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 2008 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 int cpu_intrq_setup(struct cpu *); 71 extern void cpu_intrq_register(struct cpu *); 72 extern void contig_mem_init(void); 73 extern caddr_t contig_mem_prealloc(caddr_t, pgcnt_t); 74 extern void mach_dump_buffer_init(void); 75 extern void mach_descrip_init(void); 76 extern void mach_descrip_startup_fini(void); 77 extern void mach_memscrub(void); 78 extern void mach_fpras(void); 79 extern void mach_cpu_halt_idle(void); 80 extern void mach_hw_copy_limit(void); 81 extern void load_mach_drivers(void); 82 extern void load_tod_module(void); 83 #pragma weak load_tod_module 84 85 extern int ndata_alloc_mmfsa(struct memlist *ndata); 86 #pragma weak ndata_alloc_mmfsa 87 88 extern void cif_init(void); 89 #pragma weak cif_init 90 91 extern void parse_idprom(void); 92 extern void add_vx_handler(char *, int, void (*)(cell_t *)); 93 extern void mem_config_init(void); 94 extern void memseg_remap_init(void); 95 96 extern void mach_kpm_init(void); 97 extern void pcf_init(); 98 extern int size_pse_array(pgcnt_t, int); 99 100 /* 101 * External Data: 102 */ 103 extern int vac_size; /* cache size in bytes */ 104 extern uint_t vac_mask; /* VAC alignment consistency mask */ 105 extern uint_t vac_colors; 106 107 /* 108 * Global Data Definitions: 109 */ 110 111 /* 112 * XXX - Don't port this to new architectures 113 * A 3rd party volume manager driver (vxdm) depends on the symbol romp. 114 * 'romp' has no use with a prom with an IEEE 1275 client interface. 115 * The driver doesn't use the value, but it depends on the symbol. 116 */ 117 void *romp; /* veritas driver won't load without romp 4154976 */ 118 /* 119 * Declare these as initialized data so we can patch them. 120 */ 121 pgcnt_t physmem = 0; /* memory size in pages, patch if you want less */ 122 pgcnt_t segkpsize = 123 btop(SEGKPDEFSIZE); /* size of segkp segment in pages */ 124 uint_t segmap_percent = 12; /* Size of segmap segment */ 125 126 int use_cache = 1; /* cache not reliable (605 bugs) with MP */ 127 int vac_copyback = 1; 128 char *cache_mode = NULL; 129 int use_mix = 1; 130 int prom_debug = 0; 131 132 caddr_t boot_tba; /* %tba at boot - used by kmdb */ 133 uint_t tba_taken_over = 0; 134 135 caddr_t s_text; /* start of kernel text segment */ 136 caddr_t e_text; /* end of kernel text segment */ 137 caddr_t s_data; /* start of kernel data segment */ 138 caddr_t e_data; /* end of kernel data segment */ 139 140 caddr_t modtext; /* beginning of module text */ 141 size_t modtext_sz; /* size of module text */ 142 caddr_t moddata; /* beginning of module data reserve */ 143 caddr_t e_moddata; /* end of module data reserve */ 144 145 /* 146 * End of first block of contiguous kernel in 32-bit virtual address space 147 */ 148 caddr_t econtig32; /* end of first blk of contiguous kernel */ 149 150 caddr_t ncbase; /* beginning of non-cached segment */ 151 caddr_t ncend; /* end of non-cached segment */ 152 153 size_t ndata_remain_sz; /* bytes from end of data to 4MB boundary */ 154 caddr_t nalloc_base; /* beginning of nucleus allocation */ 155 caddr_t nalloc_end; /* end of nucleus allocatable memory */ 156 caddr_t valloc_base; /* beginning of kvalloc segment */ 157 158 caddr_t kmem64_base; /* base of kernel mem segment in 64-bit space */ 159 caddr_t kmem64_end; /* end of kernel mem segment in 64-bit space */ 160 size_t kmem64_sz; /* bytes in kernel mem segment, 64-bit space */ 161 caddr_t kmem64_aligned_end; /* end of large page, overmaps 64-bit space */ 162 int kmem64_szc; /* page size code */ 163 uint64_t kmem64_pabase = (uint64_t)-1; /* physical address of kmem64_base */ 164 165 uintptr_t shm_alignment; /* VAC address consistency modulus */ 166 struct memlist *phys_install; /* Total installed physical memory */ 167 struct memlist *phys_avail; /* Available (unreserved) physical memory */ 168 struct memlist *virt_avail; /* Available (unmapped?) virtual memory */ 169 struct memlist *nopp_list; /* pages with no backing page structs */ 170 struct memlist ndata; /* memlist of nucleus allocatable memory */ 171 int memexp_flag; /* memory expansion card flag */ 172 uint64_t ecache_flushaddr; /* physical address used for flushing E$ */ 173 pgcnt_t obp_pages; /* Physical pages used by OBP */ 174 175 /* 176 * VM data structures 177 */ 178 long page_hashsz; /* Size of page hash table (power of two) */ 179 struct page *pp_base; /* Base of system page struct array */ 180 size_t pp_sz; /* Size in bytes of page struct array */ 181 struct page **page_hash; /* Page hash table */ 182 pad_mutex_t *pse_mutex; /* Locks protecting pp->p_selock */ 183 size_t pse_table_size; /* Number of mutexes in pse_mutex[] */ 184 int pse_shift; /* log2(pse_table_size) */ 185 struct seg ktextseg; /* Segment used for kernel executable image */ 186 struct seg kvalloc; /* Segment used for "valloc" mapping */ 187 struct seg kpseg; /* Segment used for pageable kernel virt mem */ 188 struct seg ktexthole; /* Segment used for nucleus text hole */ 189 struct seg kmapseg; /* Segment used for generic kernel mappings */ 190 struct seg kpmseg; /* Segment used for physical mapping */ 191 struct seg kdebugseg; /* Segment used for the kernel debugger */ 192 193 void *kpm_pp_base; /* Base of system kpm_page array */ 194 size_t kpm_pp_sz; /* Size of system kpm_page array */ 195 pgcnt_t kpm_npages; /* How many kpm pages are managed */ 196 197 struct seg *segkp = &kpseg; /* Pageable kernel virtual memory segment */ 198 struct seg *segkmap = &kmapseg; /* Kernel generic mapping segment */ 199 struct seg *segkpm = &kpmseg; /* 64bit kernel physical mapping segment */ 200 201 int segzio_fromheap = 0; /* zio allocations occur from heap */ 202 caddr_t segzio_base; /* Base address of segzio */ 203 pgcnt_t segziosize = 0; /* size of zio segment in pages */ 204 205 /* 206 * debugger pages (if allocated) 207 */ 208 struct vnode kdebugvp; 209 210 /* 211 * VA range available to the debugger 212 */ 213 const caddr_t kdi_segdebugbase = (const caddr_t)SEGDEBUGBASE; 214 const size_t kdi_segdebugsize = SEGDEBUGSIZE; 215 216 /* 217 * Segment for relocated kernel structures in 64-bit large RAM kernels 218 */ 219 struct seg kmem64; 220 221 struct memseg *memseg_free; 222 223 struct vnode unused_pages_vp; 224 225 /* 226 * VM data structures allocated early during boot. 227 */ 228 size_t pagehash_sz; 229 uint64_t memlist_sz; 230 231 char tbr_wr_addr_inited = 0; 232 233 caddr_t mpo_heap32_buf = NULL; 234 size_t mpo_heap32_bufsz = 0; 235 236 /* 237 * Static Routines: 238 */ 239 static int ndata_alloc_memseg(struct memlist *, size_t); 240 static void memlist_new(uint64_t, uint64_t, struct memlist **); 241 static void memlist_add(uint64_t, uint64_t, 242 struct memlist **, struct memlist **); 243 static void kphysm_init(void); 244 static void kvm_init(void); 245 static void install_kmem64_tte(void); 246 247 static void startup_init(void); 248 static void startup_memlist(void); 249 static void startup_modules(void); 250 static void startup_bop_gone(void); 251 static void startup_vm(void); 252 static void startup_end(void); 253 static void setup_cage_params(void); 254 static void startup_create_io_node(void); 255 256 static pgcnt_t npages; 257 static struct memlist *memlist; 258 void *memlist_end; 259 260 static pgcnt_t bop_alloc_pages; 261 static caddr_t hblk_base; 262 uint_t hblk_alloc_dynamic = 0; 263 uint_t hblk1_min = H1MIN; 264 265 266 /* 267 * Hooks for unsupported platforms and down-rev firmware 268 */ 269 int iam_positron(void); 270 #pragma weak iam_positron 271 static void do_prom_version_check(void); 272 273 /* 274 * After receiving a thermal interrupt, this is the number of seconds 275 * to delay before shutting off the system, assuming 276 * shutdown fails. Use /etc/system to change the delay if this isn't 277 * large enough. 278 */ 279 int thermal_powerdown_delay = 1200; 280 281 /* 282 * Used to hold off page relocations into the cage until OBP has completed 283 * its boot-time handoff of its resources to the kernel. 284 */ 285 int page_relocate_ready = 0; 286 287 /* 288 * Enable some debugging messages concerning memory usage... 289 */ 290 #ifdef DEBUGGING_MEM 291 static int debugging_mem; 292 static void 293 printmemlist(char *title, struct memlist *list) 294 { 295 if (!debugging_mem) 296 return; 297 298 printf("%s\n", title); 299 300 while (list) { 301 prom_printf("\taddr = 0x%x %8x, size = 0x%x %8x\n", 302 (uint32_t)(list->address >> 32), (uint32_t)list->address, 303 (uint32_t)(list->size >> 32), (uint32_t)(list->size)); 304 list = list->next; 305 } 306 } 307 308 void 309 printmemseg(struct memseg *memseg) 310 { 311 if (!debugging_mem) 312 return; 313 314 printf("memseg\n"); 315 316 while (memseg) { 317 prom_printf("\tpage = 0x%p, epage = 0x%p, " 318 "pfn = 0x%x, epfn = 0x%x\n", 319 memseg->pages, memseg->epages, 320 memseg->pages_base, memseg->pages_end); 321 memseg = memseg->next; 322 } 323 } 324 325 #define debug_pause(str) halt((str)) 326 #define MPRINTF(str) if (debugging_mem) prom_printf((str)) 327 #define MPRINTF1(str, a) if (debugging_mem) prom_printf((str), (a)) 328 #define MPRINTF2(str, a, b) if (debugging_mem) prom_printf((str), (a), (b)) 329 #define MPRINTF3(str, a, b, c) \ 330 if (debugging_mem) prom_printf((str), (a), (b), (c)) 331 #else /* DEBUGGING_MEM */ 332 #define MPRINTF(str) 333 #define MPRINTF1(str, a) 334 #define MPRINTF2(str, a, b) 335 #define MPRINTF3(str, a, b, c) 336 #endif /* DEBUGGING_MEM */ 337 338 339 /* 340 * 341 * Kernel's Virtual Memory Layout. 342 * /-----------------------\ 343 * 0xFFFFFFFF.FFFFFFFF -| |- 344 * | OBP's virtual page | 345 * | tables | 346 * 0xFFFFFFFC.00000000 -|-----------------------|- 347 * : : 348 * : : 349 * -|-----------------------|- 350 * | segzio | (base and size vary) 351 * 0xFFFFFE00.00000000 -|-----------------------|- 352 * | | Ultrasparc I/II support 353 * | segkpm segment | up to 2TB of physical 354 * | (64-bit kernel ONLY) | memory, VAC has 2 colors 355 * | | 356 * 0xFFFFFA00.00000000 -|-----------------------|- 2TB segkpm alignment 357 * : : 358 * : : 359 * 0xFFFFF810.00000000 -|-----------------------|- hole_end 360 * | | ^ 361 * | UltraSPARC I/II call | | 362 * | bug requires an extra | | 363 * | 4 GB of space between | | 364 * | hole and used RAM | | 365 * | | | 366 * 0xFFFFF800.00000000 -|-----------------------|- | 367 * | | | 368 * | Virtual Address Hole | UltraSPARC 369 * | on UltraSPARC I/II | I/II * ONLY * 370 * | | | 371 * 0x00000800.00000000 -|-----------------------|- | 372 * | | | 373 * | UltraSPARC I/II call | | 374 * | bug requires an extra | | 375 * | 4 GB of space between | | 376 * | hole and used RAM | | 377 * | | v 378 * 0x000007FF.00000000 -|-----------------------|- hole_start ----- 379 * : : ^ 380 * : : | 381 * |-----------------------| | 382 * | | | 383 * | ecache flush area | | 384 * | (twice largest e$) | | 385 * | | | 386 * 0x00000XXX.XXX00000 -|-----------------------|- kmem64_ | 387 * | overmapped area | alignend_end | 388 * | (kmem64_alignsize | | 389 * | boundary) | | 390 * 0x00000XXX.XXXXXXXX -|-----------------------|- kmem64_end | 391 * | | | 392 * | 64-bit kernel ONLY | | 393 * | | | 394 * | kmem64 segment | | 395 * | | | 396 * | (Relocated extra HME | Approximately 397 * | block allocations, | 1 TB of virtual 398 * | memnode freelists, | address space 399 * | HME hash buckets, | | 400 * | mml_table, kpmp_table,| | 401 * | page_t array and | | 402 * | hashblock pool to | | 403 * | avoid hard-coded | | 404 * | 32-bit vaddr | | 405 * | limitations) | | 406 * | | v 407 * 0x00000700.00000000 -|-----------------------|- SYSLIMIT (kmem64_base) 408 * | | 409 * | segkmem segment | (SYSLIMIT - SYSBASE = 4TB) 410 * | | 411 * 0x00000300.00000000 -|-----------------------|- SYSBASE 412 * : : 413 * : : 414 * -|-----------------------|- 415 * | | 416 * | segmap segment | SEGMAPSIZE (1/8th physmem, 417 * | | 256G MAX) 418 * 0x000002a7.50000000 -|-----------------------|- SEGMAPBASE 419 * : : 420 * : : 421 * -|-----------------------|- 422 * | | 423 * | segkp | SEGKPSIZE (2GB) 424 * | | 425 * | | 426 * 0x000002a1.00000000 -|-----------------------|- SEGKPBASE 427 * | | 428 * 0x000002a0.00000000 -|-----------------------|- MEMSCRUBBASE 429 * | | (SEGKPBASE - 0x400000) 430 * 0x0000029F.FFE00000 -|-----------------------|- ARGSBASE 431 * | | (MEMSCRUBBASE - NCARGS) 432 * 0x0000029F.FFD80000 -|-----------------------|- PPMAPBASE 433 * | | (ARGSBASE - PPMAPSIZE) 434 * 0x0000029F.FFD00000 -|-----------------------|- PPMAP_FAST_BASE 435 * | | 436 * 0x0000029F.FF980000 -|-----------------------|- PIOMAPBASE 437 * | | 438 * 0x0000029F.FF580000 -|-----------------------|- NARG_BASE 439 * : : 440 * : : 441 * 0x00000000.FFFFFFFF -|-----------------------|- OFW_END_ADDR 442 * | | 443 * | OBP | 444 * | | 445 * 0x00000000.F0000000 -|-----------------------|- OFW_START_ADDR 446 * | kmdb | 447 * 0x00000000.EDD00000 -|-----------------------|- SEGDEBUGBASE 448 * : : 449 * : : 450 * 0x00000000.7c000000 -|-----------------------|- SYSLIMIT32 451 * | | 452 * | segkmem32 segment | (SYSLIMIT32 - SYSBASE32 = 453 * | | ~64MB) 454 * 0x00000000.70002000 -|-----------------------| 455 * | panicbuf | 456 * 0x00000000.70000000 -|-----------------------|- SYSBASE32 457 * | boot-time | 458 * | temporary space | 459 * 0x00000000.4C000000 -|-----------------------|- BOOTTMPBASE 460 * : : 461 * : : 462 * | | 463 * |-----------------------|- econtig32 464 * | vm structures | 465 * 0x00000000.01C00000 |-----------------------|- nalloc_end 466 * | TSBs | 467 * |-----------------------|- end/nalloc_base 468 * | kernel data & bss | 469 * 0x00000000.01800000 -|-----------------------| 470 * : nucleus text hole : 471 * 0x00000000.01400000 -|-----------------------| 472 * : : 473 * |-----------------------| 474 * | module text | 475 * |-----------------------|- e_text/modtext 476 * | kernel text | 477 * |-----------------------| 478 * | trap table (48k) | 479 * 0x00000000.01000000 -|-----------------------|- KERNELBASE 480 * | reserved for trapstat |} TSTAT_TOTAL_SIZE 481 * |-----------------------| 482 * | | 483 * | invalid | 484 * | | 485 * 0x00000000.00000000 _|_______________________| 486 * 487 * 488 * 489 * 32-bit User Virtual Memory Layout. 490 * /-----------------------\ 491 * | | 492 * | invalid | 493 * | | 494 * 0xFFC00000 -|-----------------------|- USERLIMIT 495 * | user stack | 496 * : : 497 * : : 498 * : : 499 * | user data | 500 * -|-----------------------|- 501 * | user text | 502 * 0x00002000 -|-----------------------|- 503 * | invalid | 504 * 0x00000000 _|_______________________| 505 * 506 * 507 * 508 * 64-bit User Virtual Memory Layout. 509 * /-----------------------\ 510 * | | 511 * | invalid | 512 * | | 513 * 0xFFFFFFFF.80000000 -|-----------------------|- USERLIMIT 514 * | user stack | 515 * : : 516 * : : 517 * : : 518 * | user data | 519 * -|-----------------------|- 520 * | user text | 521 * 0x00000000.01000000 -|-----------------------|- 522 * | invalid | 523 * 0x00000000.00000000 _|_______________________| 524 */ 525 526 extern caddr_t ecache_init_scrub_flush_area(caddr_t alloc_base); 527 extern uint64_t ecache_flush_address(void); 528 529 #pragma weak load_platform_modules 530 #pragma weak plat_startup_memlist 531 #pragma weak ecache_init_scrub_flush_area 532 #pragma weak ecache_flush_address 533 534 535 /* 536 * By default the DR Cage is enabled for maximum OS 537 * MPSS performance. Users needing to disable the cage mechanism 538 * can set this variable to zero via /etc/system. 539 * Disabling the cage on systems supporting Dynamic Reconfiguration (DR) 540 * will result in loss of DR functionality. 541 * Platforms wishing to disable kernel Cage by default 542 * should do so in their set_platform_defaults() routine. 543 */ 544 int kernel_cage_enable = 1; 545 546 static void 547 setup_cage_params(void) 548 { 549 void (*func)(void); 550 551 func = (void (*)(void))kobj_getsymvalue("set_platform_cage_params", 0); 552 if (func != NULL) { 553 (*func)(); 554 return; 555 } 556 557 if (kernel_cage_enable == 0) { 558 return; 559 } 560 kcage_range_init(phys_avail, KCAGE_DOWN, total_pages / 256); 561 562 if (kcage_on) { 563 cmn_err(CE_NOTE, "!Kernel Cage is ENABLED"); 564 } else { 565 cmn_err(CE_NOTE, "!Kernel Cage is DISABLED"); 566 } 567 568 } 569 570 /* 571 * Machine-dependent startup code 572 */ 573 void 574 startup(void) 575 { 576 startup_init(); 577 if (&startup_platform) 578 startup_platform(); 579 startup_memlist(); 580 startup_modules(); 581 setup_cage_params(); 582 startup_bop_gone(); 583 startup_vm(); 584 startup_end(); 585 } 586 587 struct regs sync_reg_buf; 588 uint64_t sync_tt; 589 590 void 591 sync_handler(void) 592 { 593 struct panic_trap_info ti; 594 int i; 595 596 /* 597 * Prevent trying to talk to the other CPUs since they are 598 * sitting in the prom and won't reply. 599 */ 600 for (i = 0; i < NCPU; i++) { 601 if ((i != CPU->cpu_id) && CPU_XCALL_READY(i)) { 602 cpu[i]->cpu_flags &= ~CPU_READY; 603 cpu[i]->cpu_flags |= CPU_QUIESCED; 604 CPUSET_DEL(cpu_ready_set, cpu[i]->cpu_id); 605 } 606 } 607 608 /* 609 * We've managed to get here without going through the 610 * normal panic code path. Try and save some useful 611 * information. 612 */ 613 if (!panicstr && (curthread->t_panic_trap == NULL)) { 614 ti.trap_type = sync_tt; 615 ti.trap_regs = &sync_reg_buf; 616 ti.trap_addr = NULL; 617 ti.trap_mmu_fsr = 0x0; 618 619 curthread->t_panic_trap = &ti; 620 } 621 622 /* 623 * If we're re-entering the panic path, update the signature 624 * block so that the SC knows we're in the second part of panic. 625 */ 626 if (panicstr) 627 CPU_SIGNATURE(OS_SIG, SIGST_EXIT, SIGSUBST_DUMP, -1); 628 629 nopanicdebug = 1; /* do not perform debug_enter() prior to dump */ 630 panic("sync initiated"); 631 } 632 633 634 static void 635 startup_init(void) 636 { 637 /* 638 * We want to save the registers while we're still in OBP 639 * so that we know they haven't been fiddled with since. 640 * (In principle, OBP can't change them just because it 641 * makes a callback, but we'd rather not depend on that 642 * behavior.) 643 */ 644 char sync_str[] = 645 "warning @ warning off : sync " 646 "%%tl-c %%tstate h# %p x! " 647 "%%g1 h# %p x! %%g2 h# %p x! %%g3 h# %p x! " 648 "%%g4 h# %p x! %%g5 h# %p x! %%g6 h# %p x! " 649 "%%g7 h# %p x! %%o0 h# %p x! %%o1 h# %p x! " 650 "%%o2 h# %p x! %%o3 h# %p x! %%o4 h# %p x! " 651 "%%o5 h# %p x! %%o6 h# %p x! %%o7 h# %p x! " 652 "%%tl-c %%tpc h# %p x! %%tl-c %%tnpc h# %p x! " 653 "%%y h# %p l! %%tl-c %%tt h# %p x! " 654 "sync ; warning !"; 655 656 /* 657 * 20 == num of %p substrings 658 * 16 == max num of chars %p will expand to. 659 */ 660 char bp[sizeof (sync_str) + 16 * 20]; 661 662 /* 663 * Initialize ptl1 stack for the 1st CPU. 664 */ 665 ptl1_init_cpu(&cpu0); 666 667 /* 668 * Initialize the address map for cache consistent mappings 669 * to random pages; must be done after vac_size is set. 670 */ 671 ppmapinit(); 672 673 /* 674 * Initialize the PROM callback handler. 675 */ 676 init_vx_handler(); 677 678 /* 679 * have prom call sync_callback() to handle the sync and 680 * save some useful information which will be stored in the 681 * core file later. 682 */ 683 (void) sprintf((char *)bp, sync_str, 684 (void *)&sync_reg_buf.r_tstate, (void *)&sync_reg_buf.r_g1, 685 (void *)&sync_reg_buf.r_g2, (void *)&sync_reg_buf.r_g3, 686 (void *)&sync_reg_buf.r_g4, (void *)&sync_reg_buf.r_g5, 687 (void *)&sync_reg_buf.r_g6, (void *)&sync_reg_buf.r_g7, 688 (void *)&sync_reg_buf.r_o0, (void *)&sync_reg_buf.r_o1, 689 (void *)&sync_reg_buf.r_o2, (void *)&sync_reg_buf.r_o3, 690 (void *)&sync_reg_buf.r_o4, (void *)&sync_reg_buf.r_o5, 691 (void *)&sync_reg_buf.r_o6, (void *)&sync_reg_buf.r_o7, 692 (void *)&sync_reg_buf.r_pc, (void *)&sync_reg_buf.r_npc, 693 (void *)&sync_reg_buf.r_y, (void *)&sync_tt); 694 prom_interpret(bp, 0, 0, 0, 0, 0); 695 add_vx_handler("sync", 1, (void (*)(cell_t *))sync_handler); 696 } 697 698 699 size_t 700 calc_pp_sz(pgcnt_t npages) 701 { 702 703 return (npages * sizeof (struct page)); 704 } 705 706 size_t 707 calc_kpmpp_sz(pgcnt_t npages) 708 { 709 710 kpm_pgshft = (kpm_smallpages == 0) ? MMU_PAGESHIFT4M : MMU_PAGESHIFT; 711 kpm_pgsz = 1ull << kpm_pgshft; 712 kpm_pgoff = kpm_pgsz - 1; 713 kpmp2pshft = kpm_pgshft - PAGESHIFT; 714 kpmpnpgs = 1 << kpmp2pshft; 715 716 if (kpm_smallpages == 0) { 717 /* 718 * Avoid fragmentation problems in kphysm_init() 719 * by allocating for all of physical memory 720 */ 721 kpm_npages = ptokpmpr(physinstalled); 722 return (kpm_npages * sizeof (kpm_page_t)); 723 } else { 724 kpm_npages = npages; 725 return (kpm_npages * sizeof (kpm_spage_t)); 726 } 727 } 728 729 size_t 730 calc_pagehash_sz(pgcnt_t npages) 731 { 732 733 /* 734 * The page structure hash table size is a power of 2 735 * such that the average hash chain length is PAGE_HASHAVELEN. 736 */ 737 page_hashsz = npages / PAGE_HASHAVELEN; 738 page_hashsz = 1 << highbit(page_hashsz); 739 return (page_hashsz * sizeof (struct page *)); 740 } 741 742 void 743 alloc_kmem64(caddr_t base, caddr_t end) 744 { 745 int i; 746 caddr_t aligned_end = NULL; 747 748 /* 749 * Make one large memory alloc after figuring out the 64-bit size. This 750 * will enable use of the largest page size appropriate for the system 751 * architecture. 752 */ 753 ASSERT(mmu_exported_pagesize_mask & (1 << TTE8K)); 754 ASSERT(IS_P2ALIGNED(base, TTEBYTES(max_bootlp_tteszc))); 755 for (i = max_bootlp_tteszc; i >= TTE8K; i--) { 756 size_t alloc_size, alignsize; 757 #if !defined(C_OBP) 758 unsigned long long pa; 759 #endif /* !C_OBP */ 760 761 if ((mmu_exported_pagesize_mask & (1 << i)) == 0) 762 continue; 763 alignsize = TTEBYTES(i); 764 kmem64_szc = i; 765 766 /* limit page size for small memory */ 767 if (mmu_btop(alignsize) > (npages >> 2)) 768 continue; 769 770 aligned_end = (caddr_t)roundup((uintptr_t)end, alignsize); 771 alloc_size = aligned_end - base; 772 #if !defined(C_OBP) 773 if (prom_allocate_phys(alloc_size, alignsize, &pa) == 0) { 774 if (prom_claim_virt(alloc_size, base) != (caddr_t)-1) { 775 kmem64_pabase = pa; 776 kmem64_aligned_end = aligned_end; 777 install_kmem64_tte(); 778 break; 779 } else { 780 prom_free_phys(alloc_size, pa); 781 } 782 } 783 #else /* !C_OBP */ 784 if (prom_alloc(base, alloc_size, alignsize) == base) { 785 kmem64_pabase = va_to_pa(kmem64_base); 786 kmem64_aligned_end = aligned_end; 787 break; 788 } 789 #endif /* !C_OBP */ 790 if (i == TTE8K) { 791 prom_panic("kmem64 allocation failure"); 792 } 793 } 794 ASSERT(aligned_end != NULL); 795 } 796 797 static prom_memlist_t *boot_physinstalled, *boot_physavail, *boot_virtavail; 798 static size_t boot_physinstalled_len, boot_physavail_len, boot_virtavail_len; 799 800 #define IVSIZE roundup(((MAXIVNUM * sizeof (intr_vec_t *)) + \ 801 (MAX_RSVD_IV * sizeof (intr_vec_t)) + \ 802 (MAX_RSVD_IVX * sizeof (intr_vecx_t))), PAGESIZE) 803 804 #if !defined(C_OBP) 805 /* 806 * Install a temporary tte handler in OBP for kmem64 area. 807 * 808 * We map kmem64 area with large pages before the trap table is taken 809 * over. Since OBP makes 8K mappings, it can create 8K tlb entries in 810 * the same area. Duplicate tlb entries with different page sizes 811 * cause unpredicatble behavior. To avoid this, we don't create 812 * kmem64 mappings via BOP_ALLOC (ends up as prom_alloc() call to 813 * OBP). Instead, we manage translations with a temporary va>tte-data 814 * handler (kmem64-tte). This handler is replaced by unix-tte when 815 * the trap table is taken over. 816 * 817 * The temporary handler knows the physical address of the kmem64 818 * area. It uses the prom's pgmap@ Forth word for other addresses. 819 * 820 * We have to use BOP_ALLOC() method for C-OBP platforms because 821 * pgmap@ is not defined in C-OBP. C-OBP is only used on serengeti 822 * sun4u platforms. On sun4u we flush tlb after trap table is taken 823 * over if we use large pages for kernel heap and kmem64. Since sun4u 824 * prom (unlike sun4v) calls va>tte-data first for client address 825 * translation prom's ttes for kmem64 can't get into TLB even if we 826 * later switch to prom's trap table again. C-OBP uses 4M pages for 827 * client mappings when possible so on all platforms we get the 828 * benefit from large mappings for kmem64 area immediately during 829 * boot. 830 * 831 * pseudo code: 832 * if (context != 0) { 833 * return false 834 * } else if (miss_va in range[kmem64_base, kmem64_end)) { 835 * tte = tte_template + 836 * (((miss_va & pagemask) - kmem64_base)); 837 * return tte, true 838 * } else { 839 * return pgmap@ result 840 * } 841 */ 842 char kmem64_obp_str[] = 843 "h# %lx constant kmem64-base " 844 "h# %lx constant kmem64-end " 845 "h# %lx constant kmem64-pagemask " 846 "h# %lx constant kmem64-template " 847 848 ": kmem64-tte ( addr cnum -- false | tte-data true ) " 849 " if ( addr ) " 850 " drop false exit then ( false ) " 851 " dup kmem64-base kmem64-end within if ( addr ) " 852 " kmem64-pagemask and ( addr' ) " 853 " kmem64-base - ( addr' ) " 854 " kmem64-template + ( tte ) " 855 " true ( tte true ) " 856 " else ( addr ) " 857 " pgmap@ ( tte ) " 858 " dup 0< if true else drop false then ( tte true | false ) " 859 " then ( tte true | false ) " 860 "; " 861 862 "' kmem64-tte is va>tte-data " 863 ; 864 865 static void 866 install_kmem64_tte() 867 { 868 char b[sizeof (kmem64_obp_str) + (4 * 16)]; 869 tte_t tte; 870 871 PRM_DEBUG(kmem64_pabase); 872 PRM_DEBUG(kmem64_szc); 873 sfmmu_memtte(&tte, kmem64_pabase >> MMU_PAGESHIFT, 874 PROC_DATA | HAT_NOSYNC, kmem64_szc); 875 PRM_DEBUG(tte.ll); 876 (void) sprintf(b, kmem64_obp_str, 877 kmem64_base, kmem64_end, TTE_PAGEMASK(kmem64_szc), tte.ll); 878 ASSERT(strlen(b) < sizeof (b)); 879 prom_interpret(b, 0, 0, 0, 0, 0); 880 } 881 #endif /* !C_OBP */ 882 883 /* 884 * As OBP takes up some RAM when the system boots, pages will already be "lost" 885 * to the system and reflected in npages by the time we see it. 886 * 887 * We only want to allocate kernel structures in the 64-bit virtual address 888 * space on systems with enough RAM to make the overhead of keeping track of 889 * an extra kernel memory segment worthwhile. 890 * 891 * Since OBP has already performed its memory allocations by this point, if we 892 * have more than MINMOVE_RAM_MB MB of RAM left free, go ahead and map 893 * memory in the 64-bit virtual address space; otherwise keep allocations 894 * contiguous with we've mapped so far in the 32-bit virtual address space. 895 */ 896 #define MINMOVE_RAM_MB ((size_t)1900) 897 #define MB_TO_BYTES(mb) ((mb) * 1048576ul) 898 #define BYTES_TO_MB(b) ((b) / 1048576ul) 899 900 pgcnt_t tune_npages = (pgcnt_t) 901 (MB_TO_BYTES(MINMOVE_RAM_MB)/ (size_t)MMU_PAGESIZE); 902 903 #pragma weak page_set_colorequiv_arr_cpu 904 extern void page_set_colorequiv_arr_cpu(void); 905 extern void page_set_colorequiv_arr(void); 906 907 static pgcnt_t ramdisk_npages; 908 static struct memlist *old_phys_avail; 909 910 kcage_dir_t kcage_startup_dir = KCAGE_DOWN; 911 912 static void 913 startup_memlist(void) 914 { 915 size_t hmehash_sz, pagelist_sz, tt_sz; 916 size_t psetable_sz; 917 caddr_t alloc_base; 918 caddr_t memspace; 919 struct memlist *cur; 920 size_t syslimit = (size_t)SYSLIMIT; 921 size_t sysbase = (size_t)SYSBASE; 922 923 /* 924 * Initialize enough of the system to allow kmem_alloc to work by 925 * calling boot to allocate its memory until the time that 926 * kvm_init is completed. The page structs are allocated after 927 * rounding up end to the nearest page boundary; the memsegs are 928 * initialized and the space they use comes from the kernel heap. 929 * With appropriate initialization, they can be reallocated later 930 * to a size appropriate for the machine's configuration. 931 * 932 * At this point, memory is allocated for things that will never 933 * need to be freed, this used to be "valloced". This allows a 934 * savings as the pages don't need page structures to describe 935 * them because them will not be managed by the vm system. 936 */ 937 938 /* 939 * We're loaded by boot with the following configuration (as 940 * specified in the sun4u/conf/Mapfile): 941 * 942 * text: 4 MB chunk aligned on a 4MB boundary 943 * data & bss: 4 MB chunk aligned on a 4MB boundary 944 * 945 * These two chunks will eventually be mapped by 2 locked 4MB 946 * ttes and will represent the nucleus of the kernel. This gives 947 * us some free space that is already allocated, some or all of 948 * which is made available to kernel module text. 949 * 950 * The free space in the data-bss chunk is used for nucleus 951 * allocatable data structures and we reserve it using the 952 * nalloc_base and nalloc_end variables. This space is currently 953 * being used for hat data structures required for tlb miss 954 * handling operations. We align nalloc_base to a l2 cache 955 * linesize because this is the line size the hardware uses to 956 * maintain cache coherency. 957 * 512K is carved out for module data. 958 */ 959 960 moddata = (caddr_t)roundup((uintptr_t)e_data, MMU_PAGESIZE); 961 e_moddata = moddata + MODDATA; 962 nalloc_base = e_moddata; 963 964 nalloc_end = (caddr_t)roundup((uintptr_t)nalloc_base, MMU_PAGESIZE4M); 965 valloc_base = nalloc_base; 966 967 /* 968 * Calculate the start of the data segment. 969 */ 970 if (((uintptr_t)e_moddata & MMU_PAGEMASK4M) != (uintptr_t)s_data) 971 prom_panic("nucleus data overflow"); 972 973 PRM_DEBUG(moddata); 974 PRM_DEBUG(nalloc_base); 975 PRM_DEBUG(nalloc_end); 976 977 /* 978 * Remember any slop after e_text so we can give it to the modules. 979 */ 980 PRM_DEBUG(e_text); 981 modtext = (caddr_t)roundup((uintptr_t)e_text, MMU_PAGESIZE); 982 if (((uintptr_t)e_text & MMU_PAGEMASK4M) != (uintptr_t)s_text) 983 prom_panic("nucleus text overflow"); 984 modtext_sz = (caddr_t)roundup((uintptr_t)modtext, MMU_PAGESIZE4M) - 985 modtext; 986 PRM_DEBUG(modtext); 987 PRM_DEBUG(modtext_sz); 988 989 init_boot_memlists(); 990 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 991 &boot_physavail, &boot_physavail_len, 992 &boot_virtavail, &boot_virtavail_len); 993 994 /* 995 * Remember what the physically available highest page is 996 * so that dumpsys works properly, and find out how much 997 * memory is installed. 998 */ 999 installed_top_size_memlist_array(boot_physinstalled, 1000 boot_physinstalled_len, &physmax, &physinstalled); 1001 PRM_DEBUG(physinstalled); 1002 PRM_DEBUG(physmax); 1003 1004 /* Fill out memory nodes config structure */ 1005 startup_build_mem_nodes(boot_physinstalled, boot_physinstalled_len); 1006 1007 /* 1008 * npages is the maximum of available physical memory possible. 1009 * (ie. it will never be more than this) 1010 * 1011 * When we boot from a ramdisk, the ramdisk memory isn't free, so 1012 * using phys_avail will underestimate what will end up being freed. 1013 * A better initial guess is just total memory minus the kernel text 1014 */ 1015 npages = physinstalled - btop(MMU_PAGESIZE4M); 1016 1017 /* 1018 * First allocate things that can go in the nucleus data page 1019 * (fault status, TSBs, dmv, CPUs) 1020 */ 1021 ndata_alloc_init(&ndata, (uintptr_t)nalloc_base, (uintptr_t)nalloc_end); 1022 1023 if ((&ndata_alloc_mmfsa != NULL) && (ndata_alloc_mmfsa(&ndata) != 0)) 1024 cmn_err(CE_PANIC, "no more nucleus memory after mfsa alloc"); 1025 1026 if (ndata_alloc_tsbs(&ndata, npages) != 0) 1027 cmn_err(CE_PANIC, "no more nucleus memory after tsbs alloc"); 1028 1029 if (ndata_alloc_dmv(&ndata) != 0) 1030 cmn_err(CE_PANIC, "no more nucleus memory after dmv alloc"); 1031 1032 if (ndata_alloc_page_mutexs(&ndata) != 0) 1033 cmn_err(CE_PANIC, 1034 "no more nucleus memory after page free lists alloc"); 1035 1036 if (ndata_alloc_hat(&ndata, npages) != 0) 1037 cmn_err(CE_PANIC, "no more nucleus memory after hat alloc"); 1038 1039 if (ndata_alloc_memseg(&ndata, boot_physavail_len) != 0) 1040 cmn_err(CE_PANIC, "no more nucleus memory after memseg alloc"); 1041 1042 /* 1043 * WARNING WARNING WARNING WARNING WARNING WARNING WARNING 1044 * 1045 * There are comments all over the SFMMU code warning of dire 1046 * consequences if the TSBs are moved out of 32-bit space. This 1047 * is largely because the asm code uses "sethi %hi(addr)"-type 1048 * instructions which will not provide the expected result if the 1049 * address is a 64-bit one. 1050 * 1051 * WARNING WARNING WARNING WARNING WARNING WARNING WARNING 1052 */ 1053 alloc_base = (caddr_t)roundup((uintptr_t)nalloc_end, MMU_PAGESIZE); 1054 PRM_DEBUG(alloc_base); 1055 1056 alloc_base = sfmmu_ktsb_alloc(alloc_base); 1057 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, ecache_alignsize); 1058 PRM_DEBUG(alloc_base); 1059 1060 /* 1061 * Allocate IOMMU TSB array. We do this here so that the physical 1062 * memory gets deducted from the PROM's physical memory list. 1063 */ 1064 alloc_base = iommu_tsb_init(alloc_base); 1065 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, ecache_alignsize); 1066 PRM_DEBUG(alloc_base); 1067 1068 /* 1069 * Allow for an early allocation of physically contiguous memory. 1070 */ 1071 alloc_base = contig_mem_prealloc(alloc_base, npages); 1072 1073 /* 1074 * Platforms like Starcat and OPL need special structures assigned in 1075 * 32-bit virtual address space because their probing routines execute 1076 * FCode, and FCode can't handle 64-bit virtual addresses... 1077 */ 1078 if (&plat_startup_memlist) { 1079 alloc_base = plat_startup_memlist(alloc_base); 1080 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1081 ecache_alignsize); 1082 PRM_DEBUG(alloc_base); 1083 } 1084 1085 /* 1086 * Save off where the contiguous allocations to date have ended 1087 * in econtig32. 1088 */ 1089 econtig32 = alloc_base; 1090 PRM_DEBUG(econtig32); 1091 if (econtig32 > (caddr_t)KERNEL_LIMIT32) 1092 cmn_err(CE_PANIC, "econtig32 too big"); 1093 1094 pp_sz = calc_pp_sz(npages); 1095 PRM_DEBUG(pp_sz); 1096 if (kpm_enable) { 1097 kpm_pp_sz = calc_kpmpp_sz(npages); 1098 PRM_DEBUG(kpm_pp_sz); 1099 } 1100 1101 hmehash_sz = calc_hmehash_sz(npages); 1102 PRM_DEBUG(hmehash_sz); 1103 1104 pagehash_sz = calc_pagehash_sz(npages); 1105 PRM_DEBUG(pagehash_sz); 1106 1107 pagelist_sz = calc_free_pagelist_sz(); 1108 PRM_DEBUG(pagelist_sz); 1109 1110 #ifdef TRAPTRACE 1111 tt_sz = calc_traptrace_sz(); 1112 PRM_DEBUG(tt_sz); 1113 #else 1114 tt_sz = 0; 1115 #endif /* TRAPTRACE */ 1116 1117 /* 1118 * Place the array that protects pp->p_selock in the kmem64 wad. 1119 */ 1120 pse_shift = size_pse_array(npages, max_ncpus); 1121 PRM_DEBUG(pse_shift); 1122 pse_table_size = 1 << pse_shift; 1123 PRM_DEBUG(pse_table_size); 1124 psetable_sz = roundup( 1125 pse_table_size * sizeof (pad_mutex_t), ecache_alignsize); 1126 PRM_DEBUG(psetable_sz); 1127 1128 /* 1129 * Now allocate the whole wad 1130 */ 1131 kmem64_sz = pp_sz + kpm_pp_sz + hmehash_sz + pagehash_sz + 1132 pagelist_sz + tt_sz + psetable_sz; 1133 kmem64_sz = roundup(kmem64_sz, PAGESIZE); 1134 kmem64_base = (caddr_t)syslimit; 1135 kmem64_end = kmem64_base + kmem64_sz; 1136 alloc_kmem64(kmem64_base, kmem64_end); 1137 if (kmem64_aligned_end > (hole_start ? hole_start : kpm_vbase)) 1138 cmn_err(CE_PANIC, "not enough kmem64 space"); 1139 PRM_DEBUG(kmem64_base); 1140 PRM_DEBUG(kmem64_end); 1141 PRM_DEBUG(kmem64_aligned_end); 1142 1143 /* 1144 * ... and divy it up 1145 */ 1146 alloc_base = kmem64_base; 1147 1148 if (kpm_smallpages == 0) { 1149 npages -= kmem64_sz / (PAGESIZE + sizeof (struct page)); 1150 } else { 1151 npages -= kmem64_sz / (PAGESIZE + sizeof (struct page) + 1152 sizeof (kpm_spage_t)); 1153 } 1154 1155 pp_base = (page_t *)alloc_base; 1156 pp_sz = npages * sizeof (struct page); 1157 alloc_base += pp_sz; 1158 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, ecache_alignsize); 1159 PRM_DEBUG(pp_base); 1160 PRM_DEBUG(npages); 1161 1162 if (kpm_enable) { 1163 kpm_pp_base = alloc_base; 1164 if (kpm_smallpages == 0) { 1165 /* kpm_npages based on physinstalled, don't reset */ 1166 kpm_pp_sz = kpm_npages * sizeof (kpm_page_t); 1167 } else { 1168 kpm_npages = ptokpmpr(npages); 1169 kpm_pp_sz = kpm_npages * sizeof (kpm_spage_t); 1170 } 1171 alloc_base += kpm_pp_sz; 1172 alloc_base = 1173 (caddr_t)roundup((uintptr_t)alloc_base, ecache_alignsize); 1174 PRM_DEBUG(kpm_pp_base); 1175 } 1176 1177 alloc_base = alloc_hmehash(alloc_base); 1178 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, ecache_alignsize); 1179 PRM_DEBUG(alloc_base); 1180 1181 page_hash = (page_t **)alloc_base; 1182 alloc_base += pagehash_sz; 1183 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, ecache_alignsize); 1184 PRM_DEBUG(page_hash); 1185 1186 alloc_base = alloc_page_freelists(alloc_base); 1187 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, ecache_alignsize); 1188 PRM_DEBUG(alloc_base); 1189 1190 #ifdef TRAPTRACE 1191 ttrace_buf = alloc_base; 1192 alloc_base += tt_sz; 1193 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, ecache_alignsize); 1194 PRM_DEBUG(alloc_base); 1195 #endif /* TRAPTRACE */ 1196 1197 pse_mutex = (pad_mutex_t *)alloc_base; 1198 alloc_base += psetable_sz; 1199 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, ecache_alignsize); 1200 PRM_DEBUG(alloc_base); 1201 1202 /* adjust kmem64_end to what we really allocated */ 1203 kmem64_end = (caddr_t)roundup((uintptr_t)alloc_base, PAGESIZE); 1204 kmem64_sz = kmem64_end - kmem64_base; 1205 1206 if (&ecache_init_scrub_flush_area) { 1207 alloc_base = ecache_init_scrub_flush_area(kmem64_aligned_end); 1208 ASSERT(alloc_base <= (hole_start ? hole_start : kpm_vbase)); 1209 } 1210 1211 /* 1212 * If physmem is patched to be non-zero, use it instead of 1213 * the monitor value unless physmem is larger than the total 1214 * amount of memory on hand. 1215 */ 1216 if (physmem == 0 || physmem > npages) 1217 physmem = npages; 1218 1219 /* 1220 * root_is_ramdisk is set via /etc/system when the ramdisk miniroot 1221 * is mounted as root. This memory is held down by OBP and unlike 1222 * the stub boot_archive is never released. 1223 * 1224 * In order to get things sized correctly on lower memory 1225 * machines (where the memory used by the ramdisk represents 1226 * a significant portion of memory), physmem is adjusted. 1227 * 1228 * This is done by subtracting the ramdisk_size which is set 1229 * to the size of the ramdisk (in Kb) in /etc/system at the 1230 * time the miniroot archive is constructed. 1231 */ 1232 if (root_is_ramdisk == B_TRUE) { 1233 ramdisk_npages = (ramdisk_size * 1024) / PAGESIZE; 1234 physmem -= ramdisk_npages; 1235 } 1236 1237 if (kpm_enable && (ndata_alloc_kpm(&ndata, kpm_npages) != 0)) 1238 cmn_err(CE_PANIC, "no more nucleus memory after kpm alloc"); 1239 1240 /* 1241 * Allocate space for the interrupt vector table. 1242 */ 1243 memspace = prom_alloc((caddr_t)intr_vec_table, IVSIZE, MMU_PAGESIZE); 1244 if (memspace != (caddr_t)intr_vec_table) 1245 prom_panic("interrupt vector table allocation failure"); 1246 1247 /* 1248 * Between now and when we finish copying in the memory lists, 1249 * allocations happen so the space gets fragmented and the 1250 * lists longer. Leave enough space for lists twice as 1251 * long as we have now; then roundup to a pagesize. 1252 */ 1253 memlist_sz = sizeof (struct memlist) * (prom_phys_installed_len() + 1254 prom_phys_avail_len() + prom_virt_avail_len()); 1255 memlist_sz *= 2; 1256 memlist_sz = roundup(memlist_sz, PAGESIZE); 1257 memspace = ndata_alloc(&ndata, memlist_sz, ecache_alignsize); 1258 if (memspace == NULL) 1259 cmn_err(CE_PANIC, "no more nucleus memory after memlist alloc"); 1260 1261 memlist = (struct memlist *)memspace; 1262 memlist_end = (char *)memspace + memlist_sz; 1263 PRM_DEBUG(memlist); 1264 PRM_DEBUG(memlist_end); 1265 1266 PRM_DEBUG(sysbase); 1267 PRM_DEBUG(syslimit); 1268 kernelheap_init((void *)sysbase, (void *)syslimit, 1269 (caddr_t)sysbase + PAGESIZE, NULL, NULL); 1270 1271 /* 1272 * Take the most current snapshot we can by calling mem-update. 1273 */ 1274 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1275 &boot_physavail, &boot_physavail_len, 1276 &boot_virtavail, &boot_virtavail_len); 1277 1278 /* 1279 * Remove the space used by prom_alloc from the kernel heap 1280 * plus the area actually used by the OBP (if any) 1281 * ignoring virtual addresses in virt_avail, above syslimit. 1282 */ 1283 virt_avail = memlist; 1284 copy_memlist(boot_virtavail, boot_virtavail_len, &memlist); 1285 1286 for (cur = virt_avail; cur->next; cur = cur->next) { 1287 uint64_t range_base, range_size; 1288 1289 if ((range_base = cur->address + cur->size) < (uint64_t)sysbase) 1290 continue; 1291 if (range_base >= (uint64_t)syslimit) 1292 break; 1293 /* 1294 * Limit the range to end at syslimit. 1295 */ 1296 range_size = MIN(cur->next->address, 1297 (uint64_t)syslimit) - range_base; 1298 (void) vmem_xalloc(heap_arena, (size_t)range_size, PAGESIZE, 1299 0, 0, (void *)range_base, (void *)(range_base + range_size), 1300 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 1301 } 1302 1303 phys_avail = memlist; 1304 copy_memlist(boot_physavail, boot_physavail_len, &memlist); 1305 1306 /* 1307 * Add any extra memory at the end of the ndata region if there's at 1308 * least a page to add. There might be a few more pages available in 1309 * the middle of the ndata region, but for now they are ignored. 1310 */ 1311 nalloc_base = ndata_extra_base(&ndata, MMU_PAGESIZE, nalloc_end); 1312 if (nalloc_base == NULL) 1313 nalloc_base = nalloc_end; 1314 ndata_remain_sz = nalloc_end - nalloc_base; 1315 1316 /* 1317 * Copy physinstalled list into kernel space. 1318 */ 1319 phys_install = memlist; 1320 copy_memlist(boot_physinstalled, boot_physinstalled_len, &memlist); 1321 1322 /* 1323 * Create list of physical addrs we don't need pp's for: 1324 * kernel text 4M page 1325 * kernel data 4M page - ndata_remain_sz 1326 * kmem64 pages 1327 * 1328 * NB if adding any pages here, make sure no kpm page 1329 * overlaps can occur (see ASSERTs in kphysm_memsegs) 1330 */ 1331 nopp_list = memlist; 1332 memlist_new(va_to_pa(s_text), MMU_PAGESIZE4M, &memlist); 1333 memlist_add(va_to_pa(s_data), MMU_PAGESIZE4M - ndata_remain_sz, 1334 &memlist, &nopp_list); 1335 memlist_add(kmem64_pabase, kmem64_sz, &memlist, &nopp_list); 1336 1337 if ((caddr_t)memlist > (memspace + memlist_sz)) 1338 prom_panic("memlist overflow"); 1339 1340 /* 1341 * Size the pcf array based on the number of cpus in the box at 1342 * boot time. 1343 */ 1344 pcf_init(); 1345 1346 /* 1347 * Initialize the page structures from the memory lists. 1348 */ 1349 kphysm_init(); 1350 1351 availrmem_initial = availrmem = freemem; 1352 PRM_DEBUG(availrmem); 1353 1354 /* 1355 * Some of the locks depend on page_hashsz being set! 1356 * kmem_init() depends on this; so, keep it here. 1357 */ 1358 page_lock_init(); 1359 1360 /* 1361 * Initialize kernel memory allocator. 1362 */ 1363 kmem_init(); 1364 1365 /* 1366 * Factor in colorequiv to check additional 'equivalent' bins 1367 */ 1368 if (&page_set_colorequiv_arr_cpu != NULL) 1369 page_set_colorequiv_arr_cpu(); 1370 else 1371 page_set_colorequiv_arr(); 1372 1373 /* 1374 * Initialize bp_mapin(). 1375 */ 1376 bp_init(shm_alignment, HAT_STRICTORDER); 1377 1378 /* 1379 * Reserve space for panicbuf, intr_vec_table, reserved interrupt 1380 * vector data structures and MPO mblock structs from the 32-bit heap. 1381 */ 1382 (void) vmem_xalloc(heap32_arena, PANICBUFSIZE, PAGESIZE, 0, 0, 1383 panicbuf, panicbuf + PANICBUFSIZE, 1384 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 1385 1386 (void) vmem_xalloc(heap32_arena, IVSIZE, PAGESIZE, 0, 0, 1387 intr_vec_table, (caddr_t)intr_vec_table + IVSIZE, 1388 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 1389 1390 if (mpo_heap32_bufsz > (size_t)0) { 1391 (void) vmem_xalloc(heap32_arena, mpo_heap32_bufsz, 1392 PAGESIZE, 0, 0, mpo_heap32_buf, 1393 mpo_heap32_buf + mpo_heap32_bufsz, 1394 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 1395 } 1396 mem_config_init(); 1397 } 1398 1399 static void 1400 startup_modules(void) 1401 { 1402 int nhblk1, nhblk8; 1403 size_t nhblksz; 1404 pgcnt_t pages_per_hblk; 1405 size_t hme8blk_sz, hme1blk_sz; 1406 1407 /* 1408 * Let the platforms have a chance to change default 1409 * values before reading system file. 1410 */ 1411 if (&set_platform_defaults) 1412 set_platform_defaults(); 1413 1414 /* 1415 * Calculate default settings of system parameters based upon 1416 * maxusers, yet allow to be overridden via the /etc/system file. 1417 */ 1418 param_calc(0); 1419 1420 mod_setup(); 1421 1422 /* 1423 * If this is a positron, complain and halt. 1424 */ 1425 if (&iam_positron && iam_positron()) { 1426 cmn_err(CE_WARN, "This hardware platform is not supported" 1427 " by this release of Solaris.\n"); 1428 #ifdef DEBUG 1429 prom_enter_mon(); /* Type 'go' to resume */ 1430 cmn_err(CE_WARN, "Booting an unsupported platform.\n"); 1431 cmn_err(CE_WARN, "Booting with down-rev firmware.\n"); 1432 1433 #else /* DEBUG */ 1434 halt(0); 1435 #endif /* DEBUG */ 1436 } 1437 1438 /* 1439 * If we are running firmware that isn't 64-bit ready 1440 * then complain and halt. 1441 */ 1442 do_prom_version_check(); 1443 1444 /* 1445 * Initialize system parameters 1446 */ 1447 param_init(); 1448 1449 /* 1450 * maxmem is the amount of physical memory we're playing with. 1451 */ 1452 maxmem = physmem; 1453 1454 /* Set segkp limits. */ 1455 ncbase = kdi_segdebugbase; 1456 ncend = kdi_segdebugbase; 1457 1458 /* 1459 * Initialize the hat layer. 1460 */ 1461 hat_init(); 1462 1463 /* 1464 * Initialize segment management stuff. 1465 */ 1466 seg_init(); 1467 1468 /* 1469 * Create the va>tte handler, so the prom can understand 1470 * kernel translations. The handler is installed later, just 1471 * as we are about to take over the trap table from the prom. 1472 */ 1473 create_va_to_tte(); 1474 1475 /* 1476 * Load the forthdebugger (optional) 1477 */ 1478 forthdebug_init(); 1479 1480 /* 1481 * Create OBP node for console input callbacks 1482 * if it is needed. 1483 */ 1484 startup_create_io_node(); 1485 1486 if (modloadonly("fs", "specfs") == -1) 1487 halt("Can't load specfs"); 1488 1489 if (modloadonly("fs", "devfs") == -1) 1490 halt("Can't load devfs"); 1491 1492 if (modloadonly("misc", "swapgeneric") == -1) 1493 halt("Can't load swapgeneric"); 1494 1495 (void) modloadonly("sys", "lbl_edition"); 1496 1497 dispinit(); 1498 1499 /* 1500 * Infer meanings to the members of the idprom buffer. 1501 */ 1502 parse_idprom(); 1503 1504 /* Read cluster configuration data. */ 1505 clconf_init(); 1506 1507 setup_ddi(); 1508 1509 /* 1510 * Lets take this opportunity to load the root device. 1511 */ 1512 if (loadrootmodules() != 0) 1513 debug_enter("Can't load the root filesystem"); 1514 1515 /* 1516 * Load tod driver module for the tod part found on this system. 1517 * Recompute the cpu frequency/delays based on tod as tod part 1518 * tends to keep time more accurately. 1519 */ 1520 if (&load_tod_module) 1521 load_tod_module(); 1522 1523 /* 1524 * Allow platforms to load modules which might 1525 * be needed after bootops are gone. 1526 */ 1527 if (&load_platform_modules) 1528 load_platform_modules(); 1529 1530 setcpudelay(); 1531 1532 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1533 &boot_physavail, &boot_physavail_len, 1534 &boot_virtavail, &boot_virtavail_len); 1535 1536 /* 1537 * Calculation and allocation of hmeblks needed to remap 1538 * the memory allocated by PROM till now. 1539 * Overestimate the number of hblk1 elements by assuming 1540 * worst case of TTE64K mappings. 1541 * sfmmu_hblk_alloc will panic if this calculation is wrong. 1542 */ 1543 bop_alloc_pages = btopr(kmem64_end - kmem64_base); 1544 pages_per_hblk = btop(HMEBLK_SPAN(TTE64K)); 1545 bop_alloc_pages = roundup(bop_alloc_pages, pages_per_hblk); 1546 nhblk1 = bop_alloc_pages / pages_per_hblk + hblk1_min; 1547 1548 bop_alloc_pages = size_virtalloc(boot_virtavail, boot_virtavail_len); 1549 1550 /* sfmmu_init_nucleus_hblks expects properly aligned data structures */ 1551 hme8blk_sz = roundup(HME8BLK_SZ, sizeof (int64_t)); 1552 hme1blk_sz = roundup(HME1BLK_SZ, sizeof (int64_t)); 1553 1554 bop_alloc_pages += btopr(nhblk1 * hme1blk_sz); 1555 1556 pages_per_hblk = btop(HMEBLK_SPAN(TTE8K)); 1557 nhblk8 = 0; 1558 while (bop_alloc_pages > 1) { 1559 bop_alloc_pages = roundup(bop_alloc_pages, pages_per_hblk); 1560 nhblk8 += bop_alloc_pages /= pages_per_hblk; 1561 bop_alloc_pages *= hme8blk_sz; 1562 bop_alloc_pages = btopr(bop_alloc_pages); 1563 } 1564 nhblk8 += 2; 1565 1566 /* 1567 * Since hblk8's can hold up to 64k of mappings aligned on a 64k 1568 * boundary, the number of hblk8's needed to map the entries in the 1569 * boot_virtavail list needs to be adjusted to take this into 1570 * consideration. Thus, we need to add additional hblk8's since it 1571 * is possible that an hblk8 will not have all 8 slots used due to 1572 * alignment constraints. Since there were boot_virtavail_len entries 1573 * in that list, we need to add that many hblk8's to the number 1574 * already calculated to make sure we don't underestimate. 1575 */ 1576 nhblk8 += boot_virtavail_len; 1577 nhblksz = nhblk8 * hme8blk_sz + nhblk1 * hme1blk_sz; 1578 1579 /* Allocate in pagesize chunks */ 1580 nhblksz = roundup(nhblksz, MMU_PAGESIZE); 1581 hblk_base = kmem_zalloc(nhblksz, KM_SLEEP); 1582 sfmmu_init_nucleus_hblks(hblk_base, nhblksz, nhblk8, nhblk1); 1583 } 1584 1585 static void 1586 startup_bop_gone(void) 1587 { 1588 1589 /* 1590 * Destroy the MD initialized at startup 1591 * The startup initializes the MD framework 1592 * using prom and BOP alloc free it now. 1593 */ 1594 mach_descrip_startup_fini(); 1595 1596 /* 1597 * We're done with prom allocations. 1598 */ 1599 bop_fini(); 1600 1601 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1602 &boot_physavail, &boot_physavail_len, 1603 &boot_virtavail, &boot_virtavail_len); 1604 1605 /* 1606 * setup physically contiguous area twice as large as the ecache. 1607 * this is used while doing displacement flush of ecaches 1608 */ 1609 if (&ecache_flush_address) { 1610 ecache_flushaddr = ecache_flush_address(); 1611 if (ecache_flushaddr == (uint64_t)-1) { 1612 cmn_err(CE_PANIC, 1613 "startup: no memory to set ecache_flushaddr"); 1614 } 1615 } 1616 1617 /* 1618 * Virtual available next. 1619 */ 1620 ASSERT(virt_avail != NULL); 1621 memlist_free_list(virt_avail); 1622 virt_avail = memlist; 1623 copy_memlist(boot_virtavail, boot_virtavail_len, &memlist); 1624 1625 } 1626 1627 1628 /* 1629 * startup_fixup_physavail - called from mach_sfmmu.c after the final 1630 * allocations have been performed. We can't call it in startup_bop_gone 1631 * since later operations can cause obp to allocate more memory. 1632 */ 1633 void 1634 startup_fixup_physavail(void) 1635 { 1636 struct memlist *cur; 1637 size_t kmem64_overmap_size = kmem64_aligned_end - kmem64_end; 1638 1639 PRM_DEBUG(kmem64_overmap_size); 1640 1641 /* 1642 * take the most current snapshot we can by calling mem-update 1643 */ 1644 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1645 &boot_physavail, &boot_physavail_len, 1646 &boot_virtavail, &boot_virtavail_len); 1647 1648 /* 1649 * Copy phys_avail list, again. 1650 * Both the kernel/boot and the prom have been allocating 1651 * from the original list we copied earlier. 1652 */ 1653 cur = memlist; 1654 copy_memlist(boot_physavail, boot_physavail_len, &memlist); 1655 1656 /* 1657 * Add any unused kmem64 memory from overmapped page 1658 * (Note: va_to_pa does not work for kmem64_end) 1659 */ 1660 if (kmem64_overmap_size) { 1661 memlist_add(kmem64_pabase + (kmem64_end - kmem64_base), 1662 kmem64_overmap_size, &memlist, &cur); 1663 } 1664 1665 /* 1666 * Add any extra memory after e_data we added to the phys_avail list 1667 * back to the old list. 1668 */ 1669 if (ndata_remain_sz >= MMU_PAGESIZE) 1670 memlist_add(va_to_pa(nalloc_base), 1671 (uint64_t)ndata_remain_sz, &memlist, &cur); 1672 1673 /* 1674 * There isn't any bounds checking on the memlist area 1675 * so ensure it hasn't overgrown. 1676 */ 1677 if ((caddr_t)memlist > (caddr_t)memlist_end) 1678 cmn_err(CE_PANIC, "startup: memlist size exceeded"); 1679 1680 /* 1681 * The kernel removes the pages that were allocated for it from 1682 * the freelist, but we now have to find any -extra- pages that 1683 * the prom has allocated for it's own book-keeping, and remove 1684 * them from the freelist too. sigh. 1685 */ 1686 sync_memlists(phys_avail, cur); 1687 1688 ASSERT(phys_avail != NULL); 1689 1690 old_phys_avail = phys_avail; 1691 phys_avail = cur; 1692 } 1693 1694 void 1695 update_kcage_ranges(uint64_t addr, uint64_t len) 1696 { 1697 pfn_t base = btop(addr); 1698 pgcnt_t num = btop(len); 1699 int rv; 1700 1701 rv = kcage_range_add(base, num, kcage_startup_dir); 1702 1703 if (rv == ENOMEM) { 1704 cmn_err(CE_WARN, "%ld megabytes not available to kernel cage", 1705 (len == 0 ? 0 : BYTES_TO_MB(len))); 1706 } else if (rv != 0) { 1707 /* catch this in debug kernels */ 1708 ASSERT(0); 1709 1710 cmn_err(CE_WARN, "unexpected kcage_range_add" 1711 " return value %d", rv); 1712 } 1713 } 1714 1715 static void 1716 startup_vm(void) 1717 { 1718 size_t i; 1719 struct segmap_crargs a; 1720 struct segkpm_crargs b; 1721 1722 uint64_t avmem; 1723 caddr_t va; 1724 pgcnt_t max_phys_segkp; 1725 int mnode; 1726 1727 extern int use_brk_lpg, use_stk_lpg; 1728 1729 /* 1730 * get prom's mappings, create hments for them and switch 1731 * to the kernel context. 1732 */ 1733 hat_kern_setup(); 1734 1735 /* 1736 * Take over trap table 1737 */ 1738 setup_trap_table(); 1739 1740 /* 1741 * Install the va>tte handler, so that the prom can handle 1742 * misses and understand the kernel table layout in case 1743 * we need call into the prom. 1744 */ 1745 install_va_to_tte(); 1746 1747 /* 1748 * Set a flag to indicate that the tba has been taken over. 1749 */ 1750 tba_taken_over = 1; 1751 1752 /* initialize MMU primary context register */ 1753 mmu_init_kcontext(); 1754 1755 /* 1756 * The boot cpu can now take interrupts, x-calls, x-traps 1757 */ 1758 CPUSET_ADD(cpu_ready_set, CPU->cpu_id); 1759 CPU->cpu_flags |= (CPU_READY | CPU_ENABLE | CPU_EXISTS); 1760 1761 /* 1762 * Set a flag to tell write_scb_int() that it can access V_TBR_WR_ADDR. 1763 */ 1764 tbr_wr_addr_inited = 1; 1765 1766 /* 1767 * Initialize VM system, and map kernel address space. 1768 */ 1769 kvm_init(); 1770 1771 ASSERT(old_phys_avail != NULL && phys_avail != NULL); 1772 if (kernel_cage_enable) { 1773 diff_memlists(phys_avail, old_phys_avail, update_kcage_ranges); 1774 } 1775 memlist_free_list(old_phys_avail); 1776 1777 /* 1778 * If the following is true, someone has patched 1779 * phsymem to be less than the number of pages that 1780 * the system actually has. Remove pages until system 1781 * memory is limited to the requested amount. Since we 1782 * have allocated page structures for all pages, we 1783 * correct the amount of memory we want to remove 1784 * by the size of the memory used to hold page structures 1785 * for the non-used pages. 1786 */ 1787 if (physmem + ramdisk_npages < npages) { 1788 pgcnt_t diff, off; 1789 struct page *pp; 1790 struct seg kseg; 1791 1792 cmn_err(CE_WARN, "limiting physmem to %ld pages", physmem); 1793 1794 off = 0; 1795 diff = npages - (physmem + ramdisk_npages); 1796 diff -= mmu_btopr(diff * sizeof (struct page)); 1797 kseg.s_as = &kas; 1798 while (diff--) { 1799 pp = page_create_va(&unused_pages_vp, (offset_t)off, 1800 MMU_PAGESIZE, PG_WAIT | PG_EXCL, 1801 &kseg, (caddr_t)off); 1802 if (pp == NULL) 1803 cmn_err(CE_PANIC, "limited physmem too much!"); 1804 page_io_unlock(pp); 1805 page_downgrade(pp); 1806 availrmem--; 1807 off += MMU_PAGESIZE; 1808 } 1809 } 1810 1811 /* 1812 * When printing memory, show the total as physmem less 1813 * that stolen by a debugger. 1814 */ 1815 cmn_err(CE_CONT, "?mem = %ldK (0x%lx000)\n", 1816 (ulong_t)(physinstalled) << (PAGESHIFT - 10), 1817 (ulong_t)(physinstalled) << (PAGESHIFT - 12)); 1818 1819 avmem = (uint64_t)freemem << PAGESHIFT; 1820 cmn_err(CE_CONT, "?avail mem = %lld\n", (unsigned long long)avmem); 1821 1822 /* 1823 * For small memory systems disable automatic large pages. 1824 */ 1825 if (physmem < privm_lpg_min_physmem) { 1826 use_brk_lpg = 0; 1827 use_stk_lpg = 0; 1828 } 1829 1830 /* 1831 * Perform platform specific freelist processing 1832 */ 1833 if (&plat_freelist_process) { 1834 for (mnode = 0; mnode < max_mem_nodes; mnode++) 1835 if (mem_node_config[mnode].exists) 1836 plat_freelist_process(mnode); 1837 } 1838 1839 /* 1840 * Initialize the segkp segment type. We position it 1841 * after the configured tables and buffers (whose end 1842 * is given by econtig) and before V_WKBASE_ADDR. 1843 * Also in this area is segkmap (size SEGMAPSIZE). 1844 */ 1845 1846 /* XXX - cache alignment? */ 1847 va = (caddr_t)SEGKPBASE; 1848 ASSERT(((uintptr_t)va & PAGEOFFSET) == 0); 1849 1850 max_phys_segkp = (physmem * 2); 1851 1852 if (segkpsize < btop(SEGKPMINSIZE) || segkpsize > btop(SEGKPMAXSIZE)) { 1853 segkpsize = btop(SEGKPDEFSIZE); 1854 cmn_err(CE_WARN, "Illegal value for segkpsize. " 1855 "segkpsize has been reset to %ld pages", segkpsize); 1856 } 1857 1858 i = ptob(MIN(segkpsize, max_phys_segkp)); 1859 1860 rw_enter(&kas.a_lock, RW_WRITER); 1861 if (seg_attach(&kas, va, i, segkp) < 0) 1862 cmn_err(CE_PANIC, "startup: cannot attach segkp"); 1863 if (segkp_create(segkp) != 0) 1864 cmn_err(CE_PANIC, "startup: segkp_create failed"); 1865 rw_exit(&kas.a_lock); 1866 1867 /* 1868 * kpm segment 1869 */ 1870 segmap_kpm = kpm_enable && 1871 segmap_kpm && PAGESIZE == MAXBSIZE; 1872 1873 if (kpm_enable) { 1874 rw_enter(&kas.a_lock, RW_WRITER); 1875 1876 /* 1877 * The segkpm virtual range range is larger than the 1878 * actual physical memory size and also covers gaps in 1879 * the physical address range for the following reasons: 1880 * . keep conversion between segkpm and physical addresses 1881 * simple, cheap and unambiguous. 1882 * . avoid extension/shrink of the the segkpm in case of DR. 1883 * . avoid complexity for handling of virtual addressed 1884 * caches, segkpm and the regular mapping scheme must be 1885 * kept in sync wrt. the virtual color of mapped pages. 1886 * Any accesses to virtual segkpm ranges not backed by 1887 * physical memory will fall through the memseg pfn hash 1888 * and will be handled in segkpm_fault. 1889 * Additional kpm_size spaces needed for vac alias prevention. 1890 */ 1891 if (seg_attach(&kas, kpm_vbase, kpm_size * vac_colors, 1892 segkpm) < 0) 1893 cmn_err(CE_PANIC, "cannot attach segkpm"); 1894 1895 b.prot = PROT_READ | PROT_WRITE; 1896 b.nvcolors = shm_alignment >> MMU_PAGESHIFT; 1897 1898 if (segkpm_create(segkpm, (caddr_t)&b) != 0) 1899 panic("segkpm_create segkpm"); 1900 1901 rw_exit(&kas.a_lock); 1902 1903 mach_kpm_init(); 1904 } 1905 1906 if (!segzio_fromheap) { 1907 size_t size; 1908 size_t physmem_b = mmu_ptob(physmem); 1909 1910 /* size is in bytes, segziosize is in pages */ 1911 if (segziosize == 0) { 1912 size = physmem_b; 1913 } else { 1914 size = mmu_ptob(segziosize); 1915 } 1916 1917 if (size < SEGZIOMINSIZE) { 1918 size = SEGZIOMINSIZE; 1919 } else if (size > SEGZIOMAXSIZE) { 1920 size = SEGZIOMAXSIZE; 1921 /* 1922 * On 64-bit x86, we only have 2TB of KVA. This exists 1923 * for parity with x86. 1924 * 1925 * SEGZIOMAXSIZE is capped at 512gb so that segzio 1926 * doesn't consume all of KVA. However, if we have a 1927 * system that has more thant 512gb of physical memory, 1928 * we can actually consume about half of the difference 1929 * between 512gb and the rest of the available physical 1930 * memory. 1931 */ 1932 if (physmem_b > SEGZIOMAXSIZE) { 1933 size += (physmem_b - SEGZIOMAXSIZE) / 2; 1934 } 1935 } 1936 segziosize = mmu_btop(roundup(size, MMU_PAGESIZE)); 1937 /* put the base of the ZIO segment after the kpm segment */ 1938 segzio_base = kpm_vbase + (kpm_size * vac_colors); 1939 PRM_DEBUG(segziosize); 1940 PRM_DEBUG(segzio_base); 1941 1942 /* 1943 * On some platforms, kvm_init is called after the kpm 1944 * sizes have been determined. On SPARC, kvm_init is called 1945 * before, so we have to attach the kzioseg after kvm is 1946 * initialized, otherwise we'll try to allocate from the boot 1947 * area since the kernel heap hasn't yet been configured. 1948 */ 1949 rw_enter(&kas.a_lock, RW_WRITER); 1950 1951 (void) seg_attach(&kas, segzio_base, mmu_ptob(segziosize), 1952 &kzioseg); 1953 (void) segkmem_zio_create(&kzioseg); 1954 1955 /* create zio area covering new segment */ 1956 segkmem_zio_init(segzio_base, mmu_ptob(segziosize)); 1957 1958 rw_exit(&kas.a_lock); 1959 } 1960 1961 1962 /* 1963 * Now create generic mapping segment. This mapping 1964 * goes SEGMAPSIZE beyond SEGMAPBASE. But if the total 1965 * virtual address is greater than the amount of free 1966 * memory that is available, then we trim back the 1967 * segment size to that amount 1968 */ 1969 va = (caddr_t)SEGMAPBASE; 1970 1971 /* 1972 * 1201049: segkmap base address must be MAXBSIZE aligned 1973 */ 1974 ASSERT(((uintptr_t)va & MAXBOFFSET) == 0); 1975 1976 /* 1977 * Set size of segmap to percentage of freemem at boot, 1978 * but stay within the allowable range 1979 * Note we take percentage before converting from pages 1980 * to bytes to avoid an overflow on 32-bit kernels. 1981 */ 1982 i = mmu_ptob((freemem * segmap_percent) / 100); 1983 1984 if (i < MINMAPSIZE) 1985 i = MINMAPSIZE; 1986 1987 if (i > MIN(SEGMAPSIZE, mmu_ptob(freemem))) 1988 i = MIN(SEGMAPSIZE, mmu_ptob(freemem)); 1989 1990 i &= MAXBMASK; /* 1201049: segkmap size must be MAXBSIZE aligned */ 1991 1992 rw_enter(&kas.a_lock, RW_WRITER); 1993 if (seg_attach(&kas, va, i, segkmap) < 0) 1994 cmn_err(CE_PANIC, "cannot attach segkmap"); 1995 1996 a.prot = PROT_READ | PROT_WRITE; 1997 a.shmsize = shm_alignment; 1998 a.nfreelist = 0; /* use segmap driver defaults */ 1999 2000 if (segmap_create(segkmap, (caddr_t)&a) != 0) 2001 panic("segmap_create segkmap"); 2002 rw_exit(&kas.a_lock); 2003 2004 segdev_init(); 2005 } 2006 2007 static void 2008 startup_end(void) 2009 { 2010 if ((caddr_t)memlist > (caddr_t)memlist_end) 2011 panic("memlist overflow 2"); 2012 memlist_free_block((caddr_t)memlist, 2013 ((caddr_t)memlist_end - (caddr_t)memlist)); 2014 memlist = NULL; 2015 2016 /* enable page_relocation since OBP is now done */ 2017 page_relocate_ready = 1; 2018 2019 /* 2020 * Perform tasks that get done after most of the VM 2021 * initialization has been done but before the clock 2022 * and other devices get started. 2023 */ 2024 kern_setup1(); 2025 2026 /* 2027 * Intialize the VM arenas for allocating physically 2028 * contiguus memory chunk for interrupt queues snd 2029 * allocate/register boot cpu's queues, if any and 2030 * allocate dump buffer for sun4v systems to store 2031 * extra crash information during crash dump 2032 */ 2033 contig_mem_init(); 2034 mach_descrip_init(); 2035 2036 if (cpu_intrq_setup(CPU)) { 2037 cmn_err(CE_PANIC, "cpu%d: setup failed", CPU->cpu_id); 2038 } 2039 cpu_intrq_register(CPU); 2040 mach_htraptrace_setup(CPU->cpu_id); 2041 mach_htraptrace_configure(CPU->cpu_id); 2042 mach_dump_buffer_init(); 2043 2044 /* 2045 * Initialize interrupt related stuff 2046 */ 2047 cpu_intr_alloc(CPU, NINTR_THREADS); 2048 2049 (void) splzs(); /* allow hi clock ints but not zs */ 2050 2051 /* 2052 * Initialize errors. 2053 */ 2054 error_init(); 2055 2056 /* 2057 * Note that we may have already used kernel bcopy before this 2058 * point - but if you really care about this, adb the use_hw_* 2059 * variables to 0 before rebooting. 2060 */ 2061 mach_hw_copy_limit(); 2062 2063 /* 2064 * Install the "real" preemption guards before DDI services 2065 * are available. 2066 */ 2067 (void) prom_set_preprom(kern_preprom); 2068 (void) prom_set_postprom(kern_postprom); 2069 CPU->cpu_m.mutex_ready = 1; 2070 2071 /* 2072 * Initialize segnf (kernel support for non-faulting loads). 2073 */ 2074 segnf_init(); 2075 2076 /* 2077 * Configure the root devinfo node. 2078 */ 2079 configure(); /* set up devices */ 2080 mach_cpu_halt_idle(); 2081 } 2082 2083 2084 void 2085 post_startup(void) 2086 { 2087 #ifdef PTL1_PANIC_DEBUG 2088 extern void init_ptl1_thread(void); 2089 #endif /* PTL1_PANIC_DEBUG */ 2090 extern void abort_sequence_init(void); 2091 2092 /* 2093 * Set the system wide, processor-specific flags to be passed 2094 * to userland via the aux vector for performance hints and 2095 * instruction set extensions. 2096 */ 2097 bind_hwcap(); 2098 2099 /* 2100 * Startup memory scrubber (if any) 2101 */ 2102 mach_memscrub(); 2103 2104 /* 2105 * Allocate soft interrupt to handle abort sequence. 2106 */ 2107 abort_sequence_init(); 2108 2109 /* 2110 * Configure the rest of the system. 2111 * Perform forceloading tasks for /etc/system. 2112 */ 2113 (void) mod_sysctl(SYS_FORCELOAD, NULL); 2114 /* 2115 * ON4.0: Force /proc module in until clock interrupt handle fixed 2116 * ON4.0: This must be fixed or restated in /etc/systems. 2117 */ 2118 (void) modload("fs", "procfs"); 2119 2120 /* load machine class specific drivers */ 2121 load_mach_drivers(); 2122 2123 /* load platform specific drivers */ 2124 if (&load_platform_drivers) 2125 load_platform_drivers(); 2126 2127 /* load vis simulation module, if we are running w/fpu off */ 2128 if (!fpu_exists) { 2129 if (modload("misc", "vis") == -1) 2130 halt("Can't load vis"); 2131 } 2132 2133 mach_fpras(); 2134 2135 maxmem = freemem; 2136 2137 #ifdef PTL1_PANIC_DEBUG 2138 init_ptl1_thread(); 2139 #endif /* PTL1_PANIC_DEBUG */ 2140 } 2141 2142 #ifdef PTL1_PANIC_DEBUG 2143 int ptl1_panic_test = 0; 2144 int ptl1_panic_xc_one_test = 0; 2145 int ptl1_panic_xc_all_test = 0; 2146 int ptl1_panic_xt_one_test = 0; 2147 int ptl1_panic_xt_all_test = 0; 2148 kthread_id_t ptl1_thread_p = NULL; 2149 kcondvar_t ptl1_cv; 2150 kmutex_t ptl1_mutex; 2151 int ptl1_recurse_count_threshold = 0x40; 2152 int ptl1_recurse_trap_threshold = 0x3d; 2153 extern void ptl1_recurse(int, int); 2154 extern void ptl1_panic_xt(int, int); 2155 2156 /* 2157 * Called once per second by timeout() to wake up 2158 * the ptl1_panic thread to see if it should cause 2159 * a trap to the ptl1_panic() code. 2160 */ 2161 /* ARGSUSED */ 2162 static void 2163 ptl1_wakeup(void *arg) 2164 { 2165 mutex_enter(&ptl1_mutex); 2166 cv_signal(&ptl1_cv); 2167 mutex_exit(&ptl1_mutex); 2168 } 2169 2170 /* 2171 * ptl1_panic cross call function: 2172 * Needed because xc_one() and xc_some() can pass 2173 * 64 bit args but ptl1_recurse() expects ints. 2174 */ 2175 static void 2176 ptl1_panic_xc(void) 2177 { 2178 ptl1_recurse(ptl1_recurse_count_threshold, 2179 ptl1_recurse_trap_threshold); 2180 } 2181 2182 /* 2183 * The ptl1 thread waits for a global flag to be set 2184 * and uses the recurse thresholds to set the stack depth 2185 * to cause a ptl1_panic() directly via a call to ptl1_recurse 2186 * or indirectly via the cross call and cross trap functions. 2187 * 2188 * This is useful testing stack overflows and normal 2189 * ptl1_panic() states with a know stack frame. 2190 * 2191 * ptl1_recurse() is an asm function in ptl1_panic.s that 2192 * sets the {In, Local, Out, and Global} registers to a 2193 * know state on the stack and just prior to causing a 2194 * test ptl1_panic trap. 2195 */ 2196 static void 2197 ptl1_thread(void) 2198 { 2199 mutex_enter(&ptl1_mutex); 2200 while (ptl1_thread_p) { 2201 cpuset_t other_cpus; 2202 int cpu_id; 2203 int my_cpu_id; 2204 int target_cpu_id; 2205 int target_found; 2206 2207 if (ptl1_panic_test) { 2208 ptl1_recurse(ptl1_recurse_count_threshold, 2209 ptl1_recurse_trap_threshold); 2210 } 2211 2212 /* 2213 * Find potential targets for x-call and x-trap, 2214 * if any exist while preempt is disabled we 2215 * start a ptl1_panic if requested via a 2216 * globals. 2217 */ 2218 kpreempt_disable(); 2219 my_cpu_id = CPU->cpu_id; 2220 other_cpus = cpu_ready_set; 2221 CPUSET_DEL(other_cpus, CPU->cpu_id); 2222 target_found = 0; 2223 if (!CPUSET_ISNULL(other_cpus)) { 2224 /* 2225 * Pick the first one 2226 */ 2227 for (cpu_id = 0; cpu_id < NCPU; cpu_id++) { 2228 if (cpu_id == my_cpu_id) 2229 continue; 2230 2231 if (CPU_XCALL_READY(cpu_id)) { 2232 target_cpu_id = cpu_id; 2233 target_found = 1; 2234 break; 2235 } 2236 } 2237 ASSERT(target_found); 2238 2239 if (ptl1_panic_xc_one_test) { 2240 xc_one(target_cpu_id, 2241 (xcfunc_t *)ptl1_panic_xc, 0, 0); 2242 } 2243 if (ptl1_panic_xc_all_test) { 2244 xc_some(other_cpus, 2245 (xcfunc_t *)ptl1_panic_xc, 0, 0); 2246 } 2247 if (ptl1_panic_xt_one_test) { 2248 xt_one(target_cpu_id, 2249 (xcfunc_t *)ptl1_panic_xt, 0, 0); 2250 } 2251 if (ptl1_panic_xt_all_test) { 2252 xt_some(other_cpus, 2253 (xcfunc_t *)ptl1_panic_xt, 0, 0); 2254 } 2255 } 2256 kpreempt_enable(); 2257 (void) timeout(ptl1_wakeup, NULL, hz); 2258 (void) cv_wait(&ptl1_cv, &ptl1_mutex); 2259 } 2260 mutex_exit(&ptl1_mutex); 2261 } 2262 2263 /* 2264 * Called during early startup to create the ptl1_thread 2265 */ 2266 void 2267 init_ptl1_thread(void) 2268 { 2269 ptl1_thread_p = thread_create(NULL, 0, ptl1_thread, NULL, 0, 2270 &p0, TS_RUN, 0); 2271 } 2272 #endif /* PTL1_PANIC_DEBUG */ 2273 2274 2275 static void 2276 memlist_new(uint64_t start, uint64_t len, struct memlist **memlistp) 2277 { 2278 struct memlist *new; 2279 2280 new = *memlistp; 2281 new->address = start; 2282 new->size = len; 2283 *memlistp = new + 1; 2284 } 2285 2286 /* 2287 * Add to a memory list. 2288 * start = start of new memory segment 2289 * len = length of new memory segment in bytes 2290 * memlistp = pointer to array of available memory segment structures 2291 * curmemlistp = memory list to which to add segment. 2292 */ 2293 static void 2294 memlist_add(uint64_t start, uint64_t len, struct memlist **memlistp, 2295 struct memlist **curmemlistp) 2296 { 2297 struct memlist *new = *memlistp; 2298 2299 memlist_new(start, len, memlistp); 2300 memlist_insert(new, curmemlistp); 2301 } 2302 2303 static int 2304 ndata_alloc_memseg(struct memlist *ndata, size_t avail) 2305 { 2306 int nseg; 2307 size_t memseg_sz; 2308 struct memseg *msp; 2309 2310 /* 2311 * The memseg list is for the chunks of physical memory that 2312 * will be managed by the vm system. The number calculated is 2313 * a guess as boot may fragment it more when memory allocations 2314 * are made before kphysm_init(). 2315 */ 2316 memseg_sz = (avail + 10) * sizeof (struct memseg); 2317 memseg_sz = roundup(memseg_sz, PAGESIZE); 2318 nseg = memseg_sz / sizeof (struct memseg); 2319 msp = ndata_alloc(ndata, memseg_sz, ecache_alignsize); 2320 if (msp == NULL) 2321 return (1); 2322 PRM_DEBUG(memseg_free); 2323 2324 while (nseg--) { 2325 msp->next = memseg_free; 2326 memseg_free = msp; 2327 msp++; 2328 } 2329 return (0); 2330 } 2331 2332 /* 2333 * In the case of architectures that support dynamic addition of 2334 * memory at run-time there are two cases where memsegs need to 2335 * be initialized and added to the memseg list. 2336 * 1) memsegs that are constructed at startup. 2337 * 2) memsegs that are constructed at run-time on 2338 * hot-plug capable architectures. 2339 * This code was originally part of the function kphysm_init(). 2340 */ 2341 2342 static void 2343 memseg_list_add(struct memseg *memsegp) 2344 { 2345 struct memseg **prev_memsegp; 2346 pgcnt_t num; 2347 2348 /* insert in memseg list, decreasing number of pages order */ 2349 2350 num = MSEG_NPAGES(memsegp); 2351 2352 for (prev_memsegp = &memsegs; *prev_memsegp; 2353 prev_memsegp = &((*prev_memsegp)->next)) { 2354 if (num > MSEG_NPAGES(*prev_memsegp)) 2355 break; 2356 } 2357 2358 memsegp->next = *prev_memsegp; 2359 *prev_memsegp = memsegp; 2360 2361 if (kpm_enable) { 2362 memsegp->nextpa = (memsegp->next) ? 2363 va_to_pa(memsegp->next) : MSEG_NULLPTR_PA; 2364 2365 if (prev_memsegp != &memsegs) { 2366 struct memseg *msp; 2367 msp = (struct memseg *)((caddr_t)prev_memsegp - 2368 offsetof(struct memseg, next)); 2369 msp->nextpa = va_to_pa(memsegp); 2370 } else { 2371 memsegspa = va_to_pa(memsegs); 2372 } 2373 } 2374 } 2375 2376 /* 2377 * PSM add_physmem_cb(). US-II and newer processors have some 2378 * flavor of the prefetch capability implemented. We exploit 2379 * this capability for optimum performance. 2380 */ 2381 #define PREFETCH_BYTES 64 2382 2383 void 2384 add_physmem_cb(page_t *pp, pfn_t pnum) 2385 { 2386 extern void prefetch_page_w(void *); 2387 2388 pp->p_pagenum = pnum; 2389 2390 /* 2391 * Prefetch one more page_t into E$. To prevent future 2392 * mishaps with the sizeof(page_t) changing on us, we 2393 * catch this on debug kernels if we can't bring in the 2394 * entire hpage with 2 PREFETCH_BYTES reads. See 2395 * also, sun4u/cpu/cpu_module.c 2396 */ 2397 /*LINTED*/ 2398 ASSERT(sizeof (page_t) <= 2*PREFETCH_BYTES); 2399 prefetch_page_w((char *)pp); 2400 } 2401 2402 /* 2403 * Find memseg with given pfn 2404 */ 2405 static struct memseg * 2406 memseg_find(pfn_t base, pfn_t *next) 2407 { 2408 struct memseg *seg; 2409 2410 if (next != NULL) 2411 *next = LONG_MAX; 2412 for (seg = memsegs; seg != NULL; seg = seg->next) { 2413 if (base >= seg->pages_base && base < seg->pages_end) 2414 return (seg); 2415 if (next != NULL && seg->pages_base > base && 2416 seg->pages_base < *next) 2417 *next = seg->pages_base; 2418 } 2419 return (NULL); 2420 } 2421 2422 extern struct vnode prom_ppages; 2423 2424 /* 2425 * Put page allocated by OBP on prom_ppages 2426 */ 2427 static void 2428 kphysm_erase(uint64_t addr, uint64_t len) 2429 { 2430 struct page *pp; 2431 struct memseg *seg; 2432 pfn_t base = btop(addr), next; 2433 pgcnt_t num = btop(len); 2434 2435 while (num != 0) { 2436 pgcnt_t off, left; 2437 2438 seg = memseg_find(base, &next); 2439 if (seg == NULL) { 2440 if (next == LONG_MAX) 2441 break; 2442 left = MIN(next - base, num); 2443 base += left, num -= left; 2444 continue; 2445 } 2446 off = base - seg->pages_base; 2447 pp = seg->pages + off; 2448 left = num - MIN(num, (seg->pages_end - seg->pages_base) - off); 2449 while (num != left) { 2450 /* 2451 * init it, lock it, and hashin on prom_pages vp. 2452 * 2453 * XXX vnode offsets on the prom_ppages vnode 2454 * are page numbers (gack) for >32 bit 2455 * physical memory machines. 2456 */ 2457 add_physmem_cb(pp, base); 2458 if (page_trylock(pp, SE_EXCL) == 0) 2459 cmn_err(CE_PANIC, "prom page locked"); 2460 (void) page_hashin(pp, &prom_ppages, 2461 (offset_t)base, NULL); 2462 (void) page_pp_lock(pp, 0, 1); 2463 pp++, base++, num--; 2464 } 2465 } 2466 } 2467 2468 static page_t *ppnext; 2469 static pgcnt_t ppleft; 2470 2471 static void *kpm_ppnext; 2472 static pgcnt_t kpm_ppleft; 2473 2474 /* 2475 * Create a memseg 2476 */ 2477 static void 2478 kphysm_memseg(uint64_t addr, uint64_t len) 2479 { 2480 pfn_t base = btop(addr); 2481 pgcnt_t num = btop(len); 2482 struct memseg *seg; 2483 2484 seg = memseg_free; 2485 memseg_free = seg->next; 2486 ASSERT(seg != NULL); 2487 2488 seg->pages = ppnext; 2489 seg->epages = ppnext + num; 2490 seg->pages_base = base; 2491 seg->pages_end = base + num; 2492 ppnext += num; 2493 ppleft -= num; 2494 2495 if (kpm_enable) { 2496 pgcnt_t kpnum = ptokpmpr(num); 2497 2498 if (kpnum > kpm_ppleft) 2499 panic("kphysm_memseg: kpm_pp overflow"); 2500 seg->pagespa = va_to_pa(seg->pages); 2501 seg->epagespa = va_to_pa(seg->epages); 2502 seg->kpm_pbase = kpmptop(ptokpmp(base)); 2503 seg->kpm_nkpmpgs = kpnum; 2504 /* 2505 * In the kpm_smallpage case, the kpm array 2506 * is 1-1 wrt the page array 2507 */ 2508 if (kpm_smallpages) { 2509 kpm_spage_t *kpm_pp = kpm_ppnext; 2510 2511 kpm_ppnext = kpm_pp + kpnum; 2512 seg->kpm_spages = kpm_pp; 2513 seg->kpm_pagespa = va_to_pa(seg->kpm_spages); 2514 } else { 2515 kpm_page_t *kpm_pp = kpm_ppnext; 2516 2517 kpm_ppnext = kpm_pp + kpnum; 2518 seg->kpm_pages = kpm_pp; 2519 seg->kpm_pagespa = va_to_pa(seg->kpm_pages); 2520 /* ASSERT no kpm overlaps */ 2521 ASSERT( 2522 memseg_find(base - pmodkpmp(base), NULL) == NULL); 2523 ASSERT(memseg_find( 2524 roundup(base + num, kpmpnpgs) - 1, NULL) == NULL); 2525 } 2526 kpm_ppleft -= num; 2527 } 2528 2529 memseg_list_add(seg); 2530 } 2531 2532 /* 2533 * Add range to free list 2534 */ 2535 void 2536 kphysm_add(uint64_t addr, uint64_t len, int reclaim) 2537 { 2538 struct page *pp; 2539 struct memseg *seg; 2540 pfn_t base = btop(addr); 2541 pgcnt_t num = btop(len); 2542 2543 seg = memseg_find(base, NULL); 2544 ASSERT(seg != NULL); 2545 pp = seg->pages + (base - seg->pages_base); 2546 2547 if (reclaim) { 2548 struct page *rpp = pp; 2549 struct page *lpp = pp + num; 2550 2551 /* 2552 * page should be locked on prom_ppages 2553 * unhash and unlock it 2554 */ 2555 while (rpp < lpp) { 2556 ASSERT(PAGE_EXCL(rpp) && rpp->p_vnode == &prom_ppages); 2557 page_pp_unlock(rpp, 0, 1); 2558 page_hashout(rpp, NULL); 2559 page_unlock(rpp); 2560 rpp++; 2561 } 2562 } 2563 2564 /* 2565 * add_physmem() initializes the PSM part of the page 2566 * struct by calling the PSM back with add_physmem_cb(). 2567 * In addition it coalesces pages into larger pages as 2568 * it initializes them. 2569 */ 2570 add_physmem(pp, num, base); 2571 } 2572 2573 /* 2574 * kphysm_init() tackles the problem of initializing physical memory. 2575 */ 2576 static void 2577 kphysm_init(void) 2578 { 2579 struct memlist *pmem; 2580 2581 ASSERT(page_hash != NULL && page_hashsz != 0); 2582 2583 ppnext = pp_base; 2584 ppleft = npages; 2585 kpm_ppnext = kpm_pp_base; 2586 kpm_ppleft = kpm_npages; 2587 2588 /* 2589 * installed pages not on nopp_memlist go in memseg list 2590 */ 2591 diff_memlists(phys_install, nopp_list, kphysm_memseg); 2592 2593 /* 2594 * Free the avail list 2595 */ 2596 for (pmem = phys_avail; pmem != NULL; pmem = pmem->next) 2597 kphysm_add(pmem->address, pmem->size, 0); 2598 2599 /* 2600 * Erase pages that aren't available 2601 */ 2602 diff_memlists(phys_install, phys_avail, kphysm_erase); 2603 2604 build_pfn_hash(); 2605 } 2606 2607 /* 2608 * Kernel VM initialization. 2609 * Assumptions about kernel address space ordering: 2610 * (1) gap (user space) 2611 * (2) kernel text 2612 * (3) kernel data/bss 2613 * (4) gap 2614 * (5) kernel data structures 2615 * (6) gap 2616 * (7) debugger (optional) 2617 * (8) monitor 2618 * (9) gap (possibly null) 2619 * (10) dvma 2620 * (11) devices 2621 */ 2622 static void 2623 kvm_init(void) 2624 { 2625 /* 2626 * Put the kernel segments in kernel address space. 2627 */ 2628 rw_enter(&kas.a_lock, RW_WRITER); 2629 as_avlinit(&kas); 2630 2631 (void) seg_attach(&kas, (caddr_t)KERNELBASE, 2632 (size_t)(e_moddata - KERNELBASE), &ktextseg); 2633 (void) segkmem_create(&ktextseg); 2634 2635 (void) seg_attach(&kas, (caddr_t)(KERNELBASE + MMU_PAGESIZE4M), 2636 (size_t)(MMU_PAGESIZE4M), &ktexthole); 2637 (void) segkmem_create(&ktexthole); 2638 2639 (void) seg_attach(&kas, (caddr_t)valloc_base, 2640 (size_t)(econtig32 - valloc_base), &kvalloc); 2641 (void) segkmem_create(&kvalloc); 2642 2643 if (kmem64_base) { 2644 (void) seg_attach(&kas, (caddr_t)kmem64_base, 2645 (size_t)(kmem64_end - kmem64_base), &kmem64); 2646 (void) segkmem_create(&kmem64); 2647 } 2648 2649 /* 2650 * We're about to map out /boot. This is the beginning of the 2651 * system resource management transition. We can no longer 2652 * call into /boot for I/O or memory allocations. 2653 */ 2654 (void) seg_attach(&kas, kernelheap, ekernelheap - kernelheap, &kvseg); 2655 (void) segkmem_create(&kvseg); 2656 hblk_alloc_dynamic = 1; 2657 2658 /* 2659 * we need to preallocate pages for DR operations before enabling large 2660 * page kernel heap because of memseg_remap_init() hat_unload() hack. 2661 */ 2662 memseg_remap_init(); 2663 2664 /* at this point we are ready to use large page heap */ 2665 segkmem_heap_lp_init(); 2666 2667 (void) seg_attach(&kas, (caddr_t)SYSBASE32, SYSLIMIT32 - SYSBASE32, 2668 &kvseg32); 2669 (void) segkmem_create(&kvseg32); 2670 2671 /* 2672 * Create a segment for the debugger. 2673 */ 2674 (void) seg_attach(&kas, kdi_segdebugbase, kdi_segdebugsize, &kdebugseg); 2675 (void) segkmem_create(&kdebugseg); 2676 2677 rw_exit(&kas.a_lock); 2678 } 2679 2680 char obp_tte_str[] = 2681 "h# %x constant MMU_PAGESHIFT " 2682 "h# %x constant TTE8K " 2683 "h# %x constant SFHME_SIZE " 2684 "h# %x constant SFHME_TTE " 2685 "h# %x constant HMEBLK_TAG " 2686 "h# %x constant HMEBLK_NEXT " 2687 "h# %x constant HMEBLK_MISC " 2688 "h# %x constant HMEBLK_HME1 " 2689 "h# %x constant NHMENTS " 2690 "h# %x constant HBLK_SZMASK " 2691 "h# %x constant HBLK_RANGE_SHIFT " 2692 "h# %x constant HMEBP_HBLK " 2693 "h# %x constant HMEBUCKET_SIZE " 2694 "h# %x constant HTAG_SFMMUPSZ " 2695 "h# %x constant HTAG_BSPAGE_SHIFT " 2696 "h# %x constant HTAG_REHASH_SHIFT " 2697 "h# %x constant SFMMU_INVALID_SHMERID " 2698 "h# %x constant mmu_hashcnt " 2699 "h# %p constant uhme_hash " 2700 "h# %p constant khme_hash " 2701 "h# %x constant UHMEHASH_SZ " 2702 "h# %x constant KHMEHASH_SZ " 2703 "h# %p constant KCONTEXT " 2704 "h# %p constant KHATID " 2705 "h# %x constant ASI_MEM " 2706 2707 ": PHYS-X@ ( phys -- data ) " 2708 " ASI_MEM spacex@ " 2709 "; " 2710 2711 ": PHYS-W@ ( phys -- data ) " 2712 " ASI_MEM spacew@ " 2713 "; " 2714 2715 ": PHYS-L@ ( phys -- data ) " 2716 " ASI_MEM spaceL@ " 2717 "; " 2718 2719 ": TTE_PAGE_SHIFT ( ttesz -- hmeshift ) " 2720 " 3 * MMU_PAGESHIFT + " 2721 "; " 2722 2723 ": TTE_IS_VALID ( ttep -- flag ) " 2724 " PHYS-X@ 0< " 2725 "; " 2726 2727 ": HME_HASH_SHIFT ( ttesz -- hmeshift ) " 2728 " dup TTE8K = if " 2729 " drop HBLK_RANGE_SHIFT " 2730 " else " 2731 " TTE_PAGE_SHIFT " 2732 " then " 2733 "; " 2734 2735 ": HME_HASH_BSPAGE ( addr hmeshift -- bspage ) " 2736 " tuck >> swap MMU_PAGESHIFT - << " 2737 "; " 2738 2739 ": HME_HASH_FUNCTION ( sfmmup addr hmeshift -- hmebp ) " 2740 " >> over xor swap ( hash sfmmup ) " 2741 " KHATID <> if ( hash ) " 2742 " UHMEHASH_SZ and ( bucket ) " 2743 " HMEBUCKET_SIZE * uhme_hash + ( hmebp ) " 2744 " else ( hash ) " 2745 " KHMEHASH_SZ and ( bucket ) " 2746 " HMEBUCKET_SIZE * khme_hash + ( hmebp ) " 2747 " then ( hmebp ) " 2748 "; " 2749 2750 ": HME_HASH_TABLE_SEARCH " 2751 " ( sfmmup hmebp hblktag -- sfmmup null | sfmmup hmeblkp ) " 2752 " >r hmebp_hblk + phys-x@ begin ( sfmmup hmeblkp ) ( r: hblktag ) " 2753 " dup if ( sfmmup hmeblkp ) ( r: hblktag ) " 2754 " dup hmeblk_tag + phys-x@ r@ = if ( sfmmup hmeblkp ) " 2755 " dup hmeblk_tag + 8 + phys-x@ 2 pick = if " 2756 " true ( sfmmup hmeblkp true ) ( r: hblktag ) " 2757 " else " 2758 " hmeblk_next + phys-x@ false " 2759 " ( sfmmup hmeblkp false ) ( r: hblktag ) " 2760 " then " 2761 " else " 2762 " hmeblk_next + phys-x@ false " 2763 " ( sfmmup hmeblkp false ) ( r: hblktag ) " 2764 " then " 2765 " else " 2766 " true " 2767 " then " 2768 " until r> drop " 2769 "; " 2770 2771 ": HME_HASH_TAG ( sfmmup rehash addr -- hblktag ) " 2772 " over HME_HASH_SHIFT HME_HASH_BSPAGE ( sfmmup rehash bspage ) " 2773 " HTAG_BSPAGE_SHIFT << ( sfmmup rehash htag-bspage )" 2774 " swap HTAG_REHASH_SHIFT << or ( sfmmup htag-bspage-rehash )" 2775 " SFMMU_INVALID_SHMERID or nip ( hblktag ) " 2776 "; " 2777 2778 ": HBLK_TO_TTEP ( hmeblkp addr -- ttep ) " 2779 " over HMEBLK_MISC + PHYS-L@ HBLK_SZMASK and ( hmeblkp addr ttesz ) " 2780 " TTE8K = if ( hmeblkp addr ) " 2781 " MMU_PAGESHIFT >> NHMENTS 1- and ( hmeblkp hme-index ) " 2782 " else ( hmeblkp addr ) " 2783 " drop 0 ( hmeblkp 0 ) " 2784 " then ( hmeblkp hme-index ) " 2785 " SFHME_SIZE * + HMEBLK_HME1 + ( hmep ) " 2786 " SFHME_TTE + ( ttep ) " 2787 "; " 2788 2789 ": unix-tte ( addr cnum -- false | tte-data true ) " 2790 " KCONTEXT = if ( addr ) " 2791 " KHATID ( addr khatid ) " 2792 " else ( addr ) " 2793 " drop false exit ( false ) " 2794 " then " 2795 " ( addr khatid ) " 2796 " mmu_hashcnt 1+ 1 do ( addr sfmmup ) " 2797 " 2dup swap i HME_HASH_SHIFT " 2798 "( addr sfmmup sfmmup addr hmeshift ) " 2799 " HME_HASH_FUNCTION ( addr sfmmup hmebp ) " 2800 " over i 4 pick " 2801 "( addr sfmmup hmebp sfmmup rehash addr ) " 2802 " HME_HASH_TAG ( addr sfmmup hmebp hblktag ) " 2803 " HME_HASH_TABLE_SEARCH " 2804 "( addr sfmmup { null | hmeblkp } ) " 2805 " ?dup if ( addr sfmmup hmeblkp ) " 2806 " nip swap HBLK_TO_TTEP ( ttep ) " 2807 " dup TTE_IS_VALID if ( valid-ttep ) " 2808 " PHYS-X@ true ( tte-data true ) " 2809 " else ( invalid-tte ) " 2810 " drop false ( false ) " 2811 " then ( false | tte-data true ) " 2812 " unloop exit ( false | tte-data true ) " 2813 " then ( addr sfmmup ) " 2814 " loop ( addr sfmmup ) " 2815 " 2drop false ( false ) " 2816 "; " 2817 ; 2818 2819 void 2820 create_va_to_tte(void) 2821 { 2822 char *bp; 2823 extern int khmehash_num, uhmehash_num; 2824 extern struct hmehash_bucket *khme_hash, *uhme_hash; 2825 2826 #define OFFSET(type, field) ((uintptr_t)(&((type *)0)->field)) 2827 2828 bp = (char *)kobj_zalloc(MMU_PAGESIZE, KM_SLEEP); 2829 2830 /* 2831 * Teach obp how to parse our sw ttes. 2832 */ 2833 (void) sprintf(bp, obp_tte_str, 2834 MMU_PAGESHIFT, 2835 TTE8K, 2836 sizeof (struct sf_hment), 2837 OFFSET(struct sf_hment, hme_tte), 2838 OFFSET(struct hme_blk, hblk_tag), 2839 OFFSET(struct hme_blk, hblk_nextpa), 2840 OFFSET(struct hme_blk, hblk_misc), 2841 OFFSET(struct hme_blk, hblk_hme), 2842 NHMENTS, 2843 HBLK_SZMASK, 2844 HBLK_RANGE_SHIFT, 2845 OFFSET(struct hmehash_bucket, hmeh_nextpa), 2846 sizeof (struct hmehash_bucket), 2847 HTAG_SFMMUPSZ, 2848 HTAG_BSPAGE_SHIFT, 2849 HTAG_REHASH_SHIFT, 2850 SFMMU_INVALID_SHMERID, 2851 mmu_hashcnt, 2852 (caddr_t)va_to_pa((caddr_t)uhme_hash), 2853 (caddr_t)va_to_pa((caddr_t)khme_hash), 2854 UHMEHASH_SZ, 2855 KHMEHASH_SZ, 2856 KCONTEXT, 2857 KHATID, 2858 ASI_MEM); 2859 prom_interpret(bp, 0, 0, 0, 0, 0); 2860 2861 kobj_free(bp, MMU_PAGESIZE); 2862 } 2863 2864 void 2865 install_va_to_tte(void) 2866 { 2867 /* 2868 * advise prom that he can use unix-tte 2869 */ 2870 prom_interpret("' unix-tte is va>tte-data", 0, 0, 0, 0, 0); 2871 } 2872 2873 /* 2874 * Here we add "device-type=console" for /os-io node, for currently 2875 * our kernel console output only supports displaying text and 2876 * performing cursor-positioning operations (through kernel framebuffer 2877 * driver) and it doesn't support other functionalities required for a 2878 * standard "display" device as specified in 1275 spec. The main missing 2879 * interface defined by the 1275 spec is "draw-logo". 2880 * also see the comments above prom_stdout_is_framebuffer(). 2881 */ 2882 static char *create_node = 2883 "\" /\" find-device " 2884 "new-device " 2885 "\" os-io\" device-name " 2886 "\" "OBP_DISPLAY_CONSOLE"\" device-type " 2887 ": cb-r/w ( adr,len method$ -- #read/#written ) " 2888 " 2>r swap 2 2r> ['] $callback catch if " 2889 " 2drop 3drop 0 " 2890 " then " 2891 "; " 2892 ": read ( adr,len -- #read ) " 2893 " \" read\" ['] cb-r/w catch if 2drop 2drop -2 exit then " 2894 " ( retN ... ret1 N ) " 2895 " ?dup if " 2896 " swap >r 1- 0 ?do drop loop r> " 2897 " else " 2898 " -2 " 2899 " then " 2900 "; " 2901 ": write ( adr,len -- #written ) " 2902 " \" write\" ['] cb-r/w catch if 2drop 2drop 0 exit then " 2903 " ( retN ... ret1 N ) " 2904 " ?dup if " 2905 " swap >r 1- 0 ?do drop loop r> " 2906 " else " 2907 " 0 " 2908 " then " 2909 "; " 2910 ": poll-tty ( -- ) ; " 2911 ": install-abort ( -- ) ['] poll-tty d# 10 alarm ; " 2912 ": remove-abort ( -- ) ['] poll-tty 0 alarm ; " 2913 ": cb-give/take ( $method -- ) " 2914 " 0 -rot ['] $callback catch ?dup if " 2915 " >r 2drop 2drop r> throw " 2916 " else " 2917 " 0 ?do drop loop " 2918 " then " 2919 "; " 2920 ": give ( -- ) \" exit-input\" cb-give/take ; " 2921 ": take ( -- ) \" enter-input\" cb-give/take ; " 2922 ": open ( -- ok? ) true ; " 2923 ": close ( -- ) ; " 2924 "finish-device " 2925 "device-end "; 2926 2927 /* 2928 * Create the OBP input/output node (FCode serial driver). 2929 * It is needed for both USB console keyboard and for 2930 * the kernel terminal emulator. It is too early to check for a 2931 * kernel console compatible framebuffer now, so we create this 2932 * so that we're ready if we need to enable kernel terminal emulation. 2933 * 2934 * When the USB software takes over the input device at the time 2935 * consconfig runs, OBP's stdin is redirected to this node. 2936 * Whenever the FORTH user interface is used after this switch, 2937 * the node will call back into the kernel for console input. 2938 * If a serial device such as ttya or a UART with a Type 5 keyboard 2939 * attached is used, OBP takes over the serial device when the system 2940 * goes to the debugger after the system is booted. This sharing 2941 * of the relatively simple serial device is difficult but possible. 2942 * Sharing the USB host controller is impossible due its complexity. 2943 * 2944 * Similarly to USB keyboard input redirection, after consconfig_dacf 2945 * configures a kernel console framebuffer as the standard output 2946 * device, OBP's stdout is switched to to vector through the 2947 * /os-io node into the kernel terminal emulator. 2948 */ 2949 static void 2950 startup_create_io_node(void) 2951 { 2952 prom_interpret(create_node, 0, 0, 0, 0, 0); 2953 } 2954 2955 2956 static void 2957 do_prom_version_check(void) 2958 { 2959 int i; 2960 pnode_t node; 2961 char buf[64]; 2962 static char drev[] = "Down-rev firmware detected%s\n" 2963 "\tPlease upgrade to the following minimum version:\n" 2964 "\t\t%s\n"; 2965 2966 i = prom_version_check(buf, sizeof (buf), &node); 2967 2968 if (i == PROM_VER64_OK) 2969 return; 2970 2971 if (i == PROM_VER64_UPGRADE) { 2972 cmn_err(CE_WARN, drev, "", buf); 2973 2974 #ifdef DEBUG 2975 prom_enter_mon(); /* Type 'go' to continue */ 2976 cmn_err(CE_WARN, "Booting with down-rev firmware\n"); 2977 return; 2978 #else 2979 halt(0); 2980 #endif 2981 } 2982 2983 /* 2984 * The other possibility is that this is a server running 2985 * good firmware, but down-rev firmware was detected on at 2986 * least one other cpu board. We just complain if we see 2987 * that. 2988 */ 2989 cmn_err(CE_WARN, drev, " on one or more CPU boards", buf); 2990 } 2991 2992 2993 /* 2994 * Must be defined in platform dependent code. 2995 */ 2996 extern caddr_t modtext; 2997 extern size_t modtext_sz; 2998 extern caddr_t moddata; 2999 3000 #define HEAPTEXT_ARENA(addr) \ 3001 ((uintptr_t)(addr) < KERNELBASE + 2 * MMU_PAGESIZE4M ? 0 : \ 3002 (((uintptr_t)(addr) - HEAPTEXT_BASE) / \ 3003 (HEAPTEXT_MAPPED + HEAPTEXT_UNMAPPED) + 1)) 3004 3005 #define HEAPTEXT_OVERSIZED(addr) \ 3006 ((uintptr_t)(addr) >= HEAPTEXT_BASE + HEAPTEXT_SIZE - HEAPTEXT_OVERSIZE) 3007 3008 vmem_t *texthole_source[HEAPTEXT_NARENAS]; 3009 vmem_t *texthole_arena[HEAPTEXT_NARENAS]; 3010 kmutex_t texthole_lock; 3011 3012 char kern_bootargs[OBP_MAXPATHLEN]; 3013 3014 void 3015 kobj_vmem_init(vmem_t **text_arena, vmem_t **data_arena) 3016 { 3017 uintptr_t addr, limit; 3018 3019 addr = HEAPTEXT_BASE; 3020 limit = addr + HEAPTEXT_SIZE - HEAPTEXT_OVERSIZE; 3021 3022 /* 3023 * Before we initialize the text_arena, we want to punch holes in the 3024 * underlying heaptext_arena. This guarantees that for any text 3025 * address we can find a text hole less than HEAPTEXT_MAPPED away. 3026 */ 3027 for (; addr + HEAPTEXT_UNMAPPED <= limit; 3028 addr += HEAPTEXT_MAPPED + HEAPTEXT_UNMAPPED) { 3029 (void) vmem_xalloc(heaptext_arena, HEAPTEXT_UNMAPPED, PAGESIZE, 3030 0, 0, (void *)addr, (void *)(addr + HEAPTEXT_UNMAPPED), 3031 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 3032 } 3033 3034 /* 3035 * Allocate one page at the oversize to break up the text region 3036 * from the oversized region. 3037 */ 3038 (void) vmem_xalloc(heaptext_arena, PAGESIZE, PAGESIZE, 0, 0, 3039 (void *)limit, (void *)(limit + PAGESIZE), 3040 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 3041 3042 *text_arena = vmem_create("module_text", modtext_sz ? modtext : NULL, 3043 modtext_sz, sizeof (uintptr_t), segkmem_alloc, segkmem_free, 3044 heaptext_arena, 0, VM_SLEEP); 3045 *data_arena = vmem_create("module_data", moddata, MODDATA, 1, 3046 segkmem_alloc, segkmem_free, heap32_arena, 0, VM_SLEEP); 3047 } 3048 3049 caddr_t 3050 kobj_text_alloc(vmem_t *arena, size_t size) 3051 { 3052 caddr_t rval, better; 3053 3054 /* 3055 * First, try a sleeping allocation. 3056 */ 3057 rval = vmem_alloc(arena, size, VM_SLEEP | VM_BESTFIT); 3058 3059 if (size >= HEAPTEXT_MAPPED || !HEAPTEXT_OVERSIZED(rval)) 3060 return (rval); 3061 3062 /* 3063 * We didn't get the area that we wanted. We're going to try to do an 3064 * allocation with explicit constraints. 3065 */ 3066 better = vmem_xalloc(arena, size, sizeof (uintptr_t), 0, 0, NULL, 3067 (void *)(HEAPTEXT_BASE + HEAPTEXT_SIZE - HEAPTEXT_OVERSIZE), 3068 VM_NOSLEEP | VM_BESTFIT); 3069 3070 if (better != NULL) { 3071 /* 3072 * That worked. Free our first attempt and return. 3073 */ 3074 vmem_free(arena, rval, size); 3075 return (better); 3076 } 3077 3078 /* 3079 * That didn't work; we'll have to return our first attempt. 3080 */ 3081 return (rval); 3082 } 3083 3084 caddr_t 3085 kobj_texthole_alloc(caddr_t addr, size_t size) 3086 { 3087 int arena = HEAPTEXT_ARENA(addr); 3088 char c[30]; 3089 uintptr_t base; 3090 3091 if (HEAPTEXT_OVERSIZED(addr)) { 3092 /* 3093 * If this is an oversized allocation, there is no text hole 3094 * available for it; return NULL. 3095 */ 3096 return (NULL); 3097 } 3098 3099 mutex_enter(&texthole_lock); 3100 3101 if (texthole_arena[arena] == NULL) { 3102 ASSERT(texthole_source[arena] == NULL); 3103 3104 if (arena == 0) { 3105 texthole_source[0] = vmem_create("module_text_holesrc", 3106 (void *)(KERNELBASE + MMU_PAGESIZE4M), 3107 MMU_PAGESIZE4M, PAGESIZE, NULL, NULL, NULL, 3108 0, VM_SLEEP); 3109 } else { 3110 base = HEAPTEXT_BASE + 3111 (arena - 1) * (HEAPTEXT_MAPPED + HEAPTEXT_UNMAPPED); 3112 3113 (void) snprintf(c, sizeof (c), 3114 "heaptext_holesrc_%d", arena); 3115 3116 texthole_source[arena] = vmem_create(c, (void *)base, 3117 HEAPTEXT_UNMAPPED, PAGESIZE, NULL, NULL, NULL, 3118 0, VM_SLEEP); 3119 } 3120 3121 (void) snprintf(c, sizeof (c), "heaptext_hole_%d", arena); 3122 3123 texthole_arena[arena] = vmem_create(c, NULL, 0, 3124 sizeof (uint32_t), segkmem_alloc_permanent, segkmem_free, 3125 texthole_source[arena], 0, VM_SLEEP); 3126 } 3127 3128 mutex_exit(&texthole_lock); 3129 3130 ASSERT(texthole_arena[arena] != NULL); 3131 ASSERT(arena >= 0 && arena < HEAPTEXT_NARENAS); 3132 return (vmem_alloc(texthole_arena[arena], size, 3133 VM_BESTFIT | VM_NOSLEEP)); 3134 } 3135 3136 void 3137 kobj_texthole_free(caddr_t addr, size_t size) 3138 { 3139 int arena = HEAPTEXT_ARENA(addr); 3140 3141 ASSERT(arena >= 0 && arena < HEAPTEXT_NARENAS); 3142 ASSERT(texthole_arena[arena] != NULL); 3143 vmem_free(texthole_arena[arena], addr, size); 3144 } 3145 3146 void 3147 release_bootstrap(void) 3148 { 3149 if (&cif_init) 3150 cif_init(); 3151 } 3152