping track of 730 * an extra kernel memory segment worthwhile. 731 * 732 * Since OBP has already performed its memory allocations by this point, if we 733 * have more than MINMOVE_RAM_MB MB of RAM left free, go ahead and map 734 * memory in the 64-bit virtual address space; otherwise keep allocations 735 * contiguous with we've mapped so far in the 32-bit virtual address space. 736 */ 737 #define MINMOVE_RAM_MB ((size_t)1900) 738 #define MB_TO_BYTES(mb) ((mb) * 1048576ul) 739 740 pgcnt_t tune_npages = (pgcnt_t) 741 (MB_TO_BYTES(MINMOVE_RAM_MB)/ (size_t)MMU_PAGESIZE); 742 743 static void 744 startup_memlist(void) 745 { 746 size_t alloc_sz; 747 size_t ctrs_sz; 748 caddr_t alloc_base; 749 caddr_t ctrs_base, ctrs_end; 750 caddr_t memspace; 751 caddr_t va; 752 int memblocks = 0; 753 struct memlist *cur; 754 size_t syslimit = (size_t)SYSLIMIT; 755 size_t sysbase = (size_t)SYSBASE; 756 int alloc_alignsize = MMU_PAGESIZE; 757 extern void page_coloring_init(void); 758 759 /* 760 * Initialize enough of the system to allow kmem_alloc to work by 761 * calling boot to allocate its memory until the time that 762 * kvm_init is completed. The page structs are allocated after 763 * rounding up end to the nearest page boundary; the memsegs are 764 * initialized and the space they use comes from the kernel heap. 765 * With appropriate initialization, they can be reallocated later 766 * to a size appropriate for the machine's configuration. 767 * 768 * At this point, memory is allocated for things that will never 769 * need to be freed, this used to be "valloced". This allows a 770 * savings as the pages don't need page structures to describe 771 * them because them will not be managed by the vm system. 772 */ 773 774 /* 775 * We're loaded by boot with the following configuration (as 776 * specified in the sun4u/conf/Mapfile): 777 * 778 * text: 4 MB chunk aligned on a 4MB boundary 779 * data & bss: 4 MB chunk aligned on a 4MB boundary 780 * 781 * These two chunks will eventually be mapped by 2 locked 4MB 782 * ttes and will represent the nucleus of the kernel. This gives 783 * us some free space that is already allocated, some or all of 784 * which is made available to kernel module text. 785 * 786 * The free space in the data-bss chunk is used for nucleus 787 * allocatable data structures and we reserve it using the 788 * nalloc_base and nalloc_end variables. This space is currently 789 * being used for hat data structures required for tlb miss 790 * handling operations. We align nalloc_base to a l2 cache 791 * linesize because this is the line size the hardware uses to 792 * maintain cache coherency. 793 * 256K is carved out for module data. 794 */ 795 796 nalloc_base = (caddr_t)roundup((uintptr_t)e_data, MMU_PAGESIZE); 797 moddata = nalloc_base; 798 e_moddata = nalloc_base + MODDATA; 799 nalloc_base = e_moddata; 800 801 nalloc_end = (caddr_t)roundup((uintptr_t)nalloc_base, MMU_PAGESIZE4M); 802 valloc_base = nalloc_base; 803 804 /* 805 * Calculate the start of the data segment. 806 */ 807 sdata = (caddr_t)((uintptr_t)e_data & MMU_PAGEMASK4M); 808 809 PRM_DEBUG(moddata); 810 PRM_DEBUG(nalloc_base); 811 PRM_DEBUG(nalloc_end); 812 PRM_DEBUG(sdata); 813 814 /* 815 * Remember any slop after e_text so we can give it to the modules. 816 */ 817 PRM_DEBUG(e_text); 818 modtext = (caddr_t)roundup((uintptr_t)e_text, MMU_PAGESIZE); 819 if (((uintptr_t)modtext & MMU_PAGEMASK4M) != (uintptr_t)s_text) 820 panic("nucleus text overflow"); 821 modtext_sz = (caddr_t)roundup((uintptr_t)modtext, MMU_PAGESIZE4M) - 822 modtext; 823 PRM_DEBUG(modtext); 824 PRM_DEBUG(modtext_sz); 825 826 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 827 &boot_physavail, &boot_physavail_len, 828 &boot_virtavail, &boot_virtavail_len); 829 /* 830 * Remember what the physically available highest page is 831 * so that dumpsys works properly, and find out how much 832 * memory is installed. 833 */ 834 installed_top_size_memlist_array(boot_physinstalled, 835 boot_physinstalled_len, &physmax, &physinstalled); 836 PRM_DEBUG(physinstalled); 837 PRM_DEBUG(physmax); 838 839 /* Fill out memory nodes config structure */ 840 startup_build_mem_nodes(boot_physinstalled, boot_physinstalled_len); 841 842 /* 843 * Get the list of physically available memory to size 844 * the number of page structures needed. 845 */ 846 size_physavail(boot_physavail, boot_physavail_len, &npages, &memblocks); 847 /* 848 * This first snap shot of npages can represent the pages used 849 * by OBP's text and data approximately. This is used in the 850 * the calculation of the kernel size 851 */ 852 obp_pages = physinstalled - npages; 853 854 855 /* 856 * On small-memory systems (<MODTEXT_SM_SIZE MB, currently 256MB), the 857 * in-nucleus module text is capped to MODTEXT_SM_CAP bytes (currently 858 * 2MB) and any excess pages are put on physavail. The assumption is 859 * that small-memory systems will need more pages more than they'll 860 * need efficiently-mapped module texts. 861 */ 862 if ((physinstalled < mmu_btop(MODTEXT_SM_SIZE << 20)) && 863 modtext_sz > MODTEXT_SM_CAP) { 864 extra_etpg = mmu_btop(modtext_sz - MODTEXT_SM_CAP); 865 modtext_sz = MODTEXT_SM_CAP; 866 } else 867 extra_etpg = 0; 868 PRM_DEBUG(extra_etpg); 869 PRM_DEBUG(modtext_sz); 870 extra_etva = modtext + modtext_sz; 871 PRM_DEBUG(extra_etva); 872 873 /* 874 * Account for any pages after e_text and e_data. 875 */ 876 npages += extra_etpg; 877 npages += mmu_btopr(nalloc_end - nalloc_base); 878 PRM_DEBUG(npages); 879 880 /* 881 * npages is the maximum of available physical memory possible. 882 * (ie. it will never be more than this) 883 */ 884 885 /* 886 * initialize the nucleus memory allocator. 887 */ 888 ndata_alloc_init(&ndata, (uintptr_t)nalloc_base, (uintptr_t)nalloc_end); 889 890 /* 891 * Allocate mmu fault status area from the nucleus data area. 892 */ 893 if ((&ndata_alloc_mmfsa != NULL) && (ndata_alloc_mmfsa(&ndata) != 0)) 894 cmn_err(CE_PANIC, "no more nucleus memory after mfsa alloc"); 895 896 /* 897 * Allocate kernel TSBs from the nucleus data area. 898 */ 899 if (ndata_alloc_tsbs(&ndata, npages) != 0) 900 cmn_err(CE_PANIC, "no more nucleus memory after tsbs alloc"); 901 902 /* 903 * Allocate dmv dispatch table from the nucleus data area. 904 */ 905 if (ndata_alloc_dmv(&ndata) != 0) 906 cmn_err(CE_PANIC, "no more nucleus memory after dmv alloc"); 907 908 909 page_coloring_init(); 910 911 /* 912 * Allocate page_freelists bin headers for memnode 0 from the 913 * nucleus data area. 914 */ 915 if (ndata_alloc_page_freelists(&ndata, 0) != 0) 916 cmn_err(CE_PANIC, 917 "no more nucleus memory after page free lists alloc"); 918 919 if (kpm_enable) { 920 kpm_init(); 921 /* 922 * kpm page space -- Update kpm_npages and make the 923 * same assumption about fragmenting as it is done 924 * for memseg_sz. 925 */ 926 kpm_npages_setup(memblocks + 4); 927 } 928 929 /* 930 * Allocate hat related structs from the nucleus data area. 931 */ 932 if (ndata_alloc_hat(&ndata, npages, kpm_npages) != 0) 933 cmn_err(CE_PANIC, "no more nucleus memory after hat alloc"); 934 935 /* 936 * We want to do the BOP_ALLOCs before the real allocation of page 937 * structs in order to not have to allocate page structs for this 938 * memory. We need to calculate a virtual address because we want 939 * the page structs to come before other allocations in virtual address 940 * space. This is so some (if not all) of page structs can actually 941 * live in the nucleus. 942 */ 943 944 /* 945 * WARNING WARNING WARNING WARNING WARNING WARNING WARNING 946 * 947 * There are comments all over the SFMMU code warning of dire 948 * consequences if the TSBs are moved out of 32-bit space. This 949 * is largely because the asm code uses "sethi %hi(addr)"-type 950 * instructions which will not provide the expected result if the 951 * address is a 64-bit one. 952 * 953 * WARNING WARNING WARNING WARNING WARNING WARNING WARNING 954 */ 955 alloc_base = (caddr_t)roundup((uintptr_t)nalloc_end, MMU_PAGESIZE); 956 alloc_base = sfmmu_ktsb_alloc(alloc_base); 957 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, ecache_alignsize); 958 PRM_DEBUG(alloc_base); 959 960 /* 961 * Allocate IOMMU TSB array. We do this here so that the physical 962 * memory gets deducted from the PROM's physical memory list. 963 */ 964 alloc_base = iommu_tsb_init(alloc_base); 965 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 966 ecache_alignsize); 967 PRM_DEBUG(alloc_base); 968 969 /* 970 * Platforms like Starcat and OPL need special structures assigned in 971 * 32-bit virtual address space because their probing routines execute 972 * FCode, and FCode can't handle 64-bit virtual addresses... 973 */ 974 if (&plat_startup_memlist) { 975 alloc_base = plat_startup_memlist(alloc_base); 976 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 977 ecache_alignsize); 978 PRM_DEBUG(alloc_base); 979 } 980 981 /* 982 * If we have enough memory, use 4M pages for alignment because it 983 * greatly reduces the number of TLB misses we take albeit at the cost 984 * of possible RAM wastage (degenerate case of 4 MB - MMU_PAGESIZE per 985 * allocation.) Still, the speedup on large memory systems (e.g. > 64 986 * GB) is quite noticeable, so it is worth the effort to do if we can. 987 * 988 * Note, however, that this speedup will only occur if the boot PROM 989 * uses the largest possible MMU page size possible to map memory 990 * requests that are properly aligned and sized (for example, a request 991 * for a multiple of 4MB of memory aligned to a 4MB boundary will 992 * result in a mapping using a 4MB MMU page.) 993 * 994 * Even then, the large page mappings will only speed things up until 995 * the startup process proceeds a bit further, as when 996 * sfmmu_map_prom_mappings() copies page mappings from the PROM to the 997 * kernel it remaps everything but the TSBs using 8K pages anyway... 998 * 999 * At some point in the future, sfmmu_map_prom_mappings() will be 1000 * rewritten to copy memory mappings to the kernel using the same MMU 1001 * page sizes the PROM used. When that occurs, if the PROM did use 1002 * large MMU pages to map memory, the alignment/sizing work we're 1003 * doing now should give us a nice extra performance boost, albeit at 1004 * the cost of greater RAM usage... 1005 */ 1006 alloc_alignsize = ((npages >= tune_npages) ? MMU_PAGESIZE4M : 1007 MMU_PAGESIZE); 1008 1009 PRM_DEBUG(tune_npages); 1010 PRM_DEBUG(alloc_alignsize); 1011 1012 /* 1013 * Save off where the contiguous allocations to date have ended 1014 * in econtig32. 1015 */ 1016 econtig32 = alloc_base; 1017 PRM_DEBUG(econtig32); 1018 1019 if (econtig32 > (caddr_t)KERNEL_LIMIT32) 1020 cmn_err(CE_PANIC, "econtig32 too big"); 1021 1022 /* 1023 * To avoid memory allocation collisions in the 32-bit virtual address 1024 * space, make allocations from this point forward in 64-bit virtual 1025 * address space starting at syslimit and working up. Also use the 1026 * alignment specified by alloc_alignsize, as we may be able to save 1027 * ourselves TLB misses by using larger page sizes if they're 1028 * available. 1029 * 1030 * All this is needed because on large memory systems, the default 1031 * Solaris allocations will collide with SYSBASE32, which is hard 1032 * coded to be at the virtual address 0x78000000. Therefore, on 64-bit 1033 * kernels, move the allocations to a location in the 64-bit virtual 1034 * address space space, allowing those structures to grow without 1035 * worry. 1036 * 1037 * On current CPUs we'll run out of physical memory address bits before 1038 * we need to worry about the allocations running into anything else in 1039 * VM or the virtual address holes on US-I and II, as there's currently 1040 * about 1 TB of addressable space before the US-I/II VA hole. 1041 */ 1042 kmem64_base = (caddr_t)syslimit; 1043 PRM_DEBUG(kmem64_base); 1044 1045 alloc_base = (caddr_t)roundup((uintptr_t)kmem64_base, alloc_alignsize); 1046 1047 /* 1048 * If KHME and/or UHME hash buckets won't fit in the nucleus, allocate 1049 * them here. 1050 */ 1051 if (khme_hash == NULL || uhme_hash == NULL) { 1052 /* 1053 * alloc_hme_buckets() will align alloc_base properly before 1054 * assigning the hash buckets, so we don't need to do it 1055 * before the call... 1056 */ 1057 alloc_base = alloc_hme_buckets(alloc_base, alloc_alignsize); 1058 1059 PRM_DEBUG(alloc_base); 1060 PRM_DEBUG(khme_hash); 1061 PRM_DEBUG(uhme_hash); 1062 } 1063 1064 /* 1065 * Allocate the remaining page freelists. NUMA systems can 1066 * have lots of page freelists, one per node, which quickly 1067 * outgrow the amount of nucleus memory available. 1068 */ 1069 if (max_mem_nodes > 1) { 1070 int mnode; 1071 caddr_t alloc_start = alloc_base; 1072 1073 for (mnode = 1; mnode < max_mem_nodes; mnode++) { 1074 alloc_base = alloc_page_freelists(mnode, alloc_base, 1075 ecache_alignsize); 1076 } 1077 1078 if (alloc_base > alloc_start) { 1079 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1080 alloc_alignsize); 1081 if ((caddr_t)BOP_ALLOC(bootops, alloc_start, 1082 alloc_base - alloc_start, 1083 alloc_alignsize) != alloc_start) 1084 cmn_err(CE_PANIC, 1085 "Unable to alloc page freelists\n"); 1086 } 1087 1088 PRM_DEBUG(alloc_base); 1089 } 1090 1091 if (!mml_table) { 1092 size_t mmltable_sz; 1093 1094 /* 1095 * We need to allocate the mml_table here because there 1096 * was not enough space within the nucleus. 1097 */ 1098 mmltable_sz = sizeof (kmutex_t) * mml_table_sz; 1099 alloc_sz = roundup(mmltable_sz, alloc_alignsize); 1100 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1101 alloc_alignsize); 1102 1103 if ((mml_table = (kmutex_t *)BOP_ALLOC(bootops, alloc_base, 1104 alloc_sz, alloc_alignsize)) != (kmutex_t *)alloc_base) 1105 panic("mml_table alloc failure"); 1106 1107 alloc_base += alloc_sz; 1108 PRM_DEBUG(mml_table); 1109 PRM_DEBUG(alloc_base); 1110 } 1111 1112 if (kpm_enable && !(kpmp_table || kpmp_stable)) { 1113 size_t kpmptable_sz; 1114 caddr_t table; 1115 1116 /* 1117 * We need to allocate either kpmp_table or kpmp_stable here 1118 * because there was not enough space within the nucleus. 1119 */ 1120 kpmptable_sz = (kpm_smallpages == 0) ? 1121 sizeof (kpm_hlk_t) * kpmp_table_sz : 1122 sizeof (kpm_shlk_t) * kpmp_stable_sz; 1123 1124 alloc_sz = roundup(kpmptable_sz, alloc_alignsize); 1125 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1126 alloc_alignsize); 1127 1128 table = BOP_ALLOC(bootops, alloc_base, alloc_sz, 1129 alloc_alignsize); 1130 1131 if (table != alloc_base) 1132 panic("kpmp_table or kpmp_stable alloc failure"); 1133 1134 if (kpm_smallpages == 0) { 1135 kpmp_table = (kpm_hlk_t *)table; 1136 PRM_DEBUG(kpmp_table); 1137 } else { 1138 kpmp_stable = (kpm_shlk_t *)table; 1139 PRM_DEBUG(kpmp_stable); 1140 } 1141 1142 alloc_base += alloc_sz; 1143 PRM_DEBUG(alloc_base); 1144 } 1145 1146 if (&ecache_init_scrub_flush_area) { 1147 /* 1148 * Pass alloc_base directly, as the routine itself is 1149 * responsible for any special alignment requirements... 1150 */ 1151 alloc_base = ecache_init_scrub_flush_area(alloc_base); 1152 PRM_DEBUG(alloc_base); 1153 } 1154 1155 /* 1156 * Take the most current snapshot we can by calling mem-update. 1157 */ 1158 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1159 &boot_physavail, &boot_physavail_len, 1160 &boot_virtavail, &boot_virtavail_len); 1161 1162 /* 1163 * Reset npages and memblocks based on boot_physavail list. 1164 */ 1165 size_physavail(boot_physavail, boot_physavail_len, &npages, &memblocks); 1166 PRM_DEBUG(npages); 1167 1168 /* 1169 * Account for extra memory after e_text. 1170 */ 1171 npages += extra_etpg; 1172 1173 /* 1174 * Calculate the largest free memory chunk in the nucleus data area. 1175 * We need to figure out if page structs can fit in there or not. 1176 * We also make sure enough page structs get created for any physical 1177 * memory we might be returning to the system. 1178 */ 1179 ndata_remain_sz = ndata_maxsize(&ndata); 1180 PRM_DEBUG(ndata_remain_sz); 1181 1182 pp_sz = sizeof (struct page) * npages; 1183 1184 /* 1185 * Here's a nice bit of code based on somewhat recursive logic: 1186 * 1187 * If the page array would fit within the nucleus, we want to 1188 * add npages to cover any extra memory we may be returning back 1189 * to the system. 1190 * 1191 * HOWEVER, the page array is sized by calculating the size of 1192 * (struct page * npages), as are the pagehash table, ctrs and 1193 * memseg_list, so the very act of performing the calculation below may 1194 * in fact make the array large enough that it no longer fits in the 1195 * nucleus, meaning there would now be a much larger area of the 1196 * nucleus free that should really be added to npages, which would 1197 * make the page array that much larger, and so on. 1198 * 1199 * This also ignores the memory possibly used in the nucleus for the 1200 * the page hash, ctrs and memseg list and the fact that whether they 1201 * fit there or not varies with the npages calculation below, but we 1202 * don't even factor them into the equation at this point; perhaps we 1203 * should or perhaps we should just take the approach that the few 1204 * extra pages we could add via this calculation REALLY aren't worth 1205 * the hassle... 1206 */ 1207 if (ndata_remain_sz > pp_sz) { 1208 size_t spare = ndata_spare(&ndata, pp_sz, ecache_alignsize); 1209 1210 npages += mmu_btop(spare); 1211 1212 pp_sz = npages * sizeof (struct page); 1213 1214 pp_base = ndata_alloc(&ndata, pp_sz, ecache_alignsize); 1215 } 1216 1217 /* 1218 * If physmem is patched to be non-zero, use it instead of 1219 * the monitor value unless physmem is larger than the total 1220 * amount of memory on hand. 1221 */ 1222 if (physmem == 0 || physmem > npages) 1223 physmem = npages; 1224 1225 /* 1226 * If pp_base is NULL that means the routines above have determined 1227 * the page array will not fit in the nucleus; we'll have to 1228 * BOP_ALLOC() ourselves some space for them. 1229 */ 1230 if (pp_base == NULL) { 1231 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1232 alloc_alignsize); 1233 1234 alloc_sz = roundup(pp_sz, alloc_alignsize); 1235 1236 if ((pp_base = (struct page *)BOP_ALLOC(bootops, 1237 alloc_base, alloc_sz, alloc_alignsize)) != 1238 (struct page *)alloc_base) 1239 panic("page alloc failure"); 1240 1241 alloc_base += alloc_sz; 1242 } 1243 1244 /* 1245 * The page structure hash table size is a power of 2 1246 * such that the average hash chain length is PAGE_HASHAVELEN. 1247 */ 1248 page_hashsz = npages / PAGE_HASHAVELEN; 1249 page_hashsz = 1 << highbit((ulong_t)page_hashsz); 1250 pagehash_sz = sizeof (struct page *) * page_hashsz; 1251 1252 /* 1253 * We want to TRY to fit the page structure hash table, 1254 * the page size free list counters, the memseg list and 1255 * and the kpm page space in the nucleus if possible. 1256 * 1257 * alloc_sz counts how much memory needs to be allocated by 1258 * BOP_ALLOC(). 1259 */ 1260 page_hash = ndata_alloc(&ndata, pagehash_sz, ecache_alignsize); 1261 1262 alloc_sz = (page_hash == NULL ? pagehash_sz : 0); 1263 1264 /* 1265 * Size up per page size free list counters. 1266 */ 1267 ctrs_sz = page_ctrs_sz(); 1268 ctrs_base = ndata_alloc(&ndata, ctrs_sz, ecache_alignsize); 1269 1270 if (ctrs_base == NULL) 1271 alloc_sz = roundup(alloc_sz, ecache_alignsize) + ctrs_sz; 1272 1273 /* 1274 * The memseg list is for the chunks of physical memory that 1275 * will be managed by the vm system. The number calculated is 1276 * a guess as boot may fragment it more when memory allocations 1277 * are made before kphysm_init(). Currently, there are two 1278 * allocations before then, so we assume each causes fragmen- 1279 * tation, and add a couple more for good measure. 1280 */ 1281 memseg_sz = sizeof (struct memseg) * (memblocks + 4); 1282 memseg_base = ndata_alloc(&ndata, memseg_sz, ecache_alignsize); 1283 1284 if (memseg_base == NULL) 1285 alloc_sz = roundup(alloc_sz, ecache_alignsize) + memseg_sz; 1286 1287 1288 if (kpm_enable) { 1289 /* 1290 * kpm page space -- Update kpm_npages and make the 1291 * same assumption about fragmenting as it is done 1292 * for memseg_sz above. 1293 */ 1294 kpm_npages_setup(memblocks + 4); 1295 kpm_pp_sz = (kpm_smallpages == 0) ? 1296 kpm_npages * sizeof (kpm_page_t): 1297 kpm_npages * sizeof (kpm_spage_t); 1298 1299 kpm_pp_base = (uintptr_t)ndata_alloc(&ndata, kpm_pp_sz, 1300 ecache_alignsize); 1301 1302 if (kpm_pp_base == NULL) 1303 alloc_sz = roundup(alloc_sz, ecache_alignsize) + 1304 kpm_pp_sz; 1305 } 1306 1307 if (alloc_sz > 0) { 1308 uintptr_t bop_base; 1309 1310 /* 1311 * We need extra memory allocated through BOP_ALLOC. 1312 */ 1313 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1314 alloc_alignsize); 1315 1316 alloc_sz = roundup(alloc_sz, alloc_alignsize); 1317 1318 if ((bop_base = (uintptr_t)BOP_ALLOC(bootops, alloc_base, 1319 alloc_sz, alloc_alignsize)) != (uintptr_t)alloc_base) 1320 panic("system page struct alloc failure"); 1321 1322 alloc_base += alloc_sz; 1323 1324 if (page_hash == NULL) { 1325 page_hash = (struct page **)bop_base; 1326 bop_base = roundup(bop_base + pagehash_sz, 1327 ecache_alignsize); 1328 } 1329 1330 if (ctrs_base == NULL) { 1331 ctrs_base = (caddr_t)bop_base; 1332 bop_base = roundup(bop_base + ctrs_sz, 1333 ecache_alignsize); 1334 } 1335 1336 if (memseg_base == NULL) { 1337 memseg_base = (struct memseg *)bop_base; 1338 bop_base = roundup(bop_base + memseg_sz, 1339 ecache_alignsize); 1340 } 1341 1342 if (kpm_enable && kpm_pp_base == NULL) { 1343 kpm_pp_base = (uintptr_t)bop_base; 1344 bop_base = roundup(bop_base + kpm_pp_sz, 1345 ecache_alignsize); 1346 } 1347 1348 ASSERT(bop_base <= (uintptr_t)alloc_base); 1349 } 1350 1351 /* 1352 * Initialize per page size free list counters. 1353 */ 1354 ctrs_end = page_ctrs_alloc(ctrs_base); 1355 ASSERT(ctrs_base + ctrs_sz >= ctrs_end); 1356 1357 PRM_DEBUG(page_hash); 1358 PRM_DEBUG(memseg_base); 1359 PRM_DEBUG(kpm_pp_base); 1360 PRM_DEBUG(kpm_pp_sz); 1361 PRM_DEBUG(pp_base); 1362 PRM_DEBUG(pp_sz); 1363 PRM_DEBUG(alloc_base); 1364 1365 #ifdef TRAPTRACE 1366 /* 1367 * Allocate trap trace buffer last so as not to affect 1368 * the 4M alignments of the allocations above on V9 SPARCs... 1369 */ 1370 alloc_base = trap_trace_alloc(alloc_base); 1371 PRM_DEBUG(alloc_base); 1372 #endif /* TRAPTRACE */ 1373 1374 if (kmem64_base) { 1375 /* 1376 * Set the end of the kmem64 segment for V9 SPARCs, if 1377 * appropriate... 1378 */ 1379 kmem64_end = (caddr_t)roundup((uintptr_t)alloc_base, 1380 alloc_alignsize); 1381 1382 PRM_DEBUG(kmem64_base); 1383 PRM_DEBUG(kmem64_end); 1384 } 1385 1386 /* 1387 * Allocate space for the interrupt vector table and also for the 1388 * reserved interrupt vector data structures. 1389 */ 1390 memspace = (caddr_t)BOP_ALLOC(bootops, (caddr_t)intr_vec_table, 1391 IVSIZE, MMU_PAGESIZE); 1392 if (memspace != (caddr_t)intr_vec_table) 1393 panic("interrupt vector table allocation failure"); 1394 1395 /* 1396 * The memory lists from boot are allocated from the heap arena 1397 * so that later they can be freed and/or reallocated. 1398 */ 1399 if (BOP_GETPROP(bootops, "extent", &memlist_sz) == -1) 1400 panic("could not retrieve property \"extent\""); 1401 1402 /* 1403 * Between now and when we finish copying in the memory lists, 1404 * allocations happen so the space gets fragmented and the 1405 * lists longer. Leave enough space for lists twice as long 1406 * as what boot says it has now; roundup to a pagesize. 1407 * Also add space for the final phys-avail copy in the fixup 1408 * routine. 1409 */ 1410 va = (caddr_t)(sysbase + PAGESIZE + PANICBUFSIZE + 1411 roundup(IVSIZE, MMU_PAGESIZE)); 1412 memlist_sz *= 4; 1413 memlist_sz = roundup(memlist_sz, MMU_PAGESIZE); 1414 memspace = (caddr_t)BOP_ALLOC(bootops, va, memlist_sz, BO_NO_ALIGN); 1415 if (memspace == NULL) 1416 halt("Boot allocation failed."); 1417 1418 memlist = (struct memlist *)memspace; 1419 memlist_end = (char *)memspace + memlist_sz; 1420 1421 PRM_DEBUG(memlist); 1422 PRM_DEBUG(memlist_end); 1423 PRM_DEBUG(sysbase); 1424 PRM_DEBUG(syslimit); 1425 1426 kernelheap_init((void *)sysbase, (void *)syslimit, 1427 (caddr_t)sysbase + PAGESIZE, NULL, NULL); 1428 1429 /* 1430 * Take the most current snapshot we can by calling mem-update. 1431 */ 1432 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1433 &boot_physavail, &boot_physavail_len, 1434 &boot_virtavail, &boot_virtavail_len); 1435 1436 /* 1437 * Remove the space used by BOP_ALLOC from the kernel heap 1438 * plus the area actually used by the OBP (if any) 1439 * ignoring virtual addresses in virt_avail, above syslimit. 1440 */ 1441 virt_avail = memlist; 1442 copy_memlist(boot_virtavail, boot_virtavail_len, &memlist); 1443 1444 for (cur = virt_avail; cur->next; cur = cur->next) { 1445 uint64_t range_base, range_size; 1446 1447 if ((range_base = cur->address + cur->size) < (uint64_t)sysbase) 1448 continue; 1449 if (range_base >= (uint64_t)syslimit) 1450 break; 1451 /* 1452 * Limit the range to end at syslimit. 1453 */ 1454 range_size = MIN(cur->next->address, 1455 (uint64_t)syslimit) - range_base; 1456 (void) vmem_xalloc(heap_arena, (size_t)range_size, PAGESIZE, 1457 0, 0, (void *)range_base, (void *)(range_base + range_size), 1458 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 1459 } 1460 1461 phys_avail = memlist; 1462 (void) copy_physavail(boot_physavail, boot_physavail_len, 1463 &memlist, 0, 0); 1464 1465 /* 1466 * Add any extra memory after e_text to the phys_avail list, as long 1467 * as there's at least a page to add. 1468 */ 1469 if (extra_etpg) 1470 memlist_add(va_to_pa(extra_etva), mmu_ptob(extra_etpg), 1471 &memlist, &phys_avail); 1472 1473 /* 1474 * Add any extra memory after e_data to the phys_avail list as long 1475 * as there's at least a page to add. Usually, there isn't any, 1476 * since extra HME blocks typically get allocated there first before 1477 * using RAM elsewhere. 1478 */ 1479 if ((nalloc_base = ndata_extra_base(&ndata, MMU_PAGESIZE)) == NULL) 1480 nalloc_base = nalloc_end; 1481 ndata_remain_sz = nalloc_end - nalloc_base; 1482 1483 if (ndata_remain_sz >= MMU_PAGESIZE) 1484 memlist_add(va_to_pa(nalloc_base), 1485 (uint64_t)ndata_remain_sz, &memlist, &phys_avail); 1486 1487 PRM_DEBUG(memlist); 1488 PRM_DEBUG(memlist_sz); 1489 PRM_DEBUG(memspace); 1490 1491 if ((caddr_t)memlist > (memspace + memlist_sz)) 1492 panic("memlist overflow"); 1493 1494 PRM_DEBUG(pp_base); 1495 PRM_DEBUG(memseg_base); 1496 PRM_DEBUG(npages); 1497 1498 /* 1499 * Initialize the page structures from the memory lists. 1500 */ 1501 kphysm_init(pp_base, memseg_base, npages, kpm_pp_base, kpm_npages); 1502 1503 availrmem_initial = availrmem = freemem; 1504 PRM_DEBUG(availrmem); 1505 1506 /* 1507 * Some of the locks depend on page_hashsz being set! 1508 * kmem_init() depends on this; so, keep it here. 1509 */ 1510 page_lock_init(); 1511 1512 /* 1513 * Initialize kernel memory allocator. 1514 */ 1515 kmem_init(); 1516 1517 /* 1518 * Initialize bp_mapin(). 1519 */ 1520 bp_init(shm_alignment, HAT_STRICTORDER); 1521 1522 /* 1523 * Reserve space for panicbuf, intr_vec_table and reserved interrupt 1524 * vector data structures from the 32-bit heap. 1525 */ 1526 (void) vmem_xalloc(heap32_arena, PANICBUFSIZE, PAGESIZE, 0, 0, 1527 panicbuf, panicbuf + PANICBUFSIZE, 1528 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 1529 1530 (void) vmem_xalloc(heap32_arena, IVSIZE, PAGESIZE, 0, 0, 1531 intr_vec_table, (caddr_t)intr_vec_table + IVSIZE, 1532 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 1533 1534 mem_config_init(); 1535 } 1536 1537 static void 1538 startup_modules(void) 1539 { 1540 int proplen, nhblk1, nhblk8; 1541 size_t nhblksz; 1542 pgcnt_t hblk_pages, pages_per_hblk; 1543 size_t hme8blk_sz, hme1blk_sz; 1544 1545 /* 1546 * Log any optional messages from the boot program 1547 */ 1548 proplen = (size_t)BOP_GETPROPLEN(bootops, "boot-message"); 1549 if (proplen > 0) { 1550 char *msg; 1551 size_t len = (size_t)proplen; 1552 1553 msg = kmem_zalloc(len, KM_SLEEP); 1554 (void) BOP_GETPROP(bootops, "boot-message", msg); 1555 cmn_err(CE_CONT, "?%s\n", msg); 1556 kmem_free(msg, len); 1557 } 1558 1559 /* 1560 * Let the platforms have a chance to change default 1561 * values before reading system file. 1562 */ 1563 if (&set_platform_defaults) 1564 set_platform_defaults(); 1565 1566 /* 1567 * Calculate default settings of system parameters based upon 1568 * maxusers, yet allow to be overridden via the /etc/system file. 1569 */ 1570 param_calc(0); 1571 1572 mod_setup(); 1573 1574 /* 1575 * If this is a positron, complain and halt. 1576 */ 1577 if (&iam_positron && iam_positron()) { 1578 cmn_err(CE_WARN, "This hardware platform is not supported" 1579 " by this release of Solaris.\n"); 1580 #ifdef DEBUG 1581 prom_enter_mon(); /* Type 'go' to resume */ 1582 cmn_err(CE_WARN, "Booting an unsupported platform.\n"); 1583 cmn_err(CE_WARN, "Booting with down-rev firmware.\n"); 1584 1585 #else /* DEBUG */ 1586 halt(0); 1587 #endif /* DEBUG */ 1588 } 1589 1590 /* 1591 * If we are running firmware that isn't 64-bit ready 1592 * then complain and halt. 1593 */ 1594 do_prom_version_check(); 1595 1596 /* 1597 * Initialize system parameters 1598 */ 1599 param_init(); 1600 1601 /* 1602 * maxmem is the amount of physical memory we're playing with. 1603 */ 1604 maxmem = physmem; 1605 1606 /* Set segkp limits. */ 1607 ncbase = (caddr_t)SEGDEBUGBASE; 1608 ncend = (caddr_t)SEGDEBUGBASE; 1609 1610 /* 1611 * Initialize the hat layer. 1612 */ 1613 hat_init(); 1614 1615 /* 1616 * Initialize segment management stuff. 1617 */ 1618 seg_init(); 1619 1620 /* 1621 * Create the va>tte handler, so the prom can understand 1622 * kernel translations. The handler is installed later, just 1623 * as we are about to take over the trap table from the prom. 1624 */ 1625 create_va_to_tte(); 1626 1627 /* 1628 * Load the forthdebugger (optional) 1629 */ 1630 forthdebug_init(); 1631 1632 /* 1633 * Create OBP node for console input callbacks 1634 * if it is needed. 1635 */ 1636 startup_create_io_node(); 1637 1638 if (modloadonly("fs", "specfs") == -1) 1639 halt("Can't load specfs"); 1640 1641 if (modloadonly("fs", "devfs") == -1) 1642 halt("Can't load devfs"); 1643 1644 if (modloadonly("misc", "swapgeneric") == -1) 1645 halt("Can't load swapgeneric"); 1646 1647 (void) modloadonly("sys", "lbl_edition"); 1648 1649 dispinit(); 1650 1651 /* 1652 * Infer meanings to the members of the idprom buffer. 1653 */ 1654 parse_idprom(); 1655 1656 /* Read cluster configuration data. */ 1657 clconf_init(); 1658 1659 setup_ddi(); 1660 1661 /* 1662 * Lets take this opportunity to load the root device. 1663 */ 1664 if (loadrootmodules() != 0) 1665 debug_enter("Can't load the root filesystem"); 1666 1667 /* 1668 * Load tod driver module for the tod part found on this system. 1669 * Recompute the cpu frequency/delays based on tod as tod part 1670 * tends to keep time more accurately. 1671 */ 1672 if (&load_tod_module) 1673 load_tod_module(); 1674 1675 /* 1676 * Allow platforms to load modules which might 1677 * be needed after bootops are gone. 1678 */ 1679 if (&load_platform_modules) 1680 load_platform_modules(); 1681 1682 setcpudelay(); 1683 1684 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1685 &boot_physavail, &boot_physavail_len, 1686 &boot_virtavail, &boot_virtavail_len); 1687 1688 bop_alloc_pages = size_virtalloc(boot_virtavail, boot_virtavail_len); 1689 1690 /* 1691 * Calculation and allocation of hmeblks needed to remap 1692 * the memory allocated by PROM till now: 1693 * 1694 * (1) calculate how much virtual memory has been bop_alloc'ed. 1695 * (2) roundup this memory to span of hme8blk, i.e. 64KB 1696 * (3) calculate number of hme8blk's needed to remap this memory 1697 * (4) calculate amount of memory that's consumed by these hme8blk's 1698 * (5) add memory calculated in steps (2) and (4) above. 1699 * (6) roundup this memory to span of hme8blk, i.e. 64KB 1700 * (7) calculate number of hme8blk's needed to remap this memory 1701 * (8) calculate amount of memory that's consumed by these hme8blk's 1702 * (9) allocate additional hme1blk's to hold large mappings. 1703 * H8TOH1 determines this. The current SWAG gives enough hblk1's 1704 * to remap everything with 4M mappings. 1705 * (10) account for partially used hblk8's due to non-64K aligned 1706 * PROM mapping entries. 1707 * (11) add memory calculated in steps (8), (9), and (10) above. 1708 * (12) kmem_zalloc the memory calculated in (11); since segkmem 1709 * is not ready yet, this gets bop_alloc'ed. 1710 * (13) there will be very few bop_alloc's after this point before 1711 * trap table takes over 1712 */ 1713 1714 /* sfmmu_init_nucleus_hblks expects properly aligned data structures. */ 1715 hme8blk_sz = roundup(HME8BLK_SZ, sizeof (int64_t)); 1716 hme1blk_sz = roundup(HME1BLK_SZ, sizeof (int64_t)); 1717 1718 pages_per_hblk = btop(HMEBLK_SPAN(TTE8K)); 1719 bop_alloc_pages = roundup(bop_alloc_pages, pages_per_hblk); 1720 nhblk8 = bop_alloc_pages / pages_per_hblk; 1721 nhblk1 = roundup(nhblk8, H8TOH1) / H8TOH1; 1722 hblk_pages = btopr(nhblk8 * hme8blk_sz + nhblk1 * hme1blk_sz); 1723 bop_alloc_pages += hblk_pages; 1724 bop_alloc_pages = roundup(bop_alloc_pages, pages_per_hblk); 1725 nhblk8 = bop_alloc_pages / pages_per_hblk; 1726 nhblk1 = roundup(nhblk8, H8TOH1) / H8TOH1; 1727 if (nhblk1 < hblk1_min) 1728 nhblk1 = hblk1_min; 1729 if (nhblk8 < hblk8_min) 1730 nhblk8 = hblk8_min; 1731 1732 /* 1733 * Since hblk8's can hold up to 64k of mappings aligned on a 64k 1734 * boundary, the number of hblk8's needed to map the entries in the 1735 * boot_virtavail list needs to be adjusted to take this into 1736 * consideration. Thus, we need to add additional hblk8's since it 1737 * is possible that an hblk8 will not have all 8 slots used due to 1738 * alignment constraints. Since there were boot_virtavail_len entries 1739 * in that list, we need to add that many hblk8's to the number 1740 * already calculated to make sure we don't underestimate. 1741 */ 1742 nhblk8 += boot_virtavail_len; 1743 nhblksz = nhblk8 * hme8blk_sz + nhblk1 * hme1blk_sz; 1744 1745 /* Allocate in pagesize chunks */ 1746 nhblksz = roundup(nhblksz, MMU_PAGESIZE); 1747 hblk_base = kmem_zalloc(nhblksz, KM_SLEEP); 1748 sfmmu_init_nucleus_hblks(hblk_base, nhblksz, nhblk8, nhblk1); 1749 } 1750 1751 static void 1752 startup_bop_gone(void) 1753 { 1754 extern int bop_io_quiesced; 1755 1756 /* 1757 * Destroy the MD initialized at startup 1758 * The startup initializes the MD framework 1759 * using prom and BOP alloc free it now. 1760 */ 1761 mach_descrip_startup_fini(); 1762 1763 /* 1764 * Call back into boot and release boots resources. 1765 */ 1766 BOP_QUIESCE_IO(bootops); 1767 bop_io_quiesced = 1; 1768 1769 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1770 &boot_physavail, &boot_physavail_len, 1771 &boot_virtavail, &boot_virtavail_len); 1772 /* 1773 * Copy physinstalled list into kernel space. 1774 */ 1775 phys_install = memlist; 1776 copy_memlist(boot_physinstalled, boot_physinstalled_len, &memlist); 1777 1778 /* 1779 * setup physically contiguous area twice as large as the ecache. 1780 * this is used while doing displacement flush of ecaches 1781 */ 1782 if (&ecache_flush_address) { 1783 ecache_flushaddr = ecache_flush_address(); 1784 if (ecache_flushaddr == (uint64_t)-1) { 1785 cmn_err(CE_PANIC, 1786 "startup: no memory to set ecache_flushaddr"); 1787 } 1788 } 1789 1790 /* 1791 * Virtual available next. 1792 */ 1793 ASSERT(virt_avail != NULL); 1794 memlist_free_list(virt_avail); 1795 virt_avail = memlist; 1796 copy_memlist(boot_virtavail, boot_virtavail_len, &memlist); 1797 1798 /* 1799 * Last chance to ask our booter questions .. 1800 */ 1801 } 1802 1803 1804 /* 1805 * startup_fixup_physavail - called from mach_sfmmu.c after the final 1806 * allocations have been performed. We can't call it in startup_bop_gone 1807 * since later operations can cause obp to allocate more memory. 1808 */ 1809 void 1810 startup_fixup_physavail(void) 1811 { 1812 struct memlist *cur; 1813 1814 /* 1815 * take the most current snapshot we can by calling mem-update 1816 */ 1817 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1818 &boot_physavail, &boot_physavail_len, 1819 &boot_virtavail, &boot_virtavail_len); 1820 1821 /* 1822 * Copy phys_avail list, again. 1823 * Both the kernel/boot and the prom have been allocating 1824 * from the original list we copied earlier. 1825 */ 1826 cur = memlist; 1827 (void) copy_physavail(boot_physavail, boot_physavail_len, 1828 &memlist, 0, 0); 1829 1830 /* 1831 * Add any extra memory after e_text we added to the phys_avail list 1832 * back to the old list. 1833 */ 1834 if (extra_etpg) 1835 memlist_add(va_to_pa(extra_etva), mmu_ptob(extra_etpg), 1836 &memlist, &cur); 1837 if (ndata_remain_sz >= MMU_PAGESIZE) 1838 memlist_add(va_to_pa(nalloc_base), 1839 (uint64_t)ndata_remain_sz, &memlist, &cur); 1840 1841 /* 1842 * There isn't any bounds checking on the memlist area 1843 * so ensure it hasn't overgrown. 1844 */ 1845 if ((caddr_t)memlist > (caddr_t)memlist_end) 1846 cmn_err(CE_PANIC, "startup: memlist size exceeded"); 1847 1848 /* 1849 * The kernel removes the pages that were allocated for it from 1850 * the freelist, but we now have to find any -extra- pages that 1851 * the prom has allocated for it's own book-keeping, and remove 1852 * them from the freelist too. sigh. 1853 */ 1854 fix_prom_pages(phys_avail, cur); 1855 1856 ASSERT(phys_avail != NULL); 1857 memlist_free_list(phys_avail); 1858 phys_avail = cur; 1859 1860 /* 1861 * We're done with boot. Just after this point in time, boot 1862 * gets unmapped, so we can no longer rely on its services. 1863 * Zero the bootops to indicate this fact. 1864 */ 1865 bootops = (struct bootops *)NULL; 1866 BOOTOPS_GONE(); 1867 } 1868 1869 static void 1870 startup_vm(void) 1871 { 1872 size_t i; 1873 struct segmap_crargs a; 1874 struct segkpm_crargs b; 1875 1876 uint64_t avmem; 1877 caddr_t va; 1878 pgcnt_t max_phys_segkp; 1879 int mnode; 1880 1881 extern int exec_lpg_disable, use_brk_lpg, use_stk_lpg, use_zmap_lpg; 1882 1883 /* 1884 * get prom's mappings, create hments for them and switch 1885 * to the kernel context. 1886 */ 1887 hat_kern_setup(); 1888 1889 /* 1890 * Take over trap table 1891 */ 1892 setup_trap_table(); 1893 1894 /* 1895 * Install the va>tte handler, so that the prom can handle 1896 * misses and understand the kernel table layout in case 1897 * we need call into the prom. 1898 */ 1899 install_va_to_tte(); 1900 1901 /* 1902 * Set a flag to indicate that the tba has been taken over. 1903 */ 1904 tba_taken_over = 1; 1905 1906 /* initialize MMU primary context register */ 1907 mmu_init_kcontext(); 1908 1909 /* 1910 * The boot cpu can now take interrupts, x-calls, x-traps 1911 */ 1912 CPUSET_ADD(cpu_ready_set, CPU->cpu_id); 1913 CPU->cpu_flags |= (CPU_READY | CPU_ENABLE | CPU_EXISTS); 1914 1915 /* 1916 * Set a flag to tell write_scb_int() that it can access V_TBR_WR_ADDR. 1917 */ 1918 tbr_wr_addr_inited = 1; 1919 1920 /* 1921 * Initialize VM system, and map kernel address space. 1922 */ 1923 kvm_init(); 1924 1925 /* 1926 * XXX4U: previously, we initialized and turned on 1927 * the caches at this point. But of course we have 1928 * nothing to do, as the prom has already done this 1929 * for us -- main memory must be E$able at all times. 1930 */ 1931 1932 /* 1933 * If the following is true, someone has patched 1934 * phsymem to be less than the number of pages that 1935 * the system actually has. Remove pages until system 1936 * memory is limited to the requested amount. Since we 1937 * have allocated page structures for all pages, we 1938 * correct the amount of memory we want to remove 1939 * by the size of the memory used to hold page structures 1940 * for the non-used pages. 1941 */ 1942 if (physmem < npages) { 1943 pgcnt_t diff, off; 1944 struct page *pp; 1945 struct seg kseg; 1946 1947 cmn_err(CE_WARN, "limiting physmem to %ld pages", physmem); 1948 1949 off = 0; 1950 diff = npages - physmem; 1951 diff -= mmu_btopr(diff * sizeof (struct page)); 1952 kseg.s_as = &kas; 1953 while (diff--) { 1954 pp = page_create_va(&unused_pages_vp, (offset_t)off, 1955 MMU_PAGESIZE, PG_WAIT | PG_EXCL, 1956 &kseg, (caddr_t)off); 1957 if (pp == NULL) 1958 cmn_err(CE_PANIC, "limited physmem too much!"); 1959 page_io_unlock(pp); 1960 page_downgrade(pp); 1961 availrmem--; 1962 off += MMU_PAGESIZE; 1963 } 1964 } 1965 1966 /* 1967 * When printing memory, show the total as physmem less 1968 * that stolen by a debugger. 1969 */ 1970 cmn_err(CE_CONT, "?mem = %ldK (0x%lx000)\n", 1971 (ulong_t)(physinstalled) << (PAGESHIFT - 10), 1972 (ulong_t)(physinstalled) << (PAGESHIFT - 12)); 1973 1974 avmem = (uint64_t)freemem << PAGESHIFT; 1975 cmn_err(CE_CONT, "?avail mem = %lld\n", (unsigned long long)avmem); 1976 1977 /* For small memory systems disable automatic large pages. */ 1978 if (physmem < auto_lpg_min_physmem) { 1979 exec_lpg_disable = 1; 1980 use_brk_lpg = 0; 1981 use_stk_lpg = 0; 1982 use_zmap_lpg = 0; 1983 } 1984 1985 /* 1986 * Perform platform specific freelist processing 1987 */ 1988 if (&plat_freelist_process) { 1989 for (mnode = 0; mnode < max_mem_nodes; mnode++) 1990 if (mem_node_config[mnode].exists) 1991 plat_freelist_process(mnode); 1992 } 1993 1994 /* 1995 * Initialize the segkp segment type. We position it 1996 * after the configured tables and buffers (whose end 1997 * is given by econtig) and before V_WKBASE_ADDR. 1998 * Also in this area is segkmap (size SEGMAPSIZE). 1999 */ 2000 2001 /* XXX - cache alignment? */ 2002 va = (caddr_t)SEGKPBASE; 2003 ASSERT(((uintptr_t)va & PAGEOFFSET) == 0); 2004 2005 max_phys_segkp = (physmem * 2); 2006 2007 if (segkpsize < btop(SEGKPMINSIZE) || segkpsize > btop(SEGKPMAXSIZE)) { 2008 segkpsize = btop(SEGKPDEFSIZE); 2009 cmn_err(CE_WARN, "Illegal value for segkpsize. " 2010 "segkpsize has been reset to %ld pages", segkpsize); 2011 } 2012 2013 i = ptob(MIN(segkpsize, max_phys_segkp)); 2014 2015 rw_enter(&kas.a_lock, RW_WRITER); 2016 if (seg_attach(&kas, va, i, segkp) < 0) 2017 cmn_err(CE_PANIC, "startup: cannot attach segkp"); 2018 if (segkp_create(segkp) != 0) 2019 cmn_err(CE_PANIC, "startup: segkp_create failed"); 2020 rw_exit(&kas.a_lock); 2021 2022 /* 2023 * kpm segment 2024 */ 2025 segmap_kpm = kpm_enable && 2026 segmap_kpm && PAGESIZE == MAXBSIZE; 2027 2028 if (kpm_enable) { 2029 rw_enter(&kas.a_lock, RW_WRITER); 2030 2031 /* 2032 * The segkpm virtual range range is larger than the 2033 * actual physical memory size and also covers gaps in 2034 * the physical address range for the following reasons: 2035 * . keep conversion between segkpm and physical addresses 2036 * simple, cheap and unambiguous. 2037 * . avoid extension/shrink of the the segkpm in case of DR. 2038 * . avoid complexity for handling of virtual addressed 2039 * caches, segkpm and the regular mapping scheme must be 2040 * kept in sync wrt. the virtual color of mapped pages. 2041 * Any accesses to virtual segkpm ranges not backed by 2042 * physical memory will fall through the memseg pfn hash 2043 * and will be handled in segkpm_fault. 2044 * Additional kpm_size spaces needed for vac alias prevention. 2045 */ 2046 if (seg_attach(&kas, kpm_vbase, kpm_size * vac_colors, 2047 segkpm) < 0) 2048 cmn_err(CE_PANIC, "cannot attach segkpm"); 2049 2050 b.prot = PROT_READ | PROT_WRITE; 2051 b.nvcolors = shm_alignment >> MMU_PAGESHIFT; 2052 2053 if (segkpm_create(segkpm, (caddr_t)&b) != 0) 2054 panic("segkpm_create segkpm"); 2055 2056 rw_exit(&kas.a_lock); 2057 2058 mach_kpm_init(); 2059 } 2060 2061 /* 2062 * Now create generic mapping segment. This mapping 2063 * goes SEGMAPSIZE beyond SEGMAPBASE. But if the total 2064 * virtual address is greater than the amount of free 2065 * memory that is available, then we trim back the 2066 * segment size to that amount 2067 */ 2068 va = (caddr_t)SEGMAPBASE; 2069 2070 /* 2071 * 1201049: segkmap base address must be MAXBSIZE aligned 2072 */ 2073 ASSERT(((uintptr_t)va & MAXBOFFSET) == 0); 2074 2075 /* 2076 * Set size of segmap to percentage of freemem at boot, 2077 * but stay within the allowable range 2078 * Note we take percentage before converting from pages 2079 * to bytes to avoid an overflow on 32-bit kernels. 2080 */ 2081 i = mmu_ptob((freemem * segmap_percent) / 100); 2082 2083 if (i < MINMAPSIZE) 2084 i = MINMAPSIZE; 2085 2086 if (i > MIN(SEGMAPSIZE, mmu_ptob(freemem))) 2087 i = MIN(SEGMAPSIZE, mmu_ptob(freemem)); 2088 2089 i &= MAXBMASK; /* 1201049: segkmap size must be MAXBSIZE aligned */ 2090 2091 rw_enter(&kas.a_lock, RW_WRITER); 2092 if (seg_attach(&kas, va, i, segkmap) < 0) 2093 cmn_err(CE_PANIC, "cannot attach segkmap"); 2094 2095 a.prot = PROT_READ | PROT_WRITE; 2096 a.shmsize = shm_alignment; 2097 a.nfreelist = 0; /* use segmap driver defaults */ 2098 2099 if (segmap_create(segkmap, (caddr_t)&a) != 0) 2100 panic("segmap_create segkmap"); 2101 rw_exit(&kas.a_lock); 2102 2103 segdev_init(); 2104 } 2105 2106 static void 2107 startup_end(void) 2108 { 2109 if ((caddr_t)memlist > (caddr_t)memlist_end) 2110 panic("memlist overflow 2"); 2111 memlist_free_block((caddr_t)memlist, 2112 ((caddr_t)memlist_end - (caddr_t)memlist)); 2113 memlist = NULL; 2114 2115 /* enable page_relocation since OBP is now done */ 2116 page_relocate_ready = 1; 2117 2118 /* 2119 * Perform tasks that get done after most of the VM 2120 * initialization has been done but before the clock 2121 * and other devices get started. 2122 */ 2123 kern_setup1(); 2124 2125 /* 2126 * Intialize the VM arenas for allocating physically 2127 * contiguus memory chunk for interrupt queues snd 2128 * allocate/register boot cpu's queues, if any and 2129 * allocate dump buffer for sun4v systems to store 2130 * extra crash information during crash dump 2131 */ 2132 contig_mem_init(); 2133 mach_descrip_init(); 2134 cpu_intrq_setup(CPU); 2135 cpu_intrq_register(CPU); 2136 mach_htraptrace_setup(CPU->cpu_id); 2137 mach_htraptrace_configure(CPU->cpu_id); 2138 mach_dump_buffer_init(); 2139 2140 /* 2141 * Initialize interrupt related stuff 2142 */ 2143 cpu_intr_alloc(CPU, NINTR_THREADS); 2144 2145 (void) splzs(); /* allow hi clock ints but not zs */ 2146 2147 /* 2148 * Initialize errors. 2149 */ 2150 error_init(); 2151 2152 /* 2153 * Note that we may have already used kernel bcopy before this 2154 * point - but if you really care about this, adb the use_hw_* 2155 * variables to 0 before rebooting. 2156 */ 2157 mach_hw_copy_limit(); 2158 2159 /* 2160 * Install the "real" preemption guards before DDI services 2161 * are available. 2162 */ 2163 (void) prom_set_preprom(kern_preprom); 2164 (void) prom_set_postprom(kern_postprom); 2165 CPU->cpu_m.mutex_ready = 1; 2166 2167 /* 2168 * Initialize segnf (kernel support for non-faulting loads). 2169 */ 2170 segnf_init(); 2171 2172 /* 2173 * Configure the root devinfo node. 2174 */ 2175 configure(); /* set up devices */ 2176 mach_cpu_halt_idle(); 2177 } 2178 2179 2180 void 2181 post_startup(void) 2182 { 2183 #ifdef PTL1_PANIC_DEBUG 2184 extern void init_ptl1_thread(void); 2185 #endif /* PTL1_PANIC_DEBUG */ 2186 extern void abort_sequence_init(void); 2187 2188 /* 2189 * Set the system wide, processor-specific flags to be passed 2190 * to userland via the aux vector for performance hints and 2191 * instruction set extensions. 2192 */ 2193 bind_hwcap(); 2194 2195 /* 2196 * Startup memory scrubber (if any) 2197 */ 2198 mach_memscrub(); 2199 2200 /* 2201 * Allocate soft interrupt to handle abort sequence. 2202 */ 2203 abort_sequence_init(); 2204 2205 /* 2206 * Configure the rest of the system. 2207 * Perform forceloading tasks for /etc/system. 2208 */ 2209 (void) mod_sysctl(SYS_FORCELOAD, NULL); 2210 /* 2211 * ON4.0: Force /proc module in until clock interrupt handle fixed 2212 * ON4.0: This must be fixed or restated in /etc/systems. 2213 */ 2214 (void) modload("fs", "procfs"); 2215 2216 /* load machine class specific drivers */ 2217 load_mach_drivers(); 2218 2219 /* load platform specific drivers */ 2220 if (&load_platform_drivers) 2221 load_platform_drivers(); 2222 2223 /* load vis simulation module, if we are running w/fpu off */ 2224 if (!fpu_exists) { 2225 if (modload("misc", "vis") == -1) 2226 halt("Can't load vis"); 2227 } 2228 2229 mach_fpras(); 2230 2231 maxmem = freemem; 2232 2233 #ifdef PTL1_PANIC_DEBUG 2234 init_ptl1_thread(); 2235 #endif /* PTL1_PANIC_DEBUG */ 2236 2237 if (&cif_init) 2238 cif_init(); 2239 } 2240 2241 #ifdef PTL1_PANIC_DEBUG 2242 int ptl1_panic_test = 0; 2243 int ptl1_panic_xc_one_test = 0; 2244 int ptl1_panic_xc_all_test = 0; 2245 int ptl1_panic_xt_one_test = 0; 2246 int ptl1_panic_xt_all_test = 0; 2247 kthread_id_t ptl1_thread_p = NULL; 2248 kcondvar_t ptl1_cv; 2249 kmutex_t ptl1_mutex; 2250 int ptl1_recurse_count_threshold = 0x40; 2251 int ptl1_recurse_trap_threshold = 0x3d; 2252 extern void ptl1_recurse(int, int); 2253 extern void ptl1_panic_xt(int, int); 2254 2255 /* 2256 * Called once per second by timeout() to wake up 2257 * the ptl1_panic thread to see if it should cause 2258 * a trap to the ptl1_panic() code. 2259 */ 2260 /* ARGSUSED */ 2261 static void 2262 ptl1_wakeup(void *arg) 2263 { 2264 mutex_enter(&ptl1_mutex); 2265 cv_signal(&ptl1_cv); 2266 mutex_exit(&ptl1_mutex); 2267 } 2268 2269 /* 2270 * ptl1_panic cross call function: 2271 * Needed because xc_one() and xc_some() can pass 2272 * 64 bit args but ptl1_recurse() expects ints. 2273 */ 2274 static void 2275 ptl1_panic_xc(void) 2276 { 2277 ptl1_recurse(ptl1_recurse_count_threshold, 2278 ptl1_recurse_trap_threshold); 2279 } 2280 2281 /* 2282 * The ptl1 thread waits for a global flag to be set 2283 * and uses the recurse thresholds to set the stack depth 2284 * to cause a ptl1_panic() directly via a call to ptl1_recurse 2285 * or indirectly via the cross call and cross trap functions. 2286 * 2287 * This is useful testing stack overflows and normal 2288 * ptl1_panic() states with a know stack frame. 2289 * 2290 * ptl1_recurse() is an asm function in ptl1_panic.s that 2291 * sets the {In, Local, Out, and Global} registers to a 2292 * know state on the stack and just prior to causing a 2293 * test ptl1_panic trap. 2294 */ 2295 static void 2296 ptl1_thread(void) 2297 { 2298 mutex_enter(&ptl1_mutex); 2299 while (ptl1_thread_p) { 2300 cpuset_t other_cpus; 2301 int cpu_id; 2302 int my_cpu_id; 2303 int target_cpu_id; 2304 int target_found; 2305 2306 if (ptl1_panic_test) { 2307 ptl1_recurse(ptl1_recurse_count_threshold, 2308 ptl1_recurse_trap_threshold); 2309 } 2310 2311 /* 2312 * Find potential targets for x-call and x-trap, 2313 * if any exist while preempt is disabled we 2314 * start a ptl1_panic if requested via a 2315 * globals. 2316 */ 2317 kpreempt_disable(); 2318 my_cpu_id = CPU->cpu_id; 2319 other_cpus = cpu_ready_set; 2320 CPUSET_DEL(other_cpus, CPU->cpu_id); 2321 target_found = 0; 2322 if (!CPUSET_ISNULL(other_cpus)) { 2323 /* 2324 * Pick the first one 2325 */ 2326 for (cpu_id = 0; cpu_id < NCPU; cpu_id++) { 2327 if (cpu_id == my_cpu_id) 2328 continue; 2329 2330 if (CPU_XCALL_READY(cpu_id)) { 2331 target_cpu_id = cpu_id; 2332 target_found = 1; 2333 break; 2334 } 2335 } 2336 ASSERT(target_found); 2337 2338 if (ptl1_panic_xc_one_test) { 2339 xc_one(target_cpu_id, 2340 (xcfunc_t *)ptl1_panic_xc, 0, 0); 2341 } 2342 if (ptl1_panic_xc_all_test) { 2343 xc_some(other_cpus, 2344 (xcfunc_t *)ptl1_panic_xc, 0, 0); 2345 } 2346 if (ptl1_panic_xt_one_test) { 2347 xt_one(target_cpu_id, 2348 (xcfunc_t *)ptl1_panic_xt, 0, 0); 2349 } 2350 if (ptl1_panic_xt_all_test) { 2351 xt_some(other_cpus, 2352 (xcfunc_t *)ptl1_panic_xt, 0, 0); 2353 } 2354 } 2355 kpreempt_enable(); 2356 (void) timeout(ptl1_wakeup, NULL, hz); 2357 (void) cv_wait(&ptl1_cv, &ptl1_mutex); 2358 } 2359 mutex_exit(&ptl1_mutex); 2360 } 2361 2362 /* 2363 * Called during early startup to create the ptl1_thread 2364 */ 2365 void 2366 init_ptl1_thread(void) 2367 { 2368 ptl1_thread_p = thread_create(NULL, 0, ptl1_thread, NULL, 0, 2369 &p0, TS_RUN, 0); 2370 } 2371 #endif /* PTL1_PANIC_DEBUG */ 2372 2373 2374 /* 2375 * Add to a memory list. 2376 * start = start of new memory segment 2377 * len = length of new memory segment in bytes 2378 * memlistp = pointer to array of available memory segment structures 2379 * curmemlistp = memory list to which to add segment. 2380 */ 2381 static void 2382 memlist_add(uint64_t start, uint64_t len, struct memlist **memlistp, 2383 struct memlist **curmemlistp) 2384 { 2385 struct memlist *new; 2386 2387 new = *memlistp; 2388 new->address = start; 2389 new->size = len; 2390 *memlistp = new + 1; 2391 2392 memlist_insert(new, curmemlistp); 2393 } 2394 2395 /* 2396 * In the case of architectures that support dynamic addition of 2397 * memory at run-time there are two cases where memsegs need to 2398 * be initialized and added to the memseg list. 2399 * 1) memsegs that are constructed at startup. 2400 * 2) memsegs that are constructed at run-time on 2401 * hot-plug capable architectures. 2402 * This code was originally part of the function kphysm_init(). 2403 */ 2404 2405 static void 2406 memseg_list_add(struct memseg *memsegp) 2407 { 2408 struct memseg **prev_memsegp; 2409 pgcnt_t num; 2410 2411 /* insert in memseg list, decreasing number of pages order */ 2412 2413 num = MSEG_NPAGES(memsegp); 2414 2415 for (prev_memsegp = &memsegs; *prev_memsegp; 2416 prev_memsegp = &((*prev_memsegp)->next)) { 2417 if (num > MSEG_NPAGES(*prev_memsegp)) 2418 break; 2419 } 2420 2421 memsegp->next = *prev_memsegp; 2422 *prev_memsegp = memsegp; 2423 2424 if (kpm_enable) { 2425 memsegp->nextpa = (memsegp->next) ? 2426 va_to_pa(memsegp->next) : MSEG_NULLPTR_PA; 2427 2428 if (prev_memsegp != &memsegs) { 2429 struct memseg *msp; 2430 msp = (struct memseg *)((caddr_t)prev_memsegp - 2431 offsetof(struct memseg, next)); 2432 msp->nextpa = va_to_pa(memsegp); 2433 } else { 2434 memsegspa = va_to_pa(memsegs); 2435 } 2436 } 2437 } 2438 2439 /* 2440 * PSM add_physmem_cb(). US-II and newer processors have some 2441 * flavor of the prefetch capability implemented. We exploit 2442 * this capability for optimum performance. 2443 */ 2444 #define PREFETCH_BYTES 64 2445 2446 void 2447 add_physmem_cb(page_t *pp, pfn_t pnum) 2448 { 2449 extern void prefetch_page_w(void *); 2450 2451 pp->p_pagenum = pnum; 2452 2453 /* 2454 * Prefetch one more page_t into E$. To prevent future 2455 * mishaps with the sizeof(page_t) changing on us, we 2456 * catch this on debug kernels if we can't bring in the 2457 * entire hpage with 2 PREFETCH_BYTES reads. See 2458 * also, sun4u/cpu/cpu_module.c 2459 */ 2460 /*LINTED*/ 2461 ASSERT(sizeof (page_t) <= 2*PREFETCH_BYTES); 2462 prefetch_page_w((char *)pp); 2463 } 2464 2465 /* 2466 * kphysm_init() tackles the problem of initializing physical memory. 2467 * The old startup made some assumptions about the kernel living in 2468 * physically contiguous space which is no longer valid. 2469 */ 2470 static void 2471 kphysm_init(page_t *pp, struct memseg *memsegp, pgcnt_t npages, 2472 uintptr_t kpm_pp, pgcnt_t kpm_npages) 2473 { 2474 struct memlist *pmem; 2475 struct memseg *msp; 2476 pfn_t base; 2477 pgcnt_t num; 2478 pfn_t lastseg_pages_end = 0; 2479 pgcnt_t nelem_used = 0; 2480 2481 ASSERT(page_hash != NULL && page_hashsz != 0); 2482 2483 msp = memsegp; 2484 for (pmem = phys_avail; pmem && npages; pmem = pmem->next) { 2485 2486 /* 2487 * Build the memsegs entry 2488 */ 2489 num = btop(pmem->size); 2490 if (num > npages) 2491 num = npages; 2492 npages -= num; 2493 base = btop(pmem->address); 2494 2495 msp->pages = pp; 2496 msp->epages = pp + num; 2497 msp->pages_base = base; 2498 msp->pages_end = base + num; 2499 2500 if (kpm_enable) { 2501 pfn_t pbase_a; 2502 pfn_t pend_a; 2503 pfn_t prev_pend_a; 2504 pgcnt_t nelem; 2505 2506 msp->pagespa = va_to_pa(pp); 2507 msp->epagespa = va_to_pa(pp + num); 2508 pbase_a = kpmptop(ptokpmp(base)); 2509 pend_a = kpmptop(ptokpmp(base + num - 1)) + kpmpnpgs; 2510 nelem = ptokpmp(pend_a - pbase_a); 2511 msp->kpm_nkpmpgs = nelem; 2512 msp->kpm_pbase = pbase_a; 2513 if (lastseg_pages_end) { 2514 /* 2515 * Assume phys_avail is in ascending order 2516 * of physical addresses. 2517 */ 2518 ASSERT(base + num > lastseg_pages_end); 2519 prev_pend_a = kpmptop( 2520 ptokpmp(lastseg_pages_end - 1)) + kpmpnpgs; 2521 2522 if (prev_pend_a > pbase_a) { 2523 /* 2524 * Overlap, more than one memseg may 2525 * point to the same kpm_page range. 2526 */ 2527 if (kpm_smallpages == 0) { 2528 msp->kpm_pages = 2529 (kpm_page_t *)kpm_pp - 1; 2530 kpm_pp = (uintptr_t) 2531 ((kpm_page_t *)kpm_pp 2532 + nelem - 1); 2533 } else { 2534 msp->kpm_spages = 2535 (kpm_spage_t *)kpm_pp - 1; 2536 kpm_pp = (uintptr_t) 2537 ((kpm_spage_t *)kpm_pp 2538 + nelem - 1); 2539 } 2540 nelem_used += nelem - 1; 2541 2542 } else { 2543 if (kpm_smallpages == 0) { 2544 msp->kpm_pages = 2545 (kpm_page_t *)kpm_pp; 2546 kpm_pp = (uintptr_t) 2547 ((kpm_page_t *)kpm_pp 2548 + nelem); 2549 } else { 2550 msp->kpm_spages = 2551 (kpm_spage_t *)kpm_pp; 2552 kpm_pp = (uintptr_t) 2553 ((kpm_spage_t *) 2554 kpm_pp + nelem); 2555 } 2556 nelem_used += nelem; 2557 } 2558 2559 } else { 2560 if (kpm_smallpages == 0) { 2561 msp->kpm_pages = (kpm_page_t *)kpm_pp; 2562 kpm_pp = (uintptr_t) 2563 ((kpm_page_t *)kpm_pp + nelem); 2564 } else { 2565 msp->kpm_spages = (kpm_spage_t *)kpm_pp; 2566 kpm_pp = (uintptr_t) 2567 ((kpm_spage_t *)kpm_pp + nelem); 2568 } 2569 nelem_used = nelem; 2570 } 2571 2572 if (nelem_used > kpm_npages) 2573 panic("kphysm_init: kpm_pp overflow\n"); 2574 2575 msp->kpm_pagespa = va_to_pa(msp->kpm_pages); 2576 lastseg_pages_end = msp->pages_end; 2577 } 2578 2579 memseg_list_add(msp); 2580 2581 /* 2582 * add_physmem() initializes the PSM part of the page 2583 * struct by calling the PSM back with add_physmem_cb(). 2584 * In addition it coalesces pages into larger pages as 2585 * it initializes them. 2586 */ 2587 add_physmem(pp, num, base); 2588 pp += num; 2589 msp++; 2590 } 2591 2592 build_pfn_hash(); 2593 } 2594 2595 /* 2596 * Kernel VM initialization. 2597 * Assumptions about kernel address space ordering: 2598 * (1) gap (user space) 2599 * (2) kernel text 2600 * (3) kernel data/bss 2601 * (4) gap 2602 * (5) kernel data structures 2603 * (6) gap 2604 * (7) debugger (optional) 2605 * (8) monitor 2606 * (9) gap (possibly null) 2607 * (10) dvma 2608 * (11) devices 2609 */ 2610 static void 2611 kvm_init(void) 2612 { 2613 /* 2614 * Put the kernel segments in kernel address space. 2615 */ 2616 rw_enter(&kas.a_lock, RW_WRITER); 2617 as_avlinit(&kas); 2618 2619 (void) seg_attach(&kas, (caddr_t)KERNELBASE, 2620 (size_t)(e_moddata - KERNELBASE), &ktextseg); 2621 (void) segkmem_create(&ktextseg); 2622 2623 (void) seg_attach(&kas, (caddr_t)(KERNELBASE + MMU_PAGESIZE4M), 2624 (size_t)(MMU_PAGESIZE4M), &ktexthole); 2625 (void) segkmem_create(&ktexthole); 2626 2627 (void) seg_attach(&kas, (caddr_t)valloc_base, 2628 (size_t)(econtig32 - valloc_base), &kvalloc); 2629 (void) segkmem_create(&kvalloc); 2630 2631 if (kmem64_base) { 2632 (void) seg_attach(&kas, (caddr_t)kmem64_base, 2633 (size_t)(kmem64_end - kmem64_base), &kmem64); 2634 (void) segkmem_create(&kmem64); 2635 } 2636 2637 /* 2638 * We're about to map out /boot. This is the beginning of the 2639 * system resource management transition. We can no longer 2640 * call into /boot for I/O or memory allocations. 2641 */ 2642 (void) seg_attach(&kas, kernelheap, ekernelheap - kernelheap, &kvseg); 2643 (void) segkmem_create(&kvseg); 2644 hblk_alloc_dynamic = 1; 2645 2646 /* 2647 * we need to preallocate pages for DR operations before enabling large 2648 * page kernel heap because of memseg_remap_init() hat_unload() hack. 2649 */ 2650 memseg_remap_init(); 2651 2652 /* at this point we are ready to use large page heap */ 2653 segkmem_heap_lp_init(); 2654 2655 (void) seg_attach(&kas, (caddr_t)SYSBASE32, SYSLIMIT32 - SYSBASE32, 2656 &kvseg32); 2657 (void) segkmem_create(&kvseg32); 2658 2659 /* 2660 * Create a segment for the debugger. 2661 */ 2662 (void) seg_attach(&kas, (caddr_t)SEGDEBUGBASE, (size_t)SEGDEBUGSIZE, 2663 &kdebugseg); 2664 (void) segkmem_create(&kdebugseg); 2665 2666 rw_exit(&kas.a_lock); 2667 } 2668 2669 char obp_tte_str[] = 2670 "h# %x constant MMU_PAGESHIFT " 2671 "h# %x constant TTE8K " 2672 "h# %x constant SFHME_SIZE " 2673 "h# %x constant SFHME_TTE " 2674 "h# %x constant HMEBLK_TAG " 2675 "h# %x constant HMEBLK_NEXT " 2676 "h# %x constant HMEBLK_MISC " 2677 "h# %x constant HMEBLK_HME1 " 2678 "h# %x constant NHMENTS " 2679 "h# %x constant HBLK_SZMASK " 2680 "h# %x constant HBLK_RANGE_SHIFT " 2681 "h# %x constant HMEBP_HBLK " 2682 "h# %x constant HMEBUCKET_SIZE " 2683 "h# %x constant HTAG_SFMMUPSZ " 2684 "h# %x constant HTAG_REHASHSZ " 2685 "h# %x constant mmu_hashcnt " 2686 "h# %p constant uhme_hash " 2687 "h# %p constant khme_hash " 2688 "h# %x constant UHMEHASH_SZ " 2689 "h# %x constant KHMEHASH_SZ " 2690 "h# %p constant KCONTEXT " 2691 "h# %p constant KHATID " 2692 "h# %x constant ASI_MEM " 2693 2694 ": PHYS-X@ ( phys -- data ) " 2695 " ASI_MEM spacex@ " 2696 "; " 2697 2698 ": PHYS-W@ ( phys -- data ) " 2699 " ASI_MEM spacew@ " 2700 "; " 2701 2702 ": PHYS-L@ ( phys -- data ) " 2703 " ASI_MEM spaceL@ " 2704 "; " 2705 2706 ": TTE_PAGE_SHIFT ( ttesz -- hmeshift ) " 2707 " 3 * MMU_PAGESHIFT + " 2708 "; " 2709 2710 ": TTE_IS_VALID ( ttep -- flag ) " 2711 " PHYS-X@ 0< " 2712 "; " 2713 2714 ": HME_HASH_SHIFT ( ttesz -- hmeshift ) " 2715 " dup TTE8K = if " 2716 " drop HBLK_RANGE_SHIFT " 2717 " else " 2718 " TTE_PAGE_SHIFT " 2719 " then " 2720 "; " 2721 2722 ": HME_HASH_BSPAGE ( addr hmeshift -- bspage ) " 2723 " tuck >> swap MMU_PAGESHIFT - << " 2724 "; " 2725 2726 ": HME_HASH_FUNCTION ( sfmmup addr hmeshift -- hmebp ) " 2727 " >> over xor swap ( hash sfmmup ) " 2728 " KHATID <> if ( hash ) " 2729 " UHMEHASH_SZ and ( bucket ) " 2730 " HMEBUCKET_SIZE * uhme_hash + ( hmebp ) " 2731 " else ( hash ) " 2732 " KHMEHASH_SZ and ( bucket ) " 2733 " HMEBUCKET_SIZE * khme_hash + ( hmebp ) " 2734 " then ( hmebp ) " 2735 "; " 2736 2737 ": HME_HASH_TABLE_SEARCH " 2738 " ( sfmmup hmebp hblktag -- sfmmup null | sfmmup hmeblkp ) " 2739 " >r hmebp_hblk + phys-x@ begin ( sfmmup hmeblkp ) ( r: hblktag ) " 2740 " dup if ( sfmmup hmeblkp ) ( r: hblktag ) " 2741 " dup hmeblk_tag + phys-x@ r@ = if ( sfmmup hmeblkp ) " 2742 " dup hmeblk_tag + 8 + phys-x@ 2 pick = if " 2743 " true ( sfmmup hmeblkp true ) ( r: hblktag ) " 2744 " else " 2745 " hmeblk_next + phys-x@ false " 2746 " ( sfmmup hmeblkp false ) ( r: hblktag ) " 2747 " then " 2748 " else " 2749 " hmeblk_next + phys-x@ false " 2750 " ( sfmmup hmeblkp false ) ( r: hblktag ) " 2751 " then " 2752 " else " 2753 " true " 2754 " then " 2755 " until r> drop " 2756 "; " 2757 2758 ": HME_HASH_TAG ( sfmmup rehash addr -- hblktag ) " 2759 " over HME_HASH_SHIFT HME_HASH_BSPAGE ( sfmmup rehash bspage ) " 2760 " HTAG_REHASHSZ << or nip ( hblktag ) " 2761 "; " 2762 2763 ": HBLK_TO_TTEP ( hmeblkp addr -- ttep ) " 2764 " over HMEBLK_MISC + PHYS-L@ HBLK_SZMASK and ( hmeblkp addr ttesz ) " 2765 " TTE8K = if ( hmeblkp addr ) " 2766 " MMU_PAGESHIFT >> NHMENTS 1- and ( hmeblkp hme-index ) " 2767 " else ( hmeblkp addr ) " 2768 " drop 0 ( hmeblkp 0 ) " 2769 " then ( hmeblkp hme-index ) " 2770 " SFHME_SIZE * + HMEBLK_HME1 + ( hmep ) " 2771 " SFHME_TTE + ( ttep ) " 2772 "; " 2773 2774 ": unix-tte ( addr cnum -- false | tte-data true ) " 2775 " KCONTEXT = if ( addr ) " 2776 " KHATID ( addr khatid ) " 2777 " else ( addr ) " 2778 " drop false exit ( false ) " 2779 " then " 2780 " ( addr khatid ) " 2781 " mmu_hashcnt 1+ 1 do ( addr sfmmup ) " 2782 " 2dup swap i HME_HASH_SHIFT " 2783 "( addr sfmmup sfmmup addr hmeshift ) " 2784 " HME_HASH_FUNCTION ( addr sfmmup hmebp ) " 2785 " over i 4 pick " 2786 "( addr sfmmup hmebp sfmmup rehash addr ) " 2787 " HME_HASH_TAG ( addr sfmmup hmebp hblktag ) " 2788 " HME_HASH_TABLE_SEARCH " 2789 "( addr sfmmup { null | hmeblkp } ) " 2790 " ?dup if ( addr sfmmup hmeblkp ) " 2791 " nip swap HBLK_TO_TTEP ( ttep ) " 2792 " dup TTE_IS_VALID if ( valid-ttep ) " 2793 " PHYS-X@ true ( tte-data true ) " 2794 " else ( invalid-tte ) " 2795 " drop false ( false ) " 2796 " then ( false | tte-data true ) " 2797 " unloop exit ( false | tte-data true ) " 2798 " then ( addr sfmmup ) " 2799 " loop ( addr sfmmup ) " 2800 " 2drop false ( false ) " 2801 "; " 2802 ; 2803 2804 void 2805 create_va_to_tte(void) 2806 { 2807 char *bp; 2808 extern int khmehash_num, uhmehash_num; 2809 extern struct hmehash_bucket *khme_hash, *uhme_hash; 2810 2811 #define OFFSET(type, field) ((uintptr_t)(&((type *)0)->field)) 2812 2813 bp = (char *)kobj_zalloc(MMU_PAGESIZE, KM_SLEEP); 2814 2815 /* 2816 * Teach obp how to parse our sw ttes. 2817 */ 2818 (void) sprintf(bp, obp_tte_str, 2819 MMU_PAGESHIFT, 2820 TTE8K, 2821 sizeof (struct sf_hment), 2822 OFFSET(struct sf_hment, hme_tte), 2823 OFFSET(struct hme_blk, hblk_tag), 2824 OFFSET(struct hme_blk, hblk_nextpa), 2825 OFFSET(struct hme_blk, hblk_misc), 2826 OFFSET(struct hme_blk, hblk_hme), 2827 NHMENTS, 2828 HBLK_SZMASK, 2829 HBLK_RANGE_SHIFT, 2830 OFFSET(struct hmehash_bucket, hmeh_nextpa), 2831 sizeof (struct hmehash_bucket), 2832 HTAG_SFMMUPSZ, 2833 HTAG_REHASHSZ, 2834 mmu_hashcnt, 2835 (caddr_t)va_to_pa((caddr_t)uhme_hash), 2836 (caddr_t)va_to_pa((caddr_t)khme_hash), 2837 UHMEHASH_SZ, 2838 KHMEHASH_SZ, 2839 KCONTEXT, 2840 KHATID, 2841 ASI_MEM); 2842 prom_interpret(bp, 0, 0, 0, 0, 0); 2843 2844 kobj_free(bp, MMU_PAGESIZE); 2845 } 2846 2847 void 2848 install_va_to_tte(void) 2849 { 2850 /* 2851 * advise prom that he can use unix-tte 2852 */ 2853 prom_interpret("' unix-tte is va>tte-data", 0, 0, 0, 0, 0); 2854 } 2855 2856 2857 /* 2858 * Because kmdb links prom_stdout_is_framebuffer into its own 2859 * module, we add "device-type=display" here for /os-io node, so that 2860 * prom_stdout_is_framebuffer still works corrrectly after /os-io node 2861 * is registered into OBP. 2862 */ 2863 static char *create_node = 2864 "\" /\" find-device " 2865 "new-device " 2866 "\" os-io\" device-name " 2867 "\" display\" device-type " 2868 ": cb-r/w ( adr,len method$ -- #read/#written ) " 2869 " 2>r swap 2 2r> ['] $callback catch if " 2870 " 2drop 3drop 0 " 2871 " then " 2872 "; " 2873 ": read ( adr,len -- #read ) " 2874 " \" read\" ['] cb-r/w catch if 2drop 2drop -2 exit then " 2875 " ( retN ... ret1 N ) " 2876 " ?dup if " 2877 " swap >r 1- 0 ?do drop loop r> " 2878 " else " 2879 " -2 " 2880 " then " 2881 "; " 2882 ": write ( adr,len -- #written ) " 2883 " \" write\" ['] cb-r/w catch if 2drop 2drop 0 exit then " 2884 " ( retN ... ret1 N ) " 2885 " ?dup if " 2886 " swap >r 1- 0 ?do drop loop r> " 2887 " else " 2888 " 0 " 2889 " then " 2890 "; " 2891 ": poll-tty ( -- ) ; " 2892 ": install-abort ( -- ) ['] poll-tty d# 10 alarm ; " 2893 ": remove-abort ( -- ) ['] poll-tty 0 alarm ; " 2894 ": cb-give/take ( $method -- ) " 2895 " 0 -rot ['] $callback catch ?dup if " 2896 " >r 2drop 2drop r> throw " 2897 " else " 2898 " 0 ?do drop loop " 2899 " then " 2900 "; " 2901 ": give ( -- ) \" exit-input\" cb-give/take ; " 2902 ": take ( -- ) \" enter-input\" cb-give/take ; " 2903 ": open ( -- ok? ) true ; " 2904 ": close ( -- ) ; " 2905 "finish-device " 2906 "device-end "; 2907 2908 /* 2909 * Create the OBP input/output node (FCode serial driver). 2910 * It is needed for both USB console keyboard and for 2911 * the kernel terminal emulator. It is too early to check for a 2912 * kernel console compatible framebuffer now, so we create this 2913 * so that we're ready if we need to enable kernel terminal emulation. 2914 * 2915 * When the USB software takes over the input device at the time 2916 * consconfig runs, OBP's stdin is redirected to this node. 2917 * Whenever the FORTH user interface is used after this switch, 2918 * the node will call back into the kernel for console input. 2919 * If a serial device such as ttya or a UART with a Type 5 keyboard 2920 * attached is used, OBP takes over the serial device when the system 2921 * goes to the debugger after the system is booted. This sharing 2922 * of the relatively simple serial device is difficult but possible. 2923 * Sharing the USB host controller is impossible due its complexity. 2924 * 2925 * Similarly to USB keyboard input redirection, after consconfig_dacf 2926 * configures a kernel console framebuffer as the standard output 2927 * device, OBP's stdout is switched to to vector through the 2928 * /os-io node into the kernel terminal emulator. 2929 */ 2930 static void 2931 startup_create_io_node(void) 2932 { 2933 prom_interpret(create_node, 0, 0, 0, 0, 0); 2934 } 2935 2936 2937 static void 2938 do_prom_version_check(void) 2939 { 2940 int i; 2941 pnode_t node; 2942 char buf[64]; 2943 static char drev[] = "Down-rev firmware detected%s\n" 2944 "\tPlease upgrade to the following minimum version:\n" 2945 "\t\t%s\n"; 2946 2947 i = prom_version_check(buf, sizeof (buf), &node); 2948 2949 if (i == PROM_VER64_OK) 2950 return; 2951 2952 if (i == PROM_VER64_UPGRADE) { 2953 cmn_err(CE_WARN, drev, "", buf); 2954 2955 #ifdef DEBUG 2956 prom_enter_mon(); /* Type 'go' to continue */ 2957 cmn_err(CE_WARN, "Booting with down-rev firmware\n"); 2958 return; 2959 #else 2960 halt(0); 2961 #endif 2962 } 2963 2964 /* 2965 * The other possibility is that this is a server running 2966 * good firmware, but down-rev firmware was detected on at 2967 * least one other cpu board. We just complain if we see 2968 * that. 2969 */ 2970 cmn_err(CE_WARN, drev, " on one or more CPU boards", buf); 2971 } 2972 2973 static void 2974 kpm_init() 2975 { 2976 kpm_pgshft = (kpm_smallpages == 0) ? MMU_PAGESHIFT4M : MMU_PAGESHIFT; 2977 kpm_pgsz = 1ull << kpm_pgshft; 2978 kpm_pgoff = kpm_pgsz - 1; 2979 kpmp2pshft = kpm_pgshft - PAGESHIFT; 2980 kpmpnpgs = 1 << kpmp2pshft; 2981 ASSERT(((uintptr_t)kpm_vbase & (kpm_pgsz - 1)) == 0); 2982 } 2983 2984 void 2985 kpm_npages_setup(int memblocks) 2986 { 2987 /* 2988 * npages can be scattered in a maximum of 'memblocks' 2989 */ 2990 kpm_npages = ptokpmpr(npages) + memblocks; 2991 } 2992 2993 /* 2994 * Must be defined in platform dependent code. 2995 */ 2996 extern caddr_t modtext; 2997 extern size_t modtext_sz; 2998 extern caddr_t moddata; 2999 3000 #define HEAPTEXT_ARENA(addr) \ 3001 ((uintptr_t)(addr) < KERNELBASE + 2 * MMU_PAGESIZE4M ? 0 : \ 3002 (((uintptr_t)(addr) - HEAPTEXT_BASE) / \ 3003 (HEAPTEXT_MAPPED + HEAPTEXT_UNMAPPED) + 1)) 3004 3005 #define HEAPTEXT_OVERSIZED(addr) \ 3006 ((uintptr_t)(addr) >= HEAPTEXT_BASE + HEAPTEXT_SIZE - HEAPTEXT_OVERSIZE) 3007 3008 vmem_t *texthole_source[HEAPTEXT_NARENAS]; 3009 vmem_t *texthole_arena[HEAPTEXT_NARENAS]; 3010 kmutex_t texthole_lock; 3011 3012 char kern_bootargs[OBP_MAXPATHLEN]; 3013 3014 void 3015 kobj_vmem_init(vmem_t **text_arena, vmem_t **data_arena) 3016 { 3017 uintptr_t addr, limit; 3018 3019 addr = HEAPTEXT_BASE; 3020 limit = addr + HEAPTEXT_SIZE - HEAPTEXT_OVERSIZE; 3021 3022 /* 3023 * Before we initialize the text_arena, we want to punch holes in the 3024 * underlying heaptext_arena. This guarantees that for any text 3025 * address we can find a text hole less than HEAPTEXT_MAPPED away. 3026 */ 3027 for (; addr + HEAPTEXT_UNMAPPED <= limit; 3028 addr += HEAPTEXT_MAPPED + HEAPTEXT_UNMAPPED) { 3029 (void) vmem_xalloc(heaptext_arena, HEAPTEXT_UNMAPPED, PAGESIZE, 3030 0, 0, (void *)addr, (void *)(addr + HEAPTEXT_UNMAPPED), 3031 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 3032 } 3033 3034 /* 3035 * Allocate one page at the oversize to break up the text region 3036 * from the oversized region. 3037 */ 3038 (void) vmem_xalloc(heaptext_arena, PAGESIZE, PAGESIZE, 0, 0, 3039 (void *)limit, (void *)(limit + PAGESIZE), 3040 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 3041 3042 *text_arena = vmem_create("module_text", modtext, modtext_sz, 3043 sizeof (uintptr_t), segkmem_alloc, segkmem_free, 3044 heaptext_arena, 0, VM_SLEEP); 3045 *data_arena = vmem_create("module_data", moddata, MODDATA, 1, 3046 segkmem_alloc, segkmem_free, heap32_arena, 0, VM_SLEEP); 3047 } 3048 3049 caddr_t 3050 kobj_text_alloc(vmem_t *arena, size_t size) 3051 { 3052 caddr_t rval, better; 3053 3054 /* 3055 * First, try a sleeping allocation. 3056 */ 3057 rval = vmem_alloc(arena, size, VM_SLEEP | VM_BESTFIT); 3058 3059 if (size >= HEAPTEXT_MAPPED || !HEAPTEXT_OVERSIZED(rval)) 3060 return (rval); 3061 3062 /* 3063 * We didn't get the area that we wanted. We're going to try to do an 3064 * allocation with explicit constraints. 3065 */ 3066 better = vmem_xalloc(arena, size, sizeof (uintptr_t), 0, 0, NULL, 3067 (void *)(HEAPTEXT_BASE + HEAPTEXT_SIZE - HEAPTEXT_OVERSIZE), 3068 VM_NOSLEEP | VM_BESTFIT); 3069 3070 if (better != NULL) { 3071 /* 3072 * That worked. Free our first attempt and return. 3073 */ 3074 vmem_free(arena, rval, size); 3075 return (better); 3076 } 3077 3078 /* 3079 * That didn't work; we'll have to return our first attempt. 3080 */ 3081 return (rval); 3082 } 3083 3084 caddr_t 3085 kobj_texthole_alloc(caddr_t addr, size_t size) 3086 { 3087 int arena = HEAPTEXT_ARENA(addr); 3088 char c[30]; 3089 uintptr_t base; 3090 3091 if (HEAPTEXT_OVERSIZED(addr)) { 3092 /* 3093 * If this is an oversized allocation, there is no text hole 3094 * available for it; return NULL. 3095 */ 3096 return (NULL); 3097 } 3098 3099 mutex_enter(&texthole_lock); 3100 3101 if (texthole_arena[arena] == NULL) { 3102 ASSERT(texthole_source[arena] == NULL); 3103 3104 if (arena == 0) { 3105 texthole_source[0] = vmem_create("module_text_holesrc", 3106 (void *)(KERNELBASE + MMU_PAGESIZE4M), 3107 MMU_PAGESIZE4M, PAGESIZE, NULL, NULL, NULL, 3108 0, VM_SLEEP); 3109 } else { 3110 base = HEAPTEXT_BASE + 3111 (arena - 1) * (HEAPTEXT_MAPPED + HEAPTEXT_UNMAPPED); 3112 3113 (void) snprintf(c, sizeof (c), 3114 "heaptext_holesrc_%d", arena); 3115 3116 texthole_source[arena] = vmem_create(c, (void *)base, 3117 HEAPTEXT_UNMAPPED, PAGESIZE, NULL, NULL, NULL, 3118 0, VM_SLEEP); 3119 } 3120 3121 (void) snprintf(c, sizeof (c), "heaptext_hole_%d", arena); 3122 3123 texthole_arena[arena] = vmem_create(c, NULL, 0, 3124 sizeof (uint32_t), segkmem_alloc_permanent, segkmem_free, 3125 texthole_source[arena], 0, VM_SLEEP); 3126 } 3127 3128 mutex_exit(&texthole_lock); 3129 3130 ASSERT(texthole_arena[arena] != NULL); 3131 ASSERT(arena >= 0 && arena < HEAPTEXT_NARENAS); 3132 return (vmem_alloc(texthole_arena[arena], size, 3133 VM_BESTFIT | VM_NOSLEEP)); 3134 } 3135 3136 void 3137 kobj_texthole_free(caddr_t addr, size_t size) 3138 { 3139 int arena = HEAPTEXT_ARENA(addr); 3140 3141 ASSERT(arena >= 0 && arena < HEAPTEXT_NARENAS); 3142 ASSERT(texthole_arena[arena] != NULL); 3143 vmem_free(texthole_arena[arena], addr, size); 3144 } 3145