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 2007 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 int segzio_fromheap = 0; /* zio allocations occur from heap */ 196 caddr_t segzio_base; /* Base address of segzio */ 197 pgcnt_t segziosize = 0; /* size of zio segment in pages */ 198 199 /* 200 * debugger pages (if allocated) 201 */ 202 struct vnode kdebugvp; 203 204 /* 205 * VA range available to the debugger 206 */ 207 const caddr_t kdi_segdebugbase = (const caddr_t)SEGDEBUGBASE; 208 const size_t kdi_segdebugsize = SEGDEBUGSIZE; 209 210 /* 211 * Segment for relocated kernel structures in 64-bit large RAM kernels 212 */ 213 struct seg kmem64; 214 215 struct memseg *memseg_base; 216 size_t memseg_sz; /* Used to translate a va to page */ 217 struct vnode unused_pages_vp; 218 219 /* 220 * VM data structures allocated early during boot. 221 */ 222 size_t pagehash_sz; 223 uint64_t memlist_sz; 224 225 char tbr_wr_addr_inited = 0; 226 227 228 /* 229 * Static Routines: 230 */ 231 static void memlist_add(uint64_t, uint64_t, struct memlist **, 232 struct memlist **); 233 static void kphysm_init(page_t *, struct memseg *, pgcnt_t, uintptr_t, 234 pgcnt_t); 235 static void kvm_init(void); 236 237 static void startup_init(void); 238 static void startup_memlist(void); 239 static void startup_modules(void); 240 static void startup_bop_gone(void); 241 static void startup_vm(void); 242 static void startup_end(void); 243 static void setup_cage_params(void); 244 static void startup_create_io_node(void); 245 246 static pgcnt_t npages; 247 static struct memlist *memlist; 248 void *memlist_end; 249 250 static pgcnt_t bop_alloc_pages; 251 static caddr_t hblk_base; 252 uint_t hblk_alloc_dynamic = 0; 253 uint_t hblk1_min = H1MIN; 254 uint_t hblk8_min; 255 256 257 /* 258 * Hooks for unsupported platforms and down-rev firmware 259 */ 260 int iam_positron(void); 261 #pragma weak iam_positron 262 static void do_prom_version_check(void); 263 static void kpm_init(void); 264 static void kpm_npages_setup(int); 265 static void kpm_memseg_init(void); 266 267 /* 268 * After receiving a thermal interrupt, this is the number of seconds 269 * to delay before shutting off the system, assuming 270 * shutdown fails. Use /etc/system to change the delay if this isn't 271 * large enough. 272 */ 273 int thermal_powerdown_delay = 1200; 274 275 /* 276 * Used to hold off page relocations into the cage until OBP has completed 277 * its boot-time handoff of its resources to the kernel. 278 */ 279 int page_relocate_ready = 0; 280 281 /* 282 * Enable some debugging messages concerning memory usage... 283 */ 284 #ifdef DEBUGGING_MEM 285 static int debugging_mem; 286 static void 287 printmemlist(char *title, struct memlist *list) 288 { 289 if (!debugging_mem) 290 return; 291 292 printf("%s\n", title); 293 294 while (list) { 295 prom_printf("\taddr = 0x%x %8x, size = 0x%x %8x\n", 296 (uint32_t)(list->address >> 32), (uint32_t)list->address, 297 (uint32_t)(list->size >> 32), (uint32_t)(list->size)); 298 list = list->next; 299 } 300 } 301 302 void 303 printmemseg(struct memseg *memseg) 304 { 305 if (!debugging_mem) 306 return; 307 308 printf("memseg\n"); 309 310 while (memseg) { 311 prom_printf("\tpage = 0x%p, epage = 0x%p, " 312 "pfn = 0x%x, epfn = 0x%x\n", 313 memseg->pages, memseg->epages, 314 memseg->pages_base, memseg->pages_end); 315 memseg = memseg->next; 316 } 317 } 318 319 #define debug_pause(str) halt((str)) 320 #define MPRINTF(str) if (debugging_mem) prom_printf((str)) 321 #define MPRINTF1(str, a) if (debugging_mem) prom_printf((str), (a)) 322 #define MPRINTF2(str, a, b) if (debugging_mem) prom_printf((str), (a), (b)) 323 #define MPRINTF3(str, a, b, c) \ 324 if (debugging_mem) prom_printf((str), (a), (b), (c)) 325 #else /* DEBUGGING_MEM */ 326 #define MPRINTF(str) 327 #define MPRINTF1(str, a) 328 #define MPRINTF2(str, a, b) 329 #define MPRINTF3(str, a, b, c) 330 #endif /* DEBUGGING_MEM */ 331 332 /* Simple message to indicate that the bootops pointer has been zeroed */ 333 #ifdef DEBUG 334 static int bootops_gone_on = 0; 335 #define BOOTOPS_GONE() \ 336 if (bootops_gone_on) \ 337 prom_printf("The bootops vec is zeroed now!\n"); 338 #else 339 #define BOOTOPS_GONE() 340 #endif /* DEBUG */ 341 342 /* 343 * Monitor pages may not be where this says they are. 344 * and the debugger may not be there either. 345 * 346 * Note that 'pages' here are *physical* pages, which are 8k on sun4u. 347 * 348 * Physical memory layout 349 * (not necessarily contiguous) 350 * (THIS IS SOMEWHAT WRONG) 351 * /-----------------------\ 352 * | monitor pages | 353 * availmem -|-----------------------| 354 * | | 355 * | page pool | 356 * | | 357 * |-----------------------| 358 * | configured tables | 359 * | buffers | 360 * firstaddr -|-----------------------| 361 * | hat data structures | 362 * |-----------------------| 363 * | kernel data, bss | 364 * |-----------------------| 365 * | interrupt stack | 366 * |-----------------------| 367 * | kernel text (RO) | 368 * |-----------------------| 369 * | trap table (4k) | 370 * |-----------------------| 371 * page 1 | panicbuf | 372 * |-----------------------| 373 * page 0 | reclaimed | 374 * |_______________________| 375 * 376 * 377 * 378 * Kernel's Virtual Memory Layout. 379 * /-----------------------\ 380 * 0xFFFFFFFF.FFFFFFFF -| |- 381 * | OBP's virtual page | 382 * | tables | 383 * 0xFFFFFFFC.00000000 -|-----------------------|- 384 * : : 385 * : : 386 * -|-----------------------|- 387 * | segzio | (base and size vary) 388 * 0xFFFFFE00.00000000 -|-----------------------|- 389 * | | Ultrasparc I/II support 390 * | segkpm segment | up to 2TB of physical 391 * | (64-bit kernel ONLY) | memory, VAC has 2 colors 392 * | | 393 * 0xFFFFFA00.00000000 -|-----------------------|- 2TB segkpm alignment 394 * : : 395 * : : 396 * 0xFFFFF810.00000000 -|-----------------------|- hole_end 397 * | | ^ 398 * | UltraSPARC I/II call | | 399 * | bug requires an extra | | 400 * | 4 GB of space between | | 401 * | hole and used RAM | | 402 * | | | 403 * 0xFFFFF800.00000000 -|-----------------------|- | 404 * | | | 405 * | Virtual Address Hole | UltraSPARC 406 * | on UltraSPARC I/II | I/II * ONLY * 407 * | | | 408 * 0x00000800.00000000 -|-----------------------|- | 409 * | | | 410 * | UltraSPARC I/II call | | 411 * | bug requires an extra | | 412 * | 4 GB of space between | | 413 * | hole and used RAM | | 414 * | | v 415 * 0x000007FF.00000000 -|-----------------------|- hole_start ----- 416 * : : ^ 417 * : : | 418 * 0x00000XXX.XXXXXXXX -|-----------------------|- kmem64_end | 419 * | | | 420 * | 64-bit kernel ONLY | | 421 * | | | 422 * | kmem64 segment | | 423 * | | | 424 * | (Relocated extra HME | Approximately 425 * | block allocations, | 1 TB of virtual 426 * | memnode freelists, | address space 427 * | HME hash buckets, | | 428 * | mml_table, kpmp_table,| | 429 * | page_t array and | | 430 * | hashblock pool to | | 431 * | avoid hard-coded | | 432 * | 32-bit vaddr | | 433 * | limitations) | | 434 * | | v 435 * 0x00000700.00000000 -|-----------------------|- SYSLIMIT (kmem64_base) 436 * | | 437 * | segkmem segment | (SYSLIMIT - SYSBASE = 4TB) 438 * | | 439 * 0x00000300.00000000 -|-----------------------|- SYSBASE 440 * : : 441 * : : 442 * -|-----------------------|- 443 * | | 444 * | segmap segment | SEGMAPSIZE (1/8th physmem, 445 * | | 256G MAX) 446 * 0x000002a7.50000000 -|-----------------------|- SEGMAPBASE 447 * : : 448 * : : 449 * -|-----------------------|- 450 * | | 451 * | segkp | SEGKPSIZE (2GB) 452 * | | 453 * | | 454 * 0x000002a1.00000000 -|-----------------------|- SEGKPBASE 455 * | | 456 * 0x000002a0.00000000 -|-----------------------|- MEMSCRUBBASE 457 * | | (SEGKPBASE - 0x400000) 458 * 0x0000029F.FFE00000 -|-----------------------|- ARGSBASE 459 * | | (MEMSCRUBBASE - NCARGS) 460 * 0x0000029F.FFD80000 -|-----------------------|- PPMAPBASE 461 * | | (ARGSBASE - PPMAPSIZE) 462 * 0x0000029F.FFD00000 -|-----------------------|- PPMAP_FAST_BASE 463 * | | 464 * 0x0000029F.FF980000 -|-----------------------|- PIOMAPBASE 465 * | | 466 * 0x0000029F.FF580000 -|-----------------------|- NARG_BASE 467 * : : 468 * : : 469 * 0x00000000.FFFFFFFF -|-----------------------|- OFW_END_ADDR 470 * | | 471 * | OBP | 472 * | | 473 * 0x00000000.F0000000 -|-----------------------|- OFW_START_ADDR 474 * | kmdb | 475 * 0x00000000.EDD00000 -|-----------------------|- SEGDEBUGBASE 476 * : : 477 * : : 478 * 0x00000000.7c000000 -|-----------------------|- SYSLIMIT32 479 * | | 480 * | segkmem32 segment | (SYSLIMIT32 - SYSBASE32 = 481 * | | ~64MB) 482 * 0x00000000.78002000 -|-----------------------| 483 * | panicbuf | 484 * 0x00000000.78000000 -|-----------------------|- SYSBASE32 485 * : : 486 * : : 487 * | | 488 * |-----------------------|- econtig32 489 * | vm structures | 490 * 0x00000000.01C00000 |-----------------------|- nalloc_end 491 * | TSBs | 492 * |-----------------------|- end/nalloc_base 493 * | kernel data & bss | 494 * 0x00000000.01800000 -|-----------------------| 495 * : nucleus text hole : 496 * 0x00000000.01400000 -|-----------------------| 497 * : : 498 * |-----------------------| 499 * | module text | 500 * |-----------------------|- e_text/modtext 501 * | kernel text | 502 * |-----------------------| 503 * | trap table (48k) | 504 * 0x00000000.01000000 -|-----------------------|- KERNELBASE 505 * | reserved for trapstat |} TSTAT_TOTAL_SIZE 506 * |-----------------------| 507 * | | 508 * | invalid | 509 * | | 510 * 0x00000000.00000000 _|_______________________| 511 * 512 * 513 * 514 * 32-bit User Virtual Memory Layout. 515 * /-----------------------\ 516 * | | 517 * | invalid | 518 * | | 519 * 0xFFC00000 -|-----------------------|- USERLIMIT 520 * | user stack | 521 * : : 522 * : : 523 * : : 524 * | user data | 525 * -|-----------------------|- 526 * | user text | 527 * 0x00002000 -|-----------------------|- 528 * | invalid | 529 * 0x00000000 _|_______________________| 530 * 531 * 532 * 533 * 64-bit User Virtual Memory Layout. 534 * /-----------------------\ 535 * | | 536 * | invalid | 537 * | | 538 * 0xFFFFFFFF.80000000 -|-----------------------|- USERLIMIT 539 * | user stack | 540 * : : 541 * : : 542 * : : 543 * | user data | 544 * -|-----------------------|- 545 * | user text | 546 * 0x00000000.00100000 -|-----------------------|- 547 * | invalid | 548 * 0x00000000.00000000 _|_______________________| 549 */ 550 551 extern caddr_t ecache_init_scrub_flush_area(caddr_t alloc_base); 552 extern uint64_t ecache_flush_address(void); 553 554 #pragma weak load_platform_modules 555 #pragma weak plat_startup_memlist 556 #pragma weak ecache_init_scrub_flush_area 557 #pragma weak ecache_flush_address 558 559 560 /* 561 * By default the DR Cage is enabled for maximum OS 562 * MPSS performance. Users needing to disable the cage mechanism 563 * can set this variable to zero via /etc/system. 564 * Disabling the cage on systems supporting Dynamic Reconfiguration (DR) 565 * will result in loss of DR functionality. 566 * Platforms wishing to disable kernel Cage by default 567 * should do so in their set_platform_defaults() routine. 568 */ 569 int kernel_cage_enable = 1; 570 571 static void 572 setup_cage_params(void) 573 { 574 void (*func)(void); 575 576 func = (void (*)(void))kobj_getsymvalue("set_platform_cage_params", 0); 577 if (func != NULL) { 578 (*func)(); 579 return; 580 } 581 582 if (kernel_cage_enable == 0) { 583 return; 584 } 585 kcage_range_lock(); 586 if (kcage_range_init(phys_avail, 1) == 0) { 587 kcage_init(total_pages / 256); 588 } 589 kcage_range_unlock(); 590 591 if (kcage_on) { 592 cmn_err(CE_NOTE, "!Kernel Cage is ENABLED"); 593 } else { 594 cmn_err(CE_NOTE, "!Kernel Cage is DISABLED"); 595 } 596 597 } 598 599 /* 600 * Machine-dependent startup code 601 */ 602 void 603 startup(void) 604 { 605 startup_init(); 606 if (&startup_platform) 607 startup_platform(); 608 startup_memlist(); 609 startup_modules(); 610 setup_cage_params(); 611 startup_bop_gone(); 612 startup_vm(); 613 startup_end(); 614 } 615 616 struct regs sync_reg_buf; 617 uint64_t sync_tt; 618 619 void 620 sync_handler(void) 621 { 622 struct trap_info ti; 623 int i; 624 625 /* 626 * Prevent trying to talk to the other CPUs since they are 627 * sitting in the prom and won't reply. 628 */ 629 for (i = 0; i < NCPU; i++) { 630 if ((i != CPU->cpu_id) && CPU_XCALL_READY(i)) { 631 cpu[i]->cpu_flags &= ~CPU_READY; 632 cpu[i]->cpu_flags |= CPU_QUIESCED; 633 CPUSET_DEL(cpu_ready_set, cpu[i]->cpu_id); 634 } 635 } 636 637 /* 638 * We've managed to get here without going through the 639 * normal panic code path. Try and save some useful 640 * information. 641 */ 642 if (!panicstr && (curthread->t_panic_trap == NULL)) { 643 ti.trap_type = sync_tt; 644 ti.trap_regs = &sync_reg_buf; 645 ti.trap_addr = NULL; 646 ti.trap_mmu_fsr = 0x0; 647 648 curthread->t_panic_trap = &ti; 649 } 650 651 /* 652 * If we're re-entering the panic path, update the signature 653 * block so that the SC knows we're in the second part of panic. 654 */ 655 if (panicstr) 656 CPU_SIGNATURE(OS_SIG, SIGST_EXIT, SIGSUBST_DUMP, -1); 657 658 nopanicdebug = 1; /* do not perform debug_enter() prior to dump */ 659 panic("sync initiated"); 660 } 661 662 663 static void 664 startup_init(void) 665 { 666 /* 667 * We want to save the registers while we're still in OBP 668 * so that we know they haven't been fiddled with since. 669 * (In principle, OBP can't change them just because it 670 * makes a callback, but we'd rather not depend on that 671 * behavior.) 672 */ 673 char sync_str[] = 674 "warning @ warning off : sync " 675 "%%tl-c %%tstate h# %p x! " 676 "%%g1 h# %p x! %%g2 h# %p x! %%g3 h# %p x! " 677 "%%g4 h# %p x! %%g5 h# %p x! %%g6 h# %p x! " 678 "%%g7 h# %p x! %%o0 h# %p x! %%o1 h# %p x! " 679 "%%o2 h# %p x! %%o3 h# %p x! %%o4 h# %p x! " 680 "%%o5 h# %p x! %%o6 h# %p x! %%o7 h# %p x! " 681 "%%tl-c %%tpc h# %p x! %%tl-c %%tnpc h# %p x! " 682 "%%y h# %p l! %%tl-c %%tt h# %p x! " 683 "sync ; warning !"; 684 685 /* 686 * 20 == num of %p substrings 687 * 16 == max num of chars %p will expand to. 688 */ 689 char bp[sizeof (sync_str) + 16 * 20]; 690 691 (void) check_boot_version(BOP_GETVERSION(bootops)); 692 693 /* 694 * Initialize ptl1 stack for the 1st CPU. 695 */ 696 ptl1_init_cpu(&cpu0); 697 698 /* 699 * Initialize the address map for cache consistent mappings 700 * to random pages; must be done after vac_size is set. 701 */ 702 ppmapinit(); 703 704 /* 705 * Initialize the PROM callback handler. 706 */ 707 init_vx_handler(); 708 709 /* 710 * have prom call sync_callback() to handle the sync and 711 * save some useful information which will be stored in the 712 * core file later. 713 */ 714 (void) sprintf((char *)bp, sync_str, 715 (void *)&sync_reg_buf.r_tstate, (void *)&sync_reg_buf.r_g1, 716 (void *)&sync_reg_buf.r_g2, (void *)&sync_reg_buf.r_g3, 717 (void *)&sync_reg_buf.r_g4, (void *)&sync_reg_buf.r_g5, 718 (void *)&sync_reg_buf.r_g6, (void *)&sync_reg_buf.r_g7, 719 (void *)&sync_reg_buf.r_o0, (void *)&sync_reg_buf.r_o1, 720 (void *)&sync_reg_buf.r_o2, (void *)&sync_reg_buf.r_o3, 721 (void *)&sync_reg_buf.r_o4, (void *)&sync_reg_buf.r_o5, 722 (void *)&sync_reg_buf.r_o6, (void *)&sync_reg_buf.r_o7, 723 (void *)&sync_reg_buf.r_pc, (void *)&sync_reg_buf.r_npc, 724 (void *)&sync_reg_buf.r_y, (void *)&sync_tt); 725 prom_interpret(bp, 0, 0, 0, 0, 0); 726 add_vx_handler("sync", 1, (void (*)(cell_t *))sync_handler); 727 } 728 729 static u_longlong_t *boot_physinstalled, *boot_physavail, *boot_virtavail; 730 static size_t boot_physinstalled_len, boot_physavail_len, boot_virtavail_len; 731 732 #define IVSIZE ((MAXIVNUM * sizeof (intr_vec_t *)) + \ 733 (MAX_RSVD_IV * sizeof (intr_vec_t)) + \ 734 (MAX_RSVD_IVX * sizeof (intr_vecx_t))) 735 736 /* 737 * As OBP takes up some RAM when the system boots, pages will already be "lost" 738 * to the system and reflected in npages by the time we see it. 739 * 740 * We only want to allocate kernel structures in the 64-bit virtual address 741 * space on systems with enough RAM to make the overhead of keeping track of 742 * an extra kernel memory segment worthwhile. 743 * 744 * Since OBP has already performed its memory allocations by this point, if we 745 * have more than MINMOVE_RAM_MB MB of RAM left free, go ahead and map 746 * memory in the 64-bit virtual address space; otherwise keep allocations 747 * contiguous with we've mapped so far in the 32-bit virtual address space. 748 */ 749 #define MINMOVE_RAM_MB ((size_t)1900) 750 #define MB_TO_BYTES(mb) ((mb) * 1048576ul) 751 752 pgcnt_t tune_npages = (pgcnt_t) 753 (MB_TO_BYTES(MINMOVE_RAM_MB)/ (size_t)MMU_PAGESIZE); 754 755 static void 756 startup_memlist(void) 757 { 758 size_t alloc_sz; 759 size_t ctrs_sz; 760 caddr_t alloc_base; 761 caddr_t ctrs_base, ctrs_end; 762 caddr_t memspace; 763 caddr_t va; 764 int memblocks = 0; 765 struct memlist *cur; 766 size_t syslimit = (size_t)SYSLIMIT; 767 size_t sysbase = (size_t)SYSBASE; 768 int alloc_alignsize = MMU_PAGESIZE; 769 extern void page_coloring_init(void); 770 771 /* 772 * Initialize enough of the system to allow kmem_alloc to work by 773 * calling boot to allocate its memory until the time that 774 * kvm_init is completed. The page structs are allocated after 775 * rounding up end to the nearest page boundary; the memsegs are 776 * initialized and the space they use comes from the kernel heap. 777 * With appropriate initialization, they can be reallocated later 778 * to a size appropriate for the machine's configuration. 779 * 780 * At this point, memory is allocated for things that will never 781 * need to be freed, this used to be "valloced". This allows a 782 * savings as the pages don't need page structures to describe 783 * them because them will not be managed by the vm system. 784 */ 785 786 /* 787 * We're loaded by boot with the following configuration (as 788 * specified in the sun4u/conf/Mapfile): 789 * 790 * text: 4 MB chunk aligned on a 4MB boundary 791 * data & bss: 4 MB chunk aligned on a 4MB boundary 792 * 793 * These two chunks will eventually be mapped by 2 locked 4MB 794 * ttes and will represent the nucleus of the kernel. This gives 795 * us some free space that is already allocated, some or all of 796 * which is made available to kernel module text. 797 * 798 * The free space in the data-bss chunk is used for nucleus 799 * allocatable data structures and we reserve it using the 800 * nalloc_base and nalloc_end variables. This space is currently 801 * being used for hat data structures required for tlb miss 802 * handling operations. We align nalloc_base to a l2 cache 803 * linesize because this is the line size the hardware uses to 804 * maintain cache coherency. 805 * 256K is carved out for module data. 806 */ 807 808 nalloc_base = (caddr_t)roundup((uintptr_t)e_data, MMU_PAGESIZE); 809 moddata = nalloc_base; 810 e_moddata = nalloc_base + MODDATA; 811 nalloc_base = e_moddata; 812 813 nalloc_end = (caddr_t)roundup((uintptr_t)nalloc_base, MMU_PAGESIZE4M); 814 valloc_base = nalloc_base; 815 816 /* 817 * Calculate the start of the data segment. 818 */ 819 sdata = (caddr_t)((uintptr_t)e_data & MMU_PAGEMASK4M); 820 821 PRM_DEBUG(moddata); 822 PRM_DEBUG(nalloc_base); 823 PRM_DEBUG(nalloc_end); 824 PRM_DEBUG(sdata); 825 826 /* 827 * Remember any slop after e_text so we can give it to the modules. 828 */ 829 PRM_DEBUG(e_text); 830 modtext = (caddr_t)roundup((uintptr_t)e_text, MMU_PAGESIZE); 831 if (((uintptr_t)modtext & MMU_PAGEMASK4M) != (uintptr_t)s_text) 832 panic("nucleus text overflow"); 833 modtext_sz = (caddr_t)roundup((uintptr_t)modtext, MMU_PAGESIZE4M) - 834 modtext; 835 PRM_DEBUG(modtext); 836 PRM_DEBUG(modtext_sz); 837 838 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 839 &boot_physavail, &boot_physavail_len, 840 &boot_virtavail, &boot_virtavail_len); 841 /* 842 * Remember what the physically available highest page is 843 * so that dumpsys works properly, and find out how much 844 * memory is installed. 845 */ 846 installed_top_size_memlist_array(boot_physinstalled, 847 boot_physinstalled_len, &physmax, &physinstalled); 848 PRM_DEBUG(physinstalled); 849 PRM_DEBUG(physmax); 850 851 /* Fill out memory nodes config structure */ 852 startup_build_mem_nodes(boot_physinstalled, boot_physinstalled_len); 853 854 /* 855 * Get the list of physically available memory to size 856 * the number of page structures needed. 857 */ 858 size_physavail(boot_physavail, boot_physavail_len, &npages, &memblocks); 859 /* 860 * This first snap shot of npages can represent the pages used 861 * by OBP's text and data approximately. This is used in the 862 * the calculation of the kernel size 863 */ 864 obp_pages = physinstalled - npages; 865 866 867 /* 868 * On small-memory systems (<MODTEXT_SM_SIZE MB, currently 256MB), the 869 * in-nucleus module text is capped to MODTEXT_SM_CAP bytes (currently 870 * 2MB) and any excess pages are put on physavail. The assumption is 871 * that small-memory systems will need more pages more than they'll 872 * need efficiently-mapped module texts. 873 */ 874 if ((physinstalled < mmu_btop(MODTEXT_SM_SIZE << 20)) && 875 modtext_sz > MODTEXT_SM_CAP) { 876 extra_etpg = mmu_btop(modtext_sz - MODTEXT_SM_CAP); 877 modtext_sz = MODTEXT_SM_CAP; 878 } else 879 extra_etpg = 0; 880 PRM_DEBUG(extra_etpg); 881 PRM_DEBUG(modtext_sz); 882 extra_etva = modtext + modtext_sz; 883 PRM_DEBUG(extra_etva); 884 885 /* 886 * Account for any pages after e_text and e_data. 887 */ 888 npages += extra_etpg; 889 npages += mmu_btopr(nalloc_end - nalloc_base); 890 PRM_DEBUG(npages); 891 892 /* 893 * npages is the maximum of available physical memory possible. 894 * (ie. it will never be more than this) 895 */ 896 897 /* 898 * initialize the nucleus memory allocator. 899 */ 900 ndata_alloc_init(&ndata, (uintptr_t)nalloc_base, (uintptr_t)nalloc_end); 901 902 /* 903 * Allocate mmu fault status area from the nucleus data area. 904 */ 905 if ((&ndata_alloc_mmfsa != NULL) && (ndata_alloc_mmfsa(&ndata) != 0)) 906 cmn_err(CE_PANIC, "no more nucleus memory after mfsa alloc"); 907 908 /* 909 * Allocate kernel TSBs from the nucleus data area. 910 */ 911 if (ndata_alloc_tsbs(&ndata, npages) != 0) 912 cmn_err(CE_PANIC, "no more nucleus memory after tsbs alloc"); 913 914 /* 915 * Allocate dmv dispatch table from the nucleus data area. 916 */ 917 if (ndata_alloc_dmv(&ndata) != 0) 918 cmn_err(CE_PANIC, "no more nucleus memory after dmv alloc"); 919 920 921 page_coloring_init(); 922 923 /* 924 * Allocate page_freelists bin headers for memnode 0 from the 925 * nucleus data area. 926 */ 927 if (ndata_alloc_page_freelists(&ndata, 0) != 0) 928 cmn_err(CE_PANIC, 929 "no more nucleus memory after page free lists alloc"); 930 931 if (kpm_enable) { 932 kpm_init(); 933 /* 934 * kpm page space -- Update kpm_npages and make the 935 * same assumption about fragmenting as it is done 936 * for memseg_sz. 937 */ 938 kpm_npages_setup(memblocks + 4); 939 } 940 941 /* 942 * Allocate hat related structs from the nucleus data area. 943 */ 944 if (ndata_alloc_hat(&ndata, npages, kpm_npages) != 0) 945 cmn_err(CE_PANIC, "no more nucleus memory after hat alloc"); 946 947 /* 948 * We want to do the BOP_ALLOCs before the real allocation of page 949 * structs in order to not have to allocate page structs for this 950 * memory. We need to calculate a virtual address because we want 951 * the page structs to come before other allocations in virtual address 952 * space. This is so some (if not all) of page structs can actually 953 * live in the nucleus. 954 */ 955 956 /* 957 * WARNING WARNING WARNING WARNING WARNING WARNING WARNING 958 * 959 * There are comments all over the SFMMU code warning of dire 960 * consequences if the TSBs are moved out of 32-bit space. This 961 * is largely because the asm code uses "sethi %hi(addr)"-type 962 * instructions which will not provide the expected result if the 963 * address is a 64-bit one. 964 * 965 * WARNING WARNING WARNING WARNING WARNING WARNING WARNING 966 */ 967 alloc_base = (caddr_t)roundup((uintptr_t)nalloc_end, MMU_PAGESIZE); 968 alloc_base = sfmmu_ktsb_alloc(alloc_base); 969 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, ecache_alignsize); 970 PRM_DEBUG(alloc_base); 971 972 /* 973 * Allocate IOMMU TSB array. We do this here so that the physical 974 * memory gets deducted from the PROM's physical memory list. 975 */ 976 alloc_base = iommu_tsb_init(alloc_base); 977 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 978 ecache_alignsize); 979 PRM_DEBUG(alloc_base); 980 981 /* 982 * Platforms like Starcat and OPL need special structures assigned in 983 * 32-bit virtual address space because their probing routines execute 984 * FCode, and FCode can't handle 64-bit virtual addresses... 985 */ 986 if (&plat_startup_memlist) { 987 alloc_base = plat_startup_memlist(alloc_base); 988 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 989 ecache_alignsize); 990 PRM_DEBUG(alloc_base); 991 } 992 993 /* 994 * If we have enough memory, use 4M pages for alignment because it 995 * greatly reduces the number of TLB misses we take albeit at the cost 996 * of possible RAM wastage (degenerate case of 4 MB - MMU_PAGESIZE per 997 * allocation.) Still, the speedup on large memory systems (e.g. > 64 998 * GB) is quite noticeable, so it is worth the effort to do if we can. 999 * 1000 * Note, however, that this speedup will only occur if the boot PROM 1001 * uses the largest possible MMU page size possible to map memory 1002 * requests that are properly aligned and sized (for example, a request 1003 * for a multiple of 4MB of memory aligned to a 4MB boundary will 1004 * result in a mapping using a 4MB MMU page.) 1005 * 1006 * Even then, the large page mappings will only speed things up until 1007 * the startup process proceeds a bit further, as when 1008 * sfmmu_map_prom_mappings() copies page mappings from the PROM to the 1009 * kernel it remaps everything but the TSBs using 8K pages anyway... 1010 * 1011 * At some point in the future, sfmmu_map_prom_mappings() will be 1012 * rewritten to copy memory mappings to the kernel using the same MMU 1013 * page sizes the PROM used. When that occurs, if the PROM did use 1014 * large MMU pages to map memory, the alignment/sizing work we're 1015 * doing now should give us a nice extra performance boost, albeit at 1016 * the cost of greater RAM usage... 1017 */ 1018 alloc_alignsize = ((npages >= tune_npages) ? MMU_PAGESIZE4M : 1019 MMU_PAGESIZE); 1020 1021 PRM_DEBUG(tune_npages); 1022 PRM_DEBUG(alloc_alignsize); 1023 1024 /* 1025 * Save off where the contiguous allocations to date have ended 1026 * in econtig32. 1027 */ 1028 econtig32 = alloc_base; 1029 PRM_DEBUG(econtig32); 1030 1031 if (econtig32 > (caddr_t)KERNEL_LIMIT32) 1032 cmn_err(CE_PANIC, "econtig32 too big"); 1033 1034 /* 1035 * To avoid memory allocation collisions in the 32-bit virtual address 1036 * space, make allocations from this point forward in 64-bit virtual 1037 * address space starting at syslimit and working up. Also use the 1038 * alignment specified by alloc_alignsize, as we may be able to save 1039 * ourselves TLB misses by using larger page sizes if they're 1040 * available. 1041 * 1042 * All this is needed because on large memory systems, the default 1043 * Solaris allocations will collide with SYSBASE32, which is hard 1044 * coded to be at the virtual address 0x78000000. Therefore, on 64-bit 1045 * kernels, move the allocations to a location in the 64-bit virtual 1046 * address space space, allowing those structures to grow without 1047 * worry. 1048 * 1049 * On current CPUs we'll run out of physical memory address bits before 1050 * we need to worry about the allocations running into anything else in 1051 * VM or the virtual address holes on US-I and II, as there's currently 1052 * about 1 TB of addressable space before the US-I/II VA hole. 1053 */ 1054 kmem64_base = (caddr_t)syslimit; 1055 PRM_DEBUG(kmem64_base); 1056 1057 alloc_base = (caddr_t)roundup((uintptr_t)kmem64_base, alloc_alignsize); 1058 1059 /* 1060 * If KHME and/or UHME hash buckets won't fit in the nucleus, allocate 1061 * them here. 1062 */ 1063 if (khme_hash == NULL || uhme_hash == NULL) { 1064 /* 1065 * alloc_hme_buckets() will align alloc_base properly before 1066 * assigning the hash buckets, so we don't need to do it 1067 * before the call... 1068 */ 1069 alloc_base = alloc_hme_buckets(alloc_base, alloc_alignsize); 1070 1071 PRM_DEBUG(alloc_base); 1072 PRM_DEBUG(khme_hash); 1073 PRM_DEBUG(uhme_hash); 1074 } 1075 1076 /* 1077 * Allocate the remaining page freelists. NUMA systems can 1078 * have lots of page freelists, one per node, which quickly 1079 * outgrow the amount of nucleus memory available. 1080 */ 1081 if (max_mem_nodes > 1) { 1082 int mnode; 1083 caddr_t alloc_start = alloc_base; 1084 1085 for (mnode = 1; mnode < max_mem_nodes; mnode++) { 1086 alloc_base = alloc_page_freelists(mnode, alloc_base, 1087 ecache_alignsize); 1088 } 1089 1090 if (alloc_base > alloc_start) { 1091 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1092 alloc_alignsize); 1093 if ((caddr_t)BOP_ALLOC(bootops, alloc_start, 1094 alloc_base - alloc_start, 1095 alloc_alignsize) != alloc_start) 1096 cmn_err(CE_PANIC, 1097 "Unable to alloc page freelists\n"); 1098 } 1099 1100 PRM_DEBUG(alloc_base); 1101 } 1102 1103 if (!mml_table) { 1104 size_t mmltable_sz; 1105 1106 /* 1107 * We need to allocate the mml_table here because there 1108 * was not enough space within the nucleus. 1109 */ 1110 mmltable_sz = sizeof (kmutex_t) * mml_table_sz; 1111 alloc_sz = roundup(mmltable_sz, alloc_alignsize); 1112 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1113 alloc_alignsize); 1114 1115 if ((mml_table = (kmutex_t *)BOP_ALLOC(bootops, alloc_base, 1116 alloc_sz, alloc_alignsize)) != (kmutex_t *)alloc_base) 1117 panic("mml_table alloc failure"); 1118 1119 alloc_base += alloc_sz; 1120 PRM_DEBUG(mml_table); 1121 PRM_DEBUG(alloc_base); 1122 } 1123 1124 if (kpm_enable && !(kpmp_table || kpmp_stable)) { 1125 size_t kpmptable_sz; 1126 caddr_t table; 1127 1128 /* 1129 * We need to allocate either kpmp_table or kpmp_stable here 1130 * because there was not enough space within the nucleus. 1131 */ 1132 kpmptable_sz = (kpm_smallpages == 0) ? 1133 sizeof (kpm_hlk_t) * kpmp_table_sz : 1134 sizeof (kpm_shlk_t) * kpmp_stable_sz; 1135 1136 alloc_sz = roundup(kpmptable_sz, alloc_alignsize); 1137 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1138 alloc_alignsize); 1139 1140 table = BOP_ALLOC(bootops, alloc_base, alloc_sz, 1141 alloc_alignsize); 1142 1143 if (table != alloc_base) 1144 panic("kpmp_table or kpmp_stable alloc failure"); 1145 1146 if (kpm_smallpages == 0) { 1147 kpmp_table = (kpm_hlk_t *)table; 1148 PRM_DEBUG(kpmp_table); 1149 } else { 1150 kpmp_stable = (kpm_shlk_t *)table; 1151 PRM_DEBUG(kpmp_stable); 1152 } 1153 1154 alloc_base += alloc_sz; 1155 PRM_DEBUG(alloc_base); 1156 } 1157 1158 if (&ecache_init_scrub_flush_area) { 1159 /* 1160 * Pass alloc_base directly, as the routine itself is 1161 * responsible for any special alignment requirements... 1162 */ 1163 alloc_base = ecache_init_scrub_flush_area(alloc_base); 1164 PRM_DEBUG(alloc_base); 1165 } 1166 1167 /* 1168 * Take the most current snapshot we can by calling mem-update. 1169 */ 1170 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1171 &boot_physavail, &boot_physavail_len, 1172 &boot_virtavail, &boot_virtavail_len); 1173 1174 /* 1175 * Reset npages and memblocks based on boot_physavail list. 1176 */ 1177 size_physavail(boot_physavail, boot_physavail_len, &npages, &memblocks); 1178 PRM_DEBUG(npages); 1179 1180 /* 1181 * Account for extra memory after e_text. 1182 */ 1183 npages += extra_etpg; 1184 1185 /* 1186 * Calculate the largest free memory chunk in the nucleus data area. 1187 * We need to figure out if page structs can fit in there or not. 1188 * We also make sure enough page structs get created for any physical 1189 * memory we might be returning to the system. 1190 */ 1191 ndata_remain_sz = ndata_maxsize(&ndata); 1192 PRM_DEBUG(ndata_remain_sz); 1193 1194 pp_sz = sizeof (struct page) * npages; 1195 1196 /* 1197 * Here's a nice bit of code based on somewhat recursive logic: 1198 * 1199 * If the page array would fit within the nucleus, we want to 1200 * add npages to cover any extra memory we may be returning back 1201 * to the system. 1202 * 1203 * HOWEVER, the page array is sized by calculating the size of 1204 * (struct page * npages), as are the pagehash table, ctrs and 1205 * memseg_list, so the very act of performing the calculation below may 1206 * in fact make the array large enough that it no longer fits in the 1207 * nucleus, meaning there would now be a much larger area of the 1208 * nucleus free that should really be added to npages, which would 1209 * make the page array that much larger, and so on. 1210 * 1211 * This also ignores the memory possibly used in the nucleus for the 1212 * the page hash, ctrs and memseg list and the fact that whether they 1213 * fit there or not varies with the npages calculation below, but we 1214 * don't even factor them into the equation at this point; perhaps we 1215 * should or perhaps we should just take the approach that the few 1216 * extra pages we could add via this calculation REALLY aren't worth 1217 * the hassle... 1218 */ 1219 if (ndata_remain_sz > pp_sz) { 1220 size_t spare = ndata_spare(&ndata, pp_sz, ecache_alignsize); 1221 1222 npages += mmu_btop(spare); 1223 1224 pp_sz = npages * sizeof (struct page); 1225 1226 pp_base = ndata_alloc(&ndata, pp_sz, ecache_alignsize); 1227 } 1228 1229 /* 1230 * If physmem is patched to be non-zero, use it instead of 1231 * the monitor value unless physmem is larger than the total 1232 * amount of memory on hand. 1233 */ 1234 if (physmem == 0 || physmem > npages) 1235 physmem = npages; 1236 1237 /* 1238 * If pp_base is NULL that means the routines above have determined 1239 * the page array will not fit in the nucleus; we'll have to 1240 * BOP_ALLOC() ourselves some space for them. 1241 */ 1242 if (pp_base == NULL) { 1243 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1244 alloc_alignsize); 1245 1246 alloc_sz = roundup(pp_sz, alloc_alignsize); 1247 1248 if ((pp_base = (struct page *)BOP_ALLOC(bootops, 1249 alloc_base, alloc_sz, alloc_alignsize)) != 1250 (struct page *)alloc_base) 1251 panic("page alloc failure"); 1252 1253 alloc_base += alloc_sz; 1254 } 1255 1256 /* 1257 * The page structure hash table size is a power of 2 1258 * such that the average hash chain length is PAGE_HASHAVELEN. 1259 */ 1260 page_hashsz = npages / PAGE_HASHAVELEN; 1261 page_hashsz = 1 << highbit((ulong_t)page_hashsz); 1262 pagehash_sz = sizeof (struct page *) * page_hashsz; 1263 1264 /* 1265 * We want to TRY to fit the page structure hash table, 1266 * the page size free list counters, the memseg list and 1267 * and the kpm page space in the nucleus if possible. 1268 * 1269 * alloc_sz counts how much memory needs to be allocated by 1270 * BOP_ALLOC(). 1271 */ 1272 page_hash = ndata_alloc(&ndata, pagehash_sz, ecache_alignsize); 1273 1274 alloc_sz = (page_hash == NULL ? pagehash_sz : 0); 1275 1276 /* 1277 * Size up per page size free list counters. 1278 */ 1279 ctrs_sz = page_ctrs_sz(); 1280 ctrs_base = ndata_alloc(&ndata, ctrs_sz, ecache_alignsize); 1281 1282 if (ctrs_base == NULL) 1283 alloc_sz = roundup(alloc_sz, ecache_alignsize) + ctrs_sz; 1284 1285 /* 1286 * The memseg list is for the chunks of physical memory that 1287 * will be managed by the vm system. The number calculated is 1288 * a guess as boot may fragment it more when memory allocations 1289 * are made before kphysm_init(). Currently, there are two 1290 * allocations before then, so we assume each causes fragmen- 1291 * tation, and add a couple more for good measure. 1292 */ 1293 memseg_sz = sizeof (struct memseg) * (memblocks + 4); 1294 memseg_base = ndata_alloc(&ndata, memseg_sz, ecache_alignsize); 1295 1296 if (memseg_base == NULL) 1297 alloc_sz = roundup(alloc_sz, ecache_alignsize) + memseg_sz; 1298 1299 1300 if (kpm_enable) { 1301 /* 1302 * kpm page space -- Update kpm_npages and make the 1303 * same assumption about fragmenting as it is done 1304 * for memseg_sz above. 1305 */ 1306 kpm_npages_setup(memblocks + 4); 1307 kpm_pp_sz = (kpm_smallpages == 0) ? 1308 kpm_npages * sizeof (kpm_page_t): 1309 kpm_npages * sizeof (kpm_spage_t); 1310 1311 kpm_pp_base = (uintptr_t)ndata_alloc(&ndata, kpm_pp_sz, 1312 ecache_alignsize); 1313 1314 if (kpm_pp_base == NULL) 1315 alloc_sz = roundup(alloc_sz, ecache_alignsize) + 1316 kpm_pp_sz; 1317 } 1318 1319 if (alloc_sz > 0) { 1320 uintptr_t bop_base; 1321 1322 /* 1323 * We need extra memory allocated through BOP_ALLOC. 1324 */ 1325 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1326 alloc_alignsize); 1327 1328 alloc_sz = roundup(alloc_sz, alloc_alignsize); 1329 1330 if ((bop_base = (uintptr_t)BOP_ALLOC(bootops, alloc_base, 1331 alloc_sz, alloc_alignsize)) != (uintptr_t)alloc_base) 1332 panic("system page struct alloc failure"); 1333 1334 alloc_base += alloc_sz; 1335 1336 if (page_hash == NULL) { 1337 page_hash = (struct page **)bop_base; 1338 bop_base = roundup(bop_base + pagehash_sz, 1339 ecache_alignsize); 1340 } 1341 1342 if (ctrs_base == NULL) { 1343 ctrs_base = (caddr_t)bop_base; 1344 bop_base = roundup(bop_base + ctrs_sz, 1345 ecache_alignsize); 1346 } 1347 1348 if (memseg_base == NULL) { 1349 memseg_base = (struct memseg *)bop_base; 1350 bop_base = roundup(bop_base + memseg_sz, 1351 ecache_alignsize); 1352 } 1353 1354 if (kpm_enable && kpm_pp_base == NULL) { 1355 kpm_pp_base = (uintptr_t)bop_base; 1356 bop_base = roundup(bop_base + kpm_pp_sz, 1357 ecache_alignsize); 1358 } 1359 1360 ASSERT(bop_base <= (uintptr_t)alloc_base); 1361 } 1362 1363 /* 1364 * Initialize per page size free list counters. 1365 */ 1366 ctrs_end = page_ctrs_alloc(ctrs_base); 1367 ASSERT(ctrs_base + ctrs_sz >= ctrs_end); 1368 1369 PRM_DEBUG(page_hash); 1370 PRM_DEBUG(memseg_base); 1371 PRM_DEBUG(kpm_pp_base); 1372 PRM_DEBUG(kpm_pp_sz); 1373 PRM_DEBUG(pp_base); 1374 PRM_DEBUG(pp_sz); 1375 PRM_DEBUG(alloc_base); 1376 1377 #ifdef TRAPTRACE 1378 /* 1379 * Allocate trap trace buffer last so as not to affect 1380 * the 4M alignments of the allocations above on V9 SPARCs... 1381 */ 1382 alloc_base = trap_trace_alloc(alloc_base); 1383 PRM_DEBUG(alloc_base); 1384 #endif /* TRAPTRACE */ 1385 1386 if (kmem64_base) { 1387 /* 1388 * Set the end of the kmem64 segment for V9 SPARCs, if 1389 * appropriate... 1390 */ 1391 kmem64_end = (caddr_t)roundup((uintptr_t)alloc_base, 1392 alloc_alignsize); 1393 1394 PRM_DEBUG(kmem64_base); 1395 PRM_DEBUG(kmem64_end); 1396 } 1397 1398 /* 1399 * Allocate space for the interrupt vector table and also for the 1400 * reserved interrupt vector data structures. 1401 */ 1402 memspace = (caddr_t)BOP_ALLOC(bootops, (caddr_t)intr_vec_table, 1403 IVSIZE, MMU_PAGESIZE); 1404 if (memspace != (caddr_t)intr_vec_table) 1405 panic("interrupt vector table allocation failure"); 1406 1407 /* 1408 * The memory lists from boot are allocated from the heap arena 1409 * so that later they can be freed and/or reallocated. 1410 */ 1411 if (BOP_GETPROP(bootops, "extent", &memlist_sz) == -1) 1412 panic("could not retrieve property \"extent\""); 1413 1414 /* 1415 * Between now and when we finish copying in the memory lists, 1416 * allocations happen so the space gets fragmented and the 1417 * lists longer. Leave enough space for lists twice as long 1418 * as what boot says it has now; roundup to a pagesize. 1419 * Also add space for the final phys-avail copy in the fixup 1420 * routine. 1421 */ 1422 va = (caddr_t)(sysbase + PAGESIZE + PANICBUFSIZE + 1423 roundup(IVSIZE, MMU_PAGESIZE)); 1424 memlist_sz *= 4; 1425 memlist_sz = roundup(memlist_sz, MMU_PAGESIZE); 1426 memspace = (caddr_t)BOP_ALLOC(bootops, va, memlist_sz, BO_NO_ALIGN); 1427 if (memspace == NULL) 1428 halt("Boot allocation failed."); 1429 1430 memlist = (struct memlist *)memspace; 1431 memlist_end = (char *)memspace + memlist_sz; 1432 1433 PRM_DEBUG(memlist); 1434 PRM_DEBUG(memlist_end); 1435 PRM_DEBUG(sysbase); 1436 PRM_DEBUG(syslimit); 1437 1438 kernelheap_init((void *)sysbase, (void *)syslimit, 1439 (caddr_t)sysbase + PAGESIZE, NULL, NULL); 1440 1441 /* 1442 * Take the most current snapshot we can by calling mem-update. 1443 */ 1444 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1445 &boot_physavail, &boot_physavail_len, 1446 &boot_virtavail, &boot_virtavail_len); 1447 1448 /* 1449 * Remove the space used by BOP_ALLOC from the kernel heap 1450 * plus the area actually used by the OBP (if any) 1451 * ignoring virtual addresses in virt_avail, above syslimit. 1452 */ 1453 virt_avail = memlist; 1454 copy_memlist(boot_virtavail, boot_virtavail_len, &memlist); 1455 1456 for (cur = virt_avail; cur->next; cur = cur->next) { 1457 uint64_t range_base, range_size; 1458 1459 if ((range_base = cur->address + cur->size) < (uint64_t)sysbase) 1460 continue; 1461 if (range_base >= (uint64_t)syslimit) 1462 break; 1463 /* 1464 * Limit the range to end at syslimit. 1465 */ 1466 range_size = MIN(cur->next->address, 1467 (uint64_t)syslimit) - range_base; 1468 (void) vmem_xalloc(heap_arena, (size_t)range_size, PAGESIZE, 1469 0, 0, (void *)range_base, (void *)(range_base + range_size), 1470 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 1471 } 1472 1473 phys_avail = memlist; 1474 (void) copy_physavail(boot_physavail, boot_physavail_len, 1475 &memlist, 0, 0); 1476 1477 /* 1478 * Add any extra memory after e_text to the phys_avail list, as long 1479 * as there's at least a page to add. 1480 */ 1481 if (extra_etpg) 1482 memlist_add(va_to_pa(extra_etva), mmu_ptob(extra_etpg), 1483 &memlist, &phys_avail); 1484 1485 /* 1486 * Add any extra memory after e_data to the phys_avail list as long 1487 * as there's at least a page to add. Usually, there isn't any, 1488 * since extra HME blocks typically get allocated there first before 1489 * using RAM elsewhere. 1490 */ 1491 if ((nalloc_base = ndata_extra_base(&ndata, MMU_PAGESIZE)) == NULL) 1492 nalloc_base = nalloc_end; 1493 ndata_remain_sz = nalloc_end - nalloc_base; 1494 1495 if (ndata_remain_sz >= MMU_PAGESIZE) 1496 memlist_add(va_to_pa(nalloc_base), 1497 (uint64_t)ndata_remain_sz, &memlist, &phys_avail); 1498 1499 PRM_DEBUG(memlist); 1500 PRM_DEBUG(memlist_sz); 1501 PRM_DEBUG(memspace); 1502 1503 if ((caddr_t)memlist > (memspace + memlist_sz)) 1504 panic("memlist overflow"); 1505 1506 PRM_DEBUG(pp_base); 1507 PRM_DEBUG(memseg_base); 1508 PRM_DEBUG(npages); 1509 1510 /* 1511 * Initialize the page structures from the memory lists. 1512 */ 1513 kphysm_init(pp_base, memseg_base, npages, kpm_pp_base, kpm_npages); 1514 1515 availrmem_initial = availrmem = freemem; 1516 PRM_DEBUG(availrmem); 1517 1518 /* 1519 * Some of the locks depend on page_hashsz being set! 1520 * kmem_init() depends on this; so, keep it here. 1521 */ 1522 page_lock_init(); 1523 1524 /* 1525 * Initialize kernel memory allocator. 1526 */ 1527 kmem_init(); 1528 1529 /* 1530 * Initialize bp_mapin(). 1531 */ 1532 bp_init(shm_alignment, HAT_STRICTORDER); 1533 1534 /* 1535 * Reserve space for panicbuf, intr_vec_table and reserved interrupt 1536 * vector data structures from the 32-bit heap. 1537 */ 1538 (void) vmem_xalloc(heap32_arena, PANICBUFSIZE, PAGESIZE, 0, 0, 1539 panicbuf, panicbuf + PANICBUFSIZE, 1540 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 1541 1542 (void) vmem_xalloc(heap32_arena, IVSIZE, PAGESIZE, 0, 0, 1543 intr_vec_table, (caddr_t)intr_vec_table + IVSIZE, 1544 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 1545 1546 mem_config_init(); 1547 } 1548 1549 static void 1550 startup_modules(void) 1551 { 1552 int proplen, nhblk1, nhblk8; 1553 size_t nhblksz; 1554 pgcnt_t hblk_pages, pages_per_hblk; 1555 size_t hme8blk_sz, hme1blk_sz; 1556 1557 /* 1558 * Log any optional messages from the boot program 1559 */ 1560 proplen = (size_t)BOP_GETPROPLEN(bootops, "boot-message"); 1561 if (proplen > 0) { 1562 char *msg; 1563 size_t len = (size_t)proplen; 1564 1565 msg = kmem_zalloc(len, KM_SLEEP); 1566 (void) BOP_GETPROP(bootops, "boot-message", msg); 1567 cmn_err(CE_CONT, "?%s\n", msg); 1568 kmem_free(msg, len); 1569 } 1570 1571 /* 1572 * Let the platforms have a chance to change default 1573 * values before reading system file. 1574 */ 1575 if (&set_platform_defaults) 1576 set_platform_defaults(); 1577 1578 /* 1579 * Calculate default settings of system parameters based upon 1580 * maxusers, yet allow to be overridden via the /etc/system file. 1581 */ 1582 param_calc(0); 1583 1584 mod_setup(); 1585 1586 /* 1587 * If this is a positron, complain and halt. 1588 */ 1589 if (&iam_positron && iam_positron()) { 1590 cmn_err(CE_WARN, "This hardware platform is not supported" 1591 " by this release of Solaris.\n"); 1592 #ifdef DEBUG 1593 prom_enter_mon(); /* Type 'go' to resume */ 1594 cmn_err(CE_WARN, "Booting an unsupported platform.\n"); 1595 cmn_err(CE_WARN, "Booting with down-rev firmware.\n"); 1596 1597 #else /* DEBUG */ 1598 halt(0); 1599 #endif /* DEBUG */ 1600 } 1601 1602 /* 1603 * If we are running firmware that isn't 64-bit ready 1604 * then complain and halt. 1605 */ 1606 do_prom_version_check(); 1607 1608 /* 1609 * Initialize system parameters 1610 */ 1611 param_init(); 1612 1613 /* 1614 * maxmem is the amount of physical memory we're playing with. 1615 */ 1616 maxmem = physmem; 1617 1618 /* Set segkp limits. */ 1619 ncbase = kdi_segdebugbase; 1620 ncend = kdi_segdebugbase; 1621 1622 /* 1623 * Initialize the hat layer. 1624 */ 1625 hat_init(); 1626 1627 /* 1628 * Initialize segment management stuff. 1629 */ 1630 seg_init(); 1631 1632 /* 1633 * Create the va>tte handler, so the prom can understand 1634 * kernel translations. The handler is installed later, just 1635 * as we are about to take over the trap table from the prom. 1636 */ 1637 create_va_to_tte(); 1638 1639 /* 1640 * Load the forthdebugger (optional) 1641 */ 1642 forthdebug_init(); 1643 1644 /* 1645 * Create OBP node for console input callbacks 1646 * if it is needed. 1647 */ 1648 startup_create_io_node(); 1649 1650 if (modloadonly("fs", "specfs") == -1) 1651 halt("Can't load specfs"); 1652 1653 if (modloadonly("fs", "devfs") == -1) 1654 halt("Can't load devfs"); 1655 1656 if (modloadonly("misc", "swapgeneric") == -1) 1657 halt("Can't load swapgeneric"); 1658 1659 (void) modloadonly("sys", "lbl_edition"); 1660 1661 dispinit(); 1662 1663 /* 1664 * Infer meanings to the members of the idprom buffer. 1665 */ 1666 parse_idprom(); 1667 1668 /* Read cluster configuration data. */ 1669 clconf_init(); 1670 1671 setup_ddi(); 1672 1673 /* 1674 * Lets take this opportunity to load the root device. 1675 */ 1676 if (loadrootmodules() != 0) 1677 debug_enter("Can't load the root filesystem"); 1678 1679 /* 1680 * Load tod driver module for the tod part found on this system. 1681 * Recompute the cpu frequency/delays based on tod as tod part 1682 * tends to keep time more accurately. 1683 */ 1684 if (&load_tod_module) 1685 load_tod_module(); 1686 1687 /* 1688 * Allow platforms to load modules which might 1689 * be needed after bootops are gone. 1690 */ 1691 if (&load_platform_modules) 1692 load_platform_modules(); 1693 1694 setcpudelay(); 1695 1696 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1697 &boot_physavail, &boot_physavail_len, 1698 &boot_virtavail, &boot_virtavail_len); 1699 1700 bop_alloc_pages = size_virtalloc(boot_virtavail, boot_virtavail_len); 1701 1702 /* 1703 * Calculation and allocation of hmeblks needed to remap 1704 * the memory allocated by PROM till now: 1705 * 1706 * (1) calculate how much virtual memory has been bop_alloc'ed. 1707 * (2) roundup this memory to span of hme8blk, i.e. 64KB 1708 * (3) calculate number of hme8blk's needed to remap this memory 1709 * (4) calculate amount of memory that's consumed by these hme8blk's 1710 * (5) add memory calculated in steps (2) and (4) above. 1711 * (6) roundup this memory to span of hme8blk, i.e. 64KB 1712 * (7) calculate number of hme8blk's needed to remap this memory 1713 * (8) calculate amount of memory that's consumed by these hme8blk's 1714 * (9) allocate additional hme1blk's to hold large mappings. 1715 * H8TOH1 determines this. The current SWAG gives enough hblk1's 1716 * to remap everything with 4M mappings. 1717 * (10) account for partially used hblk8's due to non-64K aligned 1718 * PROM mapping entries. 1719 * (11) add memory calculated in steps (8), (9), and (10) above. 1720 * (12) kmem_zalloc the memory calculated in (11); since segkmem 1721 * is not ready yet, this gets bop_alloc'ed. 1722 * (13) there will be very few bop_alloc's after this point before 1723 * trap table takes over 1724 */ 1725 1726 /* sfmmu_init_nucleus_hblks expects properly aligned data structures. */ 1727 hme8blk_sz = roundup(HME8BLK_SZ, sizeof (int64_t)); 1728 hme1blk_sz = roundup(HME1BLK_SZ, sizeof (int64_t)); 1729 1730 pages_per_hblk = btop(HMEBLK_SPAN(TTE8K)); 1731 bop_alloc_pages = roundup(bop_alloc_pages, pages_per_hblk); 1732 nhblk8 = bop_alloc_pages / pages_per_hblk; 1733 nhblk1 = roundup(nhblk8, H8TOH1) / H8TOH1; 1734 hblk_pages = btopr(nhblk8 * hme8blk_sz + nhblk1 * hme1blk_sz); 1735 bop_alloc_pages += hblk_pages; 1736 bop_alloc_pages = roundup(bop_alloc_pages, pages_per_hblk); 1737 nhblk8 = bop_alloc_pages / pages_per_hblk; 1738 nhblk1 = roundup(nhblk8, H8TOH1) / H8TOH1; 1739 if (nhblk1 < hblk1_min) 1740 nhblk1 = hblk1_min; 1741 if (nhblk8 < hblk8_min) 1742 nhblk8 = hblk8_min; 1743 1744 /* 1745 * Since hblk8's can hold up to 64k of mappings aligned on a 64k 1746 * boundary, the number of hblk8's needed to map the entries in the 1747 * boot_virtavail list needs to be adjusted to take this into 1748 * consideration. Thus, we need to add additional hblk8's since it 1749 * is possible that an hblk8 will not have all 8 slots used due to 1750 * alignment constraints. Since there were boot_virtavail_len entries 1751 * in that list, we need to add that many hblk8's to the number 1752 * already calculated to make sure we don't underestimate. 1753 */ 1754 nhblk8 += boot_virtavail_len; 1755 nhblksz = nhblk8 * hme8blk_sz + nhblk1 * hme1blk_sz; 1756 1757 /* Allocate in pagesize chunks */ 1758 nhblksz = roundup(nhblksz, MMU_PAGESIZE); 1759 hblk_base = kmem_zalloc(nhblksz, KM_SLEEP); 1760 sfmmu_init_nucleus_hblks(hblk_base, nhblksz, nhblk8, nhblk1); 1761 } 1762 1763 static void 1764 startup_bop_gone(void) 1765 { 1766 extern int bop_io_quiesced; 1767 1768 /* 1769 * Destroy the MD initialized at startup 1770 * The startup initializes the MD framework 1771 * using prom and BOP alloc free it now. 1772 */ 1773 mach_descrip_startup_fini(); 1774 1775 /* 1776 * Call back into boot and release boots resources. 1777 */ 1778 BOP_QUIESCE_IO(bootops); 1779 bop_io_quiesced = 1; 1780 1781 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1782 &boot_physavail, &boot_physavail_len, 1783 &boot_virtavail, &boot_virtavail_len); 1784 /* 1785 * Copy physinstalled list into kernel space. 1786 */ 1787 phys_install = memlist; 1788 copy_memlist(boot_physinstalled, boot_physinstalled_len, &memlist); 1789 1790 /* 1791 * setup physically contiguous area twice as large as the ecache. 1792 * this is used while doing displacement flush of ecaches 1793 */ 1794 if (&ecache_flush_address) { 1795 ecache_flushaddr = ecache_flush_address(); 1796 if (ecache_flushaddr == (uint64_t)-1) { 1797 cmn_err(CE_PANIC, 1798 "startup: no memory to set ecache_flushaddr"); 1799 } 1800 } 1801 1802 /* 1803 * Virtual available next. 1804 */ 1805 ASSERT(virt_avail != NULL); 1806 memlist_free_list(virt_avail); 1807 virt_avail = memlist; 1808 copy_memlist(boot_virtavail, boot_virtavail_len, &memlist); 1809 1810 /* 1811 * Last chance to ask our booter questions .. 1812 */ 1813 } 1814 1815 1816 /* 1817 * startup_fixup_physavail - called from mach_sfmmu.c after the final 1818 * allocations have been performed. We can't call it in startup_bop_gone 1819 * since later operations can cause obp to allocate more memory. 1820 */ 1821 void 1822 startup_fixup_physavail(void) 1823 { 1824 struct memlist *cur; 1825 1826 /* 1827 * take the most current snapshot we can by calling mem-update 1828 */ 1829 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1830 &boot_physavail, &boot_physavail_len, 1831 &boot_virtavail, &boot_virtavail_len); 1832 1833 /* 1834 * Copy phys_avail list, again. 1835 * Both the kernel/boot and the prom have been allocating 1836 * from the original list we copied earlier. 1837 */ 1838 cur = memlist; 1839 (void) copy_physavail(boot_physavail, boot_physavail_len, 1840 &memlist, 0, 0); 1841 1842 /* 1843 * Add any extra memory after e_text we added to the phys_avail list 1844 * back to the old list. 1845 */ 1846 if (extra_etpg) 1847 memlist_add(va_to_pa(extra_etva), mmu_ptob(extra_etpg), 1848 &memlist, &cur); 1849 if (ndata_remain_sz >= MMU_PAGESIZE) 1850 memlist_add(va_to_pa(nalloc_base), 1851 (uint64_t)ndata_remain_sz, &memlist, &cur); 1852 1853 /* 1854 * There isn't any bounds checking on the memlist area 1855 * so ensure it hasn't overgrown. 1856 */ 1857 if ((caddr_t)memlist > (caddr_t)memlist_end) 1858 cmn_err(CE_PANIC, "startup: memlist size exceeded"); 1859 1860 /* 1861 * The kernel removes the pages that were allocated for it from 1862 * the freelist, but we now have to find any -extra- pages that 1863 * the prom has allocated for it's own book-keeping, and remove 1864 * them from the freelist too. sigh. 1865 */ 1866 fix_prom_pages(phys_avail, cur); 1867 1868 ASSERT(phys_avail != NULL); 1869 memlist_free_list(phys_avail); 1870 phys_avail = cur; 1871 1872 /* 1873 * We're done with boot. Just after this point in time, boot 1874 * gets unmapped, so we can no longer rely on its services. 1875 * Zero the bootops to indicate this fact. 1876 */ 1877 bootops = (struct bootops *)NULL; 1878 BOOTOPS_GONE(); 1879 } 1880 1881 static void 1882 startup_vm(void) 1883 { 1884 size_t i; 1885 struct segmap_crargs a; 1886 struct segkpm_crargs b; 1887 1888 uint64_t avmem; 1889 caddr_t va; 1890 pgcnt_t max_phys_segkp; 1891 int mnode; 1892 1893 extern int use_brk_lpg, use_stk_lpg; 1894 1895 /* 1896 * get prom's mappings, create hments for them and switch 1897 * to the kernel context. 1898 */ 1899 hat_kern_setup(); 1900 1901 /* 1902 * Take over trap table 1903 */ 1904 setup_trap_table(); 1905 1906 /* 1907 * Install the va>tte handler, so that the prom can handle 1908 * misses and understand the kernel table layout in case 1909 * we need call into the prom. 1910 */ 1911 install_va_to_tte(); 1912 1913 /* 1914 * Set a flag to indicate that the tba has been taken over. 1915 */ 1916 tba_taken_over = 1; 1917 1918 /* initialize MMU primary context register */ 1919 mmu_init_kcontext(); 1920 1921 /* 1922 * The boot cpu can now take interrupts, x-calls, x-traps 1923 */ 1924 CPUSET_ADD(cpu_ready_set, CPU->cpu_id); 1925 CPU->cpu_flags |= (CPU_READY | CPU_ENABLE | CPU_EXISTS); 1926 1927 /* 1928 * Set a flag to tell write_scb_int() that it can access V_TBR_WR_ADDR. 1929 */ 1930 tbr_wr_addr_inited = 1; 1931 1932 /* 1933 * Initialize VM system, and map kernel address space. 1934 */ 1935 kvm_init(); 1936 1937 /* 1938 * XXX4U: previously, we initialized and turned on 1939 * the caches at this point. But of course we have 1940 * nothing to do, as the prom has already done this 1941 * for us -- main memory must be E$able at all times. 1942 */ 1943 1944 /* 1945 * If the following is true, someone has patched 1946 * phsymem to be less than the number of pages that 1947 * the system actually has. Remove pages until system 1948 * memory is limited to the requested amount. Since we 1949 * have allocated page structures for all pages, we 1950 * correct the amount of memory we want to remove 1951 * by the size of the memory used to hold page structures 1952 * for the non-used pages. 1953 */ 1954 if (physmem < npages) { 1955 pgcnt_t diff, off; 1956 struct page *pp; 1957 struct seg kseg; 1958 1959 cmn_err(CE_WARN, "limiting physmem to %ld pages", physmem); 1960 1961 off = 0; 1962 diff = npages - physmem; 1963 diff -= mmu_btopr(diff * sizeof (struct page)); 1964 kseg.s_as = &kas; 1965 while (diff--) { 1966 pp = page_create_va(&unused_pages_vp, (offset_t)off, 1967 MMU_PAGESIZE, PG_WAIT | PG_EXCL, 1968 &kseg, (caddr_t)off); 1969 if (pp == NULL) 1970 cmn_err(CE_PANIC, "limited physmem too much!"); 1971 page_io_unlock(pp); 1972 page_downgrade(pp); 1973 availrmem--; 1974 off += MMU_PAGESIZE; 1975 } 1976 } 1977 1978 /* 1979 * When printing memory, show the total as physmem less 1980 * that stolen by a debugger. 1981 */ 1982 cmn_err(CE_CONT, "?mem = %ldK (0x%lx000)\n", 1983 (ulong_t)(physinstalled) << (PAGESHIFT - 10), 1984 (ulong_t)(physinstalled) << (PAGESHIFT - 12)); 1985 1986 avmem = (uint64_t)freemem << PAGESHIFT; 1987 cmn_err(CE_CONT, "?avail mem = %lld\n", (unsigned long long)avmem); 1988 1989 /* 1990 * For small memory systems disable automatic large pages. 1991 */ 1992 if (physmem < privm_lpg_min_physmem) { 1993 use_brk_lpg = 0; 1994 use_stk_lpg = 0; 1995 } 1996 1997 /* 1998 * Perform platform specific freelist processing 1999 */ 2000 if (&plat_freelist_process) { 2001 for (mnode = 0; mnode < max_mem_nodes; mnode++) 2002 if (mem_node_config[mnode].exists) 2003 plat_freelist_process(mnode); 2004 } 2005 2006 /* 2007 * Initialize the segkp segment type. We position it 2008 * after the configured tables and buffers (whose end 2009 * is given by econtig) and before V_WKBASE_ADDR. 2010 * Also in this area is segkmap (size SEGMAPSIZE). 2011 */ 2012 2013 /* XXX - cache alignment? */ 2014 va = (caddr_t)SEGKPBASE; 2015 ASSERT(((uintptr_t)va & PAGEOFFSET) == 0); 2016 2017 max_phys_segkp = (physmem * 2); 2018 2019 if (segkpsize < btop(SEGKPMINSIZE) || segkpsize > btop(SEGKPMAXSIZE)) { 2020 segkpsize = btop(SEGKPDEFSIZE); 2021 cmn_err(CE_WARN, "Illegal value for segkpsize. " 2022 "segkpsize has been reset to %ld pages", segkpsize); 2023 } 2024 2025 i = ptob(MIN(segkpsize, max_phys_segkp)); 2026 2027 rw_enter(&kas.a_lock, RW_WRITER); 2028 if (seg_attach(&kas, va, i, segkp) < 0) 2029 cmn_err(CE_PANIC, "startup: cannot attach segkp"); 2030 if (segkp_create(segkp) != 0) 2031 cmn_err(CE_PANIC, "startup: segkp_create failed"); 2032 rw_exit(&kas.a_lock); 2033 2034 /* 2035 * kpm segment 2036 */ 2037 segmap_kpm = kpm_enable && 2038 segmap_kpm && PAGESIZE == MAXBSIZE; 2039 2040 if (kpm_enable) { 2041 rw_enter(&kas.a_lock, RW_WRITER); 2042 2043 /* 2044 * The segkpm virtual range range is larger than the 2045 * actual physical memory size and also covers gaps in 2046 * the physical address range for the following reasons: 2047 * . keep conversion between segkpm and physical addresses 2048 * simple, cheap and unambiguous. 2049 * . avoid extension/shrink of the the segkpm in case of DR. 2050 * . avoid complexity for handling of virtual addressed 2051 * caches, segkpm and the regular mapping scheme must be 2052 * kept in sync wrt. the virtual color of mapped pages. 2053 * Any accesses to virtual segkpm ranges not backed by 2054 * physical memory will fall through the memseg pfn hash 2055 * and will be handled in segkpm_fault. 2056 * Additional kpm_size spaces needed for vac alias prevention. 2057 */ 2058 if (seg_attach(&kas, kpm_vbase, kpm_size * vac_colors, 2059 segkpm) < 0) 2060 cmn_err(CE_PANIC, "cannot attach segkpm"); 2061 2062 b.prot = PROT_READ | PROT_WRITE; 2063 b.nvcolors = shm_alignment >> MMU_PAGESHIFT; 2064 2065 if (segkpm_create(segkpm, (caddr_t)&b) != 0) 2066 panic("segkpm_create segkpm"); 2067 2068 rw_exit(&kas.a_lock); 2069 2070 mach_kpm_init(); 2071 } 2072 2073 if (!segzio_fromheap) { 2074 size_t size; 2075 size_t maxsize; 2076 2077 /* size is in bytes, segziosize is in pages */ 2078 if (segziosize == 0) { 2079 size = mmu_ptob(physmem / 2); 2080 } else { 2081 size = mmu_ptob(segziosize); 2082 } 2083 2084 /* maxsize is 3/4th physmem */ 2085 maxsize = mmu_ptob(physmem) - mmu_ptob(physmem / 4); 2086 2087 if (size < SEGZIOMINSIZE) { 2088 size = SEGZIOMINSIZE; 2089 } else if (size > maxsize) { 2090 size = maxsize; 2091 } 2092 segziosize = mmu_btop(roundup(size, MMU_PAGESIZE)); 2093 /* put the base of the ZIO segment after the kpm segment */ 2094 segzio_base = kpm_vbase + (kpm_size * vac_colors); 2095 PRM_DEBUG(segziosize); 2096 PRM_DEBUG(segzio_base); 2097 2098 /* 2099 * On some platforms, kvm_init is called after the kpm 2100 * sizes have been determined. On SPARC, kvm_init is called 2101 * before, so we have to attach the kzioseg after kvm is 2102 * initialized, otherwise we'll try to allocate from the boot 2103 * area since the kernel heap hasn't yet been configured. 2104 */ 2105 rw_enter(&kas.a_lock, RW_WRITER); 2106 2107 (void) seg_attach(&kas, segzio_base, mmu_ptob(segziosize), 2108 &kzioseg); 2109 (void) segkmem_zio_create(&kzioseg); 2110 2111 /* create zio area covering new segment */ 2112 segkmem_zio_init(segzio_base, mmu_ptob(segziosize)); 2113 2114 rw_exit(&kas.a_lock); 2115 } 2116 2117 2118 /* 2119 * Now create generic mapping segment. This mapping 2120 * goes SEGMAPSIZE beyond SEGMAPBASE. But if the total 2121 * virtual address is greater than the amount of free 2122 * memory that is available, then we trim back the 2123 * segment size to that amount 2124 */ 2125 va = (caddr_t)SEGMAPBASE; 2126 2127 /* 2128 * 1201049: segkmap base address must be MAXBSIZE aligned 2129 */ 2130 ASSERT(((uintptr_t)va & MAXBOFFSET) == 0); 2131 2132 /* 2133 * Set size of segmap to percentage of freemem at boot, 2134 * but stay within the allowable range 2135 * Note we take percentage before converting from pages 2136 * to bytes to avoid an overflow on 32-bit kernels. 2137 */ 2138 i = mmu_ptob((freemem * segmap_percent) / 100); 2139 2140 if (i < MINMAPSIZE) 2141 i = MINMAPSIZE; 2142 2143 if (i > MIN(SEGMAPSIZE, mmu_ptob(freemem))) 2144 i = MIN(SEGMAPSIZE, mmu_ptob(freemem)); 2145 2146 i &= MAXBMASK; /* 1201049: segkmap size must be MAXBSIZE aligned */ 2147 2148 rw_enter(&kas.a_lock, RW_WRITER); 2149 if (seg_attach(&kas, va, i, segkmap) < 0) 2150 cmn_err(CE_PANIC, "cannot attach segkmap"); 2151 2152 a.prot = PROT_READ | PROT_WRITE; 2153 a.shmsize = shm_alignment; 2154 a.nfreelist = 0; /* use segmap driver defaults */ 2155 2156 if (segmap_create(segkmap, (caddr_t)&a) != 0) 2157 panic("segmap_create segkmap"); 2158 rw_exit(&kas.a_lock); 2159 2160 segdev_init(); 2161 } 2162 2163 static void 2164 startup_end(void) 2165 { 2166 if ((caddr_t)memlist > (caddr_t)memlist_end) 2167 panic("memlist overflow 2"); 2168 memlist_free_block((caddr_t)memlist, 2169 ((caddr_t)memlist_end - (caddr_t)memlist)); 2170 memlist = NULL; 2171 2172 /* enable page_relocation since OBP is now done */ 2173 page_relocate_ready = 1; 2174 2175 /* 2176 * Perform tasks that get done after most of the VM 2177 * initialization has been done but before the clock 2178 * and other devices get started. 2179 */ 2180 kern_setup1(); 2181 2182 /* 2183 * Intialize the VM arenas for allocating physically 2184 * contiguus memory chunk for interrupt queues snd 2185 * allocate/register boot cpu's queues, if any and 2186 * allocate dump buffer for sun4v systems to store 2187 * extra crash information during crash dump 2188 */ 2189 contig_mem_init(); 2190 mach_descrip_init(); 2191 cpu_intrq_setup(CPU); 2192 cpu_intrq_register(CPU); 2193 mach_htraptrace_setup(CPU->cpu_id); 2194 mach_htraptrace_configure(CPU->cpu_id); 2195 mach_dump_buffer_init(); 2196 2197 /* 2198 * Initialize interrupt related stuff 2199 */ 2200 cpu_intr_alloc(CPU, NINTR_THREADS); 2201 2202 (void) splzs(); /* allow hi clock ints but not zs */ 2203 2204 /* 2205 * Initialize errors. 2206 */ 2207 error_init(); 2208 2209 /* 2210 * Note that we may have already used kernel bcopy before this 2211 * point - but if you really care about this, adb the use_hw_* 2212 * variables to 0 before rebooting. 2213 */ 2214 mach_hw_copy_limit(); 2215 2216 /* 2217 * Install the "real" preemption guards before DDI services 2218 * are available. 2219 */ 2220 (void) prom_set_preprom(kern_preprom); 2221 (void) prom_set_postprom(kern_postprom); 2222 CPU->cpu_m.mutex_ready = 1; 2223 2224 /* 2225 * Initialize segnf (kernel support for non-faulting loads). 2226 */ 2227 segnf_init(); 2228 2229 /* 2230 * Configure the root devinfo node. 2231 */ 2232 configure(); /* set up devices */ 2233 mach_cpu_halt_idle(); 2234 } 2235 2236 2237 void 2238 post_startup(void) 2239 { 2240 #ifdef PTL1_PANIC_DEBUG 2241 extern void init_ptl1_thread(void); 2242 #endif /* PTL1_PANIC_DEBUG */ 2243 extern void abort_sequence_init(void); 2244 2245 /* 2246 * Set the system wide, processor-specific flags to be passed 2247 * to userland via the aux vector for performance hints and 2248 * instruction set extensions. 2249 */ 2250 bind_hwcap(); 2251 2252 /* 2253 * Startup memory scrubber (if any) 2254 */ 2255 mach_memscrub(); 2256 2257 /* 2258 * Allocate soft interrupt to handle abort sequence. 2259 */ 2260 abort_sequence_init(); 2261 2262 /* 2263 * Configure the rest of the system. 2264 * Perform forceloading tasks for /etc/system. 2265 */ 2266 (void) mod_sysctl(SYS_FORCELOAD, NULL); 2267 /* 2268 * ON4.0: Force /proc module in until clock interrupt handle fixed 2269 * ON4.0: This must be fixed or restated in /etc/systems. 2270 */ 2271 (void) modload("fs", "procfs"); 2272 2273 /* load machine class specific drivers */ 2274 load_mach_drivers(); 2275 2276 /* load platform specific drivers */ 2277 if (&load_platform_drivers) 2278 load_platform_drivers(); 2279 2280 /* load vis simulation module, if we are running w/fpu off */ 2281 if (!fpu_exists) { 2282 if (modload("misc", "vis") == -1) 2283 halt("Can't load vis"); 2284 } 2285 2286 mach_fpras(); 2287 2288 maxmem = freemem; 2289 2290 #ifdef PTL1_PANIC_DEBUG 2291 init_ptl1_thread(); 2292 #endif /* PTL1_PANIC_DEBUG */ 2293 2294 if (&cif_init) 2295 cif_init(); 2296 } 2297 2298 #ifdef PTL1_PANIC_DEBUG 2299 int ptl1_panic_test = 0; 2300 int ptl1_panic_xc_one_test = 0; 2301 int ptl1_panic_xc_all_test = 0; 2302 int ptl1_panic_xt_one_test = 0; 2303 int ptl1_panic_xt_all_test = 0; 2304 kthread_id_t ptl1_thread_p = NULL; 2305 kcondvar_t ptl1_cv; 2306 kmutex_t ptl1_mutex; 2307 int ptl1_recurse_count_threshold = 0x40; 2308 int ptl1_recurse_trap_threshold = 0x3d; 2309 extern void ptl1_recurse(int, int); 2310 extern void ptl1_panic_xt(int, int); 2311 2312 /* 2313 * Called once per second by timeout() to wake up 2314 * the ptl1_panic thread to see if it should cause 2315 * a trap to the ptl1_panic() code. 2316 */ 2317 /* ARGSUSED */ 2318 static void 2319 ptl1_wakeup(void *arg) 2320 { 2321 mutex_enter(&ptl1_mutex); 2322 cv_signal(&ptl1_cv); 2323 mutex_exit(&ptl1_mutex); 2324 } 2325 2326 /* 2327 * ptl1_panic cross call function: 2328 * Needed because xc_one() and xc_some() can pass 2329 * 64 bit args but ptl1_recurse() expects ints. 2330 */ 2331 static void 2332 ptl1_panic_xc(void) 2333 { 2334 ptl1_recurse(ptl1_recurse_count_threshold, 2335 ptl1_recurse_trap_threshold); 2336 } 2337 2338 /* 2339 * The ptl1 thread waits for a global flag to be set 2340 * and uses the recurse thresholds to set the stack depth 2341 * to cause a ptl1_panic() directly via a call to ptl1_recurse 2342 * or indirectly via the cross call and cross trap functions. 2343 * 2344 * This is useful testing stack overflows and normal 2345 * ptl1_panic() states with a know stack frame. 2346 * 2347 * ptl1_recurse() is an asm function in ptl1_panic.s that 2348 * sets the {In, Local, Out, and Global} registers to a 2349 * know state on the stack and just prior to causing a 2350 * test ptl1_panic trap. 2351 */ 2352 static void 2353 ptl1_thread(void) 2354 { 2355 mutex_enter(&ptl1_mutex); 2356 while (ptl1_thread_p) { 2357 cpuset_t other_cpus; 2358 int cpu_id; 2359 int my_cpu_id; 2360 int target_cpu_id; 2361 int target_found; 2362 2363 if (ptl1_panic_test) { 2364 ptl1_recurse(ptl1_recurse_count_threshold, 2365 ptl1_recurse_trap_threshold); 2366 } 2367 2368 /* 2369 * Find potential targets for x-call and x-trap, 2370 * if any exist while preempt is disabled we 2371 * start a ptl1_panic if requested via a 2372 * globals. 2373 */ 2374 kpreempt_disable(); 2375 my_cpu_id = CPU->cpu_id; 2376 other_cpus = cpu_ready_set; 2377 CPUSET_DEL(other_cpus, CPU->cpu_id); 2378 target_found = 0; 2379 if (!CPUSET_ISNULL(other_cpus)) { 2380 /* 2381 * Pick the first one 2382 */ 2383 for (cpu_id = 0; cpu_id < NCPU; cpu_id++) { 2384 if (cpu_id == my_cpu_id) 2385 continue; 2386 2387 if (CPU_XCALL_READY(cpu_id)) { 2388 target_cpu_id = cpu_id; 2389 target_found = 1; 2390 break; 2391 } 2392 } 2393 ASSERT(target_found); 2394 2395 if (ptl1_panic_xc_one_test) { 2396 xc_one(target_cpu_id, 2397 (xcfunc_t *)ptl1_panic_xc, 0, 0); 2398 } 2399 if (ptl1_panic_xc_all_test) { 2400 xc_some(other_cpus, 2401 (xcfunc_t *)ptl1_panic_xc, 0, 0); 2402 } 2403 if (ptl1_panic_xt_one_test) { 2404 xt_one(target_cpu_id, 2405 (xcfunc_t *)ptl1_panic_xt, 0, 0); 2406 } 2407 if (ptl1_panic_xt_all_test) { 2408 xt_some(other_cpus, 2409 (xcfunc_t *)ptl1_panic_xt, 0, 0); 2410 } 2411 } 2412 kpreempt_enable(); 2413 (void) timeout(ptl1_wakeup, NULL, hz); 2414 (void) cv_wait(&ptl1_cv, &ptl1_mutex); 2415 } 2416 mutex_exit(&ptl1_mutex); 2417 } 2418 2419 /* 2420 * Called during early startup to create the ptl1_thread 2421 */ 2422 void 2423 init_ptl1_thread(void) 2424 { 2425 ptl1_thread_p = thread_create(NULL, 0, ptl1_thread, NULL, 0, 2426 &p0, TS_RUN, 0); 2427 } 2428 #endif /* PTL1_PANIC_DEBUG */ 2429 2430 2431 /* 2432 * Add to a memory list. 2433 * start = start of new memory segment 2434 * len = length of new memory segment in bytes 2435 * memlistp = pointer to array of available memory segment structures 2436 * curmemlistp = memory list to which to add segment. 2437 */ 2438 static void 2439 memlist_add(uint64_t start, uint64_t len, struct memlist **memlistp, 2440 struct memlist **curmemlistp) 2441 { 2442 struct memlist *new; 2443 2444 new = *memlistp; 2445 new->address = start; 2446 new->size = len; 2447 *memlistp = new + 1; 2448 2449 memlist_insert(new, curmemlistp); 2450 } 2451 2452 /* 2453 * In the case of architectures that support dynamic addition of 2454 * memory at run-time there are two cases where memsegs need to 2455 * be initialized and added to the memseg list. 2456 * 1) memsegs that are constructed at startup. 2457 * 2) memsegs that are constructed at run-time on 2458 * hot-plug capable architectures. 2459 * This code was originally part of the function kphysm_init(). 2460 */ 2461 2462 static void 2463 memseg_list_add(struct memseg *memsegp) 2464 { 2465 struct memseg **prev_memsegp; 2466 pgcnt_t num; 2467 2468 /* insert in memseg list, decreasing number of pages order */ 2469 2470 num = MSEG_NPAGES(memsegp); 2471 2472 for (prev_memsegp = &memsegs; *prev_memsegp; 2473 prev_memsegp = &((*prev_memsegp)->next)) { 2474 if (num > MSEG_NPAGES(*prev_memsegp)) 2475 break; 2476 } 2477 2478 memsegp->next = *prev_memsegp; 2479 *prev_memsegp = memsegp; 2480 2481 if (kpm_enable) { 2482 memsegp->nextpa = (memsegp->next) ? 2483 va_to_pa(memsegp->next) : MSEG_NULLPTR_PA; 2484 2485 if (prev_memsegp != &memsegs) { 2486 struct memseg *msp; 2487 msp = (struct memseg *)((caddr_t)prev_memsegp - 2488 offsetof(struct memseg, next)); 2489 msp->nextpa = va_to_pa(memsegp); 2490 } else { 2491 memsegspa = va_to_pa(memsegs); 2492 } 2493 } 2494 } 2495 2496 /* 2497 * PSM add_physmem_cb(). US-II and newer processors have some 2498 * flavor of the prefetch capability implemented. We exploit 2499 * this capability for optimum performance. 2500 */ 2501 #define PREFETCH_BYTES 64 2502 2503 void 2504 add_physmem_cb(page_t *pp, pfn_t pnum) 2505 { 2506 extern void prefetch_page_w(void *); 2507 2508 pp->p_pagenum = pnum; 2509 2510 /* 2511 * Prefetch one more page_t into E$. To prevent future 2512 * mishaps with the sizeof(page_t) changing on us, we 2513 * catch this on debug kernels if we can't bring in the 2514 * entire hpage with 2 PREFETCH_BYTES reads. See 2515 * also, sun4u/cpu/cpu_module.c 2516 */ 2517 /*LINTED*/ 2518 ASSERT(sizeof (page_t) <= 2*PREFETCH_BYTES); 2519 prefetch_page_w((char *)pp); 2520 } 2521 2522 /* 2523 * kphysm_init() tackles the problem of initializing physical memory. 2524 * The old startup made some assumptions about the kernel living in 2525 * physically contiguous space which is no longer valid. 2526 */ 2527 static void 2528 kphysm_init(page_t *pp, struct memseg *memsegp, pgcnt_t npages, 2529 uintptr_t kpm_pp, pgcnt_t kpm_npages) 2530 { 2531 struct memlist *pmem; 2532 struct memseg *msp; 2533 pfn_t base; 2534 pgcnt_t num; 2535 pfn_t lastseg_pages_end = 0; 2536 pgcnt_t nelem_used = 0; 2537 2538 ASSERT(page_hash != NULL && page_hashsz != 0); 2539 2540 msp = memsegp; 2541 for (pmem = phys_avail; pmem && npages; pmem = pmem->next) { 2542 2543 /* 2544 * Build the memsegs entry 2545 */ 2546 num = btop(pmem->size); 2547 if (num > npages) 2548 num = npages; 2549 npages -= num; 2550 base = btop(pmem->address); 2551 2552 msp->pages = pp; 2553 msp->epages = pp + num; 2554 msp->pages_base = base; 2555 msp->pages_end = base + num; 2556 2557 if (kpm_enable) { 2558 pfn_t pbase_a; 2559 pfn_t pend_a; 2560 pfn_t prev_pend_a; 2561 pgcnt_t nelem; 2562 2563 msp->pagespa = va_to_pa(pp); 2564 msp->epagespa = va_to_pa(pp + num); 2565 pbase_a = kpmptop(ptokpmp(base)); 2566 pend_a = kpmptop(ptokpmp(base + num - 1)) + kpmpnpgs; 2567 nelem = ptokpmp(pend_a - pbase_a); 2568 msp->kpm_nkpmpgs = nelem; 2569 msp->kpm_pbase = pbase_a; 2570 if (lastseg_pages_end) { 2571 /* 2572 * Assume phys_avail is in ascending order 2573 * of physical addresses. 2574 */ 2575 ASSERT(base + num > lastseg_pages_end); 2576 prev_pend_a = kpmptop( 2577 ptokpmp(lastseg_pages_end - 1)) + kpmpnpgs; 2578 2579 if (prev_pend_a > pbase_a) { 2580 /* 2581 * Overlap, more than one memseg may 2582 * point to the same kpm_page range. 2583 */ 2584 if (kpm_smallpages == 0) { 2585 msp->kpm_pages = 2586 (kpm_page_t *)kpm_pp - 1; 2587 kpm_pp = (uintptr_t) 2588 ((kpm_page_t *)kpm_pp 2589 + nelem - 1); 2590 } else { 2591 msp->kpm_spages = 2592 (kpm_spage_t *)kpm_pp - 1; 2593 kpm_pp = (uintptr_t) 2594 ((kpm_spage_t *)kpm_pp 2595 + nelem - 1); 2596 } 2597 nelem_used += nelem - 1; 2598 2599 } else { 2600 if (kpm_smallpages == 0) { 2601 msp->kpm_pages = 2602 (kpm_page_t *)kpm_pp; 2603 kpm_pp = (uintptr_t) 2604 ((kpm_page_t *)kpm_pp 2605 + nelem); 2606 } else { 2607 msp->kpm_spages = 2608 (kpm_spage_t *)kpm_pp; 2609 kpm_pp = (uintptr_t) 2610 ((kpm_spage_t *) 2611 kpm_pp + nelem); 2612 } 2613 nelem_used += nelem; 2614 } 2615 2616 } else { 2617 if (kpm_smallpages == 0) { 2618 msp->kpm_pages = (kpm_page_t *)kpm_pp; 2619 kpm_pp = (uintptr_t) 2620 ((kpm_page_t *)kpm_pp + nelem); 2621 } else { 2622 msp->kpm_spages = (kpm_spage_t *)kpm_pp; 2623 kpm_pp = (uintptr_t) 2624 ((kpm_spage_t *)kpm_pp + nelem); 2625 } 2626 nelem_used = nelem; 2627 } 2628 2629 if (nelem_used > kpm_npages) 2630 panic("kphysm_init: kpm_pp overflow\n"); 2631 2632 msp->kpm_pagespa = va_to_pa(msp->kpm_pages); 2633 lastseg_pages_end = msp->pages_end; 2634 } 2635 2636 memseg_list_add(msp); 2637 2638 /* 2639 * add_physmem() initializes the PSM part of the page 2640 * struct by calling the PSM back with add_physmem_cb(). 2641 * In addition it coalesces pages into larger pages as 2642 * it initializes them. 2643 */ 2644 add_physmem(pp, num, base); 2645 pp += num; 2646 msp++; 2647 } 2648 2649 build_pfn_hash(); 2650 } 2651 2652 /* 2653 * Kernel VM initialization. 2654 * Assumptions about kernel address space ordering: 2655 * (1) gap (user space) 2656 * (2) kernel text 2657 * (3) kernel data/bss 2658 * (4) gap 2659 * (5) kernel data structures 2660 * (6) gap 2661 * (7) debugger (optional) 2662 * (8) monitor 2663 * (9) gap (possibly null) 2664 * (10) dvma 2665 * (11) devices 2666 */ 2667 static void 2668 kvm_init(void) 2669 { 2670 /* 2671 * Put the kernel segments in kernel address space. 2672 */ 2673 rw_enter(&kas.a_lock, RW_WRITER); 2674 as_avlinit(&kas); 2675 2676 (void) seg_attach(&kas, (caddr_t)KERNELBASE, 2677 (size_t)(e_moddata - KERNELBASE), &ktextseg); 2678 (void) segkmem_create(&ktextseg); 2679 2680 (void) seg_attach(&kas, (caddr_t)(KERNELBASE + MMU_PAGESIZE4M), 2681 (size_t)(MMU_PAGESIZE4M), &ktexthole); 2682 (void) segkmem_create(&ktexthole); 2683 2684 (void) seg_attach(&kas, (caddr_t)valloc_base, 2685 (size_t)(econtig32 - valloc_base), &kvalloc); 2686 (void) segkmem_create(&kvalloc); 2687 2688 if (kmem64_base) { 2689 (void) seg_attach(&kas, (caddr_t)kmem64_base, 2690 (size_t)(kmem64_end - kmem64_base), &kmem64); 2691 (void) segkmem_create(&kmem64); 2692 } 2693 2694 /* 2695 * We're about to map out /boot. This is the beginning of the 2696 * system resource management transition. We can no longer 2697 * call into /boot for I/O or memory allocations. 2698 */ 2699 (void) seg_attach(&kas, kernelheap, ekernelheap - kernelheap, &kvseg); 2700 (void) segkmem_create(&kvseg); 2701 hblk_alloc_dynamic = 1; 2702 2703 /* 2704 * we need to preallocate pages for DR operations before enabling large 2705 * page kernel heap because of memseg_remap_init() hat_unload() hack. 2706 */ 2707 memseg_remap_init(); 2708 2709 /* at this point we are ready to use large page heap */ 2710 segkmem_heap_lp_init(); 2711 2712 (void) seg_attach(&kas, (caddr_t)SYSBASE32, SYSLIMIT32 - SYSBASE32, 2713 &kvseg32); 2714 (void) segkmem_create(&kvseg32); 2715 2716 /* 2717 * Create a segment for the debugger. 2718 */ 2719 (void) seg_attach(&kas, kdi_segdebugbase, kdi_segdebugsize, &kdebugseg); 2720 (void) segkmem_create(&kdebugseg); 2721 2722 rw_exit(&kas.a_lock); 2723 } 2724 2725 char obp_tte_str[] = 2726 "h# %x constant MMU_PAGESHIFT " 2727 "h# %x constant TTE8K " 2728 "h# %x constant SFHME_SIZE " 2729 "h# %x constant SFHME_TTE " 2730 "h# %x constant HMEBLK_TAG " 2731 "h# %x constant HMEBLK_NEXT " 2732 "h# %x constant HMEBLK_MISC " 2733 "h# %x constant HMEBLK_HME1 " 2734 "h# %x constant NHMENTS " 2735 "h# %x constant HBLK_SZMASK " 2736 "h# %x constant HBLK_RANGE_SHIFT " 2737 "h# %x constant HMEBP_HBLK " 2738 "h# %x constant HMEBUCKET_SIZE " 2739 "h# %x constant HTAG_SFMMUPSZ " 2740 "h# %x constant HTAG_REHASHSZ " 2741 "h# %x constant mmu_hashcnt " 2742 "h# %p constant uhme_hash " 2743 "h# %p constant khme_hash " 2744 "h# %x constant UHMEHASH_SZ " 2745 "h# %x constant KHMEHASH_SZ " 2746 "h# %p constant KCONTEXT " 2747 "h# %p constant KHATID " 2748 "h# %x constant ASI_MEM " 2749 2750 ": PHYS-X@ ( phys -- data ) " 2751 " ASI_MEM spacex@ " 2752 "; " 2753 2754 ": PHYS-W@ ( phys -- data ) " 2755 " ASI_MEM spacew@ " 2756 "; " 2757 2758 ": PHYS-L@ ( phys -- data ) " 2759 " ASI_MEM spaceL@ " 2760 "; " 2761 2762 ": TTE_PAGE_SHIFT ( ttesz -- hmeshift ) " 2763 " 3 * MMU_PAGESHIFT + " 2764 "; " 2765 2766 ": TTE_IS_VALID ( ttep -- flag ) " 2767 " PHYS-X@ 0< " 2768 "; " 2769 2770 ": HME_HASH_SHIFT ( ttesz -- hmeshift ) " 2771 " dup TTE8K = if " 2772 " drop HBLK_RANGE_SHIFT " 2773 " else " 2774 " TTE_PAGE_SHIFT " 2775 " then " 2776 "; " 2777 2778 ": HME_HASH_BSPAGE ( addr hmeshift -- bspage ) " 2779 " tuck >> swap MMU_PAGESHIFT - << " 2780 "; " 2781 2782 ": HME_HASH_FUNCTION ( sfmmup addr hmeshift -- hmebp ) " 2783 " >> over xor swap ( hash sfmmup ) " 2784 " KHATID <> if ( hash ) " 2785 " UHMEHASH_SZ and ( bucket ) " 2786 " HMEBUCKET_SIZE * uhme_hash + ( hmebp ) " 2787 " else ( hash ) " 2788 " KHMEHASH_SZ and ( bucket ) " 2789 " HMEBUCKET_SIZE * khme_hash + ( hmebp ) " 2790 " then ( hmebp ) " 2791 "; " 2792 2793 ": HME_HASH_TABLE_SEARCH " 2794 " ( sfmmup hmebp hblktag -- sfmmup null | sfmmup hmeblkp ) " 2795 " >r hmebp_hblk + phys-x@ begin ( sfmmup hmeblkp ) ( r: hblktag ) " 2796 " dup if ( sfmmup hmeblkp ) ( r: hblktag ) " 2797 " dup hmeblk_tag + phys-x@ r@ = if ( sfmmup hmeblkp ) " 2798 " dup hmeblk_tag + 8 + phys-x@ 2 pick = if " 2799 " true ( sfmmup hmeblkp true ) ( r: hblktag ) " 2800 " else " 2801 " hmeblk_next + phys-x@ false " 2802 " ( sfmmup hmeblkp false ) ( r: hblktag ) " 2803 " then " 2804 " else " 2805 " hmeblk_next + phys-x@ false " 2806 " ( sfmmup hmeblkp false ) ( r: hblktag ) " 2807 " then " 2808 " else " 2809 " true " 2810 " then " 2811 " until r> drop " 2812 "; " 2813 2814 ": HME_HASH_TAG ( sfmmup rehash addr -- hblktag ) " 2815 " over HME_HASH_SHIFT HME_HASH_BSPAGE ( sfmmup rehash bspage ) " 2816 " HTAG_REHASHSZ << or nip ( hblktag ) " 2817 "; " 2818 2819 ": HBLK_TO_TTEP ( hmeblkp addr -- ttep ) " 2820 " over HMEBLK_MISC + PHYS-L@ HBLK_SZMASK and ( hmeblkp addr ttesz ) " 2821 " TTE8K = if ( hmeblkp addr ) " 2822 " MMU_PAGESHIFT >> NHMENTS 1- and ( hmeblkp hme-index ) " 2823 " else ( hmeblkp addr ) " 2824 " drop 0 ( hmeblkp 0 ) " 2825 " then ( hmeblkp hme-index ) " 2826 " SFHME_SIZE * + HMEBLK_HME1 + ( hmep ) " 2827 " SFHME_TTE + ( ttep ) " 2828 "; " 2829 2830 ": unix-tte ( addr cnum -- false | tte-data true ) " 2831 " KCONTEXT = if ( addr ) " 2832 " KHATID ( addr khatid ) " 2833 " else ( addr ) " 2834 " drop false exit ( false ) " 2835 " then " 2836 " ( addr khatid ) " 2837 " mmu_hashcnt 1+ 1 do ( addr sfmmup ) " 2838 " 2dup swap i HME_HASH_SHIFT " 2839 "( addr sfmmup sfmmup addr hmeshift ) " 2840 " HME_HASH_FUNCTION ( addr sfmmup hmebp ) " 2841 " over i 4 pick " 2842 "( addr sfmmup hmebp sfmmup rehash addr ) " 2843 " HME_HASH_TAG ( addr sfmmup hmebp hblktag ) " 2844 " HME_HASH_TABLE_SEARCH " 2845 "( addr sfmmup { null | hmeblkp } ) " 2846 " ?dup if ( addr sfmmup hmeblkp ) " 2847 " nip swap HBLK_TO_TTEP ( ttep ) " 2848 " dup TTE_IS_VALID if ( valid-ttep ) " 2849 " PHYS-X@ true ( tte-data true ) " 2850 " else ( invalid-tte ) " 2851 " drop false ( false ) " 2852 " then ( false | tte-data true ) " 2853 " unloop exit ( false | tte-data true ) " 2854 " then ( addr sfmmup ) " 2855 " loop ( addr sfmmup ) " 2856 " 2drop false ( false ) " 2857 "; " 2858 ; 2859 2860 void 2861 create_va_to_tte(void) 2862 { 2863 char *bp; 2864 extern int khmehash_num, uhmehash_num; 2865 extern struct hmehash_bucket *khme_hash, *uhme_hash; 2866 2867 #define OFFSET(type, field) ((uintptr_t)(&((type *)0)->field)) 2868 2869 bp = (char *)kobj_zalloc(MMU_PAGESIZE, KM_SLEEP); 2870 2871 /* 2872 * Teach obp how to parse our sw ttes. 2873 */ 2874 (void) sprintf(bp, obp_tte_str, 2875 MMU_PAGESHIFT, 2876 TTE8K, 2877 sizeof (struct sf_hment), 2878 OFFSET(struct sf_hment, hme_tte), 2879 OFFSET(struct hme_blk, hblk_tag), 2880 OFFSET(struct hme_blk, hblk_nextpa), 2881 OFFSET(struct hme_blk, hblk_misc), 2882 OFFSET(struct hme_blk, hblk_hme), 2883 NHMENTS, 2884 HBLK_SZMASK, 2885 HBLK_RANGE_SHIFT, 2886 OFFSET(struct hmehash_bucket, hmeh_nextpa), 2887 sizeof (struct hmehash_bucket), 2888 HTAG_SFMMUPSZ, 2889 HTAG_REHASHSZ, 2890 mmu_hashcnt, 2891 (caddr_t)va_to_pa((caddr_t)uhme_hash), 2892 (caddr_t)va_to_pa((caddr_t)khme_hash), 2893 UHMEHASH_SZ, 2894 KHMEHASH_SZ, 2895 KCONTEXT, 2896 KHATID, 2897 ASI_MEM); 2898 prom_interpret(bp, 0, 0, 0, 0, 0); 2899 2900 kobj_free(bp, MMU_PAGESIZE); 2901 } 2902 2903 void 2904 install_va_to_tte(void) 2905 { 2906 /* 2907 * advise prom that he can use unix-tte 2908 */ 2909 prom_interpret("' unix-tte is va>tte-data", 0, 0, 0, 0, 0); 2910 } 2911 2912 2913 /* 2914 * Because kmdb links prom_stdout_is_framebuffer into its own 2915 * module, we add "device-type=display" here for /os-io node, so that 2916 * prom_stdout_is_framebuffer still works corrrectly after /os-io node 2917 * is registered into OBP. 2918 */ 2919 static char *create_node = 2920 "\" /\" find-device " 2921 "new-device " 2922 "\" os-io\" device-name " 2923 "\" display\" device-type " 2924 ": cb-r/w ( adr,len method$ -- #read/#written ) " 2925 " 2>r swap 2 2r> ['] $callback catch if " 2926 " 2drop 3drop 0 " 2927 " then " 2928 "; " 2929 ": read ( adr,len -- #read ) " 2930 " \" read\" ['] cb-r/w catch if 2drop 2drop -2 exit then " 2931 " ( retN ... ret1 N ) " 2932 " ?dup if " 2933 " swap >r 1- 0 ?do drop loop r> " 2934 " else " 2935 " -2 " 2936 " then " 2937 "; " 2938 ": write ( adr,len -- #written ) " 2939 " \" write\" ['] cb-r/w catch if 2drop 2drop 0 exit then " 2940 " ( retN ... ret1 N ) " 2941 " ?dup if " 2942 " swap >r 1- 0 ?do drop loop r> " 2943 " else " 2944 " 0 " 2945 " then " 2946 "; " 2947 ": poll-tty ( -- ) ; " 2948 ": install-abort ( -- ) ['] poll-tty d# 10 alarm ; " 2949 ": remove-abort ( -- ) ['] poll-tty 0 alarm ; " 2950 ": cb-give/take ( $method -- ) " 2951 " 0 -rot ['] $callback catch ?dup if " 2952 " >r 2drop 2drop r> throw " 2953 " else " 2954 " 0 ?do drop loop " 2955 " then " 2956 "; " 2957 ": give ( -- ) \" exit-input\" cb-give/take ; " 2958 ": take ( -- ) \" enter-input\" cb-give/take ; " 2959 ": open ( -- ok? ) true ; " 2960 ": close ( -- ) ; " 2961 "finish-device " 2962 "device-end "; 2963 2964 /* 2965 * Create the OBP input/output node (FCode serial driver). 2966 * It is needed for both USB console keyboard and for 2967 * the kernel terminal emulator. It is too early to check for a 2968 * kernel console compatible framebuffer now, so we create this 2969 * so that we're ready if we need to enable kernel terminal emulation. 2970 * 2971 * When the USB software takes over the input device at the time 2972 * consconfig runs, OBP's stdin is redirected to this node. 2973 * Whenever the FORTH user interface is used after this switch, 2974 * the node will call back into the kernel for console input. 2975 * If a serial device such as ttya or a UART with a Type 5 keyboard 2976 * attached is used, OBP takes over the serial device when the system 2977 * goes to the debugger after the system is booted. This sharing 2978 * of the relatively simple serial device is difficult but possible. 2979 * Sharing the USB host controller is impossible due its complexity. 2980 * 2981 * Similarly to USB keyboard input redirection, after consconfig_dacf 2982 * configures a kernel console framebuffer as the standard output 2983 * device, OBP's stdout is switched to to vector through the 2984 * /os-io node into the kernel terminal emulator. 2985 */ 2986 static void 2987 startup_create_io_node(void) 2988 { 2989 prom_interpret(create_node, 0, 0, 0, 0, 0); 2990 } 2991 2992 2993 static void 2994 do_prom_version_check(void) 2995 { 2996 int i; 2997 pnode_t node; 2998 char buf[64]; 2999 static char drev[] = "Down-rev firmware detected%s\n" 3000 "\tPlease upgrade to the following minimum version:\n" 3001 "\t\t%s\n"; 3002 3003 i = prom_version_check(buf, sizeof (buf), &node); 3004 3005 if (i == PROM_VER64_OK) 3006 return; 3007 3008 if (i == PROM_VER64_UPGRADE) { 3009 cmn_err(CE_WARN, drev, "", buf); 3010 3011 #ifdef DEBUG 3012 prom_enter_mon(); /* Type 'go' to continue */ 3013 cmn_err(CE_WARN, "Booting with down-rev firmware\n"); 3014 return; 3015 #else 3016 halt(0); 3017 #endif 3018 } 3019 3020 /* 3021 * The other possibility is that this is a server running 3022 * good firmware, but down-rev firmware was detected on at 3023 * least one other cpu board. We just complain if we see 3024 * that. 3025 */ 3026 cmn_err(CE_WARN, drev, " on one or more CPU boards", buf); 3027 } 3028 3029 static void 3030 kpm_init() 3031 { 3032 kpm_pgshft = (kpm_smallpages == 0) ? MMU_PAGESHIFT4M : MMU_PAGESHIFT; 3033 kpm_pgsz = 1ull << kpm_pgshft; 3034 kpm_pgoff = kpm_pgsz - 1; 3035 kpmp2pshft = kpm_pgshft - PAGESHIFT; 3036 kpmpnpgs = 1 << kpmp2pshft; 3037 ASSERT(((uintptr_t)kpm_vbase & (kpm_pgsz - 1)) == 0); 3038 } 3039 3040 void 3041 kpm_npages_setup(int memblocks) 3042 { 3043 /* 3044 * npages can be scattered in a maximum of 'memblocks' 3045 */ 3046 kpm_npages = ptokpmpr(npages) + memblocks; 3047 } 3048 3049 /* 3050 * Must be defined in platform dependent code. 3051 */ 3052 extern caddr_t modtext; 3053 extern size_t modtext_sz; 3054 extern caddr_t moddata; 3055 3056 #define HEAPTEXT_ARENA(addr) \ 3057 ((uintptr_t)(addr) < KERNELBASE + 2 * MMU_PAGESIZE4M ? 0 : \ 3058 (((uintptr_t)(addr) - HEAPTEXT_BASE) / \ 3059 (HEAPTEXT_MAPPED + HEAPTEXT_UNMAPPED) + 1)) 3060 3061 #define HEAPTEXT_OVERSIZED(addr) \ 3062 ((uintptr_t)(addr) >= HEAPTEXT_BASE + HEAPTEXT_SIZE - HEAPTEXT_OVERSIZE) 3063 3064 vmem_t *texthole_source[HEAPTEXT_NARENAS]; 3065 vmem_t *texthole_arena[HEAPTEXT_NARENAS]; 3066 kmutex_t texthole_lock; 3067 3068 char kern_bootargs[OBP_MAXPATHLEN]; 3069 3070 void 3071 kobj_vmem_init(vmem_t **text_arena, vmem_t **data_arena) 3072 { 3073 uintptr_t addr, limit; 3074 3075 addr = HEAPTEXT_BASE; 3076 limit = addr + HEAPTEXT_SIZE - HEAPTEXT_OVERSIZE; 3077 3078 /* 3079 * Before we initialize the text_arena, we want to punch holes in the 3080 * underlying heaptext_arena. This guarantees that for any text 3081 * address we can find a text hole less than HEAPTEXT_MAPPED away. 3082 */ 3083 for (; addr + HEAPTEXT_UNMAPPED <= limit; 3084 addr += HEAPTEXT_MAPPED + HEAPTEXT_UNMAPPED) { 3085 (void) vmem_xalloc(heaptext_arena, HEAPTEXT_UNMAPPED, PAGESIZE, 3086 0, 0, (void *)addr, (void *)(addr + HEAPTEXT_UNMAPPED), 3087 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 3088 } 3089 3090 /* 3091 * Allocate one page at the oversize to break up the text region 3092 * from the oversized region. 3093 */ 3094 (void) vmem_xalloc(heaptext_arena, PAGESIZE, PAGESIZE, 0, 0, 3095 (void *)limit, (void *)(limit + PAGESIZE), 3096 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 3097 3098 *text_arena = vmem_create("module_text", modtext, modtext_sz, 3099 sizeof (uintptr_t), segkmem_alloc, segkmem_free, 3100 heaptext_arena, 0, VM_SLEEP); 3101 *data_arena = vmem_create("module_data", moddata, MODDATA, 1, 3102 segkmem_alloc, segkmem_free, heap32_arena, 0, VM_SLEEP); 3103 } 3104 3105 caddr_t 3106 kobj_text_alloc(vmem_t *arena, size_t size) 3107 { 3108 caddr_t rval, better; 3109 3110 /* 3111 * First, try a sleeping allocation. 3112 */ 3113 rval = vmem_alloc(arena, size, VM_SLEEP | VM_BESTFIT); 3114 3115 if (size >= HEAPTEXT_MAPPED || !HEAPTEXT_OVERSIZED(rval)) 3116 return (rval); 3117 3118 /* 3119 * We didn't get the area that we wanted. We're going to try to do an 3120 * allocation with explicit constraints. 3121 */ 3122 better = vmem_xalloc(arena, size, sizeof (uintptr_t), 0, 0, NULL, 3123 (void *)(HEAPTEXT_BASE + HEAPTEXT_SIZE - HEAPTEXT_OVERSIZE), 3124 VM_NOSLEEP | VM_BESTFIT); 3125 3126 if (better != NULL) { 3127 /* 3128 * That worked. Free our first attempt and return. 3129 */ 3130 vmem_free(arena, rval, size); 3131 return (better); 3132 } 3133 3134 /* 3135 * That didn't work; we'll have to return our first attempt. 3136 */ 3137 return (rval); 3138 } 3139 3140 caddr_t 3141 kobj_texthole_alloc(caddr_t addr, size_t size) 3142 { 3143 int arena = HEAPTEXT_ARENA(addr); 3144 char c[30]; 3145 uintptr_t base; 3146 3147 if (HEAPTEXT_OVERSIZED(addr)) { 3148 /* 3149 * If this is an oversized allocation, there is no text hole 3150 * available for it; return NULL. 3151 */ 3152 return (NULL); 3153 } 3154 3155 mutex_enter(&texthole_lock); 3156 3157 if (texthole_arena[arena] == NULL) { 3158 ASSERT(texthole_source[arena] == NULL); 3159 3160 if (arena == 0) { 3161 texthole_source[0] = vmem_create("module_text_holesrc", 3162 (void *)(KERNELBASE + MMU_PAGESIZE4M), 3163 MMU_PAGESIZE4M, PAGESIZE, NULL, NULL, NULL, 3164 0, VM_SLEEP); 3165 } else { 3166 base = HEAPTEXT_BASE + 3167 (arena - 1) * (HEAPTEXT_MAPPED + HEAPTEXT_UNMAPPED); 3168 3169 (void) snprintf(c, sizeof (c), 3170 "heaptext_holesrc_%d", arena); 3171 3172 texthole_source[arena] = vmem_create(c, (void *)base, 3173 HEAPTEXT_UNMAPPED, PAGESIZE, NULL, NULL, NULL, 3174 0, VM_SLEEP); 3175 } 3176 3177 (void) snprintf(c, sizeof (c), "heaptext_hole_%d", arena); 3178 3179 texthole_arena[arena] = vmem_create(c, NULL, 0, 3180 sizeof (uint32_t), segkmem_alloc_permanent, segkmem_free, 3181 texthole_source[arena], 0, VM_SLEEP); 3182 } 3183 3184 mutex_exit(&texthole_lock); 3185 3186 ASSERT(texthole_arena[arena] != NULL); 3187 ASSERT(arena >= 0 && arena < HEAPTEXT_NARENAS); 3188 return (vmem_alloc(texthole_arena[arena], size, 3189 VM_BESTFIT | VM_NOSLEEP)); 3190 } 3191 3192 void 3193 kobj_texthole_free(caddr_t addr, size_t size) 3194 { 3195 int arena = HEAPTEXT_ARENA(addr); 3196 3197 ASSERT(arena >= 0 && arena < HEAPTEXT_NARENAS); 3198 ASSERT(texthole_arena[arena] != NULL); 3199 vmem_free(texthole_arena[arena], addr, size); 3200 } 3201