1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 22 /* 23 * Copyright 2007 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 #pragma ident "%Z%%M% %I% %E% SMI" 28 29 #include <sys/machsystm.h> 30 #include <sys/archsystm.h> 31 #include <sys/vm.h> 32 #include <sys/cpu.h> 33 #include <sys/atomic.h> 34 #include <sys/reboot.h> 35 #include <sys/kdi.h> 36 #include <sys/bootconf.h> 37 #include <sys/memlist_plat.h> 38 #include <sys/memlist_impl.h> 39 #include <sys/prom_plat.h> 40 #include <sys/prom_isa.h> 41 #include <sys/autoconf.h> 42 #include <sys/intreg.h> 43 #include <sys/ivintr.h> 44 #include <sys/fpu/fpusystm.h> 45 #include <sys/iommutsb.h> 46 #include <vm/vm_dep.h> 47 #include <vm/seg_dev.h> 48 #include <vm/seg_kmem.h> 49 #include <vm/seg_kpm.h> 50 #include <vm/seg_map.h> 51 #include <vm/seg_kp.h> 52 #include <sys/sysconf.h> 53 #include <vm/hat_sfmmu.h> 54 #include <sys/kobj.h> 55 #include <sys/sun4asi.h> 56 #include <sys/clconf.h> 57 #include <sys/platform_module.h> 58 #include <sys/panic.h> 59 #include <sys/cpu_sgnblk_defs.h> 60 #include <sys/clock.h> 61 #include <sys/cmn_err.h> 62 #include <sys/promif.h> 63 #include <sys/prom_debug.h> 64 #include <sys/traptrace.h> 65 #include <sys/memnode.h> 66 #include <sys/mem_cage.h> 67 #include <sys/mmu.h> 68 69 extern void setup_trap_table(void); 70 extern 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 98 /* 99 * External Data: 100 */ 101 extern int vac_size; /* cache size in bytes */ 102 extern uint_t vac_mask; /* VAC alignment consistency mask */ 103 extern uint_t vac_colors; 104 105 /* 106 * Global Data Definitions: 107 */ 108 109 /* 110 * XXX - Don't port this to new architectures 111 * A 3rd party volume manager driver (vxdm) depends on the symbol romp. 112 * 'romp' has no use with a prom with an IEEE 1275 client interface. 113 * The driver doesn't use the value, but it depends on the symbol. 114 */ 115 void *romp; /* veritas driver won't load without romp 4154976 */ 116 /* 117 * Declare these as initialized data so we can patch them. 118 */ 119 pgcnt_t physmem = 0; /* memory size in pages, patch if you want less */ 120 pgcnt_t segkpsize = 121 btop(SEGKPDEFSIZE); /* size of segkp segment in pages */ 122 uint_t segmap_percent = 12; /* Size of segmap segment */ 123 124 int use_cache = 1; /* cache not reliable (605 bugs) with MP */ 125 int vac_copyback = 1; 126 char *cache_mode = NULL; 127 int use_mix = 1; 128 int prom_debug = 0; 129 130 struct bootops *bootops = 0; /* passed in from boot in %o2 */ 131 caddr_t boot_tba; /* %tba at boot - used by kmdb */ 132 uint_t tba_taken_over = 0; 133 134 caddr_t s_text; /* start of kernel text segment */ 135 caddr_t e_text; /* end of kernel text segment */ 136 caddr_t s_data; /* start of kernel data segment */ 137 caddr_t e_data; /* end of kernel data segment */ 138 139 caddr_t modtext; /* beginning of module text */ 140 size_t modtext_sz; /* size of module text */ 141 caddr_t moddata; /* beginning of module data reserve */ 142 caddr_t e_moddata; /* end of module data reserve */ 143 144 /* 145 * End of first block of contiguous kernel in 32-bit virtual address space 146 */ 147 caddr_t econtig32; /* end of first blk of contiguous kernel */ 148 149 caddr_t ncbase; /* beginning of non-cached segment */ 150 caddr_t ncend; /* end of non-cached segment */ 151 caddr_t sdata; /* beginning of data segment */ 152 153 caddr_t extra_etva; /* beginning of unused nucleus text */ 154 pgcnt_t extra_etpg; /* number of pages of unused nucleus text */ 155 156 size_t ndata_remain_sz; /* bytes from end of data to 4MB boundary */ 157 caddr_t nalloc_base; /* beginning of nucleus allocation */ 158 caddr_t nalloc_end; /* end of nucleus allocatable memory */ 159 caddr_t valloc_base; /* beginning of kvalloc segment */ 160 161 caddr_t kmem64_base; /* base of kernel mem segment in 64-bit space */ 162 caddr_t kmem64_end; /* end of kernel mem segment in 64-bit space */ 163 caddr_t kmem64_aligned_end; /* end of large page, overmaps 64-bit space */ 164 int kmem64_alignsize; /* page size for mem segment in 64-bit space */ 165 int kmem64_szc; /* page size code */ 166 uint64_t kmem64_pabase = (uint64_t)-1; /* physical address of kmem64_base */ 167 168 uintptr_t shm_alignment; /* VAC address consistency modulus */ 169 struct memlist *phys_install; /* Total installed physical memory */ 170 struct memlist *phys_avail; /* Available (unreserved) physical memory */ 171 struct memlist *virt_avail; /* Available (unmapped?) virtual memory */ 172 struct memlist ndata; /* memlist of nucleus allocatable memory */ 173 int memexp_flag; /* memory expansion card flag */ 174 uint64_t ecache_flushaddr; /* physical address used for flushing E$ */ 175 pgcnt_t obp_pages; /* Physical pages used by OBP */ 176 177 /* 178 * VM data structures 179 */ 180 long page_hashsz; /* Size of page hash table (power of two) */ 181 struct page *pp_base; /* Base of system page struct array */ 182 size_t pp_sz; /* Size in bytes of page struct array */ 183 struct page **page_hash; /* Page hash table */ 184 struct seg ktextseg; /* Segment used for kernel executable image */ 185 struct seg kvalloc; /* Segment used for "valloc" mapping */ 186 struct seg kpseg; /* Segment used for pageable kernel virt mem */ 187 struct seg ktexthole; /* Segment used for nucleus text hole */ 188 struct seg kmapseg; /* Segment used for generic kernel mappings */ 189 struct seg kpmseg; /* Segment used for physical mapping */ 190 struct seg kdebugseg; /* Segment used for the kernel debugger */ 191 192 uintptr_t kpm_pp_base; /* Base of system kpm_page array */ 193 size_t kpm_pp_sz; /* Size of system kpm_page array */ 194 pgcnt_t kpm_npages; /* How many kpm pages are managed */ 195 196 struct seg *segkp = &kpseg; /* Pageable kernel virtual memory segment */ 197 struct seg *segkmap = &kmapseg; /* Kernel generic mapping segment */ 198 struct seg *segkpm = &kpmseg; /* 64bit kernel physical mapping segment */ 199 200 int segzio_fromheap = 0; /* zio allocations occur from heap */ 201 caddr_t segzio_base; /* Base address of segzio */ 202 pgcnt_t segziosize = 0; /* size of zio segment in pages */ 203 204 /* 205 * debugger pages (if allocated) 206 */ 207 struct vnode kdebugvp; 208 209 /* 210 * VA range available to the debugger 211 */ 212 const caddr_t kdi_segdebugbase = (const caddr_t)SEGDEBUGBASE; 213 const size_t kdi_segdebugsize = SEGDEBUGSIZE; 214 215 /* 216 * Segment for relocated kernel structures in 64-bit large RAM kernels 217 */ 218 struct seg kmem64; 219 220 struct memseg *memseg_base; 221 size_t memseg_sz; /* Used to translate a va to page */ 222 struct vnode unused_pages_vp; 223 224 /* 225 * VM data structures allocated early during boot. 226 */ 227 size_t pagehash_sz; 228 uint64_t memlist_sz; 229 230 char tbr_wr_addr_inited = 0; 231 232 233 /* 234 * Static Routines: 235 */ 236 static void memlist_add(uint64_t, uint64_t, struct memlist **, 237 struct memlist **); 238 static void kphysm_init(page_t *, struct memseg *, pgcnt_t, uintptr_t, 239 pgcnt_t); 240 static void kvm_init(void); 241 242 static void startup_init(void); 243 static void startup_memlist(void); 244 static void startup_modules(void); 245 static void startup_bop_gone(void); 246 static void startup_vm(void); 247 static void startup_end(void); 248 static void setup_cage_params(void); 249 static void startup_create_io_node(void); 250 251 static pgcnt_t npages; 252 static struct memlist *memlist; 253 void *memlist_end; 254 255 static pgcnt_t bop_alloc_pages; 256 static caddr_t hblk_base; 257 uint_t hblk_alloc_dynamic = 0; 258 uint_t hblk1_min = H1MIN; 259 260 261 /* 262 * Hooks for unsupported platforms and down-rev firmware 263 */ 264 int iam_positron(void); 265 #pragma weak iam_positron 266 static void do_prom_version_check(void); 267 static void kpm_init(void); 268 static void kpm_npages_setup(int); 269 static void kpm_memseg_init(void); 270 271 /* 272 * After receiving a thermal interrupt, this is the number of seconds 273 * to delay before shutting off the system, assuming 274 * shutdown fails. Use /etc/system to change the delay if this isn't 275 * large enough. 276 */ 277 int thermal_powerdown_delay = 1200; 278 279 /* 280 * Used to hold off page relocations into the cage until OBP has completed 281 * its boot-time handoff of its resources to the kernel. 282 */ 283 int page_relocate_ready = 0; 284 285 /* 286 * Enable some debugging messages concerning memory usage... 287 */ 288 #ifdef DEBUGGING_MEM 289 static int debugging_mem; 290 static void 291 printmemlist(char *title, struct memlist *list) 292 { 293 if (!debugging_mem) 294 return; 295 296 printf("%s\n", title); 297 298 while (list) { 299 prom_printf("\taddr = 0x%x %8x, size = 0x%x %8x\n", 300 (uint32_t)(list->address >> 32), (uint32_t)list->address, 301 (uint32_t)(list->size >> 32), (uint32_t)(list->size)); 302 list = list->next; 303 } 304 } 305 306 void 307 printmemseg(struct memseg *memseg) 308 { 309 if (!debugging_mem) 310 return; 311 312 printf("memseg\n"); 313 314 while (memseg) { 315 prom_printf("\tpage = 0x%p, epage = 0x%p, " 316 "pfn = 0x%x, epfn = 0x%x\n", 317 memseg->pages, memseg->epages, 318 memseg->pages_base, memseg->pages_end); 319 memseg = memseg->next; 320 } 321 } 322 323 #define debug_pause(str) halt((str)) 324 #define MPRINTF(str) if (debugging_mem) prom_printf((str)) 325 #define MPRINTF1(str, a) if (debugging_mem) prom_printf((str), (a)) 326 #define MPRINTF2(str, a, b) if (debugging_mem) prom_printf((str), (a), (b)) 327 #define MPRINTF3(str, a, b, c) \ 328 if (debugging_mem) prom_printf((str), (a), (b), (c)) 329 #else /* DEBUGGING_MEM */ 330 #define MPRINTF(str) 331 #define MPRINTF1(str, a) 332 #define MPRINTF2(str, a, b) 333 #define MPRINTF3(str, a, b, c) 334 #endif /* DEBUGGING_MEM */ 335 336 /* Simple message to indicate that the bootops pointer has been zeroed */ 337 #ifdef DEBUG 338 static int bootops_gone_on = 0; 339 #define BOOTOPS_GONE() \ 340 if (bootops_gone_on) \ 341 prom_printf("The bootops vec is zeroed now!\n"); 342 #else 343 #define BOOTOPS_GONE() 344 #endif /* DEBUG */ 345 346 /* 347 * Monitor pages may not be where this says they are. 348 * and the debugger may not be there either. 349 * 350 * Note that 'pages' here are *physical* pages, which are 8k on sun4u. 351 * 352 * Physical memory layout 353 * (not necessarily contiguous) 354 * (THIS IS SOMEWHAT WRONG) 355 * /-----------------------\ 356 * | monitor pages | 357 * availmem -|-----------------------| 358 * | | 359 * | page pool | 360 * | | 361 * |-----------------------| 362 * | configured tables | 363 * | buffers | 364 * firstaddr -|-----------------------| 365 * | hat data structures | 366 * |-----------------------| 367 * | kernel data, bss | 368 * |-----------------------| 369 * | interrupt stack | 370 * |-----------------------| 371 * | kernel text (RO) | 372 * |-----------------------| 373 * | trap table (4k) | 374 * |-----------------------| 375 * page 1 | panicbuf | 376 * |-----------------------| 377 * page 0 | reclaimed | 378 * |_______________________| 379 * 380 * 381 * 382 * Kernel's Virtual Memory Layout. 383 * /-----------------------\ 384 * 0xFFFFFFFF.FFFFFFFF -| |- 385 * | OBP's virtual page | 386 * | tables | 387 * 0xFFFFFFFC.00000000 -|-----------------------|- 388 * : : 389 * : : 390 * -|-----------------------|- 391 * | segzio | (base and size vary) 392 * 0xFFFFFE00.00000000 -|-----------------------|- 393 * | | Ultrasparc I/II support 394 * | segkpm segment | up to 2TB of physical 395 * | (64-bit kernel ONLY) | memory, VAC has 2 colors 396 * | | 397 * 0xFFFFFA00.00000000 -|-----------------------|- 2TB segkpm alignment 398 * : : 399 * : : 400 * 0xFFFFF810.00000000 -|-----------------------|- hole_end 401 * | | ^ 402 * | UltraSPARC I/II call | | 403 * | bug requires an extra | | 404 * | 4 GB of space between | | 405 * | hole and used RAM | | 406 * | | | 407 * 0xFFFFF800.00000000 -|-----------------------|- | 408 * | | | 409 * | Virtual Address Hole | UltraSPARC 410 * | on UltraSPARC I/II | I/II * ONLY * 411 * | | | 412 * 0x00000800.00000000 -|-----------------------|- | 413 * | | | 414 * | UltraSPARC I/II call | | 415 * | bug requires an extra | | 416 * | 4 GB of space between | | 417 * | hole and used RAM | | 418 * | | v 419 * 0x000007FF.00000000 -|-----------------------|- hole_start ----- 420 * : : ^ 421 * : : | 422 * 0x00000XXX.XXX00000 -|-----------------------|- kmem64_ | 423 * | overmapped area | alignend_end | 424 * | (kmem64_alignsize | | 425 * | boundary) | | 426 * 0x00000XXX.XXXXXXXX -|-----------------------|- kmem64_end | 427 * | | | 428 * | 64-bit kernel ONLY | | 429 * | | | 430 * | kmem64 segment | | 431 * | | | 432 * | (Relocated extra HME | Approximately 433 * | block allocations, | 1 TB of virtual 434 * | memnode freelists, | address space 435 * | HME hash buckets, | | 436 * | mml_table, kpmp_table,| | 437 * | page_t array and | | 438 * | hashblock pool to | | 439 * | avoid hard-coded | | 440 * | 32-bit vaddr | | 441 * | limitations) | | 442 * | | v 443 * 0x00000700.00000000 -|-----------------------|- SYSLIMIT (kmem64_base) 444 * | | 445 * | segkmem segment | (SYSLIMIT - SYSBASE = 4TB) 446 * | | 447 * 0x00000300.00000000 -|-----------------------|- SYSBASE 448 * : : 449 * : : 450 * -|-----------------------|- 451 * | | 452 * | segmap segment | SEGMAPSIZE (1/8th physmem, 453 * | | 256G MAX) 454 * 0x000002a7.50000000 -|-----------------------|- SEGMAPBASE 455 * : : 456 * : : 457 * -|-----------------------|- 458 * | | 459 * | segkp | SEGKPSIZE (2GB) 460 * | | 461 * | | 462 * 0x000002a1.00000000 -|-----------------------|- SEGKPBASE 463 * | | 464 * 0x000002a0.00000000 -|-----------------------|- MEMSCRUBBASE 465 * | | (SEGKPBASE - 0x400000) 466 * 0x0000029F.FFE00000 -|-----------------------|- ARGSBASE 467 * | | (MEMSCRUBBASE - NCARGS) 468 * 0x0000029F.FFD80000 -|-----------------------|- PPMAPBASE 469 * | | (ARGSBASE - PPMAPSIZE) 470 * 0x0000029F.FFD00000 -|-----------------------|- PPMAP_FAST_BASE 471 * | | 472 * 0x0000029F.FF980000 -|-----------------------|- PIOMAPBASE 473 * | | 474 * 0x0000029F.FF580000 -|-----------------------|- NARG_BASE 475 * : : 476 * : : 477 * 0x00000000.FFFFFFFF -|-----------------------|- OFW_END_ADDR 478 * | | 479 * | OBP | 480 * | | 481 * 0x00000000.F0000000 -|-----------------------|- OFW_START_ADDR 482 * | kmdb | 483 * 0x00000000.EDD00000 -|-----------------------|- SEGDEBUGBASE 484 * : : 485 * : : 486 * 0x00000000.7c000000 -|-----------------------|- SYSLIMIT32 487 * | | 488 * | segkmem32 segment | (SYSLIMIT32 - SYSBASE32 = 489 * | | ~64MB) 490 * 0x00000000.78002000 -|-----------------------| 491 * | panicbuf | 492 * 0x00000000.78000000 -|-----------------------|- SYSBASE32 493 * : : 494 * : : 495 * | | 496 * |-----------------------|- econtig32 497 * | vm structures | 498 * 0x00000000.01C00000 |-----------------------|- nalloc_end 499 * | TSBs | 500 * |-----------------------|- end/nalloc_base 501 * | kernel data & bss | 502 * 0x00000000.01800000 -|-----------------------| 503 * : nucleus text hole : 504 * 0x00000000.01400000 -|-----------------------| 505 * : : 506 * |-----------------------| 507 * | module text | 508 * |-----------------------|- e_text/modtext 509 * | kernel text | 510 * |-----------------------| 511 * | trap table (48k) | 512 * 0x00000000.01000000 -|-----------------------|- KERNELBASE 513 * | reserved for trapstat |} TSTAT_TOTAL_SIZE 514 * |-----------------------| 515 * | | 516 * | invalid | 517 * | | 518 * 0x00000000.00000000 _|_______________________| 519 * 520 * 521 * 522 * 32-bit User Virtual Memory Layout. 523 * /-----------------------\ 524 * | | 525 * | invalid | 526 * | | 527 * 0xFFC00000 -|-----------------------|- USERLIMIT 528 * | user stack | 529 * : : 530 * : : 531 * : : 532 * | user data | 533 * -|-----------------------|- 534 * | user text | 535 * 0x00002000 -|-----------------------|- 536 * | invalid | 537 * 0x00000000 _|_______________________| 538 * 539 * 540 * 541 * 64-bit User Virtual Memory Layout. 542 * /-----------------------\ 543 * | | 544 * | invalid | 545 * | | 546 * 0xFFFFFFFF.80000000 -|-----------------------|- USERLIMIT 547 * | user stack | 548 * : : 549 * : : 550 * : : 551 * | user data | 552 * -|-----------------------|- 553 * | user text | 554 * 0x00000000.00100000 -|-----------------------|- 555 * | invalid | 556 * 0x00000000.00000000 _|_______________________| 557 */ 558 559 extern caddr_t ecache_init_scrub_flush_area(caddr_t alloc_base); 560 extern uint64_t ecache_flush_address(void); 561 562 #pragma weak load_platform_modules 563 #pragma weak plat_startup_memlist 564 #pragma weak ecache_init_scrub_flush_area 565 #pragma weak ecache_flush_address 566 567 568 /* 569 * By default the DR Cage is enabled for maximum OS 570 * MPSS performance. Users needing to disable the cage mechanism 571 * can set this variable to zero via /etc/system. 572 * Disabling the cage on systems supporting Dynamic Reconfiguration (DR) 573 * will result in loss of DR functionality. 574 * Platforms wishing to disable kernel Cage by default 575 * should do so in their set_platform_defaults() routine. 576 */ 577 int kernel_cage_enable = 1; 578 579 static void 580 setup_cage_params(void) 581 { 582 void (*func)(void); 583 584 func = (void (*)(void))kobj_getsymvalue("set_platform_cage_params", 0); 585 if (func != NULL) { 586 (*func)(); 587 return; 588 } 589 590 if (kernel_cage_enable == 0) { 591 return; 592 } 593 kcage_range_lock(); 594 if (kcage_range_init(phys_avail, 1) == 0) { 595 kcage_init(total_pages / 256); 596 } 597 kcage_range_unlock(); 598 599 if (kcage_on) { 600 cmn_err(CE_NOTE, "!Kernel Cage is ENABLED"); 601 } else { 602 cmn_err(CE_NOTE, "!Kernel Cage is DISABLED"); 603 } 604 605 } 606 607 /* 608 * Machine-dependent startup code 609 */ 610 void 611 startup(void) 612 { 613 startup_init(); 614 if (&startup_platform) 615 startup_platform(); 616 startup_memlist(); 617 startup_modules(); 618 setup_cage_params(); 619 startup_bop_gone(); 620 startup_vm(); 621 startup_end(); 622 } 623 624 struct regs sync_reg_buf; 625 uint64_t sync_tt; 626 627 void 628 sync_handler(void) 629 { 630 struct trap_info ti; 631 int i; 632 633 /* 634 * Prevent trying to talk to the other CPUs since they are 635 * sitting in the prom and won't reply. 636 */ 637 for (i = 0; i < NCPU; i++) { 638 if ((i != CPU->cpu_id) && CPU_XCALL_READY(i)) { 639 cpu[i]->cpu_flags &= ~CPU_READY; 640 cpu[i]->cpu_flags |= CPU_QUIESCED; 641 CPUSET_DEL(cpu_ready_set, cpu[i]->cpu_id); 642 } 643 } 644 645 /* 646 * We've managed to get here without going through the 647 * normal panic code path. Try and save some useful 648 * information. 649 */ 650 if (!panicstr && (curthread->t_panic_trap == NULL)) { 651 ti.trap_type = sync_tt; 652 ti.trap_regs = &sync_reg_buf; 653 ti.trap_addr = NULL; 654 ti.trap_mmu_fsr = 0x0; 655 656 curthread->t_panic_trap = &ti; 657 } 658 659 /* 660 * If we're re-entering the panic path, update the signature 661 * block so that the SC knows we're in the second part of panic. 662 */ 663 if (panicstr) 664 CPU_SIGNATURE(OS_SIG, SIGST_EXIT, SIGSUBST_DUMP, -1); 665 666 nopanicdebug = 1; /* do not perform debug_enter() prior to dump */ 667 panic("sync initiated"); 668 } 669 670 671 static void 672 startup_init(void) 673 { 674 /* 675 * We want to save the registers while we're still in OBP 676 * so that we know they haven't been fiddled with since. 677 * (In principle, OBP can't change them just because it 678 * makes a callback, but we'd rather not depend on that 679 * behavior.) 680 */ 681 char sync_str[] = 682 "warning @ warning off : sync " 683 "%%tl-c %%tstate h# %p x! " 684 "%%g1 h# %p x! %%g2 h# %p x! %%g3 h# %p x! " 685 "%%g4 h# %p x! %%g5 h# %p x! %%g6 h# %p x! " 686 "%%g7 h# %p x! %%o0 h# %p x! %%o1 h# %p x! " 687 "%%o2 h# %p x! %%o3 h# %p x! %%o4 h# %p x! " 688 "%%o5 h# %p x! %%o6 h# %p x! %%o7 h# %p x! " 689 "%%tl-c %%tpc h# %p x! %%tl-c %%tnpc h# %p x! " 690 "%%y h# %p l! %%tl-c %%tt h# %p x! " 691 "sync ; warning !"; 692 693 /* 694 * 20 == num of %p substrings 695 * 16 == max num of chars %p will expand to. 696 */ 697 char bp[sizeof (sync_str) + 16 * 20]; 698 699 (void) check_boot_version(BOP_GETVERSION(bootops)); 700 701 /* 702 * Initialize ptl1 stack for the 1st CPU. 703 */ 704 ptl1_init_cpu(&cpu0); 705 706 /* 707 * Initialize the address map for cache consistent mappings 708 * to random pages; must be done after vac_size is set. 709 */ 710 ppmapinit(); 711 712 /* 713 * Initialize the PROM callback handler. 714 */ 715 init_vx_handler(); 716 717 /* 718 * have prom call sync_callback() to handle the sync and 719 * save some useful information which will be stored in the 720 * core file later. 721 */ 722 (void) sprintf((char *)bp, sync_str, 723 (void *)&sync_reg_buf.r_tstate, (void *)&sync_reg_buf.r_g1, 724 (void *)&sync_reg_buf.r_g2, (void *)&sync_reg_buf.r_g3, 725 (void *)&sync_reg_buf.r_g4, (void *)&sync_reg_buf.r_g5, 726 (void *)&sync_reg_buf.r_g6, (void *)&sync_reg_buf.r_g7, 727 (void *)&sync_reg_buf.r_o0, (void *)&sync_reg_buf.r_o1, 728 (void *)&sync_reg_buf.r_o2, (void *)&sync_reg_buf.r_o3, 729 (void *)&sync_reg_buf.r_o4, (void *)&sync_reg_buf.r_o5, 730 (void *)&sync_reg_buf.r_o6, (void *)&sync_reg_buf.r_o7, 731 (void *)&sync_reg_buf.r_pc, (void *)&sync_reg_buf.r_npc, 732 (void *)&sync_reg_buf.r_y, (void *)&sync_tt); 733 prom_interpret(bp, 0, 0, 0, 0, 0); 734 add_vx_handler("sync", 1, (void (*)(cell_t *))sync_handler); 735 } 736 737 static u_longlong_t *boot_physinstalled, *boot_physavail, *boot_virtavail; 738 static size_t boot_physinstalled_len, boot_physavail_len, boot_virtavail_len; 739 740 #define IVSIZE ((MAXIVNUM * sizeof (intr_vec_t *)) + \ 741 (MAX_RSVD_IV * sizeof (intr_vec_t)) + \ 742 (MAX_RSVD_IVX * sizeof (intr_vecx_t))) 743 744 #if !defined(C_OBP) 745 /* 746 * Install a temporary tte handler in OBP for kmem64 area. 747 * 748 * We map kmem64 area with large pages before the trap table is taken 749 * over. Since OBP makes 8K mappings, it can create 8K tlb entries in 750 * the same area. Duplicate tlb entries with different page sizes 751 * cause unpredicatble behavior. To avoid this, we don't create 752 * kmem64 mappings via BOP_ALLOC (ends up as prom_alloc() call to 753 * OBP). Instead, we manage translations with a temporary va>tte-data 754 * handler (kmem64-tte). This handler is replaced by unix-tte when 755 * the trap table is taken over. 756 * 757 * The temporary handler knows the physical address of the kmem64 758 * area. It uses the prom's pgmap@ Forth word for other addresses. 759 * 760 * We have to use BOP_ALLOC() method for C-OBP platforms because 761 * pgmap@ is not defined in C-OBP. C-OBP is only used on serengeti 762 * sun4u platforms. On sun4u we flush tlb after trap table is taken 763 * over if we use large pages for kernel heap and kmem64. Since sun4u 764 * prom (unlike sun4v) calls va>tte-data first for client address 765 * translation prom's ttes for kmem64 can't get into TLB even if we 766 * later switch to prom's trap table again. C-OBP uses 4M pages for 767 * client mappings when possible so on all platforms we get the 768 * benefit from large mappings for kmem64 area immediately during 769 * boot. 770 * 771 * pseudo code: 772 * if (context != 0) { 773 * return false 774 * } else if (miss_va in range[kmem64_base, kmem64_end)) { 775 * tte = tte_template + 776 * (((miss_va & pagemask) - kmem64_base)); 777 * return tte, true 778 * } else { 779 * return pgmap@ result 780 * } 781 */ 782 char kmem64_obp_str[] = 783 "h# %lx constant kmem64_base " 784 "h# %lx constant kmem64_end " 785 "h# %lx constant kmem64_pagemask " 786 "h# %lx constant kmem64_template " 787 788 ": kmem64-tte ( addr cnum -- false | tte-data true ) " 789 " if ( addr ) " 790 " drop false exit then ( false ) " 791 " dup kmem64_base kmem64_end within if ( addr ) " 792 " kmem64_pagemask and ( addr' ) " 793 " kmem64_base - ( addr' ) " 794 " kmem64_template + ( tte ) " 795 " true ( tte true ) " 796 " else ( addr ) " 797 " pgmap@ ( tte ) " 798 " dup 0< if true else drop false then ( tte true | false ) " 799 " then ( tte true | false ) " 800 "; " 801 802 "' kmem64-tte is va>tte-data " 803 ; 804 805 void 806 install_kmem64_tte() 807 { 808 char b[sizeof (kmem64_obp_str) + (4 * 16)]; 809 tte_t tte; 810 811 PRM_DEBUG(kmem64_pabase); 812 PRM_DEBUG(kmem64_szc); 813 sfmmu_memtte(&tte, kmem64_pabase >> MMU_PAGESHIFT, 814 PROC_DATA | HAT_NOSYNC, kmem64_szc); 815 PRM_DEBUG(tte.ll); 816 (void) sprintf(b, kmem64_obp_str, 817 kmem64_base, kmem64_end, TTE_PAGEMASK(kmem64_szc), tte.ll); 818 ASSERT(strlen(b) < sizeof (b)); 819 prom_interpret(b, 0, 0, 0, 0, 0); 820 } 821 #endif /* !C_OBP */ 822 823 /* 824 * As OBP takes up some RAM when the system boots, pages will already be "lost" 825 * to the system and reflected in npages by the time we see it. 826 * 827 * We only want to allocate kernel structures in the 64-bit virtual address 828 * space on systems with enough RAM to make the overhead of keeping track of 829 * an extra kernel memory segment worthwhile. 830 * 831 * Since OBP has already performed its memory allocations by this point, if we 832 * have more than MINMOVE_RAM_MB MB of RAM left free, go ahead and map 833 * memory in the 64-bit virtual address space; otherwise keep allocations 834 * contiguous with we've mapped so far in the 32-bit virtual address space. 835 */ 836 #define MINMOVE_RAM_MB ((size_t)1900) 837 #define MB_TO_BYTES(mb) ((mb) * 1048576ul) 838 839 pgcnt_t tune_npages = (pgcnt_t) 840 (MB_TO_BYTES(MINMOVE_RAM_MB)/ (size_t)MMU_PAGESIZE); 841 842 #pragma weak page_set_colorequiv_arr_cpu 843 extern void page_set_colorequiv_arr_cpu(void); 844 845 static void 846 startup_memlist(void) 847 { 848 size_t alloc_sz; 849 size_t ctrs_sz; 850 caddr_t alloc_base; 851 caddr_t ctrs_base, ctrs_end; 852 caddr_t memspace; 853 caddr_t va; 854 int memblocks = 0; 855 struct memlist *cur; 856 size_t syslimit = (size_t)SYSLIMIT; 857 size_t sysbase = (size_t)SYSBASE; 858 int alloc_alignsize = ecache_alignsize; 859 int i; 860 extern void page_coloring_init(void); 861 extern void page_set_colorequiv_arr(void); 862 863 /* 864 * Initialize enough of the system to allow kmem_alloc to work by 865 * calling boot to allocate its memory until the time that 866 * kvm_init is completed. The page structs are allocated after 867 * rounding up end to the nearest page boundary; the memsegs are 868 * initialized and the space they use comes from the kernel heap. 869 * With appropriate initialization, they can be reallocated later 870 * to a size appropriate for the machine's configuration. 871 * 872 * At this point, memory is allocated for things that will never 873 * need to be freed, this used to be "valloced". This allows a 874 * savings as the pages don't need page structures to describe 875 * them because them will not be managed by the vm system. 876 */ 877 878 /* 879 * We're loaded by boot with the following configuration (as 880 * specified in the sun4u/conf/Mapfile): 881 * 882 * text: 4 MB chunk aligned on a 4MB boundary 883 * data & bss: 4 MB chunk aligned on a 4MB boundary 884 * 885 * These two chunks will eventually be mapped by 2 locked 4MB 886 * ttes and will represent the nucleus of the kernel. This gives 887 * us some free space that is already allocated, some or all of 888 * which is made available to kernel module text. 889 * 890 * The free space in the data-bss chunk is used for nucleus 891 * allocatable data structures and we reserve it using the 892 * nalloc_base and nalloc_end variables. This space is currently 893 * being used for hat data structures required for tlb miss 894 * handling operations. We align nalloc_base to a l2 cache 895 * linesize because this is the line size the hardware uses to 896 * maintain cache coherency. 897 * 256K is carved out for module data. 898 */ 899 900 nalloc_base = (caddr_t)roundup((uintptr_t)e_data, MMU_PAGESIZE); 901 moddata = nalloc_base; 902 e_moddata = nalloc_base + MODDATA; 903 nalloc_base = e_moddata; 904 905 nalloc_end = (caddr_t)roundup((uintptr_t)nalloc_base, MMU_PAGESIZE4M); 906 valloc_base = nalloc_base; 907 908 /* 909 * Calculate the start of the data segment. 910 */ 911 sdata = (caddr_t)((uintptr_t)e_data & MMU_PAGEMASK4M); 912 913 PRM_DEBUG(moddata); 914 PRM_DEBUG(nalloc_base); 915 PRM_DEBUG(nalloc_end); 916 PRM_DEBUG(sdata); 917 918 /* 919 * Remember any slop after e_text so we can give it to the modules. 920 */ 921 PRM_DEBUG(e_text); 922 modtext = (caddr_t)roundup((uintptr_t)e_text, MMU_PAGESIZE); 923 if (((uintptr_t)e_text & MMU_PAGEMASK4M) != (uintptr_t)s_text) 924 prom_panic("nucleus text overflow"); 925 modtext_sz = (caddr_t)roundup((uintptr_t)modtext, MMU_PAGESIZE4M) - 926 modtext; 927 PRM_DEBUG(modtext); 928 PRM_DEBUG(modtext_sz); 929 930 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 931 &boot_physavail, &boot_physavail_len, 932 &boot_virtavail, &boot_virtavail_len); 933 /* 934 * Remember what the physically available highest page is 935 * so that dumpsys works properly, and find out how much 936 * memory is installed. 937 */ 938 installed_top_size_memlist_array(boot_physinstalled, 939 boot_physinstalled_len, &physmax, &physinstalled); 940 PRM_DEBUG(physinstalled); 941 PRM_DEBUG(physmax); 942 943 /* Fill out memory nodes config structure */ 944 startup_build_mem_nodes(boot_physinstalled, boot_physinstalled_len); 945 946 /* 947 * Get the list of physically available memory to size 948 * the number of page structures needed. 949 */ 950 size_physavail(boot_physavail, boot_physavail_len, &npages, &memblocks); 951 /* 952 * This first snap shot of npages can represent the pages used 953 * by OBP's text and data approximately. This is used in the 954 * the calculation of the kernel size 955 */ 956 obp_pages = physinstalled - npages; 957 958 959 /* 960 * On small-memory systems (<MODTEXT_SM_SIZE MB, currently 256MB), the 961 * in-nucleus module text is capped to MODTEXT_SM_CAP bytes (currently 962 * 2MB) and any excess pages are put on physavail. The assumption is 963 * that small-memory systems will need more pages more than they'll 964 * need efficiently-mapped module texts. 965 */ 966 if ((physinstalled < mmu_btop(MODTEXT_SM_SIZE << 20)) && 967 modtext_sz > MODTEXT_SM_CAP) { 968 extra_etpg = mmu_btop(modtext_sz - MODTEXT_SM_CAP); 969 modtext_sz = MODTEXT_SM_CAP; 970 extra_etva = modtext + modtext_sz; 971 } 972 973 PRM_DEBUG(extra_etpg); 974 PRM_DEBUG(modtext_sz); 975 PRM_DEBUG(extra_etva); 976 977 /* 978 * Account for any pages after e_text and e_data. 979 */ 980 npages += extra_etpg; 981 npages += mmu_btopr(nalloc_end - nalloc_base); 982 PRM_DEBUG(npages); 983 984 /* 985 * npages is the maximum of available physical memory possible. 986 * (ie. it will never be more than this) 987 */ 988 989 /* 990 * initialize the nucleus memory allocator. 991 */ 992 ndata_alloc_init(&ndata, (uintptr_t)nalloc_base, (uintptr_t)nalloc_end); 993 994 /* 995 * Allocate mmu fault status area from the nucleus data area. 996 */ 997 if ((&ndata_alloc_mmfsa != NULL) && (ndata_alloc_mmfsa(&ndata) != 0)) 998 cmn_err(CE_PANIC, "no more nucleus memory after mfsa alloc"); 999 1000 /* 1001 * Allocate kernel TSBs from the nucleus data area. 1002 */ 1003 if (ndata_alloc_tsbs(&ndata, npages) != 0) 1004 cmn_err(CE_PANIC, "no more nucleus memory after tsbs alloc"); 1005 1006 /* 1007 * Allocate dmv dispatch table from the nucleus data area. 1008 */ 1009 if (ndata_alloc_dmv(&ndata) != 0) 1010 cmn_err(CE_PANIC, "no more nucleus memory after dmv alloc"); 1011 1012 1013 page_coloring_init(); 1014 1015 /* 1016 * Allocate page_freelists bin headers for memnode 0 from the 1017 * nucleus data area. 1018 */ 1019 if (ndata_alloc_page_freelists(&ndata, 0) != 0) 1020 cmn_err(CE_PANIC, 1021 "no more nucleus memory after page free lists alloc"); 1022 1023 if (kpm_enable) { 1024 kpm_init(); 1025 /* 1026 * kpm page space -- Update kpm_npages and make the 1027 * same assumption about fragmenting as it is done 1028 * for memseg_sz. 1029 */ 1030 kpm_npages_setup(memblocks + 4); 1031 } 1032 1033 /* 1034 * Allocate hat related structs from the nucleus data area. 1035 */ 1036 if (ndata_alloc_hat(&ndata, npages, kpm_npages) != 0) 1037 cmn_err(CE_PANIC, "no more nucleus memory after hat alloc"); 1038 1039 /* 1040 * We want to do the BOP_ALLOCs before the real allocation of page 1041 * structs in order to not have to allocate page structs for this 1042 * memory. We need to calculate a virtual address because we want 1043 * the page structs to come before other allocations in virtual address 1044 * space. This is so some (if not all) of page structs can actually 1045 * live in the nucleus. 1046 */ 1047 1048 /* 1049 * WARNING WARNING WARNING WARNING WARNING WARNING WARNING 1050 * 1051 * There are comments all over the SFMMU code warning of dire 1052 * consequences if the TSBs are moved out of 32-bit space. This 1053 * is largely because the asm code uses "sethi %hi(addr)"-type 1054 * instructions which will not provide the expected result if the 1055 * address is a 64-bit one. 1056 * 1057 * WARNING WARNING WARNING WARNING WARNING WARNING WARNING 1058 */ 1059 alloc_base = (caddr_t)roundup((uintptr_t)nalloc_end, MMU_PAGESIZE); 1060 alloc_base = sfmmu_ktsb_alloc(alloc_base); 1061 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, ecache_alignsize); 1062 PRM_DEBUG(alloc_base); 1063 1064 /* 1065 * Allocate IOMMU TSB array. We do this here so that the physical 1066 * memory gets deducted from the PROM's physical memory list. 1067 */ 1068 alloc_base = iommu_tsb_init(alloc_base); 1069 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1070 ecache_alignsize); 1071 PRM_DEBUG(alloc_base); 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 1092 if (econtig32 > (caddr_t)KERNEL_LIMIT32) 1093 cmn_err(CE_PANIC, "econtig32 too big"); 1094 1095 /* 1096 * To avoid memory allocation collisions in the 32-bit virtual address 1097 * space, make allocations from this point forward in 64-bit virtual 1098 * address space starting at syslimit and working up. 1099 * 1100 * All this is needed because on large memory systems, the default 1101 * Solaris allocations will collide with SYSBASE32, which is hard 1102 * coded to be at the virtual address 0x78000000. Therefore, on 64-bit 1103 * kernels, move the allocations to a location in the 64-bit virtual 1104 * address space space, allowing those structures to grow without 1105 * worry. 1106 * 1107 * On current CPUs we'll run out of physical memory address bits before 1108 * we need to worry about the allocations running into anything else in 1109 * VM or the virtual address holes on US-I and II, as there's currently 1110 * about 1 TB of addressable space before the US-I/II VA hole. 1111 */ 1112 kmem64_base = (caddr_t)syslimit; 1113 PRM_DEBUG(kmem64_base); 1114 1115 /* 1116 * Allocate addresses, but not physical memory. None of these locations 1117 * can be touched until physical memory is allocated below. 1118 */ 1119 alloc_base = kmem64_base; 1120 1121 /* 1122 * If KHME and/or UHME hash buckets won't fit in the nucleus, allocate 1123 * them here. 1124 */ 1125 if (khme_hash == NULL || uhme_hash == NULL) { 1126 /* 1127 * alloc_hme_buckets() will align alloc_base properly before 1128 * assigning the hash buckets, so we don't need to do it 1129 * before the call... 1130 */ 1131 alloc_base = alloc_hme_buckets(alloc_base, alloc_alignsize); 1132 1133 PRM_DEBUG(alloc_base); 1134 PRM_DEBUG(khme_hash); 1135 PRM_DEBUG(uhme_hash); 1136 } 1137 1138 /* 1139 * Allow for an early allocation of physically contiguous memory. 1140 */ 1141 alloc_base = contig_mem_prealloc(alloc_base, npages); 1142 1143 /* 1144 * Allocate the remaining page freelists. NUMA systems can 1145 * have lots of page freelists, one per node, which quickly 1146 * outgrow the amount of nucleus memory available. 1147 */ 1148 if (max_mem_nodes > 1) { 1149 int mnode; 1150 1151 for (mnode = 1; mnode < max_mem_nodes; mnode++) { 1152 alloc_base = alloc_page_freelists(mnode, alloc_base, 1153 ecache_alignsize); 1154 } 1155 PRM_DEBUG(alloc_base); 1156 } 1157 1158 if (!mml_table) { 1159 size_t mmltable_sz; 1160 1161 /* 1162 * We need to allocate the mml_table here because there 1163 * was not enough space within the nucleus. 1164 */ 1165 mmltable_sz = sizeof (kmutex_t) * mml_table_sz; 1166 alloc_sz = roundup(mmltable_sz, alloc_alignsize); 1167 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1168 alloc_alignsize); 1169 mml_table = (kmutex_t *)alloc_base; 1170 alloc_base += alloc_sz; 1171 PRM_DEBUG(mml_table); 1172 PRM_DEBUG(alloc_base); 1173 } 1174 1175 if (kpm_enable && !(kpmp_table || kpmp_stable)) { 1176 size_t kpmptable_sz; 1177 caddr_t table; 1178 1179 /* 1180 * We need to allocate either kpmp_table or kpmp_stable here 1181 * because there was not enough space within the nucleus. 1182 */ 1183 kpmptable_sz = (kpm_smallpages == 0) ? 1184 sizeof (kpm_hlk_t) * kpmp_table_sz : 1185 sizeof (kpm_shlk_t) * kpmp_stable_sz; 1186 1187 alloc_sz = roundup(kpmptable_sz, alloc_alignsize); 1188 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1189 alloc_alignsize); 1190 1191 table = alloc_base; 1192 1193 if (kpm_smallpages == 0) { 1194 kpmp_table = (kpm_hlk_t *)table; 1195 PRM_DEBUG(kpmp_table); 1196 } else { 1197 kpmp_stable = (kpm_shlk_t *)table; 1198 PRM_DEBUG(kpmp_stable); 1199 } 1200 1201 alloc_base += alloc_sz; 1202 PRM_DEBUG(alloc_base); 1203 } 1204 1205 if (&ecache_init_scrub_flush_area) { 1206 /* 1207 * Pass alloc_base directly, as the routine itself is 1208 * responsible for any special alignment requirements... 1209 */ 1210 alloc_base = ecache_init_scrub_flush_area(alloc_base); 1211 PRM_DEBUG(alloc_base); 1212 } 1213 1214 /* 1215 * Take the most current snapshot we can by calling mem-update. 1216 */ 1217 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1218 &boot_physavail, &boot_physavail_len, 1219 &boot_virtavail, &boot_virtavail_len); 1220 1221 /* 1222 * Reset npages and memblocks based on boot_physavail list. 1223 */ 1224 size_physavail(boot_physavail, boot_physavail_len, &npages, &memblocks); 1225 PRM_DEBUG(npages); 1226 1227 /* 1228 * Account for extra memory after e_text. 1229 */ 1230 npages += extra_etpg; 1231 1232 /* 1233 * Calculate the largest free memory chunk in the nucleus data area. 1234 * We need to figure out if page structs can fit in there or not. 1235 * We also make sure enough page structs get created for any physical 1236 * memory we might be returning to the system. 1237 */ 1238 ndata_remain_sz = ndata_maxsize(&ndata); 1239 PRM_DEBUG(ndata_remain_sz); 1240 1241 pp_sz = sizeof (struct page) * npages; 1242 1243 /* 1244 * Here's a nice bit of code based on somewhat recursive logic: 1245 * 1246 * If the page array would fit within the nucleus, we want to 1247 * add npages to cover any extra memory we may be returning back 1248 * to the system. 1249 * 1250 * HOWEVER, the page array is sized by calculating the size of 1251 * (struct page * npages), as are the pagehash table, ctrs and 1252 * memseg_list, so the very act of performing the calculation below may 1253 * in fact make the array large enough that it no longer fits in the 1254 * nucleus, meaning there would now be a much larger area of the 1255 * nucleus free that should really be added to npages, which would 1256 * make the page array that much larger, and so on. 1257 * 1258 * This also ignores the memory possibly used in the nucleus for the 1259 * the page hash, ctrs and memseg list and the fact that whether they 1260 * fit there or not varies with the npages calculation below, but we 1261 * don't even factor them into the equation at this point; perhaps we 1262 * should or perhaps we should just take the approach that the few 1263 * extra pages we could add via this calculation REALLY aren't worth 1264 * the hassle... 1265 */ 1266 if (ndata_remain_sz > pp_sz) { 1267 size_t spare = ndata_spare(&ndata, pp_sz, ecache_alignsize); 1268 1269 npages += mmu_btop(spare); 1270 1271 pp_sz = npages * sizeof (struct page); 1272 1273 pp_base = ndata_alloc(&ndata, pp_sz, ecache_alignsize); 1274 } 1275 1276 /* 1277 * If physmem is patched to be non-zero, use it instead of 1278 * the monitor value unless physmem is larger than the total 1279 * amount of memory on hand. 1280 */ 1281 if (physmem == 0 || physmem > npages) 1282 physmem = npages; 1283 1284 /* 1285 * If pp_base is NULL that means the routines above have determined 1286 * the page array will not fit in the nucleus; we'll have to 1287 * BOP_ALLOC() ourselves some space for them. 1288 */ 1289 if (pp_base == NULL) { 1290 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1291 alloc_alignsize); 1292 alloc_sz = roundup(pp_sz, alloc_alignsize); 1293 1294 pp_base = (struct page *)alloc_base; 1295 1296 alloc_base += alloc_sz; 1297 } 1298 1299 /* 1300 * The page structure hash table size is a power of 2 1301 * such that the average hash chain length is PAGE_HASHAVELEN. 1302 */ 1303 page_hashsz = npages / PAGE_HASHAVELEN; 1304 page_hashsz = 1 << highbit((ulong_t)page_hashsz); 1305 pagehash_sz = sizeof (struct page *) * page_hashsz; 1306 1307 /* 1308 * We want to TRY to fit the page structure hash table, 1309 * the page size free list counters, the memseg list and 1310 * and the kpm page space in the nucleus if possible. 1311 * 1312 * alloc_sz counts how much memory needs to be allocated by 1313 * BOP_ALLOC(). 1314 */ 1315 page_hash = ndata_alloc(&ndata, pagehash_sz, ecache_alignsize); 1316 1317 alloc_sz = (page_hash == NULL ? pagehash_sz : 0); 1318 1319 /* 1320 * Size up per page size free list counters. 1321 */ 1322 ctrs_sz = page_ctrs_sz(); 1323 ctrs_base = ndata_alloc(&ndata, ctrs_sz, ecache_alignsize); 1324 1325 if (ctrs_base == NULL) 1326 alloc_sz = roundup(alloc_sz, ecache_alignsize) + ctrs_sz; 1327 1328 /* 1329 * The memseg list is for the chunks of physical memory that 1330 * will be managed by the vm system. The number calculated is 1331 * a guess as boot may fragment it more when memory allocations 1332 * are made before kphysm_init(). Currently, there are two 1333 * allocations before then, so we assume each causes fragmen- 1334 * tation, and add a couple more for good measure. 1335 */ 1336 memseg_sz = sizeof (struct memseg) * (memblocks + 4); 1337 memseg_base = ndata_alloc(&ndata, memseg_sz, ecache_alignsize); 1338 1339 if (memseg_base == NULL) 1340 alloc_sz = roundup(alloc_sz, ecache_alignsize) + memseg_sz; 1341 1342 1343 if (kpm_enable) { 1344 /* 1345 * kpm page space -- Update kpm_npages and make the 1346 * same assumption about fragmenting as it is done 1347 * for memseg_sz above. 1348 */ 1349 kpm_npages_setup(memblocks + 4); 1350 kpm_pp_sz = (kpm_smallpages == 0) ? 1351 kpm_npages * sizeof (kpm_page_t): 1352 kpm_npages * sizeof (kpm_spage_t); 1353 1354 kpm_pp_base = (uintptr_t)ndata_alloc(&ndata, kpm_pp_sz, 1355 ecache_alignsize); 1356 1357 if (kpm_pp_base == NULL) 1358 alloc_sz = roundup(alloc_sz, ecache_alignsize) + 1359 kpm_pp_sz; 1360 } 1361 1362 if (alloc_sz > 0) { 1363 uintptr_t bop_base; 1364 1365 /* 1366 * We need extra memory allocated through BOP_ALLOC. 1367 */ 1368 alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, 1369 alloc_alignsize); 1370 1371 alloc_sz = roundup(alloc_sz, alloc_alignsize); 1372 1373 bop_base = (uintptr_t)alloc_base; 1374 1375 alloc_base += alloc_sz; 1376 1377 if (page_hash == NULL) { 1378 page_hash = (struct page **)bop_base; 1379 bop_base = roundup(bop_base + pagehash_sz, 1380 ecache_alignsize); 1381 } 1382 1383 if (ctrs_base == NULL) { 1384 ctrs_base = (caddr_t)bop_base; 1385 bop_base = roundup(bop_base + ctrs_sz, 1386 ecache_alignsize); 1387 } 1388 1389 if (memseg_base == NULL) { 1390 memseg_base = (struct memseg *)bop_base; 1391 bop_base = roundup(bop_base + memseg_sz, 1392 ecache_alignsize); 1393 } 1394 1395 if (kpm_enable && kpm_pp_base == NULL) { 1396 kpm_pp_base = (uintptr_t)bop_base; 1397 bop_base = roundup(bop_base + kpm_pp_sz, 1398 ecache_alignsize); 1399 } 1400 1401 ASSERT(bop_base <= (uintptr_t)alloc_base); 1402 } 1403 1404 PRM_DEBUG(page_hash); 1405 PRM_DEBUG(memseg_base); 1406 PRM_DEBUG(kpm_pp_base); 1407 PRM_DEBUG(kpm_pp_sz); 1408 PRM_DEBUG(pp_base); 1409 PRM_DEBUG(pp_sz); 1410 PRM_DEBUG(alloc_base); 1411 1412 #ifdef TRAPTRACE 1413 alloc_base = trap_trace_alloc(alloc_base); 1414 PRM_DEBUG(alloc_base); 1415 #endif /* TRAPTRACE */ 1416 1417 /* 1418 * In theory it's possible that kmem64 chunk is 0 sized 1419 * (on very small machines). Check for that. 1420 */ 1421 if (alloc_base == kmem64_base) { 1422 kmem64_base = NULL; 1423 kmem64_end = NULL; 1424 kmem64_aligned_end = NULL; 1425 goto kmem64_alloced; 1426 } 1427 1428 /* 1429 * Allocate kmem64 memory. 1430 * Round up to end of large page and overmap. 1431 * kmem64_end..kmem64_aligned_end is added to memory list for reuse 1432 */ 1433 kmem64_end = (caddr_t)roundup((uintptr_t)alloc_base, 1434 MMU_PAGESIZE); 1435 1436 /* 1437 * Make one large memory alloc after figuring out the 64-bit size. This 1438 * will enable use of the largest page size appropriate for the system 1439 * architecture. 1440 */ 1441 ASSERT(mmu_exported_pagesize_mask & (1 << TTE8K)); 1442 ASSERT(IS_P2ALIGNED(kmem64_base, TTEBYTES(max_bootlp_tteszc))); 1443 for (i = max_bootlp_tteszc; i >= TTE8K; i--) { 1444 size_t asize; 1445 #if !defined(C_OBP) 1446 unsigned long long pa; 1447 #endif /* !C_OBP */ 1448 1449 if ((mmu_exported_pagesize_mask & (1 << i)) == 0) 1450 continue; 1451 kmem64_alignsize = TTEBYTES(i); 1452 kmem64_szc = i; 1453 1454 /* limit page size for small memory */ 1455 if (mmu_btop(kmem64_alignsize) > (npages >> 2)) 1456 continue; 1457 1458 kmem64_aligned_end = (caddr_t)roundup((uintptr_t)kmem64_end, 1459 kmem64_alignsize); 1460 asize = kmem64_aligned_end - kmem64_base; 1461 #if !defined(C_OBP) 1462 if (prom_allocate_phys(asize, kmem64_alignsize, &pa) == 0) { 1463 if (prom_claim_virt(asize, kmem64_base) != 1464 (caddr_t)-1) { 1465 kmem64_pabase = pa; 1466 install_kmem64_tte(); 1467 break; 1468 } else { 1469 prom_free_phys(asize, pa); 1470 } 1471 } 1472 #else /* !C_OBP */ 1473 if ((caddr_t)BOP_ALLOC(bootops, kmem64_base, asize, 1474 kmem64_alignsize) == kmem64_base) { 1475 kmem64_pabase = va_to_pa(kmem64_base); 1476 break; 1477 } 1478 #endif /* !C_OBP */ 1479 if (i == TTE8K) { 1480 prom_panic("kmem64 allocation failure"); 1481 } 1482 } 1483 1484 PRM_DEBUG(kmem64_base); 1485 PRM_DEBUG(kmem64_end); 1486 PRM_DEBUG(kmem64_aligned_end); 1487 PRM_DEBUG(kmem64_alignsize); 1488 1489 /* 1490 * Now set pa using saved va from above. 1491 */ 1492 if (&ecache_init_scrub_flush_area) { 1493 (void) ecache_init_scrub_flush_area(NULL); 1494 } 1495 1496 kmem64_alloced: 1497 1498 /* 1499 * Initialize per page size free list counters. 1500 */ 1501 ctrs_end = page_ctrs_alloc(ctrs_base); 1502 ASSERT(ctrs_base + ctrs_sz >= ctrs_end); 1503 1504 /* 1505 * Allocate space for the interrupt vector table and also for the 1506 * reserved interrupt vector data structures. 1507 */ 1508 memspace = (caddr_t)BOP_ALLOC(bootops, (caddr_t)intr_vec_table, 1509 IVSIZE, MMU_PAGESIZE); 1510 if (memspace != (caddr_t)intr_vec_table) 1511 prom_panic("interrupt vector table allocation failure"); 1512 1513 /* 1514 * The memory lists from boot are allocated from the heap arena 1515 * so that later they can be freed and/or reallocated. 1516 */ 1517 if (BOP_GETPROP(bootops, "extent", &memlist_sz) == -1) 1518 prom_panic("could not retrieve property \"extent\""); 1519 1520 /* 1521 * Between now and when we finish copying in the memory lists, 1522 * allocations happen so the space gets fragmented and the 1523 * lists longer. Leave enough space for lists twice as long 1524 * as what boot says it has now; roundup to a pagesize. 1525 * Also add space for the final phys-avail copy in the fixup 1526 * routine. 1527 */ 1528 va = (caddr_t)(sysbase + PAGESIZE + PANICBUFSIZE + 1529 roundup(IVSIZE, MMU_PAGESIZE)); 1530 memlist_sz *= 4; 1531 memlist_sz = roundup(memlist_sz, MMU_PAGESIZE); 1532 memspace = (caddr_t)BOP_ALLOC(bootops, va, memlist_sz, BO_NO_ALIGN); 1533 if (memspace == NULL) 1534 halt("Boot allocation failed."); 1535 1536 memlist = (struct memlist *)memspace; 1537 memlist_end = (char *)memspace + memlist_sz; 1538 1539 PRM_DEBUG(memlist); 1540 PRM_DEBUG(memlist_end); 1541 PRM_DEBUG(sysbase); 1542 PRM_DEBUG(syslimit); 1543 1544 kernelheap_init((void *)sysbase, (void *)syslimit, 1545 (caddr_t)sysbase + PAGESIZE, NULL, NULL); 1546 1547 /* 1548 * Take the most current snapshot we can by calling mem-update. 1549 */ 1550 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1551 &boot_physavail, &boot_physavail_len, 1552 &boot_virtavail, &boot_virtavail_len); 1553 1554 /* 1555 * Remove the space used by BOP_ALLOC from the kernel heap 1556 * plus the area actually used by the OBP (if any) 1557 * ignoring virtual addresses in virt_avail, above syslimit. 1558 */ 1559 virt_avail = memlist; 1560 copy_memlist(boot_virtavail, boot_virtavail_len, &memlist); 1561 1562 for (cur = virt_avail; cur->next; cur = cur->next) { 1563 uint64_t range_base, range_size; 1564 1565 if ((range_base = cur->address + cur->size) < (uint64_t)sysbase) 1566 continue; 1567 if (range_base >= (uint64_t)syslimit) 1568 break; 1569 /* 1570 * Limit the range to end at syslimit. 1571 */ 1572 range_size = MIN(cur->next->address, 1573 (uint64_t)syslimit) - range_base; 1574 (void) vmem_xalloc(heap_arena, (size_t)range_size, PAGESIZE, 1575 0, 0, (void *)range_base, (void *)(range_base + range_size), 1576 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 1577 } 1578 1579 phys_avail = memlist; 1580 (void) copy_physavail(boot_physavail, boot_physavail_len, 1581 &memlist, 0, 0); 1582 1583 /* 1584 * Add any unused kmem64 memory from overmapped page 1585 * (Note: va_to_pa does not work for kmem64_end) 1586 */ 1587 if (kmem64_end < kmem64_aligned_end) { 1588 uint64_t overlap_size = kmem64_aligned_end - kmem64_end; 1589 uint64_t overlap_pa = kmem64_pabase + 1590 (kmem64_end - kmem64_base); 1591 1592 PRM_DEBUG(overlap_pa); 1593 PRM_DEBUG(overlap_size); 1594 memlist_add(overlap_pa, overlap_size, &memlist, &phys_avail); 1595 } 1596 1597 /* 1598 * Add any extra memory after e_text to the phys_avail list, as long 1599 * as there's at least a page to add. 1600 */ 1601 if (extra_etpg) 1602 memlist_add(va_to_pa(extra_etva), mmu_ptob(extra_etpg), 1603 &memlist, &phys_avail); 1604 1605 /* 1606 * Add any extra memory at the end of the ndata region if there's at 1607 * least a page to add. There might be a few more pages available in 1608 * the middle of the ndata region, but for now they are ignored. 1609 */ 1610 nalloc_base = ndata_extra_base(&ndata, MMU_PAGESIZE, nalloc_end); 1611 if (nalloc_base == NULL) 1612 nalloc_base = nalloc_end; 1613 ndata_remain_sz = nalloc_end - nalloc_base; 1614 1615 if (ndata_remain_sz >= MMU_PAGESIZE) 1616 memlist_add(va_to_pa(nalloc_base), 1617 (uint64_t)ndata_remain_sz, &memlist, &phys_avail); 1618 1619 PRM_DEBUG(memlist); 1620 PRM_DEBUG(memlist_sz); 1621 PRM_DEBUG(memspace); 1622 1623 if ((caddr_t)memlist > (memspace + memlist_sz)) 1624 prom_panic("memlist overflow"); 1625 1626 PRM_DEBUG(pp_base); 1627 PRM_DEBUG(memseg_base); 1628 PRM_DEBUG(npages); 1629 1630 /* 1631 * Initialize the page structures from the memory lists. 1632 */ 1633 kphysm_init(pp_base, memseg_base, npages, kpm_pp_base, kpm_npages); 1634 1635 availrmem_initial = availrmem = freemem; 1636 PRM_DEBUG(availrmem); 1637 1638 /* 1639 * Some of the locks depend on page_hashsz being set! 1640 * kmem_init() depends on this; so, keep it here. 1641 */ 1642 page_lock_init(); 1643 1644 /* 1645 * Initialize kernel memory allocator. 1646 */ 1647 kmem_init(); 1648 1649 /* 1650 * Factor in colorequiv to check additional 'equivalent' bins 1651 */ 1652 if (&page_set_colorequiv_arr_cpu != NULL) 1653 page_set_colorequiv_arr_cpu(); 1654 else 1655 page_set_colorequiv_arr(); 1656 1657 /* 1658 * Initialize bp_mapin(). 1659 */ 1660 bp_init(shm_alignment, HAT_STRICTORDER); 1661 1662 /* 1663 * Reserve space for panicbuf, intr_vec_table and reserved interrupt 1664 * vector data structures from the 32-bit heap. 1665 */ 1666 (void) vmem_xalloc(heap32_arena, PANICBUFSIZE, PAGESIZE, 0, 0, 1667 panicbuf, panicbuf + PANICBUFSIZE, 1668 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 1669 1670 (void) vmem_xalloc(heap32_arena, IVSIZE, PAGESIZE, 0, 0, 1671 intr_vec_table, (caddr_t)intr_vec_table + IVSIZE, 1672 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 1673 1674 mem_config_init(); 1675 } 1676 1677 static void 1678 startup_modules(void) 1679 { 1680 int proplen, nhblk1, nhblk8; 1681 size_t nhblksz; 1682 pgcnt_t pages_per_hblk; 1683 size_t hme8blk_sz, hme1blk_sz; 1684 1685 /* 1686 * Log any optional messages from the boot program 1687 */ 1688 proplen = (size_t)BOP_GETPROPLEN(bootops, "boot-message"); 1689 if (proplen > 0) { 1690 char *msg; 1691 size_t len = (size_t)proplen; 1692 1693 msg = kmem_zalloc(len, KM_SLEEP); 1694 (void) BOP_GETPROP(bootops, "boot-message", msg); 1695 cmn_err(CE_CONT, "?%s\n", msg); 1696 kmem_free(msg, len); 1697 } 1698 1699 /* 1700 * Let the platforms have a chance to change default 1701 * values before reading system file. 1702 */ 1703 if (&set_platform_defaults) 1704 set_platform_defaults(); 1705 1706 /* 1707 * Calculate default settings of system parameters based upon 1708 * maxusers, yet allow to be overridden via the /etc/system file. 1709 */ 1710 param_calc(0); 1711 1712 mod_setup(); 1713 1714 /* 1715 * If this is a positron, complain and halt. 1716 */ 1717 if (&iam_positron && iam_positron()) { 1718 cmn_err(CE_WARN, "This hardware platform is not supported" 1719 " by this release of Solaris.\n"); 1720 #ifdef DEBUG 1721 prom_enter_mon(); /* Type 'go' to resume */ 1722 cmn_err(CE_WARN, "Booting an unsupported platform.\n"); 1723 cmn_err(CE_WARN, "Booting with down-rev firmware.\n"); 1724 1725 #else /* DEBUG */ 1726 halt(0); 1727 #endif /* DEBUG */ 1728 } 1729 1730 /* 1731 * If we are running firmware that isn't 64-bit ready 1732 * then complain and halt. 1733 */ 1734 do_prom_version_check(); 1735 1736 /* 1737 * Initialize system parameters 1738 */ 1739 param_init(); 1740 1741 /* 1742 * maxmem is the amount of physical memory we're playing with. 1743 */ 1744 maxmem = physmem; 1745 1746 /* Set segkp limits. */ 1747 ncbase = kdi_segdebugbase; 1748 ncend = kdi_segdebugbase; 1749 1750 /* 1751 * Initialize the hat layer. 1752 */ 1753 hat_init(); 1754 1755 /* 1756 * Initialize segment management stuff. 1757 */ 1758 seg_init(); 1759 1760 /* 1761 * Create the va>tte handler, so the prom can understand 1762 * kernel translations. The handler is installed later, just 1763 * as we are about to take over the trap table from the prom. 1764 */ 1765 create_va_to_tte(); 1766 1767 /* 1768 * Load the forthdebugger (optional) 1769 */ 1770 forthdebug_init(); 1771 1772 /* 1773 * Create OBP node for console input callbacks 1774 * if it is needed. 1775 */ 1776 startup_create_io_node(); 1777 1778 if (modloadonly("fs", "specfs") == -1) 1779 halt("Can't load specfs"); 1780 1781 if (modloadonly("fs", "devfs") == -1) 1782 halt("Can't load devfs"); 1783 1784 if (modloadonly("misc", "swapgeneric") == -1) 1785 halt("Can't load swapgeneric"); 1786 1787 (void) modloadonly("sys", "lbl_edition"); 1788 1789 dispinit(); 1790 1791 /* 1792 * Infer meanings to the members of the idprom buffer. 1793 */ 1794 parse_idprom(); 1795 1796 /* Read cluster configuration data. */ 1797 clconf_init(); 1798 1799 setup_ddi(); 1800 1801 /* 1802 * Lets take this opportunity to load the root device. 1803 */ 1804 if (loadrootmodules() != 0) 1805 debug_enter("Can't load the root filesystem"); 1806 1807 /* 1808 * Load tod driver module for the tod part found on this system. 1809 * Recompute the cpu frequency/delays based on tod as tod part 1810 * tends to keep time more accurately. 1811 */ 1812 if (&load_tod_module) 1813 load_tod_module(); 1814 1815 /* 1816 * Allow platforms to load modules which might 1817 * be needed after bootops are gone. 1818 */ 1819 if (&load_platform_modules) 1820 load_platform_modules(); 1821 1822 setcpudelay(); 1823 1824 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1825 &boot_physavail, &boot_physavail_len, 1826 &boot_virtavail, &boot_virtavail_len); 1827 1828 /* 1829 * Calculation and allocation of hmeblks needed to remap 1830 * the memory allocated by PROM till now. 1831 * Overestimate the number of hblk1 elements by assuming 1832 * worst case of TTE64K mappings. 1833 * sfmmu_hblk_alloc will panic if this calculation is wrong. 1834 */ 1835 bop_alloc_pages = btopr(kmem64_end - kmem64_base); 1836 pages_per_hblk = btop(HMEBLK_SPAN(TTE64K)); 1837 bop_alloc_pages = roundup(bop_alloc_pages, pages_per_hblk); 1838 nhblk1 = bop_alloc_pages / pages_per_hblk + hblk1_min; 1839 1840 bop_alloc_pages = size_virtalloc(boot_virtavail, boot_virtavail_len); 1841 1842 /* sfmmu_init_nucleus_hblks expects properly aligned data structures */ 1843 hme8blk_sz = roundup(HME8BLK_SZ, sizeof (int64_t)); 1844 hme1blk_sz = roundup(HME1BLK_SZ, sizeof (int64_t)); 1845 1846 bop_alloc_pages += btopr(nhblk1 * hme1blk_sz); 1847 1848 pages_per_hblk = btop(HMEBLK_SPAN(TTE8K)); 1849 nhblk8 = 0; 1850 while (bop_alloc_pages > 1) { 1851 bop_alloc_pages = roundup(bop_alloc_pages, pages_per_hblk); 1852 nhblk8 += bop_alloc_pages /= pages_per_hblk; 1853 bop_alloc_pages *= hme8blk_sz; 1854 bop_alloc_pages = btopr(bop_alloc_pages); 1855 } 1856 nhblk8 += 2; 1857 1858 /* 1859 * Since hblk8's can hold up to 64k of mappings aligned on a 64k 1860 * boundary, the number of hblk8's needed to map the entries in the 1861 * boot_virtavail list needs to be adjusted to take this into 1862 * consideration. Thus, we need to add additional hblk8's since it 1863 * is possible that an hblk8 will not have all 8 slots used due to 1864 * alignment constraints. Since there were boot_virtavail_len entries 1865 * in that list, we need to add that many hblk8's to the number 1866 * already calculated to make sure we don't underestimate. 1867 */ 1868 nhblk8 += boot_virtavail_len; 1869 nhblksz = nhblk8 * hme8blk_sz + nhblk1 * hme1blk_sz; 1870 1871 /* Allocate in pagesize chunks */ 1872 nhblksz = roundup(nhblksz, MMU_PAGESIZE); 1873 hblk_base = kmem_zalloc(nhblksz, KM_SLEEP); 1874 sfmmu_init_nucleus_hblks(hblk_base, nhblksz, nhblk8, nhblk1); 1875 } 1876 1877 static void 1878 startup_bop_gone(void) 1879 { 1880 extern int bop_io_quiesced; 1881 1882 /* 1883 * Destroy the MD initialized at startup 1884 * The startup initializes the MD framework 1885 * using prom and BOP alloc free it now. 1886 */ 1887 mach_descrip_startup_fini(); 1888 1889 /* 1890 * Call back into boot and release boots resources. 1891 */ 1892 BOP_QUIESCE_IO(bootops); 1893 bop_io_quiesced = 1; 1894 1895 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1896 &boot_physavail, &boot_physavail_len, 1897 &boot_virtavail, &boot_virtavail_len); 1898 /* 1899 * Copy physinstalled list into kernel space. 1900 */ 1901 phys_install = memlist; 1902 copy_memlist(boot_physinstalled, boot_physinstalled_len, &memlist); 1903 1904 /* 1905 * setup physically contiguous area twice as large as the ecache. 1906 * this is used while doing displacement flush of ecaches 1907 */ 1908 if (&ecache_flush_address) { 1909 ecache_flushaddr = ecache_flush_address(); 1910 if (ecache_flushaddr == (uint64_t)-1) { 1911 cmn_err(CE_PANIC, 1912 "startup: no memory to set ecache_flushaddr"); 1913 } 1914 } 1915 1916 /* 1917 * Virtual available next. 1918 */ 1919 ASSERT(virt_avail != NULL); 1920 memlist_free_list(virt_avail); 1921 virt_avail = memlist; 1922 copy_memlist(boot_virtavail, boot_virtavail_len, &memlist); 1923 1924 /* 1925 * Last chance to ask our booter questions .. 1926 */ 1927 } 1928 1929 1930 /* 1931 * startup_fixup_physavail - called from mach_sfmmu.c after the final 1932 * allocations have been performed. We can't call it in startup_bop_gone 1933 * since later operations can cause obp to allocate more memory. 1934 */ 1935 void 1936 startup_fixup_physavail(void) 1937 { 1938 struct memlist *cur; 1939 size_t kmem64_overmap_size = kmem64_aligned_end - kmem64_end; 1940 1941 PRM_DEBUG(kmem64_overmap_size); 1942 1943 /* 1944 * take the most current snapshot we can by calling mem-update 1945 */ 1946 copy_boot_memlists(&boot_physinstalled, &boot_physinstalled_len, 1947 &boot_physavail, &boot_physavail_len, 1948 &boot_virtavail, &boot_virtavail_len); 1949 1950 /* 1951 * Copy phys_avail list, again. 1952 * Both the kernel/boot and the prom have been allocating 1953 * from the original list we copied earlier. 1954 */ 1955 cur = memlist; 1956 (void) copy_physavail(boot_physavail, boot_physavail_len, 1957 &memlist, 0, 0); 1958 1959 /* 1960 * Add any unused kmem64 memory from overmapped page 1961 * (Note: va_to_pa does not work for kmem64_end) 1962 */ 1963 if (kmem64_overmap_size) { 1964 memlist_add(kmem64_pabase + (kmem64_end - kmem64_base), 1965 kmem64_overmap_size, 1966 &memlist, &cur); 1967 } 1968 1969 /* 1970 * Add any extra memory after e_text we added to the phys_avail list 1971 * back to the old list. 1972 */ 1973 if (extra_etpg) 1974 memlist_add(va_to_pa(extra_etva), mmu_ptob(extra_etpg), 1975 &memlist, &cur); 1976 if (ndata_remain_sz >= MMU_PAGESIZE) 1977 memlist_add(va_to_pa(nalloc_base), 1978 (uint64_t)ndata_remain_sz, &memlist, &cur); 1979 1980 /* 1981 * There isn't any bounds checking on the memlist area 1982 * so ensure it hasn't overgrown. 1983 */ 1984 if ((caddr_t)memlist > (caddr_t)memlist_end) 1985 cmn_err(CE_PANIC, "startup: memlist size exceeded"); 1986 1987 /* 1988 * The kernel removes the pages that were allocated for it from 1989 * the freelist, but we now have to find any -extra- pages that 1990 * the prom has allocated for it's own book-keeping, and remove 1991 * them from the freelist too. sigh. 1992 */ 1993 fix_prom_pages(phys_avail, cur); 1994 1995 ASSERT(phys_avail != NULL); 1996 memlist_free_list(phys_avail); 1997 phys_avail = cur; 1998 1999 /* 2000 * We're done with boot. Just after this point in time, boot 2001 * gets unmapped, so we can no longer rely on its services. 2002 * Zero the bootops to indicate this fact. 2003 */ 2004 bootops = (struct bootops *)NULL; 2005 BOOTOPS_GONE(); 2006 } 2007 2008 static void 2009 startup_vm(void) 2010 { 2011 size_t i; 2012 struct segmap_crargs a; 2013 struct segkpm_crargs b; 2014 2015 uint64_t avmem; 2016 caddr_t va; 2017 pgcnt_t max_phys_segkp; 2018 int mnode; 2019 2020 extern int use_brk_lpg, use_stk_lpg; 2021 2022 /* 2023 * get prom's mappings, create hments for them and switch 2024 * to the kernel context. 2025 */ 2026 hat_kern_setup(); 2027 2028 /* 2029 * Take over trap table 2030 */ 2031 setup_trap_table(); 2032 2033 /* 2034 * Install the va>tte handler, so that the prom can handle 2035 * misses and understand the kernel table layout in case 2036 * we need call into the prom. 2037 */ 2038 install_va_to_tte(); 2039 2040 /* 2041 * Set a flag to indicate that the tba has been taken over. 2042 */ 2043 tba_taken_over = 1; 2044 2045 /* initialize MMU primary context register */ 2046 mmu_init_kcontext(); 2047 2048 /* 2049 * The boot cpu can now take interrupts, x-calls, x-traps 2050 */ 2051 CPUSET_ADD(cpu_ready_set, CPU->cpu_id); 2052 CPU->cpu_flags |= (CPU_READY | CPU_ENABLE | CPU_EXISTS); 2053 2054 /* 2055 * Set a flag to tell write_scb_int() that it can access V_TBR_WR_ADDR. 2056 */ 2057 tbr_wr_addr_inited = 1; 2058 2059 /* 2060 * Initialize VM system, and map kernel address space. 2061 */ 2062 kvm_init(); 2063 2064 /* 2065 * XXX4U: previously, we initialized and turned on 2066 * the caches at this point. But of course we have 2067 * nothing to do, as the prom has already done this 2068 * for us -- main memory must be E$able at all times. 2069 */ 2070 2071 /* 2072 * If the following is true, someone has patched 2073 * phsymem to be less than the number of pages that 2074 * the system actually has. Remove pages until system 2075 * memory is limited to the requested amount. Since we 2076 * have allocated page structures for all pages, we 2077 * correct the amount of memory we want to remove 2078 * by the size of the memory used to hold page structures 2079 * for the non-used pages. 2080 */ 2081 if (physmem < npages) { 2082 pgcnt_t diff, off; 2083 struct page *pp; 2084 struct seg kseg; 2085 2086 cmn_err(CE_WARN, "limiting physmem to %ld pages", physmem); 2087 2088 off = 0; 2089 diff = npages - physmem; 2090 diff -= mmu_btopr(diff * sizeof (struct page)); 2091 kseg.s_as = &kas; 2092 while (diff--) { 2093 pp = page_create_va(&unused_pages_vp, (offset_t)off, 2094 MMU_PAGESIZE, PG_WAIT | PG_EXCL, 2095 &kseg, (caddr_t)off); 2096 if (pp == NULL) 2097 cmn_err(CE_PANIC, "limited physmem too much!"); 2098 page_io_unlock(pp); 2099 page_downgrade(pp); 2100 availrmem--; 2101 off += MMU_PAGESIZE; 2102 } 2103 } 2104 2105 /* 2106 * When printing memory, show the total as physmem less 2107 * that stolen by a debugger. 2108 */ 2109 cmn_err(CE_CONT, "?mem = %ldK (0x%lx000)\n", 2110 (ulong_t)(physinstalled) << (PAGESHIFT - 10), 2111 (ulong_t)(physinstalled) << (PAGESHIFT - 12)); 2112 2113 avmem = (uint64_t)freemem << PAGESHIFT; 2114 cmn_err(CE_CONT, "?avail mem = %lld\n", (unsigned long long)avmem); 2115 2116 /* 2117 * For small memory systems disable automatic large pages. 2118 */ 2119 if (physmem < privm_lpg_min_physmem) { 2120 use_brk_lpg = 0; 2121 use_stk_lpg = 0; 2122 } 2123 2124 /* 2125 * Perform platform specific freelist processing 2126 */ 2127 if (&plat_freelist_process) { 2128 for (mnode = 0; mnode < max_mem_nodes; mnode++) 2129 if (mem_node_config[mnode].exists) 2130 plat_freelist_process(mnode); 2131 } 2132 2133 /* 2134 * Initialize the segkp segment type. We position it 2135 * after the configured tables and buffers (whose end 2136 * is given by econtig) and before V_WKBASE_ADDR. 2137 * Also in this area is segkmap (size SEGMAPSIZE). 2138 */ 2139 2140 /* XXX - cache alignment? */ 2141 va = (caddr_t)SEGKPBASE; 2142 ASSERT(((uintptr_t)va & PAGEOFFSET) == 0); 2143 2144 max_phys_segkp = (physmem * 2); 2145 2146 if (segkpsize < btop(SEGKPMINSIZE) || segkpsize > btop(SEGKPMAXSIZE)) { 2147 segkpsize = btop(SEGKPDEFSIZE); 2148 cmn_err(CE_WARN, "Illegal value for segkpsize. " 2149 "segkpsize has been reset to %ld pages", segkpsize); 2150 } 2151 2152 i = ptob(MIN(segkpsize, max_phys_segkp)); 2153 2154 rw_enter(&kas.a_lock, RW_WRITER); 2155 if (seg_attach(&kas, va, i, segkp) < 0) 2156 cmn_err(CE_PANIC, "startup: cannot attach segkp"); 2157 if (segkp_create(segkp) != 0) 2158 cmn_err(CE_PANIC, "startup: segkp_create failed"); 2159 rw_exit(&kas.a_lock); 2160 2161 /* 2162 * kpm segment 2163 */ 2164 segmap_kpm = kpm_enable && 2165 segmap_kpm && PAGESIZE == MAXBSIZE; 2166 2167 if (kpm_enable) { 2168 rw_enter(&kas.a_lock, RW_WRITER); 2169 2170 /* 2171 * The segkpm virtual range range is larger than the 2172 * actual physical memory size and also covers gaps in 2173 * the physical address range for the following reasons: 2174 * . keep conversion between segkpm and physical addresses 2175 * simple, cheap and unambiguous. 2176 * . avoid extension/shrink of the the segkpm in case of DR. 2177 * . avoid complexity for handling of virtual addressed 2178 * caches, segkpm and the regular mapping scheme must be 2179 * kept in sync wrt. the virtual color of mapped pages. 2180 * Any accesses to virtual segkpm ranges not backed by 2181 * physical memory will fall through the memseg pfn hash 2182 * and will be handled in segkpm_fault. 2183 * Additional kpm_size spaces needed for vac alias prevention. 2184 */ 2185 if (seg_attach(&kas, kpm_vbase, kpm_size * vac_colors, 2186 segkpm) < 0) 2187 cmn_err(CE_PANIC, "cannot attach segkpm"); 2188 2189 b.prot = PROT_READ | PROT_WRITE; 2190 b.nvcolors = shm_alignment >> MMU_PAGESHIFT; 2191 2192 if (segkpm_create(segkpm, (caddr_t)&b) != 0) 2193 panic("segkpm_create segkpm"); 2194 2195 rw_exit(&kas.a_lock); 2196 2197 mach_kpm_init(); 2198 } 2199 2200 if (!segzio_fromheap) { 2201 size_t size; 2202 size_t physmem_b = mmu_ptob(physmem); 2203 2204 /* size is in bytes, segziosize is in pages */ 2205 if (segziosize == 0) { 2206 size = physmem_b; 2207 } else { 2208 size = mmu_ptob(segziosize); 2209 } 2210 2211 if (size < SEGZIOMINSIZE) { 2212 size = SEGZIOMINSIZE; 2213 } else if (size > SEGZIOMAXSIZE) { 2214 size = SEGZIOMAXSIZE; 2215 /* 2216 * On 64-bit x86, we only have 2TB of KVA. This exists 2217 * for parity with x86. 2218 * 2219 * SEGZIOMAXSIZE is capped at 512gb so that segzio 2220 * doesn't consume all of KVA. However, if we have a 2221 * system that has more thant 512gb of physical memory, 2222 * we can actually consume about half of the difference 2223 * between 512gb and the rest of the available physical 2224 * memory. 2225 */ 2226 if (physmem_b > SEGZIOMAXSIZE) { 2227 size += (physmem_b - SEGZIOMAXSIZE) / 2; 2228 } 2229 } 2230 segziosize = mmu_btop(roundup(size, MMU_PAGESIZE)); 2231 /* put the base of the ZIO segment after the kpm segment */ 2232 segzio_base = kpm_vbase + (kpm_size * vac_colors); 2233 PRM_DEBUG(segziosize); 2234 PRM_DEBUG(segzio_base); 2235 2236 /* 2237 * On some platforms, kvm_init is called after the kpm 2238 * sizes have been determined. On SPARC, kvm_init is called 2239 * before, so we have to attach the kzioseg after kvm is 2240 * initialized, otherwise we'll try to allocate from the boot 2241 * area since the kernel heap hasn't yet been configured. 2242 */ 2243 rw_enter(&kas.a_lock, RW_WRITER); 2244 2245 (void) seg_attach(&kas, segzio_base, mmu_ptob(segziosize), 2246 &kzioseg); 2247 (void) segkmem_zio_create(&kzioseg); 2248 2249 /* create zio area covering new segment */ 2250 segkmem_zio_init(segzio_base, mmu_ptob(segziosize)); 2251 2252 rw_exit(&kas.a_lock); 2253 } 2254 2255 2256 /* 2257 * Now create generic mapping segment. This mapping 2258 * goes SEGMAPSIZE beyond SEGMAPBASE. But if the total 2259 * virtual address is greater than the amount of free 2260 * memory that is available, then we trim back the 2261 * segment size to that amount 2262 */ 2263 va = (caddr_t)SEGMAPBASE; 2264 2265 /* 2266 * 1201049: segkmap base address must be MAXBSIZE aligned 2267 */ 2268 ASSERT(((uintptr_t)va & MAXBOFFSET) == 0); 2269 2270 /* 2271 * Set size of segmap to percentage of freemem at boot, 2272 * but stay within the allowable range 2273 * Note we take percentage before converting from pages 2274 * to bytes to avoid an overflow on 32-bit kernels. 2275 */ 2276 i = mmu_ptob((freemem * segmap_percent) / 100); 2277 2278 if (i < MINMAPSIZE) 2279 i = MINMAPSIZE; 2280 2281 if (i > MIN(SEGMAPSIZE, mmu_ptob(freemem))) 2282 i = MIN(SEGMAPSIZE, mmu_ptob(freemem)); 2283 2284 i &= MAXBMASK; /* 1201049: segkmap size must be MAXBSIZE aligned */ 2285 2286 rw_enter(&kas.a_lock, RW_WRITER); 2287 if (seg_attach(&kas, va, i, segkmap) < 0) 2288 cmn_err(CE_PANIC, "cannot attach segkmap"); 2289 2290 a.prot = PROT_READ | PROT_WRITE; 2291 a.shmsize = shm_alignment; 2292 a.nfreelist = 0; /* use segmap driver defaults */ 2293 2294 if (segmap_create(segkmap, (caddr_t)&a) != 0) 2295 panic("segmap_create segkmap"); 2296 rw_exit(&kas.a_lock); 2297 2298 segdev_init(); 2299 } 2300 2301 static void 2302 startup_end(void) 2303 { 2304 if ((caddr_t)memlist > (caddr_t)memlist_end) 2305 panic("memlist overflow 2"); 2306 memlist_free_block((caddr_t)memlist, 2307 ((caddr_t)memlist_end - (caddr_t)memlist)); 2308 memlist = NULL; 2309 2310 /* enable page_relocation since OBP is now done */ 2311 page_relocate_ready = 1; 2312 2313 /* 2314 * Perform tasks that get done after most of the VM 2315 * initialization has been done but before the clock 2316 * and other devices get started. 2317 */ 2318 kern_setup1(); 2319 2320 /* 2321 * Intialize the VM arenas for allocating physically 2322 * contiguus memory chunk for interrupt queues snd 2323 * allocate/register boot cpu's queues, if any and 2324 * allocate dump buffer for sun4v systems to store 2325 * extra crash information during crash dump 2326 */ 2327 contig_mem_init(); 2328 mach_descrip_init(); 2329 2330 if (cpu_intrq_setup(CPU)) { 2331 cmn_err(CE_PANIC, "cpu%d: setup failed", CPU->cpu_id); 2332 } 2333 cpu_intrq_register(CPU); 2334 mach_htraptrace_setup(CPU->cpu_id); 2335 mach_htraptrace_configure(CPU->cpu_id); 2336 mach_dump_buffer_init(); 2337 2338 /* 2339 * Initialize interrupt related stuff 2340 */ 2341 cpu_intr_alloc(CPU, NINTR_THREADS); 2342 2343 (void) splzs(); /* allow hi clock ints but not zs */ 2344 2345 /* 2346 * Initialize errors. 2347 */ 2348 error_init(); 2349 2350 /* 2351 * Note that we may have already used kernel bcopy before this 2352 * point - but if you really care about this, adb the use_hw_* 2353 * variables to 0 before rebooting. 2354 */ 2355 mach_hw_copy_limit(); 2356 2357 /* 2358 * Install the "real" preemption guards before DDI services 2359 * are available. 2360 */ 2361 (void) prom_set_preprom(kern_preprom); 2362 (void) prom_set_postprom(kern_postprom); 2363 CPU->cpu_m.mutex_ready = 1; 2364 2365 /* 2366 * Initialize segnf (kernel support for non-faulting loads). 2367 */ 2368 segnf_init(); 2369 2370 /* 2371 * Configure the root devinfo node. 2372 */ 2373 configure(); /* set up devices */ 2374 mach_cpu_halt_idle(); 2375 } 2376 2377 2378 void 2379 post_startup(void) 2380 { 2381 #ifdef PTL1_PANIC_DEBUG 2382 extern void init_ptl1_thread(void); 2383 #endif /* PTL1_PANIC_DEBUG */ 2384 extern void abort_sequence_init(void); 2385 2386 /* 2387 * Set the system wide, processor-specific flags to be passed 2388 * to userland via the aux vector for performance hints and 2389 * instruction set extensions. 2390 */ 2391 bind_hwcap(); 2392 2393 /* 2394 * Startup memory scrubber (if any) 2395 */ 2396 mach_memscrub(); 2397 2398 /* 2399 * Allocate soft interrupt to handle abort sequence. 2400 */ 2401 abort_sequence_init(); 2402 2403 /* 2404 * Configure the rest of the system. 2405 * Perform forceloading tasks for /etc/system. 2406 */ 2407 (void) mod_sysctl(SYS_FORCELOAD, NULL); 2408 /* 2409 * ON4.0: Force /proc module in until clock interrupt handle fixed 2410 * ON4.0: This must be fixed or restated in /etc/systems. 2411 */ 2412 (void) modload("fs", "procfs"); 2413 2414 /* load machine class specific drivers */ 2415 load_mach_drivers(); 2416 2417 /* load platform specific drivers */ 2418 if (&load_platform_drivers) 2419 load_platform_drivers(); 2420 2421 /* load vis simulation module, if we are running w/fpu off */ 2422 if (!fpu_exists) { 2423 if (modload("misc", "vis") == -1) 2424 halt("Can't load vis"); 2425 } 2426 2427 mach_fpras(); 2428 2429 maxmem = freemem; 2430 2431 #ifdef PTL1_PANIC_DEBUG 2432 init_ptl1_thread(); 2433 #endif /* PTL1_PANIC_DEBUG */ 2434 2435 if (&cif_init) 2436 cif_init(); 2437 } 2438 2439 #ifdef PTL1_PANIC_DEBUG 2440 int ptl1_panic_test = 0; 2441 int ptl1_panic_xc_one_test = 0; 2442 int ptl1_panic_xc_all_test = 0; 2443 int ptl1_panic_xt_one_test = 0; 2444 int ptl1_panic_xt_all_test = 0; 2445 kthread_id_t ptl1_thread_p = NULL; 2446 kcondvar_t ptl1_cv; 2447 kmutex_t ptl1_mutex; 2448 int ptl1_recurse_count_threshold = 0x40; 2449 int ptl1_recurse_trap_threshold = 0x3d; 2450 extern void ptl1_recurse(int, int); 2451 extern void ptl1_panic_xt(int, int); 2452 2453 /* 2454 * Called once per second by timeout() to wake up 2455 * the ptl1_panic thread to see if it should cause 2456 * a trap to the ptl1_panic() code. 2457 */ 2458 /* ARGSUSED */ 2459 static void 2460 ptl1_wakeup(void *arg) 2461 { 2462 mutex_enter(&ptl1_mutex); 2463 cv_signal(&ptl1_cv); 2464 mutex_exit(&ptl1_mutex); 2465 } 2466 2467 /* 2468 * ptl1_panic cross call function: 2469 * Needed because xc_one() and xc_some() can pass 2470 * 64 bit args but ptl1_recurse() expects ints. 2471 */ 2472 static void 2473 ptl1_panic_xc(void) 2474 { 2475 ptl1_recurse(ptl1_recurse_count_threshold, 2476 ptl1_recurse_trap_threshold); 2477 } 2478 2479 /* 2480 * The ptl1 thread waits for a global flag to be set 2481 * and uses the recurse thresholds to set the stack depth 2482 * to cause a ptl1_panic() directly via a call to ptl1_recurse 2483 * or indirectly via the cross call and cross trap functions. 2484 * 2485 * This is useful testing stack overflows and normal 2486 * ptl1_panic() states with a know stack frame. 2487 * 2488 * ptl1_recurse() is an asm function in ptl1_panic.s that 2489 * sets the {In, Local, Out, and Global} registers to a 2490 * know state on the stack and just prior to causing a 2491 * test ptl1_panic trap. 2492 */ 2493 static void 2494 ptl1_thread(void) 2495 { 2496 mutex_enter(&ptl1_mutex); 2497 while (ptl1_thread_p) { 2498 cpuset_t other_cpus; 2499 int cpu_id; 2500 int my_cpu_id; 2501 int target_cpu_id; 2502 int target_found; 2503 2504 if (ptl1_panic_test) { 2505 ptl1_recurse(ptl1_recurse_count_threshold, 2506 ptl1_recurse_trap_threshold); 2507 } 2508 2509 /* 2510 * Find potential targets for x-call and x-trap, 2511 * if any exist while preempt is disabled we 2512 * start a ptl1_panic if requested via a 2513 * globals. 2514 */ 2515 kpreempt_disable(); 2516 my_cpu_id = CPU->cpu_id; 2517 other_cpus = cpu_ready_set; 2518 CPUSET_DEL(other_cpus, CPU->cpu_id); 2519 target_found = 0; 2520 if (!CPUSET_ISNULL(other_cpus)) { 2521 /* 2522 * Pick the first one 2523 */ 2524 for (cpu_id = 0; cpu_id < NCPU; cpu_id++) { 2525 if (cpu_id == my_cpu_id) 2526 continue; 2527 2528 if (CPU_XCALL_READY(cpu_id)) { 2529 target_cpu_id = cpu_id; 2530 target_found = 1; 2531 break; 2532 } 2533 } 2534 ASSERT(target_found); 2535 2536 if (ptl1_panic_xc_one_test) { 2537 xc_one(target_cpu_id, 2538 (xcfunc_t *)ptl1_panic_xc, 0, 0); 2539 } 2540 if (ptl1_panic_xc_all_test) { 2541 xc_some(other_cpus, 2542 (xcfunc_t *)ptl1_panic_xc, 0, 0); 2543 } 2544 if (ptl1_panic_xt_one_test) { 2545 xt_one(target_cpu_id, 2546 (xcfunc_t *)ptl1_panic_xt, 0, 0); 2547 } 2548 if (ptl1_panic_xt_all_test) { 2549 xt_some(other_cpus, 2550 (xcfunc_t *)ptl1_panic_xt, 0, 0); 2551 } 2552 } 2553 kpreempt_enable(); 2554 (void) timeout(ptl1_wakeup, NULL, hz); 2555 (void) cv_wait(&ptl1_cv, &ptl1_mutex); 2556 } 2557 mutex_exit(&ptl1_mutex); 2558 } 2559 2560 /* 2561 * Called during early startup to create the ptl1_thread 2562 */ 2563 void 2564 init_ptl1_thread(void) 2565 { 2566 ptl1_thread_p = thread_create(NULL, 0, ptl1_thread, NULL, 0, 2567 &p0, TS_RUN, 0); 2568 } 2569 #endif /* PTL1_PANIC_DEBUG */ 2570 2571 2572 /* 2573 * Add to a memory list. 2574 * start = start of new memory segment 2575 * len = length of new memory segment in bytes 2576 * memlistp = pointer to array of available memory segment structures 2577 * curmemlistp = memory list to which to add segment. 2578 */ 2579 static void 2580 memlist_add(uint64_t start, uint64_t len, struct memlist **memlistp, 2581 struct memlist **curmemlistp) 2582 { 2583 struct memlist *new; 2584 2585 new = *memlistp; 2586 new->address = start; 2587 new->size = len; 2588 *memlistp = new + 1; 2589 2590 memlist_insert(new, curmemlistp); 2591 } 2592 2593 /* 2594 * In the case of architectures that support dynamic addition of 2595 * memory at run-time there are two cases where memsegs need to 2596 * be initialized and added to the memseg list. 2597 * 1) memsegs that are constructed at startup. 2598 * 2) memsegs that are constructed at run-time on 2599 * hot-plug capable architectures. 2600 * This code was originally part of the function kphysm_init(). 2601 */ 2602 2603 static void 2604 memseg_list_add(struct memseg *memsegp) 2605 { 2606 struct memseg **prev_memsegp; 2607 pgcnt_t num; 2608 2609 /* insert in memseg list, decreasing number of pages order */ 2610 2611 num = MSEG_NPAGES(memsegp); 2612 2613 for (prev_memsegp = &memsegs; *prev_memsegp; 2614 prev_memsegp = &((*prev_memsegp)->next)) { 2615 if (num > MSEG_NPAGES(*prev_memsegp)) 2616 break; 2617 } 2618 2619 memsegp->next = *prev_memsegp; 2620 *prev_memsegp = memsegp; 2621 2622 if (kpm_enable) { 2623 memsegp->nextpa = (memsegp->next) ? 2624 va_to_pa(memsegp->next) : MSEG_NULLPTR_PA; 2625 2626 if (prev_memsegp != &memsegs) { 2627 struct memseg *msp; 2628 msp = (struct memseg *)((caddr_t)prev_memsegp - 2629 offsetof(struct memseg, next)); 2630 msp->nextpa = va_to_pa(memsegp); 2631 } else { 2632 memsegspa = va_to_pa(memsegs); 2633 } 2634 } 2635 } 2636 2637 /* 2638 * PSM add_physmem_cb(). US-II and newer processors have some 2639 * flavor of the prefetch capability implemented. We exploit 2640 * this capability for optimum performance. 2641 */ 2642 #define PREFETCH_BYTES 64 2643 2644 void 2645 add_physmem_cb(page_t *pp, pfn_t pnum) 2646 { 2647 extern void prefetch_page_w(void *); 2648 2649 pp->p_pagenum = pnum; 2650 2651 /* 2652 * Prefetch one more page_t into E$. To prevent future 2653 * mishaps with the sizeof(page_t) changing on us, we 2654 * catch this on debug kernels if we can't bring in the 2655 * entire hpage with 2 PREFETCH_BYTES reads. See 2656 * also, sun4u/cpu/cpu_module.c 2657 */ 2658 /*LINTED*/ 2659 ASSERT(sizeof (page_t) <= 2*PREFETCH_BYTES); 2660 prefetch_page_w((char *)pp); 2661 } 2662 2663 /* 2664 * kphysm_init() tackles the problem of initializing physical memory. 2665 * The old startup made some assumptions about the kernel living in 2666 * physically contiguous space which is no longer valid. 2667 */ 2668 static void 2669 kphysm_init(page_t *pp, struct memseg *memsegp, pgcnt_t npages, 2670 uintptr_t kpm_pp, pgcnt_t kpm_npages) 2671 { 2672 struct memlist *pmem; 2673 struct memseg *msp; 2674 pfn_t base; 2675 pgcnt_t num; 2676 pfn_t lastseg_pages_end = 0; 2677 pgcnt_t nelem_used = 0; 2678 2679 ASSERT(page_hash != NULL && page_hashsz != 0); 2680 2681 msp = memsegp; 2682 for (pmem = phys_avail; pmem && npages; pmem = pmem->next) { 2683 2684 /* 2685 * Build the memsegs entry 2686 */ 2687 num = btop(pmem->size); 2688 if (num > npages) 2689 num = npages; 2690 npages -= num; 2691 base = btop(pmem->address); 2692 2693 msp->pages = pp; 2694 msp->epages = pp + num; 2695 msp->pages_base = base; 2696 msp->pages_end = base + num; 2697 2698 if (kpm_enable) { 2699 pfn_t pbase_a; 2700 pfn_t pend_a; 2701 pfn_t prev_pend_a; 2702 pgcnt_t nelem; 2703 2704 msp->pagespa = va_to_pa(pp); 2705 msp->epagespa = va_to_pa(pp + num); 2706 pbase_a = kpmptop(ptokpmp(base)); 2707 pend_a = kpmptop(ptokpmp(base + num - 1)) + kpmpnpgs; 2708 nelem = ptokpmp(pend_a - pbase_a); 2709 msp->kpm_nkpmpgs = nelem; 2710 msp->kpm_pbase = pbase_a; 2711 if (lastseg_pages_end) { 2712 /* 2713 * Assume phys_avail is in ascending order 2714 * of physical addresses. 2715 */ 2716 ASSERT(base + num > lastseg_pages_end); 2717 prev_pend_a = kpmptop( 2718 ptokpmp(lastseg_pages_end - 1)) + kpmpnpgs; 2719 2720 if (prev_pend_a > pbase_a) { 2721 /* 2722 * Overlap, more than one memseg may 2723 * point to the same kpm_page range. 2724 */ 2725 if (kpm_smallpages == 0) { 2726 msp->kpm_pages = 2727 (kpm_page_t *)kpm_pp - 1; 2728 kpm_pp = (uintptr_t) 2729 ((kpm_page_t *)kpm_pp 2730 + nelem - 1); 2731 } else { 2732 msp->kpm_spages = 2733 (kpm_spage_t *)kpm_pp - 1; 2734 kpm_pp = (uintptr_t) 2735 ((kpm_spage_t *)kpm_pp 2736 + nelem - 1); 2737 } 2738 nelem_used += nelem - 1; 2739 2740 } else { 2741 if (kpm_smallpages == 0) { 2742 msp->kpm_pages = 2743 (kpm_page_t *)kpm_pp; 2744 kpm_pp = (uintptr_t) 2745 ((kpm_page_t *)kpm_pp 2746 + nelem); 2747 } else { 2748 msp->kpm_spages = 2749 (kpm_spage_t *)kpm_pp; 2750 kpm_pp = (uintptr_t) 2751 ((kpm_spage_t *) 2752 kpm_pp + nelem); 2753 } 2754 nelem_used += nelem; 2755 } 2756 2757 } else { 2758 if (kpm_smallpages == 0) { 2759 msp->kpm_pages = (kpm_page_t *)kpm_pp; 2760 kpm_pp = (uintptr_t) 2761 ((kpm_page_t *)kpm_pp + nelem); 2762 } else { 2763 msp->kpm_spages = (kpm_spage_t *)kpm_pp; 2764 kpm_pp = (uintptr_t) 2765 ((kpm_spage_t *)kpm_pp + nelem); 2766 } 2767 nelem_used = nelem; 2768 } 2769 2770 if (nelem_used > kpm_npages) 2771 panic("kphysm_init: kpm_pp overflow\n"); 2772 2773 msp->kpm_pagespa = va_to_pa(msp->kpm_pages); 2774 lastseg_pages_end = msp->pages_end; 2775 } 2776 2777 memseg_list_add(msp); 2778 2779 /* 2780 * add_physmem() initializes the PSM part of the page 2781 * struct by calling the PSM back with add_physmem_cb(). 2782 * In addition it coalesces pages into larger pages as 2783 * it initializes them. 2784 */ 2785 add_physmem(pp, num, base); 2786 pp += num; 2787 msp++; 2788 } 2789 2790 build_pfn_hash(); 2791 } 2792 2793 /* 2794 * Kernel VM initialization. 2795 * Assumptions about kernel address space ordering: 2796 * (1) gap (user space) 2797 * (2) kernel text 2798 * (3) kernel data/bss 2799 * (4) gap 2800 * (5) kernel data structures 2801 * (6) gap 2802 * (7) debugger (optional) 2803 * (8) monitor 2804 * (9) gap (possibly null) 2805 * (10) dvma 2806 * (11) devices 2807 */ 2808 static void 2809 kvm_init(void) 2810 { 2811 /* 2812 * Put the kernel segments in kernel address space. 2813 */ 2814 rw_enter(&kas.a_lock, RW_WRITER); 2815 as_avlinit(&kas); 2816 2817 (void) seg_attach(&kas, (caddr_t)KERNELBASE, 2818 (size_t)(e_moddata - KERNELBASE), &ktextseg); 2819 (void) segkmem_create(&ktextseg); 2820 2821 (void) seg_attach(&kas, (caddr_t)(KERNELBASE + MMU_PAGESIZE4M), 2822 (size_t)(MMU_PAGESIZE4M), &ktexthole); 2823 (void) segkmem_create(&ktexthole); 2824 2825 (void) seg_attach(&kas, (caddr_t)valloc_base, 2826 (size_t)(econtig32 - valloc_base), &kvalloc); 2827 (void) segkmem_create(&kvalloc); 2828 2829 if (kmem64_base) { 2830 (void) seg_attach(&kas, (caddr_t)kmem64_base, 2831 (size_t)(kmem64_end - kmem64_base), &kmem64); 2832 (void) segkmem_create(&kmem64); 2833 } 2834 2835 /* 2836 * We're about to map out /boot. This is the beginning of the 2837 * system resource management transition. We can no longer 2838 * call into /boot for I/O or memory allocations. 2839 */ 2840 (void) seg_attach(&kas, kernelheap, ekernelheap - kernelheap, &kvseg); 2841 (void) segkmem_create(&kvseg); 2842 hblk_alloc_dynamic = 1; 2843 2844 /* 2845 * we need to preallocate pages for DR operations before enabling large 2846 * page kernel heap because of memseg_remap_init() hat_unload() hack. 2847 */ 2848 memseg_remap_init(); 2849 2850 /* at this point we are ready to use large page heap */ 2851 segkmem_heap_lp_init(); 2852 2853 (void) seg_attach(&kas, (caddr_t)SYSBASE32, SYSLIMIT32 - SYSBASE32, 2854 &kvseg32); 2855 (void) segkmem_create(&kvseg32); 2856 2857 /* 2858 * Create a segment for the debugger. 2859 */ 2860 (void) seg_attach(&kas, kdi_segdebugbase, kdi_segdebugsize, &kdebugseg); 2861 (void) segkmem_create(&kdebugseg); 2862 2863 rw_exit(&kas.a_lock); 2864 } 2865 2866 char obp_tte_str[] = 2867 "h# %x constant MMU_PAGESHIFT " 2868 "h# %x constant TTE8K " 2869 "h# %x constant SFHME_SIZE " 2870 "h# %x constant SFHME_TTE " 2871 "h# %x constant HMEBLK_TAG " 2872 "h# %x constant HMEBLK_NEXT " 2873 "h# %x constant HMEBLK_MISC " 2874 "h# %x constant HMEBLK_HME1 " 2875 "h# %x constant NHMENTS " 2876 "h# %x constant HBLK_SZMASK " 2877 "h# %x constant HBLK_RANGE_SHIFT " 2878 "h# %x constant HMEBP_HBLK " 2879 "h# %x constant HMEBUCKET_SIZE " 2880 "h# %x constant HTAG_SFMMUPSZ " 2881 "h# %x constant HTAG_REHASHSZ " 2882 "h# %x constant mmu_hashcnt " 2883 "h# %p constant uhme_hash " 2884 "h# %p constant khme_hash " 2885 "h# %x constant UHMEHASH_SZ " 2886 "h# %x constant KHMEHASH_SZ " 2887 "h# %p constant KCONTEXT " 2888 "h# %p constant KHATID " 2889 "h# %x constant ASI_MEM " 2890 2891 ": PHYS-X@ ( phys -- data ) " 2892 " ASI_MEM spacex@ " 2893 "; " 2894 2895 ": PHYS-W@ ( phys -- data ) " 2896 " ASI_MEM spacew@ " 2897 "; " 2898 2899 ": PHYS-L@ ( phys -- data ) " 2900 " ASI_MEM spaceL@ " 2901 "; " 2902 2903 ": TTE_PAGE_SHIFT ( ttesz -- hmeshift ) " 2904 " 3 * MMU_PAGESHIFT + " 2905 "; " 2906 2907 ": TTE_IS_VALID ( ttep -- flag ) " 2908 " PHYS-X@ 0< " 2909 "; " 2910 2911 ": HME_HASH_SHIFT ( ttesz -- hmeshift ) " 2912 " dup TTE8K = if " 2913 " drop HBLK_RANGE_SHIFT " 2914 " else " 2915 " TTE_PAGE_SHIFT " 2916 " then " 2917 "; " 2918 2919 ": HME_HASH_BSPAGE ( addr hmeshift -- bspage ) " 2920 " tuck >> swap MMU_PAGESHIFT - << " 2921 "; " 2922 2923 ": HME_HASH_FUNCTION ( sfmmup addr hmeshift -- hmebp ) " 2924 " >> over xor swap ( hash sfmmup ) " 2925 " KHATID <> if ( hash ) " 2926 " UHMEHASH_SZ and ( bucket ) " 2927 " HMEBUCKET_SIZE * uhme_hash + ( hmebp ) " 2928 " else ( hash ) " 2929 " KHMEHASH_SZ and ( bucket ) " 2930 " HMEBUCKET_SIZE * khme_hash + ( hmebp ) " 2931 " then ( hmebp ) " 2932 "; " 2933 2934 ": HME_HASH_TABLE_SEARCH " 2935 " ( sfmmup hmebp hblktag -- sfmmup null | sfmmup hmeblkp ) " 2936 " >r hmebp_hblk + phys-x@ begin ( sfmmup hmeblkp ) ( r: hblktag ) " 2937 " dup if ( sfmmup hmeblkp ) ( r: hblktag ) " 2938 " dup hmeblk_tag + phys-x@ r@ = if ( sfmmup hmeblkp ) " 2939 " dup hmeblk_tag + 8 + phys-x@ 2 pick = if " 2940 " true ( sfmmup hmeblkp true ) ( r: hblktag ) " 2941 " else " 2942 " hmeblk_next + phys-x@ false " 2943 " ( sfmmup hmeblkp false ) ( r: hblktag ) " 2944 " then " 2945 " else " 2946 " hmeblk_next + phys-x@ false " 2947 " ( sfmmup hmeblkp false ) ( r: hblktag ) " 2948 " then " 2949 " else " 2950 " true " 2951 " then " 2952 " until r> drop " 2953 "; " 2954 2955 ": HME_HASH_TAG ( sfmmup rehash addr -- hblktag ) " 2956 " over HME_HASH_SHIFT HME_HASH_BSPAGE ( sfmmup rehash bspage ) " 2957 " HTAG_REHASHSZ << or nip ( hblktag ) " 2958 "; " 2959 2960 ": HBLK_TO_TTEP ( hmeblkp addr -- ttep ) " 2961 " over HMEBLK_MISC + PHYS-L@ HBLK_SZMASK and ( hmeblkp addr ttesz ) " 2962 " TTE8K = if ( hmeblkp addr ) " 2963 " MMU_PAGESHIFT >> NHMENTS 1- and ( hmeblkp hme-index ) " 2964 " else ( hmeblkp addr ) " 2965 " drop 0 ( hmeblkp 0 ) " 2966 " then ( hmeblkp hme-index ) " 2967 " SFHME_SIZE * + HMEBLK_HME1 + ( hmep ) " 2968 " SFHME_TTE + ( ttep ) " 2969 "; " 2970 2971 ": unix-tte ( addr cnum -- false | tte-data true ) " 2972 " KCONTEXT = if ( addr ) " 2973 " KHATID ( addr khatid ) " 2974 " else ( addr ) " 2975 " drop false exit ( false ) " 2976 " then " 2977 " ( addr khatid ) " 2978 " mmu_hashcnt 1+ 1 do ( addr sfmmup ) " 2979 " 2dup swap i HME_HASH_SHIFT " 2980 "( addr sfmmup sfmmup addr hmeshift ) " 2981 " HME_HASH_FUNCTION ( addr sfmmup hmebp ) " 2982 " over i 4 pick " 2983 "( addr sfmmup hmebp sfmmup rehash addr ) " 2984 " HME_HASH_TAG ( addr sfmmup hmebp hblktag ) " 2985 " HME_HASH_TABLE_SEARCH " 2986 "( addr sfmmup { null | hmeblkp } ) " 2987 " ?dup if ( addr sfmmup hmeblkp ) " 2988 " nip swap HBLK_TO_TTEP ( ttep ) " 2989 " dup TTE_IS_VALID if ( valid-ttep ) " 2990 " PHYS-X@ true ( tte-data true ) " 2991 " else ( invalid-tte ) " 2992 " drop false ( false ) " 2993 " then ( false | tte-data true ) " 2994 " unloop exit ( false | tte-data true ) " 2995 " then ( addr sfmmup ) " 2996 " loop ( addr sfmmup ) " 2997 " 2drop false ( false ) " 2998 "; " 2999 ; 3000 3001 void 3002 create_va_to_tte(void) 3003 { 3004 char *bp; 3005 extern int khmehash_num, uhmehash_num; 3006 extern struct hmehash_bucket *khme_hash, *uhme_hash; 3007 3008 #define OFFSET(type, field) ((uintptr_t)(&((type *)0)->field)) 3009 3010 bp = (char *)kobj_zalloc(MMU_PAGESIZE, KM_SLEEP); 3011 3012 /* 3013 * Teach obp how to parse our sw ttes. 3014 */ 3015 (void) sprintf(bp, obp_tte_str, 3016 MMU_PAGESHIFT, 3017 TTE8K, 3018 sizeof (struct sf_hment), 3019 OFFSET(struct sf_hment, hme_tte), 3020 OFFSET(struct hme_blk, hblk_tag), 3021 OFFSET(struct hme_blk, hblk_nextpa), 3022 OFFSET(struct hme_blk, hblk_misc), 3023 OFFSET(struct hme_blk, hblk_hme), 3024 NHMENTS, 3025 HBLK_SZMASK, 3026 HBLK_RANGE_SHIFT, 3027 OFFSET(struct hmehash_bucket, hmeh_nextpa), 3028 sizeof (struct hmehash_bucket), 3029 HTAG_SFMMUPSZ, 3030 HTAG_REHASHSZ, 3031 mmu_hashcnt, 3032 (caddr_t)va_to_pa((caddr_t)uhme_hash), 3033 (caddr_t)va_to_pa((caddr_t)khme_hash), 3034 UHMEHASH_SZ, 3035 KHMEHASH_SZ, 3036 KCONTEXT, 3037 KHATID, 3038 ASI_MEM); 3039 prom_interpret(bp, 0, 0, 0, 0, 0); 3040 3041 kobj_free(bp, MMU_PAGESIZE); 3042 } 3043 3044 void 3045 install_va_to_tte(void) 3046 { 3047 /* 3048 * advise prom that he can use unix-tte 3049 */ 3050 prom_interpret("' unix-tte is va>tte-data", 0, 0, 0, 0, 0); 3051 } 3052 3053 /* 3054 * Here we add "device-type=console" for /os-io node, for currently 3055 * our kernel console output only supports displaying text and 3056 * performing cursor-positioning operations (through kernel framebuffer 3057 * driver) and it doesn't support other functionalities required for a 3058 * standard "display" device as specified in 1275 spec. The main missing 3059 * interface defined by the 1275 spec is "draw-logo". 3060 * also see the comments above prom_stdout_is_framebuffer(). 3061 */ 3062 static char *create_node = 3063 "\" /\" find-device " 3064 "new-device " 3065 "\" os-io\" device-name " 3066 "\" "OBP_DISPLAY_CONSOLE"\" device-type " 3067 ": cb-r/w ( adr,len method$ -- #read/#written ) " 3068 " 2>r swap 2 2r> ['] $callback catch if " 3069 " 2drop 3drop 0 " 3070 " then " 3071 "; " 3072 ": read ( adr,len -- #read ) " 3073 " \" read\" ['] cb-r/w catch if 2drop 2drop -2 exit then " 3074 " ( retN ... ret1 N ) " 3075 " ?dup if " 3076 " swap >r 1- 0 ?do drop loop r> " 3077 " else " 3078 " -2 " 3079 " then " 3080 "; " 3081 ": write ( adr,len -- #written ) " 3082 " \" write\" ['] cb-r/w catch if 2drop 2drop 0 exit then " 3083 " ( retN ... ret1 N ) " 3084 " ?dup if " 3085 " swap >r 1- 0 ?do drop loop r> " 3086 " else " 3087 " 0 " 3088 " then " 3089 "; " 3090 ": poll-tty ( -- ) ; " 3091 ": install-abort ( -- ) ['] poll-tty d# 10 alarm ; " 3092 ": remove-abort ( -- ) ['] poll-tty 0 alarm ; " 3093 ": cb-give/take ( $method -- ) " 3094 " 0 -rot ['] $callback catch ?dup if " 3095 " >r 2drop 2drop r> throw " 3096 " else " 3097 " 0 ?do drop loop " 3098 " then " 3099 "; " 3100 ": give ( -- ) \" exit-input\" cb-give/take ; " 3101 ": take ( -- ) \" enter-input\" cb-give/take ; " 3102 ": open ( -- ok? ) true ; " 3103 ": close ( -- ) ; " 3104 "finish-device " 3105 "device-end "; 3106 3107 /* 3108 * Create the OBP input/output node (FCode serial driver). 3109 * It is needed for both USB console keyboard and for 3110 * the kernel terminal emulator. It is too early to check for a 3111 * kernel console compatible framebuffer now, so we create this 3112 * so that we're ready if we need to enable kernel terminal emulation. 3113 * 3114 * When the USB software takes over the input device at the time 3115 * consconfig runs, OBP's stdin is redirected to this node. 3116 * Whenever the FORTH user interface is used after this switch, 3117 * the node will call back into the kernel for console input. 3118 * If a serial device such as ttya or a UART with a Type 5 keyboard 3119 * attached is used, OBP takes over the serial device when the system 3120 * goes to the debugger after the system is booted. This sharing 3121 * of the relatively simple serial device is difficult but possible. 3122 * Sharing the USB host controller is impossible due its complexity. 3123 * 3124 * Similarly to USB keyboard input redirection, after consconfig_dacf 3125 * configures a kernel console framebuffer as the standard output 3126 * device, OBP's stdout is switched to to vector through the 3127 * /os-io node into the kernel terminal emulator. 3128 */ 3129 static void 3130 startup_create_io_node(void) 3131 { 3132 prom_interpret(create_node, 0, 0, 0, 0, 0); 3133 } 3134 3135 3136 static void 3137 do_prom_version_check(void) 3138 { 3139 int i; 3140 pnode_t node; 3141 char buf[64]; 3142 static char drev[] = "Down-rev firmware detected%s\n" 3143 "\tPlease upgrade to the following minimum version:\n" 3144 "\t\t%s\n"; 3145 3146 i = prom_version_check(buf, sizeof (buf), &node); 3147 3148 if (i == PROM_VER64_OK) 3149 return; 3150 3151 if (i == PROM_VER64_UPGRADE) { 3152 cmn_err(CE_WARN, drev, "", buf); 3153 3154 #ifdef DEBUG 3155 prom_enter_mon(); /* Type 'go' to continue */ 3156 cmn_err(CE_WARN, "Booting with down-rev firmware\n"); 3157 return; 3158 #else 3159 halt(0); 3160 #endif 3161 } 3162 3163 /* 3164 * The other possibility is that this is a server running 3165 * good firmware, but down-rev firmware was detected on at 3166 * least one other cpu board. We just complain if we see 3167 * that. 3168 */ 3169 cmn_err(CE_WARN, drev, " on one or more CPU boards", buf); 3170 } 3171 3172 static void 3173 kpm_init() 3174 { 3175 kpm_pgshft = (kpm_smallpages == 0) ? MMU_PAGESHIFT4M : MMU_PAGESHIFT; 3176 kpm_pgsz = 1ull << kpm_pgshft; 3177 kpm_pgoff = kpm_pgsz - 1; 3178 kpmp2pshft = kpm_pgshft - PAGESHIFT; 3179 kpmpnpgs = 1 << kpmp2pshft; 3180 ASSERT(((uintptr_t)kpm_vbase & (kpm_pgsz - 1)) == 0); 3181 } 3182 3183 void 3184 kpm_npages_setup(int memblocks) 3185 { 3186 /* 3187 * npages can be scattered in a maximum of 'memblocks' 3188 */ 3189 kpm_npages = ptokpmpr(npages) + memblocks; 3190 } 3191 3192 /* 3193 * Must be defined in platform dependent code. 3194 */ 3195 extern caddr_t modtext; 3196 extern size_t modtext_sz; 3197 extern caddr_t moddata; 3198 3199 #define HEAPTEXT_ARENA(addr) \ 3200 ((uintptr_t)(addr) < KERNELBASE + 2 * MMU_PAGESIZE4M ? 0 : \ 3201 (((uintptr_t)(addr) - HEAPTEXT_BASE) / \ 3202 (HEAPTEXT_MAPPED + HEAPTEXT_UNMAPPED) + 1)) 3203 3204 #define HEAPTEXT_OVERSIZED(addr) \ 3205 ((uintptr_t)(addr) >= HEAPTEXT_BASE + HEAPTEXT_SIZE - HEAPTEXT_OVERSIZE) 3206 3207 vmem_t *texthole_source[HEAPTEXT_NARENAS]; 3208 vmem_t *texthole_arena[HEAPTEXT_NARENAS]; 3209 kmutex_t texthole_lock; 3210 3211 char kern_bootargs[OBP_MAXPATHLEN]; 3212 3213 void 3214 kobj_vmem_init(vmem_t **text_arena, vmem_t **data_arena) 3215 { 3216 uintptr_t addr, limit; 3217 3218 addr = HEAPTEXT_BASE; 3219 limit = addr + HEAPTEXT_SIZE - HEAPTEXT_OVERSIZE; 3220 3221 /* 3222 * Before we initialize the text_arena, we want to punch holes in the 3223 * underlying heaptext_arena. This guarantees that for any text 3224 * address we can find a text hole less than HEAPTEXT_MAPPED away. 3225 */ 3226 for (; addr + HEAPTEXT_UNMAPPED <= limit; 3227 addr += HEAPTEXT_MAPPED + HEAPTEXT_UNMAPPED) { 3228 (void) vmem_xalloc(heaptext_arena, HEAPTEXT_UNMAPPED, PAGESIZE, 3229 0, 0, (void *)addr, (void *)(addr + HEAPTEXT_UNMAPPED), 3230 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 3231 } 3232 3233 /* 3234 * Allocate one page at the oversize to break up the text region 3235 * from the oversized region. 3236 */ 3237 (void) vmem_xalloc(heaptext_arena, PAGESIZE, PAGESIZE, 0, 0, 3238 (void *)limit, (void *)(limit + PAGESIZE), 3239 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 3240 3241 *text_arena = vmem_create("module_text", modtext_sz ? modtext : NULL, 3242 modtext_sz, sizeof (uintptr_t), segkmem_alloc, segkmem_free, 3243 heaptext_arena, 0, VM_SLEEP); 3244 *data_arena = vmem_create("module_data", moddata, MODDATA, 1, 3245 segkmem_alloc, segkmem_free, heap32_arena, 0, VM_SLEEP); 3246 } 3247 3248 caddr_t 3249 kobj_text_alloc(vmem_t *arena, size_t size) 3250 { 3251 caddr_t rval, better; 3252 3253 /* 3254 * First, try a sleeping allocation. 3255 */ 3256 rval = vmem_alloc(arena, size, VM_SLEEP | VM_BESTFIT); 3257 3258 if (size >= HEAPTEXT_MAPPED || !HEAPTEXT_OVERSIZED(rval)) 3259 return (rval); 3260 3261 /* 3262 * We didn't get the area that we wanted. We're going to try to do an 3263 * allocation with explicit constraints. 3264 */ 3265 better = vmem_xalloc(arena, size, sizeof (uintptr_t), 0, 0, NULL, 3266 (void *)(HEAPTEXT_BASE + HEAPTEXT_SIZE - HEAPTEXT_OVERSIZE), 3267 VM_NOSLEEP | VM_BESTFIT); 3268 3269 if (better != NULL) { 3270 /* 3271 * That worked. Free our first attempt and return. 3272 */ 3273 vmem_free(arena, rval, size); 3274 return (better); 3275 } 3276 3277 /* 3278 * That didn't work; we'll have to return our first attempt. 3279 */ 3280 return (rval); 3281 } 3282 3283 caddr_t 3284 kobj_texthole_alloc(caddr_t addr, size_t size) 3285 { 3286 int arena = HEAPTEXT_ARENA(addr); 3287 char c[30]; 3288 uintptr_t base; 3289 3290 if (HEAPTEXT_OVERSIZED(addr)) { 3291 /* 3292 * If this is an oversized allocation, there is no text hole 3293 * available for it; return NULL. 3294 */ 3295 return (NULL); 3296 } 3297 3298 mutex_enter(&texthole_lock); 3299 3300 if (texthole_arena[arena] == NULL) { 3301 ASSERT(texthole_source[arena] == NULL); 3302 3303 if (arena == 0) { 3304 texthole_source[0] = vmem_create("module_text_holesrc", 3305 (void *)(KERNELBASE + MMU_PAGESIZE4M), 3306 MMU_PAGESIZE4M, PAGESIZE, NULL, NULL, NULL, 3307 0, VM_SLEEP); 3308 } else { 3309 base = HEAPTEXT_BASE + 3310 (arena - 1) * (HEAPTEXT_MAPPED + HEAPTEXT_UNMAPPED); 3311 3312 (void) snprintf(c, sizeof (c), 3313 "heaptext_holesrc_%d", arena); 3314 3315 texthole_source[arena] = vmem_create(c, (void *)base, 3316 HEAPTEXT_UNMAPPED, PAGESIZE, NULL, NULL, NULL, 3317 0, VM_SLEEP); 3318 } 3319 3320 (void) snprintf(c, sizeof (c), "heaptext_hole_%d", arena); 3321 3322 texthole_arena[arena] = vmem_create(c, NULL, 0, 3323 sizeof (uint32_t), segkmem_alloc_permanent, segkmem_free, 3324 texthole_source[arena], 0, VM_SLEEP); 3325 } 3326 3327 mutex_exit(&texthole_lock); 3328 3329 ASSERT(texthole_arena[arena] != NULL); 3330 ASSERT(arena >= 0 && arena < HEAPTEXT_NARENAS); 3331 return (vmem_alloc(texthole_arena[arena], size, 3332 VM_BESTFIT | VM_NOSLEEP)); 3333 } 3334 3335 void 3336 kobj_texthole_free(caddr_t addr, size_t size) 3337 { 3338 int arena = HEAPTEXT_ARENA(addr); 3339 3340 ASSERT(arena >= 0 && arena < HEAPTEXT_NARENAS); 3341 ASSERT(texthole_arena[arena] != NULL); 3342 vmem_free(texthole_arena[arena], addr, size); 3343 } 3344